linux/drivers/nvdimm/btt.c

1748 lines
43 KiB
C
Raw Normal View History

// SPDX-License-Identifier: GPL-2.0-only
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
/*
* Block Translation Table
* Copyright (c) 2014-2015, Intel Corporation.
*/
#include <linux/highmem.h>
#include <linux/debugfs.h>
#include <linux/blkdev.h>
#include <linux/module.h>
#include <linux/device.h>
#include <linux/mutex.h>
#include <linux/hdreg.h>
#include <linux/genhd.h>
#include <linux/sizes.h>
#include <linux/ndctl.h>
#include <linux/fs.h>
#include <linux/nd.h>
#include <linux/backing-dev.h>
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
#include "btt.h"
#include "nd.h"
enum log_ent_request {
LOG_NEW_ENT = 0,
LOG_OLD_ENT
};
static struct device *to_dev(struct arena_info *arena)
{
return &arena->nd_btt->dev;
}
libnvdimm, btt: rework error clearing Clearing errors or badblocks during a BTT write requires sending an ACPI DSM, which means potentially sleeping. Since a BTT IO happens in atomic context (preemption disabled, spinlocks may be held), we cannot perform error clearing in the course of an IO. Due to this error clearing for BTT IOs has hitherto been disabled. In this patch we move error clearing out of the atomic section, and thus re-enable error clearing with BTTs. When we are about to add a block to the free list, we check if it was previously marked as an error, and if it was, we add it to the freelist, but also set a flag that says error clearing will be required. We then drop the lane (ending the atomic context), and send a zero buffer so that the error can be cleared. The error flag in the free list is protected by the nd 'lane', and is set only be a thread while it holds that lane. When the error is cleared, the flag is cleared, but while holding a mutex for that freelist index. When writing, we check for two things - 1/ If the freelist mutex is held or if the error flag is set. If so, this is an error block that is being (or about to be) cleared. 2/ If the block is a known badblock based on nsio->bb The second check is required because the BTT map error flag for a map entry only gets set when an error LBA is read. If we write to a new location that may not have the map error flag set, but still might be in the region's badblock list, we can trigger an EIO on the write, which is undesirable and completely avoidable. Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Toshi Kani <toshi.kani@hpe.com> Cc: Dan Williams <dan.j.williams@intel.com> Signed-off-by: Vishal Verma <vishal.l.verma@intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2017-08-31 09:36:03 +08:00
static u64 adjust_initial_offset(struct nd_btt *nd_btt, u64 offset)
{
return offset + nd_btt->initial_offset;
}
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
static int arena_read_bytes(struct arena_info *arena, resource_size_t offset,
void *buf, size_t n, unsigned long flags)
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
{
struct nd_btt *nd_btt = arena->nd_btt;
struct nd_namespace_common *ndns = nd_btt->ndns;
/* arena offsets may be shifted from the base of the device */
libnvdimm, btt: rework error clearing Clearing errors or badblocks during a BTT write requires sending an ACPI DSM, which means potentially sleeping. Since a BTT IO happens in atomic context (preemption disabled, spinlocks may be held), we cannot perform error clearing in the course of an IO. Due to this error clearing for BTT IOs has hitherto been disabled. In this patch we move error clearing out of the atomic section, and thus re-enable error clearing with BTTs. When we are about to add a block to the free list, we check if it was previously marked as an error, and if it was, we add it to the freelist, but also set a flag that says error clearing will be required. We then drop the lane (ending the atomic context), and send a zero buffer so that the error can be cleared. The error flag in the free list is protected by the nd 'lane', and is set only be a thread while it holds that lane. When the error is cleared, the flag is cleared, but while holding a mutex for that freelist index. When writing, we check for two things - 1/ If the freelist mutex is held or if the error flag is set. If so, this is an error block that is being (or about to be) cleared. 2/ If the block is a known badblock based on nsio->bb The second check is required because the BTT map error flag for a map entry only gets set when an error LBA is read. If we write to a new location that may not have the map error flag set, but still might be in the region's badblock list, we can trigger an EIO on the write, which is undesirable and completely avoidable. Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Toshi Kani <toshi.kani@hpe.com> Cc: Dan Williams <dan.j.williams@intel.com> Signed-off-by: Vishal Verma <vishal.l.verma@intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2017-08-31 09:36:03 +08:00
offset = adjust_initial_offset(nd_btt, offset);
return nvdimm_read_bytes(ndns, offset, buf, n, flags);
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
}
static int arena_write_bytes(struct arena_info *arena, resource_size_t offset,
void *buf, size_t n, unsigned long flags)
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
{
struct nd_btt *nd_btt = arena->nd_btt;
struct nd_namespace_common *ndns = nd_btt->ndns;
/* arena offsets may be shifted from the base of the device */
libnvdimm, btt: rework error clearing Clearing errors or badblocks during a BTT write requires sending an ACPI DSM, which means potentially sleeping. Since a BTT IO happens in atomic context (preemption disabled, spinlocks may be held), we cannot perform error clearing in the course of an IO. Due to this error clearing for BTT IOs has hitherto been disabled. In this patch we move error clearing out of the atomic section, and thus re-enable error clearing with BTTs. When we are about to add a block to the free list, we check if it was previously marked as an error, and if it was, we add it to the freelist, but also set a flag that says error clearing will be required. We then drop the lane (ending the atomic context), and send a zero buffer so that the error can be cleared. The error flag in the free list is protected by the nd 'lane', and is set only be a thread while it holds that lane. When the error is cleared, the flag is cleared, but while holding a mutex for that freelist index. When writing, we check for two things - 1/ If the freelist mutex is held or if the error flag is set. If so, this is an error block that is being (or about to be) cleared. 2/ If the block is a known badblock based on nsio->bb The second check is required because the BTT map error flag for a map entry only gets set when an error LBA is read. If we write to a new location that may not have the map error flag set, but still might be in the region's badblock list, we can trigger an EIO on the write, which is undesirable and completely avoidable. Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Toshi Kani <toshi.kani@hpe.com> Cc: Dan Williams <dan.j.williams@intel.com> Signed-off-by: Vishal Verma <vishal.l.verma@intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2017-08-31 09:36:03 +08:00
offset = adjust_initial_offset(nd_btt, offset);
return nvdimm_write_bytes(ndns, offset, buf, n, flags);
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
}
static int btt_info_write(struct arena_info *arena, struct btt_sb *super)
{
int ret;
/*
* infooff and info2off should always be at least 512B aligned.
* We rely on that to make sure rw_bytes does error clearing
* correctly, so make sure that is the case.
*/
dev_WARN_ONCE(to_dev(arena), !IS_ALIGNED(arena->infooff, 512),
"arena->infooff: %#llx is unaligned\n", arena->infooff);
dev_WARN_ONCE(to_dev(arena), !IS_ALIGNED(arena->info2off, 512),
"arena->info2off: %#llx is unaligned\n", arena->info2off);
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
ret = arena_write_bytes(arena, arena->info2off, super,
sizeof(struct btt_sb), 0);
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
if (ret)
return ret;
return arena_write_bytes(arena, arena->infooff, super,
sizeof(struct btt_sb), 0);
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
}
static int btt_info_read(struct arena_info *arena, struct btt_sb *super)
{
return arena_read_bytes(arena, arena->infooff, super,
sizeof(struct btt_sb), 0);
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
}
/*
* 'raw' version of btt_map write
* Assumptions:
* mapping is in little-endian
* mapping contains 'E' and 'Z' flags as desired
*/
static int __btt_map_write(struct arena_info *arena, u32 lba, __le32 mapping,
unsigned long flags)
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
{
u64 ns_off = arena->mapoff + (lba * MAP_ENT_SIZE);
if (unlikely(lba >= arena->external_nlba))
dev_err_ratelimited(to_dev(arena),
"%s: lba %#x out of range (max: %#x)\n",
__func__, lba, arena->external_nlba);
return arena_write_bytes(arena, ns_off, &mapping, MAP_ENT_SIZE, flags);
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
}
static int btt_map_write(struct arena_info *arena, u32 lba, u32 mapping,
u32 z_flag, u32 e_flag, unsigned long rwb_flags)
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
{
u32 ze;
__le32 mapping_le;
/*
* This 'mapping' is supposed to be just the LBA mapping, without
* any flags set, so strip the flag bits.
*/
mapping = ent_lba(mapping);
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
ze = (z_flag << 1) + e_flag;
switch (ze) {
case 0:
/*
* We want to set neither of the Z or E flags, and
* in the actual layout, this means setting the bit
* positions of both to '1' to indicate a 'normal'
* map entry
*/
mapping |= MAP_ENT_NORMAL;
break;
case 1:
mapping |= (1 << MAP_ERR_SHIFT);
break;
case 2:
mapping |= (1 << MAP_TRIM_SHIFT);
break;
default:
/*
* The case where Z and E are both sent in as '1' could be
* construed as a valid 'normal' case, but we decide not to,
* to avoid confusion
*/
dev_err_ratelimited(to_dev(arena),
"Invalid use of Z and E flags\n");
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
return -EIO;
}
mapping_le = cpu_to_le32(mapping);
return __btt_map_write(arena, lba, mapping_le, rwb_flags);
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
}
static int btt_map_read(struct arena_info *arena, u32 lba, u32 *mapping,
int *trim, int *error, unsigned long rwb_flags)
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
{
int ret;
__le32 in;
u32 raw_mapping, postmap, ze, z_flag, e_flag;
u64 ns_off = arena->mapoff + (lba * MAP_ENT_SIZE);
if (unlikely(lba >= arena->external_nlba))
dev_err_ratelimited(to_dev(arena),
"%s: lba %#x out of range (max: %#x)\n",
__func__, lba, arena->external_nlba);
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
ret = arena_read_bytes(arena, ns_off, &in, MAP_ENT_SIZE, rwb_flags);
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
if (ret)
return ret;
raw_mapping = le32_to_cpu(in);
z_flag = ent_z_flag(raw_mapping);
e_flag = ent_e_flag(raw_mapping);
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
ze = (z_flag << 1) + e_flag;
postmap = ent_lba(raw_mapping);
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
/* Reuse the {z,e}_flag variables for *trim and *error */
z_flag = 0;
e_flag = 0;
switch (ze) {
case 0:
/* Initial state. Return postmap = premap */
*mapping = lba;
break;
case 1:
*mapping = postmap;
e_flag = 1;
break;
case 2:
*mapping = postmap;
z_flag = 1;
break;
case 3:
*mapping = postmap;
break;
default:
return -EIO;
}
if (trim)
*trim = z_flag;
if (error)
*error = e_flag;
return ret;
}
static int btt_log_group_read(struct arena_info *arena, u32 lane,
struct log_group *log)
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
{
return arena_read_bytes(arena,
arena->logoff + (lane * LOG_GRP_SIZE), log,
LOG_GRP_SIZE, 0);
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
}
static struct dentry *debugfs_root;
static void arena_debugfs_init(struct arena_info *a, struct dentry *parent,
int idx)
{
char dirname[32];
struct dentry *d;
/* If for some reason, parent bttN was not created, exit */
if (!parent)
return;
snprintf(dirname, 32, "arena%d", idx);
d = debugfs_create_dir(dirname, parent);
if (IS_ERR_OR_NULL(d))
return;
a->debugfs_dir = d;
debugfs_create_x64("size", S_IRUGO, d, &a->size);
debugfs_create_x64("external_lba_start", S_IRUGO, d,
&a->external_lba_start);
debugfs_create_x32("internal_nlba", S_IRUGO, d, &a->internal_nlba);
debugfs_create_u32("internal_lbasize", S_IRUGO, d,
&a->internal_lbasize);
debugfs_create_x32("external_nlba", S_IRUGO, d, &a->external_nlba);
debugfs_create_u32("external_lbasize", S_IRUGO, d,
&a->external_lbasize);
debugfs_create_u32("nfree", S_IRUGO, d, &a->nfree);
debugfs_create_u16("version_major", S_IRUGO, d, &a->version_major);
debugfs_create_u16("version_minor", S_IRUGO, d, &a->version_minor);
debugfs_create_x64("nextoff", S_IRUGO, d, &a->nextoff);
debugfs_create_x64("infooff", S_IRUGO, d, &a->infooff);
debugfs_create_x64("dataoff", S_IRUGO, d, &a->dataoff);
debugfs_create_x64("mapoff", S_IRUGO, d, &a->mapoff);
debugfs_create_x64("logoff", S_IRUGO, d, &a->logoff);
debugfs_create_x64("info2off", S_IRUGO, d, &a->info2off);
debugfs_create_x32("flags", S_IRUGO, d, &a->flags);
debugfs_create_u32("log_index_0", S_IRUGO, d, &a->log_index[0]);
debugfs_create_u32("log_index_1", S_IRUGO, d, &a->log_index[1]);
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
}
static void btt_debugfs_init(struct btt *btt)
{
int i = 0;
struct arena_info *arena;
btt->debugfs_dir = debugfs_create_dir(dev_name(&btt->nd_btt->dev),
debugfs_root);
if (IS_ERR_OR_NULL(btt->debugfs_dir))
return;
list_for_each_entry(arena, &btt->arena_list, list) {
arena_debugfs_init(arena, btt->debugfs_dir, i);
i++;
}
}
static u32 log_seq(struct log_group *log, int log_idx)
{
return le32_to_cpu(log->ent[log_idx].seq);
}
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
/*
* This function accepts two log entries, and uses the
* sequence number to find the 'older' entry.
