Instead of keeping two levels of indirection for requests types, fold it
all into the operations.  The little caveat here is that previously
cmd_type only applied to struct request, while the request and bio op
fields were set to plain REQ_OP_READ/WRITE even for passthrough
operations.

Instead this patch adds new REQ_OP_* for SCSI passthrough and driver
private requests, althought it has to add two for each so that we
can communicate the data in/out nature of the request.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NJens Axboe <axboe@fb.com>

This can be used to check for fs vs non-fs requests and basically
removes all knowledge of BLOCK_PC specific from the block layer,
as well as preparing for removing the cmd_type field in struct request.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NJens Axboe <axboe@fb.com>

And require all drivers that want to support BLOCK_PC to allocate it
as the first thing of their private data.  To support this the legacy
IDE and BSG code is switched to set cmd_size on their queues to let
the block layer allocate the additional space.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NJens Axboe <axboe@fb.com>

This mirrors the blk-mq capabilities to allocate extra drivers-specific
data behind struct request by setting a cmd_size field, as well as having
a constructor / destructor for it.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NMartin K. Petersen <martin.petersen@oracle.com>
Reviewed-by: NHannes Reinecke <hare@suse.com>
Signed-off-by: NJens Axboe <axboe@fb.com>

Return an errno value instead of the passed in queue so that the callers
don't have to keep track of two queues, and move the assignment of the
request_fn and lock to the caller as passing them as argument doesn't
simplify anything.  While we're at it also remove two pointless NULL
assignments, given that the request structure is zeroed on allocation.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NBart Van Assche <bart.vanassche@sandisk.com>
Reviewed-by: NMartin K. Petersen <martin.petersen@oracle.com>
Reviewed-by: NHannes Reinecke <hare@suse.com>
Signed-off-by: NJens Axboe <axboe@fb.com>

If we have both multiple hardware queues and shared tag map between
devices, we need to ensure that we propagate the hardware queue
restart bit higher up. This is because we can get into a situation
where we don't have any IO pending on a hardware queue, yet we fail
getting a tag to start new IO. If that happens, it's not enough to
mark the hardware queue as needing a restart, we need to bubble
that up to the higher level queue as well.
Signed-off-by: NJens Axboe <axboe@fb.com>
Reviewed-by: NOmar Sandoval <osandov@fb.com>
Tested-by: NHannes Reinecke <hare@suse.com>

In preparation for putting blk-mq debugging information in debugfs,
create a directory tree mirroring the one in sysfs:

    # tree -d /sys/kernel/debug/block
    /sys/kernel/debug/block
    |-- nvme0n1
    |   `-- mq
    |       |-- 0
    |       |   `-- cpu0
    |       |-- 1
    |       |   `-- cpu1
    |       |-- 2
    |       |   `-- cpu2
    |       `-- 3
    |           `-- cpu3
    `-- vda
        `-- mq
            `-- 0
                |-- cpu0
                |-- cpu1
                |-- cpu2
                `-- cpu3

Also add the scaffolding for the actual files that will go in here,
either under the hardware queue or software queue directories.
Reviewed-by: NHannes Reinecke <hare@suse.com>
Signed-off-by: NOmar Sandoval <osandov@fb.com>
Signed-off-by: NJens Axboe <axboe@fb.com>

This adds a set of hooks that intercepts the blk-mq path of
allocating/inserting/issuing/completing requests, allowing
us to develop a scheduler within that framework.

We reuse the existing elevator scheduler API on the registration
side, but augment that with the scheduler flagging support for
the blk-mq interfce, and with a separate set of ops hooks for MQ
devices.

We split driver and scheduler tags, so we can run the scheduling
independently of device queue depth.
Signed-off-by: NJens Axboe <axboe@fb.com>
Reviewed-by: NBart Van Assche <bart.vanassche@sandisk.com>
Reviewed-by: NOmar Sandoval <osandov@fb.com>

Add a helper to calculate the actual data transfer size for special
payload requests.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NHannes Reinecke <hare@suse.com>
Signed-off-by: NJens Axboe <axboe@fb.com>

All block device data fields and functions returning a number of 512B
sectors are by convention named xxx_sectors while names in the form
xxx_size are generally used for a number of bytes. The blk_queue_zone_size
and bdev_zone_size functions were not following this convention so rename
them.

