No need to differentiate fabrics from pci/loop, also lower
it to 32 as we don't really need 256 inflight admin commands.
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Martin K. Petersen <martin.petersen@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Max Gurtovoy <maxg@mellanox.com>
Signed-off-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
This adds support for Directives in NVMe, particular for the Streams
directive. Support for Directives is a new feature in NVMe 1.3. It
allows a user to pass in information about where to store the data, so
that it the device can do so most effiently. If an application is
managing and writing data with different life times, mixing differently
retentioned data onto the same locations on flash can cause write
amplification to grow. This, in turn, will reduce performance and life
time of the device.
Reviewed-by: Martin K. Petersen <martin.petersen@oracle.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The NVMe 1.3 spec introduces Namespace Optimal IO Boundaries (NOIOB),
which standardizes the stripe mechanism we currently have quirks for.
This patch implements the necessary logic to handle this new feature.
Signed-off-by: Scott Bauer <scott.bauer@intel.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Add the new to NVMe 1.3 fields EDSTT, DSTO, FWUG, HCTMA, MNTMT, MXTMT,
and SANICAP into the idenfity controller data structure.
Signed-off-by: Guan Junxiong <guanjunxiong@huawei.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Allow overriding the announced NVMe Version of a via configfs.
This is particularly helpful when debugging new features for the host
or target side without bumping the hard coded version (as the target
might not be fully compliant to the announced version yet).
Signed-off-by: Johannes Thumshirn <jthumshirn@suse.de>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Reviewed-by: Guan Junxiong <guanjunxiong@huawei.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Johannes Thumshirn <jthumshirn@suse.de>
Reviewed-by: Max Gurtovoy <maxg@mellanox.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Use NVME_IDENTIFY_DATA_SIZE define instead of hard coding the magic
4096 value.
Signed-off-by: Johannes Thumshirn <jthumshirn@suse.de>
Reviewed-by: Max Gurtovoy <maxg@mellanox.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Hannes Reinecke <hare@suse.com>
[hch: converted three more users]
Signed-off-by: Christoph Hellwig <hch@lst.de>
The existing driver initially maps 8192 bytes of BAR0 which is
intended to cover doorbells of admin SQ and CQ. However, if a
large stride, e.g. 10, is used, the doorbell of admin CQ will
be out of 8192 bytes. Consequently, a page fault will be raised
when the admin CQ doorbell is accessed in nvme_configure_admin_queue().
This patch fixes this issue by remapping BAR0 before accessing
admin CQ doorbell if the initial mapping is not enough.
Signed-off-by: Xu Yu <yu.a.xu@intel.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Arnav Dawn <a.dawn@samsung.com>
[hch: split from a larger patch, new changelog]
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>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Amir Goldstein <amir73il@gmail.com>
Reviewed-by: Andy Shevchenko <andriy.shevchenko@linux.intel.com>
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 enum values for QPTYPE, PRTYPE and CMS are off by 1 from the
values defined in figure 42 of the NVM Express over Fabrics 1.0:
http://www.nvmexpress.org/wp-content/uploads/NVMe_over_Fabrics_1_0_Gold_20160605-1.pdf
Fix our enums to match the final spec.
Signed-off-by: Roland Dreier <roland@purestorage.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
NVMe devices can advertise multiple power states. These states can
be either "operational" (the device is fully functional but possibly
slow) or "non-operational" (the device is asleep until woken up).
Some devices can automatically enter a non-operational state when
idle for a specified amount of time and then automatically wake back
up when needed.
The hardware configuration is a table. For each state, an entry in
the table indicates the next deeper non-operational state, if any,
to autonomously transition to and the idle time required before
transitioning.
This patch teaches the driver to program APST so that each successive
non-operational state will be entered after an idle time equal to 100%
of the total latency (entry plus exit) associated with that state.
The maximum acceptable latency is controlled using dev_pm_qos
(e.g. power/pm_qos_latency_tolerance_us in sysfs); non-operational
states with total latency greater than this value will not be used.
As a special case, setting the latency tolerance to 0 will disable
APST entirely. On hardware without APST support, the sysfs file will
not be exposed.
The latency tolerance for newly-probed devices is set by the module
parameter nvme_core.default_ps_max_latency_us.
In theory, the device can expose "default" APST table, but this
doesn't seem to function correctly on my device (Samsung 950), nor
does it seem particularly useful. There is also an optional
mechanism by which a configuration can be "saved" so it will be
automatically loaded on reset. This can be configured from
userspace, but it doesn't seem useful to support in the driver.
On my laptop, enabling APST seems to save nearly 1W.
The hardware tables can be decoded in userspace with nvme-cli.
'nvme id-ctrl /dev/nvmeN' will show the power state table and
'nvme get-feature -f 0x0c -H /dev/nvme0' will show the current APST
configuration.
This feature is quirked off on a known-buggy Samsung device.
Signed-off-by: Andy Lutomirski <luto@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Jens Axboe <axboe@fb.com>
This patch defines CNS field as 8-bit field and avoids cpu_to/from_le
conversions.
