Currently, blk-mq uses a percpu_counter to keep track of how many
usages are in flight. The percpu_counter is drained while freezing to
ensure that no usage is left in-flight after freezing is complete.
blk_mq_queue_enter/exit() and blk_mq_[un]freeze_queue() implement this
per-cpu gating mechanism.
This type of code has relatively high chance of subtle bugs which are
extremely difficult to trigger and it's way too hairy to be open coded
in blk-mq. percpu_ref can serve the same purpose after the recent
changes. This patch replaces the open-coded per-cpu usage counting
and draining mechanism with percpu_ref.
blk_mq_queue_enter() performs tryget_live on the ref and exit()
performs put. blk_mq_freeze_queue() kills the ref and waits until the
reference count reaches zero. blk_mq_unfreeze_queue() revives the ref
and wakes up the waiters.
Signed-off-by: Tejun Heo <tj@kernel.org>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Nicholas A. Bellinger <nab@linux-iscsi.org>
Cc: Kent Overstreet <kmo@daterainc.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Keeping __blk_mq_drain_queue() as a separate function doesn't buy us
anything and it's gonna be further simplified. Let's flatten it into
its caller.
This patch doesn't make any functional change.
Signed-off-by: Tejun Heo <tj@kernel.org>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Nicholas A. Bellinger <nab@linux-iscsi.org>
Signed-off-by: Jens Axboe <axboe@fb.com>
blk_mq freezing is entangled with generic bypassing which bypasses
blkcg and io scheduler and lets IO requests fall through the block
layer to the drivers in FIFO order. This allows forward progress on
IOs with the advanced features disabled so that those features can be
configured or altered without worrying about stalling IO which may
lead to deadlock through memory allocation.
However, generic bypassing doesn't quite fit blk-mq. blk-mq currently
doesn't make use of blkcg or ioscheds and it maps bypssing to
freezing, which blocks request processing and drains all the in-flight
ones. This causes problems as bypassing assumes that request
processing is online. blk-mq works around this by conditionally
allowing request processing for the problem case - during queue
initialization.
Another weirdity is that except for during queue cleanup, bypassing
started on the generic side prevents blk-mq from processing new
requests but doesn't drain the in-flight ones. This shouldn't break
anything but again highlights that something isn't quite right here.
The root cause is conflating blk-mq freezing and generic bypassing
which are two different mechanisms. The only intersecting purpose
that they serve is during queue cleanup. Let's properly separate
blk-mq freezing from generic bypassing and simply use it where
necessary.
* request_queue->mq_freeze_depth is added and
blk_mq_[un]freeze_queue() now operate on this counter instead of
->bypass_depth. The replacement for QUEUE_FLAG_BYPASS isn't added
but the counter is tested directly. This will be further updated by
later changes.
* blk_mq_drain_queue() is dropped and "__" prefix is dropped from
blk_mq_freeze_queue(). Queue cleanup path now calls
blk_mq_freeze_queue() directly.
* blk_queue_enter()'s fast path condition is simplified to simply
check @q->mq_freeze_depth. Previously, the condition was
!blk_queue_dying(q) &&
(!blk_queue_bypass(q) || !blk_queue_init_done(q))
mq_freeze_depth is incremented right after dying is set and
blk_queue_init_done() exception isn't necessary as blk-mq doesn't
start frozen, which only leaves the blk_queue_bypass() test which
can be replaced by @q->mq_freeze_depth test.
This change simplifies the code and reduces confusion in the area.
Signed-off-by: Tejun Heo <tj@kernel.org>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Nicholas A. Bellinger <nab@linux-iscsi.org>
Signed-off-by: Jens Axboe <axboe@fb.com>
Currently, both blk_queue_bypass_start() and blk_mq_freeze_queue()
skip queue draining if bypass_depth was already above zero. The
assumption is that the one which bumped the bypass_depth should have
performed draining already; however, there's nothing which prevents a
new instance of bypassing/freezing from starting before the previous
one finishes draining. The current code may allow the later
bypassing/freezing instances to complete while there still are
in-flight requests which haven't finished draining.