No functional change is introduced by this patch.
Signed-off-by: NDamien Le Moal <damien.lemoal@wdc.com>

Collapsed the two patches, they were nonsensically split and broke
bisection.
Signed-off-by: NJens Axboe <axboe@fb.com>

We never change it, make that clear.
Signed-off-by: NJens Axboe <axboe@fb.com>
Reviewed-by: NBart Van Assche <bart.vanassche@sandisk.com>

If the last bvec of the 1st bio and the 1st bvec of the next
bio are physically contigious, and the latter can be merged
to last segment of the 1st bio, we should think they don't
violate sg gap(or virt boundary) limit.

Both Vitaly and Dexuan reported lots of unmergeable small bios
are observed when running mkfs on Hyper-V virtual storage, and
performance becomes quite low. This patch fixes that performance
issue.

The same issue should exist on NVMe, since it sets virt boundary too.
Reported-by: NVitaly Kuznetsov <vkuznets@redhat.com>
Reported-by: NDexuan Cui <decui@microsoft.com>
Tested-by: NDexuan Cui <decui@microsoft.com>
Cc: Keith Busch <keith.busch@intel.com>
Signed-off-by: NMing Lei <ming.lei@canonical.com>
Signed-off-by: NJens Axboe <axboe@fb.com>

Signed-off-by: NRitesh Harjani <riteshh@codeaurora.org>
Reviewed-by: NBart Van Assche <bart.vanassche@sandisk.com>
Signed-off-by: NJens Axboe <axboe@fb.com>

Instead of allocating a single unused biovec for discard requests, send
them down without any payload.  Instead we allow the driver to add a
"special" payload using a biovec embedded into struct request (unioned
over other fields never used while in the driver), and overloading
the number of segments for this case.

This has a couple of advantages:

 - we don't have to allocate the bio_vec
 - the amount of special casing for discard requests in the block
   layer is significantly reduced
 - using this same scheme for other request types is trivial,
   which will be important for implementing the new WRITE_ZEROES
   op on devices where it actually requires a payload (e.g. SCSI)
 - we can get rid of playing games with the request length, as
   we'll never touch it and completions will work just fine
 - it will allow us to support ranged discard operations in the
   future by merging non-contiguous discard bios into a single
   request
 - last but not least it removes a lot of code

This patch is the common base for my WIP series for ranges discards and to
remove discard_zeroes_data in favor of always using REQ_OP_WRITE_ZEROES,
so it would be good to get it in quickly.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NJens Axboe <axboe@fb.com>

This adds a new block layer operation to zero out a range of
LBAs. This allows to implement zeroing for devices that don't use
either discard with a predictable zero pattern or WRITE SAME of zeroes.
The prominent example of that is NVMe with the Write Zeroes command,
but in the future, this should also help with improving the way
zeroing discards work. For this operation, suitable entry is exported in
sysfs which indicate the number of maximum bytes allowed in one
write zeroes operation by the device.
Signed-off-by: NChaitanya Kulkarni <chaitanya.kulkarni@hgst.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NJens Axboe <axboe@fb.com>

Similar to __blkdev_issue_discard this variant allows submitting
the final bio asynchronously and chaining multiple ranges
into a single completion.
Signed-off-by: NChaitanya Kulkarni <chaitanya.kulkarni@hgst.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NJens Axboe <axboe@fb.com>

With compilers which follow the C99 standard (like modern versions of
gcc and clang), "extern inline" does the opposite thing from older
versions of gcc (emits code for an externally linkable version of the
inline function).

"static inline" does the intended behavior in all cases instead.

Description taken from commit 6d91857d ("staging, rtl8192e,
LLVMLinux: Change extern inline to static inline").

