openwrt/staging/blogic.git
7 years agolightnvm: pblk: cleanup unnecessary code
Javier González [Mon, 26 Jun 2017 09:57:25 +0000 (11:57 +0200)]
lightnvm: pblk: cleanup unnecessary code

Cleanup unnecessary headers and code lines.

Signed-off-by: Javier González <javier@cnexlabs.com>
Signed-off-by: Matias Bjørling <matias@cnexlabs.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
7 years agolightnvm: pblk: set metadata list for all I/Os
Javier González [Mon, 26 Jun 2017 09:57:24 +0000 (11:57 +0200)]
lightnvm: pblk: set metadata list for all I/Os

Set a dma area for all I/Os in order to read/write from/to the metadata
stored on the per-sector out-of-bound area.

Signed-off-by: Javier González <javier@cnexlabs.com>
Signed-off-by: Matias Bjørling <matias@cnexlabs.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
7 years agolightnvm: pblk: choose optimal victim GC line
Javier González [Mon, 26 Jun 2017 09:57:23 +0000 (11:57 +0200)]
lightnvm: pblk: choose optimal victim GC line

At the moment, we separate the closed lines on three different list
based on their number of valid sectors. GC recycles lines from each list
based on capacity. Lines from each list are taken in a FIFO fashion.

Since the number of lines is limited (it corresponds to the number of
blocks in a LUN, which is somewhere between 1000-2000), we can afford
scanning the lists to choose the optimal line to be recycled. This helps
specially in lines with a high number of valid sectors.

If the number of blocks per LUN increases, we will consider a more
efficient policy.

Signed-off-by: Javier González <javier@cnexlabs.com>
Signed-off-by: Matias Bjørling <matias@cnexlabs.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
7 years agolightnvm: pblk: decouple bad block from line alloc
Javier González [Mon, 26 Jun 2017 09:57:22 +0000 (11:57 +0200)]
lightnvm: pblk: decouple bad block from line alloc

Decouple bad block discovery from line allocation logic. This allows to
return meaningful error codes in case of bad block discovery failure.

Signed-off-by: Javier González <javier@cnexlabs.com>
Signed-off-by: Matias Bjørling <matias@cnexlabs.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
7 years agolightnvm: pblk: simplify meta. memory allocation
Javier González [Mon, 26 Jun 2017 09:57:21 +0000 (11:57 +0200)]
lightnvm: pblk: simplify meta. memory allocation

smeta size will always be suitable for a kmalloc allocation. Simplify
the code and leave the vmalloc fallback only for emeta, where the pblk
configuration has an impact.

Signed-off-by: Javier González <javier@cnexlabs.com>
Signed-off-by: Matias Bjørling <matias@cnexlabs.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
7 years agolightnvm: pblk: issue multiplane reads if possible
Javier González [Mon, 26 Jun 2017 09:57:20 +0000 (11:57 +0200)]
lightnvm: pblk: issue multiplane reads if possible

If a read request is sequential and its size aligns with a
multi-plane page size, use the multi-plane hint to process the I/O in
parallel in the controller.

Signed-off-by: Javier González <javier@cnexlabs.com>
Signed-off-by: Matias Bjørling <matias@cnexlabs.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
7 years agolightnvm: pblk: delete redundant buffer pointer
Javier González [Mon, 26 Jun 2017 09:57:19 +0000 (11:57 +0200)]
lightnvm: pblk: delete redundant buffer pointer

After refactoring the metadata path, the backpointer controlling
synced I/Os in a line becomes unnecessary; metadata is scheduled
on the write thread, thus we know when the end of the line is reached
and act on it directly.

Signed-off-by: Javier González <javier@cnexlabs.com>
Signed-off-by: Matias Bjørling <matias@cnexlabs.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
7 years agolightnvm: pblk: delete redundant debug line stat
Javier González [Mon, 26 Jun 2017 09:57:18 +0000 (11:57 +0200)]
lightnvm: pblk: delete redundant debug line stat

Remove a legacy variable that helped verifying the consistency of the
run-time metadata for the free line list. With the new metadata layout,
this check is no longer necessary.

Signed-off-by: Javier González <javier@cnexlabs.com>
Signed-off-by: Matias Bjørling <matias@cnexlabs.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
7 years agolightnvm: pblk: sched. metadata on write thread
Javier González [Mon, 26 Jun 2017 09:57:17 +0000 (11:57 +0200)]
lightnvm: pblk: sched. metadata on write thread

At the moment, line metadata is persisted on a separate work queue, that
is kicked each time that a line is closed. The assumption when designing
this was that freeing the write thread from creating a new write request
was better than the potential impact of writes colliding on the media
(user I/O and metadata I/O). Experimentation has proven that this
assumption is wrong; collision can cause up to 25% of bandwidth and
introduce long tail latencies on the write thread, which potentially
cause user write threads to spend more time spinning to get a free entry
on the write buffer.

This patch moves the metadata logic to the write thread. When a line is
closed, remaining metadata is written in memory and is placed on a
metadata queue. The write thread then takes the metadata corresponding
to the previous line, creates the write request and schedules it to
minimize collisions on the media. Using this approach, we see that we
can saturate the media's bandwidth, which helps reducing both write
latencies and the spinning time for user writer threads.

Signed-off-by: Javier González <javier@cnexlabs.com>
Signed-off-by: Matias Bjørling <matias@cnexlabs.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
7 years agolightnvm: pblk: rename read request pool
Javier González [Mon, 26 Jun 2017 22:27:13 +0000 (16:27 -0600)]
lightnvm: pblk: rename read request pool

Read requests allocate some extra memory to store its per I/O context.
Instead of requiring yet another memory pool for other type of requests,
generalize this context allocation (and change naming accordingly).

Signed-off-by: Javier González <javier@cnexlabs.com>
Signed-off-by: Matias Bjørling <matias@cnexlabs.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
7 years agolightnvm: pblk: generalize erase path
Javier González [Mon, 26 Jun 2017 09:57:15 +0000 (11:57 +0200)]
lightnvm: pblk: generalize erase path

Erase I/Os are scheduled with the following goals in mind: (i) minimize
LUNs collisions with write I/Os, and (ii) even out the price of erasing
on every write, instead of putting all the burden on when garbage
collection runs. This works well on the current design, but is specific
to the default mapping algorithm.

This patch generalizes the erase path so that other mapping algorithms
can select an arbitrary line to be erased instead. It also gets rid of
the erase semaphore since it creates jittering for user writes.

Signed-off-by: Javier González <javier@cnexlabs.com>
Signed-off-by: Matias Bjørling <matias@cnexlabs.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
7 years agolightnvm: pblk: expose max sec per write on sysfs
Javier González [Mon, 26 Jun 2017 09:57:14 +0000 (11:57 +0200)]
lightnvm: pblk: expose max sec per write on sysfs

Allow to configure the number of maximum sectors per write command
through sysfs. This makes it easier to tune write command sizes for
different controller configurations.

