If we have shared tags enabled, then every IO completion will trigger
a full loop of every queue belonging to a tag set, and every hardware
queue for each of those queues, even if nothing needs to be done.
This causes a massive performance regression if you have a lot of
shared devices.
Instead of doing this huge full scan on every IO, add an atomic
counter to the main queue that tracks how many hardware queues have
been marked as needing a restart. With that, we can avoid looking for
restartable queues, if we don't have to.
Max reports that this restores performance. Before this patch, 4K
IOPS was limited to 22-23K IOPS. With the patch, we are running at
950-970K IOPS.
Fixes: 6d8c6c0f97 ("blk-mq: Restart a single queue if tag sets are shared")
Reported-by: Max Gurtovoy <maxg@mellanox.com>
Tested-by: Max Gurtovoy <maxg@mellanox.com>
Reviewed-by: Bart Van Assche <bart.vanassche@sandisk.com>
Tested-by: Bart Van Assche <bart.vanassche@wdc.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Document the locking assumptions in functions that modify
blk_mq_ctx.rq_list to make it easier for humans to verify
this code.
Signed-off-by: Bart Van Assche <bart.vanassche@sandisk.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Cc: Hannes Reinecke <hare@suse.com>
Cc: Omar Sandoval <osandov@fb.com>
Cc: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
It is required that no dispatch can happen any more once
blk_mq_quiesce_queue() returns, and we don't have such requirement
on APIs of stopping queue.
But blk_mq_quiesce_queue() still may not block/drain dispatch in the
the case of BLK_MQ_S_START_ON_RUN, so use the new introduced flag of
QUEUE_FLAG_QUIESCED and evaluate it inside RCU read-side critical
sections for fixing this issue.
Also blk_mq_quiesce_queue() is implemented via stopping queue, which
limits its uses, and easy to cause race, because any queue restart in
other paths may break blk_mq_quiesce_queue(). With the introduced
flag of QUEUE_FLAG_QUIESCED, we don't need to depend on stopping queue
for quiescing any more.
Signed-off-by: Ming Lei <ming.lei@redhat.com>
Reviewed-by: Bart Van Assche <Bart.VanAssche@sandisk.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
blk_mq_sched_assign_ioc now only handles the assigned of the ioc if
the schedule needs it (bfq only at the moment). The caller to the
per-request initializer is moved out so that it can be merged with
a similar call for the kyber I/O scheduler.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Having these as separate helpers in a header really does not help
readability, or my chances to refactor this code sanely.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Having them out of line in blk-mq-sched.c just makes the code flow
unnecessarily complicated.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Because what the per-sw-queue bio merge does is basically same with
scheduler's .bio_merge(), this patch makes per-sw-queue bio merge
as the default .bio_merge if no scheduler is used or io scheduler
doesn't provide .bio_merge().
Signed-off-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
This provides the infrastructure for schedulers to expose their internal
state through debugfs. We add a list of queue attributes and a list of
hctx attributes to struct elevator_type and wire them up when switching
schedulers.
Signed-off-by: Omar Sandoval <osandov@fb.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Add missing seq_file.h header in blk-mq-debugfs.h
Signed-off-by: Jens Axboe <axboe@fb.com>
We have update the troublesome driver (mtip32xx) to deal with this
appropriately. So kill the hack that bypassed scheduler allocation
and insertion for reserved requests.
Reviewed-by: Ming Lei <ming.lei@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Tested-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
At least one driver, mtip32xx, has a hard coded dependency on
the value of the reserved tag used for internal commands. While
that should really be fixed up, for now let's ensure that we just
bypass the scheduler tags an allocation marked as reserved. They
are used for house keeping or error handling, so we can safely
ignore them in the scheduler.
Tested-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
commit c13660a08c ("blk-mq-sched: change ->dispatch_requests()
to ->dispatch_request()") removed the last user of this function.
Hence also remove the function itself.
Signed-off-by: Bart Van Assche <bart.vanassche@sandisk.com>
Cc: Omar Sandoval <osandov@fb.com>
Cc: Hannes Reinecke <hare@suse.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Schedulers need to be informed when a hardware queue is added or removed
at runtime so they can allocate/free per-hardware queue data. So,
replace the blk_mq_sched_init_hctx_data() helper, which only makes sense
at init time, with .init_hctx() and .exit_hctx() hooks.
Signed-off-by: Omar Sandoval <osandov@fb.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
To improve scalability, if hardware queues are shared, restart
a single hardware queue in round-robin fashion. Rename
blk_mq_sched_restart_queues() to reflect the new semantics.
