CPU hotplug handling for blk-mq (blk_mq_queue_reinit) acquires
all_q_mutex in blk_mq_queue_reinit_notify() and then removes sysfs
entries by blk_mq_sysfs_unregister(). Removing sysfs entry needs to
be blocked until the active reference of the kernfs_node to be zero.
On the other hand, reading blk_mq_hw_sysfs_cpu sysfs entry (e.g.
/sys/block/nullb0/mq/0/cpu_list) acquires all_q_mutex in
blk_mq_hw_sysfs_cpus_show().
If these happen at the same time, a deadlock can happen. Because one
can wait for the active reference to be zero with holding all_q_mutex,
and the other tries to acquire all_q_mutex with holding the active
reference.
The reason that all_q_mutex is acquired in blk_mq_hw_sysfs_cpus_show()
is to avoid reading an imcomplete hctx->cpumask. Since reading sysfs
entry for blk-mq needs to acquire q->sysfs_lock, we can avoid deadlock
and reading an imcomplete hctx->cpumask by protecting q->sysfs_lock
while hctx->cpumask is being updated.
Signed-off-by: Akinobu Mita <akinobu.mita@gmail.com>
Reviewed-by: Ming Lei <tom.leiming@gmail.com>
Cc: Ming Lei <tom.leiming@gmail.com>
Cc: Wanpeng Li <wanpeng.li@hotmail.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
There is a race between cpu hotplug handling and adding/deleting
gendisk for blk-mq, where both are trying to register and unregister
the same sysfs entries.
null_add_dev
--> blk_mq_init_queue
--> blk_mq_init_allocated_queue
--> add to 'all_q_list' (*)
--> add_disk
--> blk_register_queue
--> blk_mq_register_disk (++)
null_del_dev
--> del_gendisk
--> blk_unregister_queue
--> blk_mq_unregister_disk (--)
--> blk_cleanup_queue
--> blk_mq_free_queue
--> del from 'all_q_list' (*)
blk_mq_queue_reinit
--> blk_mq_sysfs_unregister (-)
--> blk_mq_sysfs_register (+)
While the request queue is added to 'all_q_list' (*),
blk_mq_queue_reinit() can be called for the queue anytime by CPU
hotplug callback. But blk_mq_sysfs_unregister (-) and
blk_mq_sysfs_register (+) in blk_mq_queue_reinit must not be called
before blk_mq_register_disk (++) and after blk_mq_unregister_disk (--)
is finished. Because '/sys/block/*/mq/' is not exists.
There has already been BLK_MQ_F_SYSFS_UP flag in hctx->flags which can
be used to track these sysfs stuff, but it is only fixing this issue
partially.
In order to fix it completely, we just need per-queue flag instead of
per-hctx flag with appropriate locking. So this introduces
q->mq_sysfs_init_done which is properly protected with all_q_mutex.
Also, we need to ensure that blk_mq_map_swqueue() is called with
all_q_mutex is held. Since hctx->nr_ctx is reset temporarily and
updated in blk_mq_map_swqueue(), so we should avoid
blk_mq_register_hctx() seeing the temporary hctx->nr_ctx value
in CPU hotplug handling or adding/deleting gendisk .
Signed-off-by: Akinobu Mita <akinobu.mita@gmail.com>
Reviewed-by: Ming Lei <tom.leiming@gmail.com>
Cc: Ming Lei <tom.leiming@gmail.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
There may be lots of pending requests so that the buffer of PAGE_SIZE
can't hold them at all.
One typical example is scsi-mq, the queue depth(.can_queue) of
scsi_host and blk-mq is quite big but scsi_device's queue_depth
is a bit small(.cmd_per_lun), then it is quite easy to have lots
of pending requests in hw queue.
This patch fixes the following warning and the related memory
destruction.
