nbd: make sure request completion won't concurrent

commit cddce0116058 ("nbd: Aovid double completion of a request")
try to fix that nbd_clear_que() and recv_work() can complete a
request concurrently. However, the problem still exists:

t1                    t2                     t3

nbd_disconnect_and_put
 flush_workqueue
                      recv_work
                       blk_mq_complete_request
                        blk_mq_complete_request_remote -> this is true
                         WRITE_ONCE(rq->state, MQ_RQ_COMPLETE)
                          blk_mq_raise_softirq
                                             blk_done_softirq
                                              blk_complete_reqs
                                               nbd_complete_rq
                                                blk_mq_end_request
                                                 blk_mq_free_request
                                                  WRITE_ONCE(rq->state, MQ_RQ_IDLE)
  nbd_clear_que
   blk_mq_tagset_busy_iter
    nbd_clear_req
                                                   __blk_mq_free_request
                                                    blk_mq_put_tag
     blk_mq_complete_request -> complete again

There are three places where request can be completed in nbd:
recv_work(), nbd_clear_que() and nbd_xmit_timeout(). Since they
all hold cmd->lock before completing the request, it's easy to
avoid the problem by setting and checking a cmd flag.

Signed-off-by: Yu Kuai <yukuai3@huawei.com>
Reviewed-by: Ming Lei <ming.lei@redhat.com>
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Link: https://lore.kernel.org/r/20210916093350.1410403-3-yukuai3@huawei.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>

nbd: don't handle response without a corresponding request message

While handling a response message from server, nbd_read_stat() will
try to get request by tag, and then complete the request. However,
this is problematic if nbd haven't sent a corresponding request
message:

t1                      t2
                        submit_bio
                         nbd_queue_rq
                          blk_mq_start_request
recv_work
 nbd_read_stat
  blk_mq_tag_to_rq
 blk_mq_complete_request
                          nbd_send_cmd

Thus add a new cmd flag 'NBD_CMD_INFLIGHT', it will be set in
nbd_send_cmd() and checked in nbd_read_stat().

Noted that this patch can't fix that blk_mq_tag_to_rq() might
return a freed request, and this will be fixed in following
patches.

Signed-off-by: Yu Kuai <yukuai3@huawei.com>
Reviewed-by: Ming Lei <ming.lei@redhat.com>
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Link: https://lore.kernel.org/r/20210916093350.1410403-2-yukuai3@huawei.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>

mtip32xx: Remove redundant 'flush_workqueue()' calls

'destroy_workqueue()' already drains the queue before destroying it, so
there is no need to flush it explicitly.

Remove the redundant 'flush_workqueue()' calls.

This was generated with coccinelle:

@@
expression E;
@@
-  flush_workqueue(E);
destroy_workqueue(E);

Signed-off-by: Christophe JAILLET <christophe.jaillet@wanadoo.fr>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Link: https://lore.kernel.org/r/0fea349c808c6cfbf549b0e33701320c7860c8b7.1634234221.git.christophe.jaillet@wanadoo.fr
Signed-off-by: Jens Axboe <axboe@kernel.dk>

md: update superblock after changing rdev flags in state_store

When the in memory flag is changed, we need to persist the change in the
rdev superblock flags. This is needed for "writemostly" and "failfast".

Reviewed-by: Li Feng <fengli@smartx.com>
Signed-off-by: Xiao Ni <xni@redhat.com>
Signed-off-by: Song Liu <songliubraving@fb.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

md: remove unused argument from md_new_event

Actually, mddev is not used by md_new_event.

Signed-off-by: Guoqing Jiang <guoqing.jiang@linux.dev>
Signed-off-by: Song Liu <songliubraving@fb.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

md/raid5: call roundup_pow_of_two in raid5_run

Let's call roundup_pow_of_two here instead of open code.

Signed-off-by: Guoqing Jiang <guoqing.jiang@linux.dev>
Signed-off-by: Song Liu <songliubraving@fb.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

md/raid1: use rdev in raid1_write_request directly

We already get rdev from conf->mirrors[i].rdev at the beginning of the
loop, so just use it.

Signed-off-by: Guoqing Jiang <guoqing.jiang@linux.dev>
Signed-off-by: Song Liu <songliubraving@fb.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

md/raid1: only allocate write behind bio for WriteMostly device

Commit 6607cd319b6b91bff94e90f798a61c031650b514 ("raid1: ensure write
behind bio has less than BIO_MAX_VECS sectors") tried to guarantee the
size of behind bio is not bigger than BIO_MAX_VECS sectors.

Unfortunately the same calltrace still could happen since an array could
enable write-behind without write mostly device.

To match the manpage of mdadm (which says "write-behind is only attempted
on drives marked as write-mostly"), we need to check WriteMostly flag to
avoid such unexpected behavior.

