This sets the queue's max segment size to match the device's
capabilities. The default of 128 is usable until a device's transfer
capability exceeds 512k, assuming a device page size of 4k. Many nvme
devices exceed that transfer limit, so this lets the block layer know what
kind of commands it to allow to form rather than unnecessarily split them.
One additional segment is added to account for a transfer that may start
in the middle of a page.
Signed-off-by: Keith Busch <keith.busch@intel.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
Uses div_u64 for u64 division and round_down, a bitwise operation,
instead of rounddown, which uses a modulus.
Signed-off-by: Jon Derrick <jonathan.derrick@intel.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Some controllers have a controller-side memory buffer available for use
for submissions, completions, lists, or data.
If a CMB is available, the entire CMB will be ioremapped and it will
attempt to map the IO SQes onto the CMB. The queues will be shrunk as
needed. The CMB will not be used if the queue depth is shrunk below some
threshold where it may have reduced performance over a larger queue
in system memory.
Signed-off-by: Jon Derrick <jonathan.derrick@intel.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
This patch changes sq_cmd writers to instead create their command on
the stack. __nvme_submit_cmd copies the sq entry to the queue and writes
the doorbell.
Signed-off-by: Jon Derrick <jonathan.derrick@intel.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
Some drivers use it now, others just set the limits field manually.
But in preparation for splitting this into a hard and soft limit,
ensure that they all call the proper function for setting the hw
limit for discards.
Reviewed-by: Jeff Moyer <jmoyer@redhat.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
This patch has the driver automatically reread partitions if a namespace
has a separate metadata format. Previously revalidating a disk was
sufficient to get the correct capacity set on such formatted drives,
but partitions that may exist would not have been surfaced.
Reported-by: Paul Grabinar <paul.grabinar@ranbarg.com>
Signed-off-by: Keith Busch <keith.busch@intel.com>
Cc: Matthew Wilcox <willy@linux.intel.com>
Tested-by: Paul Grabinar <paul.grabinar@ranbarg.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Fixes an issue when queue_reuest_irq fails in nvme_setup_io_queues. This
patch initializes all vectors to -1 and resets the vector to -1 in the
case of a failure in queue_request_irq. This avoids the free_irq in
nvme_suspend_queue if the queue did not get an irq.
Signed-off-by: Jon Derrick <jonathan.derrick@intel.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
gcc-4.4.4 (and possibly other versions) fail the compile when initializers
are used with anonymous unions. Work around this.
drivers/block/nvme-core.c: In function 'nvme_identify_ctrl':
drivers/block/nvme-core.c:1163: error: unknown field 'identify' specified in initializer
drivers/block/nvme-core.c:1163: warning: missing braces around initializer
drivers/block/nvme-core.c:1163: warning: (near initialization for 'c.<anonymous>')
drivers/block/nvme-core.c:1164: error: unknown field 'identify' specified in initializer
drivers/block/nvme-core.c:1164: warning: excess elements in struct initializer
drivers/block/nvme-core.c:1164: warning: (near initialization for 'c')
...
This patch has no effect on text size with gcc-4.8.2.
Fixes: d29ec8241c ("nvme: submit internal commands through the block layer")
Cc: Christoph Hellwig <hch@lst.de>
Cc: Jens Axboe <axboe@fb.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Jens Axboe <axboe@fb.com>
Move gendisk deletion before controller shutdown so filesystem may sync
dirty pages. Before, this would deadlock trying to allocate requests
on frozen queues that are about to be deleted.
Signed-off-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
This fixes an infinite device reset loop that may occur on devices that
fail initialization. If the drive fails to become ready for any reason
that does not involve an admin command timeout, the probe task should
assume the drive is unavailable and remove it from the topology. In
the case an admin command times out during device probing, the driver's
existing reset action will handle removing the drive.
Signed-off-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
This unifies probe and resume so they both may be scheduled in the same
way. This is necessary for error handling that may occur during device
initialization since the task to cleanup the device wouldn't be able to
run if it is blocked on device initialization.
Signed-off-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Synchronized commands do different things for timed out commands
vs. controller returned errors.
Signed-off-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Don't release block queue and tagging resoureces if the driver never
got them in the first place. This can happen if the controller fails to
become ready, if memory wasn't available to allocate a tagset or admin
queue, or if the resources were released as part of error recovery.
Signed-off-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Pull block driver updates from Jens Axboe:
"This contains:
- a few race fixes for null_blk, from Akinobu Mita.