[ 359.025101] fill_read_buffer: blk_mq_hw_sysfs_show+0x0/0x7d returned bad count^M
[ 359.055595] irq event stamp: 15537^M
[ 359.055606] general protection fault: 0000 [#1] PREEMPT SMP DEBUG_PAGEALLOC ^M
[ 359.055614] Dumping ftrace buffer:^M
[ 359.055660] (ftrace buffer empty)^M
[ 359.055672] Modules linked in: nbd ipv6 kvm_intel kvm serio_raw^M
[ 359.055678] CPU: 4 PID: 21631 Comm: stress-ng-sysfs Not tainted 4.2.0-rc5-next-20150805 #434^M
[ 359.055679] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS Bochs 01/01/2011^M
[ 359.055682] task: ffff8802161cc000 ti: ffff88021b4a8000 task.ti: ffff88021b4a8000^M
[ 359.055693] RIP: 0010:[<ffffffff811541c5>] [<ffffffff811541c5>] __kmalloc+0xe8/0x152^M
Cc: <stable@vger.kernel.org>
Signed-off-by: Ming Lei <ming.lei@canonical.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Add a variant of blk_mq_init_queue that allows a previously allocated
queue to be initialized. blk_mq_init_allocated_queue models
blk_init_allocated_queue -- which was also created for DM's use.
DM's approach to device creation requires a placeholder request_queue be
allocated for use with alloc_dev() but the decision about what type of
request_queue will be ultimately created is deferred until all component
devices referenced in the DM table are processed to determine the table
type (request-based, blk-mq request-based, or bio-based).
Also, because of DM's late finalization of the request_queue type
the call to blk_mq_register_disk() doesn't happen during alloc_dev().
Must export blk_mq_register_disk() so that DM can backfill the 'mq' dir
once the blk-mq queue is fully allocated.
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
Reviewed-by: Ming Lei <ming.lei@canonical.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
This reverts commit 76d697d107.
The commit 76d697d107 causes general protection fault
reported from Bart Van Assche:
https://lkml.org/lkml/2015/1/28/334
Reported-by: Bart Van Assche <bart.vanassche@sandisk.com>
Signed-off-by: Ming Lei <ming.lei@canonical.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
The kobject memory shouldn't have been freed before the kobject
is released because driver core can access it freely before its
release.
This patch frees hctx in its release callback. For ctx, they
share one single per-cpu variable which is associated with
the request queue, so free ctx in q->mq_kobj's release handler.
Signed-off-by: Sasha Levin <sasha.levin@oracle.com>
(fix ctx kobjects)
Signed-off-by: Ming Lei <ming.lei@canonical.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
When a CPU is hotplugged, the current blk-mq spews a warning like:
kobject '(null)' (ffffe8ffffc8b5d8): tried to add an uninitialized object, something is seriously wrong.
CPU: 1 PID: 1386 Comm: systemd-udevd Not tainted 3.18.0-rc7-2.g088d59b-default #1
Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.7.5-20140531_171129-lamiak 04/01/2014
0000000000000000 0000000000000002 ffffffff81605f07 ffffe8ffffc8b5d8
ffffffff8132c7a0 ffff88023341d370 0000000000000020 ffff8800bb05bd58
ffff8800bb05bd08 000000000000a0a0 000000003f441940 0000000000000007
Call Trace:
[<ffffffff81005306>] dump_trace+0x86/0x330
[<ffffffff81005644>] show_stack_log_lvl+0x94/0x170
[<ffffffff81006d21>] show_stack+0x21/0x50
[<ffffffff81605f07>] dump_stack+0x41/0x51
[<ffffffff8132c7a0>] kobject_add+0xa0/0xb0
[<ffffffff8130aee1>] blk_mq_register_hctx+0x91/0xb0
[<ffffffff8130b82e>] blk_mq_sysfs_register+0x3e/0x60
[<ffffffff81309298>] blk_mq_queue_reinit_notify+0xf8/0x190
[<ffffffff8107cfdc>] notifier_call_chain+0x4c/0x70
[<ffffffff8105fd23>] cpu_notify+0x23/0x50
[<ffffffff81060037>] _cpu_up+0x157/0x170
[<ffffffff810600d9>] cpu_up+0x89/0xb0
[<ffffffff815fa5b5>] cpu_subsys_online+0x35/0x80
[<ffffffff814323cd>] device_online+0x5d/0xa0
[<ffffffff81432485>] online_store+0x75/0x80
[<ffffffff81236a5a>] kernfs_fop_write+0xda/0x150
[<ffffffff811c5532>] vfs_write+0xb2/0x1f0
[<ffffffff811c5f42>] SyS_write+0x42/0xb0
[<ffffffff8160c4ed>] system_call_fastpath+0x16/0x1b
[<00007f0132fb24e0>] 0x7f0132fb24e0
This is indeed because of an uninitialized kobject for blk_mq_ctx.
