The tags used for an IO scheduler are currently per hctx.
As such, when q->nr_hw_queues grows, so does the request queue total IO
scheduler tag depth.
This may cause problems for SCSI MQ HBAs whose total driver depth is
fixed.
Ming and Yanhui report higher CPU usage and lower throughput in scenarios
where the fixed total driver tag depth is appreciably lower than the total
scheduler tag depth:
https://lore.kernel.org/linux-block/440dfcfc-1a2c-bd98-1161-cec4d78c6dfc@huawei.com/T/#mc0d6d4f95275a2743d1c8c3e4dc9ff6c9aa3a76b
In that scenario, since the scheduler tag is got first, much contention
is introduced since a driver tag may not be available after we have got
the sched tag.
Improve this scenario by introducing request queue-wide tags for when
a tagset-wide sbitmap is used. The static sched requests are still
allocated per hctx, as requests are initialised per hctx, as in
blk_mq_init_request(..., hctx_idx, ...) ->
set->ops->init_request(.., hctx_idx, ...).
For simplicity of resizing the request queue sbitmap when updating the
request queue depth, just init at the max possible size, so we don't need
to deal with the possibly with swapping out a new sbitmap for old if
we need to grow.
Signed-off-by: John Garry <john.garry@huawei.com>
Reviewed-by: Ming Lei <ming.lei@redhat.com>
Link: https://lore.kernel.org/r/1620907258-30910-3-git-send-email-john.garry@huawei.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
__blk_mq_sched_bio_merge() gets the ctx and hctx for the current CPU and
passes the hctx to ->bio_merge(). kyber_bio_merge() then gets the ctx
for the current CPU again and uses that to get the corresponding Kyber
context in the passed hctx. However, the thread may be preempted between
the two calls to blk_mq_get_ctx(), and the ctx returned the second time
may no longer correspond to the passed hctx. This "works" accidentally
most of the time, but it can cause us to read garbage if the second ctx
came from an hctx with more ctx's than the first one (i.e., if
ctx->index_hw[hctx->type] > hctx->nr_ctx).
This manifested as this UBSAN array index out of bounds error reported
by Jakub:
UBSAN: array-index-out-of-bounds in ../kernel/locking/qspinlock.c:130:9
index 13106 is out of range for type 'long unsigned int [128]'
Call Trace:
dump_stack+0xa4/0xe5
ubsan_epilogue+0x5/0x40
__ubsan_handle_out_of_bounds.cold.13+0x2a/0x34
queued_spin_lock_slowpath+0x476/0x480
do_raw_spin_lock+0x1c2/0x1d0
kyber_bio_merge+0x112/0x180
blk_mq_submit_bio+0x1f5/0x1100
submit_bio_noacct+0x7b0/0x870
submit_bio+0xc2/0x3a0
btrfs_map_bio+0x4f0/0x9d0
btrfs_submit_data_bio+0x24e/0x310
submit_one_bio+0x7f/0xb0
submit_extent_page+0xc4/0x440
__extent_writepage_io+0x2b8/0x5e0
__extent_writepage+0x28d/0x6e0
extent_write_cache_pages+0x4d7/0x7a0
extent_writepages+0xa2/0x110
do_writepages+0x8f/0x180
__writeback_single_inode+0x99/0x7f0
writeback_sb_inodes+0x34e/0x790
__writeback_inodes_wb+0x9e/0x120
wb_writeback+0x4d2/0x660
wb_workfn+0x64d/0xa10
process_one_work+0x53a/0xa80
worker_thread+0x69/0x5b0
kthread+0x20b/0x240
ret_from_fork+0x1f/0x30
Only Kyber uses the hctx, so fix it by passing the request_queue to
->bio_merge() instead. BFQ and mq-deadline just use that, and Kyber can
map the queues itself to avoid the mismatch.
Fixes: a6088845c2 ("block: kyber: make kyber more friendly with merging")
Reported-by: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: Omar Sandoval <osandov@fb.com>
Link: https://lore.kernel.org/r/c7598605401a48d5cfeadebb678abd10af22b83f.1620691329.git.osandov@fb.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
This series consists of the usual driver updates (ufs, target, tcmu,
smartpqi, lpfc, zfcp, qla2xxx, mpt3sas, pm80xx). The major core
change is using a sbitmap instead of an atomic for queue tracking.
Signed-off-by: James E.J. Bottomley <jejb@linux.ibm.com>
-----BEGIN PGP SIGNATURE-----
iJwEABMIAEQWIQTnYEDbdso9F2cI+arnQslM7pishQUCYInvqCYcamFtZXMuYm90
dG9tbGV5QGhhbnNlbnBhcnRuZXJzaGlwLmNvbQAKCRDnQslM7pishYh2AP0SgqqL
WYZRT2oiyBOKD28v+ceOSiXvgjPlqABwVMC0BAEAn29/wNCxyvzZ1k/b0iPJ4M+S
klkSxLzXKQLzJBgdK5w=
=p5B/
-----END PGP SIGNATURE-----
Merge tag 'scsi-misc' of git://git.kernel.org/pub/scm/linux/kernel/git/jejb/scsi
Pull SCSI updates from James Bottomley:
"This consists of the usual driver updates (ufs, target, tcmu,
smartpqi, lpfc, zfcp, qla2xxx, mpt3sas, pm80xx).