Remove blk_mq_sched_mark_restart_queue() because this function
has no callers. Remove flag QUEUE_FLAG_RESTART because this
patch removes the code that uses this flag.
Signed-off-by: Bart Van Assche <bart.vanassche@sandisk.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Hannes Reinecke <hare@suse.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
In elevator_switch(), if blk_mq_init_sched() fails, we attempt to fall
back to the original scheduler. However, at this point, we've already
torn down the original scheduler's tags, so this causes a crash. Doing
the fallback like the legacy elevator path is much harder for mq, so fix
it by just falling back to none, instead.
Signed-off-by: Omar Sandoval <osandov@fb.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
If a new hardware queue is added at runtime, we don't allocate scheduler
tags for it, leading to a crash. This hooks up the scheduler framework
to blk_mq_{init,exit}_hctx() to make sure everything gets properly
initialized/freed.
Signed-off-by: Omar Sandoval <osandov@fb.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Preparation cleanup for the next couple of fixes, push
blk_mq_sched_setup() and e->ops.mq.init_sched() into a helper.
Signed-off-by: Omar Sandoval <osandov@fb.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
While dispatching requests, if we fail to get a driver tag, we mark the
hardware queue as waiting for a tag and put the requests on a
hctx->dispatch list to be run later when a driver tag is freed. However,
blk_mq_dispatch_rq_list() may dispatch requests from multiple hardware
queues if using a single-queue scheduler with a multiqueue device. If
blk_mq_get_driver_tag() fails, it doesn't update the hardware queue we
are processing. This means we end up using the hardware queue of the
previous request, which may or may not be the same as that of the
current request. If it isn't, the wrong hardware queue will end up
waiting for a tag, and the requests will be on the wrong dispatch list,
leading to a hang.
The fix is twofold:
1. Make sure we save which hardware queue we were trying to get a
request for in blk_mq_get_driver_tag() regardless of whether it
succeeds or not.
2. Make blk_mq_dispatch_rq_list() take a request_queue instead of a
blk_mq_hw_queue to make it clear that it must handle multiple
hardware queues, since I've already messed this up on a couple of
occasions.
This didn't appear in testing with nvme and mq-deadline because nvme has
more driver tags than the default number of scheduler tags. However,
with the blk_mq_update_nr_hw_queues() fix, it showed up with nbd.
Signed-off-by: Omar Sandoval <osandov@fb.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
No functional difference, it just makes a little more sense to update
the tag map where we actually allocate the tag.
Signed-off-by: Omar Sandoval <osandov@fb.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Tested-by: Sagi Grimberg <sagi@grimberg.me>
blk_mq_alloc_request_hctx() allocates a driver request directly, unlike
its blk_mq_alloc_request() counterpart. It also crashes because it
doesn't update the tags->rqs map.
Fix it by making it allocate a scheduler request.
Reported-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Omar Sandoval <osandov@fb.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Tested-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
Modified by me to also check at driver tag allocation time if the
original request was reserved, so we can be sure to allocate a
properly reserved tag at that point in time, too.
Signed-off-by: Jens Axboe <axboe@fb.com>
In blk_mq_sched_dispatch_requests(), we call blk_mq_sched_mark_restart()
after we dispatch requests left over on our hardware queue dispatch
list. This is so we'll go back and dispatch requests from the scheduler.
In this case, it's only necessary to restart the hardware queue that we
are running; there's no reason to run other hardware queues just because
we are using shared tags.
So, split out blk_mq_sched_mark_restart() into two operations, one for
just the hardware queue and one for the whole request queue. The core
code only needs the hctx variant, but I/O schedulers will want to use
both.
This also requires adjusting blk_mq_sched_restart_queues() to always
check the queue restart flag, not just when using shared tags.
Signed-off-by: Omar Sandoval <osandov@fb.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
The wording in the entries were poor and not understandable
by even deities. Kill the selection for default block scheduler,
and impose a policy with sane defaults.
Architected-by: Linus Torvalds <torvalds@linux-foundation.org>
Reviewed-by: Omar Sandoval <osandov@fb.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Usually we don't ask the scheduler for work, if we already have
leftovers on the dispatch list. This is done to leave work on
the scheduler side for as long as possible, for proper merging.
But if we do have work leftover but didn't dispatch anything,
then we should ask the scheduler since we could potentially
issue requests from that.
Signed-off-by: Jens Axboe <axboe@fb.com>
Reviewed-by: Omar Sandoval <osandov@fb.com>
If we are currently out of driver tags, we don't want to add a
new flush (without a tag) to the head of the requeue list. We
want to add it to the back, behind the others that are
potentially also waiting for a tag.
