CMB doesn't get unmapped until removal while getting remapped on every
reset. Add the unmapping and sysfs file removal to the reset path in
nvme_pci_disable to match the mapping path in nvme_pci_enable.
Fixes: 202021c1a ("nvme : Add sysfs entry for NVMe CMBs when appropriate")
Signed-off-by: Jon Derrick <jonathan.derrick@intel.com>
Acked-by: Keith Busch <keith.busch@intel.com>
Reviewed-By: Stephen Bates <sbates@raithlin.com>
Cc: <stable@vger.kernel.org> # 4.9+
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
Free up kmalloc allocated memory if failure happens while handling L2P
table transfer in nvme_nvm_get_l2p_tbl.
Fixes: 8e79b5cb ("lightnvm: move block provisioning to targets")
Signed-off-by: Rakesh Pandit <rakesh@tuxera.com>
Reviewed-by: Javier González <javier@cnexlabs.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
With gcc 4.1.2:
drivers/nvme/host/lightnvm.c: In function ‘nvme_nvm_submit_io’:
drivers/nvme/host/lightnvm.c:498: warning: ‘rq’ is used uninitialized in this function
Indeed, since commit 2e13f33a24 ("lightnvm: create cmd before
allocating request"), the request is passed to nvme_nvm_rqtocmd() before
it is allocated.
Fortunately, as of commit 91276162de ("lightnvm: refactor end_io
functions for sync"), nvme_nvm_rqtocmd () no longer uses the passed
request, so this warning is a false positive.
Drop the unused parameter to clean up the code and kill the warning.
Fixes: 2e13f33a24 ("lightnvm: create cmd before allocating request")
Fixes: 91276162de ("lightnvm: refactor end_io functions for sync")
Signed-off-by: Geert Uytterhoeven <geert@linux-m68k.org>
Signed-off-by: Jens Axboe <axboe@fb.com>
Create nvme command before allocating a request using
nvme_alloc_request, which uses the command direction. Up until now, the
command has been zeroized, so all commands have been allocated as a
read operation.
Signed-off-by: Javier González <javier@cnexlabs.com>
Reviewed-by: Matias Bjørling <matias@cnexlabs.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
Remove the request_idx parameter, which can't be used safely now that we
support I/O schedulers with blk-mq. Except for a superflous check in
mtip32xx it was unused anyway.
Also pass the tag_set instead of just the driver data - this allows drivers
to avoid some code duplication in a follow on cleanup.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
Christoph writes:
"A couple more updates for 4.12. The biggest pile is fc and lpfc
updates from James, but there are various small fixes and cleanups as
well."
Fixes up a few merge issues, and also a warning in
lpfc_nvmet_rcv_unsol_abort() if CONFIG_NVME_TARGET_FC isn't enabled.
Signed-off-by: Jens Axboe <axboe@fb.com>
This function just returns the same error code and sense data as
the default statement in the switch in the caller.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Keith Busch <keith.busch@intel.com>
The current command submission code uses a sector-based value when
considering the maximum number of blocks per command. With a
4k-formatted namespace and a command exceeding max hardware limits, this
calculation doesn't split IOs which should be split and fails in the
nvme layer. This patch fixes that calculation and enables IO splitting
in these circumstances.
Signed-off-by: Jon Derrick <jonathan.derrick@intel.com>
Reviewed-by: Jens Axboe <axboe@fb.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Do not call nvmf_free_options() from the nvme_fc_ctlr destructor if
nvme_fc_create_ctrl() returns an error, because nvmf_create_ctrl()
frees the options when an error is returned.
Signed-off-by: Ewan D. Milne <emilne@redhat.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
We're probably going to be stuck quirking APST off on an over-broad
range of devices for 4.11. Let's make it easy to override the quirk
for testing.
Signed-off-by: Andy Lutomirski <luto@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
Debugging APST is currently a bit of a pain. This gives optional
simple log messages that describe the APST state.
The easiest way to use this is probably with the nvme_core.dyndbg=+p
module parameter.
Signed-off-by: Andy Lutomirski <luto@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
There was a typo in the description of the timeout heuristic.