* It also updates the sequence number in this old entry to
* make it the 'new' one if the mark_flag is set.
* Finally, it returns which of the entries was the older one.
*
* TODO The logic feels a bit kludge-y. make it better..
*/
static int btt_log_get_old(struct arena_info *a, struct log_group *log)
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
{
int idx0 = a->log_index[0];
int idx1 = a->log_index[1];
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
int old;
/*
* the first ever time this is seen, the entry goes into [0]
* the next time, the following logic works out to put this
* (next) entry into [1]
*/
if (log_seq(log, idx0) == 0) {
log->ent[idx0].seq = cpu_to_le32(1);
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
return 0;
}
if (log_seq(log, idx0) == log_seq(log, idx1))
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
return -EINVAL;
if (log_seq(log, idx0) + log_seq(log, idx1) > 5)
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
return -EINVAL;
if (log_seq(log, idx0) < log_seq(log, idx1)) {
if ((log_seq(log, idx1) - log_seq(log, idx0)) == 1)
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
old = 0;
else
old = 1;
} else {
if ((log_seq(log, idx0) - log_seq(log, idx1)) == 1)
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
old = 1;
else
old = 0;
}
return old;
}
/*
* This function copies the desired (old/new) log entry into ent if
* it is not NULL. It returns the sub-slot number (0 or 1)
* where the desired log entry was found. Negative return values
* indicate errors.
*/
static int btt_log_read(struct arena_info *arena, u32 lane,
struct log_entry *ent, int old_flag)
{
int ret;
int old_ent, ret_ent;
struct log_group log;
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
ret = btt_log_group_read(arena, lane, &log);
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
if (ret)
return -EIO;
old_ent = btt_log_get_old(arena, &log);
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
if (old_ent < 0 || old_ent > 1) {
dev_err(to_dev(arena),
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
"log corruption (%d): lane %d seq [%d, %d]\n",
old_ent, lane, log.ent[arena->log_index[0]].seq,
log.ent[arena->log_index[1]].seq);
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
/* TODO set error state? */
return -EIO;
}
ret_ent = (old_flag ? old_ent : (1 - old_ent));
if (ent != NULL)
memcpy(ent, &log.ent[arena->log_index[ret_ent]], LOG_ENT_SIZE);
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
return ret_ent;
}
/*
* This function commits a log entry to media
* It does _not_ prepare the freelist entry for the next write
* btt_flog_write is the wrapper for updating the freelist elements
*/
static int __btt_log_write(struct arena_info *arena, u32 lane,
u32 sub, struct log_entry *ent, unsigned long flags)
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
{
int ret;
u32 group_slot = arena->log_index[sub];
unsigned int log_half = LOG_ENT_SIZE / 2;
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
void *src = ent;
u64 ns_off;
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
ns_off = arena->logoff + (lane * LOG_GRP_SIZE) +
(group_slot * LOG_ENT_SIZE);
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
/* split the 16B write into atomic, durable halves */
ret = arena_write_bytes(arena, ns_off, src, log_half, flags);
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
if (ret)
return ret;
ns_off += log_half;
src += log_half;
return arena_write_bytes(arena, ns_off, src, log_half, flags);
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
}
static int btt_flog_write(struct arena_info *arena, u32 lane, u32 sub,
struct log_entry *ent)
{
int ret;
ret = __btt_log_write(arena, lane, sub, ent, NVDIMM_IO_ATOMIC);
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
if (ret)
return ret;
/* prepare the next free entry */
arena->freelist[lane].sub = 1 - arena->freelist[lane].sub;
if (++(arena->freelist[lane].seq) == 4)
arena->freelist[lane].seq = 1;
if (ent_e_flag(le32_to_cpu(ent->old_map)))
libnvdimm, btt: rework error clearing Clearing errors or badblocks during a BTT write requires sending an ACPI DSM, which means potentially sleeping. Since a BTT IO happens in atomic context (preemption disabled, spinlocks may be held), we cannot perform error clearing in the course of an IO. Due to this error clearing for BTT IOs has hitherto been disabled. In this patch we move error clearing out of the atomic section, and thus re-enable error clearing with BTTs. When we are about to add a block to the free list, we check if it was previously marked as an error, and if it was, we add it to the freelist, but also set a flag that says error clearing will be required. We then drop the lane (ending the atomic context), and send a zero buffer so that the error can be cleared. The error flag in the free list is protected by the nd 'lane', and is set only be a thread while it holds that lane. When the error is cleared, the flag is cleared, but while holding a mutex for that freelist index. When writing, we check for two things - 1/ If the freelist mutex is held or if the error flag is set. If so, this is an error block that is being (or about to be) cleared. 2/ If the block is a known badblock based on nsio->bb The second check is required because the BTT map error flag for a map entry only gets set when an error LBA is read. If we write to a new location that may not have the map error flag set, but still might be in the region's badblock list, we can trigger an EIO on the write, which is undesirable and completely avoidable. Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Toshi Kani <toshi.kani@hpe.com> Cc: Dan Williams <dan.j.williams@intel.com> Signed-off-by: Vishal Verma <vishal.l.verma@intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2017-08-31 09:36:03 +08:00
arena->freelist[lane].has_err = 1;
arena->freelist[lane].block = ent_lba(le32_to_cpu(ent->old_map));
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
return ret;
}
/*
* This function initializes the BTT map to the initial state, which is
* all-zeroes, and indicates an identity mapping
*/
static int btt_map_init(struct arena_info *arena)
{
int ret = -EINVAL;
void *zerobuf;
size_t offset = 0;
size_t chunk_size = SZ_2M;
size_t mapsize = arena->logoff - arena->mapoff;
zerobuf = kzalloc(chunk_size, GFP_KERNEL);
if (!zerobuf)
return -ENOMEM;
/*
* mapoff should always be at least 512B aligned. We rely on that to
* make sure rw_bytes does error clearing correctly, so make sure that
* is the case.
*/
dev_WARN_ONCE(to_dev(arena), !IS_ALIGNED(arena->mapoff, 512),
"arena->mapoff: %#llx is unaligned\n", arena->mapoff);
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
while (mapsize) {
size_t size = min(mapsize, chunk_size);
dev_WARN_ONCE(to_dev(arena), size < 512,
"chunk size: %#zx is unaligned\n", size);
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
ret = arena_write_bytes(arena, arena->mapoff + offset, zerobuf,
size, 0);
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
if (ret)
goto free;
offset += size;
mapsize -= size;
cond_resched();
}
free:
kfree(zerobuf);
return ret;
}
/*
* This function initializes the BTT log with 'fake' entries pointing
* to the initial reserved set of blocks as being free
*/
static int btt_log_init(struct arena_info *arena)
{
size_t logsize = arena->info2off - arena->logoff;
size_t chunk_size = SZ_4K, offset = 0;
struct log_entry ent;
void *zerobuf;
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
int ret;
u32 i;
zerobuf = kzalloc(chunk_size, GFP_KERNEL);
if (!zerobuf)
return -ENOMEM;
/*
* logoff should always be at least 512B aligned. We rely on that to
* make sure rw_bytes does error clearing correctly, so make sure that
* is the case.