The blk_mq_ctx kobjects are initialized in blk_mq_sysfs_init(), but it
goes loop over hctx_for_each_ctx(), i.e. it initializes only for
online CPUs. Thus, when a CPU is hotplugged, the ctx for the newly
onlined CPU is registered without initialization.
This patch fixes the issue by initializing the all ctx kobjects
belonging to each queue.
Bugzilla: https://bugzilla.novell.com/show_bug.cgi?id=908794
Cc: <stable@vger.kernel.org>
Signed-off-by: Takashi Iwai <tiwai@suse.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
blk-mq uses percpu_ref for its usage counter which tracks the number
of in-flight commands and used to synchronously drain the queue on
freeze. percpu_ref shutdown takes measureable wallclock time as it
involves a sched RCU grace period. This means that draining a blk-mq
takes measureable wallclock time. One would think that this shouldn't
matter as queue shutdown should be a rare event which takes place
asynchronously w.r.t. userland.
Unfortunately, SCSI probing involves synchronously setting up and then
tearing down a lot of request_queues back-to-back for non-existent
LUNs. This means that SCSI probing may take above ten seconds when
scsi-mq is used.
[ 0.949892] scsi host0: Virtio SCSI HBA
[ 1.007864] scsi 0:0:0:0: Direct-Access QEMU QEMU HARDDISK 1.1. PQ: 0 ANSI: 5
[ 1.021299] scsi 0:0:1:0: Direct-Access QEMU QEMU HARDDISK 1.1. PQ: 0 ANSI: 5
[ 1.520356] tsc: Refined TSC clocksource calibration: 2491.910 MHz
<stall>
[ 16.186549] sd 0:0:0:0: Attached scsi generic sg0 type 0
[ 16.190478] sd 0:0:1:0: Attached scsi generic sg1 type 0
[ 16.194099] osd: LOADED open-osd 0.2.1
[ 16.203202] sd 0:0:0:0: [sda] 31457280 512-byte logical blocks: (16.1 GB/15.0 GiB)
[ 16.208478] sd 0:0:0:0: [sda] Write Protect is off
[ 16.211439] sd 0:0:0:0: [sda] Write cache: enabled, read cache: enabled, doesn't support DPO or FUA
[ 16.218771] sd 0:0:1:0: [sdb] 31457280 512-byte logical blocks: (16.1 GB/15.0 GiB)
[ 16.223264] sd 0:0:1:0: [sdb] Write Protect is off
[ 16.225682] sd 0:0:1:0: [sdb] Write cache: enabled, read cache: enabled, doesn't support DPO or FUA
This is also the reason why request_queues start in bypass mode which
is ended on blk_register_queue() as shutting down a fully functional
queue also involves a RCU grace period and the queues for non-existent
SCSI devices never reach registration.
blk-mq basically needs to do the same thing - start the mq in a
degraded mode which is faster to shut down and then make it fully
functional only after the queue reaches registration. percpu_ref
recently grew facilities to force atomic operation until explicitly
switched to percpu mode, which can be used for this purpose. This
patch makes blk-mq initialize q->mq_usage_counter in atomic mode and
switch it to percpu mode only once blk_register_queue() is reached.
Note that this issue was previously worked around by 0a30288da1
("blk-mq, percpu_ref: implement a kludge for SCSI blk-mq stall during
probe") for v3.17. The temp fix was reverted in preparation of adding
persistent atomic mode to percpu_ref by 9eca80461a ("Revert "blk-mq,
percpu_ref: implement a kludge for SCSI blk-mq stall during probe"").
This patch and the prerequisite percpu_ref changes will be merged
during v3.18 devel cycle.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reported-by: Christoph Hellwig <hch@infradead.org>
Link: http://lkml.kernel.org/g/20140919113815.GA10791@lst.de
Fixes: add703fda9 ("blk-mq: use percpu_ref for mq usage count")
Reviewed-by: Kent Overstreet <kmo@daterainc.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Currently blk-mq registers all the hardware queues in sysfs,
regardless of whether it uses them (e.g. they have CPU mappings)
or not. The unused hardware queues lack the cpux/ directories,
and the other sysfs entries (like active, pending, etc) are all
zeroes.
Change this so that sysfs correctly reflects the current mappings
of the hardware queues.
Signed-off-by: Jens Axboe <axboe@fb.com>
This adds support for active queue tracking, meaning that the
blk-mq tagging maintains a count of active users of a tag set.
