If we run out of driver tags, we currently treat shared and non-shared
tags the same - both cases hook into the tag waitqueue. This is a bit
more costly than it needs to be on unshared tags, since we have to both
grab the hctx lock, and the waitqueue lock (and disable interrupts).
For the non-shared case, we can simply mark the queue as needing a
restart.
Split blk_mq_dispatch_wait_add() to account for both cases, and
rename it to blk_mq_mark_tag_wait() to better reflect what it
does now.
Without this patch, shared and non-shared performance is about the same
with 4 fio thread hammering on a single null_blk device (~410K, at 75%
sys). With the patch, the shared case is the same, but the non-shared
tags case runs at 431K at 71% sys.
Reviewed-by: Omar Sandoval <osandov@fb.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Currently we are inconsistent in when we decide to run the queue. Using
blk_mq_run_hw_queues() we check if the hctx has pending IO before
running it, but we don't do that from the individual queue run function,
blk_mq_run_hw_queue(). This results in a lot of extra and pointless
queue runs, potentially, on flush requests and (much worse) on tag
starvation situations. This is observable just looking at top output,
with lots of kworkers active. For the !async runs, it just adds to the
CPU overhead of blk-mq.
Move the has-pending check into the run function instead of having
callers do it.
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Several block layer and NVMe core functions accept a combination
of BLK_MQ_REQ_* flags through the 'flags' argument but there is
no verification at compile time whether the right type of block
layer flags is passed. Make it possible for sparse to verify this.
This patch does not change any functionality.
Signed-off-by: Bart Van Assche <bart.vanassche@wdc.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Tested-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Cc: linux-nvme@lists.infradead.org
Cc: Christoph Hellwig <hch@lst.de>
Cc: Johannes Thumshirn <jthumshirn@suse.de>
Cc: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The contexts from which a SCSI device can be quiesced or resumed are:
* Writing into /sys/class/scsi_device/*/device/state.
* SCSI parallel (SPI) domain validation.
* The SCSI device power management methods. See also scsi_bus_pm_ops.
It is essential during suspend and resume that neither the filesystem
state nor the filesystem metadata in RAM changes. This is why while
the hibernation image is being written or restored that SCSI devices
are quiesced. The SCSI core quiesces devices through scsi_device_quiesce()
and scsi_device_resume(). In the SDEV_QUIESCE state execution of
non-preempt requests is deferred. This is realized by returning
BLKPREP_DEFER from inside scsi_prep_state_check() for quiesced SCSI
devices. Avoid that a full queue prevents power management requests
to be submitted by deferring allocation of non-preempt requests for
devices in the quiesced state. This patch has been tested by running
the following commands and by verifying that after each resume the
fio job was still running:
for ((i=0; i<10; i++)); do
(
cd /sys/block/md0/md &&
while true; do
[ "$(<sync_action)" = "idle" ] && echo check > sync_action
sleep 1
done
) &
pids=($!)
for d in /sys/class/block/sd*[a-z]; do
bdev=${d#/sys/class/block/}
hcil=$(readlink "$d/device")
hcil=${hcil#../../../}
echo 4 > "$d/queue/nr_requests"
echo 1 > "/sys/class/scsi_device/$hcil/device/queue_depth"
fio --name="$bdev" --filename="/dev/$bdev" --buffered=0 --bs=512 \
--rw=randread --ioengine=libaio --numjobs=4 --iodepth=16 \
--iodepth_batch=1 --thread --loops=$((2**31)) &
pids+=($!)
done
sleep 1
echo "$(date) Hibernating ..." >>hibernate-test-log.txt
systemctl hibernate
sleep 10
kill "${pids[@]}"
echo idle > /sys/block/md0/md/sync_action
wait
echo "$(date) Done." >>hibernate-test-log.txt
done
Reported-by: Oleksandr Natalenko <oleksandr@natalenko.name>
References: "I/O hangs after resuming from suspend-to-ram" (https://marc.info/?l=linux-block&m=150340235201348).
Signed-off-by: Bart Van Assche <bart.vanassche@wdc.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Tested-by: Martin Steigerwald <martin@lichtvoll.de>
Tested-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Cc: Martin K. Petersen <martin.petersen@oracle.com>
Cc: Ming Lei <ming.lei@redhat.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
This patch makes it possible to pause request allocation for
the legacy block layer by calling blk_mq_freeze_queue() and
blk_mq_unfreeze_queue().