The major core change is using a sbitmap instead of an atomic for
queue tracking"
* tag 'scsi-misc' of git://git.kernel.org/pub/scm/linux/kernel/git/jejb/scsi: (412 commits)
scsi: target: tcm_fc: Fix a kernel-doc header
scsi: target: Shorten ALUA error messages
scsi: target: Fix two format specifiers
scsi: target: Compare explicitly with SAM_STAT_GOOD
scsi: sd: Introduce a new local variable in sd_check_events()
scsi: dc395x: Open-code status_byte(u8) calls
scsi: 53c700: Open-code status_byte(u8) calls
scsi: smartpqi: Remove unused functions
scsi: qla4xxx: Remove an unused function
scsi: myrs: Remove unused functions
scsi: myrb: Remove unused functions
scsi: mpt3sas: Fix two kernel-doc headers
scsi: fcoe: Suppress a compiler warning
scsi: libfc: Fix a format specifier
scsi: aacraid: Remove an unused function
scsi: core: Introduce enum scsi_disposition
scsi: core: Modify the scsi_send_eh_cmnd() return value for the SDEV_BLOCK case
scsi: core: Rename scsi_softirq_done() into scsi_complete()
scsi: core: Remove an incorrect comment
scsi: core: Make the scsi_alloc_sgtables() documentation more accurate
...
list_sort() internally casts the comparison function passed to it
to a different type with constant struct list_head pointers, and
uses this pointer to call the functions, which trips indirect call
Control-Flow Integrity (CFI) checking.
Instead of removing the consts, this change defines the
list_cmp_func_t type and changes the comparison function types of
all list_sort() callers to use const pointers, thus avoiding type
mismatches.
Suggested-by: Nick Desaulniers <ndesaulniers@google.com>
Signed-off-by: Sami Tolvanen <samitolvanen@google.com>
Reviewed-by: Nick Desaulniers <ndesaulniers@google.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Kees Cook <keescook@chromium.org>
Tested-by: Nick Desaulniers <ndesaulniers@google.com>
Tested-by: Nathan Chancellor <nathan@kernel.org>
Signed-off-by: Kees Cook <keescook@chromium.org>
Link: https://lore.kernel.org/r/20210408182843.1754385-10-samitolvanen@google.com
SCSI uses a global atomic variable to track queue depth for each
LUN/request queue.
This doesn't scale well when there are lots of CPU cores and the disk is
very fast. It has been observed that IOPS is affected a lot by tracking
queue depth via sdev->device_busy in the I/O path.
Return budget token from .get_budget callback. The budget token can be
passed to driver so that we can replace the atomic variable with
sbitmap_queue and alleviate the scaling problems that way.
Link: https://lore.kernel.org/r/20210122023317.687987-9-ming.lei@redhat.com
Cc: Omar Sandoval <osandov@fb.com>
Cc: Kashyap Desai <kashyap.desai@broadcom.com>
Cc: Sumanesh Samanta <sumanesh.samanta@broadcom.com>
Cc: Ewan D. Milne <emilne@redhat.com>
Tested-by: Sumanesh Samanta <sumanesh.samanta@broadcom.com>
Reviewed-by: Hannes Reinecke <hare@suse.de>
Signed-off-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
Commit a1ce35fa49 ("block: remove dead
elevator code") removed all users of RQF_SORTED. However it is still
defined, and there is one reference left to it (which in effect is
dead code). Clear it all up.
Signed-off-by: Jean Delvare <jdelvare@suse.de>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Ming Lei <ming.lei@redhat.com>
Cc: Omar Sandoval <osandov@fb.com>
Cc: Hannes Reinecke <hare@suse.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Get rid of the wrapper for trace_block_rq_insert() and call the function
directly.
Signed-off-by: Chaitanya Kulkarni <chaitanya.kulkarni@wdc.com>
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Reviewed-by: Damien Le Moal <damien.lemoal@wdc.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The request_queue can trivially be derived from the request.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Damien Le Moal <damien.lemoal@wdc.com>
Reviewed-by: Hannes Reinecke <hare@suse.de>
Reviewed-by: Chaitanya Kulkarni <chaitanya.kulkarni@wdc.com>
Acked-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
After commit 923218f616 ("blk-mq: don't allocate driver tag upfront
for flush rq"), blk_mq_submit_bio() will call blk_insert_flush()
directly to handle flush request rather than blk_mq_sched_insert_request()
in the case of elevator.
Then, all flush request either have set RQF_FLUSH_SEQ flag when call
blk_mq_sched_insert_request(), or have inserted into hctx->dispatch.
So, remove the dead code path.