*/
dev_WARN_ONCE(to_dev(arena), !IS_ALIGNED(arena->logoff, 512),
"arena->logoff: %#llx is unaligned\n", arena->logoff);
while (logsize) {
size_t size = min(logsize, chunk_size);
dev_WARN_ONCE(to_dev(arena), size < 512,
"chunk size: %#zx is unaligned\n", size);
ret = arena_write_bytes(arena, arena->logoff + offset, zerobuf,
size, 0);
if (ret)
goto free;
offset += size;
logsize -= size;
cond_resched();
}
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
for (i = 0; i < arena->nfree; i++) {
ent.lba = cpu_to_le32(i);
ent.old_map = cpu_to_le32(arena->external_nlba + i);
ent.new_map = cpu_to_le32(arena->external_nlba + i);
ent.seq = cpu_to_le32(LOG_SEQ_INIT);
ret = __btt_log_write(arena, i, 0, &ent, 0);
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
if (ret)
goto free;
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
}
free:
kfree(zerobuf);
return ret;
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
}
libnvdimm, btt: rework error clearing Clearing errors or badblocks during a BTT write requires sending an ACPI DSM, which means potentially sleeping. Since a BTT IO happens in atomic context (preemption disabled, spinlocks may be held), we cannot perform error clearing in the course of an IO. Due to this error clearing for BTT IOs has hitherto been disabled. In this patch we move error clearing out of the atomic section, and thus re-enable error clearing with BTTs. When we are about to add a block to the free list, we check if it was previously marked as an error, and if it was, we add it to the freelist, but also set a flag that says error clearing will be required. We then drop the lane (ending the atomic context), and send a zero buffer so that the error can be cleared. The error flag in the free list is protected by the nd 'lane', and is set only be a thread while it holds that lane. When the error is cleared, the flag is cleared, but while holding a mutex for that freelist index. When writing, we check for two things - 1/ If the freelist mutex is held or if the error flag is set. If so, this is an error block that is being (or about to be) cleared. 2/ If the block is a known badblock based on nsio->bb The second check is required because the BTT map error flag for a map entry only gets set when an error LBA is read. If we write to a new location that may not have the map error flag set, but still might be in the region's badblock list, we can trigger an EIO on the write, which is undesirable and completely avoidable. Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Toshi Kani <toshi.kani@hpe.com> Cc: Dan Williams <dan.j.williams@intel.com> Signed-off-by: Vishal Verma <vishal.l.verma@intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2017-08-31 09:36:03 +08:00
static u64 to_namespace_offset(struct arena_info *arena, u64 lba)
{
return arena->dataoff + ((u64)lba * arena->internal_lbasize);
}
static int arena_clear_freelist_error(struct arena_info *arena, u32 lane)
{
int ret = 0;
if (arena->freelist[lane].has_err) {
void *zero_page = page_address(ZERO_PAGE(0));
u32 lba = arena->freelist[lane].block;
u64 nsoff = to_namespace_offset(arena, lba);
unsigned long len = arena->sector_size;
mutex_lock(&arena->err_lock);
while (len) {
unsigned long chunk = min(len, PAGE_SIZE);
ret = arena_write_bytes(arena, nsoff, zero_page,
chunk, 0);
if (ret)
break;
len -= chunk;
nsoff += chunk;
if (len == 0)
arena->freelist[lane].has_err = 0;
}
mutex_unlock(&arena->err_lock);
}
return ret;
}
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
static int btt_freelist_init(struct arena_info *arena)
{
int new, ret;
struct log_entry log_new;
libnvdimm/btt: Fix LBA masking during 'free list' population The Linux BTT implementation assumes that log entries will never have the 'zero' flag set, and indeed it never sets that flag for log entries itself. However, the UEFI spec is ambiguous on the exact format of the LBA field of a log entry, specifically as to whether it should include the additional flag bits or not. While a zero bit doesn't make sense in the context of a log entry, other BTT implementations might still have it set. If an implementation does happen to have it set, we would happily read it in as the next block to write to for writes. Since a high bit is set, it pushes the block number out of the range of an 'arena', and we fail such a write with an EIO. Follow the robustness principle, and tolerate such implementations by stripping out the zero flag when populating the free list during initialization. Additionally, use the same stripped out entries for detection of incomplete writes and map restoration that happens at this stage. Add a sysfs file 'log_zero_flags' that indicates the ability to accept such a layout to userspace applications. This enables 'ndctl check-namespace' to recognize whether the kernel is able to handle zero flags, or whether it should attempt a fix-up under the --repair option. Cc: Dan Williams <dan.j.williams@intel.com> Reported-by: Dexuan Cui <decui@microsoft.com> Reported-by: Pedro d'Aquino Filocre F S Barbuda <pbarbuda@microsoft.com> Tested-by: Dexuan Cui <decui@microsoft.com> Signed-off-by: Vishal Verma <vishal.l.verma@intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2019-02-28 08:06:27 +08:00
u32 i, map_entry, log_oldmap, log_newmap;
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
arena->freelist = kcalloc(arena->nfree, sizeof(struct free_entry),
GFP_KERNEL);
if (!arena->freelist)
return -ENOMEM;
for (i = 0; i < arena->nfree; i++) {
new = btt_log_read(arena, i, &log_new, LOG_NEW_ENT);
if (new < 0)
return new;
libnvdimm/btt: Fix LBA masking during 'free list' population The Linux BTT implementation assumes that log entries will never have the 'zero' flag set, and indeed it never sets that flag for log entries itself. However, the UEFI spec is ambiguous on the exact format of the LBA field of a log entry, specifically as to whether it should include the additional flag bits or not. While a zero bit doesn't make sense in the context of a log entry, other BTT implementations might still have it set. If an implementation does happen to have it set, we would happily read it in as the next block to write to for writes. Since a high bit is set, it pushes the block number out of the range of an 'arena', and we fail such a write with an EIO. Follow the robustness principle, and tolerate such implementations by stripping out the zero flag when populating the free list during initialization. Additionally, use the same stripped out entries for detection of incomplete writes and map restoration that happens at this stage. Add a sysfs file 'log_zero_flags' that indicates the ability to accept such a layout to userspace applications. This enables 'ndctl check-namespace' to recognize whether the kernel is able to handle zero flags, or whether it should attempt a fix-up under the --repair option. Cc: Dan Williams <dan.j.williams@intel.com> Reported-by: Dexuan Cui <decui@microsoft.com> Reported-by: Pedro d'Aquino Filocre F S Barbuda <pbarbuda@microsoft.com> Tested-by: Dexuan Cui <decui@microsoft.com> Signed-off-by: Vishal Verma <vishal.l.verma@intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2019-02-28 08:06:27 +08:00
/* old and new map entries with any flags stripped out */
log_oldmap = ent_lba(le32_to_cpu(log_new.old_map));
log_newmap = ent_lba(le32_to_cpu(log_new.new_map));
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
/* sub points to the next one to be overwritten */
arena->freelist[i].sub = 1 - new;
arena->freelist[i].seq = nd_inc_seq(le32_to_cpu(log_new.seq));
libnvdimm/btt: Fix LBA masking during 'free list' population The Linux BTT implementation assumes that log entries will never have the 'zero' flag set, and indeed it never sets that flag for log entries itself. However, the UEFI spec is ambiguous on the exact format of the LBA field of a log entry, specifically as to whether it should include the additional flag bits or not. While a zero bit doesn't make sense in the context of a log entry, other BTT implementations might still have it set. If an implementation does happen to have it set, we would happily read it in as the next block to write to for writes. Since a high bit is set, it pushes the block number out of the range of an 'arena', and we fail such a write with an EIO. Follow the robustness principle, and tolerate such implementations by stripping out the zero flag when populating the free list during initialization. Additionally, use the same stripped out entries for detection of incomplete writes and map restoration that happens at this stage. Add a sysfs file 'log_zero_flags' that indicates the ability to accept such a layout to userspace applications. This enables 'ndctl check-namespace' to recognize whether the kernel is able to handle zero flags, or whether it should attempt a fix-up under the --repair option. Cc: Dan Williams <dan.j.williams@intel.com> Reported-by: Dexuan Cui <decui@microsoft.com> Reported-by: Pedro d'Aquino Filocre F S Barbuda <pbarbuda@microsoft.com> Tested-by: Dexuan Cui <decui@microsoft.com> Signed-off-by: Vishal Verma <vishal.l.verma@intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2019-02-28 08:06:27 +08:00
arena->freelist[i].block = log_oldmap;
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
libnvdimm, btt: rework error clearing Clearing errors or badblocks during a BTT write requires sending an ACPI DSM, which means potentially sleeping. Since a BTT IO happens in atomic context (preemption disabled, spinlocks may be held), we cannot perform error clearing in the course of an IO. Due to this error clearing for BTT IOs has hitherto been disabled. In this patch we move error clearing out of the atomic section, and thus re-enable error clearing with BTTs. When we are about to add a block to the free list, we check if it was previously marked as an error, and if it was, we add it to the freelist, but also set a flag that says error clearing will be required. We then drop the lane (ending the atomic context), and send a zero buffer so that the error can be cleared. The error flag in the free list is protected by the nd 'lane', and is set only be a thread while it holds that lane. When the error is cleared, the flag is cleared, but while holding a mutex for that freelist index. When writing, we check for two things - 1/ If the freelist mutex is held or if the error flag is set. If so, this is an error block that is being (or about to be) cleared. 2/ If the block is a known badblock based on nsio->bb The second check is required because the BTT map error flag for a map entry only gets set when an error LBA is read. If we write to a new location that may not have the map error flag set, but still might be in the region's badblock list, we can trigger an EIO on the write, which is undesirable and completely avoidable. Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Toshi Kani <toshi.kani@hpe.com> Cc: Dan Williams <dan.j.williams@intel.com> Signed-off-by: Vishal Verma <vishal.l.verma@intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2017-08-31 09:36:03 +08:00
/*
* FIXME: if error clearing fails during init, we want to make
* the BTT read-only
*/
if (ent_e_flag(le32_to_cpu(log_new.old_map)) &&
!ent_normal(le32_to_cpu(log_new.old_map))) {
libnvdimm/btt: Fix LBA masking during 'free list' population The Linux BTT implementation assumes that log entries will never have the 'zero' flag set, and indeed it never sets that flag for log entries itself. However, the UEFI spec is ambiguous on the exact format of the LBA field of a log entry, specifically as to whether it should include the additional flag bits or not. While a zero bit doesn't make sense in the context of a log entry, other BTT implementations might still have it set. If an implementation does happen to have it set, we would happily read it in as the next block to write to for writes. Since a high bit is set, it pushes the block number out of the range of an 'arena', and we fail such a write with an EIO. Follow the robustness principle, and tolerate such implementations by stripping out the zero flag when populating the free list during initialization. Additionally, use the same stripped out entries for detection of incomplete writes and map restoration that happens at this stage. Add a sysfs file 'log_zero_flags' that indicates the ability to accept such a layout to userspace applications. This enables 'ndctl check-namespace' to recognize whether the kernel is able to handle zero flags, or whether it should attempt a fix-up under the --repair option. Cc: Dan Williams <dan.j.williams@intel.com> Reported-by: Dexuan Cui <decui@microsoft.com> Reported-by: Pedro d'Aquino Filocre F S Barbuda <pbarbuda@microsoft.com> Tested-by: Dexuan Cui <decui@microsoft.com> Signed-off-by: Vishal Verma <vishal.l.verma@intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2019-02-28 08:06:27 +08:00