This allows us to maintain a notion of fairness between users,
so that we can distribute the tag depth evenly without starving
some users while allowing others to try unfair deep queues.
If sharing of a tag set is detected, each hardware queue will
track the depth of its own queue. And if this exceeds the total
depth divided by the number of active queues, the user is actively
throttled down.
The active queue count is done lazily to avoid bouncing that data
between submitter and completer. Each hardware queue gets marked
active when it allocates its first tag, and gets marked inactive
when 1) the last tag is cleared, and 2) the queue timeout grace
period has passed.
Signed-off-by: Jens Axboe <axboe@fb.com>
The blk-mq code is using it's own version of the I/O completion affinity
tunables, which causes a few issues:
- the rq_affinity sysfs file doesn't work for blk-mq devices, even if it
still is present, thus breaking existing tuning setups.
- the rq_affinity = 1 mode, which is the defauly for legacy request based
drivers isn't implemented at all.
- blk-mq drivers don't implement any completion affinity with the default
flag settings.
This patches removes the blk-mq ipi_redirect flag and sysfs file, as well
as the internal BLK_MQ_F_SHOULD_IPI flag and replaces it with code that
respects the queue-wide rq_affinity flags and also implements the
rq_affinity = 1 mode.
This means I/O completion affinity can now only be tuned block-queue wide
instead of per context, which seems more sensible to me anyway.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
Now that we have a cpu mask of CPUs that are mapped to
a specific hardware queue, we can just iterate that to
display the sysfs num-hw-queue/cpu_list file.
Signed-off-by: Jens Axboe <axboe@fb.com>
Now that we are out of initial debug/bringup mode, remove
the verbose dump of the mapping table.
Provide the mapping table in sysfs, under the hardware queue
directory, in the cpu_list file.
Signed-off-by: Jens Axboe <axboe@fb.com>
Linux currently has two models for block devices:
- The classic request_fn based approach, where drivers use struct
request units for IO. The block layer provides various helper
functionalities to let drivers share code, things like tag
management, timeout handling, queueing, etc.
- The "stacked" approach, where a driver squeezes in between the
block layer and IO submitter. Since this bypasses the IO stack,
driver generally have to manage everything themselves.
With drivers being written for new high IOPS devices, the classic
request_fn based driver doesn't work well enough. The design dates
back to when both SMP and high IOPS was rare. It has problems with
scaling to bigger machines, and runs into scaling issues even on
smaller machines when you have IOPS in the hundreds of thousands
per device.
The stacked approach is then most often selected as the model
for the driver. But this means that everybody has to re-invent
everything, and along with that we get all the problems again
that the shared approach solved.
This commit introduces blk-mq, block multi queue support. The
design is centered around per-cpu queues for queueing IO, which
then funnel down into x number of hardware submission queues.
We might have a 1:1 mapping between the two, or it might be
an N:M mapping. That all depends on what the hardware supports.
blk-mq provides various helper functions, which include:
- Scalable support for request tagging. Most devices need to
be able to uniquely identify a request both in the driver and
to the hardware. The tagging uses per-cpu caches for freed
tags, to enable cache hot reuse.
- Timeout handling without tracking request on a per-device
basis. Basically the driver should be able to get a notification,
if a request happens to fail.
- Optional support for non 1:1 mappings between issue and
submission queues. blk-mq can redirect IO completions to the
desired location.
- Support for per-request payloads. Drivers almost always need
to associate a request structure with some driver private
command structure. Drivers can tell blk-mq this at init time,
and then any request handed to the driver will have the
required size of memory associated with it.
- Support for merging of IO, and plugging. The stacked model
gets neither of these. Even for high IOPS devices, merging
sequential IO reduces per-command overhead and thus
increases bandwidth.
For now, this is provided as a potential 3rd queueing model, with
the hope being that, as it matures, it can replace both the classic
and stacked model. That would get us back to having just 1 real
model for block devices, leaving the stacked approach to dm/md
devices (as it was originally intended).
Contributions in this patch from the following people:
Shaohua Li <shli@fusionio.com>
Alexander Gordeev <agordeev@redhat.com>
Christoph Hellwig <hch@infradead.org>
Mike Christie <michaelc@cs.wisc.edu>
Matias Bjorling <m@bjorling.me>
Jeff Moyer <jmoyer@redhat.com>
Acked-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Akinobu Mita