Signed-off-by: Yufen Yu <yuyufen@huawei.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Move blk_mq_sched_try_merge to blk-merge.c, which allows to mark
a lot of the merge infrastructure static there.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Now we usually free the hctx->sched_data by e->type->ops.exit_hctx(),
and no users will use blk_mq_sched_free_hctx_data() function.
Remove it.
Signed-off-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Some SCSI HBAs (such as HPSA, megaraid, mpt3sas, hisi_sas_v3 ..) support
multiple reply queues with single hostwide tags.
In addition, these drivers want to use interrupt assignment in
pci_alloc_irq_vectors(PCI_IRQ_AFFINITY). However, as discussed in [0],
CPU hotplug may cause in-flight IO completion to not be serviced when an
interrupt is shutdown. That problem is solved in commit bf0beec060
("blk-mq: drain I/O when all CPUs in a hctx are offline").
However, to take advantage of that blk-mq feature, the HBA HW queuess are
required to be mapped to that of the blk-mq hctx's; to do that, the HBA HW
queues need to be exposed to the upper layer.
In making that transition, the per-SCSI command request tags are no
longer unique per Scsi host - they are just unique per hctx. As such, the
HBA LLDD would have to generate this tag internally, which has a certain
performance overhead.
However another problem is that blk-mq assumes the host may accept
(Scsi_host.can_queue * #hw queue) commands. In commit 6eb045e092 ("scsi:
core: avoid host-wide host_busy counter for scsi_mq"), the Scsi host busy
counter was removed, which would stop the LLDD being sent more than
.can_queue commands; however, it should still be ensured that the block
layer does not issue more than .can_queue commands to the Scsi host.
To solve this problem, introduce a shared sbitmap per blk_mq_tag_set,
which may be requested at init time.
New flag BLK_MQ_F_TAG_HCTX_SHARED should be set when requesting the
tagset to indicate whether the shared sbitmap should be used.
Even when BLK_MQ_F_TAG_HCTX_SHARED is set, a full set of tags and requests
are still allocated per hctx; the reason for this is that if tags and
requests were only allocated for a single hctx - like hctx0 - it may break
block drivers which expect a request be associated with a specific hctx,
i.e. not always hctx0. This will introduce extra memory usage.
This change is based on work originally from Ming Lei in [1] and from
Bart's suggestion in [2].
[0] https://lore.kernel.org/linux-block/alpine.DEB.2.21.1904051331270.1802@nanos.tec.linutronix.de/
[1] https://lore.kernel.org/linux-block/20190531022801.10003-1-ming.lei@redhat.com/
[2] https://lore.kernel.org/linux-block/ff77beff-5fd9-9f05-12b6-826922bace1f@huawei.com/T/#m3db0a602f095cbcbff27e9c884d6b4ae826144be
Signed-off-by: John Garry <john.garry@huawei.com>
Tested-by: Don Brace<don.brace@microsemi.com> #SCSI resv cmds patches used
Tested-by: Douglas Gilbert <dgilbert@interlog.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Pass hctx/tagset flags argument down to blk_mq_init_tags() and
blk_mq_free_tags() for selective init/free.
For now, make it include the alloc policy flag, which can be evaluated
when needed (in blk_mq_init_tags()).
Signed-off-by: John Garry <john.garry@huawei.com>
Tested-by: Douglas Gilbert <dgilbert@interlog.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The small blk_mq_attempt_merge() function is only called by
__blk_mq_sched_bio_merge(), just open code it.
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
There are lots of duplicated code when trying to merge a bio from
plug list and sw queue, we can introduce a new helper to attempt
to merge a bio, which can simplify the blk_bio_list_merge()
and blk_attempt_plug_merge().
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Move the blk_mq_bio_list_merge() into blk-merge.c and
rename it as a generic name.
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
SCHED_RESTART code path is relied to re-run queue for dispatch requests
in hctx->dispatch. Meantime the SCHED_RSTART flag is checked when adding
requests to hctx->dispatch.
memory barriers have to be used for ordering the following two pair of OPs:
1) adding requests to hctx->dispatch and checking SCHED_RESTART in
blk_mq_dispatch_rq_list()
2) clearing SCHED_RESTART and checking if there is request in hctx->dispatch
in blk_mq_sched_restart().
Without the added memory barrier, either:
1) blk_mq_sched_restart() may miss requests added to hctx->dispatch meantime
blk_mq_dispatch_rq_list() observes SCHED_RESTART, and not run queue in
dispatch side
or
2) blk_mq_dispatch_rq_list still sees SCHED_RESTART, and not run queue
in dispatch side, meantime checking if there is request in
hctx->dispatch from blk_mq_sched_restart() is missed.
IO hang in ltp/fs_fill test is reported by kernel test robot:
https://lkml.org/lkml/2020/7/26/77
Turns out it is caused by the above out-of-order OPs. And the IO hang
can't be observed any more after applying this patch.
