Obviously, IO serialization could cause the degradation of
performance a lot. In order to reduce the degradation, so a
rb interval tree is added in raid1 to speed up the check of
collision.
So, a rb root is needed in md_rdev, then abstract all the
serialize related members to a new struct (serial_in_rdev),
embed it into md_rdev.
Of course, we need to free the struct if it is not needed
anymore, so rdev/rdevs_uninit_serial are added accordingly.
And they should be called when destroty memory pool or can't
alloc memory.
And we need to consider to call mddev_destroy_serial_pool
in case serialize_policy/write-behind is disabled, bitmap
is destroyed or in __md_stop_writes.
Signed-off-by: Guoqing Jiang <guoqing.jiang@cloud.ionos.com>
Signed-off-by: Song Liu <songliubraving@fb.com>
With the new sysfs node, we can use it to control if raid1 array
wants io serialization or not. So mddev_create_serial_pool and
mddev_destroy_serial_pool are called in serialize_policy_store
to enable or disable the serialization.
Signed-off-by: Guoqing Jiang <guoqing.jiang@cloud.ionos.com>
Signed-off-by: Song Liu <songliubraving@fb.com>
1. The related resources (spin_lock, list and waitqueue) are needed for
address raid1 reorder overlap issue too, in this case, rdev is set to
NULL for mddev_create/destroy_serial_pool which implies all rdevs need
to handle these resources.
And also add "is_suspend" to mddev_destroy_serial_pool since it will
be called under suspended situation, which also makes both create and
destroy pool have same arguments.
2. Introduce rdevs_init_serial which is called if raid1 io serialization
is enabled since all rdevs need to init related stuffs.
3. rdev_init_serial and clear_bit(CollisionCheck, &rdev->flags) should
be called between suspend and resume.
No need to export mddev_create_serial_pool since it is only called in
md-mod module.
Signed-off-by: Guoqing Jiang <guoqing.jiang@cloud.ionos.com>
Signed-off-by: Song Liu <songliubraving@fb.com>
Previously, wb_info_pool and wb_list stuffs are introduced
to address potential data inconsistence issue for write
behind device.
Now rename them to serial related name, since the same
mechanism will be used to address reorder overlap write
issue for raid1.
Signed-off-by: Guoqing Jiang <guoqing.jiang@cloud.ionos.com>
Signed-off-by: Song Liu <songliubraving@fb.com>
If pers->make_request fails in md_flush_request(), the bio is lost. To
fix this, pass back a bool to indicate if the original make_request call
should continue to handle the I/O and instead of assuming the flush logic
will push it to completion.
Convert md_flush_request to return a bool and no longer calls the raid
driver's make_request function. If the return is true, then the md flush
logic has or will complete the bio and the md make_request call is done.
If false, then the md make_request function needs to keep processing like
it is a normal bio. Let the original call to md_handle_request handle any
need to retry sending the bio to the raid driver's make_request function
should it be needed.
Also mark md_flush_request and the make_request function pointer as
__must_check to issue warnings should these critical return values be
ignored.
Fixes: 2bc13b83e6 ("md: batch flush requests.")
Cc: stable@vger.kernel.org # # v4.19+
Cc: NeilBrown <neilb@suse.com>
Signed-off-by: David Jeffery <djeffery@redhat.com>
Reviewed-by: Xiao Ni <xni@redhat.com>
Signed-off-by: Song Liu <songliubraving@fb.com>
Currently md raid0/linear are not provided with any mechanism to validate
if an array member got removed or failed. The driver keeps sending BIOs
regardless of the state of array members, and kernel shows state 'clean'
in the 'array_state' sysfs attribute. This leads to the following
situation: if a raid0/linear array member is removed and the array is
mounted, some user writing to this array won't realize that errors are
happening unless they check dmesg or perform one fsync per written file.
Despite udev signaling the member device is gone, 'mdadm' cannot issue the
STOP_ARRAY ioctl successfully, given the array is mounted.
In other words, no -EIO is returned and writes (except direct ones) appear
normal. Meaning the user might think the wrote data is correctly stored in
the array, but instead garbage was written given that raid0 does stripping
(and so, it requires all its members to be working in order to not corrupt
data). For md/linear, writes to the available members will work fine, but
if the writes go to the missing member(s), it'll cause a file corruption
situation, whereas the portion of the writes to the missing devices aren't
written effectively.
This patch changes this behavior: we check if the block device's gendisk
is UP when submitting the BIO to the array member, and if it isn't, we flag
the md device as MD_BROKEN and fail subsequent I/Os to that device; a read
request to the array requiring data from a valid member is still completed.
While flagging the device as MD_BROKEN, we also show a rate-limited warning
in the kernel log.
A new array state 'broken' was added too: it mimics the state 'clean' in
every aspect, being useful only to distinguish if the array has some member
missing. We rely on the MD_BROKEN flag to put the array in the 'broken'
state. This state cannot be written in 'array_state' as it just shows
one or more members of the array are missing but acts like 'clean', it
wouldn't make sense to write it.