Also initialize nvme_command cns value explicitly to NVME_ID_CNS_NS
for readability (don't rely on the fact that NVME_ID_CNS_NS = 0).
Reviewed-by: Max Gurtovoy <maxg@mellanox.com>
Signed-off-by: Parav Pandit <parav@mellanox.com>
Reviewed-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Jens Axboe <axboe@fb.com>
We need to verify that the controller supports the security
commands before actually trying to issue them.
Signed-off-by: Scott Bauer <scott.bauer@intel.com>
[hch: moved the check so that we don't call into the OPAL code if not
supported]
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
NVMe supports up to 256 ranges per DSM command, so wire up support
for ranged discards up to that limit.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
Pull block layer updates from Jens Axboe:
"This is the main block pull request this series. Contrary to previous
release, I've kept the core and driver changes in the same branch. We
always ended up having dependencies between the two for obvious
reasons, so makes more sense to keep them together. That said, I'll
probably try and keep more topical branches going forward, especially
for cycles that end up being as busy as this one.
The major parts of this pull request is:
- Improved support for O_DIRECT on block devices, with a small
private implementation instead of using the pig that is
fs/direct-io.c. From Christoph.
- Request completion tracking in a scalable fashion. This is utilized
by two components in this pull, the new hybrid polling and the
writeback queue throttling code.
- Improved support for polling with O_DIRECT, adding a hybrid mode
that combines pure polling with an initial sleep. From me.
- Support for automatic throttling of writeback queues on the block
side. This uses feedback from the device completion latencies to
scale the queue on the block side up or down. From me.
- Support from SMR drives in the block layer and for SD. From Hannes
and Shaun.
- Multi-connection support for nbd. From Josef.
- Cleanup of request and bio flags, so we have a clear split between
which are bio (or rq) private, and which ones are shared. From
Christoph.
- A set of patches from Bart, that improve how we handle queue
stopping and starting in blk-mq.
- Support for WRITE_ZEROES from Chaitanya.
- Lightnvm updates from Javier/Matias.
- Supoort for FC for the nvme-over-fabrics code. From James Smart.
- A bunch of fixes from a whole slew of people, too many to name
here"
* 'for-4.10/block' of git://git.kernel.dk/linux-block: (182 commits)
blk-stat: fix a few cases of missing batch flushing
blk-flush: run the queue when inserting blk-mq flush
elevator: make the rqhash helpers exported
blk-mq: abstract out blk_mq_dispatch_rq_list() helper
blk-mq: add blk_mq_start_stopped_hw_queue()
block: improve handling of the magic discard payload
blk-wbt: don't throttle discard or write zeroes
nbd: use dev_err_ratelimited in io path
nbd: reset the setup task for NBD_CLEAR_SOCK
nvme-fabrics: Add FC LLDD loopback driver to test FC-NVME
nvme-fabrics: Add target support for FC transport
nvme-fabrics: Add host support for FC transport
nvme-fabrics: Add FC transport LLDD api definitions
nvme-fabrics: Add FC transport FC-NVME definitions
nvme-fabrics: Add FC transport error codes to nvme.h
Add type 0x28 NVME type code to scsi fc headers
nvme-fabrics: patch target code in prep for FC transport support
nvme-fabrics: set sqe.command_id in core not transports
parser: add u64 number parser
nvme-rdma: align to generic ib_event logging helper
...
Signed-off-by: James Smart <james.smart@broadcom.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Jay Freyensee <james_p_freyensee@linux.intel.com>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Add the command structure, optional command set support (ONCS) bit and
a new error code for the Write Zeroes command.
Signed-off-by: Chaitanya Kulkarni <chaitanya.kulkarni@hgst.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
This adds a shared per-request structure for all NVMe I/O. This structure
is embedded as the first member in all NVMe transport drivers request
private data and allows to implement common functionality between the
drivers.
The first use is to replace the current abuse of the SCSI command
passthrough fields in struct request for the NVMe command passthrough,
but it will grow a field more fields to allow implementing things
like common abort handlers in the future.
The passthrough commands are handled by having a pointer to the SQE
(struct nvme_command) in struct nvme_request, and the union of the
possible result fields, which had to be turned from an anonymous
into a named union for that purpose. This avoids having to pass
a reference to a full CQE around and thus makes checking the result
a lot more lightweight.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Ported over from nvme-cli.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Gabriel Krisman Bertazi <krisman@linux.vnet.ibm.com>
Reviewed-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
This makes life easier for nvme-cli and we don't really need the uuid
type anyway to start with.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Gabriel Krisman Bertazi <krisman@linux.vnet.ibm.com>
Reviewed-by: Jay Freyensee <james_p_freyensee@linux.intel.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Import a few updates to nvme.h from nvme-cli. This mostly includes a few
new fields and error codes, but also a few renames that so far are only
used in user space. Also one field is moved from an array of two le64
values to one of 16 u8 values so that we can more easily access it.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Keith Busch <keith.busch@intel.com>
Reviewed-by: Gabriel Krisman Bertazi <krisman@linux.vnet.ibm.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
NVMe 1.2.1 specification adds a tertiary element to the version number.
