nbd: wait uninterruptible for the dead timeout

If we have a pending signal or the user kills their application then
it'll bring down the whole device, which is less than awesome.  Instead
wait uninterruptible for the dead timeout so we're sure we gave it our
best shot.

Fixes: 560bc4b39952 ("nbd: handle dead connections")
Cc: stable@vger.kernel.org
Signed-off-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

blk-mq: don't allocate driver tag upfront for flush rq

The idea behind it is simple:

1) for none scheduler, driver tag has to be borrowed for flush rq,
   otherwise we may run out of tag, and that causes an IO hang. And
   get/put driver tag is actually noop for none, so reordering tags
   isn't necessary at all.

2) for a real I/O scheduler, we need not allocate a driver tag upfront
   for flush rq. It works just fine to follow the same approach as
   normal requests: allocate driver tag for each rq just before calling
   ->queue_rq().

One driver visible change is that the driver tag isn't shared in the
flush request sequence. That won't be a problem, since we always do that
in legacy path.

Then flush rq need not be treated specially wrt. get/put driver tag.
This cleans up the code - for instance, reorder_tags_to_front() can be
removed, and we needn't worry about request ordering in dispatch list
for avoiding I/O deadlock.

Also we have to put the driver tag before requeueing.

Signed-off-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

blk-mq: move blk_mq_put_driver_tag*() into blk-mq.h

We need this helper to put the driver tag for flush rq, since we will
not share tag in the flush request sequence in the following patch
in case that I/O scheduler is applied.

Signed-off-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

blk-mq-sched: decide how to handle flush rq via RQF_FLUSH_SEQ

In case of IO scheduler we always pre-allocate one driver tag before
calling blk_insert_flush(), and flush request will be marked as
RQF_FLUSH_SEQ once it is in flush machinery.

So if RQF_FLUSH_SEQ isn't set, we call blk_insert_flush() to handle
the request, otherwise the flush request is dispatched to ->dispatch
list directly.

This is a preparation patch for not preallocating a driver tag for flush
requests, and for not treating flush requests as a special case. This is
similar to what the legacy path does.

Signed-off-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

blk-flush: use blk_mq_request_bypass_insert()

In the following patch, we will use RQF_FLUSH_SEQ to decide:

1) if the flag isn't set, the flush rq need to be inserted via
blk_insert_flush()

2) otherwise, the flush rq need to be dispatched directly since
it is in flush machinery now.

So we use blk_mq_request_bypass_insert() for requests of bypassing
flush machinery, just like the legacy path did.

Signed-off-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: pass 'run_queue' to blk_mq_request_bypass_insert

Block flush need this function without running the queue, so add a
parameter controlling whether we run it or not.

Signed-off-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

blk-flush: don't run queue for requests bypassing flush

blk_insert_flush() should only insert request since run queue always
follows it.

In case of bypassing flush, we don't need to run queue because every
blk_insert_flush() follows one run queue.

Signed-off-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

blk-mq: put the driver tag of nxt rq before first one is requeued

When freeing the driver tag of the next rq with an I/O scheduler
configured, we get the first entry of the list. However, this can
race with requeue of a request, and we end up getting the wrong request
from the head of the list. Free the driver tag of next rq before the
failed one is requeued in the failure branch of queue_rq callback.

Signed-off-by: Jianchao Wang <jianchao.w.wang@oracle.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

blkcg: add sanity check for blkcg policy operations

blkcg policy should keep cpd/pd's alloc_fn and free_fn in pairs,
otherwise policy would register fail.

Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: weiping zhang <zhangweiping@didichuxing.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

blk-mq: don't handle failure in .get_budget

It is enough to just check if we can get the budget via .get_budget().
And we don't need to deal with device state change in .get_budget().

For SCSI, one issue to be fixed is that we have to call
scsi_mq_uninit_cmd() to free allocated ressources if SCSI device fails
to handle the request. And it isn't enough to simply call
blk_mq_end_request() to do that if this request is marked as
RQF_DONTPREP.

Fixes: 0df21c86bdbf(scsi: implement .get_budget and .put_budget for blk-mq)
Signed-off-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

SCSI: don't get target/host busy_count in scsi_mq_get_budget()

It is very expensive to atomic_inc/atomic_dec the host wide counter of
host->busy_count, and it should have been avoided via blk-mq's mechanism
of getting driver tag, which uses the more efficient way of sbitmap queue.

Also we don't check atomic_read(&sdev->device_busy) in scsi_mq_get_budget()
and don't run queue if the counter becomes zero, so IO hang may be caused
if all requests are completed just before the current SCSI device
is added to shost->starved_list.

Fixes: 0df21c86bdbf(scsi: implement .get_budget and .put_budget for blk-mq)
Reported-by: Bart Van Assche <bart.vanassche@wdc.com>
Signed-off-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: fix peeking requests during PM

We need to look for an active PM request until the next softbarrier
instead of looking for the first non-PM request.  Otherwise any cause
of request reordering might starve the PM request(s).

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

blk-mq: Make blk_mq_get_request() error path less confusing

blk_mq_get_tag() can modify data->ctx. This means that in the
error path of blk_mq_get_request() data->ctx should be passed to
blk_mq_put_ctx() instead of local_ctx. Note: since blk_mq_put_ctx()
ignores its argument, this patch does not change any functionality.

References: commit 1ad43c0078b7 ("blk-mq: don't leak preempt counter/q_usage_counter when allocating rq failed")
Reviewed-by: Ming Lei <ming.lei@redhat.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Bart Van Assche <bart.vanassche@wdc.com>
Cc: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

blk-mq: fix nr_requests wrong value when modify it from sysfs

if blk-mq use "none" io scheduler, nr_request get a wrong value when
input a number > tag_set->queue_depth. blk_mq_tag_update_depth will get
the smaller one min(nr, set->queue_depth), and then q->nr_request get a
wrong value.