With this patch, the filesystem reacts much faster to the event of missing
array member: after some I/O errors, ext4 for instance aborts the journal
and prevents corruption. Without this change, we're able to keep writing
in the disk and after a machine reboot, e2fsck shows some severe fs errors
that demand fixing. This patch was tested in ext4 and xfs filesystems, and
requires a 'mdadm' counterpart to handle the 'broken' state.
Cc: Song Liu <songliubraving@fb.com>
Reviewed-by: NeilBrown <neilb@suse.de>
Signed-off-by: Guilherme G. Piccoli <gpiccoli@canonical.com>
Signed-off-by: Song Liu <songliubraving@fb.com>
Until revalidate_disk() has completed, the size of a new md array will
appear to be zero.
So we shouldn't report, through array_state, that the array is active
until that time.
udev rules check array_state to see if the array is ready. As soon as
it appear to be zero, fsck can be run. If it find the size to be
zero, it will fail.
So add a new flag to provide an interlock between do_md_run() and
array_state_show(). This flag is set while do_md_run() is active and
it prevents array_state_show() from reporting that the array is
active.
Before do_md_run() is called, ->pers will be NULL so array is
definitely not active.
After do_md_run() is called, revalidate_disk() will have run and the
array will be completely ready.
We also move various sysfs_notify*() calls out of md_run() into
do_md_run() after MD_NOT_READY is cleared. This ensure the
information is ready before the notification is sent.
Prior to v4.12, array_state_show() was called with the
mddev->reconfig_mutex held, which provided exclusion with do_md_run().
Note that MD_NOT_READY cleared twice. This is deliberate to cover
both success and error paths with minimal noise.
Fixes: b7b17c9b67 ("md: remove mddev_lock() from md_attr_show()")
Cc: stable@vger.kernel.org (v4.12++)
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: Song Liu <songliubraving@fb.com>
When the 'last' device in a RAID1 or RAID10 reports an error,
we do not mark it as failed. This would serve little purpose
as there is no risk of losing data beyond that which is obviously
lost (as there is with RAID5), and there could be other sectors
on the device which are readable, and only readable from this device.
This in general this maximises access to data.
However the current implementation also stops an admin from removing
the last device by direct action. This is rarely useful, but in many
case is not harmful and can make automation easier by removing special
cases.
Also, if an attempt to write metadata fails the device must be marked
as faulty, else an infinite loop will result, attempting to update
the metadata on all non-faulty devices.
So add 'fail_last_dev' member to 'struct mddev', then we can bypasses
the 'last disk' checks for RAID1 and RAID10, and control the behavior
per array by change sysfs node.
Signed-off-by: NeilBrown <neilb@suse.de>
[add sysfs node for fail_last_dev by Guoqing]
Signed-off-by: Guoqing Jiang <guoqing.jiang@cloud.ionos.com>
Signed-off-by: Song Liu <songliubraving@fb.com>
Previously, we called rdev_init_wb to avoid potential data
inconsistency when array is created.
Now, we need to call the function and create mempool if a
device is added or just be flaged as "writemostly". So
mddev_create_wb_pool is introduced and called accordingly.
And for safety reason, we mark implicit GFP_NOIO allocation
scope for create mempool during mddev_suspend/mddev_resume.
And mempool should be removed conversely after remove a
member device or its's "writemostly" flag, which is done
by call mddev_destroy_wb_pool.
Reviewed-by: NeilBrown <neilb@suse.com>
Signed-off-by: Guoqing Jiang <gqjiang@suse.com>
Signed-off-by: Song Liu <songliubraving@fb.com>
For write-behind mode, we think write IO is complete once it has
reached all the non-writemostly devices. It works fine for single
queue devices.
But for multiqueue device, if there are lots of IOs come from upper
layer, then the write-behind device could issue those IOs to different
queues, depends on the each queue's delay, so there is no guarantee
that those IOs can arrive in order.
To address the issue, we need to check the collision among write
behind IOs, we can only continue without collision, otherwise wait
for the completion of previous collisioned IO.
And WBCollision is introduced for multiqueue device which is worked
under write-behind mode.
But this patch doesn't handle below cases which could have the data
inconsistency issue as well, these cases will be handled in later
patches.
1. modify max_write_behind by write backlog node.
2. add or remove array's bitmap dynamically.
3. the change of member disk.
Reviewed-by: NeilBrown <neilb@suse.com>
Signed-off-by: Guoqing Jiang <gqjiang@suse.com>
Signed-off-by: Song Liu <songliubraving@fb.com>
Based on 1 normalized pattern(s):
this program is free software you can redistribute it and or modify
it under the terms of the gnu general public license as published by
the free software foundation either version 2 or at your option any
later version you should have received a copy of the gnu general
public license for example usr src linux copying if not write to the
free software foundation inc 675 mass ave cambridge ma 02139 usa
extracted by the scancode license scanner the SPDX license identifier
GPL-2.0-or-later
has been chosen to replace the boilerplate/reference in 20 file(s).