Signed-off-by: Ming Lei <ming.lei@redhat.com>
[ bvanassche: Combined two patches into one, edited a comment and made sure
REQ_NOWAIT is handled properly in blk_old_get_request() ]
Signed-off-by: Bart Van Assche <bart.vanassche@wdc.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Tested-by: Martin Steigerwald <martin@lichtvoll.de>
Tested-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Cc: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
This patch attempts to make the case of hctx re-running on driver tag
failure more robust. Without this patch, it's pretty easy to trigger a
stall condition with shared tags. An example is using null_blk like
this:
modprobe null_blk queue_mode=2 nr_devices=4 shared_tags=1 submit_queues=1 hw_queue_depth=1
which sets up 4 devices, sharing the same tag set with a depth of 1.
Running a fio job ala:
[global]
bs=4k
rw=randread
norandommap
direct=1
ioengine=libaio
iodepth=4
[nullb0]
filename=/dev/nullb0
[nullb1]
filename=/dev/nullb1
[nullb2]
filename=/dev/nullb2
[nullb3]
filename=/dev/nullb3
will inevitably end with one or more threads being stuck waiting for a
scheduler tag. That IO is then stuck forever, until someone else
triggers a run of the queue.
Ensure that we always re-run the hardware queue, if the driver tag we
were waiting for got freed before we added our leftover request entries
back on the dispatch list.
Reviewed-by: Bart Van Assche <bart.vanassche@wdc.com>
Tested-by: Bart Van Assche <bart.vanassche@wdc.com>
Reviewed-by: Ming Lei <ming.lei@redhat.com>
Reviewed-by: Omar Sandoval <osandov@fb.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We have to put the driver tag if dispatch budget can't be got, otherwise
it might cause IO deadlock, especially in case that size of tags is very
small.
Fixes: de1482974080(blk-mq: introduce .get_budget and .put_budget in blk_mq_ops)
Signed-off-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The idea behind it is simple:
1) for none scheduler, driver tag has to be borrowed for flush rq,
otherwise we may run out of tag, and that causes an IO hang. And
get/put driver tag is actually noop for none, so reordering tags
isn't necessary at all.
2) for a real I/O scheduler, we need not allocate a driver tag upfront
for flush rq. It works just fine to follow the same approach as
normal requests: allocate driver tag for each rq just before calling
->queue_rq().
One driver visible change is that the driver tag isn't shared in the
flush request sequence. That won't be a problem, since we always do that
in legacy path.
Then flush rq need not be treated specially wrt. get/put driver tag.
This cleans up the code - for instance, reorder_tags_to_front() can be
removed, and we needn't worry about request ordering in dispatch list
for avoiding I/O deadlock.
Also we have to put the driver tag before requeueing.
Signed-off-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We need this helper to put the driver tag for flush rq, since we will
not share tag in the flush request sequence in the following patch
in case that I/O scheduler is applied.
Signed-off-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Block flush need this function without running the queue, so add a
parameter controlling whether we run it or not.
Signed-off-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
When freeing the driver tag of the next rq with an I/O scheduler
configured, we get the first entry of the list. However, this can
race with requeue of a request, and we end up getting the wrong request
from the head of the list. Free the driver tag of next rq before the
failed one is requeued in the failure branch of queue_rq callback.
Signed-off-by: Jianchao Wang <jianchao.w.wang@oracle.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
It is enough to just check if we can get the budget via .get_budget().
And we don't need to deal with device state change in .get_budget().
For SCSI, one issue to be fixed is that we have to call
scsi_mq_uninit_cmd() to free allocated ressources if SCSI device fails
to handle the request. And it isn't enough to simply call
blk_mq_end_request() to do that if this request is marked as
RQF_DONTPREP.
Fixes: 0df21c86bdbf(scsi: implement .get_budget and .put_budget for blk-mq)
Signed-off-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
blk_mq_get_tag() can modify data->ctx. This means that in the
error path of blk_mq_get_request() data->ctx should be passed to
blk_mq_put_ctx() instead of local_ctx. Note: since blk_mq_put_ctx()
ignores its argument, this patch does not change any functionality.
References: commit 1ad43c0078 ("blk-mq: don't leak preempt counter/q_usage_counter when allocating rq failed")
Reviewed-by: Ming Lei <ming.lei@redhat.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Bart Van Assche <bart.vanassche@wdc.com>
Cc: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
if blk-mq use "none" io scheduler, nr_request get a wrong value when
input a number > tag_set->queue_depth. blk_mq_tag_update_depth will get
the smaller one min(nr, set->queue_depth), and then q->nr_request get a
wrong value.
Reproduce:
echo none > /sys/block/nvme0n1/queue/scheduler
echo 1000000 > /sys/block/nvme0n1/queue/nr_requests
cat /sys/block/nvme0n1/queue/nr_requests
1000000
Signed-off-by: weiping zhang <zhangweiping@didichuxing.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
That we we can also poll non blk-mq queues. Mostly needed for
the NVMe multipath code, but could also be useful elsewhere.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
SCSI restarts its queue in scsi_end_request() automatically, so we don't
need to handle this case in blk-mq.