- a series of fixes for mtip32xx, from Asai Thambi and Selvan Mani at
Micron.
- NVMe:
* Fix for missing error return on allocation failure, from Axel
Lin.
* Code consolidation and cleanups from Christoph.
* Memory barrier addition, syncing queue count and queue
pointers. From Jon Derrick.
* Various fixes from Keith, an addition to support user
issue reset from sysfs or ioctl, and automatic namespace
rescan.
* Fix from Matias, avoiding losing some request flags when
marking the request failfast.
- small cleanups and sparse fixups for ps3vram. From Geert
Uytterhoeven and Geoff Lavand.
- s390/dasd dead code removal, from Jarod Wilson.
- a set of fixes and optimizations for loop, from Ming Lei.
- conversion to blkdev_reread_part() of loop, dasd, ndb. From Ming
Lei.
- updates to cciss. From Tomas Henzl"
* 'for-4.2/drivers' of git://git.kernel.dk/linux-block: (44 commits)
mtip32xx: Fix accessing freed memory
block: nvme-scsi: Catch kcalloc failure
NVMe: Fix IO for extended metadata formats
nvme: don't overwrite req->cmd_flags on sync cmd
mtip32xx: increase wait time for hba reset
mtip32xx: fix minor number
mtip32xx: remove unnecessary sleep in mtip_ftl_rebuild_poll()
mtip32xx: fix crash on surprise removal of the drive
mtip32xx: Abort I/O during secure erase operation
mtip32xx: fix incorrectly setting MTIP_DDF_SEC_LOCK_BIT
mtip32xx: remove unused variable 'port->allocated'
mtip32xx: fix rmmod issue
MAINTAINERS: Update ps3vram block driver
block/ps3vram: Remove obsolete reference to MTD
block/ps3vram: Fix sparse warnings
NVMe: Automatic namespace rescan
NVMe: Memory barrier before queue_count is incremented
NVMe: add sysfs and ioctl controller reset
null_blk: restart request processing on completion handler
null_blk: prevent timer handler running on a different CPU where started
...
This fixes io submit ioctl handling when using extended metadata
formats. When these formats are used, the user provides a single virtually
contiguous buffer containing both the block and metadata interleaved,
so the metadata size needs to be added to the total length and not mapped
as a separate transfer.
The command is also driver generated, so this patch does not enforce
blk-integrity extensions provide the metadata buffer.
Reported-by: Marcin Dziegielewski <marcin.dziegielewski@intel.com>
Signed-off-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
In __nvme_submit_sync_cmd, the request direction is overwritten when
the REQ_FAILFAST_DRIVER flag is set.
Signed-off-by: Matias Bjørling <m@bjorling.me>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Fixes: 75619bfa90 ("NVMe: End sync requests immediately on failure")
Signed-off-by: Jens Axboe <axboe@fb.com>
Namespaces may be dynamically allocated and deleted or attached and
detached. This has the driver rescan the device for namespace changes
after each device reset or namespace change asynchronous event.
There could potentially be many detached namespaces that we don't want
polluting /dev/ with unusable block handles, so this will delete disks
if the namespace is not active as indicated by the response from identify
namespace. This also skips adding the disk if no capacity is provisioned
to the namespace in the first place.
Signed-off-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Protects against reordering and/or preempting which would allow the
kthread to access the queue descriptor before it is set up
Signed-off-by: Jon Derrick <jonathan.derrick@intel.com>
Acked-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
We need the ability to perform an nvme controller reset as discussed on
the mailing list thread:
http://lists.infradead.org/pipermail/linux-nvme/2015-March/001585.html
This adds a sysfs entry that when written to will reset perform an NVMe
controller reset if the controller was successfully initialized in the
first place.
This also adds locking around resetting the device in the async probe
method so the driver can't schedule two resets.
Signed-off-by: Keith Busch <keith.busch@intel.com>
Cc: Brandon Schultz <brandon.schulz@hgst.com>
Cc: David Sariel <david.sariel@pmcs.com>
Updated by Jens to:
1) Merge this with the ioctl reset patch from David Sariel. The ioctl
path now shares the reset code from the sysfs path.
2) Don't flush work if we fail issuing the reset.