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Reviewed-by: Allison Randal <allison@lohutok.net>
Reviewed-by: Kate Stewart <kstewart@linuxfoundation.org>
Cc: linux-spdx@vger.kernel.org
Link: https://lkml.kernel.org/r/20190520170858.552543146@linutronix.de
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
This reverts commit 5a409b4f56.
This patch has two problems.
1/ it make multiple calls to submit_bio() from inside a make_request_fn.
The bios thus submitted will be queued on current->bio_list and not
submitted immediately. As the bios are allocated from a mempool,
this can theoretically result in a deadlock - all the pool of requests
could be in various ->bio_list queues and a subsequent mempool_alloc
could block waiting for one of them to be released.
2/ It aims to handle a case when there are many concurrent flush requests.
It handles this by submitting many requests in parallel - all of which
are identical and so most of which do nothing useful.
It would be more efficient to just send one lower-level request, but
allow that to satisfy multiple upper-level requests.
Fixes: 5a409b4f56 ("MD: fix lock contention for flush bios")
Cc: <stable@vger.kernel.org> # v4.19+
Tested-by: Xiao Ni <xni@redhat.com>
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: Song Liu <songliubraving@fb.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
For clustered raid10 scenario, we need to let all the nodes
know about that a new disk is added to the array, and the
reshape caused by add new member just need to be happened in
one node, but other nodes should know about the change.
Since reshape means read data from somewhere (which is already
used by array) and write data to unused region. Obviously, it
is awful if one node is reading data from address while another
node is writing to the same address. Considering we have
implemented suspend writes in the resyncing area, so we can
just broadcast the reading address to other nodes to avoid the
trouble.
For master node, it would call reshape_request then update sb
during the reshape period. To avoid above trouble, we call
resync_info_update to send RESYNC message in reshape_request.
Then from slave node's view, it receives two type messages:
1. RESYNCING message
Slave node add the address (where master node reading data from)
to suspend list.
2. METADATA_UPDATED message
Once slave nodes know the reshaping is started in master node,
it is time to update reshape position and call start_reshape to
follow master node's step. After reshape is done, only reshape
position is need to be updated, so the majority task of reshaping
is happened on the master node.
Reviewed-by: NeilBrown <neilb@suse.com>
Signed-off-by: Guoqing Jiang <gqjiang@suse.com>
Signed-off-by: Shaohua Li <shli@fb.com>
When resync or recovery is happening in one node,
other nodes don't show the appropriate info now.
For example, when create an array in master node
without "--assume-clean", then assemble the array
in slave nodes, you can see "resync=PENDING" when
read /proc/mdstat in slave nodes. However, the info
is confusing since "PENDING" status is introduced
for start array in read-only mode.
We introduce RESYNCING_REMOTE flag to indicate that
resync thread is running in remote node. The flags
is set when node receive RESYNCING msg. And we clear
the REMOTE flag in following cases:
1. resync or recover is finished in master node,
which means slaves receive msg with both lo
and hi are set to 0.
2. node continues resync/recovery in recover_bitmaps.
3. when resync_finish is called.
Then we show accurate information in status_resync
by check REMOTE flags and with other conditions.
Signed-off-by: Guoqing Jiang <gqjiang@suse.com>
Reviewed-by: NeilBrown <neilb@suse.com>
Signed-off-by: Shaohua Li <shli@fb.com>
Pull MD updates from Shaohua Li:
"A few fixes of MD for this merge window. Mostly bug fixes:
- raid5 stripe batch fix from Amy
- Read error handling for raid1 FailFast device from Gioh
- raid10 recovery NULL pointer dereference fix from Guoqing
- Support write hint for raid5 stripe cache from Mariusz
- Fixes for device hot add/remove from Neil and Yufen
- Improve flush bio scalability from Xiao"
* 'for-next' of git://git.kernel.org/pub/scm/linux/kernel/git/shli/md:
MD: fix lock contention for flush bios
md/raid5: Assigning NULL to sh->batch_head before testing bit R5_Overlap of a stripe
md/raid1: add error handling of read error from FailFast device
md: fix NULL dereference of mddev->pers in remove_and_add_spares()
raid5: copy write hint from origin bio to stripe
md: fix two problems with setting the "re-add" device state.
raid10: check bio in r10buf_pool_free to void NULL pointer dereference
md: fix an error code format and remove unsed bio_sector
There is a lock contention when there are many processes which send flush bios
to md device. eg. Create many lvs on one raid device and mkfs.xfs on each lv.
Now it just can handle flush request sequentially. It needs to wait mddev->flush_bio
to be NULL, otherwise get mddev->lock.
This patch remove mddev->flush_bio and handle flush bio asynchronously.
