Now ctrl state machine allows to transition from RESETTING to
RECONNECTING.  In nvme-rdma when we receive a rdma cm DISONNECTED event,
we trigger nvme_rdma_error_recovery. This happens also when we execute a
controller reset, issue a cm diconnect request and receive a cm
disconnect reply, as a result, the reset work and the error recovery work
can run concurrently.

Until now the state machine prevented from the error recovery work from
running as a result of a controller reset (RESETTING -> RECONNECTING was
not allowed).

To fix this, we adopt the FC state machine approach, we always transition
from LIVE to RESETTING and only then to RECONNECTING.  We do this both
for the error recovery work and the controller reset work:

 1. transition to RESETTING
 2. teardown the controller association
 3. transition to RECONNECTING

This will restore the protection against reset work and error recovery work
from concurrently running together.

Fixes: 3cec7f9d ("nvme: allow controller RESETTING to RECONNECTING transition")
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>
Signed-off-by: NChristoph Hellwig <hch@lst.de>

In case nvme_rdma_wait_for_cm timeout expires before we get
an established or rejected event (rdma_connect succeeded) from
rdma_cm, we end up with leaking the ib transport resources for
dedicated queue. This scenario can easily reproduced using traffic
test during port toggling.
Also, in order to protect from parallel ib queue destruction, that
may be invoked from different context's, introduce new flag that
stands for transport readiness. While we're here, protect also against
a situation that we can receive rdma_cm events during ib queue destruction.
Signed-off-by: NMax Gurtovoy <maxg@mellanox.com>
Signed-off-by: NChristoph Hellwig <hch@lst.de>

Currently, blk_mq_tagset_iter() iterate over initial hctx tags only.  If
an I/O scheduler is used, it doesn't iterate the hctx scheduler tags and
the static request aren't been updated. For example, while using NVMe
over Fabrics RDMA host, this cause us not to reinit the scheduler
requests and thus not re-register all the memory regions during the
tagset re-initialization in the reconnect flow.

This may lead to a memory registration error:

  "MEMREG for CQE 0xffff88044c14dce8 failed with status memory management operation error (6)"

With this commit we don't need to reinit the requests, and thus fix this
failure.
Signed-off-by: NIsrael Rukshin <israelr@mellanox.com>
Reviewed-by: NSagi Grimberg <sagi@grimberg.me>
Reviewed-by: NMax Gurtovoy <maxg@mellanox.com>
Signed-off-by: NChristoph Hellwig <hch@lst.de>

If we got a remote invalidation on a bogus rkey, this is a protocol error.
Fail the connection in this case.
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>
Reviewed-by: NMax Gurtovoy <maxg@mellanox.com>
Signed-off-by: NChristoph Hellwig <hch@lst.de>

We must not complete a request before the host memory region is
invalidated.  Luckily we have send with invalidate protocol support so
we usually don't need to execute it, but in case the target did not
invalidate a memory region for us, we must wait for the invalidation to
complete before unmapping host memory and completing the I/O.
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>
Reviewed-by: NMax Gurtovoy <maxg@mellanox.com>
Signed-off-by: NChristoph Hellwig <hch@lst.de>

In order to guarantee that the HCA will never get an access violation
(either from invalidated rkey or from iommu) when retrying a send
operation we must complete a request only when both send completion and
the nvme cqe has arrived. We need to set the send/recv completions flags
atomically because we might have more than a single context accessing the
request concurrently (one is cq irq-poll context and the other is
user-polling used in IOCB_HIPRI).

Only then we are safe to invalidate the rkey (if needed), unmap the host
buffers, and complete the IO.
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>
Reviewed-by: NMax Gurtovoy <maxg@mellanox.com>
Signed-off-by: NChristoph Hellwig <hch@lst.de>

The entire completions suppress mechanism is currently broken because the
HCA might retry a send operation (due to dropped ack) after the nvme
transaction has completed.