Especailly any request won't be dequeued in this case, we needn't to
worry about IO hang caused by restart vs. dispatch.
Signed-off-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
SCSI devices use host-wide tagset, and the shared driver tag space is
often quite big. However, there is also a queue depth for each lun(
.cmd_per_lun), which is often small, for example, on both lpfc and
qla2xxx, .cmd_per_lun is just 3.
So lots of requests may stay in sw queue, and we always flush all
belonging to same hw queue and dispatch them all to driver.
Unfortunately it is easy to cause queue busy because of the small
.cmd_per_lun. Once these requests are flushed out, they have to stay in
hctx->dispatch, and no bio merge can happen on these requests, and
sequential IO performance is harmed.
This patch introduces blk_mq_dequeue_from_ctx for dequeuing a request
from a sw queue, so that we can dispatch them in scheduler's way. We can
then avoid dequeueing too many requests from sw queue, since we don't
flush ->dispatch completely.
This patch improves dispatching from sw queue by using the .get_budget
and .put_budget callbacks.
Reviewed-by: Omar Sandoval <osandov@fb.com>
Signed-off-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
For SCSI devices, there is often a per-request-queue depth, which needs
to be respected before queuing one request.
Currently blk-mq always dequeues the request first, then calls
.queue_rq() to dispatch the request to lld. One obvious issue with this
approach is that I/O merging may not be successful, because when the
per-request-queue depth can't be respected, .queue_rq() has to return
BLK_STS_RESOURCE, and then this request has to stay in hctx->dispatch
list. This means it never gets a chance to be merged with other IO.
This patch introduces .get_budget and .put_budget callback in blk_mq_ops,
then we can try to get reserved budget first before dequeuing request.
If the budget for queueing I/O can't be satisfied, we don't need to
dequeue request at all. Hence the request can be left in the IO
scheduler queue, for more merging opportunities.
Signed-off-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Legacy queue sets request's request_list, mq doesn't. This makes mq does
the same thing, so we can find cgroup of a request. Note, we really
only use blkg field of request_list, it's pointless to allocate mempool
for request_list in mq case.
Signed-off-by: Shaohua Li <shli@fb.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
For memory ordering guarantees on stores, we need to ensure that
these two bits share the same byte of storage in the unsigned
long. Add a comment as to why, and a BUILD_BUG_ON() to ensure that
we don't violate this requirement.
Suggested-by: Boqun Feng <boqun.feng@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Attempt to untangle the ordering in blk-mq. The patch introducing the
single smp_mb__before_atomic() is obviously broken in that it doesn't
clearly specify a pairing barrier and an obtained guarantee.
The comment is further misleading in that it hints that the
deadline store and the COMPLETE store also need to be ordered, but
AFAICT there is no such dependency. However what does appear to be
important is the clear happening _after_ the store, and that worked by
pure accident.
This clarifies blk_mq_start_request() -- we should not get there with
STARTING set -- this simplifies the code and makes the barrier usage
sane (the old code could be read to allow not having _any_ atomic after
the barrier, in which case the barrier hasn't got anything to order). We
then also introduce the missing pairing barrier for it.
Also down-grade the barrier to smp_wmb(), this is cheaper for
PowerPC/ARM and doesn't cost anything extra on x86.
And it documents the STARTING vs COMPLETE ordering. Although I've not
been entirely successful in reverse engineering the blk-mq state
machine so there might still be more funnies around timeout vs
requeue.
If I got anything wrong, feel free to educate me by adding comments to
clarify things ;-)
Cc: Alan Stern <stern@rowland.harvard.edu>
Cc: Will Deacon <will.deacon@arm.com>
Cc: Ming Lei <tom.leiming@gmail.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Andrea Parri <parri.andrea@gmail.com>
Cc: Boqun Feng <boqun.feng@gmail.com>
Cc: Bart Van Assche <bart.vanassche@wdc.com>
Cc: "Paul E. McKenney" <paulmck@linux.vnet.ibm.com>
Fixes: 538b753418 ("blk-mq: request deadline must be visible before marking rq as started")
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
For some reason, the laptop mode IO completion notified was never wired
up for blk-mq. Ensure that we trigger the callback appropriately, to arm
the laptop mode flush timer.
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Bart Van Assche <Bart.VanAssche@wdc.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Jens Axboe
Bart Van Assche
Ming Lei
Jianchao Wang
weiping zhang
Christoph Hellwig
Shaohua Li
Peter Zijlstra