Signed-off-by: Jens Axboe <axboe@fb.com>
The driver needs to track shared tags to support multiple namespaces
that may be dynamically allocated or deleted. Relying on the first
request_queue's hctx's is not appropriate as we cannot clear outstanding
tags for all namespaces using this handle, nor can the driver easily track
all request_queue's hctx as namespaces are attached/detached. Instead,
this patch uses the nvme_dev's tagset to get the shared tag resources
instead of through a request_queue hctx.
Signed-off-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Do not retry failed sync commands so the original status may be seen
without issuing unnecessary retries.
Signed-off-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
A recent change to the ioctl handling caused a new harmless
warning in the NVMe driver on all 32-bit machines:
drivers/block/nvme-core.c: In function 'nvme_submit_io':
drivers/block/nvme-core.c:1794:29: warning: cast to pointer from integer of different size [-Wint-to-pointer-cast]
In order to shup up that warning, this introduces a new
temporary variable that uses a double cast to extract
the pointer from an __u64 structure member.
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Fixes: a67a95134f ("NVMe: Meta data handling through submit io ioctl")
Acked-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Replaces req->sense_len usage, which is not owned by the LLD, to
req->special to contain the command result for driver created commands,
and sets the result unconditionally on completion.
Signed-off-by: Keith Busch <keith.busch@intel.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Jens Axboe <axboe@fb.com>
Fixes: d29ec8241c ("nvme: submit internal commands through the block layer")
Signed-off-by: Jens Axboe <axboe@fb.com>
Use block layer queues with an internal cmd_type to submit internally
generated NVMe commands. This both simplifies the code a lot and allow
for a better structure. For example now the LighNVM code can construct
commands without knowing the details of the underlying I/O descriptors.
Or a future NVMe over network target could inject commands, as well as
could the SCSI translation and ioctl code be reused for such a beast.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
Most users want the generic device, so store that in struct nvme_dev
instead of the pci_dev. This also happens to be a nice step towards
making some code reusable for non-PCI transports.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
Note that we keep the unused timeout argument, but allow callers to
pass 0 instead of a timeout if they want the default. This will allow
adding a timeout to the pass through path later on.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
The queue_lock needs to be taken with irqs disabled. This is mostly
due to the old pre blk-mq usage pattern, but we've also picked it up
in most of the few places where we use the queue_lock with blk-mq.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
Pull block driver updates from Jens Axboe:
"This is the block driver pull request for 4.1. As with the core bits,
this is a relatively slow round. This pull request contains:
- Various fixes and cleanups for NVMe, from Alexey Khoroshilov, Chong
Yuan, myself, Keith Busch, and Murali Iyer.
- Documentation and code cleanups for nbd from Markus Pargmann.
- Change of brd maintainer to me, from Ross Zwisler. At least the
email doesn't bounce anymore then.
- Two xen-blkback fixes from Tao Chen"
* 'for-4.1/drivers' of git://git.kernel.dk/linux-block: (23 commits)
NVMe: Meta data handling through submit io ioctl
NVMe: Add translation for block limits
NVMe: Remove check for null
NVMe: Fix error handling of class_create("nvme")
xen-blkback: define pr_fmt macro to avoid the duplication of DRV_PFX
xen-blkback: enlarge the array size of blkback name
nbd: Return error pointer directly
nbd: Return error code directly
nbd: Remove fixme that was already fixed
nbd: Restructure debugging prints
nbd: Fix device bytesize type
nbd: Replace kthread_create with kthread_run
nbd: Remove kernel internal header
Documentation: nbd: Add list of module parameters
Documentation: nbd: Reformat to allow more documentation
NVMe: increase depth of admin queue
nvme: Fix PRP list calculation for non-4k system page size
NVMe: Fix blk-mq hot cpu notification
NVMe: embedded iod mask cleanup
NVMe: Freeze admin queue on device failure
...
This adds support for the extended metadata formats through the submit
IO ioctl, and simplifies the rest when using a separate metadata format.
Signed-off-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Checking fails static analysis due to additional arithmetic prior to
the NULL check. Mapping doesn't return NULL here anyway, so removing
the check.
Reported-by: Dan Carpenter <dan.carpenter@oracle.com>
Signed-off-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
class_create() returns ERR_PTR on failure,
so IS_ERR() should be used instead of check for NULL.
Found by Linux Driver Verification project (linuxtesting.org).
Signed-off-by: Alexey Khoroshilov <khoroshilov@ispras.ru>
Acked-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Usually the admin queue depth of 64 is plenty, but for some use cases we
really need it larger. Examples are use cases like MAT, where you have
to touch all of NAND for init/format like purposes. In those cases, we
see a good 2x increase with an increased queue depth.