[1]. https://bugzilla.kernel.org/show_bug.cgi?id=213181#c25

Cc: stable@vger.kernel.org # v5.12+
Cc: Jens Stutte <jens@chianterastutte.eu>
Reported-by: Jens Stutte <jens@chianterastutte.eu>
Signed-off-by: Guoqing Jiang <guoqing.jiang@linux.dev>
Signed-off-by: Song Liu <songliubraving@fb.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

md: properly unwind when failing to add the kobject in md_alloc

Add proper error handling to delete the gendisk when failing to add
the md kobject and clean up the error unwinding in general.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Song Liu <songliubraving@fb.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

md: extend disks_mutex coverage

disks_mutex is intended to serialize md_alloc.  Extended it to also cover
the kobject_uevent call and getting the sysfs dirent to help reducing
error handling complexity.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Song Liu <songliubraving@fb.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

md: add the bitmap group to the default groups for the md kobject

Replace the deprecated default_attrs with the default_groups mechanism,
and add the always visible bitmap group to the groups created add
kobject_add time.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Song Liu <songliubraving@fb.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

md: add error handling support for add_disk()

We never checked for errors on add_disk() as this function
returned void. Now that this is fixed, use the shiny new
error handling.

We just do the unwinding of what was not done before, and are
sure to unlock prior to bailing.

Signed-off-by: Luis Chamberlain <mcgrof@kernel.org>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Song Liu <songliubraving@fb.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

swim3: add missing major.h include

swim3 got this through blkdev.h previously, but blkdev.h is not including
it anymore. Include it specifically for the driver, otherwise FLOPPY_MAJOR
is undefined and breaks the compile on PPC if swim3 is configured.

Fixes: b81e0c2372e6 ("block: drop unused includes in <linux/genhd.h>")
Reported-by: Naresh Kamboju <naresh.kamboju@linaro.org>
Acked-by: Randy Dunlap <rdunlap@infradead.org> # build-tested
Signed-off-by: Jens Axboe <axboe@kernel.dk>

sx8: fix an error code in carm_init_one()

Return a negative error code here on this error path instead of
returning success.

Fixes: 637208e74a86 ("block/sx8: add error handling support for add_disk()")
Signed-off-by: Dan Carpenter <dan.carpenter@oracle.com>
Link: https://lore.kernel.org/r/20211001122722.GC2283@kili
Signed-off-by: Jens Axboe <axboe@kernel.dk>

pf: fix error codes in pf_init_unit()

Return a negative error code instead of success on these error paths.

Fixes: fb367e6baeb0 ("pf: cleanup initialization")
Signed-off-by: Dan Carpenter <dan.carpenter@oracle.com>
Link: https://lore.kernel.org/r/20211001122654.GB2283@kili
Signed-off-by: Jens Axboe <axboe@kernel.dk>

pcd: fix error codes in pcd_init_unit()

Return -ENODEV on these error paths instead of returning success.

Fixes: af761f277b7f ("pcd: cleanup initialization")
Signed-off-by: Dan Carpenter <dan.carpenter@oracle.com>
Link: https://lore.kernel.org/r/20211001122623.GA2283@kili
Signed-off-by: Jens Axboe <axboe@kernel.dk>

xtensa/platforms/iss/simdisk: add error handling support for add_disk()

We never checked for errors on add_disk() as this function
returned void. Now that this is fixed, use the shiny new
error handling.

Signed-off-by: Luis Chamberlain <mcgrof@kernel.org>
Acked-by: Max Filippov <jcmvbkbc@gmail.com>
Link: https://lore.kernel.org/r/20210927220110.1066271-7-mcgrof@kernel.org
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block/ataflop: add error handling support for add_disk()

We never checked for errors on add_disk() as this function
returned void. Now that this is fixed, use the shiny new
error handling.

Signed-off-by: Luis Chamberlain <mcgrof@kernel.org>
Link: https://lore.kernel.org/r/20210927220302.1073499-15-mcgrof@kernel.org
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block/ataflop: provide a helper for cleanup up an atari disk

Instead of using two separate code paths for cleaning up an atari disk,
use one. We take the more careful approach to check for *all* disk
types, as is done on exit. The init path didn't have that check as
the alternative disk types are only probed for later, they are not
initialized by default.

Yes, there is a shared tag for all disks.

Signed-off-by: Luis Chamberlain <mcgrof@kernel.org>
Link: https://lore.kernel.org/r/20210927220302.1073499-14-mcgrof@kernel.org
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block/ataflop: add registration bool before calling del_gendisk()

The ataflop assumes del_gendisk() is safe to call, this is only
true because add_disk() does not return a failure, but that will
change soon. And so, before we get to adding error handling for
that case, let's make sure we keep track of which disks actually
get registered. Then we use this to only call del_gendisk for them.

Signed-off-by: Luis Chamberlain <mcgrof@kernel.org>
Link: https://lore.kernel.org/r/20210927220302.1073499-13-mcgrof@kernel.org
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block/ataflop: use the blk_cleanup_disk() helper

Use the helper to replace two lines with one.