Fix it by draining regardless of bypass_depth. We still skip draining
from blk_queue_bypass_start() while the queue is initializing to avoid
introducing excessive delays during boot. INIT_DONE setting is moved
above the initial blk_queue_bypass_end() so that bypassing attempts
can't slip inbetween.
Signed-off-by: Tejun Heo <tj@kernel.org>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Nicholas A. Bellinger <nab@linux-iscsi.org>
Signed-off-by: Jens Axboe <axboe@fb.com>
blk-mq uses a percpu_counter to keep track of how many usages are in
flight. The percpu_counter is drained while freezing to ensure that
no usage is left in-flight after freezing is complete.
blk_mq_queue_enter/exit() and blk_mq_[un]freeze_queue() implement this
per-cpu gating mechanism; unfortunately, it contains a subtle bug -
smp_wmb() in blk_mq_queue_enter() doesn't prevent prevent the cpu from
fetching @q->bypass_depth before incrementing @q->mq_usage_counter and
if freezing happens inbetween the caller can slip through and freezing
can be complete while there are active users.
Use smp_mb() instead so that bypass_depth and mq_usage_counter
modifications and tests are properly interlocked.
Signed-off-by: Tejun Heo <tj@kernel.org>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Nicholas A. Bellinger <nab@linux-iscsi.org>
Signed-off-by: Jens Axboe <axboe@fb.com>
Currently it calls __blk_mq_run_hw_queue(), which depends on the
CPU placement being correct. This means it's not possible to call
blk_mq_start_hw_queues(q) from a context that is correct for all
queues, leading to triggering the
WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask));
in __blk_mq_run_hw_queue().
Reported-by: Ming Lei <tom.leiming@gmail.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
If we need to drain a queue we need to run all queues, even if they
are marked stopped to make sure the driver has a chance to error out
on all queued requests.
This fixes surprise removal with scsi-mq.
Reported-by: Bart Van Assche <bvanassche@acm.org>
Tested-by: Bart Van Assche <bvanassche@acm.org>
Signed-off-by: Jens Axboe <axboe@fb.com>
This way will become consistent with non-mq case, also
avoid to update rq->deadline twice for mq.
The comment said: "We do this early, to ensure we are on
the right CPU.", but no percpu stuff is used in blk_add_timer(),
so it isn't necessary. Even when inserting from plug list, there
is no such guarantee at all.
Signed-off-by: Ming Lei <tom.leiming@gmail.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
The blk-mq core only initializes this if io stats are enabled, since
blk-mq only reads the field in that case. But drivers could
potentially use it internally, so ensure that we always set it to
the current time when the request is allocated.
Reported-by: Ming Lei <tom.leiming@gmail.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
It'll be used in blk_mq_start_request() to set a potential timeout
for the request, so clear it to zero at alloc time to ensure that
we know if someone has set it or not.
Fixes random early timeouts on NVMe testing.
Signed-off-by: Jens Axboe <axboe@fb.com>
If the queue is going away, don't let new allocs or queueing
happen on it. Go through the normal wait process, and exit with
ENODEV in that case.
Signed-off-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
For some scsi-mq cases, the tag map can be huge. So increase the
max number of tags we support.
Additionally, don't fail with EINVAL if a user requests too many
tags. Warn that the tag depth has been adjusted down, and store
the new value inside the tag_set passed in.
Signed-off-by: Jens Axboe <axboe@fb.com>
We currently pass in the hardware queue, and get the tags from there.
But from scsi-mq, with a shared tag space, it's a lot more convenient
to pass in the blk_mq_tags instead as the hardware queue isn't always
directly available. So instead of having to re-map to a given
hardware queue from rq->mq_ctx, just pass in the tags structure.
Signed-off-by: Jens Axboe <axboe@fb.com>
When the code was collapsed to avoid duplication, the recent patch
for ensuring that a queue is idled before free was dropped, which was
added by commit 19c5d84f14.
Add back the blk_mq_tag_idle(), to ensure we don't leak a reference
to an active queue when it is freed.