Fixes: bd166ef183 ("blk-mq-sched: add framework for MQ capable IO schedulers")
Reported-by: kernel test robot <rong.a.chen@intel.com>
Signed-off-by: Ming Lei <ming.lei@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Cc: Bart Van Assche <bvanassche@acm.org>
Cc: Christoph Hellwig <hch@lst.de>
Cc: David Jeffery <djeffery@redhat.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Drop the repeated word "to".
Signed-off-by: Randy Dunlap <rdunlap@infradead.org>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: linux-block@vger.kernel.org
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Remove unnecessary local variable 'ret' in blk_mq_dispatch_hctx_list().
Signed-off-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Reviewed-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
More and more drivers want to get batching requests queued from
block layer, such as mmc, and tcp based storage drivers. Also
current in-tree users have virtio-scsi, virtio-blk and nvme.
For none, we already support batching dispatch.
But for io scheduler, every time we just take one request from scheduler
and pass the single request to blk_mq_dispatch_rq_list(). This way makes
batching dispatch not possible when io scheduler is applied. One reason
is that we don't want to hurt sequential IO performance, becasue IO
merge chance is reduced if more requests are dequeued from scheduler
queue.
Try to support batching dispatch for io scheduler by starting with the
following simple approach:
1) still make sure we can get budget before dequeueing request
2) use hctx->dispatch_busy to evaluate if queue is busy, if it is busy
we fackback to non-batching dispatch, otherwise dequeue as many as
possible requests from scheduler, and pass them to blk_mq_dispatch_rq_list().
Wrt. 2), we use similar policy for none, and turns out that SCSI SSD
performance got improved much.
In future, maybe we can develop more intelligent algorithem for batching
dispatch.
Baolin has tested this patch and found that MMC performance is improved[3].
[1] https://lore.kernel.org/linux-block/20200512075501.GF1531898@T590/#r
[2] https://lore.kernel.org/linux-block/fe6bd8b9-6ed9-b225-f80c-314746133722@grimberg.me/
[3] https://lore.kernel.org/linux-block/CADBw62o9eTQDJ9RvNgEqSpXmg6Xcq=2TxH0Hfxhp29uF2W=TXA@mail.gmail.com/
Signed-off-by: Ming Lei <ming.lei@redhat.com>
Tested-by: Baolin Wang <baolin.wang7@gmail.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Cc: Sagi Grimberg <sagi@grimberg.me>
Cc: Baolin Wang <baolin.wang7@gmail.com>
Cc: Christoph Hellwig <hch@infradead.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Pass obtained budget count to blk_mq_dispatch_rq_list(), and prepare
for supporting fully batching submission.
With the obtained budget count, it is easier to put extra budgets
in case of .queue_rq failure.
Meantime remove the old 'got_budget' parameter.
Signed-off-by: Ming Lei <ming.lei@redhat.com>
Tested-by: Baolin Wang <baolin.wang7@gmail.com>
Reviewed-by: Christoph Hellwig <hch@infradead.org>
Cc: Sagi Grimberg <sagi@grimberg.me>
Cc: Baolin Wang <baolin.wang7@gmail.com>
Cc: Christoph Hellwig <hch@infradead.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
All requests in the 'list' of blk_mq_dispatch_rq_list belong to same
hctx, so it is better to pass hctx instead of request queue, because
blk-mq's dispatch target is hctx instead of request queue.
Signed-off-by: Ming Lei <ming.lei@redhat.com>
Tested-by: Baolin Wang <baolin.wang7@gmail.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Cc: Sagi Grimberg <sagi@grimberg.me>
Cc: Baolin Wang <baolin.wang7@gmail.com>
Cc: Christoph Hellwig <hch@infradead.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
blk-mq budget is abstract from scsi's device queue depth, and it is
always per-request-queue instead of hctx.
It can be quite absurd to get a budget from one hctx, then dequeue a
request from scheduler queue, and this request may not belong to this
hctx, at least for bfq and deadline.
So fix the mess and always pass request queue to get/put budget
callback.
Signed-off-by: Ming Lei <ming.lei@redhat.com>
Tested-by: Baolin Wang <baolin.wang7@gmail.com>
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Douglas Anderson <dianders@chromium.org>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Cc: Sagi Grimberg <sagi@grimberg.me>
Cc: Baolin Wang <baolin.wang7@gmail.com>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: Douglas Anderson <dianders@chromium.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Fix sparse warnings:
block/blk-mq-sched.c:209:5: warning: symbol '__blk_mq_sched_dispatch_requests' was not declared. Should it be static?
Reported-by: Hulk Robot <hulkci@huawei.com>
Signed-off-by: Zheng Bin <zhengbin13@huawei.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Flushes bypass the I/O scheduler and get added to hctx->dispatch
in blk_mq_sched_bypass_insert. This can happen while a kworker is running
hctx->run_work work item and is past the point in
blk_mq_sched_dispatch_requests where hctx->dispatch is checked.
The blk_mq_do_dispatch_sched call is not guaranteed to end in bounded time,
because the I/O scheduler can feed an arbitrary number of commands.