Signed-off-by: Javier González <javier@cnexlabs.com>
Signed-off-by: Matias Bjørling <matias@cnexlabs.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
7 years agolightnvm: pblk: add debug stat for read cache hits
Javier González [Mon, 26 Jun 2017 09:57:13 +0000 (11:57 +0200)]
lightnvm: pblk: add debug stat for read cache hits

Add a new debug counter to measure cache hits on the read path

Signed-off-by: Javier González <javier@cnexlabs.com>
Signed-off-by: Matias Bjørling <matias@cnexlabs.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
7 years agolightnvm: pblk: spare double cpu_to_le64 calc.
Javier González [Mon, 26 Jun 2017 09:57:12 +0000 (11:57 +0200)]
lightnvm: pblk: spare double cpu_to_le64 calc.

Spare a double calculation on the fast write path.

Signed-off-by: Javier González <javier@cnexlabs.com>
Signed-off-by: Matias Bjørling <matias@cnexlabs.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
7 years agolightnvm: propagate right error code to target
Javier González [Mon, 26 Jun 2017 09:57:11 +0000 (11:57 +0200)]
lightnvm: propagate right error code to target

If nvme_alloc_request fails, propagate the right error, instead of
assuming ENOMEM.

Signed-off-by: Javier González <javier@cnexlabs.com>
Signed-off-by: Matias Bjørling <matias@cnexlabs.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
7 years agolightnvm: re-convert ppa format on I/O failure
Javier González [Mon, 26 Jun 2017 09:57:10 +0000 (11:57 +0200)]
lightnvm: re-convert ppa format on I/O failure

In case of a failure when submitting a request, convert the ppa_list
addresses to the target format so that it can interpret ppas for
recovery

Signed-off-by: Javier González <javier@cnexlabs.com>
Signed-off-by: Matias Bjørling <matias@cnexlabs.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
7 years agomtip32xx: fix up the checking for internal command failure
Jens Axboe [Fri, 23 Jun 2017 15:18:54 +0000 (09:18 -0600)]
mtip32xx: fix up the checking for internal command failure

This fixes up two commits that have touched this driver. The
command status field is now a blk_status_t, so we can't check
for < 0 and we definitely can't assume it's holding -Exxxx error
values. All we care about here is whether ->status is zero or not.
Check for that, and remove the various attempts at smart error
reporting. Just log to dmesg what command failed, and the
blk_status_t value.

Reported-by: Dan Carpenter <dan.carpenter@oracle.com>
Fixes: 2a842acab109 ("block: introduce new block status code type")
Fixes: 3f5e6a35774c ("mtip32xx: convert internal command issue to block IO path")
Signed-off-by: Jens Axboe <axboe@kernel.dk>
7 years agoMerge commit '8e8320c9315c' into for-4.13/block
Jens Axboe [Fri, 23 Jun 2017 03:55:24 +0000 (21:55 -0600)]
Merge commit '8e8320c9315c' into for-4.13/block

Pull in the fix for shared tags, as it conflicts with the pending
changes in for-4.13/block. We already pulled in v4.12-rc5 to solve
other conflicts or get fixes that went into 4.12, so not a lot
of changes in this merge.

Signed-off-by: Jens Axboe <axboe@kernel.dk>
7 years agoblk-mq: remove double set queue_num
weiping [Thu, 22 Jun 2017 15:06:56 +0000 (23:06 +0800)]
blk-mq: remove double set queue_num

hwctx's queue_num has been set prior call blk_mq_init_hctx, so no need
set it again.

Signed-off-by: weiping <zhangweiping@didichuxing.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
7 years agoblk-mq: Make it safe to quiesce and unquiesce from an interrupt handler
Bart Van Assche [Wed, 21 Jun 2017 17:55:47 +0000 (10:55 -0700)]
blk-mq: Make it safe to quiesce and unquiesce from an interrupt handler

Since blk_mq_quiesce_queue_nowait() can be called from interrupt
context, make this safe. Since this function is not in the hot
path, uninline it.

Fixes: commit f4560ffe8cec ("blk-mq: use QUEUE_FLAG_QUIESCED to quiesce queue")
Signed-off-by: Bart Van Assche <bart.vanassche@wdc.com>
Cc: Ming Lei <ming.lei@redhat.com>
Cc: Hannes Reinecke <hare@suse.com>
Cc: Martin K. Petersen <martin.petersen@oracle.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
7 years agoblock: Fix off-by-one errors in blk_status_to_errno() and print_req_error()
Bart Van Assche [Wed, 21 Jun 2017 17:55:46 +0000 (10:55 -0700)]
block: Fix off-by-one errors in blk_status_to_errno() and print_req_error()

This was detected by the smatch static analyzer.

Fixes: commit 2a842acab109 ("block: introduce new block status code type")
Signed-off-by: Bart Van Assche <bart.vanassche@wdc.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Hannes Reinecke <hare@suse.com>
Cc: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
7 years agoblock: Declare local symbols static
Bart Van Assche [Wed, 21 Jun 2017 17:55:45 +0000 (10:55 -0700)]
block: Declare local symbols static

Avoid that building with W=1 causes the compiler to complain that
a declaration for bounce_bio_set and bounce_bio_split is missing.

References: commit a8821f3f32be ("block: Improvements to bounce-buffer handling")
Signed-off-by: Bart Van Assche <bart.vanassche@wdc.com>
Cc: Neil Brown <neilb@suse.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Hannes Reinecke <hare@suse.com>
Cc: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
7 years agoblock: Add fallthrough markers to switch statements
Bart Van Assche [Wed, 21 Jun 2017 16:40:11 +0000 (09:40 -0700)]
block: Add fallthrough markers to switch statements

This patch suppresses gcc 7 warnings about falling through in switch
statements when building with W=1. From the gcc documentation: The
-Wimplicit-fallthrough=3 warning is enabled by -Wextra. See also
https://gcc.gnu.org/onlinedocs/gcc-7.1.0/gcc/Warning-Options.html.

Signed-off-by: Bart Van Assche <bart.vanassche@wdc.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
7 years agoblk-mq: fix performance regression with shared tags
Jens Axboe [Tue, 20 Jun 2017 23:56:13 +0000 (17:56 -0600)]
blk-mq: fix performance regression with shared tags

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: 6d8c6c0f97ad ("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>
7 years agobtrfs: use new block error code
Dan Carpenter [Mon, 19 Jun 2017 10:55:37 +0000 (13:55 +0300)]
btrfs: use new block error code

This function is supposed to return blk_status_t error codes now but
there was a stray -ENOMEM left behind.

Fixes: 4e4cbee93d56 ("block: switch bios to blk_status_t")
Signed-off-by: Dan Carpenter <dan.carpenter@oracle.com>
Acked-by: Christoph Hellwig <hch@lst.de>
Acked-by: David Sterba <dsterba@suse.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
7 years agoblk-mq: Warn when attempting to run a hardware queue that is not mapped
Bart Van Assche [Tue, 20 Jun 2017 18:15:49 +0000 (11:15 -0700)]
blk-mq: Warn when attempting to run a hardware queue that is not mapped

A queue must be frozen while the mapped state of a hardware queue
is changed. Additionally, any change of the mapped state is
followed by a call to blk_mq_map_swqueue() (see also
blk_mq_init_allocated_queue() and blk_mq_update_nr_hw_queues()).
Since blk_mq_map_swqueue() does not map any unmapped hardware
queue onto any software queue, no attempt will be made to run
an unmapped hardware queue. Hence issue a warning upon attempts
to run an unmapped hardware queue.