arena->freelist[i].has_err = 1;
libnvdimm, btt: rework error clearing Clearing errors or badblocks during a BTT write requires sending an ACPI DSM, which means potentially sleeping. Since a BTT IO happens in atomic context (preemption disabled, spinlocks may be held), we cannot perform error clearing in the course of an IO. Due to this error clearing for BTT IOs has hitherto been disabled. In this patch we move error clearing out of the atomic section, and thus re-enable error clearing with BTTs. When we are about to add a block to the free list, we check if it was previously marked as an error, and if it was, we add it to the freelist, but also set a flag that says error clearing will be required. We then drop the lane (ending the atomic context), and send a zero buffer so that the error can be cleared. The error flag in the free list is protected by the nd 'lane', and is set only be a thread while it holds that lane. When the error is cleared, the flag is cleared, but while holding a mutex for that freelist index. When writing, we check for two things - 1/ If the freelist mutex is held or if the error flag is set. If so, this is an error block that is being (or about to be) cleared. 2/ If the block is a known badblock based on nsio->bb The second check is required because the BTT map error flag for a map entry only gets set when an error LBA is read. If we write to a new location that may not have the map error flag set, but still might be in the region's badblock list, we can trigger an EIO on the write, which is undesirable and completely avoidable. Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Toshi Kani <toshi.kani@hpe.com> Cc: Dan Williams <dan.j.williams@intel.com> Signed-off-by: Vishal Verma <vishal.l.verma@intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2017-08-31 09:36:03 +08:00
ret = arena_clear_freelist_error(arena, i);
if (ret)
dev_err_ratelimited(to_dev(arena),
"Unable to clear known errors\n");
libnvdimm, btt: rework error clearing Clearing errors or badblocks during a BTT write requires sending an ACPI DSM, which means potentially sleeping. Since a BTT IO happens in atomic context (preemption disabled, spinlocks may be held), we cannot perform error clearing in the course of an IO. Due to this error clearing for BTT IOs has hitherto been disabled. In this patch we move error clearing out of the atomic section, and thus re-enable error clearing with BTTs. When we are about to add a block to the free list, we check if it was previously marked as an error, and if it was, we add it to the freelist, but also set a flag that says error clearing will be required. We then drop the lane (ending the atomic context), and send a zero buffer so that the error can be cleared. The error flag in the free list is protected by the nd 'lane', and is set only be a thread while it holds that lane. When the error is cleared, the flag is cleared, but while holding a mutex for that freelist index. When writing, we check for two things - 1/ If the freelist mutex is held or if the error flag is set. If so, this is an error block that is being (or about to be) cleared. 2/ If the block is a known badblock based on nsio->bb The second check is required because the BTT map error flag for a map entry only gets set when an error LBA is read. If we write to a new location that may not have the map error flag set, but still might be in the region's badblock list, we can trigger an EIO on the write, which is undesirable and completely avoidable. Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Toshi Kani <toshi.kani@hpe.com> Cc: Dan Williams <dan.j.williams@intel.com> Signed-off-by: Vishal Verma <vishal.l.verma@intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2017-08-31 09:36:03 +08:00
}
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
/* This implies a newly created or untouched flog entry */
libnvdimm/btt: Fix LBA masking during 'free list' population The Linux BTT implementation assumes that log entries will never have the 'zero' flag set, and indeed it never sets that flag for log entries itself. However, the UEFI spec is ambiguous on the exact format of the LBA field of a log entry, specifically as to whether it should include the additional flag bits or not. While a zero bit doesn't make sense in the context of a log entry, other BTT implementations might still have it set. If an implementation does happen to have it set, we would happily read it in as the next block to write to for writes. Since a high bit is set, it pushes the block number out of the range of an 'arena', and we fail such a write with an EIO. Follow the robustness principle, and tolerate such implementations by stripping out the zero flag when populating the free list during initialization. Additionally, use the same stripped out entries for detection of incomplete writes and map restoration that happens at this stage. Add a sysfs file 'log_zero_flags' that indicates the ability to accept such a layout to userspace applications. This enables 'ndctl check-namespace' to recognize whether the kernel is able to handle zero flags, or whether it should attempt a fix-up under the --repair option. Cc: Dan Williams <dan.j.williams@intel.com> Reported-by: Dexuan Cui <decui@microsoft.com> Reported-by: Pedro d'Aquino Filocre F S Barbuda <pbarbuda@microsoft.com> Tested-by: Dexuan Cui <decui@microsoft.com> Signed-off-by: Vishal Verma <vishal.l.verma@intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2019-02-28 08:06:27 +08:00
if (log_oldmap == log_newmap)
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
continue;
/* Check if map recovery is needed */
ret = btt_map_read(arena, le32_to_cpu(log_new.lba), &map_entry,
NULL, NULL, 0);
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
if (ret)
return ret;
libnvdimm/btt: Fix LBA masking during 'free list' population The Linux BTT implementation assumes that log entries will never have the 'zero' flag set, and indeed it never sets that flag for log entries itself. However, the UEFI spec is ambiguous on the exact format of the LBA field of a log entry, specifically as to whether it should include the additional flag bits or not. While a zero bit doesn't make sense in the context of a log entry, other BTT implementations might still have it set. If an implementation does happen to have it set, we would happily read it in as the next block to write to for writes. Since a high bit is set, it pushes the block number out of the range of an 'arena', and we fail such a write with an EIO. Follow the robustness principle, and tolerate such implementations by stripping out the zero flag when populating the free list during initialization. Additionally, use the same stripped out entries for detection of incomplete writes and map restoration that happens at this stage. Add a sysfs file 'log_zero_flags' that indicates the ability to accept such a layout to userspace applications. This enables 'ndctl check-namespace' to recognize whether the kernel is able to handle zero flags, or whether it should attempt a fix-up under the --repair option. Cc: Dan Williams <dan.j.williams@intel.com> Reported-by: Dexuan Cui <decui@microsoft.com> Reported-by: Pedro d'Aquino Filocre F S Barbuda <pbarbuda@microsoft.com> Tested-by: Dexuan Cui <decui@microsoft.com> Signed-off-by: Vishal Verma <vishal.l.verma@intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2019-02-28 08:06:27 +08:00
/*
* The map_entry from btt_read_map is stripped of any flag bits,
* so use the stripped out versions from the log as well for
* testing whether recovery is needed. For restoration, use the
* 'raw' version of the log entries as that captured what we
* were going to write originally.
*/
if ((log_newmap != map_entry) && (log_oldmap == map_entry)) {
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
/*
* Last transaction wrote the flog, but wasn't able
* to complete the map write. So fix up the map.
*/
ret = btt_map_write(arena, le32_to_cpu(log_new.lba),
le32_to_cpu(log_new.new_map), 0, 0, 0);
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
if (ret)
return ret;
}
}
return 0;
}
static bool ent_is_padding(struct log_entry *ent)
{
return (ent->lba == 0) && (ent->old_map == 0) && (ent->new_map == 0)
&& (ent->seq == 0);
}
/*
* Detecting valid log indices: We read a log group (see the comments in btt.h
* for a description of a 'log_group' and its 'slots'), and iterate over its
* four slots. We expect that a padding slot will be all-zeroes, and use this
* to detect a padding slot vs. an actual entry.
*
* If a log_group is in the initial state, i.e. hasn't been used since the
* creation of this BTT layout, it will have three of the four slots with
* zeroes. We skip over these log_groups for the detection of log_index. If
* all log_groups are in the initial state (i.e. the BTT has never been
* written to), it is safe to assume the 'new format' of log entries in slots
* (0, 1).
*/
static int log_set_indices(struct arena_info *arena)
{
bool idx_set = false, initial_state = true;
int ret, log_index[2] = {-1, -1};
u32 i, j, next_idx = 0;
struct log_group log;
u32 pad_count = 0;
for (i = 0; i < arena->nfree; i++) {
ret = btt_log_group_read(arena, i, &log);
if (ret < 0)
return ret;
for (j = 0; j < 4; j++) {
if (!idx_set) {
if (ent_is_padding(&log.ent[j])) {
pad_count++;
continue;
} else {
/* Skip if index has been recorded */
if ((next_idx == 1) &&
(j == log_index[0]))
continue;
/* valid entry, record index */
log_index[next_idx] = j;
next_idx++;
}
if (next_idx == 2) {
/* two valid entries found */
idx_set = true;
} else if (next_idx > 2) {
/* too many valid indices */
return -ENXIO;
}
} else {
/*
* once the indices have been set, just verify
* that all subsequent log groups are either in
* their initial state or follow the same
* indices.
*/
if (j == log_index[0]) {
/* entry must be 'valid' */
if (ent_is_padding(&log.ent[j]))
return -ENXIO;
} else if (j == log_index[1]) {
;
/*
* log_index[1] can be padding if the
* lane never got used and it is still
* in the initial state (three 'padding'
* entries)
*/
} else {
/* entry must be invalid (padding) */
if (!ent_is_padding(&log.ent[j]))
return -ENXIO;
}
}
}
/*
* If any of the log_groups have more than one valid,
* non-padding entry, then the we are no longer in the
* initial_state
*/
if (pad_count < 3)
initial_state = false;
pad_count = 0;
}
if (!initial_state && !idx_set)
return -ENXIO;
/*
* If all the entries in the log were in the initial state,
* assume new padding scheme
*/
if (initial_state)
log_index[1] = 1;
/*
* Only allow the known permutations of log/padding indices,
* i.e. (0, 1), and (0, 2)
*/
if ((log_index[0] == 0) && ((log_index[1] == 1) || (log_index[1] == 2)))
; /* known index possibilities */
else {
dev_err(to_dev(arena), "Found an unknown padding scheme\n");
return -ENXIO;
}
arena->log_index[0] = log_index[0];
arena->log_index[1] = log_index[1];
dev_dbg(to_dev(arena), "log_index_0 = %d\n", log_index[0]);
dev_dbg(to_dev(arena), "log_index_1 = %d\n", log_index[1]);
return 0;
}
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
static int btt_rtt_init(struct arena_info *arena)
{
arena->rtt = kcalloc(arena->nfree, sizeof(u32), GFP_KERNEL);
if (arena->rtt == NULL)
return -ENOMEM;
return 0;
}
static int btt_maplocks_init(struct arena_info *arena)
{
u32 i;
arena->map_locks = kcalloc(arena->nfree, sizeof(struct aligned_lock),
GFP_KERNEL);
if (!arena->map_locks)
return -ENOMEM;
for (i = 0; i < arena->nfree; i++)
spin_lock_init(&arena->map_locks[i].lock);
return 0;
}
static struct arena_info *alloc_arena(struct btt *btt, size_t size,
size_t start, size_t arena_off)
{
struct arena_info *arena;
u64 logsize, mapsize, datasize;
u64 available = size;