Reproduce:

echo none > /sys/block/nvme0n1/queue/scheduler
echo 1000000 > /sys/block/nvme0n1/queue/nr_requests
cat /sys/block/nvme0n1/queue/nr_requests
1000000

Signed-off-by: weiping zhang <zhangweiping@didichuxing.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

cdrom: hide CONFIG_CDROM menu selection

We don't need to expose this. The point is that drivers select
the uniform CDROM layer, if they need it, the user should not
have to make a conscious decision on whether to include this
separately or not.

Fixes: 2a750166a5be ("block: Rework drivers/cdrom/Makefile")
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: add a poll_fn callback to struct request_queue

That we we can also poll non blk-mq queues.  Mostly needed for
the NVMe multipath code, but could also be useful elsewhere.

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

block: introduce GENHD_FL_HIDDEN

With this flag a driver can create a gendisk that can be used for I/O
submission inside the kernel, but which is not registered as user
facing block device.  This will be useful for the NVMe multipath
implementation.

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

block: don't look at the struct device dev_t in disk_devt

The hidden gendisks introduced in the next patch need to keep the dev
field in their struct device empty so that udev won't try to create
block device nodes for them.  To support that rewrite disk_devt to
look at the major and first_minor fields in the gendisk itself instead
of looking into the struct device.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: add a blk_steal_bios helper

This helpers allows to bounce steal the uncompleted bios from a request so
that they can be reissued on another path.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: provide a direct_make_request helper

This helper allows reinserting a bio into a new queue without much
overhead, but requires all queue limits to be the same for the upper
and lower queues, and it does not provide any recursion preventions.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Javier González <javier@cnexlabs.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: add REQ_DRV bit

Set aside a bit in the request/bio flags for driver use.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: move REQ_NOWAIT

This flag should be before the operation-specific REQ_NOUNMAP bit.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

Merge branch 'nvme-4.15' of git://git.infradead.org/nvme into for-4.15/block

Pull NVMe changes from Christoph:

"Below are the currently queue nvme updates for Linux 4.15.  There are
a few more things that could make it for this merge window, but I'd
like to get things into linux-next, especially for the unlikely case
that Linus decided to cut -rc8.

Highlights:
 - support for SGLs in the PCIe driver (Chaitanya Kulkarni)
 - disable I/O schedulers for the admin queue (Israel Rukshin)
 - various Fibre Channel fixes and enhancements (James Smart)
 - various refactoring for better code sharing between transports
   (Sagi Grimberg and me)

as well as lots of little bits from various contributors."

nvme: comment typo fixed in clearing AER

a tiny typo fixed in a comment.

Signed-off-by: Minwoo Im <minwoo.im.dev@gmail.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>

skd: use ktime_get_real_seconds()

Like many storage drivers, skd uses an unsigned 32-bit number for
interchanging the current time with the firmware. This will overflow in
y2106 and is otherwise safe.

However, the get_seconds() function is generally considered deprecated
since the behavior is different between 32-bit and 64-bit architectures,
and using it may indicate a bigger problem.

To annotate that we've thought about this, let's add a comment here
and migrate to the ktime_get_real_seconds() function that consistently
returns a 64-bit number.

Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: fix CDROM dependency on BLK_DEV

After the cdrom cleanup, I get randconfig warnings for some configurations:

warning: (BLK_DEV_IDECD && BLK_DEV_SR) selects CDROM which has unmet direct dependencies (BLK_DEV)

This adds an explicit BLK_DEV dependency for both drivers. The other
drivers that select 'CDROM' already have this and don't need a change.

Fixes: 2a750166a5be ("block: Rework drivers/cdrom/Makefile")
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

nvme: Remove unused headers

Signed-off-by: Keith Busch <keith.busch@intel.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Christoph Hellwig <hch@lst.de>

nvmet: fix fatal_err_work deadlock

Below is a stack trace for an issue that was reported.

What's happening is that the nvmet layer had it's controller kato
timeout fire, which causes it to schedule its fatal error handler
via the fatal_err_work element. The error handler is invoked, which
calls the transport delete_ctrl() entry point, and as the transport
tears down the controller, nvmet_sq_destroy ends up doing the final
put on the ctlr causing it to enter its free routine. The ctlr free
routine does a cancel_work_sync() on fatal_err_work element, which
then does a flush_work and wait_for_completion. But, as the wait is
in the context of the work element being flushed, its in a catch-22
and the thread hangs.

[  326.903131] nvmet: ctrl 1 keep-alive timer (15 seconds) expired!
[  326.909832] nvmet: ctrl 1 fatal error occurred!
[  327.643100] lpfc 0000:04:00.0: 0:6313 NVMET Defer ctx release xri
x114 flg x2
[  494.582064] INFO: task kworker/0:2:243 blocked for more than 120
seconds.
[  494.589638]       Not tainted 4.14.0-rc1.James+ #1
[  494.594986] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs"
disables this message.
[  494.603718] kworker/0:2     D    0   243      2 0x80000000
[  494.609839] Workqueue: events nvmet_fatal_error_handler [nvmet]
[  494.616447] Call Trace:
[  494.619177]  __schedule+0x28d/0x890
[  494.623070]  schedule+0x36/0x80
[  494.626571]  schedule_timeout+0x1dd/0x300
[  494.631044]  ? dequeue_task_fair+0x592/0x840
[  494.635810]  ? pick_next_task_fair+0x23b/0x5c0
[  494.640756]  wait_for_completion+0x121/0x180
[  494.645521]  ? wake_up_q+0x80/0x80
[  494.649315]  flush_work+0x11d/0x1a0
[  494.653206]  ? wake_up_worker+0x30/0x30
[  494.657484]  __cancel_work_timer+0x10b/0x190
[  494.662249]  cancel_work_sync+0x10/0x20
[  494.666525]  nvmet_ctrl_put+0xa3/0x100 [nvmet]
[  494.671482]  nvmet_sq_:q+0x64/0xd0 [nvmet]
[  494.676540]  nvmet_fc_delete_target_queue+0x202/0x220 [nvmet_fc]
[  494.683245]  nvmet_fc_delete_target_assoc+0x6d/0xc0 [nvmet_fc]
[  494.689743]  nvmet_fc_delete_ctrl+0x137/0x1a0 [nvmet_fc]
[  494.695673]  nvmet_fatal_error_handler+0x30/0x40 [nvmet]
[  494.701589]  process_one_work+0x149/0x360
[  494.706064]  worker_thread+0x4d/0x3c0
[  494.710148]  kthread+0x109/0x140
[  494.713751]  ? rescuer_thread+0x380/0x380
[  494.718214]  ? kthread_park+0x60/0x60