Signed-off-by: Andy Lutomirski <luto@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
Found by sparse.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: James Smart <james.smart@broadcom.com>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Found by sparse.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: James Smart <james.smart@broadcom.com>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
That's what it's used as.
Found by sparse.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: James Smart <james.smart@broadcom.com>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
The passed in desc_len is a big endian value, so mark it as such.
Found by sparse.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: James Smart <james.smart@broadcom.com>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Found by sparse.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: James Smart <james.smart@broadcom.com>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Found by sparse.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: James Smart <james.smart@broadcom.com>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
This patch actually does quite a few things. When looking to add
controller reset support, the organization modeled after rdma was
very fragmented. rdma duplicates the reset and teardown paths and does
different things to the block layer on the two paths. The code to build
up the controller is also duplicated between the initial creation and
the reset/error recovery paths. So I decided to make this sane.
I reorganized the controller creation and teardown so that there is a
connect path and a disconnect path. Initial creation obviously uses
the connect path. Controller teardown will use the disconnect path,
followed last access code. Controller reset will use the disconnect
path to stop operation, and then the connect path to re-establish
the controller.
Along the way, several things were fixed
- aens were not properly set up. They are allocated differently from
the per-request structure on the blk queues.
- aens were oddly torn down. the prior patch corrected to abort, but
we still need to dma unmap and free relative elements.
- missed a few ref counting points: in aen completion and on i/o's
that fail
- controller initial create failure paths were still confused vs teardown
before converting to ref counting vs after we convert to refcounting.
Signed-off-by: James Smart <james.smart@broadcom.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Add abort support for aens. Commonized the op abort to apply to aen or
real ios (caused some reorg/routine movement). Abort path sets termination
flag in prep for next patch that will be watching i/o abort completion
before proceeding with controller teardown.
Now that we're aborting aens, the "exit" code that simply cleared out
their context no longer applies.
Also clarified how we detect an AEN vs a normal io - by a flag, not
by whether a rq exists or the a rqno is out of range.
Note: saw some interesting cases where if the queues are stopped and
we're waiting for the aborts, the core layer can call the complete_rq
callback for the io. So the io completion synchronizes link side completion
with possible blk layer completion under error.
Signed-off-by: James Smart <james.smart@broadcom.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Christoph Hellwig <hch@lst.de>
The code validates the command_id in the response to the original
sqe command. But prior code was using the rq->rqno as the sqe command
id. The core layer overwrites what the transport set there originally.
Use the actual sqe content.
Signed-off-by: James Smart <james.smart@broadcom.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Currently most IOs which return the nvme error codes are retried on
the other path if those IOs returns EIO from NVMe driver. This
patch let Multipath distinguish nvme media error codes and some
generic or cmd-specific nvme error codes so that multipath will
not retry those kinds of IO, to save bandwidth.
Signed-off-by: Junxiong Guan <guanjunxiong@huawei.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
If an IO timeout occurs, it's helpful to know if the controller did not
post a completion or the driver missed an interrupt. While we never expect
the latter, this patch will make it possible to tell the difference so
we don't have to guess.
Signed-off-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Tested-by: Johannes Thumshirn <jthumshirn@suse.de>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
The FC-NVME spec revised syntax to avoid comma separators.
Sync with the change in the parser for traddr on port attachments.
Signed-off-by: James Smart <james.smart@broadcom.com>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
remoteport teardown never aborted the LS opertions. Add support.
Signed-off-by: James Smart <james.smart@broadcom.com>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
Link LS's on the remoteport rather than the controller. LS's are
between nport's. Makes more sense, especially on async teardown where
the controller is torn down regardless of the LS (LS is more of a notifier
to the target of the teardown), to have them on the remoteport.
While revising ls send/done routines, issues were seen relative to
refcounting and cleanup, especially in async path. Reworked these code
paths.
Signed-off-by: James Smart <james.smart@broadcom.com>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
target transport:
----------------------
There are cases when there is a need to abort in-progress target
operations (writedata) so that controller termination or errors can
clean up. That can't happen currently as the abort is another target
op type, so it can't be used till the running one finishes (and it may
not). Solve by removing the abort op type and creating a separate
downcall from the transport to the lldd to request an io to be aborted.