Signed-off-by: Luis Chamberlain <mcgrof@kernel.org>
Link: https://lore.kernel.org/r/20210927220302.1073499-12-mcgrof@kernel.org
Signed-off-by: Jens Axboe <axboe@kernel.dk>

swim: add error handling support for add_disk()

We never checked for errors on add_disk() as this function
returned void. Now that this is fixed, use the shiny new
error handling.

Since we have a caller to do our unwinding for the disk,
and this is already dealt with safely we can re-use our
existing error path goto label which already deals with
the cleanup.

Signed-off-by: Luis Chamberlain <mcgrof@kernel.org>
Link: https://lore.kernel.org/r/20210927220302.1073499-11-mcgrof@kernel.org
Signed-off-by: Jens Axboe <axboe@kernel.dk>

swim: add a floppy registration bool which triggers del_gendisk()

Instead of calling del_gendisk() on exit alone, let's add
a registration bool to the floppy disk state, this way this can
be done on the shared caller, swim_cleanup_floppy_disk().

This will be more useful in subsequent patches. Right now, this
just shuffles functionality out to a helper in a safe way.

Signed-off-by: Luis Chamberlain <mcgrof@kernel.org>
Link: https://lore.kernel.org/r/20210927220302.1073499-10-mcgrof@kernel.org
Signed-off-by: Jens Axboe <axboe@kernel.dk>

swim: add helper for disk cleanup

Disk cleanup can be shared between exit and bringup. Use a
helper to do the work required. The only functional change at
this point is we're being overly paraoid on exit to check for
a null disk as well now, and this should be safe.

We'll later expand on this, this change just makes subsequent
changes easier to read.

Signed-off-by: Luis Chamberlain <mcgrof@kernel.org>
Link: https://lore.kernel.org/r/20210927220302.1073499-9-mcgrof@kernel.org
Signed-off-by: Jens Axboe <axboe@kernel.dk>

swim: simplify using blk_cleanup_disk() on swim_remove()

We can simplify swim_remove() by using one call instead of two,
just as other drivers do. Use that pattern.

Signed-off-by: Luis Chamberlain <mcgrof@kernel.org>
Link: https://lore.kernel.org/r/20210927220302.1073499-8-mcgrof@kernel.org
Signed-off-by: Jens Axboe <axboe@kernel.dk>

amiflop: add error handling support for add_disk()

We never checked for errors on add_disk() as this function
returned void. Now that this is fixed, use the shiny new
error handling. The caller for fd_alloc_disk() deals with
the rest of the cleanup like the tag.

Signed-off-by: Luis Chamberlain <mcgrof@kernel.org>
Link: https://lore.kernel.org/r/20210927220302.1073499-7-mcgrof@kernel.org
Signed-off-by: Jens Axboe <axboe@kernel.dk>

floppy: add error handling support for add_disk()

We never checked for errors on add_disk() as this function
returned void. Now that this is fixed, use the shiny new
error handling.

Signed-off-by: Luis Chamberlain <mcgrof@kernel.org>
Link: https://lore.kernel.org/r/20210927220302.1073499-6-mcgrof@kernel.org
Signed-off-by: Jens Axboe <axboe@kernel.dk>

floppy: fix calling platform_device_unregister() on invalid drives

platform_device_unregister() should only be called when
a respective platform_device_register() is called. However
the floppy driver currently allows failures when registring
a drive and a bail out could easily cause an invalid call
to platform_device_unregister() where it was not intended.

Fix this by adding a bool to keep track of when the platform
device was registered for a drive.

This does not fix any known panic / bug. This issue was found
through code inspection while preparing the driver to use the
up and coming support for device_add_disk() error handling.
From what I can tell from code inspection, chances of this
ever happening should be insanely small, perhaps OOM.

Signed-off-by: Luis Chamberlain <mcgrof@kernel.org>
Link: https://lore.kernel.org/r/20210927220302.1073499-5-mcgrof@kernel.org
Signed-off-by: Jens Axboe <axboe@kernel.dk>

floppy: use blk_cleanup_disk()

Use the blk_cleanup_queue() followed by put_disk() can be
replaced with blk_cleanup_disk(). No need for two separate
loops.

Signed-off-by: Luis Chamberlain <mcgrof@kernel.org>
Link: https://lore.kernel.org/r/20210927220302.1073499-4-mcgrof@kernel.org
Signed-off-by: Jens Axboe <axboe@kernel.dk>

floppy: fix add_disk() assumption on exit due to new developments

After the patch titled "floppy: use blk_mq_alloc_disk and
blk_cleanup_disk" the floppy driver was modified to allocate
the blk_mq_alloc_disk() which allocates the disk with the
queue. This is further clarified later with the patch titled
"block: remove alloc_disk and alloc_disk_node". This clarifies
that:

   Most drivers should use and have been converted to use
   blk_alloc_disk and blk_mq_alloc_disk.  Only the scsi
   ULPs and dasd still allocate a disk separately from the
   request_queue so don't bother with convenience macros for
   something that should not see significant new users and
   remove these wrappers.