Correct by having the fc transport convert to a different workq context
for the actual controller teardown which may call the cancel_work_sync.

Signed-off-by: James Smart <james.smart@broadcom.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Christoph Hellwig <hch@lst.de>

nvme-fc: add dev_loss_tmo timeout and remoteport resume support

When a remoteport is unregistered (connectivity lost), the following
actions are taken:

 - the remoteport is marked DELETED
 - the time when dev_loss_tmo would expire is set in the remoteport
 - all controllers on the remoteport are reset.

After a controller resets, it will stall in a RECONNECTING state waiting
for one of the following:

 - the controller will continue to attempt reconnect per max_retries and
   reconnect_delay.  As no remoteport connectivity, the reconnect attempt
   will immediately fail.  If max reconnects has not been reached, a new
   reconnect_delay timer will be schedule.  If the current time plus
   another reconnect_delay exceeds when dev_loss_tmo expires on the remote
   port, then the reconnect_delay will be shortend to schedule no later
   than when dev_loss_tmo expires.  If max reconnect attempts are reached
   (e.g. ctrl_loss_tmo reached) or dev_loss_tmo ix exceeded without
   connectivity, the controller is deleted.
 - the remoteport is re-registered prior to dev_loss_tmo expiring.
   The resume of the remoteport will immediately attempt to reconnect
   each of its suspended controllers.

Signed-off-by: James Smart <james.smart@broadcom.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
[hch: updated to use nvme_delete_ctrl]
Signed-off-by: Christoph Hellwig <hch@lst.de>

nvme: allow controller RESETTING to RECONNECTING transition

Transport will typically transition from LIVE to RESETTING when initially
performing a reset or recovering from an error.  Adding this transition
allows a transport to transition to RECONNECTING when it checks/waits for
connectivity then creates new transport connections and reinits the
controller.

Signed-off-by: James Smart <james.smart@broadcom.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>

nvme-fc: check connectivity before initiating reconnects

Check remoteport connectivity before initiating reconnects

Signed-off-by: James Smart <james.smart@broadcom.com>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>

nvme-fc: add a dev_loss_tmo field to the remoteport

Add a dev_loss_tmo value, paralleling the SCSI FC transport, for device
connectivity loss.

The transport initializes the value in the nvme_fc_register_remoteport()
call. If the value is not set, a default of 60s is set.

Add a new routine to the api, nvme_fc_set_remoteport_devloss() routine,
which allows the lldd to dynamically update the value on an existing
remoteport.

Signed-off-by: James Smart <james.smart@broadcom.com>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>

nvme-fc: change ctlr state assignments during reset/reconnect

Clean up some of the controller state checks and add the
RESETTING->RECONNECTING state transition.

Specifically:
- the movement of the RESETTING state change and schedule of reset_work
  to core doesn't work wiht nvme_fc_error_recovery setting state to
  RECONNECTING before attempting to reset.  Remove the state change as
  the reset request does it.
- In the rare cases where an error occurs right as we're transitioning
  to LIVE, defer the controller start actions.
- In error handling on teardown of associations while performing initial
  controller creation - avoid quiesce calls on the admin_q.  They are
  unneeded.
- Add the RESETTING->RECONNECTING transition in the reset handler.

Signed-off-by: James Smart <james.smart@broadcom.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>

nvme: flush reset_work before safely continuing with delete operation

Prevent racing controller reset and delete flows. reset_work must not
ever self-requeue so flushing it suffices.

Reported-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Christoph Hellwig <hch@lst.de>

nvme-rdma: reuse nvme_delete_ctrl when reconnect attempts expire

instead of just queueing delete work

Reported-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Christoph Hellwig <hch@lst.de>

nvme: consolidate common code from ->reset_work

No change in behavior except that the FC code cancels two work items a
little later now.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: James Smart <james.smart@broadcom.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>

nvme-rdma: remove nvme_rdma_remove_ctrl

It is only used in two places, and some of the work done by it will
be taken into common code soon.

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

nvme: move controller deletion to common code

Move the ->delete_work and the associated helpers to common code instead
of duplicating them in every driver.  This also adds the missing reference
get/put for the loop driver.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: James Smart <james.smart@broadcom.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>

nvme-fc: merge __nvme_fc_schedule_delete_work into __nvme_fc_del_ctrl

No need to have two functions doing the same thing.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: James Smart <james.smart@broadcom.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>

nvme-fc: avoid workqueue flush stalls

There's no need to wait for the full nvme_wq, which is now shared,
to flush. flush only the delete_work item.

Signed-off-by: James Smart <james.smart@broadcom.com>
Reviewed-by: Sagi Grimberg <sgi@grimberg.me>
Signed-off-by: Christoph Hellwig <hch@lst.de>

block: Rework drivers/cdrom/Makefile

Instead of referring from inside drivers/cdrom/Makefile to all the
drivers that use this driver, let these drivers select the cdrom
driver. This change makes the cdrom build code follow the approach
that is used for most other drivers, namely refer from the higher
layers to the lower layer instead of from the lower layer to the
higher layers.

Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Bart Van Assche <bart.vanassche@wdc.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

blk-mq: don't restart queue when .get_budget returns BLK_STS_RESOURCE

SCSI restarts its queue in scsi_end_request() automatically, so we don't
need to handle this case in blk-mq.

Especailly any request won't be dequeued in this case, we needn't to
worry about IO hang caused by restart vs. dispatch.

Signed-off-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

blk-mq: don't handle TAG_SHARED in restart

Now restart is used in the following cases, and TAG_SHARED is for
SCSI only.

1) .get_budget() returns BLK_STS_RESOURCE
- if resource in target/host level isn't satisfied, this SCSI device
will be added in shost->starved_list, and the whole queue will be rerun
(via SCSI's built-in RESTART) in scsi_end_request() after any request
initiated from this host/targe is completed. Forget to mention, host level
resource can't be an issue for blk-mq at all.

- the same is true if resource in the queue level isn't satisfied.

- if there isn't outstanding request on this queue, then SCSI's RESTART
can't work(blk-mq's can't work too), and the queue will be run after
SCSI_QUEUE_DELAY, and finally all starved sdevs will be handled by SCSI's
RESTART when this request is finished

2) scsi_dispatch_cmd() returns BLK_STS_RESOURCE
- if there isn't onprogressing request on this queue, the queue
will be run after SCSI_QUEUE_DELAY

- otherwise, SCSI's RESTART covers the rerun.

3) blk_mq_get_driver_tag() failed
- BLK_MQ_S_TAG_WAITING covers the cross-queue RESTART for driver
allocation.

In one word, SCSI's built-in RESTART is enough to cover the queue
rerun, and we don't need to pay special attention to TAG_SHARED wrt. restart.

In my test on scsi_debug(8 luns), this patch improves IOPS by 20% ~ 30% when
running I/O on these 8 luns concurrently.

Aslo Roman Pen reported the current RESTART is very expensive especialy
when there are lots of LUNs attached in one host, such as in his
test, RESTART causes half of IOPS be cut.

Fixes: https://marc.info/?l=linux-kernel&m=150832216727524&w=2
Fixes: 6d8c6c0f97ad ("blk-mq: Restart a single queue if tag sets are shared")
Signed-off-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

scsi: implement .get_budget and .put_budget for blk-mq

We need to tell blk-mq to reserve resources before queuing one request,
so implement these two callbacks. Then blk-mq can avoid to dequeue
request too early, and IO merging can be improved a lot.

Signed-off-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

scsi: allow passing in null rq to scsi_prep_state_check()

In the following patch, we will implement scsi_get_budget()
which need to call scsi_prep_state_check() when rq isn't
dequeued yet.

Signed-off-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

blk-mq-sched: improve dispatching from sw queue

SCSI devices use host-wide tagset, and the shared driver tag space is
often quite big. However, there is also a queue depth for each lun(
.cmd_per_lun), which is often small, for example, on both lpfc and
qla2xxx, .cmd_per_lun is just 3.

So lots of requests may stay in sw queue, and we always flush all
belonging to same hw queue and dispatch them all to driver.
Unfortunately it is easy to cause queue busy because of the small
.cmd_per_lun.  Once these requests are flushed out, they have to stay in
hctx->dispatch, and no bio merge can happen on these requests, and
sequential IO performance is harmed.

This patch introduces blk_mq_dequeue_from_ctx for dequeuing a request
from a sw queue, so that we can dispatch them in scheduler's way. We can
then avoid dequeueing too many requests from sw queue, since we don't
flush ->dispatch completely.

This patch improves dispatching from sw queue by using the .get_budget
and .put_budget callbacks.

Reviewed-by: Omar Sandoval <osandov@fb.com>
Signed-off-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

blk-mq: introduce .get_budget and .put_budget in blk_mq_ops

For SCSI devices, there is often a per-request-queue depth, which needs
to be respected before queuing one request.

Currently blk-mq always dequeues the request first, then calls
.queue_rq() to dispatch the request to lld. One obvious issue with this
approach is that I/O merging may not be successful, because when the
per-request-queue depth can't be respected, .queue_rq() has to return
BLK_STS_RESOURCE, and then this request has to stay in hctx->dispatch
list. This means it never gets a chance to be merged with other IO.

This patch introduces .get_budget and .put_budget callback in blk_mq_ops,
then we can try to get reserved budget first before dequeuing request.
If the budget for queueing I/O can't be satisfied, we don't need to
dequeue request at all. Hence the request can be left in the IO
scheduler queue, for more merging opportunities.

Signed-off-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: kyber: check if there are requests in ctx in kyber_has_work()

There may be request in sw queue, and not fetched to domain queue
yet, so check it in kyber_has_work().

Signed-off-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

sbitmap: introduce __sbitmap_for_each_set()

For blk-mq, we need to be able to iterate software queues starting
from any queue in a round robin fashion, so introduce this helper.

Reviewed-by: Omar Sandoval <osandov@fb.com>
Cc: Omar Sandoval <osandov@fb.com>
Signed-off-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

blk-mq-sched: move actual dispatching into one helper

So that it becomes easy to support to dispatch from sw queue in the
following patch.

No functional change.

Reviewed-by: Bart Van Assche <bart.vanassche@wdc.com>
Reviewed-by: Omar Sandoval <osandov@fb.com>
Suggested-by: Christoph Hellwig <hch@lst.de> # for simplifying dispatch logic
Signed-off-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

blk-mq-sched: dispatch from scheduler IFF progress is made in ->dispatch

When the hw queue is busy, we shouldn't take requests from the scheduler
queue any more, otherwise it is difficult to do IO merge.