The transport will abort ios on queue teardown or io errors. In general
the transport tries to call the lldd abort only when the io state is
idle. Meaning: ops that transmit data (readdata or rsp) will always
finish their transmit (or the lldd will see a state on the
link or initiator port that fails the transmit) and the done call for
the operation will occur. The transport will wait for the op done
upcall before calling the abort function, and as the io is idle, the
io can be cleaned up immediately after the abort call; Similarly, ios
that are not waiting for data or transmitting data must be in the nvmet
layer being processed. The transport will wait for the nvmet layer
completion before calling the abort function, and as the io is idle,
the io can be cleaned up immediately after the abort call; As for ops
that are waiting for data (writedata), they may be outstanding
indefinitely if the lldd doesn't see a condition where the initiatior
port or link is bad. In those cases, the transport will call the abort
function and wait for the lldd's op done upcall for the operation, where
it will then clean up the io.
Additionally, if a lldd receives an ABTS and matches it to an outstanding
request in the transport, A new new transport upcall was created to abort
the outstanding request in the transport. The transport expects any
outstanding op call (readdata or writedata) will completed by the lldd and
the operation upcall made. The transport doesn't act on the reported
abort (e.g. clean up the io) until an op done upcall occurs, a new op is
attempted, or the nvmet layer completes the io processing.
fcloop:
----------------------
Updated to support the new target apis.
On fcp io aborts from the initiator, the loopback context is updated to
NULL out the half that has completed. The initiator side is immediately
called after the abort request with an io completion (abort status).
On fcp io aborts from the target, the io is stopped and the initiator side
sees it as an aborted io. Target side ops, perhaps in progress while the
initiator side is done, continue but noop the data movement as there's no
structure on the initiator side to reference.
patch also contains:
----------------------
Revised lpfc to support the new abort api
commonized rsp buffer syncing and nulling of private data based on
calling paths.
errors in op done calls don't take action on the fod. They're bad
operations which implies the fod may be bad.
Signed-off-by: James Smart <james.smart@broadcom.com>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
Current design has the fcloop job struct, used for both initiator and
target processing, allocated as part of the initiator request structure.
On aborts, the initiator side (based on the request) may terminate, yet
the target side wants to continue processing. the target side can't do
that if the initiator side goes away.
Revise fcloop to allocate an independent target side structure when it
starts an io from the initiator.
Added a lock to the request struct as well to synchronize pointer updates
on abort calls.
Modified target downcalls to recognize conditions where initiator has
aborted the io (thus nulled the pointer between job structs), thus
avoid referencing sgl lists which are gone and no longer making upcalls
to the initiator.
In conditions where the targetport is no longer connected, have the
initiator return an access failure rather than simulating a command
completion.
Signed-off-by: James Smart <james.smart@broadcom.com>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
With the advent of the opdone calls changing context, the lldd can no
longer assume that once the op->done call returns for RSP operations
that the request struct is no longer being accessed.
As such, revise the lldd api for a req_release callback that the
transport will call when the job is complete. This will also be used
with abort cases.
Fixed text in api header for change in io complete semantics.
Revised lpfc to support the new req_release api.
Signed-off-by: James Smart <james.smart@broadcom.com>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
Two new feature flags were added to control whether upcalls to the
transport result in context switches or stay in the calling context.
NVMET_FCTGTFEAT_CMD_IN_ISR:
By default, if the flag is not set, the transport assumes the
lldd is in a non-isr context and in the cpu context it should be
for the io queue. As such, the cmd handler is called directly in the
calling context.
If the flag is set, indicating the upcall is an isr context, the
transport mandates a transition to a workqueue. The workqueue assigned
to the queue is used for the context.
NVMET_FCTGTFEAT_OPDONE_IN_ISR
By default, if the flag is not set, the transport assumes the
lldd is in a non-isr context and in the cpu context it should be
for the io queue. As such, the fcp operation done callback is called
directly in the calling context.
If the flag is set, indicating the upcall is an isr context, the
transport mandates a transition to a workqueue. The workqueue assigned
to the queue is used for the context.