arena = kzalloc(sizeof(struct arena_info), GFP_KERNEL);
if (!arena)
return NULL;
arena->nd_btt = btt->nd_btt;
arena->sector_size = btt->sector_size;
mutex_init(&arena->err_lock);
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
if (!size)
return arena;
arena->size = size;
arena->external_lba_start = start;
arena->external_lbasize = btt->lbasize;
arena->internal_lbasize = roundup(arena->external_lbasize,
INT_LBASIZE_ALIGNMENT);
arena->nfree = BTT_DEFAULT_NFREE;
arena->version_major = btt->nd_btt->version_major;
arena->version_minor = btt->nd_btt->version_minor;
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
if (available % BTT_PG_SIZE)
available -= (available % BTT_PG_SIZE);
/* Two pages are reserved for the super block and its copy */
available -= 2 * BTT_PG_SIZE;
/* The log takes a fixed amount of space based on nfree */
logsize = roundup(arena->nfree * LOG_GRP_SIZE, BTT_PG_SIZE);
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
available -= logsize;
/* Calculate optimal split between map and data area */
arena->internal_nlba = div_u64(available - BTT_PG_SIZE,
arena->internal_lbasize + MAP_ENT_SIZE);
arena->external_nlba = arena->internal_nlba - arena->nfree;
mapsize = roundup((arena->external_nlba * MAP_ENT_SIZE), BTT_PG_SIZE);
datasize = available - mapsize;
/* 'Absolute' values, relative to start of storage space */
arena->infooff = arena_off;
arena->dataoff = arena->infooff + BTT_PG_SIZE;
arena->mapoff = arena->dataoff + datasize;
arena->logoff = arena->mapoff + mapsize;
arena->info2off = arena->logoff + logsize;
/* Default log indices are (0,1) */
arena->log_index[0] = 0;
arena->log_index[1] = 1;
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
return arena;
}
static void free_arenas(struct btt *btt)
{
struct arena_info *arena, *next;
list_for_each_entry_safe(arena, next, &btt->arena_list, list) {
list_del(&arena->list);
kfree(arena->rtt);
kfree(arena->map_locks);
kfree(arena->freelist);
debugfs_remove_recursive(arena->debugfs_dir);
kfree(arena);
}
}
/*
* This function reads an existing valid btt superblock and
* populates the corresponding arena_info struct
*/
static void parse_arena_meta(struct arena_info *arena, struct btt_sb *super,
u64 arena_off)
{
arena->internal_nlba = le32_to_cpu(super->internal_nlba);
arena->internal_lbasize = le32_to_cpu(super->internal_lbasize);
arena->external_nlba = le32_to_cpu(super->external_nlba);
arena->external_lbasize = le32_to_cpu(super->external_lbasize);
arena->nfree = le32_to_cpu(super->nfree);
arena->version_major = le16_to_cpu(super->version_major);
arena->version_minor = le16_to_cpu(super->version_minor);
arena->nextoff = (super->nextoff == 0) ? 0 : (arena_off +
le64_to_cpu(super->nextoff));
arena->infooff = arena_off;
arena->dataoff = arena_off + le64_to_cpu(super->dataoff);
arena->mapoff = arena_off + le64_to_cpu(super->mapoff);
arena->logoff = arena_off + le64_to_cpu(super->logoff);
arena->info2off = arena_off + le64_to_cpu(super->info2off);
arena->size = (le64_to_cpu(super->nextoff) > 0)
? (le64_to_cpu(super->nextoff))
: (arena->info2off - arena->infooff + BTT_PG_SIZE);
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
arena->flags = le32_to_cpu(super->flags);
}
static int discover_arenas(struct btt *btt)
{
int ret = 0;
struct arena_info *arena;
struct btt_sb *super;
size_t remaining = btt->rawsize;
u64 cur_nlba = 0;
size_t cur_off = 0;
int num_arenas = 0;
super = kzalloc(sizeof(*super), GFP_KERNEL);
if (!super)
return -ENOMEM;
while (remaining) {
/* Alloc memory for arena */
arena = alloc_arena(btt, 0, 0, 0);
if (!arena) {
ret = -ENOMEM;
goto out_super;
}
arena->infooff = cur_off;
ret = btt_info_read(arena, super);
if (ret)
goto out;
if (!nd_btt_arena_is_valid(btt->nd_btt, super)) {
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
if (remaining == btt->rawsize) {
btt->init_state = INIT_NOTFOUND;
dev_info(to_dev(arena), "No existing arenas\n");
goto out;
} else {
dev_err(to_dev(arena),
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
"Found corrupted metadata!\n");
ret = -ENODEV;
goto out;
}
}
arena->external_lba_start = cur_nlba;
parse_arena_meta(arena, super, cur_off);
ret = log_set_indices(arena);
if (ret) {
dev_err(to_dev(arena),
"Unable to deduce log/padding indices\n");
goto out;
}
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
ret = btt_freelist_init(arena);
if (ret)
goto out;
ret = btt_rtt_init(arena);
if (ret)
goto out;
ret = btt_maplocks_init(arena);
if (ret)
goto out;
list_add_tail(&arena->list, &btt->arena_list);
remaining -= arena->size;
cur_off += arena->size;
cur_nlba += arena->external_nlba;
num_arenas++;
if (arena->nextoff == 0)
break;
}
btt->num_arenas = num_arenas;
btt->nlba = cur_nlba;
btt->init_state = INIT_READY;
kfree(super);
return ret;
out:
kfree(arena);
free_arenas(btt);
out_super:
kfree(super);
return ret;
}
static int create_arenas(struct btt *btt)
{
size_t remaining = btt->rawsize;
size_t cur_off = 0;
while (remaining) {
struct arena_info *arena;
size_t arena_size = min_t(u64, ARENA_MAX_SIZE, remaining);
remaining -= arena_size;
if (arena_size < ARENA_MIN_SIZE)
break;
arena = alloc_arena(btt, arena_size, btt->nlba, cur_off);
if (!arena) {
free_arenas(btt);
return -ENOMEM;
}
btt->nlba += arena->external_nlba;
if (remaining >= ARENA_MIN_SIZE)
arena->nextoff = arena->size;
else
arena->nextoff = 0;
cur_off += arena_size;
list_add_tail(&arena->list, &btt->arena_list);
}
return 0;
}
/*
* This function completes arena initialization by writing
* all the metadata.
* It is only called for an uninitialized arena when a write
* to that arena occurs for the first time.
*/
static int btt_arena_write_layout(struct arena_info *arena)
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
{
int ret;
u64 sum;
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
struct btt_sb *super;
struct nd_btt *nd_btt = arena->nd_btt;
const u8 *parent_uuid = nd_dev_to_uuid(&nd_btt->ndns->dev);
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
ret = btt_map_init(arena);
if (ret)
return ret;
ret = btt_log_init(arena);
if (ret)
return ret;
super = kzalloc(sizeof(struct btt_sb), GFP_NOIO);
if (!super)
return -ENOMEM;
strncpy(super->signature, BTT_SIG, BTT_SIG_LEN);
memcpy(super->uuid, nd_btt->uuid, 16);
memcpy(super->parent_uuid, parent_uuid, 16);
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
super->flags = cpu_to_le32(arena->flags);
super->version_major = cpu_to_le16(arena->version_major);
super->version_minor = cpu_to_le16(arena->version_minor);
super->external_lbasize = cpu_to_le32(arena->external_lbasize);
super->external_nlba = cpu_to_le32(arena->external_nlba);
super->internal_lbasize = cpu_to_le32(arena->internal_lbasize);
super->internal_nlba = cpu_to_le32(arena->internal_nlba);
super->nfree = cpu_to_le32(arena->nfree);
super->infosize = cpu_to_le32(sizeof(struct btt_sb));
super->nextoff = cpu_to_le64(arena->nextoff);
/*
* Subtract arena->infooff (arena start) so numbers are relative
* to 'this' arena
*/
super->dataoff = cpu_to_le64(arena->dataoff - arena->infooff);
super->mapoff = cpu_to_le64(arena->mapoff - arena->infooff);
super->logoff = cpu_to_le64(arena->logoff - arena->infooff);
super->info2off = cpu_to_le64(arena->info2off - arena->infooff);
super->flags = 0;
sum = nd_sb_checksum((struct nd_gen_sb *) super);
super->checksum = cpu_to_le64(sum);
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
ret = btt_info_write(arena, super);
kfree(super);
return ret;
}
/*
* This function completes the initialization for the BTT namespace
* such that it is ready to accept IOs
*/
static int btt_meta_init(struct btt *btt)
{
int ret = 0;
struct arena_info *arena;
mutex_lock(&btt->init_lock);
list_for_each_entry(arena, &btt->arena_list, list) {
ret = btt_arena_write_layout(arena);
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
if (ret)
goto unlock;
ret = btt_freelist_init(arena);
if (ret)
goto unlock;
ret = btt_rtt_init(arena);
if (ret)
goto unlock;
ret = btt_maplocks_init(arena);
if (ret)
goto unlock;
}
btt->init_state = INIT_READY;
unlock:
mutex_unlock(&btt->init_lock);
return ret;
}
static u32 btt_meta_size(struct btt *btt)
{
return btt->lbasize - btt->sector_size;
}
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
/*
* This function calculates the arena in which the given LBA lies
* by doing a linear walk. This is acceptable since we expect only
* a few arenas. If we have backing devices that get much larger,
* we can construct a balanced binary tree of arenas at init time
* so that this range search becomes faster.
*/
static int lba_to_arena(struct btt *btt, sector_t sector, __u32 *premap,
struct arena_info **arena)
{
struct arena_info *arena_list;
__u64 lba = div_u64(sector << SECTOR_SHIFT, btt->sector_size);
list_for_each_entry(arena_list, &btt->arena_list, list) {
if (lba < arena_list->external_nlba) {
*arena = arena_list;
*premap = lba;
return 0;
}
lba -= arena_list->external_nlba;
}
return -EIO;
}
/*
* The following (lock_map, unlock_map) are mostly just to improve
* readability, since they index into an array of locks
*/
static void lock_map(struct arena_info *arena, u32 premap)
__acquires(&arena->map_locks[idx].lock)
{
u32 idx = (premap * MAP_ENT_SIZE / L1_CACHE_BYTES) % arena->nfree;
spin_lock(&arena->map_locks[idx].lock);
}
static void unlock_map(struct arena_info *arena, u32 premap)
__releases(&arena->map_locks[idx].lock)
{
u32 idx = (premap * MAP_ENT_SIZE / L1_CACHE_BYTES) % arena->nfree;
spin_unlock(&arena->map_locks[idx].lock);
}
static int btt_data_read(struct arena_info *arena, struct page *page,
unsigned int off, u32 lba, u32 len)
{
int ret;
u64 nsoff = to_namespace_offset(arena, lba);
void *mem = kmap_atomic(page);
ret = arena_read_bytes(arena, nsoff, mem + off, len, NVDIMM_IO_ATOMIC);
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
kunmap_atomic(mem);
return ret;
}
static int btt_data_write(struct arena_info *arena, u32 lba,
struct page *page, unsigned int off, u32 len)
{
int ret;
u64 nsoff = to_namespace_offset(arena, lba);
void *mem = kmap_atomic(page);
ret = arena_write_bytes(arena, nsoff, mem + off, len, NVDIMM_IO_ATOMIC);
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
kunmap_atomic(mem);
return ret;
}
static void zero_fill_data(struct page *page, unsigned int off, u32 len)
{
void *mem = kmap_atomic(page);
memset(mem + off, 0, len);
kunmap_atomic(mem);
}
#ifdef CONFIG_BLK_DEV_INTEGRITY
static int btt_rw_integrity(struct btt *btt, struct bio_integrity_payload *bip,
struct arena_info *arena, u32 postmap, int rw)
{
unsigned int len = btt_meta_size(btt);
u64 meta_nsoff;
int ret = 0;
if (bip == NULL)
return 0;
meta_nsoff = to_namespace_offset(arena, postmap) + btt->sector_size;
while (len) {
unsigned int cur_len;
struct bio_vec bv;
void *mem;
bv = bvec_iter_bvec(bip->bip_vec, bip->bip_iter);
/*
* The 'bv' obtained from bvec_iter_bvec has its .bv_len and
* .bv_offset already adjusted for iter->bi_bvec_done, and we
* can use those directly
*/
cur_len = min(len, bv.bv_len);
mem = kmap_atomic(bv.bv_page);
if (rw)
ret = arena_write_bytes(arena, meta_nsoff,
mem + bv.bv_offset, cur_len,
NVDIMM_IO_ATOMIC);
else
ret = arena_read_bytes(arena, meta_nsoff,
mem + bv.bv_offset, cur_len,
NVDIMM_IO_ATOMIC);
kunmap_atomic(mem);
if (ret)
return ret;
len -= cur_len;
meta_nsoff += cur_len;
if (!bvec_iter_advance(bip->bip_vec, &bip->bip_iter, cur_len))
return -EIO;
}
return ret;
}
#else /* CONFIG_BLK_DEV_INTEGRITY */
static int btt_rw_integrity(struct btt *btt, struct bio_integrity_payload *bip,
struct arena_info *arena, u32 postmap, int rw)
{
return 0;
}
#endif
static int btt_read_pg(struct btt *btt, struct bio_integrity_payload *bip,
struct page *page, unsigned int off, sector_t sector,
unsigned int len)
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
{
int ret = 0;
int t_flag, e_flag;
struct arena_info *arena = NULL;
u32 lane = 0, premap, postmap;
while (len) {
u32 cur_len;
lane = nd_region_acquire_lane(btt->nd_region);
ret = lba_to_arena(btt, sector, &premap, &arena);
if (ret)
goto out_lane;
cur_len = min(btt->sector_size, len);
ret = btt_map_read(arena, premap, &postmap, &t_flag, &e_flag,
NVDIMM_IO_ATOMIC);
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
if (ret)
goto out_lane;
/*
* We loop to make sure that the post map LBA didn't change
* from under us between writing the RTT and doing the actual
* read.