In order to handle this, we signal all send completions and introduce a
separate done handler for async events as they will be handled differently
(as they don't include in-capsule data by definition).
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>
Reviewed-by: NMax Gurtovoy <maxg@mellanox.com>
Signed-off-by: NChristoph Hellwig <hch@lst.de>

When the fabrics queue is not alive and fully functional, no commands
should be allowed to pass but connect (which moves the queue to a fully
functional state). Any other command should be failed, with either
temporary status BLK_STS_RESOUCE or permanent status BLK_STS_IOERR.

This is shared across all fabrics, hence move the check to fabrics
library.
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>
Signed-off-by: NChristoph Hellwig <hch@lst.de>

The driver can handle tracking only one AEN request, so this patch
removes handling for multiple ones.
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NJames Smart  <james.smart@broadcom.com>
Signed-off-by: NKeith Busch <keith.busch@intel.com>
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

All the transports were unnecessarilly duplicating the AEN request
accounting. This patch defines everything in one place.
Signed-off-by: NKeith Busch <keith.busch@intel.com>
Reviewed-by: NGuan Junxiong <guanjunxiong@huawei.com>
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

QP object is created using rdma_cm api, therefore the destruction
should use the same api for symmetry.
Signed-off-by: NMax Gurtovoy <maxg@mellanox.com>
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

instead of just queueing delete work
Reported-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>
Signed-off-by: NChristoph Hellwig <hch@lst.de>

No change in behavior except that the FC code cancels two work items a
little later now.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NSagi Grimberg <sagi@grimberg.me>
Reviewed-by: NJames Smart <james.smart@broadcom.com>
Signed-off-by: NChristoph Hellwig <hch@lst.de>

It is only used in two places, and some of the work done by it will
be taken into common code soon.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NSagi Grimberg <sagi@grimberg.me>
Signed-off-by: NChristoph Hellwig <hch@lst.de>

Move the ->delete_work and the associated helpers to common code instead
of duplicating them in every driver.  This also adds the missing reference
get/put for the loop driver.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NSagi Grimberg <sagi@grimberg.me>
Reviewed-by: NJames Smart <james.smart@broadcom.com>
Signed-off-by: NChristoph Hellwig <hch@lst.de>

Adds support for the duplicate_connect option. When set to true,
checks whether there's an existing controller via the same target
address (traddr), target port (trsvcid), and if specified, host
address (host_traddr). Fails the connection request if there is
an existing controller.
Signed-off-by: NJames Smart <james.smart@broadcom.com>
Reviewed-by: NJohannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: NChristoph Hellwig <hch@lst.de>

Instead of allocating a separate struct device for the character device
handle embedd it into struct nvme_ctrl and use it for the main controller
refcounting.  This removes double refcounting and gets us an automatic
reference for the character device operations.  We keep ctrl->device as a
pointer for now to avoid chaning printks all over, but in the future we
could look into message printing helpers that take a controller structure
similar to what other subsystems do.

Note the delete_ctrl operation always already has a reference (either
through sysfs due this change, or because every open file on the
/dev/nvme-fabrics node has a refernece) when it is entered now, so we
don't need to do the unless_zero variant there.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NHannes Reinecke <hare@suse.com>

Signed-off-by: NNitzan Carmi <nitzanc@mellanox.com>
Reviewed-by: NMax Gurtovoy <maxg@mellanox.com>
Signed-off-by: NChristoph Hellwig <hch@lst.de>

Signed-off-by: NMax Gurtovoy <maxg@mellanox.com>
Signed-off-by: NChristoph Hellwig <hch@lst.de>

nvme_rdma_queue_is_ready() fails requests in case a queue is not
LIVE. If the controller is in RECONNECTING state, we might be in
this state for a long time (until we successfully reconnect) and
we are better off with failing the request fast. Otherwise, we
fail with BLK_STS_RESOURCE to have the block layer try again
soon.