Signed-off-by: Jens Axboe <axboe@fb.com>
Reviewed-by: Omar Sandoval <osandov@fb.com>
bio is used in bfq-mq's get_rq_priv, to get the request group. We could
pass directly the group here, but I thought that passing the bio was
more general, giving the possibility to get other pieces of information
if needed.
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@fb.com>
Switch these constants to an enum, and make let the compiler ensure that
all callers of blk_try_merge and elv_merge handle all potential values.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
If we end up doing a request-to-request merge when we have completed
a bio-to-request merge, we free the request from deep down in that
path. For blk-mq-sched, the merge path has to hold the appropriate
lock, but we don't need it for freeing the request. And in fact
holding the lock is problematic, since we are now calling the
mq sched put_rq_private() hook with the lock held. Other call paths
do not hold this lock.
Fix this inconsistency by ensuring that the caller frees a merged
request. Then we can do it outside of the lock, making it both more
efficient and fixing the blk-mq-sched problem of invoking parts of
the scheduler with an unknown lock state.
Reported-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@fb.com>
Reviewed-by: Omar Sandoval <osandov@fb.com>
There's a weird inconsistency that flushes are mostly hidden from the
scheduler, but it needs to be aware of them in ->insert_requests().
Instead of having every scheduler call blk_mq_sched_bypass_insert(),
let's do it in the common framework.
Signed-off-by: Omar Sandoval <osandov@fb.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Instead of letting the caller check this and handle the details
of inserting a flush request, put the logic in the scheduler
insertion function. This fixes direct flush insertion outside
of the usual make_request_fn calls, like from dm via
blk_insert_cloned_request().
Signed-off-by: Jens Axboe <axboe@fb.com>
This centralizes the checks for bios that needs to be go into the flush
state machine.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Martin K. Petersen <martin.petersen@oracle.com>
Reviewed-by: Bart Van Assche <bart.vanassche@sandisk.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
When we invoke dispatch_requests(), the scheduler empties everything
into the passed in list. This isn't always a good thing, since it
means that we remove items that we could have potentially merged
with.
Change the function to dispatch single requests at the time. If
we do that, we can backoff exactly at the point where the device
can't consume more IO, and leave the rest with the scheduler for
better merging and future dispatch decision making.
Signed-off-by: Jens Axboe <axboe@fb.com>
Reviewed-by: Omar Sandoval <osandov@fb.com>
Tested-by: Hannes Reinecke <hare@suse.com>
If we have both multiple hardware queues and shared tag map between
devices, we need to ensure that we propagate the hardware queue
restart bit higher up. This is because we can get into a situation
where we don't have any IO pending on a hardware queue, yet we fail
getting a tag to start new IO. If that happens, it's not enough to
mark the hardware queue as needing a restart, we need to bubble
that up to the higher level queue as well.
Signed-off-by: Jens Axboe <axboe@fb.com>
Reviewed-by: Omar Sandoval <osandov@fb.com>
Tested-by: Hannes Reinecke <hare@suse.com>
We don't trigger this from the normal IO path, since we always use
blocking allocations from there. But Bart saw it testing multipath
dm, since that is a heavy user of atomic request allocations in
the map and clone path.
Reported-by: Bart Van Assche <bart.vanassche@sandisk.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
If we come in from blk_mq_alloc_requst() with NOWAIT set in flags,
we must ensure that we don't later overwrite that in
blk_mq_sched_get_request(). Initialize alloc_data->flags before
passing it in.
Reported-by: Bart Van Assche <bart.vanassche@sandisk.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Add Kconfig entries to manage what devices get assigned an MQ
scheduler, and add a blk-mq flag for drivers to opt out of scheduling.
The latter is useful for admin type queues that still allocate a blk-mq
queue and tag set, but aren't use for normal IO.
Signed-off-by: Jens Axboe <axboe@fb.com>
Reviewed-by: Bart Van Assche <bart.vanassche@sandisk.com>
Reviewed-by: Omar Sandoval <osandov@fb.com>
This adds a set of hooks that intercepts the blk-mq path of
allocating/inserting/issuing/completing requests, allowing
us to develop a scheduler within that framework.
We reuse the existing elevator scheduler API on the registration
side, but augment that with the scheduler flagging support for
the blk-mq interfce, and with a separate set of ops hooks for MQ
devices.
We split driver and scheduler tags, so we can run the scheduling
independently of device queue depth.
Signed-off-by: Jens Axboe <axboe@fb.com>
Reviewed-by: Bart Van Assche <bart.vanassche@sandisk.com>
Reviewed-by: Omar Sandoval <osandov@fb.com>
Jens Axboe