This updates the macro and its callers to include the final number and
fixup a single place in nvmet where the version was generated manually.
Signed-off-by: Gabriel Krisman Bertazi <krisman@linux.vnet.ibm.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
NVM Express 1.2.1 section 7.9, NVMe Qualified Names, specifies that the
UUID format of NQN uses a UUID based on RFC 4122.
RFC 4122 specifies that the UUID is encoded in big-endian byte order.
Switch the NVMe over Fabrics host ID field from little-endian UUID to
big-endian UUID to match the specification.
Signed-off-by: Daniel Verkamp <daniel.verkamp@intel.com>
Reviewed-by: Jay Freyensee <james_p_freyensee@linux.intel.com>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
KAS: keep-alive support and granularity of kato in units of 100 ms
nvme_admin_keep_alive opcode: 0x18
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
The NVMe over Fabrics specification defines a protocol interface and
related extensions to NVMe that enable operation over network protocols.
The NVMe over Fabrics specification has an NVMe Transport binding for
each NVMe Transport.
This patch adds the fabrics related definitions:
- fabric specific command set and error codes
- transport addressing and binding definitions
- fabrics sgl extensions
- controller identification fabrics enhancements
- discovery log page definition
Signed-off-by: Armen Baloyan <armenx.baloyan@intel.com>
Signed-off-by: James Smart <james.smart@broadcom.com>
Signed-off-by: Jay Freyensee <james.p.freyensee@intel.com>
Signed-off-by: Ming Lin <ming.l@ssi.samsung.com>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Centralize the check if a given NVMe command reads or writes data.
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Jay Freyensee <james.p.freyensee@intel.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Reviewed-by: Jay Freyensee <james.p.freyensee@intel.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Ming Lin <ming.l@ssi.samsung.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Add get_log_page command structure and a corresponding entry in
nvme_command union
Signed-off-by: Armen Baloyan <armenx.baloyan@intel.com>
Reviewed-by: Jay Freyensee <james.p.freyensee@intel.com>
Reviewed--by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
These have been added in NVMe 1.2 and we'll need at least oaes for the
NVMe target driver.
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Jay Freyensee <james.p.freyensee@intel.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Section 3.1 gives the comment for the offset of controller registers
in the specification 1.2a.
Some are mis-copied in the header file nvme.h. Correct them.
Signed-off-by: Wang Sheng-Hui <shhuiw@foxmail.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Jens Axboe <axboe@fb.com>
This makes life easier for future non-PCI drivers where access to the
registers might be more complicated. Note that Linux drivers are
pretty evenly split between the two versions, and in fact the NVMe
driver already uses offsets for the doorbells.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Acked-by: Keith Busch <keith.busch@intel.com>
[Fixed CMBSZ offset]
Signed-off-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
The buildbot complains about this even if it doesn't generate
a a build warning. But it's an easy fix, so here we go:
Reported-by: kbuild test robot <fengguang.wu@intel.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
Currently all NVMe command and completion structures are exposed to userspace
through the uapi version of nvme.h. They are not an ABI between the kernel
and userspace, and will change in C-incompatible way for future versions of
the spec. Move them to the kernel version of the file and rename the uapi
header to nvme_ioctl.h so that userspace can easily detect the presence of
the new clean header. Nvme-cli already carries a local copy of the header,
so it won't be affected by this move.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Acked-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Add a new drivers/block/nvme.h which contains all the driver internal
interface.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Acked-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Releasing IO queues and disks was done in a work queue outside the
controller resume context to delete namespaces if the controller failed
after a resume from suspend. This is unnecessary since we can resume
a device asynchronously.
This patch makes resume use probe_work so it can directly remove
namespaces if the device is manageable but not IO capable. Since the
deleting disks was the only reason we had the convoluted "reset_workfn",
this patch removes that unnecessary indirection.
Signed-off-by: Keith Busch <keith.busch@intel.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
Dynamic namespace attachment means the namespace may be removed at any
time, so the namespace reference count can not be tied to the device
reference count. This fixes a NULL dereference if an opened namespace
is detached from a controller.
Signed-off-by: Keith Busch <keith.busch@intel.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
Controllers can perform optional subsystem resets as introduced in NVMe
1.1. This patch adds an IOCTL to trigger the subsystem reset by writing
"NVMe" to the NSSR register.
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>
Controllers part of an NVMe subsystem may be reset by any other controller
in the subsystem. If the device is capable of subsystem resets, this
patch adds detection for such events and performs appropriate controller
initialization upon subsystem reset detection.
The register bit is a RW1C type, so the driver needs to write a 1 to the
status bit to clear the subsystem reset occured bit during initialization.
Signed-off-by: Keith Busch <keith.busch@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>
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>
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>
Sagi Grimberg