This also fixes the following GCC warning when building with CONFIG_PM
disabled:

  ./include/linux/blkdev.h:1143:20: warning: no previous prototype for 'blk_set_runtime_active' [-Wmissing-prototypes]

Fixes: d07ab6d1 ("block: Add blk_set_runtime_active()")
Reviewed-by: NMika Westerberg <mika.westerberg@linux.intel.com>
Signed-off-by: NTobias Klauser <tklauser@distanz.ch>
Signed-off-by: NJens Axboe <axboe@fb.com>

The previous commit introduced the hybrid sleep/poll mode. Take
that one step further, and use the completion latencies to
automatically sleep for half the mean completion time. This is
a good approximation.

This changes the 'io_poll_delay' sysfs file a bit to expose the
various options. Depending on the value, the polling code will
behave differently:

-1	Never enter hybrid sleep mode
 0	Use half of the completion mean for the sleep delay
>0	Use this specific value as the sleep delay
Signed-off-by: NJens Axboe <axboe@fb.com>
Tested-By: NStephen Bates <sbates@raithlin.com>
Reviewed-By: NStephen Bates <sbates@raithlin.com>

This patch enables a hybrid polling mode. Instead of polling after IO
submission, we can induce an artificial delay, and then poll after that.
For example, if the IO is presumed to complete in 8 usecs from now, we
can sleep for 4 usecs, wake up, and then do our polling. This still puts
a sleep/wakeup cycle in the IO path, but instead of the wakeup happening
after the IO has completed, it'll happen before. With this hybrid
scheme, we can achieve big latency reductions while still using the same
(or less) amount of CPU.
Signed-off-by: NJens Axboe <axboe@fb.com>
Tested-By: NStephen Bates <sbates@raithlin.com>
Reviewed-By: NStephen Bates <sbates@raithlin.com>

The poll code is blk-mq specific, let's move it to blk-mq.c. This
is a prep patch for improving the polling code.
Signed-off-by: NJens Axboe <axboe@fb.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>

Enable throttling of buffered writeback to make it a lot
more smooth, and has way less impact on other system activity.
Background writeback should be, by definition, background
activity. The fact that we flush huge bundles of it at the time
means that it potentially has heavy impacts on foreground workloads,
which isn't ideal. We can't easily limit the sizes of writes that
we do, since that would impact file system layout in the presence
of delayed allocation. So just throttle back buffered writeback,
unless someone is waiting for it.

The algorithm for when to throttle takes its inspiration in the
CoDel networking scheduling algorithm. Like CoDel, blk-wb monitors
the minimum latencies of requests over a window of time. In that
window of time, if the minimum latency of any request exceeds a
given target, then a scale count is incremented and the queue depth
is shrunk. The next monitoring window is shrunk accordingly. Unlike
CoDel, if we hit a window that exhibits good behavior, then we
simply increment the scale count and re-calculate the limits for that
scale value. This prevents us from oscillating between a
close-to-ideal value and max all the time, instead remaining in the
windows where we get good behavior.

Unlike CoDel, blk-wb allows the scale count to to negative. This
happens if we primarily have writes going on. Unlike positive
scale counts, this doesn't change the size of the monitoring window.
When the heavy writers finish, blk-bw quickly snaps back to it's
stable state of a zero scale count.

The patch registers a sysfs entry, 'wb_lat_usec'. This sets the latency
target to me met. It defaults to 2 msec for non-rotational storage, and
75 msec for rotational storage. Setting this value to '0' disables
blk-wb. Generally, a user would not have to touch this setting.

We don't enable WBT on devices that are managed with CFQ, and have
a non-root block cgroup attached. If we have a proportional share setup
on this particular disk, then the wbt throttling will interfere with
that. We don't have a strong need for wbt for that case, since we will
rely on CFQ doing that for us.
Signed-off-by: NJens Axboe <axboe@fb.com>

For legacy block, we simply track them in the request queue. For
blk-mq, we track them on a per-sw queue basis, which we can then
sum up through the hardware queues and finally to a per device
state.

The stats are tracked in, roughly, 0.1s interval windows.

Add sysfs files to display the stats.