Signed-off-by: Jens Axboe <axboe@fb.com>
blk_mq_map_request() can return NULL if we fail entering the queue
(dying, or removed), in which case it has already ended IO on the
bio. So nothing more to do, except just return.
Signed-off-by: Jens Axboe <axboe@fb.com>
'struct blk_mq_ctx' is __percpu, so add the annotation
and fix the sparse warning reported from Fengguang:
[block:for-linus 2/3] block/blk-mq.h:75:16: sparse: incorrect
type in initializer (different address spaces)
Reported-by: kbuild test robot <fengguang.wu@intel.com>
Signed-off-by: Ming Lei <tom.leiming@gmail.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
blk_mq_put_ctx() has to be called before io_schedule() in
bt_get().
This patch fixes the problem by taking similar approach from
percpu_ida allocation for the situation.
Signed-off-by: Ming Lei <tom.leiming@gmail.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
We have callers outside of the blk-mq proper (like timeouts) that
want to call __blk_mq_complete_request(), so rename the function
and put the decision code for whether to use ->softirq_done_fn
or blk_mq_endio() into __blk_mq_complete_request().
This also makes the interface more logical again.
blk_mq_complete_request() attempts to atomically mark the request
completed, and calls __blk_mq_complete_request() if successful.
__blk_mq_complete_request() then just ends the request.
Signed-off-by: Jens Axboe <axboe@fb.com>
Commit 07068d5b8e added a direct-to-hw-queue mode, but this mode
needs to remember to add the request timeout handler as well.
Without it, we don't track timeouts for these requests.
Signed-off-by: Jens Axboe <axboe@fb.com>
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>
flush request is special, which borrows the tag from the parent
request. Hence blk_mq_tag_to_rq needs special handling to return
the flush request from the tag.
Signed-off-by: Shaohua Li <shli@fusionio.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
We currently clear a lot more than we need to, so make that a bit
more clever. Make some of the init dependent on features, like
only setting start_time if we are going to use it.
Signed-off-by: Jens Axboe <axboe@fb.com>
If devices are not SG starved, we waste a lot of time potentially
collapsing SG segments. Enough that 1.5% of the CPU time goes
to this, at only 400K IOPS. Add a queue flag, QUEUE_FLAG_NO_SG_MERGE,
which just returns the number of vectors in a bio instead of looping
over all segments and checking for collapsible ones.
Add a BLK_MQ_F_SG_MERGE flag so that drivers can opt-in on the sg
merging, if they so desire.
Signed-off-by: Jens Axboe <axboe@fb.com>
There is no need for drivers to control hardware context allocation
now that we do the context to node mapping in common code.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
None of the blk-mq files have an explanatory comment at the top
for what that particular file does. Add that and add appropriate
copyright notices as well.
Signed-off-by: Jens Axboe <axboe@fb.com>
We now only have one caller left and can open code it there in a cleaner
way.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
We already do a non-blocking allocation in blk_mq_map_request, no need
to repeat it. Just call __blk_mq_alloc_request to wait directly.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
The current logic for blocking tag allocation is rather confusing, as we
first allocated and then free again a tag in blk_mq_wait_for_tags, just
to attempt a non-blocking allocation and then repeat if someone else
managed to grab the tag before us.
Instead change blk_mq_alloc_request_pinned to simply do a blocking tag
allocation itself and use the request we get back from it.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
Both callers if __blk_mq_alloc_request want to initialize the request, so
lift it into the common path.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
Instead of having two almost identical copies of the same code just let
the callers pass in the reserved flag directly.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
Both the cache flush state machine and the SCSI midlayer want to submit
requests from irq context, and the current per-request requeue_work
unfortunately causes corruption due to sharing with the csd field for
flushes. Replace them with a per-request_queue list of requests to
be requeued.
Based on an earlier test by Ming Lei.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reported-by: Ming Lei <tom.leiming@gmail.com>
Tested-by: Ming Lei <tom.leiming@gmail.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Right now we export two ways of completing a request:
1) blk_mq_complete_request(). This uses an IPI (if needed) and
completes through q->softirq_done_fn(). It also works with
timeouts.
2) blk_mq_end_io(). This completes inline, and ignores any timeout
state of the request.
