Storage controllers may expose multiple block devices that share hardware
resources managed by blk-mq. This patch enhances the shared tags so a
low-level driver can access the shared resources not tied to the unshared
h/w contexts. This way the LLD can dynamically add and delete disks and
request queues without having to track all the request_queue hctx's to
iterate outstanding tags.
Signed-off-by: NKeith Busch <keith.busch@intel.com>
Signed-off-by: NJens Axboe <axboe@fb.com>

When allocating from the reserved tags pool, bt_get() is called with
a NULL hctx.  If all tags are in use, the hw queue is kicked to push
out any pending IO, potentially freeing tags, and tag allocation is
retried.  The problem is that blk_mq_run_hw_queue() doesn't check for
a NULL hctx.  So we avoid it with a simple NULL hctx test.

Tested by hammering mtip32xx with concurrent smartctl/hdparm.
Signed-off-by: NSam Bradshaw <sbradshaw@micron.com>
Signed-off-by: NSelvan Mani <smani@micron.com>
Fixes: b3223207 ("blk-mq: fix hang in bt_get()")
Cc: stable@kernel.org

Added appropriate comment.
Signed-off-by: NJens Axboe <axboe@fb.com>

If the allocation of bt->bs fails, then bt->map can be freed twice, once
in blk_mq_init_bitmap_tags() -> bt_alloc(), and once in
blk_mq_init_bitmap_tags() -> bt_free().  Fix by setting the pointer to
NULL after the first free.

Cc: <stable@vger.kernel.org>
Signed-off-by: NTony Battersby <tonyb@cybernetics.com>
Signed-off-by: NJens Axboe <axboe@fb.com>

This is the blk-mq part to support tag allocation policy. The default
allocation policy isn't changed (though it's not a strict FIFO). The new
policy is round-robin for libata. But it's a try-best implementation. If
multiple tasks are competing, the tags returned will be mixed (which is
unavoidable even with !mq, as requests from different tasks can be
mixed in queue)

Cc: Jens Axboe <axboe@fb.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Christoph Hellwig <hch@infradead.org>
Signed-off-by: NShaohua Li <shli@fb.com>
Signed-off-by: NJens Axboe <axboe@fb.com>

The blk-mq tagging tries to maintain some locality between CPUs and
the tags issued. The tags are split into groups of words, and the
words may not be fully populated. When searching for a new free tag,
blk-mq may look at partial words, hence it passes in an offset/size
to find_next_zero_bit(). However, it does that wrong, the size must
always be the full length of the number of tags in that word,
otherwise we'll potentially miss some near the end.

Another issue is when __bt_get() goes from one word set to the next.
It bumps the index, but not the last_tag associated with the
previous index. Bump that to be in the range of the new word.

Finally, clean up __bt_get() and __bt_get_word() a bit and get
rid of the goto in there, and the unnecessary 'wrap' variable.
Signed-off-by: NJens Axboe <axboe@fb.com>

If it's dying, we can't expect new request to complete and come
in an wake up other tasks waiting for requests. So after we
have marked it as dying, wake up everybody currently waiting
for a request. Once they wake, they will retry their allocation
and fail appropriately due to the state of the queue.
Tested-by: NKeith Busch <keith.busch@intel.com>
Signed-off-by: NJens Axboe <axboe@fb.com>

This reverts commit 52f7eb94.

The optimization is only really safe for a single queue, otherwise
'bs' and 'bt' can indeed change, and if we don't do a finish_wait()
for each loop, we'll potentially change the wait structure and
corrupt task wait list.
Reported-by: NJan Kara <jack@suse.cz>

Remove a superfluous finish_wait() call. Convert the two bt_wait_ptr()
calls into a single call.
Signed-off-by: NBart Van Assche <bvanassche@acm.org>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Robert Elliott <elliott@hp.com>
Cc: Ming Lei <ming.lei@canonical.com>
Cc: Alexander Gordeev <agordeev@redhat.com>
Signed-off-by: NJens Axboe <axboe@fb.com>

What we need is the following two guarantees:
* Any thread that observes the effect of the test_and_set_bit() by
  __bt_get_word() also observes the preceding addition of 'current'
  to the appropriate wait list. This is guaranteed by the semantics
  of the spin_unlock() operation performed by prepare_and_wait().
  Hence the conversion of test_and_set_bit_lock() into
  test_and_set_bit().
* The wait lists are examined by bt_clear() after the tag bit has
  been cleared. clear_bit_unlock() guarantees that any thread that
  observes that the bit has been cleared also observes the store
  operations preceding clear_bit_unlock(). However,
  clear_bit_unlock() does not prevent that the wait lists are examined
  before that the tag bit is cleared. Hence the addition of a memory
  barrier between clear_bit() and the wait list examination.
Signed-off-by: NBart Van Assche <bvanassche@acm.org>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Robert Elliott <elliott@hp.com>
Cc: Ming Lei <ming.lei@canonical.com>
Cc: Alexander Gordeev <agordeev@redhat.com>
Cc: <stable@vger.kernel.org> # v3.13+
Signed-off-by: NJens Axboe <axboe@fb.com>