And then we have the patch titled, "block: hold a request_queue
reference for the lifetime of struct gendisk" which ensures
that a queue is *always* present for sure during the entire
lifetime of a disk.

In the floppy driver's case then the disk always comes with the
queue. So even if even if the queue was cleaned up on exit, putting
the disk *is* still required, and likewise, blk_cleanup_queue() on
a null queue should not happen now as disk->queue is valid from
disk allocation time on.

Automatic backport code scrapers should hopefully not cherry pick
this patch as a stable fix candidate without full due dilligence to
ensure all the work done on the block layer to make this happen is
merged first.

Signed-off-by: Luis Chamberlain <mcgrof@kernel.org>
Link: https://lore.kernel.org/r/20210927220302.1073499-3-mcgrof@kernel.org
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block/swim3: add error handling support for add_disk()

We never checked for errors on add_disk() as this function
returned void. Now that this is fixed, use the shiny new
error handling.

Signed-off-by: Luis Chamberlain <mcgrof@kernel.org>
Link: https://lore.kernel.org/r/20210927220302.1073499-2-mcgrof@kernel.org
Signed-off-by: Jens Axboe <axboe@kernel.dk>

rbd: add add_disk() error handling

We never checked for errors on add_disk() as this function
returned void. Now that this is fixed, use the shiny new
error handling.

Signed-off-by: Luis Chamberlain <mcgrof@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

cdrom/gdrom: add error handling support for add_disk()

We never checked for errors on add_disk() as this function
returned void. Now that this is fixed, use the shiny new
error handling.

Signed-off-by: Luis Chamberlain <mcgrof@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

pf: add error handling support for add_disk()

We never checked for errors on add_disk() as this function
returned void. Now that this is fixed, use the shiny new
error handling.

Signed-off-by: Luis Chamberlain <mcgrof@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block/sx8: add error handling support for add_disk()

We never checked for errors on add_disk() as this function
returned void. Now that this is fixed, use the shiny new
error handling.

A completion is used to notify the initial probe what is
happening and so we must defer error handling on completion.
Do this by remembering the error and using the shared cleanup
function.

The tags are shared and so are hanlded later for the
driver already.

Signed-off-by: Luis Chamberlain <mcgrof@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block/rsxx: add error handling support for add_disk()

We never checked for errors on add_disk() as this function
returned void. Now that this is fixed, use the shiny new
error handling.

Signed-off-by: Luis Chamberlain <mcgrof@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

pktcdvd: add error handling support for add_disk()

We never checked for errors on add_disk() as this function
returned void. Now that this is fixed, use the shiny new
error handling.

The out_mem2 error label already does what we need so
re-use that.

Signed-off-by: Luis Chamberlain <mcgrof@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

mtip32xx: add error handling support for add_disk()

We never checked for errors on add_disk() as this function
returned void. Now that this is fixed, use the shiny new
error handling.

The read_capacity_error error label already does what we need,
so just re-use that.

Signed-off-by: Luis Chamberlain <mcgrof@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

pd: add error handling support for add_disk()

We never checked for errors on add_disk() as this function
returned void. Now that this is fixed, use the shiny new
error handling.

Signed-off-by: Luis Chamberlain <mcgrof@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

pcd: capture errors on cdrom_register()

No errors were being captured wehen cdrom_register() fails,
capture the error and return the error.

Signed-off-by: Luis Chamberlain <mcgrof@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

pcd: fix ordering of unregister_cdrom()

We first register cdrom and then we add_disk() and
so we we should likewise unregister the cdrom first and
then del_gendisk().

Signed-off-by: Luis Chamberlain <mcgrof@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

pcd: add error handling support for add_disk()

We never checked for errors on add_disk() as this function
returned void. Now that this is fixed, use the shiny new
error handling.

Signed-off-by: Luis Chamberlain <mcgrof@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

pd: cleanup initialization

Refactor the pf initialization to have a dedicated helper to initialize
a single disk.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

pf: cleanup initialization

Refactor the pf initialization to have a dedicated helper to initialize
a single disk.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

pcd: cleanup initialization

Refactor the pcd initialization to have a dedicated helper to initialize
a single disk.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

pcd: move the identify buffer into pcd_identify

No need to pass it through a bunch of functions.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

n64cart: add error handling support for add_disk()

We never checked for errors on add_disk() as this function
returned void. Now that this is fixed, use the shiny new
error handling.

Signed-off-by: Luis Chamberlain <mcgrof@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

drbd: add error handling support for add_disk()

We never checked for errors on add_disk() as this function
returned void. Now that this is fixed, use the shiny new
error handling.

Signed-off-by: Luis Chamberlain <mcgrof@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

aoe: add error handling support for add_disk()

We never checked for errors on add_disk() as this function
returned void. Now that this is fixed, use the shiny new
error handling.