*/
while (1) {
u32 new_map;
int new_t, new_e;
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
if (t_flag) {
zero_fill_data(page, off, cur_len);
goto out_lane;
}
if (e_flag) {
ret = -EIO;
goto out_lane;
}
arena->rtt[lane] = RTT_VALID | postmap;
/*
* Barrier to make sure this write is not reordered
* to do the verification map_read before the RTT store
*/
barrier();
ret = btt_map_read(arena, premap, &new_map, &new_t,
&new_e, NVDIMM_IO_ATOMIC);
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
if (ret)
goto out_rtt;
if ((postmap == new_map) && (t_flag == new_t) &&
(e_flag == new_e))
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
break;
postmap = new_map;
t_flag = new_t;
e_flag = new_e;
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
}
ret = btt_data_read(arena, page, off, postmap, cur_len);
libnvdimm, btt: rework error clearing Clearing errors or badblocks during a BTT write requires sending an ACPI DSM, which means potentially sleeping. Since a BTT IO happens in atomic context (preemption disabled, spinlocks may be held), we cannot perform error clearing in the course of an IO. Due to this error clearing for BTT IOs has hitherto been disabled. In this patch we move error clearing out of the atomic section, and thus re-enable error clearing with BTTs. When we are about to add a block to the free list, we check if it was previously marked as an error, and if it was, we add it to the freelist, but also set a flag that says error clearing will be required. We then drop the lane (ending the atomic context), and send a zero buffer so that the error can be cleared. The error flag in the free list is protected by the nd 'lane', and is set only be a thread while it holds that lane. When the error is cleared, the flag is cleared, but while holding a mutex for that freelist index. When writing, we check for two things - 1/ If the freelist mutex is held or if the error flag is set. If so, this is an error block that is being (or about to be) cleared. 2/ If the block is a known badblock based on nsio->bb The second check is required because the BTT map error flag for a map entry only gets set when an error LBA is read. If we write to a new location that may not have the map error flag set, but still might be in the region's badblock list, we can trigger an EIO on the write, which is undesirable and completely avoidable. Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Toshi Kani <toshi.kani@hpe.com> Cc: Dan Williams <dan.j.williams@intel.com> Signed-off-by: Vishal Verma <vishal.l.verma@intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2017-08-31 09:36:03 +08:00
if (ret) {
/* Media error - set the e_flag */
if (btt_map_write(arena, premap, postmap, 0, 1, NVDIMM_IO_ATOMIC))
dev_warn_ratelimited(to_dev(arena),
"Error persistently tracking bad blocks at %#x\n",
premap);
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
goto out_rtt;
libnvdimm, btt: rework error clearing Clearing errors or badblocks during a BTT write requires sending an ACPI DSM, which means potentially sleeping. Since a BTT IO happens in atomic context (preemption disabled, spinlocks may be held), we cannot perform error clearing in the course of an IO. Due to this error clearing for BTT IOs has hitherto been disabled. In this patch we move error clearing out of the atomic section, and thus re-enable error clearing with BTTs. When we are about to add a block to the free list, we check if it was previously marked as an error, and if it was, we add it to the freelist, but also set a flag that says error clearing will be required. We then drop the lane (ending the atomic context), and send a zero buffer so that the error can be cleared. The error flag in the free list is protected by the nd 'lane', and is set only be a thread while it holds that lane. When the error is cleared, the flag is cleared, but while holding a mutex for that freelist index. When writing, we check for two things - 1/ If the freelist mutex is held or if the error flag is set. If so, this is an error block that is being (or about to be) cleared. 2/ If the block is a known badblock based on nsio->bb The second check is required because the BTT map error flag for a map entry only gets set when an error LBA is read. If we write to a new location that may not have the map error flag set, but still might be in the region's badblock list, we can trigger an EIO on the write, which is undesirable and completely avoidable. Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Toshi Kani <toshi.kani@hpe.com> Cc: Dan Williams <dan.j.williams@intel.com> Signed-off-by: Vishal Verma <vishal.l.verma@intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2017-08-31 09:36:03 +08:00
}
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
if (bip) {
ret = btt_rw_integrity(btt, bip, arena, postmap, READ);
if (ret)
goto out_rtt;
}
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
arena->rtt[lane] = RTT_INVALID;
nd_region_release_lane(btt->nd_region, lane);
len -= cur_len;
off += cur_len;
sector += btt->sector_size >> SECTOR_SHIFT;
}
return 0;
out_rtt:
arena->rtt[lane] = RTT_INVALID;
out_lane:
nd_region_release_lane(btt->nd_region, lane);
return ret;
}
libnvdimm, btt: rework error clearing Clearing errors or badblocks during a BTT write requires sending an ACPI DSM, which means potentially sleeping. Since a BTT IO happens in atomic context (preemption disabled, spinlocks may be held), we cannot perform error clearing in the course of an IO. Due to this error clearing for BTT IOs has hitherto been disabled. In this patch we move error clearing out of the atomic section, and thus re-enable error clearing with BTTs. When we are about to add a block to the free list, we check if it was previously marked as an error, and if it was, we add it to the freelist, but also set a flag that says error clearing will be required. We then drop the lane (ending the atomic context), and send a zero buffer so that the error can be cleared. The error flag in the free list is protected by the nd 'lane', and is set only be a thread while it holds that lane. When the error is cleared, the flag is cleared, but while holding a mutex for that freelist index. When writing, we check for two things - 1/ If the freelist mutex is held or if the error flag is set. If so, this is an error block that is being (or about to be) cleared. 2/ If the block is a known badblock based on nsio->bb The second check is required because the BTT map error flag for a map entry only gets set when an error LBA is read. If we write to a new location that may not have the map error flag set, but still might be in the region's badblock list, we can trigger an EIO on the write, which is undesirable and completely avoidable. Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Toshi Kani <toshi.kani@hpe.com> Cc: Dan Williams <dan.j.williams@intel.com> Signed-off-by: Vishal Verma <vishal.l.verma@intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2017-08-31 09:36:03 +08:00
/*
* Normally, arena_{read,write}_bytes will take care of the initial offset
* adjustment, but in the case of btt_is_badblock, where we query is_bad_pmem,
* we need the final, raw namespace offset here
*/
static bool btt_is_badblock(struct btt *btt, struct arena_info *arena,
u32 postmap)
{
u64 nsoff = adjust_initial_offset(arena->nd_btt,
to_namespace_offset(arena, postmap));
sector_t phys_sector = nsoff >> 9;
return is_bad_pmem(btt->phys_bb, phys_sector, arena->internal_lbasize);
}
static int btt_write_pg(struct btt *btt, struct bio_integrity_payload *bip,
sector_t sector, struct page *page, unsigned int off,
unsigned int len)
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
{
int ret = 0;
struct arena_info *arena = NULL;
u32 premap = 0, old_postmap, new_postmap, lane = 0, i;
struct log_entry log;
int sub;
while (len) {
u32 cur_len;
libnvdimm, btt: rework error clearing Clearing errors or badblocks during a BTT write requires sending an ACPI DSM, which means potentially sleeping. Since a BTT IO happens in atomic context (preemption disabled, spinlocks may be held), we cannot perform error clearing in the course of an IO. Due to this error clearing for BTT IOs has hitherto been disabled. In this patch we move error clearing out of the atomic section, and thus re-enable error clearing with BTTs. When we are about to add a block to the free list, we check if it was previously marked as an error, and if it was, we add it to the freelist, but also set a flag that says error clearing will be required. We then drop the lane (ending the atomic context), and send a zero buffer so that the error can be cleared. The error flag in the free list is protected by the nd 'lane', and is set only be a thread while it holds that lane. When the error is cleared, the flag is cleared, but while holding a mutex for that freelist index. When writing, we check for two things - 1/ If the freelist mutex is held or if the error flag is set. If so, this is an error block that is being (or about to be) cleared. 2/ If the block is a known badblock based on nsio->bb The second check is required because the BTT map error flag for a map entry only gets set when an error LBA is read. If we write to a new location that may not have the map error flag set, but still might be in the region's badblock list, we can trigger an EIO on the write, which is undesirable and completely avoidable. Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Toshi Kani <toshi.kani@hpe.com> Cc: Dan Williams <dan.j.williams@intel.com> Signed-off-by: Vishal Verma <vishal.l.verma@intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2017-08-31 09:36:03 +08:00
int e_flag;
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
libnvdimm, btt: rework error clearing Clearing errors or badblocks during a BTT write requires sending an ACPI DSM, which means potentially sleeping. Since a BTT IO happens in atomic context (preemption disabled, spinlocks may be held), we cannot perform error clearing in the course of an IO. Due to this error clearing for BTT IOs has hitherto been disabled. In this patch we move error clearing out of the atomic section, and thus re-enable error clearing with BTTs. When we are about to add a block to the free list, we check if it was previously marked as an error, and if it was, we add it to the freelist, but also set a flag that says error clearing will be required. We then drop the lane (ending the atomic context), and send a zero buffer so that the error can be cleared. The error flag in the free list is protected by the nd 'lane', and is set only be a thread while it holds that lane. When the error is cleared, the flag is cleared, but while holding a mutex for that freelist index. When writing, we check for two things - 1/ If the freelist mutex is held or if the error flag is set. If so, this is an error block that is being (or about to be) cleared. 2/ If the block is a known badblock based on nsio->bb The second check is required because the BTT map error flag for a map entry only gets set when an error LBA is read. If we write to a new location that may not have the map error flag set, but still might be in the region's badblock list, we can trigger an EIO on the write, which is undesirable and completely avoidable. Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Toshi Kani <toshi.kani@hpe.com> Cc: Dan Williams <dan.j.williams@intel.com> Signed-off-by: Vishal Verma <vishal.l.verma@intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2017-08-31 09:36:03 +08:00
retry:
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
lane = nd_region_acquire_lane(btt->nd_region);
ret = lba_to_arena(btt, sector, &premap, &arena);
if (ret)
goto out_lane;
cur_len = min(btt->sector_size, len);
if ((arena->flags & IB_FLAG_ERROR_MASK) != 0) {
ret = -EIO;
goto out_lane;
}
libnvdimm, btt: rework error clearing Clearing errors or badblocks during a BTT write requires sending an ACPI DSM, which means potentially sleeping. Since a BTT IO happens in atomic context (preemption disabled, spinlocks may be held), we cannot perform error clearing in the course of an IO. Due to this error clearing for BTT IOs has hitherto been disabled. In this patch we move error clearing out of the atomic section, and thus re-enable error clearing with BTTs. When we are about to add a block to the free list, we check if it was previously marked as an error, and if it was, we add it to the freelist, but also set a flag that says error clearing will be required. We then drop the lane (ending the atomic context), and send a zero buffer so that the error can be cleared. The error flag in the free list is protected by the nd 'lane', and is set only be a thread while it holds that lane. When the error is cleared, the flag is cleared, but while holding a mutex for that freelist index. When writing, we check for two things - 1/ If the freelist mutex is held or if the error flag is set. If so, this is an error block that is being (or about to be) cleared. 2/ If the block is a known badblock based on nsio->bb The second check is required because the BTT map error flag for a map entry only gets set when an error LBA is read. If we write to a new location that may not have the map error flag set, but still might be in the region's badblock list, we can trigger an EIO on the write, which is undesirable and completely avoidable. Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Toshi Kani <toshi.kani@hpe.com> Cc: Dan Williams <dan.j.williams@intel.com> Signed-off-by: Vishal Verma <vishal.l.verma@intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2017-08-31 09:36:03 +08:00