The feature is off by default, to avoid any extra overhead. In-kernel
users of it can turn it on by setting QUEUE_FLAG_STATS in the queue
flags. We currently don't turn it on if someone just reads any of
the stats files, that is something we could add as well.
Signed-off-by: NJens Axboe <axboe@fb.com>

For blk-mq, ->nr_requests does track queue depth, at least at init
time. But for the older queue paths, it's simply a soft setting.
On top of that, it's generally larger than the hardware setting
on purpose, to allow backup of requests for merging.

Fill a hole in struct request with a 'queue_depth' member, that
drivers can call to more closely inform the block layer of the
real queue depth.
Signed-off-by: NJens Axboe <axboe@fb.com>
Reviewed-by: NJan Kara <jack@suse.cz>

Currently block plug holds up to 16 non-mergeable requests. This makes
sense if the request size is small, eg, reduce lock contention. But if
request size is big enough, we don't need to worry about lock
contention. Holding such request makes no sense and it lows the disk
utilization.

In practice, this improves 10% throughput for my raid5 sequential write
workload.

The size (128k) is arbitrary right now, but it makes sure lock
contention is small. This probably could be more intelligent, eg, check
average request size holded. Since this is mainly for sequential IO,
probably not worthy.

V2: check the last request instead of the first request, so as long as
there is one big size request we flush the plug.
Signed-off-by: NShaohua Li <shli@fb.com>
Signed-off-by: NJens Axboe <axboe@fb.com>

blk_mq_quiesce_queue() waits until ongoing .queue_rq() invocations
have finished. This function does *not* wait until all outstanding
requests have finished (this means invocation of request.end_io()).
The algorithm used by blk_mq_quiesce_queue() is as follows:
* Hold either an RCU read lock or an SRCU read lock around
  .queue_rq() calls. The former is used if .queue_rq() does not
  block and the latter if .queue_rq() may block.
* blk_mq_quiesce_queue() first calls blk_mq_stop_hw_queues()
  followed by synchronize_srcu() or synchronize_rcu(). The latter
  call waits for .queue_rq() invocations that started before
  blk_mq_quiesce_queue() was called.
* The blk_mq_hctx_stopped() calls that control whether or not
  .queue_rq() will be called are called with the (S)RCU read lock
  held. This is necessary to avoid race conditions against
  blk_mq_quiesce_queue().
Signed-off-by: NBart Van Assche <bart.vanassche@sandisk.com>
Cc: Hannes Reinecke <hare@suse.com>
Cc: Johannes Thumshirn <jthumshirn@suse.de>
Reviewed-by: NSagi Grimberg <sagi@grimberg.me>
Reviewed-by: NMing Lei <tom.leiming@gmail.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NJens Axboe <axboe@fb.com>

Now that we don't need the common flags to overflow outside the range
of a 32-bit type we can encode them the same way for both the bio and
request fields.  This in addition allows us to place the operation
first (and make some room for more ops while we're at it) and to
stop having to shift around the operation values.

In addition this allows passing around only one value in the block layer
instead of two (and eventuall also in the file systems, but we can do
that later) and thus clean up a lot of code.

Last but not least this allows decreasing the size of the cmd_flags
field in struct request to 32-bits.  Various functions passing this
value could also be updated, but I'd like to avoid the churn for now.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NJens Axboe <axboe@fb.com>

A lot of the REQ_* flags are only used on struct requests, and only of
use to the block layer and a few drivers that dig into struct request
internals.

This patch adds a new req_flags_t rq_flags field to struct request for
them, and thus dramatically shrinks the number of common requests.  It
also removes the unfortunate situation where we have to fit the fields
from the same enum into 32 bits for struct bio and 64 bits for
struct request.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NShaun Tancheff <shaun.tancheff@seagate.com>
Signed-off-by: NJens Axboe <axboe@fb.com>

Patch adds an association between iocontext ioprio and the ioprio of a
request. This is done to enable request based drivers the ability to
act on priority information stored in the request. An example being
ATA devices that support command priorities. If the ATA driver discovers
that the device supports command priorities and the request has valid
priority information indicating the request is high priority, then a high
priority command can be sent to the device. This should improve tail
latencies for high priority IO on any device that queues requests
internally and can make use of the priority information stored in the
request.