Signed-off-by: Luis Chamberlain <mcgrof@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

nbd: add error handling support for add_disk()

We never checked for errors on add_disk() as this function
returned void. Now that this is fixed, use the shiny new
error handling.

Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Hannes Reinecke <hare@suse.de>
Signed-off-by: Luis Chamberlain <mcgrof@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

loop: add error handling support for add_disk()

We never checked for errors on add_disk() as this function
returned void. Now that this is fixed, use the shiny new
error handling.

Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Hannes Reinecke <hare@suse.de>
Signed-off-by: Luis Chamberlain <mcgrof@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

null_blk: poll queue support

There's currently no way to experiment with polled IO with null_blk,
which seems like an oversight. This patch adds support for polled IO.
We keep a list of issued IOs on submit, and then process that list
when mq_ops->poll() is invoked.

A new parameter is added, poll_queues. It defaults to 1 like the
submit queues, meaning we'll have 1 poll queue available.

Fixes-by: Bart Van Assche <bvanassche@acm.org>
Fixes-by: Pavel Begunkov <asml.silence@gmail.com>
Reviewed-by: Chaitanya Kulkarni <chaitanya.kulkarni@wdc.com>
Link: https://lore.kernel.org/r/baca710d-0f2a-16e2-60bd-b105b854e0ae@kernel.dk
Signed-off-by: Jens Axboe <axboe@kernel.dk>

nvme: wire up completion batching for the IRQ path

Trivial to do now, just need our own io_comp_batch on the stack and pass
that in to the usual command completion handling.

I pondered making this dependent on how many entries we had to process,
but even for a single entry there's no discernable difference in
performance or latency. Running a sync workload over io_uring:

t/io_uring -b512 -d1 -s1 -c1 -p0 -F1 -B1 -n2 /dev/nvme1n1 /dev/nvme2n1

yields the below performance before the patch:

IOPS=254820, BW=124MiB/s, IOS/call=1/1, inflight=(1 1)
IOPS=251174, BW=122MiB/s, IOS/call=1/1, inflight=(1 1)
IOPS=250806, BW=122MiB/s, IOS/call=1/1, inflight=(1 1)

and the following after:

IOPS=255972, BW=124MiB/s, IOS/call=1/1, inflight=(1 1)
IOPS=251920, BW=123MiB/s, IOS/call=1/1, inflight=(1 1)
IOPS=251794, BW=122MiB/s, IOS/call=1/1, inflight=(1 1)

which definitely isn't slower, about the same if you factor in a bit of
variance. For peak performance workloads, benchmarking shows a 2%
improvement.

Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

io_uring: utilize the io batching infrastructure for more efficient polled IO

Wire up using an io_comp_batch for f_op->iopoll(). If the lower stack
supports it, we can handle high rates of polled IO more efficiently.

This raises the single core efficiency on my system from ~6.1M IOPS to
~6.6M IOPS running a random read workload at depth 128 on two gen2
Optane drives.

Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

nvme: add support for batched completion of polled IO

Take advantage of struct io_comp_batch, if passed in to the nvme poll
handler. If it's set, rather than complete each request individually
inline, store them in the io_comp_batch list. We only do so for requests
that will complete successfully, anything else will be completed inline as
before.

Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: add support for blk_mq_end_request_batch()

Instead of calling blk_mq_end_request() on a single request, add a helper
that takes the new struct io_comp_batch and completes any request stored
in there.

Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

sbitmap: add helper to clear a batch of tags

sbitmap currently only supports clearing tags one-by-one, add a helper
that allows the caller to pass in an array of tags to clear.

Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: add a struct io_comp_batch argument to fops->iopoll()

struct io_comp_batch contains a list head and a completion handler, which
will allow completions to more effciently completed batches of IO.

For now, no functional changes in this patch, we just define the
io_comp_batch structure and add the argument to the file_operations iopoll
handler.

Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: provide helpers for rq_list manipulation

Instead of open-coding the list additions, traversal, and removal,
provide a basic set of helpers.

Suggested-by: Christoph Hellwig <hch@infradead.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: remove some blk_mq_hw_ctx debugfs entries

Just like the blk_mq_ctx counterparts, we've got a bunch of counters
in here that are only for debugfs and are of questionnable value. They
are:

- dispatched, index of how many requests were dispatched in one go

- poll_{considered,invoked,success}, which track poll sucess rates. We're
  confident in the iopoll implementation at this point, don't bother
  tracking these.

As a bonus, this shrinks each hardware queue from 576 bytes to 512 bytes,
dropping a whole cacheline.

Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: remove debugfs blk_mq_ctx dispatched/merged/completed attributes

These were added as part of early days debugging for blk-mq, and they
are not really useful anymore. Rather than spend cycles updating them,
just get rid of them.

As a bonus, this shrinks the per-cpu software queue size from 256b
to 192b. That's a whole cacheline less.

Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: cache rq_flags inside blk_mq_rq_ctx_init()

Add a local variable for rq_flags, it helps to compile out some of
rq_flags reloads.

Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: blk_mq_rq_ctx_init cache ctx/q/hctx

We should have enough of registers in blk_mq_rq_ctx_init(), store them
in local vars, so we don't keep reloading them.

note: keeping q->elevator may look unnecessary, but it's also used
inside inlined blk_mq_tags_from_data().

Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: skip elevator fields init for non-elv queue

Don't init rq->hash and rq->rb_node in blk_mq_rq_ctx_init() if there is
no elevator. Also, move some other initialisers that imply barriers to
the end, so the compiler is free to rearrange and optimise other the
rest of them.

note: fold in a change from Jens leaving queue_list unconditional, as
it might lead to problems otherwise.

Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: store elevator state in request

Add an rq private RQF_ELV flag, which tells the block layer that this
request was initialized on a queue that has an IO scheduler attached.
This allows for faster checking in the fast path, rather than having to
deference rq->q later on.

Elevator switching does full quiesce of the queue before detaching an
IO scheduler, so it's safe to cache this in the request itself.

Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: only mark bio as tracked if it really is tracked

We set BIO_TRACKED unconditionally when rq_qos_throttle() is called, even
though we may not even have an rq_qos handler. Only mark it as TRACKED if
it really is potentially tracked.

This saves considerable time for the case where the bio isn't tracked:

     2.64%     -1.65%  [kernel.vmlinux]  [k] bio_endio

Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: improve layout of struct request

It's been a while since this was analyzed, move some members around to
better flow with the use case. Initial state up top, and queued state
after that. This improves my peak case by about 1.5%, from 7750K to
7900K IOPS.

Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: move update request helpers into blk-mq.c

For some reason we still have them in blk-core, with the rest of the
request completion being in blk-mq. That causes and out-of-line call
for each completion.

Move them into blk-mq.c instead, where they belong.

Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: remove useless caller argument to print_req_error()

We have exactly one caller of this, just get rid of adding the useless
function name to the output.

Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: don't bother iter advancing a fully done bio

If we're completing nbytes and nbytes is the size of the bio, don't bother
with calling into the iterator increment helpers. Just clear the bio
size and we're done.

Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: convert the rest of block to bdev_get_queue

Convert bdev->bd_disk->queue to bdev_get_queue(), it's uses a cached
queue pointer and so is faster.

Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/addf6ea988c04213697ba3684c853e4ed7642a39.1634219547.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: use bdev_get_queue() in blk-core.c

Convert bdev->bd_disk->queue to bdev_get_queue(), it's uses a cached
queue pointer and so is faster.

Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/efc41f880262517c8dc32f932f1b23112f21b255.1634219547.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: use bdev_get_queue() in bio.c

Convert bdev->bd_disk->queue to bdev_get_queue(), it's uses a cached
queue pointer and so is faster.

Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/85c36ea784d285a5075baa10049e6b59e15fb484.1634219547.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: use bdev_get_queue() in bdev.c

Convert bdev->bd_disk->queue to bdev_get_queue(), it's uses a cached
queue pointer and so is faster.

Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/a352936ce5d9ac719645b1e29b173d931ebcdc02.1634219547.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: cache request queue in bdev

There are tons of places where we need to get a request_queue only
having bdev, which turns into bdev->bd_disk->queue. There are probably a
hundred of such places considering inline helpers, and enough of them
are in hot paths.

Cache queue pointer in struct block_device and make use of it in
bdev_get_queue().

Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/a3bfaecdd28956f03629d0ca5c63ebc096e1c809.1634219547.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: handle fast path of bio splitting inline

The fast path is no splitting needed. Separate the handling into a
check part we can inline, and an out-of-line handling path if we do
need to split.

Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: use flags instead of bit fields for blkdev_dio

This generates a lot better code for me, and bumps performance from
7650K IOPS to 7750K IOPS. Looking at profiles for the run and running
perf diff, it confirms that we're now sending a lot less time there:

     6.38%     -2.80%  [kernel.vmlinux]  [k] blkdev_direct_IO

Taking it from the 2nd most cycle consumer to only the 9th most at
3.35% of the CPU time.

Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: cache bdev in struct file for raw bdev IO

bdev = &BDEV_I(file->f_mapping->host)->bdev

Getting struct block_device from a file requires 2 memory dereferences
as illustrated above, that takes a toll on performance, so cache it in
yet unused file->private_data. That gives a noticeable peak performance
improvement.

Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Link: https://lore.kernel.org/r/8415f9fe12e544b9da89593dfbca8de2b52efe03.1634115360.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>

nvme-multipath: enable polled I/O

Set the poll queue flag to enable polling, given that the multipath
node just dispatches the bios to a lower queue.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Tested-by: Mark Wunderlich <mark.wunderlich@intel.com>
Link: https://lore.kernel.org/r/20211012111226.760968-17-hch@lst.de
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: don't allow writing to the poll queue attribute

The poll attribute is a historic artefact from before when we had
explicit poll queues that require driver specific configuration.
Just print a warning when writing to the attribute.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Tested-by: Mark Wunderlich <mark.wunderlich@intel.com>
Link: https://lore.kernel.org/r/20211012111226.760968-16-hch@lst.de
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: switch polling to be bio based

Replace the blk_poll interface that requires the caller to keep a queue
and cookie from the submissions with polling based on the bio.

Polling for the bio itself leads to a few advantages:

 - the cookie construction can made entirely private in blk-mq.c
 - the caller does not need to remember the request_queue and cookie
   separately and thus sidesteps their lifetime issues
 - keeping the device and the cookie inside the bio allows to trivially
   support polling BIOs remapping by stacking drivers
 - a lot of code to propagate the cookie back up the submission path can
   be removed entirely.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Tested-by: Mark Wunderlich <mark.wunderlich@intel.com>
Link: https://lore.kernel.org/r/20211012111226.760968-15-hch@lst.de
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: define 'struct bvec_iter' as packed

'struct bvec_iter' is embedded into 'struct bio', define it as packed
so that we can get one extra 4bytes for other uses without expanding
bio.

'struct bvec_iter' is often allocated on stack, so making it packed
doesn't affect performance. Also I have run io_uring on both
nvme/null_blk, and not observe performance effect in this way.

Suggested-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Ming Lei <ming.lei@redhat.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Hannes Reinecke <hare@suse.de>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Tested-by: Mark Wunderlich <mark.wunderlich@intel.com>
Link: https://lore.kernel.org/r/20211012111226.760968-14-hch@lst.de
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: use SLAB_TYPESAFE_BY_RCU for the bio slab

This flags ensures that the pages will not be reused for non-bio
allocations before the end of an RCU grace period.  With that we can
safely use a RCU lookup for bio polling as long as we are fine with
occasionally polling the wrong device.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Tested-by: Mark Wunderlich <mark.wunderlich@intel.com>
Link: https://lore.kernel.org/r/20211012111226.760968-13-hch@lst.de
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: rename REQ_HIPRI to REQ_POLLED

Unlike the RWF_HIPRI userspace ABI which is intentionally kept vague,
the bio flag is specific to the polling implementation, so rename and
document it properly.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Chaitanya Kulkarni <chaitanya.kulkarni@wdc.com>
Tested-by: Mark Wunderlich <mark.wunderlich@intel.com>
Link: https://lore.kernel.org/r/20211012111226.760968-12-hch@lst.de
Signed-off-by: Jens Axboe <axboe@kernel.dk>

io_uring: don't sleep when polling for I/O

There is no point in sleeping for the expected I/O completion timeout
in the io_uring async polling model as we never poll for a specific
I/O.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Tested-by: Mark Wunderlich <mark.wunderlich@intel.com>
Link: https://lore.kernel.org/r/20211012111226.760968-11-hch@lst.de
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: replace the spin argument to blk_iopoll with a flags argument

Switch the boolean spin argument to blk_poll to passing a set of flags
instead.  This will allow to control polling behavior in a more fine
grained way.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Tested-by: Mark Wunderlich <mark.wunderlich@intel.com>
Link: https://lore.kernel.org/r/20211012111226.760968-10-hch@lst.de
[axboe: adapt to changed io_uring iopoll]
Signed-off-by: Jens Axboe <axboe@kernel.dk>

blk-mq: remove blk_qc_t_valid

Move the trivial check into the only caller.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Chaitanya Kulkarni <chaitanya.kulkarni@wdc.com>
Tested-by: Mark Wunderlich <mark.wunderlich@intel.com>
Link: https://lore.kernel.org/r/20211012111226.760968-9-hch@lst.de
Signed-off-by: Jens Axboe <axboe@kernel.dk>

blk-mq: remove blk_qc_t_to_tag and blk_qc_t_is_internal

Merge both functions into their only caller to keep the blk-mq tag to
blk_qc_t mapping as private as possible in blk-mq.c.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Tested-by: Mark Wunderlich <mark.wunderlich@intel.com>
Link: https://lore.kernel.org/r/20211012111226.760968-8-hch@lst.de
Signed-off-by: Jens Axboe <axboe@kernel.dk>

blk-mq: factor out a "classic" poll helper

Factor the code to do the classic full metal polling out of blk_poll into
a separate blk_mq_poll_classic helper.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Tested-by: Mark Wunderlich <mark.wunderlich@intel.com>
Link: https://lore.kernel.org/r/20211012111226.760968-7-hch@lst.de
Signed-off-by: Jens Axboe <axboe@kernel.dk>

blk-mq: factor out a blk_qc_to_hctx helper

Add a helper to get the hctx from a request_queue and cookie, and fold
the blk_qc_t_to_queue_num helper into it as no other callers are left.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Tested-by: Mark Wunderlich <mark.wunderlich@intel.com>
Link: https://lore.kernel.org/r/20211012111226.760968-6-hch@lst.de
Signed-off-by: Jens Axboe <axboe@kernel.dk>