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>
7 years agoblock: Constify disk_type
Bart Van Assche [Tue, 20 Jun 2017 18:15:48 +0000 (11:15 -0700)]
block: Constify disk_type

The variable 'disk_type' is never modified so constify it.

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>
7 years agoblk-mq: Document locking assumptions
Bart Van Assche [Tue, 20 Jun 2017 18:15:47 +0000 (11:15 -0700)]
blk-mq: Document locking assumptions

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>
7 years agoblock: Document what queue type each function is intended for
Bart Van Assche [Tue, 20 Jun 2017 18:15:46 +0000 (11:15 -0700)]
block: Document what queue type each function is intended for

Some functions in block/blk-core.c must only be used on blk-sq queues
while others are safe to use against any queue type. Document which
functions are intended for blk-sq queues and issue a warning if the
blk-sq API is misused. This does not only help block driver authors
but will also make it easier to remove the blk-sq code once that code
is declared obsolete.

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>
7 years agoblock: Check locking assumptions at runtime
Bart Van Assche [Tue, 20 Jun 2017 18:15:45 +0000 (11:15 -0700)]
block: Check locking assumptions at runtime

Instead of documenting the locking assumptions of most block layer
functions as a comment, use lockdep_assert_held() to verify locking
assumptions at runtime.

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>
7 years agoblock: Add a comment above queue_lockdep_assert_held()
Bart Van Assche [Tue, 20 Jun 2017 18:15:44 +0000 (11:15 -0700)]
block: Add a comment above queue_lockdep_assert_held()

Add a comment above the queue_lockdep_assert_held() macro that
explains the purpose of the q->queue_lock test.

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>
7 years agoblk-mq: Initialize .rq_flags in blk_mq_rq_ctx_init()
Bart Van Assche [Tue, 20 Jun 2017 18:15:43 +0000 (11:15 -0700)]
blk-mq: Initialize .rq_flags in blk_mq_rq_ctx_init()

Initialization of blk-mq requests is a bit weird: blk_mq_rq_ctx_init()
is called after a value has been assigned to .rq_flags and .rq_flags
is initialized in __blk_mq_finish_request(). Initialize .rq_flags in
blk_mq_rq_ctx_init() instead of relying on __blk_mq_finish_request().
Moving the initialization of .rq_flags is fine because all changes
and tests of .rq_flags occur between blk_get_request() and finishing
a request.

Signed-off-by: Bart Van Assche <bart.vanassche@sandisk.com>
Cc: 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>
7 years agoblock: Change argument type of scsi_req_init()
Bart Van Assche [Tue, 20 Jun 2017 18:15:42 +0000 (11:15 -0700)]
block: Change argument type of scsi_req_init()

Since scsi_req_init() works on a struct scsi_request, change the
argument type into struct scsi_request *.

Signed-off-by: Bart Van Assche <bart.vanassche@sandisk.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Reviewed-by: Martin K. Petersen <martin.petersen@oracle.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
7 years agoblock: Make most scsi_req_init() calls implicit
Bart Van Assche [Tue, 20 Jun 2017 18:15:41 +0000 (11:15 -0700)]
block: Make most scsi_req_init() calls implicit

Instead of explicitly calling scsi_req_init() after blk_get_request(),
call that function from inside blk_get_request(). Add an
.initialize_rq_fn() callback function to the block drivers that need
it. Merge the IDE .init_rq_fn() function into .initialize_rq_fn()
because it is too small to keep it as a separate function. Keep the
scsi_req_init() call in ide_prep_sense() because it follows a
blk_rq_init() call.

References: commit 82ed4db499b8 ("block: split scsi_request out of struct request")
Signed-off-by: Bart Van Assche <bart.vanassche@sandisk.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Hannes Reinecke <hare@suse.com>
Cc: Omar Sandoval <osandov@fb.com>
Cc: Nicholas Bellinger <nab@linux-iscsi.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
7 years agoblock: Introduce request_queue.initialize_rq_fn()
Bart Van Assche [Tue, 20 Jun 2017 18:15:40 +0000 (11:15 -0700)]
block: Introduce request_queue.initialize_rq_fn()

Several block drivers need to initialize the driver-private request
data after having called blk_get_request() and before .prep_rq_fn()
is called, e.g. when submitting a REQ_OP_SCSI_* request. Avoid that
that initialization code has to be repeated after every
blk_get_request() call by adding new callback functions to struct
request_queue and to struct blk_mq_ops.

Signed-off-by: Bart Van Assche <bart.vanassche@sandisk.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Hannes Reinecke <hare@suse.com>
Cc: Omar Sandoval <osandov@fb.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
7 years agoblock: Make request operation type argument declarations consistent
Bart Van Assche [Tue, 20 Jun 2017 18:15:39 +0000 (11:15 -0700)]
block: Make request operation type argument declarations consistent

Instead of declaring the second argument of blk_*_get_request()
as int and passing it to functions that expect an unsigned int,
declare that second argument as unsigned int. Also because of
consistency, rename that second argument from 'rw' into 'op'.
This patch does not change any functionality.

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>
7 years agoblk-mq: Reduce blk_mq_hw_ctx size
Bart Van Assche [Tue, 20 Jun 2017 18:15:38 +0000 (11:15 -0700)]
blk-mq: Reduce blk_mq_hw_ctx size

Since the srcu structure is rather large (184 bytes on an x86-64
system with kernel debugging disabled), only allocate it if needed.

Reported-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Bart Van Assche <bart.vanassche@sandisk.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Ming Lei <ming.lei@redhat.com>
Cc: Hannes Reinecke <hare@suse.com>
Cc: Omar Sandoval <osandov@fb.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
7 years agoblock: stop using bio_data() in blk_write_same_mergeable
Christoph Hellwig [Mon, 19 Jun 2017 07:24:41 +0000 (09:24 +0200)]
block: stop using bio_data() in blk_write_same_mergeable

While the Write Same page currently always is in low-level it is just
as easy and safer to just compare the page and offset directly.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Martin K. Petersen <martin.petersen@oracle.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
7 years agoblock: remove the unused bio_to_phys macro
Christoph Hellwig [Mon, 19 Jun 2017 07:24:40 +0000 (09:24 +0200)]
block: remove the unused bio_to_phys macro

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Martin K. Petersen <martin.petersen@oracle.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
7 years agonull_blk: add support for shared tags
Jens Axboe [Tue, 20 Jun 2017 20:22:01 +0000 (14:22 -0600)]
null_blk: add support for shared tags

Some storage drivers need to share tag sets between devices. It's
useful to be able to model that with null_blk, to find hangs or
performance issues.