if (btt_is_badblock(btt, arena, arena->freelist[lane].block))
arena->freelist[lane].has_err = 1;
if (mutex_is_locked(&arena->err_lock)
|| arena->freelist[lane].has_err) {
nd_region_release_lane(btt->nd_region, lane);
ret = arena_clear_freelist_error(arena, lane);
if (ret)
return ret;
/* OK to acquire a different lane/free block */
goto retry;
}
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
new_postmap = arena->freelist[lane].block;
/* Wait if the new block is being read from */
for (i = 0; i < arena->nfree; i++)
while (arena->rtt[i] == (RTT_VALID | new_postmap))
cpu_relax();
if (new_postmap >= arena->internal_nlba) {
ret = -EIO;
goto out_lane;
}
ret = btt_data_write(arena, new_postmap, page, off, cur_len);
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
if (ret)
goto out_lane;
if (bip) {
ret = btt_rw_integrity(btt, bip, arena, new_postmap,
WRITE);
if (ret)
goto out_lane;
}
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
lock_map(arena, premap);
libnvdimm, btt: rework error clearing Clearing errors or badblocks during a BTT write requires sending an ACPI DSM, which means potentially sleeping. Since a BTT IO happens in atomic context (preemption disabled, spinlocks may be held), we cannot perform error clearing in the course of an IO. Due to this error clearing for BTT IOs has hitherto been disabled. In this patch we move error clearing out of the atomic section, and thus re-enable error clearing with BTTs. When we are about to add a block to the free list, we check if it was previously marked as an error, and if it was, we add it to the freelist, but also set a flag that says error clearing will be required. We then drop the lane (ending the atomic context), and send a zero buffer so that the error can be cleared. The error flag in the free list is protected by the nd 'lane', and is set only be a thread while it holds that lane. When the error is cleared, the flag is cleared, but while holding a mutex for that freelist index. When writing, we check for two things - 1/ If the freelist mutex is held or if the error flag is set. If so, this is an error block that is being (or about to be) cleared. 2/ If the block is a known badblock based on nsio->bb The second check is required because the BTT map error flag for a map entry only gets set when an error LBA is read. If we write to a new location that may not have the map error flag set, but still might be in the region's badblock list, we can trigger an EIO on the write, which is undesirable and completely avoidable. Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Toshi Kani <toshi.kani@hpe.com> Cc: Dan Williams <dan.j.williams@intel.com> Signed-off-by: Vishal Verma <vishal.l.verma@intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2017-08-31 09:36:03 +08:00
ret = btt_map_read(arena, premap, &old_postmap, NULL, &e_flag,
NVDIMM_IO_ATOMIC);
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
if (ret)
goto out_map;
if (old_postmap >= arena->internal_nlba) {
ret = -EIO;
goto out_map;
}
libnvdimm, btt: rework error clearing Clearing errors or badblocks during a BTT write requires sending an ACPI DSM, which means potentially sleeping. Since a BTT IO happens in atomic context (preemption disabled, spinlocks may be held), we cannot perform error clearing in the course of an IO. Due to this error clearing for BTT IOs has hitherto been disabled. In this patch we move error clearing out of the atomic section, and thus re-enable error clearing with BTTs. When we are about to add a block to the free list, we check if it was previously marked as an error, and if it was, we add it to the freelist, but also set a flag that says error clearing will be required. We then drop the lane (ending the atomic context), and send a zero buffer so that the error can be cleared. The error flag in the free list is protected by the nd 'lane', and is set only be a thread while it holds that lane. When the error is cleared, the flag is cleared, but while holding a mutex for that freelist index. When writing, we check for two things - 1/ If the freelist mutex is held or if the error flag is set. If so, this is an error block that is being (or about to be) cleared. 2/ If the block is a known badblock based on nsio->bb The second check is required because the BTT map error flag for a map entry only gets set when an error LBA is read. If we write to a new location that may not have the map error flag set, but still might be in the region's badblock list, we can trigger an EIO on the write, which is undesirable and completely avoidable. Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Toshi Kani <toshi.kani@hpe.com> Cc: Dan Williams <dan.j.williams@intel.com> Signed-off-by: Vishal Verma <vishal.l.verma@intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2017-08-31 09:36:03 +08:00
if (e_flag)
set_e_flag(old_postmap);
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
log.lba = cpu_to_le32(premap);
log.old_map = cpu_to_le32(old_postmap);
log.new_map = cpu_to_le32(new_postmap);
log.seq = cpu_to_le32(arena->freelist[lane].seq);
sub = arena->freelist[lane].sub;
ret = btt_flog_write(arena, lane, sub, &log);
if (ret)
goto out_map;
ret = btt_map_write(arena, premap, new_postmap, 0, 0,
NVDIMM_IO_ATOMIC);
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
if (ret)
goto out_map;
unlock_map(arena, premap);
nd_region_release_lane(btt->nd_region, lane);
libnvdimm, btt: rework error clearing Clearing errors or badblocks during a BTT write requires sending an ACPI DSM, which means potentially sleeping. Since a BTT IO happens in atomic context (preemption disabled, spinlocks may be held), we cannot perform error clearing in the course of an IO. Due to this error clearing for BTT IOs has hitherto been disabled. In this patch we move error clearing out of the atomic section, and thus re-enable error clearing with BTTs. When we are about to add a block to the free list, we check if it was previously marked as an error, and if it was, we add it to the freelist, but also set a flag that says error clearing will be required. We then drop the lane (ending the atomic context), and send a zero buffer so that the error can be cleared. The error flag in the free list is protected by the nd 'lane', and is set only be a thread while it holds that lane. When the error is cleared, the flag is cleared, but while holding a mutex for that freelist index. When writing, we check for two things - 1/ If the freelist mutex is held or if the error flag is set. If so, this is an error block that is being (or about to be) cleared. 2/ If the block is a known badblock based on nsio->bb The second check is required because the BTT map error flag for a map entry only gets set when an error LBA is read. If we write to a new location that may not have the map error flag set, but still might be in the region's badblock list, we can trigger an EIO on the write, which is undesirable and completely avoidable. Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Toshi Kani <toshi.kani@hpe.com> Cc: Dan Williams <dan.j.williams@intel.com> Signed-off-by: Vishal Verma <vishal.l.verma@intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2017-08-31 09:36:03 +08:00
if (e_flag) {
ret = arena_clear_freelist_error(arena, lane);
if (ret)
return ret;
}
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
len -= cur_len;
off += cur_len;
sector += btt->sector_size >> SECTOR_SHIFT;
}
return 0;
out_map:
unlock_map(arena, premap);
out_lane:
nd_region_release_lane(btt->nd_region, lane);
return ret;
}
static int btt_do_bvec(struct btt *btt, struct bio_integrity_payload *bip,
struct page *page, unsigned int len, unsigned int off,
unsigned int op, sector_t sector)
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
{
int ret;
if (!op_is_write(op)) {
ret = btt_read_pg(btt, bip, page, off, sector, len);
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
flush_dcache_page(page);
} else {
flush_dcache_page(page);
ret = btt_write_pg(btt, bip, sector, page, off, len);
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
}
return ret;
}
static blk_qc_t btt_submit_bio(struct bio *bio)
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
{
struct bio_integrity_payload *bip = bio_integrity(bio);
struct btt *btt = bio->bi_disk->private_data;
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
struct bvec_iter iter;
unsigned long start;
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
struct bio_vec bvec;
int err = 0;
bool do_acct;
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
if (!bio_integrity_prep(bio))
return BLK_QC_T_NONE;
do_acct = blk_queue_io_stat(bio->bi_disk->queue);
if (do_acct)
start = bio_start_io_acct(bio);
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
bio_for_each_segment(bvec, bio, iter) {
unsigned int len = bvec.bv_len;
if (len > PAGE_SIZE || len < btt->sector_size ||
len % btt->sector_size) {
dev_err_ratelimited(&btt->nd_btt->dev,
"unaligned bio segment (len: %d)\n", len);
bio->bi_status = BLK_STS_IOERR;
break;
}
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
err = btt_do_bvec(btt, bip, bvec.bv_page, len, bvec.bv_offset,
bio_op(bio), iter.bi_sector);
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
if (err) {
dev_err(&btt->nd_btt->dev,
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
"io error in %s sector %lld, len %d,\n",
(op_is_write(bio_op(bio))) ? "WRITE" :
"READ",
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
(unsigned long long) iter.bi_sector, len);
bio->bi_status = errno_to_blk_status(err);
break;
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
}
}
if (do_acct)
bio_end_io_acct(bio, start);
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
bio_endio(bio);
return BLK_QC_T_NONE;
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
}
static int btt_rw_page(struct block_device *bdev, sector_t sector,
struct page *page, unsigned int op)
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
{
struct btt *btt = bdev->bd_disk->private_data;
int rc;
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
rc = btt_do_bvec(btt, NULL, page, thp_size(page), 0, op, sector);
if (rc == 0)
page_endio(page, op_is_write(op), 0);
return rc;
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
}
static int btt_getgeo(struct block_device *bd, struct hd_geometry *geo)
{
/* some standard values */
geo->heads = 1 << 6;
geo->sectors = 1 << 5;
geo->cylinders = get_capacity(bd->bd_disk) >> 11;
return 0;
}
static const struct block_device_operations btt_fops = {
.owner = THIS_MODULE,
.submit_bio = btt_submit_bio,
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
.rw_page = btt_rw_page,
.getgeo = btt_getgeo,
};
static int btt_blk_init(struct btt *btt)
{
struct nd_btt *nd_btt = btt->nd_btt;
struct nd_namespace_common *ndns = nd_btt->ndns;
/* create a new disk and request queue for btt */
btt->btt_queue = blk_alloc_queue(NUMA_NO_NODE);
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
if (!btt->btt_queue)
return -ENOMEM;
btt->btt_disk = alloc_disk(0);
if (!btt->btt_disk) {
blk_cleanup_queue(btt->btt_queue);
return -ENOMEM;
}
nvdimm_namespace_disk_name(ndns, btt->btt_disk->disk_name);
btt->btt_disk->first_minor = 0;
btt->btt_disk->fops = &btt_fops;
btt->btt_disk->private_data = btt;
btt->btt_disk->queue = btt->btt_queue;
btt->btt_disk->flags = GENHD_FL_EXT_DEVT;
blk_queue_logical_block_size(btt->btt_queue, btt->sector_size);
blk_queue_max_hw_sectors(btt->btt_queue, UINT_MAX);
blk_queue_flag_set(QUEUE_FLAG_NONROT, btt->btt_queue);
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
if (btt_meta_size(btt)) {
int rc = nd_integrity_init(btt->btt_disk, btt_meta_size(btt));
if (rc) {
del_gendisk(btt->btt_disk);
put_disk(btt->btt_disk);
blk_cleanup_queue(btt->btt_queue);
return rc;
}
}
set_capacity(btt->btt_disk, btt->nlba * btt->sector_size >> 9);
device_add_disk(&btt->nd_btt->dev, btt->btt_disk, NULL);
btt->nd_btt->size = btt->nlba * (u64)btt->sector_size;
nvdimm_check_and_set_ro(btt->btt_disk);
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
return 0;
}
static void btt_blk_cleanup(struct btt *btt)
{
del_gendisk(btt->btt_disk);
put_disk(btt->btt_disk);
blk_cleanup_queue(btt->btt_queue);
}
/**
* btt_init - initialize a block translation table for the given device
* @nd_btt: device with BTT geometry and backing device info
* @rawsize: raw size in bytes of the backing device
* @lbasize: lba size of the backing device
* @uuid: A uuid for the backing device - this is stored on media
* @maxlane: maximum number of parallel requests the device can handle
*
* Initialize a Block Translation Table on a backing device to provide
* single sector power fail atomicity.