io_uring: fix a layering violation in io_iopoll_req_issued

syscall-level code can't just poke into the details of the poll cookie,
which is private information of the block layer.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Link: https://lore.kernel.org/r/20211012111226.760968-5-hch@lst.de
Signed-off-by: Jens Axboe <axboe@kernel.dk>

iomap: don't try to poll multi-bio I/Os in __iomap_dio_rw

If an iocb is split into multiple bios we can't poll for both.  So don't
bother to even try to poll in that case.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Tested-by: Mark Wunderlich <mark.wunderlich@intel.com>
Link: https://lore.kernel.org/r/20211012111226.760968-4-hch@lst.de
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: don't try to poll multi-bio I/Os in __blkdev_direct_IO

If an iocb is split into multiple bios we can't poll for both.  So don't
even bother to try to poll in that case.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Link: https://lore.kernel.org/r/20211012111226.760968-3-hch@lst.de
Signed-off-by: Jens Axboe <axboe@kernel.dk>

direct-io: remove blk_poll support

The polling support in the legacy direct-io support is a little crufty.
It already doesn't support the asynchronous polling needed for io_uring
polling, and is hard to adopt to upcoming changes in the polling
interfaces.  Given that all the major file systems already use the iomap
direct I/O code, just drop the polling support.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Tested-by: Mark Wunderlich <mark.wunderlich@intel.com>
Link: https://lore.kernel.org/r/20211012111226.760968-2-hch@lst.de
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: only check previous entry for plug merge attempt

Currently we scan the entire plug list, which is potentially very
expensive. In an IOPS bound workload, we can drive about 5.6M IOPS with
merging enabled, and profiling shows that the plug merge check is the
(by far) most expensive thing we're doing:

  Overhead  Command   Shared Object     Symbol
  +   20.89%  io_uring  [kernel.vmlinux]  [k] blk_attempt_plug_merge
  +    4.98%  io_uring  [kernel.vmlinux]  [k] io_submit_sqes
  +    4.78%  io_uring  [kernel.vmlinux]  [k] blkdev_direct_IO
  +    4.61%  io_uring  [kernel.vmlinux]  [k] blk_mq_submit_bio

Instead of browsing the whole list, just check the previously inserted
entry. That is enough for a naive merge check and will catch most cases,
and for devices that need full merging, the IO scheduler attached to
such devices will do that anyway. The plug merge is meant to be an
inexpensive check to avoid getting a request, but if we repeatedly
scan the list for every single insert, it is very much not a cheap
check.

With this patch, the workload instead runs at ~7.0M IOPS, providing
a 25% improvement. Disabling merging entirely yields another 5%
improvement.

Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: move CONFIG_BLOCK guard to top Makefile

Every object under block/ depends on CONFIG_BLOCK.

Move the guard to the top Makefile since there is no point to
descend into block/ if CONFIG_BLOCK=n.

Signed-off-by: Masahiro Yamada <masahiroy@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Link: https://lore.kernel.org/r/20210927140000.866249-5-masahiroy@kernel.org
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: move menu "Partition type" to block/partitions/Kconfig

Move the menu to the relevant place.

Signed-off-by: Masahiro Yamada <masahiroy@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Link: https://lore.kernel.org/r/20210927140000.866249-4-masahiroy@kernel.org
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: simplify Kconfig files

Everything under block/ depends on BLOCK. BLOCK_HOLDER_DEPRECATED is
selected from drivers/md/Kconfig, which is entirely dependent on BLOCK.

Extend the 'if BLOCK' ... 'endif' so it covers the whole block/Kconfig.

Also, clean up the definition of BLOCK_COMPAT and BLK_MQ_PCI because
COMPAT and PCI are boolean.

Signed-off-by: Masahiro Yamada <masahiroy@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Link: https://lore.kernel.org/r/20210927140000.866249-3-masahiroy@kernel.org
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: remove redundant =y from BLK_CGROUP dependency

CONFIG_BLK_CGROUP is a boolean option, that is, its value is 'y' or 'n'.
The comparison to 'y' is redundant.

Signed-off-by: Masahiro Yamada <masahiroy@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Link: https://lore.kernel.org/r/20210927140000.866249-2-masahiroy@kernel.org
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: improve batched tag allocation

Add a blk_mq_get_tags() helper, which uses the new sbitmap API for
allocating a batch of tags all at once. This both simplifies the block
code for batched allocation, and it is also more efficient than just
doing repeated calls into __sbitmap_queue_get().

This reduces the sbitmap overhead in peak runs from ~3% to ~1% and
yields a performanc increase from 6.6M IOPS to 6.8M IOPS for a single
CPU core.

Signed-off-by: Jens Axboe <axboe@kernel.dk>