Signed-off-by: Jens Axboe <axboe@fb.com>
Acked-by: Keith Busch <keith.busch@intel.com>
PRP list calculation is supposed to be based on device's page size.
Systems with page size larger than device's page size cause corruption
to the name space as well as system memory with out this fix.
Systems like x86 might not experience this issue because it uses
PAGE_SIZE of 4K where as powerpc uses PAGE_SIZE of 64k while NVMe device's
page size varies depending upon the vendor.
Signed-off-by: Murali Iyer <mniyer@us.ibm.com>
Signed-off-by: Brian King <brking@linux.vnet.ibm.com>
Acked-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
The driver may issue commands to a device that may never return, so its
request_queue could always have active requests while the controller is
running. Waiting for the queue to freeze could block forever, which is
what blk-mq's hot cpu notification handler was doing when nvme drives
were in use.
This has the nvme driver make the asynchronous event command's tag
reserved and does not keep the request active. We can't have more than
one since the request is released back to the request_queue before the
command is completed. Having only one avoids potential tag collisions,
and reserving the tag for this purpose prevents other admin tasks from
reusing the tag.
I also couldn't think of a scenario where issuing AEN requests single
depth is worse than issuing them in batches, so I don't think we lose
anything with this change.
As an added bonus, doing it this way removes "Cancelling I/O" warnings
observed when unbinding the nvme driver from a device.
Reported-by: Yigal Korman <yigal@plexistor.com>
Signed-off-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
This fixes a race accessing an invalid address when a controller's admin
queue is in use during a reset for failure or hot removal occurs. The
admin queue will be frozen to prevent new users from entering prior to
the doorbell queue being unmapped.
Signed-off-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Driver recovery requires the device's list node to have been initialized.
Fixes: https://lkml.org/lkml/2015/3/22/262
Reported-by: Steven Noonan <steven@uplinklabs.net>
Signed-off-by: Keith Busch <keith.busch@intel.com>
Cc: Matthew Wilcox <willy@linux.intel.com>
Cc: Jens Axboe <axboe@fb.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Need to define and use appropriate functions for when BLK_DEV_INTEGRITY
is not set.
Reported-by: Fengguang Wu <fengguang.wu@intel.com>
Signed-off-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
This makes all sync commands uninterruptible and schedules without timeout
so the controller either has to post a completion or the timeout recovery
fails the command. This fixes potential memory or data corruption from
a command timing out too early or woken by a signal. Previously any DMA
buffers mapped for that command would have been released even though we
don't know what the controller is planning to do with those addresses.
Signed-off-by: Keith Busch <keith.busch@intel.com>
We don't track queues in a llist, subscribe to hot-cpu notifications,
or internally retry commands. Delete the unused artifacts.
Signed-off-by: Keith Busch <keith.busch@intel.com>
The driver has to end unreturned commands at some point even if the
controller has not provided a completion. The driver tried to be safe by
deleting IO queues prior to ending all unreturned commands. That should
cause the controller to internally abort inflight commands, but IO queue
deletion request does not have to be successful, so all bets are off. We
still have to make progress, so to be extra safe, this patch doesn't
clear a queue to release the dma mapping for a command until after the
pci device has been disabled.
This patch removes the special handling during device initialization
so controller recovery can be done all the time. This is possible since
initialization is not inlined with pci probe anymore.
Reported-by: Nilish Choudhury <nilesh.choudhury@oracle.com>
Signed-off-by: Keith Busch <keith.busch@intel.com>
This performs the longest parts of nvme device probe in scheduled work.
This speeds up probe significantly when multiple devices are in use.
Signed-off-by: Keith Busch <keith.busch@intel.com>
This creates a new class type for nvme devices to register their
management character devices with. This is so we do not rely on miscdev
to provide enough minors for as many nvme devices some people plan to
use. The previous limit was approximately 60 NVMe controllers, depending
on the platform and kernel. Now the limit is 1M, which ought to be enough
for anybody.
Since we have a new device class, it makes sense to attach the block
devices under this as well, so part of this patch moves the management
handle initialization prior to the namespaces discovery.
Signed-off-by: Keith Busch <keith.busch@intel.com>
Adds support for NVMe metadata formats and exposes block devices for
all namespaces regardless of their format. Namespace formats that are
unusable will have disk capacity set to 0, but a handle to the block
device is created to simplify device management. A namespace is not
usable when the format requires host interleave block and metadata in
single buffer, has no provisioned storage, or has better data but failed
to register with blk integrity.