Add a 'shared_tags' bool module parameter that. If that is set to
true and nr_devices is bigger than 1, all devices allocated will
share the same tag set.

Signed-off-by: Jens Axboe <axboe@kernel.dk>
7 years agobtrfs: nowait aio support
Goldwyn Rodrigues [Tue, 20 Jun 2017 12:05:49 +0000 (07:05 -0500)]
btrfs: nowait aio support

Return EAGAIN if any of the following checks fail
 + i_rwsem is not lockable
 + NODATACOW or PREALLOC is not set
 + Cannot nocow at the desired location
 + Writing beyond end of file which is not allocated

Acked-by: David Sterba <dsterba@suse.com>
Signed-off-by: Goldwyn Rodrigues <rgoldwyn@suse.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
7 years agoxfs: nowait aio support
Goldwyn Rodrigues [Tue, 20 Jun 2017 12:05:48 +0000 (07:05 -0500)]
xfs: nowait aio support

If IOCB_NOWAIT is set, bail if the i_rwsem is not lockable
immediately.

IF IOMAP_NOWAIT is set, return EAGAIN in xfs_file_iomap_begin
if it needs allocation either due to file extension, writing to a hole,
or COW or waiting for other DIOs to finish.

Return -EAGAIN if we don't have extent list in memory.

Signed-off-by: Goldwyn Rodrigues <rgoldwyn@suse.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
7 years agoext4: nowait aio support
Goldwyn Rodrigues [Tue, 20 Jun 2017 12:05:47 +0000 (07:05 -0500)]
ext4: nowait aio support

Return EAGAIN if any of the following checks fail for direct I/O:
  + i_rwsem is lockable
  + Writing beyond end of file (will trigger allocation)
  + Blocks are not allocated at the write location

Signed-off-by: Goldwyn Rodrigues <rgoldwyn@suse.com>
Reviewed-by: Jan Kara <jack@suse.cz>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
7 years agoblock: return on congested block device
Goldwyn Rodrigues [Tue, 20 Jun 2017 12:05:46 +0000 (07:05 -0500)]
block: return on congested block device

A new bio operation flag REQ_NOWAIT is introduced to identify bio's
orignating from iocb with IOCB_NOWAIT. This flag indicates
to return immediately if a request cannot be made instead
of retrying.

Stacked devices such as md (the ones with make_request_fn hooks)
currently are not supported because it may block for housekeeping.
For example, an md can have a part of the device suspended.
For this reason, only request based devices are supported.
In the future, this feature will be expanded to stacked devices
by teaching them how to handle the REQ_NOWAIT flags.

Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Goldwyn Rodrigues <rgoldwyn@suse.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
7 years agofs: Introduce IOMAP_NOWAIT
Goldwyn Rodrigues [Tue, 20 Jun 2017 12:05:45 +0000 (07:05 -0500)]
fs: Introduce IOMAP_NOWAIT

IOCB_NOWAIT translates to IOMAP_NOWAIT for iomaps.
This is used by XFS in the XFS patch.

Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Jan Kara <jack@suse.cz>
Signed-off-by: Goldwyn Rodrigues <rgoldwyn@suse.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
7 years agofs: return if direct I/O will trigger writeback
Goldwyn Rodrigues [Tue, 20 Jun 2017 12:05:44 +0000 (07:05 -0500)]
fs: return if direct I/O will trigger writeback

Find out if the I/O will trigger a wait due to writeback. If yes,
return -EAGAIN.

Return -EINVAL for buffered AIO: there are multiple causes of
delay such as page locks, dirty throttling logic, page loading
from disk etc. which cannot be taken care of.

Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Jan Kara <jack@suse.cz>
Signed-off-by: Goldwyn Rodrigues <rgoldwyn@suse.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
7 years agofs: Introduce RWF_NOWAIT and FMODE_AIO_NOWAIT
Goldwyn Rodrigues [Tue, 20 Jun 2017 12:05:43 +0000 (07:05 -0500)]
fs: Introduce RWF_NOWAIT and FMODE_AIO_NOWAIT

RWF_NOWAIT informs kernel to bail out if an AIO request will block
for reasons such as file allocations, or a writeback triggered,
or would block while allocating requests while performing
direct I/O.

RWF_NOWAIT is translated to IOCB_NOWAIT for iocb->ki_flags.

FMODE_AIO_NOWAIT is a flag which identifies the file opened is capable
of returning -EAGAIN if the AIO call will block. This must be set by
supporting filesystems in the ->open() call.

Filesystems xfs, btrfs and ext4 would be supported in the following patches.

Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Jan Kara <jack@suse.cz>
Signed-off-by: Goldwyn Rodrigues <rgoldwyn@suse.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
7 years agofs: Use RWF_* flags for AIO operations
Goldwyn Rodrigues [Tue, 20 Jun 2017 12:05:42 +0000 (07:05 -0500)]
fs: Use RWF_* flags for AIO operations

aio_rw_flags is introduced in struct iocb (using aio_reserved1) which will
carry the RWF_* flags. We cannot use aio_flags because they are not
checked for validity which may break existing applications.

Note, the only place RWF_HIPRI comes in effect is dio_await_one().
All the rest of the locations, aio code return -EIOCBQUEUED before the
checks for RWF_HIPRI.

Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Jan Kara <jack@suse.cz>
Signed-off-by: Goldwyn Rodrigues <rgoldwyn@suse.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
7 years agofs: Introduce filemap_range_has_page()
Goldwyn Rodrigues [Tue, 20 Jun 2017 12:05:41 +0000 (07:05 -0500)]
fs: Introduce filemap_range_has_page()

filemap_range_has_page() return true if the file's mapping has
a page within the range mentioned. This function will be used
to check if a write() call will cause a writeback of previous
writes.

Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Jan Kara <jack@suse.cz>
Signed-off-by: Goldwyn Rodrigues <rgoldwyn@suse.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
7 years agofs: Separate out kiocb flags setup based on RWF_* flags
Goldwyn Rodrigues [Tue, 20 Jun 2017 12:05:40 +0000 (07:05 -0500)]
fs: Separate out kiocb flags setup based on RWF_* flags

Also added RWF_SUPPORTED to encompass all flags.

Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Jan Kara <jack@suse.cz>
Signed-off-by: Goldwyn Rodrigues <rgoldwyn@suse.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
7 years agoMerge branch 'stable/for-jens-4.12' of git://git.kernel.org/pub/scm/linux/kernel...
Jens Axboe [Tue, 20 Jun 2017 13:09:27 +0000 (07:09 -0600)]
Merge branch 'stable/for-jens-4.12' of git://git./linux/kernel/git/konrad/xen into for-linus

Pull xen-blkback fixes from Konrad:

"Security and memory leak fixes in xen block driver."

7 years agonvme: host: unquiesce queue in nvme_kill_queues()
Ming Lei [Mon, 19 Jun 2017 02:21:08 +0000 (10:21 +0800)]
nvme: host: unquiesce queue in nvme_kill_queues()

When nvme_kill_queues() is run, queues may be in
quiesced state, so we forcibly unquiesce queues to avoid
blocking dispatch, and I/O hang can be avoided in
remove path.

