In the case of small NVMe-oF queue size (<32) we may enter a deadlock
caused by the fact that the IB completions aren't sent waiting for 32
and the send queue will fill up.

The error is seen as (using mlx5):
[ 2048.693355] mlx5_0:mlx5_ib_post_send:3765:(pid 7273):
[ 2048.693360] nvme nvme1: nvme_rdma_post_send failed with error code -12

This patch changes the way the signaling is done so that it depends on
the queue depth now. The magic define has been removed completely.

Cc: stable@vger.kernel.org
Signed-off-by: NMarta Rybczynska <marta.rybczynska@kalray.eu>
Signed-off-by: NSamuel Jones <sjones@kalray.eu>
Acked-by: NSagi Grimberg <sagi@grimberg.me>
Signed-off-by: NChristoph Hellwig <hch@lst.de>

Per the recommendation by Sagi on:
http://lists.infradead.org/pipermail/linux-nvme/2017-April/009261.html

Rather than waiting for reset work thread to stop queues and abort the ios,
immediately stop the queues on error detection. Reset thread will restop
the queues (as it's called on other paths), but it does not appear to have
a side effect.
Signed-off-by: NJames Smart <james.smart@broadcom.com>
Reviewed-by: NJohannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NJens Axboe <axboe@fb.com>

In order to create an association, the remoteport must be
serving either a target role or a discovery role.
Signed-off-by: NJames Smart <james.smart@broadcom.com>
Reviewed-by: NJohannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NJens Axboe <axboe@fb.com>

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: 202021c1 ("nvme : Add sysfs entry for NVMe CMBs when appropriate")
Signed-off-by: NJon Derrick <jonathan.derrick@intel.com>
Acked-by: NKeith Busch <keith.busch@intel.com>
Reviewed-By: NStephen Bates <sbates@raithlin.com>
Cc: <stable@vger.kernel.org> # 4.9+
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NJens 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: NRakesh Pandit <rakesh@tuxera.com>
Reviewed-by: NJavier González <javier@cnexlabs.com>
Signed-off-by: NJens 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 2e13f33a ("lightnvm: create cmd before
allocating request"), the request is passed to nvme_nvm_rqtocmd() before
it is allocated.

Fortunately, as of commit 91276162 ("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: 2e13f33a ("lightnvm: create cmd before allocating request")
Fixes: 91276162 ("lightnvm: refactor end_io functions for sync")
Signed-off-by: NGeert Uytterhoeven <geert@linux-m68k.org>
Signed-off-by: NJens 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: NJavier González <javier@cnexlabs.com>
Reviewed-by: NMatias Bjørling <matias@cnexlabs.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NJens 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: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NJens 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: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NKeith Busch <keith.busch@intel.com>

Found by sparse.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NMatias Bjørling <matias@cnexlabs.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: NJon Derrick <jonathan.derrick@intel.com>
Reviewed-by: NJens Axboe <axboe@fb.com>
Signed-off-by: NChristoph 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: NEwan D. Milne <emilne@redhat.com>
Signed-off-by: NChristoph 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: NAndy Lutomirski <luto@kernel.org>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NJens 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: NAndy Lutomirski <luto@kernel.org>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NJens Axboe <axboe@fb.com>

There was a typo in the description of the timeout heuristic.
Signed-off-by: NAndy Lutomirski <luto@kernel.org>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NJens Axboe <axboe@fb.com>

Found by sparse.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NJames Smart <james.smart@broadcom.com>
Reviewed-by: NJohannes 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: NJames Smart <james.smart@broadcom.com>
Signed-off-by: NChristoph 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: NJames Smart <james.smart@broadcom.com>
Reviewed-by: NSagi Grimberg <sagi@grimberg.me>
Signed-off-by: NChristoph 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: NJames Smart <james.smart@broadcom.com>
Reviewed-by: NSagi Grimberg <sagi@grimberg.me>
Signed-off-by: NChristoph 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: NJunxiong Guan <guanjunxiong@huawei.com>
Signed-off-by: NChristoph 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: NKeith Busch <keith.busch@intel.com>
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Tested-by: NJohannes Thumshirn <jthumshirn@suse.de>
Reviewed-by: NJohannes Thumshirn <jthumshirn@suse.de>

remoteport teardown never aborted the LS opertions. Add support.
Signed-off-by: NJames Smart <james.smart@broadcom.com>
Signed-off-by: NSagi 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: NJames Smart <james.smart@broadcom.com>
Signed-off-by: NSagi 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: NRob Nelson <rlnelson@google.com>
[mlin: port for upstream]
Signed-off-by: NMing Lin <mlin@kernel.org>
[koike: updated for upstream]
Signed-off-by: NHelen Koike <helen.koike@collabora.co.uk>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NKeith 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: NKeith Busch <keith.busch@intel.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NSagi 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: NAndy Lutomirski <luto@kernel.org>
Signed-off-by: NJens 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: NAndy Lutomirski <luto@kernel.org>
Signed-off-by: NJens 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: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NJohannes Thumshirn <jthumshirn@suse.de>
Reviewed-by: NBart Van Assche <Bart.VanAssche@sandisk.com>
Signed-off-by: NJens 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: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NJens 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: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NJohannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: NJens 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: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NJohannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: NJens 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: NBart Van Assche <bart.vanassche@sandisk.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Cc: Matias Bjørling <m@bjorling.me>
Cc: Adam Manzanares <adam.manzanares@wdc.com>
Signed-off-by: NJens Axboe <axboe@fb.com>

Signed-off-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NBjorn Helgaas <bhelgaas@google.com>
Reviewed-by: NThomas Gleixner <tglx@linutronix.de>
Reviewed-by: NKeith Busch <keith.busch@intel.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: NJavier González <javier@cnexlabs.com>
Signed-off-by: NMatias Bjørling <matias@cnexlabs.com>
Signed-off-by: NJens 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: NScott Bauer <scott.bauer@intel.com>
Signed-off-by: NMatias Bjørling <matias@cnexlabs.com>
Signed-off-by: NJens 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: NJavier González <javier@cnexlabs.com>
Fixed spelling error.
Signed-off-by: NMatias Bjørling <matias@cnexlabs.com>
Signed-off-by: NMatias Bjørling <matias@cnexlabs.com>
Signed-off-by: NJens 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: NScott Bauer <scott.bauer@intel.com>
Signed-off-by: NMatias Bjørling <matias@cnexlabs.com>
Signed-off-by: NJens 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: NTrapp, Darren <Darren.Trapp@cavium.com>
Reported-by: NDaniel Verkamp <daniel.verkamp@intel.com>
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NJens 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: NTrapp, Darren <Darren.Trapp@cavium.com>
Reported-by: NDaniel Verkamp <daniel.verkamp@intel.com>
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NJens Axboe <axboe@fb.com>

But now for the real NVMe Write Zeroes yet, just to get rid of the
discard abuse for zeroing.  Also rename the quirk flag to be a bit
more self-explanatory.
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>