This patch fixes the awful IO performance on some SCSI devices(lpfc,
qla2xxx, ...) when mq-deadline/kyber is used by not taking requests if
hw queue is busy.

Reviewed-by: Omar Sandoval <osandov@fb.com>
Reviewed-by: Bart Van Assche <bart.vanassche@wdc.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

MAINTAINERS: Remove Rafael from Opal maintainers.

He is no longer working on storage.

Signed-off-by: Scott Bauer <scott.bauer@intel.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

bcache: explicitly destroy mutex while exiting

mutex_destroy does nothing most of time, but it's better to call
it to make the code future proof and it also has some meaning
for like mutex debug.

As Coly pointed out in a previous review, bcache_exit() may not be
able to handle all the references properly if userspace registers
cache and backing devices right before bch_debug_init runs and
bch_debug_init failes later. So not exposing userspace interface
until everything is ready to avoid that issue.

Signed-off-by: Liang Chen <liangchen.linux@gmail.com>
Reviewed-by: Michael Lyle <mlyle@lyle.org>
Reviewed-by: Coly Li <colyli@suse.de>
Reviewed-by: Eric Wheeler <bcache@linux.ewheeler.net>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

bcache: fix wrong cache_misses statistics

Currently, Cache missed IOs are identified by s->cache_miss, but actually,
there are many situations that missed IOs are not assigned a value for
s->cache_miss in cached_dev_cache_miss(), for example, a bypassed IO
(s->iop.bypass = 1), or the cache_bio allocate failed. In these situations,
it will go to out_put or out_submit, and s->cache_miss is null, which leads
bch_mark_cache_accounting() to treat this IO as a hit IO.

[ML: applied by 3-way merge]

Signed-off-by: tang.junhui <tang.junhui@zte.com.cn>
Reviewed-by: Michael Lyle <mlyle@lyle.org>
Reviewed-by: Coly Li <colyli@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

bcache: update bucket_in_use in real time

bucket_in_use is updated in gc thread which triggered by invalidating or
writing sectors_to_gc dirty data, It's a long interval. Therefore, when we
use it to compare with the threshold, it is often not timely, which leads
to inaccurate judgment and often results in bucket depletion.

We have send a patch before, by the means of updating bucket_in_use
periodically In gc thread, which Coly thought that would lead high
latency, In this patch, we add avail_nbuckets to record the count of
available buckets, and we calculate bucket_in_use when alloc or free
bucket in real time.

[edited by ML: eliminated some whitespace errors]

Signed-off-by: Tang Junhui <tang.junhui@zte.com.cn>
Signed-off-by: Michael Lyle <mlyle@lyle.org>
Reviewed-by: Michael Lyle <mlyle@lyle.org>
Reviewed-by: Coly Li <colyli@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

bcache: convert cached_dev.count from atomic_t to refcount_t

atomic_t variables are currently used to implement reference
counters with the following properties:
 - counter is initialized to 1 using atomic_set()
 - a resource is freed upon counter reaching zero
 - once counter reaches zero, its further
   increments aren't allowed
 - counter schema uses basic atomic operations
   (set, inc, inc_not_zero, dec_and_test, etc.)

Such atomic variables should be converted to a newly provided
refcount_t type and API that prevents accidental counter overflows
and underflows. This is important since overflows and underflows
can lead to use-after-free situation and be exploitable.

The variable cached_dev.count is used as pure reference counter.
Convert it to refcount_t and fix up the operations.

Suggested-by: Kees Cook <keescook@chromium.org>
Reviewed-by: David Windsor <dwindsor@gmail.com>
Reviewed-by: Hans Liljestrand <ishkamiel@gmail.com>
Reviewed-by: Michael Lyle <mlyle@lyle.org>
Signed-off-by: Elena Reshetova <elena.reshetova@intel.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

bcache: only permit to recovery read error when cache device is clean

When bcache does read I/Os, for example in writeback or writethrough mode,
if a read request on cache device is failed, bcache will try to recovery
the request by reading from cached device. If the data on cached device is
not synced with cache device, then requester will get a stale data.

For critical storage system like database, providing stale data from
recovery may result an application level data corruption, which is
unacceptible.

With this patch, for a failed read request in writeback or writethrough
mode, recovery a recoverable read request only happens when cache device
is clean. That is to say, all data on cached device is up to update.

For other cache modes in bcache, read request will never hit
cached_dev_read_error(), they don't need this patch.

Please note, because cache mode can be switched arbitrarily in run time, a
writethrough mode might be switched from a writeback mode. Therefore
checking dc->has_data in writethrough mode still makes sense.

Changelog:
V4: Fix parens error pointed by Michael Lyle.
v3: By response from Kent Oversteet, he thinks recovering stale data is a
    bug to fix, and option to permit it is unnecessary. So this version
    the sysfs file is removed.
v2: rename sysfs entry from allow_stale_data_on_failure  to
    allow_stale_data_on_failure, and fix the confusing commit log.
v1: initial patch posted.

[small change to patch comment spelling by mlyle]

Signed-off-by: Coly Li <colyli@suse.de>
Signed-off-by: Michael Lyle <mlyle@lyle.org>
Reported-by: Arne Wolf <awolf@lenovo.com>
Reviewed-by: Michael Lyle <mlyle@lyle.org>
Cc: Kent Overstreet <kent.overstreet@gmail.com>
Cc: Nix <nix@esperi.org.uk>
Cc: Kai Krakow <hurikhan77@gmail.com>
Cc: Eric Wheeler <bcache@lists.ewheeler.net>
Cc: Junhui Tang <tang.junhui@zte.com.cn>
Cc: stable@vger.kernel.org
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: Fix a race between blk_cleanup_queue() and timeout handling

Make sure that if the timeout timer fires after a queue has been
marked "dying" that the affected requests are finished.