If __bt_get_word() is called with last_tag != 0, if the first
find_next_zero_bit() fails, if after wrap-around the
test_and_set_bit() call fails and find_next_zero_bit() succeeds,
if the next test_and_set_bit() call fails and subsequently
find_next_zero_bit() does not find a zero bit, then another
wrap-around will occur. Avoid this by introducing an additional
local variable.
Signed-off-by: NBart Van Assche <bvanassche@acm.org>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Robert Elliott <elliott@hp.com>
Cc: Ming Lei <ming.lei@canonical.com>
Cc: Alexander Gordeev <agordeev@redhat.com>
Cc: <stable@vger.kernel.org> # v3.13+
Signed-off-by: NJens Axboe <axboe@fb.com>

If we run out of tags and have to sleep, we run the hardware queue
to kick pending IO into gear. During that run, we may have completed
requests, so re-check if we have free tags before going to sleep.
Signed-off-by: NJens Axboe <axboe@fb.com>

Avoid that if there are fewer hardware queues than CPU threads that
bt_get() can hang. The symptoms of the hang were as follows:

* All tags allocated for a particular hardware queue.
* (nr_tags) pending commands for that hardware queue.
* No pending commands for the software queues associated with that
  hardware queue.
Signed-off-by: NJens Axboe <axboe@fb.com>

We only call __blk_mq_put_tag() and __blk_mq_put_reserved_tag()
from blk_mq_put_tag(), so just inline the two calls instead of
having them as separate functions.
Signed-off-by: NJens Axboe <axboe@fb.com>

The queuecommand() callback functions in SCSI low-level drivers
need to know which hardware context has been selected by the
block layer. Since this information is not available in the
request structure, and since passing the hctx pointer directly to
the queuecommand callback function would require modification of
all SCSI LLDs, add a function to the block layer that allows to
query the hardware context index.
Signed-off-by: NBart Van Assche <bvanassche@acm.org>
Acked-by: NJens Axboe <axboe@kernel.dk>
Reviewed-by: NSagi Grimberg <sagig@mellanox.com>
Reviewed-by: NMartin K. Petersen <martin.petersen@oracle.com>
Signed-off-by: NChristoph Hellwig <hch@lst.de>

Eliminate a backwards goto statement from bt_clear_tag().
Signed-off-by: NBart Van Assche <bvanassche@acm.org>
Signed-off-by: NJens Axboe <axboe@fb.com>

We currently divide the queue depth by 4 as our batch wakeup
count, but we split the wakeups over BT_WAIT_QUEUES number of
wait queues. This defaults to 8. If the product of the resulting
batch wake count and BT_WAIT_QUEUES is higher than the device
queue depth, we can get into a situation where a task goes to
sleep waiting for a request, but never gets woken up.
Reported-by: NBart Van Assche <bvanassche@acm.org>
Fixes: 4bb659b1
Cc: stable@kernel.org
Signed-off-by: NJens Axboe <axboe@fb.com>

Don't do a kmalloc from timer to handle timeouts, chances are we could be
under heavy load or similar and thus just miss out on the timeouts.
Fortunately it is very easy to just iterate over all in use tags, and doing
this properly actually cleans up the blk_mq_busy_iter API as well, and
prepares us for the next patch by passing a reserved argument to the
iterator.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NJens Axboe <axboe@fb.com>

This update fixes few issues in bt_get() function:

- list_empty(&wait.task_list) check is not protected;

- was_empty check is always true which results in *every* thread
  entering the loop resets bt_wait_state::wait_cnt counter rather
  than every bt->wake_cnt'th thread;

- 'bt_wait_state::wait_cnt' counter update is redundant, since
  it also gets reset in bt_clear_tag() function;

Cc: Christoph Hellwig <hch@infradead.org>
Cc: Ming Lei <tom.leiming@gmail.com>
Cc: Jens Axboe <axboe@kernel.dk>
Signed-off-by: NAlexander Gordeev <agordeev@redhat.com>
Signed-off-by: NJens Axboe <axboe@fb.com>

This piece of code in bt_clear_tag() function is racy:

	bs = bt_wake_ptr(bt);
	if (bs && atomic_dec_and_test(&bs->wait_cnt)) {
		atomic_set(&bs->wait_cnt, bt->wake_cnt);
 		wake_up(&bs->wait);
	}

