Even with a number of waitqueues, we can get into a situation where we
are heavily contended on the waitqueue lock. I got a report on spc1
where we're spending seconds doing this. Arguably the use case is nasty,
I reproduce it with one device and 1000 threads banging on the device.
But that doesn't mean we shouldn't be handling it better.

What ends up happening is that a thread will fail to get a tag, add
itself to the waitqueue, and subsequently get woken up when a tag is
freed - only to find itself going back to sleep on the waitqueue.

Instead of waking all threads, use an exclusive wait and wake up our
sbitmap batch count instead. This seems to work well for me (massive
improvement for this use case), and it survives basic testing. But I
haven't fully verified it yet.

An additional improvement is running the queue and checking for a new
tag BEFORE needing to add ourselves to the waitqueue.
Signed-off-by: NJens Axboe <axboe@kernel.dk>

This operation supports the use case of limiting the number of bits that
can be allocated for a given operation. Rather than setting aside some
bits at the end of the bitmap, we can set aside bits in each word of the
bitmap. This means we can keep the allocation hints spread out and
support sbitmap_resize() nicely at the cost of lower granularity for the
allowed depth.
Signed-off-by: NOmar Sandoval <osandov@fb.com>
Signed-off-by: NJens Axboe <axboe@fb.com>

<linux/kasan.h> is a low level header that is included early
in affected kernel headers. But it includes <linux/sched.h>
which complicates the cleanup of sched.h dependencies.

But kasan.h has almost no need for sched.h: its only use of
scheduler functionality is in two inline functions which are
not used very frequently - so uninline kasan_enable_current()
and kasan_disable_current().

Also add a <linux/sched.h> dependency to a .c file that depended
on kasan.h including it.

This paves the way to remove the <linux/sched.h> include from kasan.h.
Acked-by: NLinus Torvalds <torvalds@linux-foundation.org>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: linux-kernel@vger.kernel.org
Signed-off-by: NIngo Molnar <mingo@kernel.org>

This is useful debugging information that will be used in the blk-mq
debugfs directory.
Reviewed-by: NHannes Reinecke <hare@suse.com>
Signed-off-by: NOmar Sandoval <osandov@fb.com>

Changed 'weight' to 'busy'.
Signed-off-by: NJens Axboe <axboe@fb.com>

When we resize a struct sbitmap_queue, we update the wakeup batch size,
but we don't update the wait count in the struct sbq_wait_states. If we
resized down from a size which could use a bigger batch size, these
counts could be too large and cause us to miss necessary wakeups. To fix
this, update the wait counts when we resize (ensuring some careful
memory ordering so that it's safe w.r.t. concurrent clears).

This also fixes a theoretical issue where two threads could end up
bumping the wait count up by the batch size, which could also
potentially lead to hangs.
Reported-by: NMartin Raiber <martin@urbackup.org>
Fixes: e3a2b3f9 ("blk-mq: allow changing of queue depth through sysfs")
Fixes: 2971c35f ("blk-mq: bitmap tag: fix race on blk_mq_bitmap_tags::wake_cnt")
Signed-off-by: NOmar Sandoval <osandov@fb.com>
Signed-off-by: NJens Axboe <axboe@fb.com>

We always do an atomic clear_bit() right before we call sbq_wake_up(),
so we can use smp_mb__after_atomic(). While we're here, comment the
memory barriers in here a little more.
Signed-off-by: NOmar Sandoval <osandov@fb.com>
Signed-off-by: NJens Axboe <axboe@fb.com>

Variable weight is not being initialized to zero before it is
used to compute the weight sum. Ensure it is initialized to zero.

Found with static analysis with cppcheck:
[lib/sbitmap.c:177]: (error) Uninitialized variable: weight
Signed-off-by: NColin Ian King <colin.king@canonical.com>
Signed-off-by: NJens Axboe <axboe@fb.com>

If we have a bunch of high-numbered bits allocated and then we resize
the struct sbitmap_queue, when those bits get cleared, we'll update the
hint and then have to re-randomize it repeatedly. Avoid that by checking
that the cleared bit is still a valid hint. No measurable performance
difference in the common case.
Signed-off-by: NOmar Sandoval <osandov@fb.com>
Signed-off-by: NJens Axboe <axboe@fb.com>

After a struct sbitmap_queue is resized smaller, the allocation hints
may still be set to bits beyond the new depth of the bitmap. This means
that, for example, if the number of blk-mq tags is reduced through
sysfs, more requests than the nominal queue depth may be in flight.

It's tempting to fix this at resize time by doing a one-time
reinitialization of the hints, but this can race with
__sbitmap_queue_get() updating the hint. Instead, check the hint before
we use it. This caused no measurable performance difference in my
synthetic benchmarks.
Signed-off-by: NOmar Sandoval <osandov@fb.com>
Signed-off-by: NJens Axboe <axboe@fb.com>

In order to get good cache behavior from a sbitmap, we want each CPU to
stick to its own cacheline(s) as much as possible. This might happen
naturally as the bitmap gets filled up and the alloc_hint values spread
out, but we really want this behavior from the start. blk-mq apparently
intended to do this, but the code to do this was never wired up. Get rid
of the dead code and make it part of the sbitmap library.
Signed-off-by: NOmar Sandoval <osandov@fb.com>
Signed-off-by: NJens Axboe <axboe@fb.com>

Again, there's no point in passing this in every time. Make it part of
struct sbitmap_queue and clean up the API.
Signed-off-by: NOmar Sandoval <osandov@fb.com>
Signed-off-by: NJens Axboe <axboe@fb.com>

Allocating your own per-cpu allocation hint separately makes for an
awkward API. Instead, allocate the per-cpu hint as part of the struct
sbitmap_queue. There's no point for a struct sbitmap_queue without the
cache, but you can still use a bare struct sbitmap.
Signed-off-by: NOmar Sandoval <osandov@fb.com>
Signed-off-by: NJens Axboe <axboe@fb.com>

The original bt_alloc() we converted from was using kzalloc(), not
kzalloc_node(), to allocate the wait queues. This was probably an
oversight, so fix it for sbitmap_queue_init_node().
Signed-off-by: NOmar Sandoval <osandov@fb.com>
Signed-off-by: NJens Axboe <axboe@fb.com>

This is a generally useful data structure, so make it available to
anyone else who might want to use it. It's also a nice cleanup
separating the allocation logic from the rest of the tag handling logic.

The code is behind a new Kconfig option, CONFIG_SBITMAP, which is only
selected by CONFIG_BLOCK for now.

This should be a complete noop functionality-wise.
Signed-off-by: NOmar Sandoval <osandov@fb.com>
Signed-off-by: NJens Axboe <axboe@fb.com>