Reported-by: chenxiang (M) <chenxiang66@hisilicon.com>
Fixes: commit 287922eb0b18 ("block: defer timeouts to a workqueue")
Signed-off-by: Bart Van Assche <bart.vanassche@wdc.com>
Tested-by: chenxiang (M) <chenxiang66@hisilicon.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Keith Busch <keith.busch@intel.com>
Cc: Hannes Reinecke <hare@suse.com>
Cc: Ming Lei <ming.lei@redhat.com>
Cc: Johannes Thumshirn <jthumshirn@suse.de>
Cc: <stable@vger.kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

nvme-fc: remove NVME_FC_MAX_SEGMENTS

The define is an arbitrary limit to the io size on the initiator,
capping the io to 1MB-4KB.

Remove the define from the transport. I/O size will solely be limited
by the LLDD sg limits.

Signed-off-by: James Smart <james.smart@broadcom.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>

nvme-fc: add support for duplicate_connect option

Adds support for the duplicate_connect option. When set to true,
checks whether there's an existing controller via the same host port
and target port for the same host (hostnqn, hostid) to the same
subsystem. Fails the connection request if an existing controller.

Signed-off-by: James Smart <james.smart@broadcom.com>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Christoph Hellwig <hch@lst.de>

nvme-rdma: add support for duplicate_connect option

Adds support for the duplicate_connect option. When set to true,
checks whether there's an existing controller via the same target
address (traddr), target port (trsvcid), and if specified, host
address (host_traddr). Fails the connection request if there is
an existing controller.

Signed-off-by: James Smart <james.smart@broadcom.com>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Christoph Hellwig <hch@lst.de>

nvme: add helper to compare options to controller

Adds a helper function that compares the host and subsytem
specified in a connect options list vs a controller.

Signed-off-by: James Smart <james.smart@broadcom.com>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Christoph Hellwig <hch@lst.de>

nvme: add duplicate_connect option

Add the "duplicate_connect" boolean option (presence means true).
Default is false.

When false, the transport should validate whether a new controller request
is targeted for the same host transport addressing and target transport
addressing as an existing controller. If so, the new controller request
should be rejected.

When true, the callee is explicitly requesting a duplicate controller
connection to be made and the new request should be attempted.

Signed-off-by: James Smart <james.smart@broadcom.com>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Christoph Hellwig <hch@lst.de>

nvme: check for a live controller in nvme_dev_open

This is a much more sensible check than just the admin queue.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Sagi Grimberg <sagi@rimbeg.me>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>

nvme: get rid of nvme_ctrl_list

Use the core chrdev code to set up the link between the character device
and the nvme controller.  This allows us to get rid of the global list
of all controllers, and also ensures that we have both a reference to
the controller and the transport module before the open method of the
character device is called.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Sagi Grimberg <sgi@grimberg.me>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Reviewed-by: Hannes Reinecke <hare@suse.com>

nvme: switch controller refcounting to use struct device

Instead of allocating a separate struct device for the character device
handle embedd it into struct nvme_ctrl and use it for the main controller
refcounting.  This removes double refcounting and gets us an automatic
reference for the character device operations.  We keep ctrl->device as a
pointer for now to avoid chaning printks all over, but in the future we
could look into message printing helpers that take a controller structure
similar to what other subsystems do.

Note the delete_ctrl operation always already has a reference (either
through sysfs due this change, or because every open file on the
/dev/nvme-fabrics node has a refernece) when it is entered now, so we
don't need to do the unless_zero variant there.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Hannes Reinecke <hare@suse.com>

nvme: simplify nvme_open

Now that we are protected against lookup vs free races for the namespace
by using kref_get_unless_zero we don't need the hack of NULLing out the
disk private data during removal.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>

nvme: use kref_get_unless_zero in nvme_find_get_ns

For kref_get_unless_zero to protect against lookup vs free races we need
to use it in all places where we aren't guaranteed to already hold a
reference.  There is no such guarantee in nvme_find_get_ns, so switch to
kref_get_unless_zero in this function.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>

mq-deadline: add 'deadline' as a name alias

The scheduler framework now supports looking up the appropriate
scheduler with the {name,mq} tupple. We can register mq-deadline
with the alias of 'deadline', so that switching to 'deadline'
will do the right thing based on the type of driver attached to
it.

Reviewed-by: Omar Sandoval <osandov@fb.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

elevator: allow name aliases

Since we now lookup elevator types with the appropriate multiqueue
capability, allow schedulers to register with an alias alongside
the real name. This is in preparation for allowing 'mq-deadline'
to register an alias of 'deadline' as well.

Reviewed-by: Omar Sandoval <osandov@fb.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

elevator: lookup mq vs non-mq elevators

If an IO scheduler is selected via elevator= and it doesn't match
the driver in question wrt blk-mq support, then we fail to boot.

The elevator= parameter is deprecated and only supported for
non-mq devices. Augment the elevator lookup API so that we
pass in if we're looking for an mq capable scheduler or not,
so that we only ever return a valid type for the queue in
question.

Fixes: https://bugzilla.kernel.org/show_bug.cgi?id=196695
Reviewed-by: Omar Sandoval <osandov@fb.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: cope with WRITE ZEROES failing in blkdev_issue_zeroout()

sd_config_write_same() ignores ->max_ws_blocks == 0 and resets it to
permit trying WRITE SAME on older SCSI devices, unless ->no_write_same
is set.  Because REQ_OP_WRITE_ZEROES is implemented in terms of WRITE
SAME, blkdev_issue_zeroout() may fail with -EREMOTEIO:

  $ fallocate -zn -l 1k /dev/sdg
  fallocate: fallocate failed: Remote I/O error
  $ fallocate -zn -l 1k /dev/sdg  # OK
  $ fallocate -zn -l 1k /dev/sdg  # OK

The following calls succeed because sd_done() sets ->no_write_same in
response to a sense that would become BLK_STS_TARGET/-EREMOTEIO, causing
__blkdev_issue_zeroout() to fall back to generating ZERO_PAGE bios.