The namespace has to be scanned in two phases to support separate
metadata formats. The first establishes the sector size and capacity
prior to invoking add_disk. If metadata is required, the capacity will
be temporarilly set to 0 until it can be revalidated and registered with
the integrity extenstions after add_disk completes.
The driver relies on the integrity extensions to provide the metadata
buffer. NVMe requires this be a single physically contiguous region,
so only one integrity segment is allowed per command. If the metadata
is used for T10 PI, the driver provides mappings to save and restore
the reftag physical block translation. The driver provides no-op
functions for generate and verify if metadata is not used for protection
information. This way the setup is always provided by the block layer.
If a request does not supply a required metadata buffer, the command
is failed with bad address. This could only happen if a user manually
disables verify/generate on such a disk. The only exception to where
this is okay is if the controller is capable of stripping/generating
the metadata, which is possible on some types of formats.
The metadata scatter gather list now occupies the spot in the nvme_iod
that used to be used to link retryable IOD's, but we don't do that
anymore, so the field was unused.
Signed-off-by: Keith Busch <keith.busch@intel.com>
Pull block driver changes from Jens Axboe:
"This contains:
- The 4k/partition fixes for brd from Boaz/Matthew.
- A few xen front/back block fixes from David Vrabel and Roger Pau
Monne.
- Floppy changes from Takashi, cleaning the device file creation.
- Switching libata to use the new blk-mq tagging policy, removing
code (and a suboptimal implementation) from libata. This will
throw you a merge conflict, since a bug in the original libata
tagging code was fixed since this code was branched. Trivial.
From Shaohua.
- Conversion of loop to blk-mq, from Ming Lei.
- Cleanup of the io_schedule() handling in bsg from Peter Zijlstra.
He claims it improves on unreadable code, which will cost him a
beer.
- Maintainer update or NDB, now handled by Markus Pargmann.
- NVMe:
- Optimization from me that avoids a kmalloc/kfree per IO for
smaller (<= 8KB) IO. This cuts about 1% of high IOPS CPU
overhead.
- Removal of (now) dead RCU code, a relic from before NVMe was
converted to blk-mq"
* 'for-3.20/drivers' of git://git.kernel.dk/linux-block:
xen-blkback: default to X86_32 ABI on x86
xen-blkfront: fix accounting of reqs when migrating
xen-blkback,xen-blkfront: add myself as maintainer
block: Simplify bsg complete all
floppy: Avoid manual call of device_create_file()
NVMe: avoid kmalloc/kfree for smaller IO
MAINTAINERS: Update NBD maintainer
libata: make sata_sil24 use fifo tag allocator
libata: move sas ata tag allocation to libata-scsi.c
libata: use blk taging
NVMe: within nvme_free_queues(), delete RCU sychro/deferred free
null_blk: suppress invalid partition info
brd: Request from fdisk 4k alignment
brd: Fix all partitions BUGs
axonram: Fix bug in direct_access
loop: add blk-mq.h include
block: loop: don't handle REQ_FUA explicitly
block: loop: introduce lo_discard() and lo_req_flush()
block: loop: say goodby to bio
block: loop: improve performance via blk-mq
Currently we allocate an nvme_iod for each IO, which holds the
sg list, prps, and other IO related info. Set a threshold of
2 pages and/or 8KB of data, below which we can just embed this
in the per-command pdu in blk-mq. For any IO at or below
NVME_INT_PAGES and NVME_INT_BYTES, we save a kmalloc and kfree.
For higher IOPS, this saves up to 1% of CPU time.
Signed-off-by: Jens Axboe <axboe@fb.com>
Reviewed-by: Keith Busch <keith.busch@intel.com>
Converting from to blk-queue got rid of the driver's RCU
locking-on-queue, so removing unnecessary RCU locking-on-queue
artefacts.
Reviewed-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Kelly Nicole Kaoudis <kaoudis@colorado.edu>
Signed-off-by: Jens Axboe <axboe@fb.com>
This was inadvertently dropped from an earlier commit, otherwise
the check against cq_vector == -1 to prevent double free doesn't
make any sense.
Fixes: 2b25d98179
Signed-off-by: Jens Axboe <axboe@fb.com>
The queues and device need to be locked when messing with them.
Signed-off-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