Peviously we use blk_mq_start_stopped_hw_queues() as
counterpart of blk_mq_quiesce_queue(), now we have
introduced blk_mq_unquiesce_queue(), so use it explicitly.

Cc: linux-nvme@lists.infradead.org
Signed-off-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
7 years agoRevert "blk-mq: don't use sync workqueue flushing from drivers"
Ming Lei [Tue, 6 Jun 2017 15:22:10 +0000 (23:22 +0800)]
Revert "blk-mq: don't use sync workqueue flushing from drivers"

This patch reverts commit 2719aa217e0d02(blk-mq: don't use
sync workqueue flushing from drivers) because only
blk_mq_quiesce_queue() need the sync flush, and now
we don't need to stop queue any more, so revert it.

Also changes to cancel_delayed_work() in blk_mq_stop_hw_queue().

Reviewed-by: Bart Van Assche <Bart.VanAssche@sandisk.com>
Signed-off-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
7 years agoblk-mq: clarify dispatch may not be drained/blocked by stopping queue
Ming Lei [Tue, 6 Jun 2017 15:22:09 +0000 (23:22 +0800)]
blk-mq: clarify dispatch may not be drained/blocked by stopping queue

BLK_MQ_S_STOPPED may not be observed in other concurrent I/O paths,
we can't guarantee that dispatching won't happen after returning
from the APIs of stopping queue.

So clarify the fact and avoid potential misuse.

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>
7 years agoblk-mq: don't stop queue for quiescing
Ming Lei [Tue, 6 Jun 2017 15:22:08 +0000 (23:22 +0800)]
blk-mq: don't stop queue for quiescing

Queue can be started by other blk-mq APIs and can be used in
different cases, this limits uses of blk_mq_quiesce_queue()
if it is based on stopping queue, and make its usage very
difficult, especially users have to use the stop queue APIs
carefully for avoiding to break blk_mq_quiesce_queue().

We have applied the QUIESCED flag for draining and blocking
dispatch, so it isn't necessary to stop queue any more.

After stopping queue is removed, blk_mq_quiesce_queue() can
be used safely and easily, then users won't worry about queue
restarting during quiescing at all.

Reviewed-by: Bart Van Assche <Bart.VanAssche@sandisk.com>
Signed-off-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
7 years agoblk-mq: update comments on blk_mq_quiesce_queue()
Ming Lei [Tue, 6 Jun 2017 15:22:07 +0000 (23:22 +0800)]
blk-mq: update comments on blk_mq_quiesce_queue()

Actually what we want to get from blk_mq_quiesce_queue()
isn't only to wait for completion of all ongoing .queue_rq().

In the typical context of canceling requests, we need to
make sure that the following is done in the dispatch path
before starting to cancel requests:

- failed dispatched request is finished
- busy dispatched request is requeued, and the STARTED
flag is cleared

So update comment to keep code, doc and our expection consistent.

Signed-off-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
7 years agoblk-mq: use QUEUE_FLAG_QUIESCED to quiesce queue
Ming Lei [Sun, 18 Jun 2017 20:24:27 +0000 (14:24 -0600)]
blk-mq: use QUEUE_FLAG_QUIESCED to quiesce queue

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>
7 years agoblk-mq: use the introduced blk_mq_unquiesce_queue()
Ming Lei [Tue, 6 Jun 2017 15:22:04 +0000 (23:22 +0800)]
blk-mq: use the introduced blk_mq_unquiesce_queue()

blk_mq_unquiesce_queue() is used for unquiescing the
queue explicitly, so replace blk_mq_start_stopped_hw_queues()
with it.

For the scsi part, this patch takes Bart's suggestion to
switch to block quiesce/unquiesce API completely.

Cc: linux-nvme@lists.infradead.org
Cc: linux-scsi@vger.kernel.org
Cc: dm-devel@redhat.com
Reviewed-by: Bart Van Assche <Bart.VanAssche@sandisk.com>
Signed-off-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
7 years agoblk-mq: introduce blk_mq_unquiesce_queue
Ming Lei [Tue, 6 Jun 2017 15:22:03 +0000 (23:22 +0800)]
blk-mq: introduce blk_mq_unquiesce_queue

blk_mq_start_stopped_hw_queues() is used implictly
as counterpart of blk_mq_quiesce_queue() for unquiescing queue,
so we introduce blk_mq_unquiesce_queue() and make it
as counterpart of blk_mq_quiesce_queue() explicitly.

This function is for improving the current quiescing mechanism
in the following patches.

Reviewed-by: Bart Van Assche <Bart.VanAssche@sandisk.com>
Signed-off-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
7 years agoblk-mq: introduce blk_mq_quiesce_queue_nowait()
Ming Lei [Tue, 6 Jun 2017 15:22:02 +0000 (23:22 +0800)]
blk-mq: introduce blk_mq_quiesce_queue_nowait()

This patch introduces blk_mq_quiesce_queue_nowait() so
that we can workaround mpt3sas for quiescing its queue.

Once mpt3sas is fixed, we can remove this helper.

Reviewed-by: Bart Van Assche <Bart.VanAssche@sandisk.com>
Signed-off-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
7 years agoblk-mq: move blk_mq_quiesce_queue() into include/linux/blk-mq.h
Ming Lei [Tue, 6 Jun 2017 15:22:01 +0000 (23:22 +0800)]
blk-mq: move blk_mq_quiesce_queue() into include/linux/blk-mq.h

We usually put blk_mq_*() into include/linux/blk-mq.h, so
move this API into there.

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>
7 years agoblock: don't check for BIO_MAX_PAGES in blk_bio_segment_split()
NeilBrown [Sun, 18 Jun 2017 04:38:59 +0000 (14:38 +1000)]
block: don't check for BIO_MAX_PAGES in blk_bio_segment_split()

blk_bio_segment_split() makes sure bios have no more than
BIO_MAX_PAGES entries in the bi_io_vec.
This was done because bio_clone_bioset() (when given a
mempool bioset) could not handle larger io_vecs.

No driver uses bio_clone_bioset() any more, they all
use bio_clone_fast() if anything, and bio_clone_fast()
doesn't clone the bi_io_vec.

The main user of of bio_clone_bioset() at this level
is bounce.c, and bouncing now happens before blk_bio_segment_split(),
so that is not of concern.

So remove the big helpful comment and the code.

Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
7 years agoblock: remove bio_clone() and all references.
NeilBrown [Sun, 18 Jun 2017 04:38:59 +0000 (14:38 +1000)]
block: remove bio_clone() and all references.

bio_clone() is no longer used.
Only bio_clone_bioset() or bio_clone_fast().
This is for the best, as bio_clone() used fs_bio_set,
and filesystems are unlikely to want to use bio_clone().

So remove bio_clone() and all references.
This includes a fix to some incorrect documentation.

Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
7 years agobcache: use kmalloc to allocate bio in bch_data_verify()
NeilBrown [Sun, 18 Jun 2017 04:38:59 +0000 (14:38 +1000)]
bcache: use kmalloc to allocate bio in bch_data_verify()

This function allocates a bio, then a collection
of pages.  It copes with failure.