The ioprio of the request is set in blk_rq_set_prio which takes the
request and the ioc as arguments. If the ioc is valid in blk_rq_set_prio
then the iopriority of the request is set as the iopriority of the ioc.
In init_request_from_bio a check is made to see if the ioprio of the bio
is valid and if so then the request prio comes from the bio.
Signed-off-by: NAdam Manzananares <adam.manzanares@wdc.com>
Reviewed-by: NJens Axboe <axboe@kernel.dk>
Signed-off-by: NTejun Heo <tj@kernel.org>

Adds the new BLKREPORTZONE and BLKRESETZONE ioctls for respectively
obtaining the zone configuration of a zoned block device and resetting
the write pointer of sequential zones of a zoned block device.

The BLKREPORTZONE ioctl maps directly to a single call of the function
blkdev_report_zones. The zone information result is passed as an array
of struct blk_zone identical to the structure used internally for
processing the REQ_OP_ZONE_REPORT operation.  The BLKRESETZONE ioctl
maps to a single call of the blkdev_reset_zones function.
Signed-off-by: NShaun Tancheff <shaun.tancheff@seagate.com>
Signed-off-by: NDamien Le Moal <damien.lemoal@hgst.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NMartin K. Petersen <martin.petersen@oracle.com>
Reviewed-by: NHannes Reinecke <hare@suse.com>
Signed-off-by: NJens Axboe <axboe@fb.com>

Implement zoned block device zone information reporting and reset.
Zone information are reported as struct blk_zone. This implementation
does not differentiate between host-aware and host-managed device
models and is valid for both. Two functions are provided:
blkdev_report_zones for discovering the zone configuration of a
zoned block device, and blkdev_reset_zones for resetting the write
pointer of sequential zones. The helper function blk_queue_zone_size
and bdev_zone_size are also provided for, as the name suggest,
obtaining the zone size (in 512B sectors) of the zones of the device.
Signed-off-by: NHannes Reinecke <hare@suse.de>

[Damien: * Removed the zone cache
         * Implement report zones operation based on earlier proposal
           by Shaun Tancheff <shaun.tancheff@seagate.com>]
Signed-off-by: NDamien Le Moal <damien.lemoal@hgst.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NMartin K. Petersen <martin.petersen@oracle.com>
Reviewed-by: NShaun Tancheff <shaun.tancheff@seagate.com>
Tested-by: NShaun Tancheff <shaun.tancheff@seagate.com>
Signed-off-by: NJens Axboe <axboe@fb.com>

Add the zoned queue limit to indicate the zoning model of a block device.
Defined values are 0 (BLK_ZONED_NONE) for regular block devices,
1 (BLK_ZONED_HA) for host-aware zone block devices and 2 (BLK_ZONED_HM)
for host-managed zone block devices. The standards defined drive managed
model is not defined here since these block devices do not provide any
command for accessing zone information. Drive managed model devices will
be reported as BLK_ZONED_NONE.

The helper functions blk_queue_zoned_model and bdev_zoned_model return
the zoned limit and the functions blk_queue_is_zoned and bdev_is_zoned
return a boolean for callers to test if a block device is zoned.

The zoned attribute is also exported as a string to applications via
sysfs. BLK_ZONED_NONE shows as "none", BLK_ZONED_HA as "host-aware" and
BLK_ZONED_HM as "host-managed".
Signed-off-by: NDamien Le Moal <damien.lemoal@hgst.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NMartin K. Petersen <martin.petersen@oracle.com>
Reviewed-by: NShaun Tancheff <shaun.tancheff@seagate.com>
Tested-by: NShaun Tancheff <shaun.tancheff@seagate.com>
Signed-off-by: NJens Axboe <axboe@fb.com>

blk_mq_delay_kick_requeue_list() provides the ability to kick the
q->requeue_list after a specified time.  To do this the request_queue's
'requeue_work' member was changed to a delayed_work.