I did a test with command dbench -s 128 -t 300. This is the test result:
=================Without the patch============================
Operation Count AvgLat MaxLat
--------------------------------------------------
Flush 11165 167.595 5879.560
Close 107469 1.391 2231.094
LockX 384 0.003 0.019
Rename 5944 2.141 1856.001
ReadX 208121 0.003 0.074
WriteX 98259 1925.402 15204.895
Unlink 25198 13.264 3457.268
UnlockX 384 0.001 0.009
FIND_FIRST 47111 0.012 0.076
SET_FILE_INFORMATION 12966 0.007 0.065
QUERY_FILE_INFORMATION 27921 0.004 0.085
QUERY_PATH_INFORMATION 124650 0.005 5.766
QUERY_FS_INFORMATION 22519 0.003 0.053
NTCreateX 141086 4.291 2502.812
Throughput 3.7181 MB/sec (sync open) 128 clients 128 procs max_latency=15204.905 ms
=================With the patch============================
Operation Count AvgLat MaxLat
--------------------------------------------------
Flush 4500 174.134 406.398
Close 48195 0.060 467.062
LockX 256 0.003 0.029
Rename 2324 0.026 0.360
ReadX 78846 0.004 0.504
WriteX 66832 562.775 1467.037
Unlink 5516 3.665 1141.740
UnlockX 256 0.002 0.019
FIND_FIRST 16428 0.015 0.313
SET_FILE_INFORMATION 6400 0.009 0.520
QUERY_FILE_INFORMATION 17865 0.003 0.089
QUERY_PATH_INFORMATION 47060 0.078 416.299
QUERY_FS_INFORMATION 7024 0.004 0.032
NTCreateX 55921 0.854 1141.452
Throughput 11.744 MB/sec (sync open) 128 clients 128 procs max_latency=1467.041 ms
The test is done on raid1 disk with two rotational disks
V5: V4 is more complicated than the version with memory pool. So revert to the memory pool
version
V4: use address of fbio to do hash to choose free flush info.
V3:
Shaohua suggests mempool is overkill. In v3 it allocs memory during creating raid device
and uses a simple bitmap to record which resource is free.
Fix a bug from v2. It should set flush_pending to 1 at first.
V2:
Neil pointed out two problems. One is counting error problem and another is return value
when allocat memory fails.
1. counting error problem
This isn't safe. It is only safe to call rdev_dec_pending() on rdevs
that you previously called
atomic_inc(&rdev->nr_pending);
If an rdev was added to the list between the start and end of the flush,
this will do something bad.
Now it doesn't use bio_chain. It uses specified call back function for each
flush bio.
2. Returned on IO error when kmalloc fails is wrong.
I use mempool suggested by Neil in V2
3. Fixed some places pointed by Guoqing
Suggested-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Xiao Ni <xni@redhat.com>
Signed-off-by: Shaohua Li <shli@fb.com>
If no metadata devices are configured on raid1/4/5/6/10
(e.g. via dm-raid), md_write_start() unconditionally waits
for superblocks to be written thus deadlocking.
Fix introduces mddev->has_superblocks bool, defines it in md_run()
and checks for it in md_write_start() to conditionally avoid waiting.
Once on it, check for non-existing superblocks in md_super_write().
Link: https://bugzilla.kernel.org/show_bug.cgi?id=198647
Fixes: cc27b0c78c ("md: fix deadlock between mddev_suspend() and md_write_start()")
Signed-off-by: Heinz Mauelshagen <heinzm@redhat.com>
Signed-off-by: Shaohua Li <sh.li@alibaba-inc.com>
In order to provide data consistency with PPL for disks with write-back
cache enabled all data has to be flushed to disks before next PPL
entry. The disks to be flushed are marked in the bitmap. It's modified
under a mutex and it's only read after PPL io unit is submitted.
A limitation of 64 disks in the array has been introduced to keep data
structures and implementation simple. RAID5 arrays with so many disks are
not likely due to high risk of multiple disks failure. Such restriction
should not be a real life limitation.
With write-back cache disabled next PPL entry is submitted when data write
for current one completes. Data flush defers next log submission so trigger
it when there are no stripes for handling found.
As PPL assures all data is flushed to disk at request completion, just
acknowledge flush request when PPL is enabled.
Signed-off-by: Tomasz Majchrzak <tomasz.majchrzak@intel.com>
Signed-off-by: Shaohua Li <sh.li@alibaba-inc.com>
In do_md_run(), md threads should not wake up until the array is fully
initialized in md_run(). However, in raid5_run(), raid5-cache may wake
up mddev->thread to flush stripes that need to be written back. This
design doesn't break badly right now. But it could lead to bad bug in
the future.
This patch tries to resolve this problem by splitting start up work
into two personality functions, run() and start(). Tasks that do not
require the md threads should go into run(), while task that require
the md threads go into start().
r5l_load_log() is moved to raid5_start(), so it is not called until
the md threads are started in do_md_run().
Signed-off-by: Song Liu <songliubraving@fb.com>
Signed-off-by: Shaohua Li <shli@fb.com>
The '2' argument means "wake up anything that is waiting".