It currently uses a mempool() to allocate the bio,
but alloc_page() to allocate the pages.  These fail
in different ways, so the usage is inconsistent.

Change the bio_clone() to bio_clone_kmalloc()
so that no pool is used either for the bio or the pages.

Reviewed-by: Christoph Hellwig <hch@lst.de>
Acked-by: Kent Overstreet <kent.overstreet@gmail.com>
Reviewed-by : Ming Lei <ming.lei@redhat.com>
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
7 years agoxen-blkfront: remove bio splitting.
NeilBrown [Sun, 18 Jun 2017 04:38:59 +0000 (14:38 +1000)]
xen-blkfront: remove bio splitting.

bios that are re-submitted will pass through blk_queue_split() when
blk_queue_bio() is called, and this will split the bio if necessary.
There is no longer any need to do this splitting in xen-blkfront.

Acked-by: Roger Pau Monné <roger.pau@citrix.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
7 years agolightnvm/pblk-read: use bio_clone_fast()
NeilBrown [Sun, 18 Jun 2017 04:38:58 +0000 (14:38 +1000)]
lightnvm/pblk-read: use bio_clone_fast()

pblk_submit_read() uses bio_clone_bioset() but doesn't change the
io_vec, so bio_clone_fast() is a better choice.

It also uses fs_bio_set which is intended for filesystems.  Using it
in a device driver can deadlock.
So allocate a new bioset, and and use bio_clone_fast().

Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Javier González <javier@cnexlabs.com>
Tested-by: Javier González <javier@cnexlabs.com>
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
7 years agopktcdvd: use bio_clone_fast() instead of bio_clone()
NeilBrown [Sun, 18 Jun 2017 04:38:58 +0000 (14:38 +1000)]
pktcdvd: use bio_clone_fast() instead of bio_clone()

pktcdvd doesn't change the bi_io_vec of the clone bio,
so it is more efficient to use bio_clone_fast(), and not clone
the bi_io_vec.
This requires providing a bio_set, and it is safest to
provide a dedicated bio_set rather than sharing
fs_bio_set, which filesytems use.
This new bio_set, pkt_bio_set, can also be use for the bio_split()
call as the two allocations (bio_clone_fast, and bio_split) are
independent, neither can block a bio allocated by the other.

Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
7 years agodrbd: use bio_clone_fast() instead of bio_clone()
NeilBrown [Sun, 18 Jun 2017 04:38:58 +0000 (14:38 +1000)]
drbd: use bio_clone_fast() instead of bio_clone()

drbd does not modify the bi_io_vec of the cloned bio,
so there is no need to clone that part.  So bio_clone_fast()
is the better choice.
For bio_clone_fast() we need to specify a bio_set.
We could use fs_bio_set, which bio_clone() uses, or
drbd_md_io_bio_set, which drbd uses for metadata, but it is
generally best to avoid sharing bio_sets unless you can
be certain that there are no interdependencies.

So create a new bio_set, drbd_io_bio_set, and use bio_clone_fast().

Also remove a "XXX cannot fail ???" comment because it definitely
cannot fail - bio_clone_fast() doesn't fail if the GFP flags allow for
sleeping.

Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
7 years agorbd: use bio_clone_fast() instead of bio_clone()
NeilBrown [Sun, 18 Jun 2017 04:38:58 +0000 (14:38 +1000)]
rbd: use bio_clone_fast() instead of bio_clone()

bio_clone() makes a copy of the bi_io_vec, but rbd never changes that,
so there is no need for a copy.
bio_clone_fast() can be used instead, which avoids making the copy.

This requires that we provide a bio_set.  bio_clone() uses fs_bio_set,
but it isn't, in general, safe to use the same bio_set at different
levels of the stack, as that can lead to deadlocks.  As filesystems
use fs_bio_set, block devices shouldn't.

As rbd never stacks, it is safe to have a single global bio_set for
all rbd devices to use.  So allocate that when the module is
initialised, and use it with bio_clone_fast().

Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
7 years agoblock: Improvements to bounce-buffer handling
NeilBrown [Sun, 18 Jun 2017 04:38:58 +0000 (14:38 +1000)]
block: Improvements to bounce-buffer handling

Since commit 23688bf4f830 ("block: ensure to split after potentially
bouncing a bio") blk_queue_bounce() is called *before*
blk_queue_split().
This means that:
 1/ the comments blk_queue_split() about bounce buffers are
    irrelevant, and
 2/ a very large bio (more than BIO_MAX_PAGES) will no longer be
    split before it arrives at blk_queue_bounce(), leading to the
    possibility that bio_clone_bioset() will fail and a NULL
    will be dereferenced.

Separately, blk_queue_bounce() shouldn't use fs_bio_set as the bio
being copied could be from the same set, and this could lead to a
deadlock.

So:
 - allocate 2 private biosets for blk_queue_bounce, one for
   splitting enormous bios and one for cloning bios.
 - add code to split a bio that exceeds BIO_MAX_PAGES.
 - Fix up the comments in blk_queue_split()

Credit-to: Ming Lei <tom.leiming@gmail.com> (suggested using single bio_for_each_segment loop)
Reviewed-by: Ming Lei <ming.lei@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
7 years agoblk: use non-rescuing bioset for q->bio_split.
NeilBrown [Sun, 18 Jun 2017 04:38:57 +0000 (14:38 +1000)]
blk: use non-rescuing bioset for q->bio_split.

A rescuing bioset is only useful if there might be bios from
that same bioset on the bio_list_on_stack queue at a time
when bio_alloc_bioset() is called.  This never applies to
q->bio_split.

Allocations from q->bio_split are only ever made from
blk_queue_split() which is only ever called early in each of
various make_request_fn()s.  The original bio (call this A)
is then passed to generic_make_request() and is placed on
the bio_list_on_stack queue, and the bio that was allocated
from q->bio_split (B) is processed.

The processing of this may cause other bios to be passed to
generic_make_request() or may even cause the bio B itself to
be passed, possible after some prefix has been split off
(using some other bioset).

generic_make_request() now guarantees that all of these bios
(B and dependants) will be fully processed before the tail
of the original bio A gets handled.  None of these early bios
can possible trigger an allocation from the original
q->bio_split as they are either too small to require
splitting or (more likely) are destined for a different queue.

The next time that the original q->bio_split might be used
by this thread is when A is processed again, as it might
still be too big to handle directly.  By this time there
cannot be any other bios allocated from q->bio_split in the
generic_make_request() queue.  So no rescuing will ever be
needed.

Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
7 years agoblk: make the bioset rescue_workqueue optional.
NeilBrown [Sun, 18 Jun 2017 04:38:57 +0000 (14:38 +1000)]
blk: make the bioset rescue_workqueue optional.

This patch converts bioset_create() to not create a workqueue by
default, so alloctions will never trigger punt_bios_to_rescuer().  It
also introduces a new flag BIOSET_NEED_RESCUER which tells
bioset_create() to preserve the old behavior.