This means blkdev_issue_zeroout() must cope with WRITE ZEROES failing
and fall back to manually zeroing, unless BLKDEV_ZERO_NOFALLBACK is
specified.  For BLKDEV_ZERO_NOFALLBACK case, return -EOPNOTSUPP if
sd_done() has just set ->no_write_same thus indicating lack of offload
support.

Fixes: c20cfc27a473 ("block: stop using blkdev_issue_write_same for zeroing")
Cc: Hannes Reinecke <hare@suse.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Martin K. Petersen <martin.petersen@oracle.com>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: factor out __blkdev_issue_zero_pages()

blkdev_issue_zeroout() will use this in !BLKDEV_ZERO_NOFALLBACK case.

Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Martin K. Petersen <martin.petersen@oracle.com>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: move CAP_SYS_ADMIN check in blkdev_roset()

Check for CAP_SYS_ADMIN before calling into the driver, similar to
blkdev_flushbuf().  This is safer and can spare a check in the driver.

(Currently BLKROSET is overridden by md and rbd, rbd is missing the
check.  md has the check, but it covers a lot more than BLKROSET.)

Acked-by: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: Invalidate cache on discard v2

It is reasonable drop page cache on discard, otherwise that pages may
be written by writeback second later, so thin provision devices will
not be happy. This seems to be a  security leak in case of secure discard case.

Also add check for queue_discard flag on early stage.

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

lightnvm: pblk: remove leftover testing function

A previous patch inadvertently left an unused test function in the
header, kill it.

Fixes: 8bd400204bd5 ("lightnvm: pblk: cleanup unused and static functions")
Signed-off-by: Javier González <javier@cnexlabs.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

nvme-rdma: Add debug message when reaches timeout

Signed-off-by: Nitzan Carmi <nitzanc@mellanox.com>
Reviewed-by: Max Gurtovoy <maxg@mellanox.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>

nvme-rdma: align nvme_rdma_device structure

Signed-off-by: Max Gurtovoy <maxg@mellanox.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>

nvme-fc: correct io timeout behavior

The transport io timeout behavior wasn't quite correct. It ignored
that the io error handler is supposed to be synchronous so it possibly
allowed the blk request to be restarted while the io associated was
still aborting. Timeouts on reserved commands, those used for
association create, were never timing out thus they hung out forever.

To correct:
If an io is times out while a remoteport is not connected, just
restart the io timer. The lack of connectivity will simultaneously
be resetting the controller, so the reset path will abort and terminate
the io.

If an io is times out while it was marked for transport abort, just
reset the io timer. The abort process is underway and will complete
the io.

Otherwise, if an io times out, abort the io. If the abort was
unsuccessful (unlikely) give up and return not handled.

If the abort was successful, as the abort process is underway it will
terminate the io, so rather than synchronously waiting, just restart
the io timer.

Signed-off-by: James Smart <james.smart@broadcom.com>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Christoph Hellwig <hch@lst.de>

nvme-fc: correct io termination handling

The io completion handling for i/o's that are failing due to
to a transport error or association termination had issues, causing
io failures (DNR set so retries didn't kick in) or long stalls.

Change the io completion handler for the following items:

When an io has been completed due to a transport abort (based on an
exchange error) or when marked as aborted as part of an association
termination (FCOP_FLAGS_TERMIO), set the NVME completion status to
NVME_SC_ABORTED. By default, do not set DNR on the status so that a
retry can be attempted after association recreate.

In cases where an io is failed (non-successful nvme status including
aborted), if the controller is being deleted (blk_queue_dying) or
the io was part of the ios used for association creation (ctrl state
is NEW or RECONNECTING), then additionally set the DNR bit so the io
will not be retried. If the failed io was part of association creation,
the failure will tear down the partially completioned association and
typically restart a new reconnect attempt (another create association
later).

Rearranged code flow to remove a largely unneeded local variable.

Signed-off-by: James Smart <james.smart@broadcom.com>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Christoph Hellwig <hch@lst.de>

nvme-pci: add SGL support

This adds SGL support for NVMe PCIe driver, based on an earlier patch
from Rajiv Shanmugam Madeswaran <smrajiv15 at gmail.com>. This patch
refactors the original code and adds new module parameter sgl_threshold
to determine whether to use SGL or PRP for IOs.

The usage of SGLs is controlled by the sgl_threshold module parameter,
which allows to conditionally use SGLs if average request segment
size (avg_seg_size) is greater than sgl_threshold. In the original patch,
the decision of using SGLs was dependent only on the IO size,
with the new approach we consider not only IO size but also the
number of physical segments present in the IO.

We calculate avg_seg_size based on request payload bytes and number
of physical segments present in the request.

For e.g.:-

1. blk_rq_nr_phys_segments = 2 blk_rq_payload_bytes = 8k
avg_seg_size = 4K use sgl if avg_seg_size >= sgl_threshold.

2. blk_rq_nr_phys_segments = 2 blk_rq_payload_bytes = 64k
avg_seg_size = 32K use sgl if avg_seg_size >= sgl_threshold.

3. blk_rq_nr_phys_segments = 16 blk_rq_payload_bytes = 64k
avg_seg_size = 4K use sgl if avg_seg_size >= sgl_threshold.

Signed-off-by: Chaitanya Kulkarni <chaitanya.kulkarni@wdc.com>
Reviewed-by: Keith Busch <keith.busch@intel.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Christoph Hellwig <hch@lst.de>

nvme: use ida_simple_{get,remove} for the controller instance

Switch to the ida_simple_* helpers instead of opencoding them.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Keith Busch <keith.busch@intel.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>

nvmet: Change max_nsid in subsystem due to ns_disable if needed

In case we disable namespaces which has the nsid like
subsystem max_nsid we need to search for the next largest nsid
in this subsystem. If the subsystem don't has more namespaces
we set it to 0, else we take nsid from the last namespace in
namespaces list because the list is sorted while inserting.