Since we have only one hctx->run_work, the commands waiting in
hctx->dispatch will wait an arbitrary length of time for run_work to be
rerun.
A similar phenomenon exists with dispatches from the software queue.
The solution is to poll hctx->dispatch in blk_mq_do_dispatch_sched and
blk_mq_do_dispatch_ctx and return from the run_work handler and let it
rerun.
Signed-off-by: Salman Qazi <sqazi@google.com>
Reviewed-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
If ever a thread running blk-mq code tries to get budget and fails it
immediately stops doing work and assumes that whenever budget is freed
up that queues will be kicked and whatever work the thread was trying
to do will be tried again.
One path where budget is freed and queues are kicked in the normal
case can be seen in scsi_finish_command(). Specifically:
- scsi_finish_command()
- scsi_device_unbusy()
- # Decrement "device_busy", AKA release budget
- scsi_io_completion()
- scsi_end_request()
- blk_mq_run_hw_queues()
The above is all well and good. The problem comes up when a thread
claims the budget but then releases it without actually dispatching
any work. Since we didn't schedule any work we'll never run the path
of finishing work / kicking the queues.
This isn't often actually a problem which is why this issue has
existed for a while and nobody noticed. Specifically we only get into
this situation when we unexpectedly found that we weren't going to do
any work. Code that later receives new work kicks the queues. All
good, right?
The problem shows up, however, if timing is just wrong and we hit a
race. To see this race let's think about the case where we only have
a budget of 1 (only one thread can hold budget). Now imagine that a
thread got budget and then decided not to dispatch work. It's about
to call put_budget() but then the thread gets context switched out for
a long, long time. While in this state, any and all kicks of the
queue (like the when we received new work) will be no-ops because
nobody can get budget. Finally the thread holding budget gets to run
again and returns. All the normal kicks will have been no-ops and we
have an I/O stall.
As you can see from the above, you need just the right timing to see
the race. To start with, the only case it happens if we thought we
had work, actually managed to get the budget, but then actually didn't
have work. That's pretty rare to start with. Even then, there's
usually a very small amount of time between realizing that there's no
work and putting the budget. During this small amount of time new
work has to come in and the queue kick has to make it all the way to
trying to get the budget and fail. It's pretty unlikely.
One case where this could have failed is illustrated by an example of
threads running blk_mq_do_dispatch_sched():
* Threads A and B both run has_work() at the same time with the same
"hctx". Imagine has_work() is exact. There's no lock, so it's OK
if Thread A and B both get back true.
* Thread B gets interrupted for a long time right after it decides
that there is work. Maybe its CPU gets an interrupt and the
interrupt handler is slow.
* Thread A runs, get budget, dispatches work.
* Thread A's work finishes and budget is released.
* Thread B finally runs again and gets budget.
* Since Thread A already took care of the work and no new work has
come in, Thread B will get NULL from dispatch_request(). I believe
this is specifically why dispatch_request() is allowed to return
NULL in the first place if has_work() must be exact.
* Thread B will now be holding the budget and is about to call
put_budget(), but hasn't called it yet.
* Thread B gets interrupted for a long time (again). Dang interrupts.
* Now Thread C (maybe with a different "hctx" but the same queue)
comes along and runs blk_mq_do_dispatch_sched().
* Thread C won't do anything because it can't get budget.
* Finally Thread B will run again and put the budget without kicking
any queues.
Even though the example above is with blk_mq_do_dispatch_sched() I
believe the race is possible any time someone is holding budget but
doesn't do work.
Unfortunately, the unlikely has become more likely if you happen to be
using the BFQ I/O scheduler. BFQ, by design, sometimes returns "true"
for has_work() but then NULL for dispatch_request() and stays in this
state for a while (currently up to 9 ms). Suddenly you only need one
race to hit, not two races in a row. With my current setup this is
easy to reproduce in reboot tests and traces have actually shown that
we hit a race similar to the one described above.
Note that we only need to fix blk_mq_do_dispatch_sched() and
blk_mq_do_dispatch_ctx() and not the other places that put budget. In
other cases we know that we have work to do on at least one "hctx" and
code already exists to kick that "hctx"'s queue. When that work
finally finishes all the queues will be kicked using the normal flow.
One last note is that (at least in the SCSI case) budget is shared by
all "hctx"s that have the same queue. Thus we need to make sure to
kick the whole queue, not just re-run dispatching on a single "hctx".
Signed-off-by: Douglas Anderson <dianders@chromium.org>
Reviewed-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
commit 01e99aeca3 ("blk-mq: insert passthrough request into
hctx->dispatch directly") may change to add flush request to the tail
of dispatch by applying the 'add_head' parameter of
blk_mq_sched_insert_request.
Turns out this way causes performance regression on NCQ controller because
flush is non-NCQ command, which can't be queued when there is any in-flight
NCQ command. When adding flush rq to the front of hctx->dispatch, it is
easier to introduce extra time to flush rq's latency compared with adding
to the tail of dispatch queue because of S_SCHED_RESTART, then chance of
flush merge is increased, and less flush requests may be issued to
controller.