*
* Context:
* Might sleep.
*
* Returns:
* Pointer to a new struct btt on success, NULL on failure.
*/
static struct btt *btt_init(struct nd_btt *nd_btt, unsigned long long rawsize,
u32 lbasize, u8 *uuid, struct nd_region *nd_region)
{
int ret;
struct btt *btt;
libnvdimm, btt: rework error clearing Clearing errors or badblocks during a BTT write requires sending an ACPI DSM, which means potentially sleeping. Since a BTT IO happens in atomic context (preemption disabled, spinlocks may be held), we cannot perform error clearing in the course of an IO. Due to this error clearing for BTT IOs has hitherto been disabled. In this patch we move error clearing out of the atomic section, and thus re-enable error clearing with BTTs. When we are about to add a block to the free list, we check if it was previously marked as an error, and if it was, we add it to the freelist, but also set a flag that says error clearing will be required. We then drop the lane (ending the atomic context), and send a zero buffer so that the error can be cleared. The error flag in the free list is protected by the nd 'lane', and is set only be a thread while it holds that lane. When the error is cleared, the flag is cleared, but while holding a mutex for that freelist index. When writing, we check for two things - 1/ If the freelist mutex is held or if the error flag is set. If so, this is an error block that is being (or about to be) cleared. 2/ If the block is a known badblock based on nsio->bb The second check is required because the BTT map error flag for a map entry only gets set when an error LBA is read. If we write to a new location that may not have the map error flag set, but still might be in the region's badblock list, we can trigger an EIO on the write, which is undesirable and completely avoidable. Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Toshi Kani <toshi.kani@hpe.com> Cc: Dan Williams <dan.j.williams@intel.com> Signed-off-by: Vishal Verma <vishal.l.verma@intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2017-08-31 09:36:03 +08:00
struct nd_namespace_io *nsio;
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
struct device *dev = &nd_btt->dev;
btt = devm_kzalloc(dev, sizeof(struct btt), GFP_KERNEL);
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
if (!btt)
return NULL;
btt->nd_btt = nd_btt;
btt->rawsize = rawsize;
btt->lbasize = lbasize;
btt->sector_size = ((lbasize >= 4096) ? 4096 : 512);
INIT_LIST_HEAD(&btt->arena_list);
mutex_init(&btt->init_lock);
btt->nd_region = nd_region;
libnvdimm, btt: rework error clearing Clearing errors or badblocks during a BTT write requires sending an ACPI DSM, which means potentially sleeping. Since a BTT IO happens in atomic context (preemption disabled, spinlocks may be held), we cannot perform error clearing in the course of an IO. Due to this error clearing for BTT IOs has hitherto been disabled. In this patch we move error clearing out of the atomic section, and thus re-enable error clearing with BTTs. When we are about to add a block to the free list, we check if it was previously marked as an error, and if it was, we add it to the freelist, but also set a flag that says error clearing will be required. We then drop the lane (ending the atomic context), and send a zero buffer so that the error can be cleared. The error flag in the free list is protected by the nd 'lane', and is set only be a thread while it holds that lane. When the error is cleared, the flag is cleared, but while holding a mutex for that freelist index. When writing, we check for two things - 1/ If the freelist mutex is held or if the error flag is set. If so, this is an error block that is being (or about to be) cleared. 2/ If the block is a known badblock based on nsio->bb The second check is required because the BTT map error flag for a map entry only gets set when an error LBA is read. If we write to a new location that may not have the map error flag set, but still might be in the region's badblock list, we can trigger an EIO on the write, which is undesirable and completely avoidable. Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Toshi Kani <toshi.kani@hpe.com> Cc: Dan Williams <dan.j.williams@intel.com> Signed-off-by: Vishal Verma <vishal.l.verma@intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2017-08-31 09:36:03 +08:00
nsio = to_nd_namespace_io(&nd_btt->ndns->dev);
btt->phys_bb = &nsio->bb;
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
ret = discover_arenas(btt);
if (ret) {
dev_err(dev, "init: error in arena_discover: %d\n", ret);
return NULL;
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
}
if (btt->init_state != INIT_READY && nd_region->ro) {
dev_warn(dev, "%s is read-only, unable to init btt metadata\n",
dev_name(&nd_region->dev));
return NULL;
} else if (btt->init_state != INIT_READY) {
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
btt->num_arenas = (rawsize / ARENA_MAX_SIZE) +
((rawsize % ARENA_MAX_SIZE) ? 1 : 0);
dev_dbg(dev, "init: %d arenas for %llu rawsize\n",
btt->num_arenas, rawsize);
ret = create_arenas(btt);
if (ret) {
dev_info(dev, "init: create_arenas: %d\n", ret);
return NULL;
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
}
ret = btt_meta_init(btt);
if (ret) {
dev_err(dev, "init: error in meta_init: %d\n", ret);
return NULL;
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
}
}
ret = btt_blk_init(btt);
if (ret) {
dev_err(dev, "init: error in blk_init: %d\n", ret);
return NULL;
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
}
btt_debugfs_init(btt);
return btt;
}
/**
* btt_fini - de-initialize a BTT
* @btt: the BTT handle that was generated by btt_init
*
* De-initialize a Block Translation Table on device removal
*
* Context:
* Might sleep.
*/
static void btt_fini(struct btt *btt)
{
if (btt) {
btt_blk_cleanup(btt);
free_arenas(btt);
debugfs_remove_recursive(btt->debugfs_dir);
}
}
int nvdimm_namespace_attach_btt(struct nd_namespace_common *ndns)
{
struct nd_btt *nd_btt = to_nd_btt(ndns->claim);
struct nd_region *nd_region;
struct btt_sb *btt_sb;
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
struct btt *btt;
size_t size, rawsize;
int rc;
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
if (!nd_btt->uuid || !nd_btt->ndns || !nd_btt->lbasize) {
dev_dbg(&nd_btt->dev, "incomplete btt configuration\n");
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
return -ENODEV;
}
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
btt_sb = devm_kzalloc(&nd_btt->dev, sizeof(*btt_sb), GFP_KERNEL);
if (!btt_sb)
return -ENOMEM;
size = nvdimm_namespace_capacity(ndns);
rc = devm_namespace_enable(&nd_btt->dev, ndns, size);
if (rc)
return rc;
/*
* If this returns < 0, that is ok as it just means there wasn't
* an existing BTT, and we're creating a new one. We still need to
* call this as we need the version dependent fields in nd_btt to be
* set correctly based on the holder class
*/
nd_btt_version(nd_btt, ndns, btt_sb);
rawsize = size - nd_btt->initial_offset;
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
if (rawsize < ARENA_MIN_SIZE) {
dev_dbg(&nd_btt->dev, "%s must be at least %ld bytes\n",
dev_name(&ndns->dev),
ARENA_MIN_SIZE + nd_btt->initial_offset);
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
return -ENXIO;
}
nd_region = to_nd_region(nd_btt->dev.parent);
btt = btt_init(nd_btt, rawsize, nd_btt->lbasize, nd_btt->uuid,
nd_region);
if (!btt)
return -ENOMEM;
nd_btt->btt = btt;
return 0;
}
EXPORT_SYMBOL(nvdimm_namespace_attach_btt);
int nvdimm_namespace_detach_btt(struct nd_btt *nd_btt)
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
{
struct btt *btt = nd_btt->btt;
btt_fini(btt);
nd_btt->btt = NULL;
return 0;
}
EXPORT_SYMBOL(nvdimm_namespace_detach_btt);
static int __init nd_btt_init(void)
{
int rc = 0;
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
debugfs_root = debugfs_create_dir("btt", NULL);
if (IS_ERR_OR_NULL(debugfs_root))
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 16:20:32 +08:00
rc = -ENXIO;
return rc;
}
static void __exit nd_btt_exit(void)
{
debugfs_remove_recursive(debugfs_root);
}
MODULE_ALIAS_ND_DEVICE(ND_DEVICE_BTT);
MODULE_AUTHOR("Vishal Verma <vishal.l.verma@linux.intel.com>");
MODULE_LICENSE("GPL v2");
module_init(nd_btt_init);
module_exit(nd_btt_exit);