This is an inelegant part of the design and was added
to help support management of suspend_lo/suspend_hi setting.
Now that suspend_lo/hi is managed in mddev_suspend/resume,
that need is gone.
These is still a couple of places where we call 'quiesce'
with an argument of '2', but they can safely be changed to
call ->quiesce(.., 1); ->quiesce(.., 0) which
achieve the same result at the small cost of pausing IO
briefly.
This removes a small "optimization" from suspend_{hi,lo}_store,
but it isn't clear that optimization served a useful purpose.
The code now is a lot clearer.
Suggested-by: Shaohua Li <shli@kernel.org>
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: Shaohua Li <shli@fb.com>
There are various deadlocks that can occur
when a thread holds reconfig_mutex and calls
->quiesce(mddev, 1).
As some write request block waiting for
metadata to be updated (e.g. to record device
failure), and as the md thread updates the metadata
while the reconfig mutex is held, holding the mutex
can stop write requests completing, and this prevents
->quiesce(mddev, 1) from completing.
->quiesce() is now usually called from mddev_suspend(),
and it is always called with reconfig_mutex held. So
at this time it is safe for the thread to update metadata
without explicitly taking the lock.
So add 2 new flags, one which says the unlocked updates is
allowed, and one which ways it is happening. Then allow it
while the quiesce completes, and then wait for it to finish.
Reported-and-tested-by: Xiao Ni <xni@redhat.com>
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: Shaohua Li <shli@fb.com>
With commit cc27b0c78c, pers->make_request could bail out without handling
the bio. If that happens, we should retry. The commit fixes md_make_request
but not other call sites. Separate the request handling part, so other call
sites can use it.
Reported-by: Nate Dailey <nate.dailey@stratus.com>
Fix: cc27b0c78c79(md: fix deadlock between mddev_suspend() and md_write_start())
Cc: stable@vger.kernel.org
Reviewed-by: NeilBrown <neilb@suse.com>
Signed-off-by: Shaohua Li <shli@fb.com>
Pull MD updates from Shaohua Li:
"This update mainly fixes bugs:
- Make raid5 ppl support several ppl from Pawel
- Several raid5-cache bug fixes from Song
- Bitmap fixes from Neil and Me
- One raid1/10 regression fix since 4.12 from Me
- Other small fixes and cleanup"
* tag 'md/4.14-rc1' of git://git.kernel.org/pub/scm/linux/kernel/git/shli/md:
md/bitmap: disable bitmap_resize for file-backed bitmaps.
raid5-ppl: Recovery support for multiple partial parity logs
md: Runtime support for multiple ppls
md/raid0: attach correct cgroup info in bio
lib/raid6: align AVX512 constants to 512 bits, not bytes
raid5: remove raid5_build_block
md/r5cache: call mddev_lock/unlock() in r5c_journal_mode_show
md: replace seq_release_private with seq_release
md: notify about new spare disk in the container
md/raid1/10: reset bio allocated from mempool
md/raid5: release/flush io in raid5_do_work()
md/bitmap: copy correct data for bitmap super
Increase PPL area to 1MB and use it as circular buffer to store PPL. The
entry with highest generation number is the latest one. If PPL to be
written is larger then space left in a buffer, rewind the buffer to the
start (don't wrap it).
Signed-off-by: Pawel Baldysiak <pawel.baldysiak@intel.com>
Signed-off-by: Artur Paszkiewicz <artur.paszkiewicz@intel.com>
Signed-off-by: Shaohua Li <shli@fb.com>
This way we don't need a block_device structure to submit I/O. The
block_device has different life time rules from the gendisk and
request_queue and is usually only available when the block device node
is open. Other callers need to explicitly create one (e.g. the lightnvm
passthrough code, or the new nvme multipathing code).
For the actual I/O path all that we need is the gendisk, which exists
once per block device. But given that the block layer also does
partition remapping we additionally need a partition index, which is
used for said remapping in generic_make_request.
Note that all the block drivers generally want request_queue or
sometimes the gendisk, so this removes a layer of indirection all
over the stack.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
No function change, just move 'struct resync_pages' and related
helpers into raid1-10.c
Signed-off-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Shaohua Li <shli@fb.com>
bio_add_page() won't fail for resync bio, and the page index for each
bio is same, so remove it.
More importantly the 'idx' of 'struct resync_pages' is initialized in
mempool allocator function, the current way is wrong since mempool is
only responsible for allocation, we can't use that for initialization.
Suggested-by: NeilBrown <neilb@suse.com>
Reported-by: NeilBrown <neilb@suse.com>
Reported-and-tested-by: Patrick <dto@gmx.net>
Fixes: f0250618361d(md: raid10: don't use bio's vec table to manage resync pages)
Fixes: 98d30c5812c3(md: raid1: don't use bio's vec table to manage resync pages)
Cc: stable@vger.kernel.org (4.12+)
Signed-off-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Shaohua Li <shli@fb.com>
The device owns Bitmap_sync flag needs recovery
to become in sync, and read page from this type
device could get stale status.