So always insert flush request to the front of dispatch queue just like
before applying commit 01e99aeca3 ("blk-mq: insert passthrough request
into hctx->dispatch directly").
Cc: Damien Le Moal <Damien.LeMoal@wdc.com>
Cc: Shinichiro Kawasaki <shinichiro.kawasaki@wdc.com>
Reported-by: Shinichiro Kawasaki <shinichiro.kawasaki@wdc.com>
Fixes: 01e99aeca3 ("blk-mq: insert passthrough request into hctx->dispatch directly")
Signed-off-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
For some reason, device may be in one situation which can't handle
FS request, so STS_RESOURCE is always returned and the FS request
will be added to hctx->dispatch. However passthrough request may
be required at that time for fixing the problem. If passthrough
request is added to scheduler queue, there isn't any chance for
blk-mq to dispatch it given we prioritize requests in hctx->dispatch.
Then the FS IO request may never be completed, and IO hang is caused.
So passthrough request has to be added to hctx->dispatch directly
for fixing the IO hang.
Fix this issue by inserting passthrough request into hctx->dispatch
directly together withing adding FS request to the tail of
hctx->dispatch in blk_mq_dispatch_rq_list(). Actually we add FS request
to tail of hctx->dispatch at default, see blk_mq_request_bypass_insert().
Then it becomes consistent with original legacy IO request
path, in which passthrough request is always added to q->queue_head.
Cc: Dongli Zhang <dongli.zhang@oracle.com>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: Ewan D. Milne <emilne@redhat.com>
Signed-off-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Commit c48dac137a ("block: don't hold q->sysfs_lock in elevator_init_mq")
removes q->sysfs_lock from elevator_init_mq(), but forgot to deal with
lockdep_assert_held() called in blk_mq_sched_free_requests() which is
run in failure path of elevator_init_mq().
blk_mq_sched_free_requests() is called in the following 3 functions:
elevator_init_mq()
elevator_exit()
blk_cleanup_queue()
In blk_cleanup_queue(), blk_mq_sched_free_requests() is followed exactly
by 'mutex_lock(&q->sysfs_lock)'.
So moving the lockdep_assert_held() from blk_mq_sched_free_requests()
into elevator_exit() for fixing the report by syzbot.
Reported-by: syzbot+da3b7677bb913dc1b737@syzkaller.appspotmail.com
Fixed: c48dac137a ("block: don't hold q->sysfs_lock in elevator_init_mq")
Reviewed-by: Bart Van Assche <bvanassche@acm.org>
Reviewed-by: Damien Le Moal <damien.lemoal@wdc.com>
Signed-off-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
No code that occurs between blk_mq_get_ctx() and blk_mq_put_ctx() depends
on preemption being disabled for its correctness. Since removing the CPU
preemption calls does not measurably affect performance, simplify the
blk-mq code by removing the blk_mq_put_ctx() function and also by not
disabling preemption in blk_mq_get_ctx().
Cc: Hannes Reinecke <hare@suse.com>
Cc: Omar Sandoval <osandov@fb.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Bart Van Assche <bvanassche@acm.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We only need the number of segments in the blk-mq submission path.
Remove the field from struct bio, and return it from a variant of
blk_queue_split instead of that it can passed as an argument to
those functions that need the value.
This also means we stop recounting segments except for cloning
and partial segments.
To keep the number of arguments in this how path down remove
pointless struct request_queue arguments from any of the functions
that had it and grew a nr_segs argument.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
blk_mq_sched_free_requests() may be called in failure path in which
q->elevator may not be setup yet, so remove WARN_ON(!q->elevator) from
blk_mq_sched_free_requests for avoiding the false positive.
This function is actually safe to call in case of !q->elevator because
hctx->sched_tags is checked.
Cc: Bart Van Assche <bvanassche@acm.org>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Yi Zhang <yi.zhang@redhat.com>
Fixes: c3e2219216 ("block: free sched's request pool in blk_cleanup_queue")
Reported-by: syzbot+b9d0d56867048c7bcfde@syzkaller.appspotmail.com
Signed-off-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
In theory, IO scheduler belongs to request queue, and the request pool
of sched tags belongs to the request queue too.
However, the current tags allocation interfaces are re-used for both
driver tags and sched tags, and driver tags is definitely host wide,
and doesn't belong to any request queue, same with its request pool.
So we need tagset instance for freeing request of sched tags.
Meantime, blk_mq_free_tag_set() often follows blk_cleanup_queue() in case
of non-BLK_MQ_F_TAG_SHARED, this way requires that request pool of sched
tags to be freed before calling blk_mq_free_tag_set().
Commit 47cdee29ef ("block: move blk_exit_queue into __blk_release_queue")
moves blk_exit_queue into __blk_release_queue for simplying the fast
path in generic_make_request(), then causes oops during freeing requests
of sched tags in __blk_release_queue().