Reviewed-by: Max Gurtovoy <maxg@mellanox.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Roy Shterman <roys@lightbitslabs.com>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
[hch: slight refactor]
Signed-off-by: Christoph Hellwig <hch@lst.de>

nvme-loop: Add BLK_MQ_F_NO_SCHED flag to admin tag set

This flag is useful for admin queues that aren't used for normal IO.

Signed-off-by: Israel Rukshin <israelr@mellanox.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Max Gurtovoy <maxg@mellanox.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>

nvme-fc: Add BLK_MQ_F_NO_SCHED flag to admin tag set

Since commit b86dd81
"block: get rid of blk-mq default scheduler choice Kconfig entries",
when setting nr_hw_queues to 1 the admin tag set uses mq-deadline scheduler.
This flag is useful for admin queues that aren't used for normal IO.

Signed-off-by: Israel Rukshin <israelr@mellanox.com>
Reviewed-by: Max Gurtovoy <maxg@mellanox.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: James Smart <james.smart@broadcom.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>

nvme-rdma: Add BLK_MQ_F_NO_SCHED flag to admin tag set

Since commit b86dd81
"block: get rid of blk-mq default scheduler choice Kconfig entries",
when setting nr_hw_queues to 1 the admin tag set uses mq-deadline scheduler.
This flag is useful for admin queues that aren't used for normal IO.

Signed-off-by: Israel Rukshin <israelr@mellanox.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Max Gurtovoy <maxg@mellanox.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>

nvme-pci: fix typos in comments

fixed comment typos in adapter_alloc_cq() and adapter_alloc_sq().
'the the' duplications are replaced with 'that the'.

Signed-off-by: Minwoo Im <dn3108@gmail.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>

nvme-rdma: stop controller reset if the controller is deleting

If the controller is deleting (in case the user decided to delete it), we
have no point to continue reset sequence.

Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Christoph Hellwig <hch@lst.de>

nvme-rdma: change queue flag semantics DELETING -> ALLOCATED

Instead of marking we are deleting, mark we are allocated and check that
instead. This makes the logic symmetrical to connected mark check.

Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Christoph Hellwig <hch@lst.de>

nvme-rdma: Don't local invalidate if the queue is not live

No chance for the local invalidate to succeed if the queue-pair
is in error state. Most likely the target will do a remote
invalidation of our mr so not a big loss on the test_bit.

Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Christoph Hellwig <hch@lst.de>

nvme-rdma: teardown admin/io queues once on error recovery

Relying on the queue state while tearing down on every reconnect
attempt is not a good design. We should do it once in err_work
and simply try to establish the queues for each reconnect attempt.

Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Christoph Hellwig <hch@lst.de>

nvme-rdma: Check that reinit_request got a proper mr

Warn if req->mr is NULL as it should never happen.

Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Christoph Hellwig <hch@lst.de>

nvme-rdma: move assignment to declaration

No need for the extra line for trivial assignments.

Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Max Gurtovoy <maxg@mellanox.com>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Christoph Hellwig <hch@lst.de>

nvme-rdma: fix wrong logging message

Not necessarily address resolution failed.

Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Max Gurtovoy <maxg@mellanox.com>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Christoph Hellwig <hch@lst.de>

nvme-rdma: pass tagset to directly nvme_rdma_free_tagset

Instead of flagging admin/io.

Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Max Gurtovoy <maxg@mellanox.com>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Christoph Hellwig <hch@lst.de>

block: remove blk_mq_reinit_tagset

No callers left.

Reviewed-by: Jens Axboe <axboe@kernel.dk>
Reviewed-by: Bart Van Assche <bart.vanassche@wdc.com>
Reviewed-by: Max Gurtovoy <maxg@mellanox.com>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Christoph Hellwig <hch@lst.de>

nvme: introduce nvme_reinit_tagset

Move blk_mq_reinit_tagset from blk-mq to nvme core
as the only user of it. Current transports that use
it (rdma, fc) simply implement .reinit_request op.

This patch does not change any functionality.

Reviewed-by: Jens Axboe <axboe@kernel.dk>
Reviewed-by: Max Gurtovoy <maxg@mellanox.com>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Christoph Hellwig <hch@lst.de>

block: introduce blk_mq_tagset_iter

Iterator helper to apply a function on all the
tags in a given tagset. export it as it will be used
outside the block layer later on.

Reviewed-by: Bart Van Assche <bart.vanassche@wdc.com>
Reviewed-by: Jens Axboe <axboe@kernel.dk>
Reviewed-by: Max Gurtovoy <maxg@mellanox.com>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Christoph Hellwig <hch@lst.de>

kyber: fix hang on domain token wait queue

When we're getting a domain token, if we fail to get a token on our
first attempt, we put the current hardware queue on a wait queue and
then try again just in case a token was freed after our initial attempt
but before we got on the wait queue. If this second attempt succeeds, we
currently leave the hardware queue on the wait queue. Usually this is
okay; we'll just run the hardware queue one extra time when another
token is freed. However, if the hardware queue doesn't have any other
requests waiting, then when it it gets the extra wakeup, it won't have
anything to free and therefore won't wake up any other hardware queues.
If tokens are limited, then we won't make forward progress and the
device will hang.

Reported-by: Bin Zha <zhabin.zb@alibaba-inc.com>
Signed-off-by: Omar Sandoval <osandov@fb.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

nullb: fix error return code in null_init()

Fix to return error code -ENOMEM from the null_alloc_dev() error
handling case instead of 0, as done elsewhere in this function.

Fixes: 2984c8684f96 ("nullb: factor disk parameters")
Signed-off-by: Wei Yongjun <weiyongjun1@huawei.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>