Also add comments for Bitmap_sync bit per the
suggestion from Shaohua and Neil.
Previous disscussion can be found here:
https://marc.info/?t=149760428900004&r=1&w=2
Signed-off-by: Guoqing Jiang <gqjiang@suse.com>
Signed-off-by: Shaohua Li <shli@fb.com>
md devices allocate a bio_set and use it for two
distinct purposes.
mddev->bio_set is used to clone bios as part of sending
upper level requests down to lower level devices,
and it is also use for synchronous IO such as superblock
and bitmap updates, and for correcting read errors.
This multiple usage can lead to deadlocks. It is likely
that cloned bios might be queued for write and to be
waiting for a metadata update before the write can be permitted.
If the cloning exhausted mddev->bio_set, the metadata update
may not be able to proceed.
This scenario has been seen during heavy testing, with lots of IO and
lots of memory pressure.
Address this by adding a new bio_set specifically for synchronous IO.
All synchronous IO goes directly to the underlying device and is not
queued at the md level, so request using entries from the new
mddev->sync_set will complete in a timely fashion.
Requests that use mddev->bio_set will sometimes need to wait
for synchronous IO, but will no longer risk deadlocking that iO.
Also: small simplification in mddev_put(): there is no need to
wait until the spinlock is released before calling bioset_free().
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: Shaohua Li <shli@fb.com>
If mddev_suspend() races with md_write_start() we can deadlock
with mddev_suspend() waiting for the request that is currently
in md_write_start() to complete the ->make_request() call,
and md_write_start() waiting for the metadata to be updated
to mark the array as 'dirty'.
As metadata updates done by md_check_recovery() only happen then
the mddev_lock() can be claimed, and as mddev_suspend() is often
called with the lock held, these threads wait indefinitely for each
other.
We fix this by having md_write_start() abort if mddev_suspend()
is happening, and ->make_request() aborts if md_write_start()
aborted.
md_make_request() can detect this abort, decrease the ->active_io
count, and wait for mddev_suspend().
Reported-by: Nix <nix@esperi.org.uk>
Fix: 68866e425be2(MD: no sync IO while suspended)
Cc: stable@vger.kernel.org
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: Shaohua Li <shli@fb.com>
The new per-cpu counter for writes_pending is initialised in
md_alloc(), which is not called by dm-raid.
So dm-raid fails when md_write_start() is called.
Move the initialization to the personality modules
that need it. This way it is always initialised when needed,
but isn't unnecessarily initialized (requiring memory allocation)
when the personality doesn't use writes_pending.
Reported-by: Heinz Mauelshagen <heinzm@redhat.com>
Fixes: 4ad23a9764 ("MD: use per-cpu counter for writes_pending")
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: Shaohua Li <shli@fb.com>
This essentially reverts commit b5470dc5fc ("md: resolve external
metadata handling deadlock in md_allow_write") with some adjustments.
Since commit 6791875e2e ("md: make reconfig_mutex optional for writes
to md sysfs files.") changing array_state to 'active' does not use
mddev_lock() and will not cause a deadlock with md_allow_write(). This
revert simplifies userspace tools that write to sysfs attributes like
"stripe_cache_size" or "consistency_policy" because it removes the need
for special handling for external metadata arrays, checking for EAGAIN
and retrying the write.
Signed-off-by: Artur Paszkiewicz <artur.paszkiewicz@intel.com>
Signed-off-by: Shaohua Li <shli@fb.com>
Copy & paste from the REQ_OP_WRITE_SAME code.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Now resync I/O use bio's bec table to manage pages,
this way is very hacky, and may not work any more
once multipage bvec is introduced.
So introduce helpers and new data structure for
managing resync I/O pages more cleanly.
Signed-off-by: Ming Lei <tom.leiming@gmail.com>
Signed-off-by: Shaohua Li <shli@fb.com>
Both raid1 and raid10 share common resync
block size and page count, so move them into md.h.
Signed-off-by: Ming Lei <tom.leiming@gmail.com>
Signed-off-by: Shaohua Li <shli@fb.com>
The 'writes_pending' counter is used to determine when the
array is stable so that it can be marked in the superblock
as "Clean". Consequently it needs to be updated frequently
but only checked for zero occasionally. Recent changes to
raid5 cause the count to be updated even more often - once
per 4K rather than once per bio. This provided
justification for making the updates more efficient.
So we replace the atomic counter a percpu-refcount.
This can be incremented and decremented cheaply most of the
time, and can be switched to "atomic" mode when more
precise counting is needed. As it is possible for multiple
threads to want a precise count, we introduce a
"sync_checker" counter to count the number of threads
in "set_in_sync()", and only switch the refcount back
to percpu mode when that is zero.