Fix the above issue by move freeing request pool of sched tags into
blk_cleanup_queue(), this way is safe becasue queue has been frozen and no any
in-queue requests at that time. Freeing sched tags has to be kept in queue's
release handler becasue there might be un-completed dispatch activity
which might refer to sched tags.
Cc: Bart Van Assche <bvanassche@acm.org>
Cc: Christoph Hellwig <hch@lst.de>
Fixes: 47cdee29ef ("block: move blk_exit_queue into __blk_release_queue")
Tested-by: Yi Zhang <yi.zhang@redhat.com>
Reported-by: kernel test robot <rong.a.chen@intel.com>
Signed-off-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Just like aio/io_uring, we need to grab 2 refcount for queuing one
request, one is for submission, another is for completion.
If the request isn't queued from plug code path, the refcount grabbed
in generic_make_request() serves for submission. In theroy, this
refcount should have been released after the sumission(async run queue)
is done. blk_freeze_queue() works with blk_sync_queue() together
for avoiding race between cleanup queue and IO submission, given async
run queue activities are canceled because hctx->run_work is scheduled with
the refcount held, so it is fine to not hold the refcount when
running the run queue work function for dispatch IO.
However, if request is staggered into plug list, and finally queued
from plug code path, the refcount in submission side is actually missed.
And we may start to run queue after queue is removed because the queue's
kobject refcount isn't guaranteed to be grabbed in flushing plug list
context, then kernel oops is triggered, see the following race:
blk_mq_flush_plug_list():
blk_mq_sched_insert_requests()
insert requests to sw queue or scheduler queue
blk_mq_run_hw_queue
Because of concurrent run queue, all requests inserted above may be
completed before calling the above blk_mq_run_hw_queue. Then queue can
be freed during the above blk_mq_run_hw_queue().
Fixes the issue by grab .q_usage_counter before calling
blk_mq_sched_insert_requests() in blk_mq_flush_plug_list(). This way is
safe because the queue is absolutely alive before inserting request.
Cc: Dongli Zhang <dongli.zhang@oracle.com>
Cc: James Smart <james.smart@broadcom.com>
Cc: linux-scsi@vger.kernel.org,
Cc: Martin K . Petersen <martin.petersen@oracle.com>,
Cc: Christoph Hellwig <hch@lst.de>,
Cc: James E . J . Bottomley <jejb@linux.vnet.ibm.com>,
Reviewed-by: Bart Van Assche <bvanassche@acm.org>
Tested-by: James Smart <james.smart@broadcom.com>
Signed-off-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Various block layer files do not have any licensing information at all.
Add SPDX tags for the default kernel GPLv2 license to those.
Reviewed-by: Chaitanya Kulkarni <chaitanya.kulkarni@wdc.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
blk_mq_try_issue_directly() can return BLK_STS*_RESOURCE for requests that
have been queued. If that happens when blk_mq_try_issue_directly() is called
by the dm-mpath driver then dm-mpath will try to resubmit a request that is
already queued and a kernel crash follows. Since it is nontrivial to fix
blk_mq_request_issue_directly(), revert the blk_mq_request_issue_directly()
changes that went into kernel v5.0.
This patch reverts the following commits:
* d6a51a97c0 ("blk-mq: replace and kill blk_mq_request_issue_directly") # v5.0.
* 5b7a6f128a ("blk-mq: issue directly with bypass 'false' in blk_mq_sched_insert_requests") # v5.0.
* 7f556a44e6 ("blk-mq: refactor the code of issue request directly") # v5.0.
Cc: Christoph Hellwig <hch@infradead.org>
Cc: Ming Lei <ming.lei@redhat.com>
Cc: Jianchao Wang <jianchao.w.wang@oracle.com>
Cc: Hannes Reinecke <hare@suse.com>
Cc: Johannes Thumshirn <jthumshirn@suse.de>
Cc: James Smart <james.smart@broadcom.com>
Cc: Dongli Zhang <dongli.zhang@oracle.com>
Cc: Laurence Oberman <loberman@redhat.com>
Cc: <stable@vger.kernel.org>
Reported-by: Laurence Oberman <loberman@redhat.com>
Tested-by: Laurence Oberman <loberman@redhat.com>
Fixes: 7f556a44e6 ("blk-mq: refactor the code of issue request directly") # v5.0.
Signed-off-by: Bart Van Assche <bvanassche@acm.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Currently, the queue mapping result is saved in a two-dimensional
array. In the hot path, to get a hctx, we need do following:
q->queue_hw_ctx[q->tag_set->map[type].mq_map[cpu]]
This isn't very efficient. We could save the queue mapping result into
ctx directly with different hctx type, like,
ctx->hctxs[type]
Signed-off-by: Jianchao Wang <jianchao.w.wang@oracle.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
When a request is added to rq list of sw queue(ctx), the rq may be from
a different type of hctx, especially after multi queue mapping is
introduced.
So when dispach request from sw queue via blk_mq_flush_busy_ctxs() or
blk_mq_dequeue_from_ctx(), one request belonging to other queue type of
hctx can be dispatched to current hctx in case that read queue or poll
queue is enabled.