blk_mq_delay_kick_requeue_list() allows DM to defer processing requeued
requests while it doesn't make sense to immediately requeue them
(e.g. when all paths in a DM multipath have failed).
Signed-off-by: NMike Snitzer <snitzer@redhat.com>
Signed-off-by: NJens Axboe <axboe@fb.com>

Add a helper to schedule a regular struct work on a particular CPU.
Signed-off-by: NJens Axboe <axboe@fb.com>

Commit 288dab8a ("block: add a separate operation type for secure
erase") split REQ_OP_SECURE_ERASE from REQ_OP_DISCARD without considering
all the places REQ_OP_DISCARD was being used to mean either. Fix those.
Signed-off-by: NAdrian Hunter <adrian.hunter@intel.com>
Fixes: 288dab8a ("block: add a separate operation type for secure erase")
Signed-off-by: NJens Axboe <axboe@fb.com>

Commit abf54548 changed it from an 'rw' flags type to the
newer ops based interface, but now we're effectively leaking
some bdev internals to the rest of the kernel. Since we only
care about whether it's a read or a write at that level, just
pass in a bool 'is_write' parameter instead.

Then we can also move op_is_write() and friends back under
CONFIG_BLOCK protection.
Reviewed-by: NMike Christie <mchristi@redhat.com>
Signed-off-by: NJens Axboe <axboe@fb.com>

The rw_page users were not converted to use bio/req ops. As a result
bdev_write_page is not passing down REQ_OP_WRITE and the IOs will
be sent down as reads.
Signed-off-by: NMike Christie <mchristi@redhat.com>
Fixes: 4e1b2d52 ("block, fs, drivers: remove REQ_OP compat defs and related code")

Modified by me to:

1) Drop op_flags passing into ->rw_page(), as we don't use it.
2) Make op_is_write() and friends safe to use for !CONFIG_BLOCK
Signed-off-by: NJens Axboe <axboe@fb.com>

In include/linux/blkdev.h duplicate declarations of the request
struct exist.  Cleaned up by removing the second, unneeded
declaration.
Signed-off-by: NJohn Pittman <jpittman@redhat.com>
Signed-off-by: NJens Axboe <axboe@fb.com>

For a front merge, the maximum number of sectors of the
request must be checked against the front merge BIO sector,
not the current sector of the request.
Signed-off-by: NDamien Le Moal <damien.lemoal@hgst.com>
Reviewed-by: NHannes Reinecke <hare@suse.com>
Signed-off-by: NJens Axboe <axboe@fb.com>

Currently, presence of direct_access() in block_device_operations
indicates support of DAX on its block device.  Because
block_device_operations is instantiated with 'const', this DAX
capablity may not be enabled conditinally.

In preparation for supporting DAX to device-mapper devices, add
QUEUE_FLAG_DAX to request_queue flags to advertise their DAX
support.  This will allow to set the DAX capability based on how
mapped device is composed.
Signed-off-by: NToshi Kani <toshi.kani@hpe.com>
Acked-by: NDan Williams <dan.j.williams@intel.com>
Signed-off-by: NMike Snitzer <snitzer@redhat.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Ross Zwisler <ross.zwisler@linux.intel.com>
Cc: Martin Schwidefsky <schwidefsky@de.ibm.com>
Cc: Heiko Carstens <heiko.carstens@de.ibm.com>
Cc: <linux-s390@vger.kernel.org>
Signed-off-by: NJens Axboe <axboe@fb.com>

I wish the OSD code could simply use blk_rq_map_* helpers like
everyone else, but the complex nature of deciding if we have
DATA IN and/or DATA OUT buffers might make this impossible
(at least for a mere human like me).

But using blk_rq_append_bio at least allows sharing the setup code
between request with or without dat a buffers, and given that this
is the last user of blk_make_request it allows getting rid of that
somewhat awkward interface.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Acked-by: NBoaz Harrosh <ooo@electrozaur.com>
Signed-off-by: NJens Axboe <axboe@fb.com>