Updated lpfc for flags
Signed-off-by: James Smart <james.smart@broadcom.com>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
This is safer as it doesn't rely on the data being stored in
a single page in an sgl.
It also aids our effort to start phasing out users of sg_page. See [1].
For this we kmalloc some memory, copy to it and free at the end. Note:
we can't allocate this memory on the stack as the kbuild test robot
reports some frame size overflows on i386.
[1] https://lwn.net/Articles/720053/
Signed-off-by: Logan Gunthorpe <logang@deltatee.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Max Gurtovoy <maxg@mellanox.com>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
This change provides a mechanism to reduce the number of MMIO doorbell
writes for the NVMe driver. When running in a virtualized environment
like QEMU, the cost of an MMIO is quite hefy here. The main idea for
the patch is provide the device two memory location locations:
1) to store the doorbell values so they can be lookup without the doorbell
MMIO write
2) to store an event index.
I believe the doorbell value is obvious, the event index not so much.
Similar to the virtio specification, the virtual device can tell the
driver (guest OS) not to write MMIO unless you are writing past this
value.
FYI: doorbell values are written by the nvme driver (guest OS) and the
event index is written by the virtual device (host OS).
The patch implements a new admin command that will communicate where
these two memory locations reside. If the command fails, the nvme
driver will work as before without any optimizations.
Contributions:
Eric Northup <digitaleric@google.com>
Frank Swiderski <fes@google.com>
Ted Tso <tytso@mit.edu>
Keith Busch <keith.busch@intel.com>
Just to give an idea on the performance boost with the vendor
extension: Running fio [1], a stock NVMe driver I get about 200K read
IOPs with my vendor patch I get about 1000K read IOPs. This was
running with a null device i.e. the backing device simply returned
success on every read IO request.
[1] Running on a 4 core machine:
fio --time_based --name=benchmark --runtime=30
--filename=/dev/nvme0n1 --nrfiles=1 --ioengine=libaio --iodepth=32
--direct=1 --invalidate=1 --verify=0 --verify_fatal=0 --numjobs=4
--rw=randread --blocksize=4k --randrepeat=false
Signed-off-by: Rob Nelson <rlnelson@google.com>
[mlin: port for upstream]
Signed-off-by: Ming Lin <mlin@kernel.org>
[koike: updated for upstream]
Signed-off-by: Helen Koike <helen.koike@collabora.co.uk>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Keith Busch <keith.busch@intel.com>
The QPRIO field is only valid if weighted round robin arbitration is used,
and this driver doesn't enable that controller configuration option.
Signed-off-by: Keith Busch <keith.busch@intel.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
There's a report that it malfunctions with APST on.
See https://bugs.launchpad.net/ubuntu/+source/linux/+bug/1678184
Cc: Kai-Heng Feng <kai.heng.feng@canonical.com>
Signed-off-by: Andy Lutomirski <luto@kernel.org>
Signed-off-by: Jens Axboe <axboe@fb.com>
I got a couple more reports: the Samsung APST issues appears to
affect multiple 950-series devices in Dell XPS 15 9550 and Precision
5510 laptops. Change the quirk: rather than blacklisting the
firmware on the first problematic SSD that was reported, disable
APST on all 144d:a802 devices if they're installed in the two
affected Dell models. While we're at it, disable only the deepest
sleep state instead of all of them -- the reporters say that this is
sufficient to fix the problem.
(I have a device that appears to be entirely identical to one of the
affected devices, but I have a different Dell laptop, so it's not
the case that all Samsung devices with firmware BXW75D0Q are broken
under all circumstances.)
Samsung engineers have an affected system, and hopefully they'll
give us a better workaround some time soon. In the mean time, this
should minimize regressions.
See https://bugs.launchpad.net/ubuntu/+source/linux/+bug/1678184
Cc: Kai-Heng Feng <kai.heng.feng@canonical.com>
Signed-off-by: Andy Lutomirski <luto@kernel.org>
Signed-off-by: Jens Axboe <axboe@fb.com>
Now that all drivers that call blk_mq_complete_requests have a
->complete callback we can remove the direct call to blk_mq_end_request,
as well as the error argument to blk_mq_complete_request.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Reviewed-by: Bart Van Assche <Bart.VanAssche@sandisk.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Currently it's used by the lighnvm passthrough ioctl, but we'd like to make
it private in preparation of block layer specific error code. Lighnvm already
returns the real NVMe status anyway, so I think we can just limit it to
returning -EIO for any status set.