Since nothing prevents bt_wake_ptr() from returning the very
same 'bs' address on multiple CPUs, the following scenario is
possible:

    CPU1                                CPU2
    ----                                ----

0.  bs = bt_wake_ptr(bt);               bs = bt_wake_ptr(bt);
1.  atomic_dec_and_test(&bs->wait_cnt)
2.                                      atomic_dec_and_test(&bs->wait_cnt)
3.  atomic_set(&bs->wait_cnt, bt->wake_cnt);

If the decrement in [1] yields zero then for some amount of time
the decrement in [2] results in a negative/overflow value, which
is not expected. The follow-up assignment in [3] overwrites the
invalid value with the batch value (and likely prevents the issue
from being severe) which is still incorrect and should be a lesser.

Cc: Ming Lei <tom.leiming@gmail.com>
Cc: Jens Axboe <axboe@kernel.dk>
Signed-off-by: NAlexander Gordeev <agordeev@redhat.com>
Signed-off-by: NJens Axboe <axboe@fb.com>

Fix racy updates of shared blk_mq_bitmap_tags::wake_index
and blk_mq_hw_ctx::wake_index fields.

Cc: Ming Lei <tom.leiming@gmail.com>
Signed-off-by: NAlexander Gordeev <agordeev@redhat.com>
Signed-off-by: NJens 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: NMing Lei <tom.leiming@gmail.com>
Signed-off-by: NJens 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: NJens 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: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NJens Axboe <axboe@fb.com>

Export the blk-mq in-flight tag iterator for driver consumption.
This is particularly useful in exception paths or SRSI where
in-flight IOs need to be cancelled and/or reissued. The NVMe driver
conversion will use this.
Signed-off-by: NSam Bradshaw <sbradshaw@micron.com>
Signed-off-by: NMatias Bjørling <m@bjorling.me>
Signed-off-by: NJens 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: NJens Axboe <axboe@fb.com>

We will use it for the pending list in blk-mq core as well.
Signed-off-by: NJens 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: NJens Axboe <axboe@fb.com>

Both nr_cache and nr_tags arn't needed for bitmap tag anymore.
Signed-off-by: NMing Lei <tom.leiming@gmail.com>
Signed-off-by: NJens Axboe <axboe@fb.com>

The selected tag should be selected at random between 0 and
(depth - 1) with probability 1/depth, instead between 0 and
(depth - 2) with probability 1/(depth - 1).
Signed-off-by: NMing Lei <tom.leiming@gmail.com>
Signed-off-by: NJens Axboe <axboe@fb.com>

The barrier isn't necessary because both atomic_dec_and_test()
and wake_up() implicate one barrier.
Signed-off-by: NMing Lei <tom.leiming@gmail.com>
Signed-off-by: NJens Axboe <axboe@fb.com>

The unlock memory barrier need to order access to req in free
path and clearing tag bit, otherwise either request free path
may see a allocated request, or initialized request in allocate
path might be modified by the ongoing free path.
Signed-off-by: NMing Lei <tom.leiming@gmail.com>
Signed-off-by: NJens Axboe <axboe@fb.com>

For best performance, spreading tags over multiple cachelines
makes the tagging more efficient on multicore systems. But since
we have 8 * sizeof(unsigned long) tags per cacheline, we don't
always get a nice spread.

Attempt to spread the tags over at least 4 cachelines, using fewer
number of bits per unsigned long if we have to. This improves
tagging performance in setups with 32-128 tags. For higher depths,
the spread is the same as before (BITS_PER_LONG tags per cacheline).
Signed-off-by: NJens Axboe <axboe@fb.com>

blk-mq currently uses percpu_ida for tag allocation. But that only
works well if the ratio between tag space and number of CPUs is
sufficiently high. For most devices and systems, that is not the
case. The end result if that we either only utilize the tag space
partially, or we end up attempting to fully exhaust it and run
into lots of lock contention with stealing between CPUs. This is
not optimal.

This new tagging scheme is a hybrid bitmap allocator. It uses
two tricks to both be SMP friendly and allow full exhaustion
of the space:

1) We cache the last allocated (or freed) tag on a per blk-mq
   software context basis. This allows us to limit the space
   we have to search. The key element here is not caching it
   in the shared tag structure, otherwise we end up dirtying
   more shared cache lines on each allocate/free operation.

2) The tag space is split into cache line sized groups, and
   each context will start off randomly in that space. Even up
   to full utilization of the space, this divides the tag users
   efficiently into cache line groups, avoiding dirtying the same
   one both between allocators and between allocator and freeer.

This scheme shows drastically better behaviour, both on small
tag spaces but on large ones as well. It has been tested extensively
to show better performance for all the cases blk-mq cares about.
Signed-off-by: NJens Axboe <axboe@fb.com>

blk_mq_wait_for_tags() is only able to wait for "normal" tags,
not reserved tags. Pass in which one we should attempt to get
a tag for, so that waiting for reserved tags will work.