In case we are removing the controller when the admin queue
is not LIVE, we will terminate the request with BLK_STS_RESOURCE
but it happens before we call blk_mq_start_request() so the
request timeout never expires, and the queue will never get
back to LIVE (because we are removing the controller). This
causes the removal operation to block infinitly [1].

Thus, if we are removing (state DELETING), and the queue is
not LIVE, we need to fail the request permanently as there is
no chance for it to ever complete successfully.

[1]
--
sysrq: SysRq : Show Blocked State
  task                        PC stack   pid father
kworker/u66:2   D    0   440      2 0x80000000
Workqueue: nvme-wq nvme_rdma_del_ctrl_work [nvme_rdma]
Call Trace:
 __schedule+0x3e9/0xb00
 schedule+0x40/0x90
 schedule_timeout+0x221/0x580
 io_schedule_timeout+0x1e/0x50
 wait_for_completion_io_timeout+0x118/0x180
 blk_execute_rq+0x86/0xc0
 __nvme_submit_sync_cmd+0x89/0xf0
 nvmf_reg_write32+0x4b/0x90 [nvme_fabrics]
 nvme_shutdown_ctrl+0x41/0xe0
 nvme_rdma_shutdown_ctrl+0xca/0xd0 [nvme_rdma]
 nvme_rdma_remove_ctrl+0x2b/0x40 [nvme_rdma]
 nvme_rdma_del_ctrl_work+0x25/0x30 [nvme_rdma]
 process_one_work+0x1fd/0x630
 worker_thread+0x1db/0x3b0
 kthread+0x11e/0x150
 ret_from_fork+0x27/0x40
01              D    0  2868   2862 0x00000000
Call Trace:
 __schedule+0x3e9/0xb00
 schedule+0x40/0x90
 schedule_timeout+0x260/0x580
 wait_for_completion+0x108/0x170
 flush_work+0x1e0/0x270
 nvme_rdma_del_ctrl+0x5a/0x80 [nvme_rdma]
 nvme_sysfs_delete+0x2a/0x40
 dev_attr_store+0x18/0x30
 sysfs_kf_write+0x45/0x60
 kernfs_fop_write+0x124/0x1c0
 __vfs_write+0x28/0x150
 vfs_write+0xc7/0x1b0
 SyS_write+0x49/0xa0
 entry_SYSCALL_64_fastpath+0x18/0xad
--
Reported-by: NBart Van Assche <bart.vanassche@wdc.com>
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>
Signed-off-by: NChristoph Hellwig <hch@lst.de>

We should make sure to escelate allocation failures to prevent a
use-after-free in nvmf_create_ctrl.

Fixes: b28a308e ("nvme-rdma: move tagset allocation to a dedicated routine")
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>
Signed-off-by: NChristoph Hellwig <hch@lst.de>

The fact that we free the async event buffer in
nvme_rdma_destroy_admin_queue can cause us to free it
more than once because this happens in every reconnect
attempt since commit 31fdf184. we rely on the queue
state flags DELETING to avoid this for other resources.

A more complete fix is to not destroy the admin/io queues
unconditionally on every reconnect attempt, but its a bit
more extensive and will go in the next release.

Fixes: 31fdf184 ("nvme-rdma: reuse configure/destroy_admin_queue")
Reported-by: NYi Zhang <yi.zhang@redhat.com>
Reviewed-by: NJohannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>
Signed-off-by: NChristoph Hellwig <hch@lst.de>

Since commit b86dd815
"block: get rid of blk-mq default scheduler choice Kconfig entries",
when setting nr_hw_queues to 1 the admin tag set uses mq-deadline scheduler.
This flag is useful for admin queues that aren't used for normal IO.
Signed-off-by: NIsrael Rukshin <israelr@mellanox.com>
Reviewed-by: NSagi Grimberg <sagi@grimberg.me>
Reviewed-by: NMax Gurtovoy <maxg@mellanox.com>
Signed-off-by: NChristoph Hellwig <hch@lst.de>