This will need a careful audit from the lightnvm folks, though.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
We want our own clearly defined error field for NVMe passthrough commands,
and the request errors field is going away in its current form.
Just store the status and result field in the nvme_request field from
hardirq completion context (using a new helper) and then generate a
Linux errno for the block layer only when we actually need it.
Because we can't overload the status value with a negative error code
for cancelled command we now have a flags filed in struct nvme_request
that contains a bit for this condition.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
nvme_complete_async_event expects the little endian status code
including the phase bit, and a new completion handler I plan to
introduce will do so as well.
Change the status variable into the little endian format with the
phase bit used in the NVMe CQE to fix / enable this.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
This patch changes the behavior of the lightnvm driver as follows:
* REQ_FAILFAST_MASK is set for read-ahead requests.
* If no I/O priority has been set in the bio, the I/O priority is
copied from the I/O context.
* The rq_disk member is initialized if bio->bi_bdev != NULL.
* The bio sector offset is copied into req->__sector instead of
retaining the value -1 set by blk_mq_alloc_request().
* req->errors is initialized to zero.
Signed-off-by: Bart Van Assche <bart.vanassche@sandisk.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Cc: Matias Bjørling <m@bjorling.me>
Cc: Adam Manzanares <adam.manzanares@wdc.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
The NVMe I/O command control bits are 16 bytes, but is interpreted as
32 bytes in the lightnvm user I/O data path.
Signed-off-by: Javier González <javier@cnexlabs.com>
Signed-off-by: Matias Bjørling <matias@cnexlabs.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
The asserts in _nvme_nvm_check_size are not compiled due to the function
not begin called. Make sure that it is called, and also fix the wrong
sizes of asserts for nvme_nvm_addr_format, and nvme_nvm_bb_tbl, which
checked for number of bits instead of bytes.
Reported-by: Scott Bauer <scott.bauer@intel.com>
Signed-off-by: Matias Bjørling <matias@cnexlabs.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Until now erases have been submitted as synchronous commands through a
dedicated erase function. In order to enable targets implementing
asynchronous erases, refactor the erase path so that it uses the normal
async I/O submission functions. If a target requires sync I/O, it can
implement it internally. Also, adapt rrpc to use the new erase path.
Signed-off-by: Javier González <javier@cnexlabs.com>
Fixed spelling error.
Signed-off-by: Matias Bjørling <matias@cnexlabs.com>
Signed-off-by: Matias Bjørling <matias@cnexlabs.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
There are two closely named structs in lightnvm:
struct nvme_nvm_addr_format and
struct nvme_addr_format.
The first struct has 4 reserved bytes at the end, the second does not.
(gdb) p sizeof(struct nvme_nvm_addr_format)
$1 = 16
(gdb) p sizeof(struct nvm_addr_format)
$2 = 12
In the nvme_nvm_identify function we memcpy from the larger struct to the
smaller struct. We incorrectly pass the length of the larger struct
and overflow by 4 bytes, lets not do that.
Signed-off-by: Scott Bauer <scott.bauer@intel.com>
Signed-off-by: Matias Bjørling <matias@cnexlabs.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
both our sqsize and the controller MQES cap are a 0 based value,
so making it 1 based is wrong.
Reported-by: Trapp, Darren <Darren.Trapp@cavium.com>
Reported-by: Daniel Verkamp <daniel.verkamp@intel.com>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
both our sqsize and the controller MQES cap are a 0 based value,
so making it 1 based is wrong.
Reported-by: Trapp, Darren <Darren.Trapp@cavium.com>
Reported-by: Daniel Verkamp <daniel.verkamp@intel.com>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
both our sqsize and the controller MQES cap are a 0 based value,
so making it 1 based is wrong.
Reported-by: Trapp, Darren <Darren.Trapp@cavium.com>
Reported-by: Daniel Verkamp <daniel.verkamp@intel.com>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