Reserved tags are used for internal commands, which are usually
serialized. Hence no waiting generally takes place, but we should
ensure that it actually works if users need that functionality.
Signed-off-by: NJens Axboe <axboe@fb.com>

Add a new blk_mq_tag_set structure that gets set up before we initialize
the queue.  A single blk_mq_tag_set structure can be shared by multiple
queues.
Signed-off-by: NChristoph Hellwig <hch@lst.de>

Modular export of blk_mq_{alloc,free}_tagset added by me.
Signed-off-by: NJens Axboe <axboe@fb.com>

cppcheck detected following format string mismatch.
[blk-mq-tag.c:201]: (warning) %u in format string (no. 1) requires
'unsigned int' but the argument type is 'int'.

Change "cpu" from int to unsigned int, because the cpu
never become minus value.
Signed-off-by: NMasanari Iida <standby24x7@gmail.com>
Signed-off-by: NJens Axboe <axboe@fb.com>

This patch changes percpu_ida_alloc() + callers to accept task state
bitmask for prepare_to_wait() for code like target/iscsi that needs
it for interruptible sleep, that is provided in a subsequent patch.

It now expects TASK_UNINTERRUPTIBLE when the caller is able to sleep
waiting for a new tag, or TASK_RUNNING when the caller cannot sleep,
and is forced to return a negative value when no tags are available.

v2 changes:
  - Include blk-mq + tcm_fc + vhost/scsi + target/iscsi changes
  - Drop signal_pending_state() call
v3 changes:
  - Only call prepare_to_wait() + finish_wait() when != TASK_RUNNING
    (PeterZ)
Reported-by: NLinus Torvalds <torvalds@linux-foundation.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Jens Axboe <axboe@kernel.dk>
Signed-off-by: NKent Overstreet <kmo@daterainc.com>
Cc: <stable@vger.kernel.org> #3.12+
Signed-off-by: NNicholas Bellinger <nab@linux-iscsi.org>

Linux currently has two models for block devices:

- The classic request_fn based approach, where drivers use struct
  request units for IO. The block layer provides various helper
  functionalities to let drivers share code, things like tag
  management, timeout handling, queueing, etc.

- The "stacked" approach, where a driver squeezes in between the
  block layer and IO submitter. Since this bypasses the IO stack,
  driver generally have to manage everything themselves.

With drivers being written for new high IOPS devices, the classic
request_fn based driver doesn't work well enough. The design dates
back to when both SMP and high IOPS was rare. It has problems with
scaling to bigger machines, and runs into scaling issues even on
smaller machines when you have IOPS in the hundreds of thousands
per device.

The stacked approach is then most often selected as the model
for the driver. But this means that everybody has to re-invent
everything, and along with that we get all the problems again
that the shared approach solved.

This commit introduces blk-mq, block multi queue support. The
design is centered around per-cpu queues for queueing IO, which
then funnel down into x number of hardware submission queues.
We might have a 1:1 mapping between the two, or it might be
an N:M mapping. That all depends on what the hardware supports.

blk-mq provides various helper functions, which include:

- Scalable support for request tagging. Most devices need to
  be able to uniquely identify a request both in the driver and
  to the hardware. The tagging uses per-cpu caches for freed
  tags, to enable cache hot reuse.

- Timeout handling without tracking request on a per-device
  basis. Basically the driver should be able to get a notification,
  if a request happens to fail.

- Optional support for non 1:1 mappings between issue and
  submission queues. blk-mq can redirect IO completions to the
  desired location.

- Support for per-request payloads. Drivers almost always need
  to associate a request structure with some driver private
  command structure. Drivers can tell blk-mq this at init time,
  and then any request handed to the driver will have the
  required size of memory associated with it.

- Support for merging of IO, and plugging. The stacked model
  gets neither of these. Even for high IOPS devices, merging
  sequential IO reduces per-command overhead and thus
  increases bandwidth.

For now, this is provided as a potential 3rd queueing model, with
the hope being that, as it matures, it can replace both the classic
and stacked model. That would get us back to having just 1 real
model for block devices, leaving the stacked approach to dm/md
devices (as it was originally intended).

Contributions in this patch from the following people:

Shaohua Li <shli@fusionio.com>
Alexander Gordeev <agordeev@redhat.com>
Christoph Hellwig <hch@infradead.org>
Mike Christie <michaelc@cs.wisc.edu>
Matias Bjorling <m@bjorling.me>
Jeff Moyer <jmoyer@redhat.com>
Acked-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NJens Axboe <axboe@kernel.dk>