Let blk_mq_complete_request() handle non-softirq_done_fn completions
as well, by just completing inline. If a driver has enough completion
ports to place completions correctly, it need not define a
mq_ops->complete() and we can avoid an indirect function call by
doing the completion inline.
Signed-off-by: Jens Axboe <axboe@fb.com>
Drivers currently have to figure this out on their own, and they
are missing information to do it properly. The ones that did
attempt to do it, do it wrong.
So just pass in the suggested node directly to the alloc
function.
Signed-off-by: Jens Axboe <axboe@fb.com>
The percpu counter is only used for blk-mq, so move
its allocation and free inside blk-mq, and don't
allocate it for legacy queue device.
Signed-off-by: Ming Lei <tom.leiming@gmail.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
blk_mq_exit_hw_queues() and blk_mq_free_hw_queues()
are introduced to avoid code duplication.
Signed-off-by: Ming Lei <tom.leiming@gmail.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
hctx->ctx_map should have been freed inside blk_mq_free_queue().
Signed-off-by: Ming Lei <tom.leiming@gmail.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Without this we can leak the active_queues reference if a command is
freed while it is considered active.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
Currently blk-mq uses the queue timeout for all requests. But
for some commands, drivers may want to set a specific timeout
for special requests. Allow this to be passed in through
request->timeout, and use it if set.
Signed-off-by: Jens Axboe <axboe@fb.com>
We want slightly different behavior from them:
- On single queue devices, we currently use the per-process plug
for deferred IO and for merging.
- On multi queue devices, we don't use the per-process plug, but
we want to go straight to hardware for SYNC IO.
Split blk_mq_make_request() into a blk_sq_make_request() for single
queue devices, and retain blk_mq_make_request() for multi queue
devices. Then we don't need multiple checks for q->nr_hw_queues
in the request mapping.
Signed-off-by: Jens Axboe <axboe@fb.com>
Depending on the topology of the machine and the number of queues
exposed by a device, we can end up in a situation where some of
the hardware queues are unused (as in, they don't map to any
software queues). For this case, free up the memory used by the
request map, as we will not use it. This can be a substantial
amount of memory, depending on the number of queues vs CPUs and
the queue depth of the device.
Signed-off-by: Jens Axboe <axboe@fb.com>
Prepare this for the next patch which adds more smarts in the
plugging logic, so that we can save some memory.
Signed-off-by: Jens Axboe <axboe@fb.com>
In blk_mq_make_request(), do the blk_queue_nomerges() check
outside the call to blk_attempt_plug_merge() to eliminate
function call overhead when nomerges=2 (disabled)
Signed-off-by: Robert Elliott <elliott@hp.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
blk_queue_make_requests() overwrites our set value for q->nr_requests,
turning it into the default of 128. Set this appropriately after
initializing queue values in blk_queue_make_request().
Signed-off-by: Jens Axboe <axboe@fb.com>
For request_fn based devices, the block layer exports a 'nr_requests'
file through sysfs to allow adjusting of queue depth on the fly.
Currently this returns -EINVAL for blk-mq, since it's not wired up.
Wire this up for blk-mq, so that it now also always dynamic
adjustments of the allowed queue depth for any given block device
managed by blk-mq.
Signed-off-by: Jens Axboe <axboe@fb.com>
Each hardware queue has a bitmap of software queues with pending
requests. When new IO is queued on a software queue, the bit is
set, and when IO is pruned on a hardware queue run, the bit is
cleared. This causes a lot of traffic. Switch this from the regular
BITS_PER_LONG bitmap to a sparser layout, similarly to what was
done for blk-mq tagging.
20% performance increase was observed for single threaded IO, and
about 15% performanc increase on multiple threads driving the
same device.
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>
Commit c6d600c6 opened up a small race where we could attempt to
account IO completion on a request, racing with IO start accounting.
Fix this up by ensuring that we've accounted for IO start before
inserting the request.
Signed-off-by: Jens Axboe <axboe@fb.com>
Tejun Heo
Jens Axboe
Christoph Hellwig
Ming Lei
Keith Busch
Shaohua Li
Robert Elliott