This patch fixes this issue by introducing per-queue-type list.
Cc: Christoph Hellwig <hch@lst.de>
Signed-off-by: Ming Lei <ming.lei@redhat.com>
Changed by me to not use separately cacheline aligned lists, just
place them all in the same cacheline where we had just the one list
and lock before.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
For a zoned block device using mq-deadline, if a write request for a
zone is received while another write was already dispatched for the same
zone, dd_dispatch_request() will return NULL and the newly inserted
write request is kept in the scheduler queue waiting for the ongoing
zone write to complete. With this behavior, when no other request has
been dispatched, rq_list in blk_mq_sched_dispatch_requests() is empty
and blk_mq_sched_mark_restart_hctx() not called. This in turn leads to
__blk_mq_free_request() call of blk_mq_sched_restart() to not run the
queue when the already dispatched write request completes. The newly
dispatched request stays stuck in the scheduler queue until eventually
another request is submitted.
This problem does not affect SCSI disk as the SCSI stack handles queue
restart on request completion. However, this problem is can be triggered
the nullblk driver with zoned mode enabled.
Fix this by always requesting a queue restart in dd_dispatch_request()
if no request was dispatched while WRITE requests are queued.
Fixes: 5700f69178 ("mq-deadline: Introduce zone locking support")
Cc: <stable@vger.kernel.org>
Signed-off-by: Damien Le Moal <damien.lemoal@wdc.com>
Add missing export of blk_mq_sched_restart()
Signed-off-by: Jens Axboe <axboe@kernel.dk>
It is not necessary to issue request directly with bypass 'true'
in blk_mq_sched_insert_requests and handle the non-issued requests
itself. Just set bypass to 'false' and let blk_mq_try_issue_directly
handle them totally. Remove the blk_rq_can_direct_dispatch check,
because blk_mq_try_issue_directly can handle it well.If request is
direct-issued unsuccessfully, insert the reset.
Signed-off-by: Jianchao Wang <jianchao.w.wang@oracle.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
If the first request allocated and issued by a process is a passhthrough
request, we don't set up an IO context for it. Ensure that
blk_mq_sched_assign_ioc() ignores a NULL io_context.
Fixes: e2b3fa5af7 ("block: Remove bio->bi_ioc")
Reported-by: Ming Lei <ming.lei@redhat.com>
Tested-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
bio->bi_ioc is never set so always NULL. Remove references to it in
bio_disassociate_task() and in rq_ioc() and delete this field from
struct bio. With this change, rq_ioc() always returns
current->io_context without the need for a bio argument. Further
simplify the code and make it more readable by also removing this
helper, which also allows to simplify blk_mq_sched_assign_ioc() by
removing its bio argument.
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Reviewed-by: Adam Manzanares <adam.manzanares@wdc.com>
Signed-off-by: Damien Le Moal <damien.lemoal@wdc.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
With the legacy request path gone there is no good reason to keep
queue_lock as a pointer, we can always use the embedded lock now.
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Fixed floppy and blk-cgroup missing conversions and half done edits.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
It's somewhat strange to have a list insertion function that
relies on the fact that the caller has mapped things correctly.
Pass in the hardware queue directly for insertion, which makes
for a much cleaner interface and implementation.
Reviewed-by: Keith Busch <keith.busch@intel.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We call blk_mq_map_queue() a lot, at least two times for each
request per IO, sometimes more. Since we now have an indirect
call as well in that function. cache the mapping so we don't
have to re-call blk_mq_map_queue() for the same request
multiple times.
Reviewed-by: Keith Busch <keith.busch@intel.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The mapping used to be dependent on just the CPU location, but
now it's a tuple of (type, cpu) instead. This is a prep patch
for allowing a single software queue to map to multiple hardware
queues. No functional changes in this patch.
This changes the software queue count to an unsigned short
to save a bit of space. We can still support 64K-1 CPUs,
which should be enough. Add a check to catch a wrap.
Reviewed-by: Hannes Reinecke <hare@suse.com>
Reviewed-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Prep patch for being able to place request based not just on
CPU location, but also on the type of request.
Reviewed-by: Hannes Reinecke <hare@suse.com>
Reviewed-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
This is a remnant of when we had ops for both SQ and MQ
schedulers. Now it's just MQ, so get rid of the union.
Reviewed-by: Omar Sandoval <osandov@fb.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Currently, when update nr_hw_queues, IO scheduler's init_hctx will
be invoked before the mapping between ctx and hctx is adapted
correctly by blk_mq_map_swqueue. The IO scheduler init_hctx (kyber)
may depend on this mapping and get wrong result and panic finally.
A simply way to fix this is that switch the IO scheduler to 'none'
before update the nr_hw_queues, and then switch it back after
update nr_hw_queues. blk_mq_sched_init_/exit_hctx are removed due
to nobody use them any more.
Signed-off-by: Jianchao Wang <jianchao.w.wang@oracle.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