All callers of bioset_create() that are inside block device drivers,
are given the BIOSET_NEED_RESCUER flag.

biosets used by filesystems or other top-level users do not
need rescuing as the bio can never be queued behind other
bios.  This includes fs_bio_set, blkdev_dio_pool,
btrfs_bioset, xfs_ioend_bioset, and one allocated by
target_core_iblock.c.

biosets used by md/raid do not need rescuing as
their usage was recently audited and revised to never
risk deadlock.

It is hoped that most, if not all, of the remaining biosets
can end up being the non-rescued version.

Reviewed-by: Christoph Hellwig <hch@lst.de>
Credit-to: Ming Lei <ming.lei@redhat.com> (minor fixes)
Reviewed-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
7 years agoblk: replace bioset_create_nobvec() with a flags arg to bioset_create()
NeilBrown [Sun, 18 Jun 2017 04:38:57 +0000 (14:38 +1000)]
blk: replace bioset_create_nobvec() with a flags arg to bioset_create()

"flags" arguments are often seen as good API design as they allow
easy extensibility.
bioset_create_nobvec() is implemented internally as a variation in
flags passed to __bioset_create().

To support future extension, make the internal structure part of the
API.
i.e. add a 'flags' argument to bioset_create() and discard
bioset_create_nobvec().

Note that the bio_split allocations in drivers/md/raid* do not need
the bvec mempool - they should have used bioset_create_nobvec().

Suggested-by: Christoph Hellwig <hch@infradead.org>
Reviewed-by: Christoph Hellwig <hch@infradead.org>
Reviewed-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
7 years agoblk: remove bio_set arg from blk_queue_split()
NeilBrown [Sun, 18 Jun 2017 04:38:57 +0000 (14:38 +1000)]
blk: remove bio_set arg from blk_queue_split()

blk_queue_split() is always called with the last arg being q->bio_split,
where 'q' is the first arg.

Also blk_queue_split() sometimes uses the passed-in 'bs' and sometimes uses
q->bio_split.

This is inconsistent and unnecessary.  Remove the last arg and always use
q->bio_split inside blk_queue_split()

Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Ming Lei <ming.lei@redhat.com>
Credit-to: Javier González <jg@lightnvm.io> (Noticed that lightnvm was missed)
Reviewed-by: Javier González <javier@cnexlabs.com>
Tested-by: Javier González <javier@cnexlabs.com>
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
7 years agoblk-mq: remove __blk_mq_alloc_request
Christoph Hellwig [Fri, 16 Jun 2017 16:15:27 +0000 (18:15 +0200)]
blk-mq: remove __blk_mq_alloc_request

Move most code into blk_mq_rq_ctx_init, and the rest into
blk_mq_get_request.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
7 years agoblk-mq-sched: unify request prepare methods
Christoph Hellwig [Fri, 16 Jun 2017 16:15:26 +0000 (18:15 +0200)]
blk-mq-sched: unify request prepare methods

This patch makes sure we always allocate requests in the core blk-mq
code and use a common prepare_request method to initialize them for
both mq I/O schedulers.  For Kyber and additional limit_depth method
is added that is called before allocating the request.

Also because none of the intializations can really fail the new method
does not return an error - instead the bfq finish method is hardened
to deal with the no-IOC case.

Last but not least this removes the abuse of RQF_QUEUE by the blk-mq
scheduling code as RQF_ELFPRIV is all that is needed now.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
7 years agoblk-mq: refactor blk_mq_sched_assign_ioc
Christoph Hellwig [Fri, 16 Jun 2017 16:15:25 +0000 (18:15 +0200)]
blk-mq: refactor blk_mq_sched_assign_ioc

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>
7 years agobfq-iosched: fix NULL ioc check in bfq_get_rq_private
Christoph Hellwig [Fri, 16 Jun 2017 16:15:24 +0000 (18:15 +0200)]
bfq-iosched: fix NULL ioc check in bfq_get_rq_private

icq_to_bic is a container_of operation, so we need to check for NULL
before it.  Also move the check outside the spinlock while we're at
it.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
7 years agoblk-mq: streamline blk_mq_get_request
Christoph Hellwig [Fri, 16 Jun 2017 16:15:23 +0000 (18:15 +0200)]
blk-mq: streamline blk_mq_get_request

Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
7 years agoblk-mq: simplify blk_mq_free_request
Christoph Hellwig [Fri, 16 Jun 2017 16:15:22 +0000 (18:15 +0200)]
blk-mq: simplify blk_mq_free_request

Merge three functions only tail-called by blk_mq_free_request into
blk_mq_free_request.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
7 years agoblk-mq-sched: unify request finished methods
Christoph Hellwig [Fri, 16 Jun 2017 16:15:21 +0000 (18:15 +0200)]
blk-mq-sched: unify request finished methods

No need to have two different callouts of bfq vs kyber.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
7 years agoblk-mq: remove blk_mq_sched_{get,put}_rq_priv
Christoph Hellwig [Fri, 16 Jun 2017 16:15:20 +0000 (18:15 +0200)]
blk-mq: remove blk_mq_sched_{get,put}_rq_priv

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>
7 years agoblk-mq: move blk_mq_sched_{get,put}_request to blk-mq.c
Christoph Hellwig [Fri, 16 Jun 2017 16:15:19 +0000 (18:15 +0200)]
blk-mq: move blk_mq_sched_{get,put}_request to blk-mq.c

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>
7 years agoblk-mq: mark blk_mq_rq_ctx_init static
Christoph Hellwig [Fri, 16 Jun 2017 16:15:18 +0000 (18:15 +0200)]
blk-mq: mark blk_mq_rq_ctx_init static

Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
7 years agoloop: Add PF_LESS_THROTTLE to block/loop device thread.
NeilBrown [Fri, 16 Jun 2017 05:02:09 +0000 (15:02 +1000)]
loop: Add PF_LESS_THROTTLE to block/loop device thread.

When a filesystem is mounted from a loop device, writes are
throttled by balance_dirty_pages() twice: once when writing
to the filesystem and once when the loop_handle_cmd() writes
to the backing file.  This double-throttling can trigger
positive feedback loops that create significant delays.  The
throttling at the lower level is seen by the upper level as
a slow device, so it throttles extra hard.

The PF_LESS_THROTTLE flag was created to handle exactly this
circumstance, though with an NFS filesystem mounted from a
local NFS server.  It reduces the throttling on the lower
layer so that it can proceed largely unthrottled.

To demonstrate this, create a filesystem on a loop device
and write (e.g. with dd) several large files which combine
to consume significantly more than the limit set by
/proc/sys/vm/dirty_ratio or dirty_bytes.  Measure the total
time taken.

When I do this directly on a device (no loop device) the
total time for several runs (mkfs, mount, write 200 files,
umount) is fairly stable: 28-35 seconds.
When I do this over a loop device the times are much worse
and less stable.  52-460 seconds.  Half below 100seconds,
half above.
When I apply this patch, the times become stable again,
though not as fast as the no-loop-back case: 53-72 seconds.

There may be room for further improvement as the total overhead still
seems too high, but this is a big improvement.

Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Ming Lei <tom.leiming@gmail.com>
Suggested-by: Michal Hocko <mhocko@suse.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agoMerge branch 'nvme-4.13' of git://git.infradead.org/nvme into for-4.13/block
Jens Axboe [Fri, 16 Jun 2017 16:14:59 +0000 (10:14 -0600)]
Merge branch 'nvme-4.13' of git://git.infradead.org/nvme into for-4.13/block

Pull NVMe changes for 4.13 from Christoph:

Highlights:

 - UUID identifier support from Johannes
 - Lots of cleanups from Sagi
 - Host Memory Buffer support from me

And lots of cleanups and smaller fixes of course.

Note that the UUID identifier changes are based on top of the uuid tree.
I am the maintainer of that tree and will send it to Linus as soon as
4.12 is released as various other trees depend on it as well (and the
diffstat includes those changes unfortunately)

7 years agoblock: swim3: make of_device_ids const.
Arvind Yadav [Fri, 16 Jun 2017 09:54:39 +0000 (15:24 +0530)]
block: swim3: make of_device_ids const.

of_device_ids are not supposed to change at runtime. All functions
working with of_device_ids provided by <linux/of.h> work with const
of_device_ids. So mark the non-const structs as const.

File size before:
   text    data     bss     dec     hex filename
   8908    1096     624   10628    2984 drivers/block/swim3.o

File size after constify swim3_match:
   text    data     bss     dec     hex filename
   9708     296     624   10628    2984 drivers/block/swim3.o

Signed-off-by: Arvind Yadav <arvind.yadav.cs@gmail.com>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agoblock: Dedicated error code fixups
Bart Van Assche [Tue, 13 Jun 2017 15:07:33 +0000 (08:07 -0700)]
block: Dedicated error code fixups

This patch fixes two sparse warnings introduced by the "dedicated
error codes for the block layer V3" patch series. These changes
have not been tested.

Signed-off-by: Bart Van Assche <bart.vanassche@sandisk.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agonvme: implement NS Optimal IO Boundary from 1.3 Spec
Scott Bauer [Thu, 15 Jun 2017 16:44:30 +0000 (10:44 -0600)]
nvme: implement NS Optimal IO Boundary from 1.3 Spec

The NVMe 1.3 spec introduces Namespace Optimal IO Boundaries (NOIOB),
which standardizes the stripe mechanism we currently have quirks for.
This patch implements the necessary logic to handle this new feature.

Signed-off-by: Scott Bauer <scott.bauer@intel.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
7 years agonvme: don't hard code size of struct t10_pi_tuple
Sagi Grimberg [Thu, 15 Jun 2017 13:31:29 +0000 (16:31 +0300)]
nvme: don't hard code size of struct t10_pi_tuple

Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Christoph Hellwig <hch@lst.de>
7 years agonvme: no need to wait for the reset when keepalive fails
Christoph Hellwig [Mon, 12 Jun 2017 16:21:19 +0000 (18:21 +0200)]
nvme: no need to wait for the reset when keepalive fails

We don't need to wait for the reset from the delayed work item that
is kicked off when we don't get a keepalive.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reported-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
7 years agonvme: move reset workqueue handling to common code
Christoph Hellwig [Thu, 15 Jun 2017 13:41:08 +0000 (15:41 +0200)]
nvme: move reset workqueue handling to common code

This moves the nvme_reset function from the PCIe driver to common code,
renaming it to nvme_reset_ctrl in the process.  Additionally a new
helper nvme_reset_ctrl_sync is added for the case where we want to
wait for the reset.  To facilitate that the reset_work work structure is
move to the common nvme_ctrl structure and the ->reset_ctrl method is
removed.  For now the drivers initialize the reset_work with their own
callback, but longer term we should move to callouts for specific
parts of the reset process and move even more code to the core.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
7 years agonvme-pci: merge init_request methods
Christoph Hellwig [Tue, 13 Jun 2017 07:15:18 +0000 (09:15 +0200)]
nvme-pci: merge init_request methods

Now that we get the tagset passed we can have a single implementation for
the I/O and admin queues.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Max Gurtovoy <maxg@mellanox.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Christoph Hellwig <hch@lst.de>
7 years agonvme-loop: merge init_request methods
Christoph Hellwig [Tue, 13 Jun 2017 07:15:21 +0000 (09:15 +0200)]
nvme-loop: merge init_request methods

Now that we get the tagset passed we can have a single implementation for
the I/O and admin queues.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Max Gurtovoy <maxg@mellanox.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Christoph Hellwig <hch@lst.de>
7 years agonvme-fc: merge init_request methods
Christoph Hellwig [Tue, 13 Jun 2017 07:15:20 +0000 (09:15 +0200)]
nvme-fc: merge init_request methods

Now that we get the tagset passed we can have a single implementation for
the I/O and admin queues.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Max Gurtovoy <maxg@mellanox.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Christoph Hellwig <hch@lst.de>
7 years agonvme-rdma: merge init_request and exit_request methods
Christoph Hellwig [Tue, 13 Jun 2017 07:15:19 +0000 (09:15 +0200)]
nvme-rdma: merge init_request and exit_request methods

Now that we get the tagset passed we can have a single implementation for
the I/O and admin queues.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Max Gurtovoy <maxg@mellanox.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Christoph Hellwig <hch@lst.de>
7 years agonvme: move protection information check into nvme_setup_rw
Christoph Hellwig [Mon, 12 Jun 2017 16:36:32 +0000 (18:36 +0200)]
nvme: move protection information check into nvme_setup_rw

It only applies to read/write commands, and this way non-PCIe drivers
get the check as well instead of having to duplicate it when adding
metadata support.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Keith Busch <keith.busch@intel.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Christoph Hellwig <hch@lst.de>
7 years agonvme: mark shutdown_timeout static
Christoph Hellwig [Mon, 12 Jun 2017 16:30:51 +0000 (18:30 +0200)]
nvme: mark shutdown_timeout static

And open code the SHUTDOWN_TIMEOUT macro.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Christoph Hellwig <hch@lst.de>
7 years agonvme-rdma: fix error code in nvme_rdma_create_ctrl()
Dan Carpenter [Wed, 14 Jun 2017 10:46:45 +0000 (13:46 +0300)]
nvme-rdma: fix error code in nvme_rdma_create_ctrl()

We accidentally return ERR_PTR(0) which is NULL.  The caller isn't
explicitly checking for that but I couldn't immediately spot whether
this would lead to a NULL dereference.  Anyway, we can fix add an
error code easily enough.

Signed-off-by: Dan Carpenter <dan.carpenter@oracle.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
7 years agonvmf: keep track of nvmet connect error status
Guan Junxiong [Tue, 13 Jun 2017 02:51:24 +0000 (10:51 +0800)]
nvmf: keep track of nvmet connect error status

To let the host know what happends to the connection establishment,
adjust the behavior of nvmf_log_connect_error to make more connect
specifig error codes human-readble.

Signed-off-by: Guan Junxiong <guanjunxiong@huawei.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Christoph Hellwig <hch@lst.de>