We need to be careful about races between set_in_sync()
setting ->in_sync to 1, and md_write_start() setting it
to zero. md_write_start() holds the rcu_read_lock()
while checking if the refcount is in percpu mode. If
it is, then we know a switch to 'atomic' will not happen until
after we call rcu_read_unlock(), in which case set_in_sync()
will see the elevated count, and not set in_sync to 1.
If it is not in percpu mode, we take the mddev->lock to
ensure proper synchronization.
It is no longer possible to quickly check if the count is zero, which
we previously did to update a timer or to schedule the md_thread.
So now we do these every time we decrement that counter, but make
sure they are fast.
mod_timer() already optimizes the case where the timeout value doesn't
actually change. We leverage that further by always rounding off the
jiffies to the timeout value. This may delay the marking of 'clean'
slightly, but ensure we only perform atomic operation here when absolutely
needed.
md_wakeup_thread() current always calls wake_up(), even if
THREAD_WAKEUP is already set. That too can be optimised to avoid
calls to wake_up().
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: Shaohua Li <shli@fb.com>
We use md_write_start() to increase the count of pending writes, and
md_write_end() to decrement the count. We currently count bios
submitted to md/raid5. Change it count stripe_heads that a WRITE bio
has been attached to.
So now, raid5_make_request() calls md_write_start() and then
md_write_end() to keep the count elevated during the setup of the
request.
add_stripe_bio() calls md_write_start() for each stripe_head, and the
completion routines always call md_write_end(), instead of only
calling it when raid5_dec_bi_active_stripes() returns 0.
make_discard_request also calls md_write_start/end().
The parallel between md_write_{start,end} and use of bi_phys_segments
can be seen in that:
Whenever we set bi_phys_segments to 1, we now call md_write_start.
Whenever we increment it on non-read requests with
raid5_inc_bi_active_stripes(), we now call md_write_start().
Whenever we decrement bi_phys_segments on non-read requsts with
raid5_dec_bi_active_stripes(), we now call md_write_end().
This reduces our dependence on keeping a per-bio count of active
stripes in bi_phys_segments.
md_write_inc() is added which parallels md_write_start(), but requires
that a write has already been started, and is certain never to sleep.
This can be used inside a spinlocked region when adding to a write
request.
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: Shaohua Li <shli@fb.com>
Allow writing to 'consistency_policy' attribute when the array is
active. Add a new function 'change_consistency_policy' to the
md_personality operations structure to handle the change in the
personality code. Values "ppl" and "resync" are accepted and
turn PPL on and off respectively.
When enabling PPL its location and size should first be set using
'ppl_sector' and 'ppl_size' attributes and a valid PPL header should be
written at this location on each member device.
Enabling or disabling PPL is performed under a suspended array. The
raid5_reset_stripe_cache function frees the stripe cache and allocates
it again in order to allocate or free the ppl_pages for the stripes in
the stripe cache.
Signed-off-by: Artur Paszkiewicz <artur.paszkiewicz@intel.com>
Signed-off-by: Shaohua Li <shli@fb.com>
Include information about PPL location and size into mdp_superblock_1
and copy it to/from rdev. Because PPL is mutually exclusive with bitmap,
put it in place of 'bitmap_offset'. Add a new flag MD_FEATURE_PPL for
'feature_map', analogically to MD_FEATURE_BITMAP_OFFSET. Add MD_HAS_PPL
to mddev->flags to indicate that PPL is enabled on an array.
Signed-off-by: Artur Paszkiewicz <artur.paszkiewicz@intel.com>
Signed-off-by: Shaohua Li <shli@fb.com>
Previously, when node received METADATA_UPDATED msg, it just
need to wakeup mddev->thread, then md_reload_sb will be called
eventually.
We taken the asynchronous way to avoid a deadlock issue, the
deadlock issue could happen when one node is receiving the
METADATA_UPDATED msg (wants reconfig_mutex) and trying to run
the path:
md_check_recovery -> mddev_trylock(hold reconfig_mutex)
-> md_update_sb-metadata_update_start
(want EX on token however token is
got by the sending node)
Since we will support resizing for clustered raid, and we
need the metadata update handling to be synchronous so that
the initiating node can detect failure, so we need to change
the way for handling METADATA_UPDATED msg.
But, we obviously need to avoid above deadlock with the
sync way. To make this happen, we considered to not hold
reconfig_mutex to call md_reload_sb, if some other thread
has already taken reconfig_mutex and waiting for the 'token',
then process_recvd_msg() can safely call md_reload_sb()
without taking the mutex. This is because we can be certain
that no other thread will take the mutex, and we also certain
that the actions performed by md_reload_sb() won't interfere
with anything that the other thread is in the middle of.
To make this more concrete, we added a new cinfo->state bit
MD_CLUSTER_HOLDING_MUTEX_FOR_RECVD
Which is set in lock_token() just before dlm_lock_sync() is
called, and cleared just after. As lock_token() is always
called with reconfig_mutex() held (the specific case is the
resync_info_update which is distinguished well in previous
patch), if process_recvd_msg() finds that the new bit is set,
then the mutex must be held by some other thread, and it will
keep waiting.
So process_metadata_update() can call md_reload_sb() if either
mddev_trylock() succeeds, or if MD_CLUSTER_HOLDING_MUTEX_FOR_RECVD
is set. The tricky bit is what to do if neither of these apply.
We need to wait. Fortunately mddev_unlock() always calls wake_up()
on mddev->thread->wqueue. So we can get lock_token() to call
wake_up() on that when it sets the bit.
There are also some related changes inside this commit:
1. remove RELOAD_SB related codes since there are not valid anymore.
2. mddev is added into md_cluster_info then we can get mddev inside
lock_token.
3. add new parameter for lock_token to distinguish reconfig_mutex
is held or not.
And, we need to set MD_CLUSTER_HOLDING_MUTEX_FOR_RECVD in below:
1. set it before unregister thread, otherwise a deadlock could
appear if stop a resyncing array.
This is because md_unregister_thread(&cinfo->recv_thread) is
blocked by recv_daemon -> process_recvd_msg
-> process_metadata_update.
To resolve the issue, MD_CLUSTER_HOLDING_MUTEX_FOR_RECVD is
also need to be set before unregister thread.
2. set it in metadata_update_start to fix another deadlock.
a. Node A sends METADATA_UPDATED msg (held Token lock).
b. Node B wants to do resync, and is blocked since it can't
get Token lock, but MD_CLUSTER_HOLDING_MUTEX_FOR_RECVD is
not set since the callchain
(md_do_sync -> sync_request
-> resync_info_update
-> sendmsg
-> lock_comm -> lock_token)
doesn't hold reconfig_mutex.
c. Node B trys to update sb (held reconfig_mutex), but stopped
at wait_event() in metadata_update_start since we have set
MD_CLUSTER_SEND_LOCK flag in lock_comm (step 2).
d. Then Node B receives METADATA_UPDATED msg from A, of course
recv_daemon is blocked forever.
Since metadata_update_start always calls lock_token with reconfig_mutex,
we need to set MD_CLUSTER_HOLDING_MUTEX_FOR_RECVD here as well, and
lock_token don't need to set it twice unless lock_token is invoked from
lock_comm.
Finally, thanks to Neil for his great idea and help!
Reviewed-by: NeilBrown <neilb@suse.com>
Signed-off-by: Guoqing Jiang <gqjiang@suse.com>
Signed-off-by: Shaohua Li <shli@fb.com>
Firstly bio_clone_mddev() is used in raid normal I/O and isn't
in resync I/O path.
Secondly all the direct access to bvec table in raid happens on
resync I/O except for write behind of raid1, in which we still
use bio_clone() for allocating new bvec table.
So this patch replaces bio_clone() with bio_clone_fast()
in bio_clone_mddev().
Also kill bio_clone_mddev() and call bio_clone_fast() directly, as
suggested by Christoph Hellwig.
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Ming Lei <tom.leiming@gmail.com>
Signed-off-by: Shaohua Li <shli@fb.com>
This makes md do the same thing as dm for write same IO failure. Please
see 7eee4ae(dm: disable WRITE SAME if it fails) for details why we need
this.
We did a little bit different than dm. Instead of disabling writesame in
the first IO error, we disable it till next writesame IO coming after
the first IO error. This way we don't need to clone a bio.
Also reported here: https://bugzilla.kernel.org/show_bug.cgi?id=118581
Suggested-by: NeilBrown <neilb@suse.com>
Acked-by: NeilBrown <neilb@suse.com>
Signed-off-by: Shaohua Li <shli@fb.com>
Commit 6995f0b (md: takeover should clear unrelated bits) clear
unrelated bits, but it's quite fragile. To avoid error in the future,
define a macro for unsupported mddev flags for each raid type and use it
to clear unsupported mddev flags. This should be less error-prone.
Suggested-by: NeilBrown <neilb@suse.com>
Signed-off-by: Shaohua Li <shli@fb.com>
The mddev->flags are used for different purposes. There are a lot of
places we check/change the flags without masking unrelated flags, we
could check/change unrelated flags. These usage are most for superblock
write, so spearate superblock related flags. This should make the code
clearer and also fix real bugs.
Reviewed-by: NeilBrown <neilb@suse.com>
Signed-off-by: Shaohua Li <shli@fb.com>
This can only be supported on personalities which ensure
that md_error() never causes an array to enter the 'failed'
state. i.e. if marking a device Faulty would cause some
data to be inaccessible, the device is status is left as
non-Faulty. This is true for RAID1 and RAID10.
If we get a failure writing metadata but the device doesn't
fail, it must be the last device so we re-write without
FAILFAST to improve chance of success. We also flag the
device as LastDev so that future metadata updates don't
waste time on failfast writes.
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: Shaohua Li <shli@fb.com>
Guoqing Jiang