If the controller is deleting (in case the user decided to delete it), we
have no point to continue reset sequence.
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>
Reviewed-by: NJohannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: NChristoph Hellwig <hch@lst.de>

Instead of marking we are deleting, mark we are allocated and check that
instead. This makes the logic symmetrical to connected mark check.
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>
Signed-off-by: NChristoph Hellwig <hch@lst.de>

No chance for the local invalidate to succeed if the queue-pair
is in error state. Most likely the target will do a remote
invalidation of our mr so not a big loss on the test_bit.
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>
Signed-off-by: NChristoph Hellwig <hch@lst.de>

Relying on the queue state while tearing down on every reconnect
attempt is not a good design. We should do it once in err_work
and simply try to establish the queues for each reconnect attempt.
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>
Signed-off-by: NChristoph Hellwig <hch@lst.de>

Warn if req->mr is NULL as it should never happen.
Reviewed-by: NJohannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>
Signed-off-by: NChristoph Hellwig <hch@lst.de>

No need for the extra line for trivial assignments.
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>
Reviewed-by: NMax Gurtovoy <maxg@mellanox.com>
Reviewed-by: NJohannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: NChristoph Hellwig <hch@lst.de>

Not necessarily address resolution failed.
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>
Reviewed-by: NMax Gurtovoy <maxg@mellanox.com>
Reviewed-by: NJohannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: NChristoph Hellwig <hch@lst.de>

Instead of flagging admin/io.
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>
Reviewed-by: NMax Gurtovoy <maxg@mellanox.com>
Reviewed-by: NJohannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: NChristoph Hellwig <hch@lst.de>

Move blk_mq_reinit_tagset from blk-mq to nvme core
as the only user of it. Current transports that use
it (rdma, fc) simply implement .reinit_request op.

This patch does not change any functionality.
Reviewed-by: NJens Axboe <axboe@kernel.dk>
Reviewed-by: NMax Gurtovoy <maxg@mellanox.com>
Reviewed-by: NJohannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>
Signed-off-by: NChristoph Hellwig <hch@lst.de>

By calling nvme_stop_ctrl on a already failed controller will wait for the
scan work to complete (only by identify timeout expiration which is 60
seconds). This is unnecessary when we already know that the controller has
failed.
Reported-by: NYi Zhang <yizhan@redhat.com>
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

If we failed to transition to state LIVE after a successful reconnect,
then controller deletion already started. In this case there is no
point moving forward with reconnect.
Reviewed-by: NJohannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

Due to various page sizes in the system (IOMMU/device/kernel), we
set the fabrics controller page size to 4k and block layer boundaries
accordinglly. In architectures that uses different kernel page size
we'll have a mismatch to the MR page size that may cause a mapping error.
Update the MR page size to correspond to the core ctrl settings.
Signed-off-by: NMax Gurtovoy <maxg@mellanox.com>
Reviewed-by: NSagi Grimberg <sagi@grimberg.me>
Signed-off-by: NChristoph Hellwig <hch@lst.de>

This patch slightly improves performance (mainly for small block sizes).
Signed-off-by: NMax Gurtovoy <maxg@mellanox.com>
Signed-off-by: NChristoph Hellwig <hch@lst.de>

To make the nvme_rdma_configure_admin_queue generic in preparation of
moving it to common code.
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>
Signed-off-by: NChristoph Hellwig <hch@lst.de>

No need to queue an extra work to indirect controller removal, just call the
ctrl remove routine.
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>
Signed-off-by: NChristoph Hellwig <hch@lst.de>

This should pair with nvme_rdma_stop_queue.  While this is not a complete
inverse, it still pairs up pretty well because in fabrics we don't have a
disconnect capsule (yet) but we simply teardown the transport association.
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>
Signed-off-by: NChristoph Hellwig <hch@lst.de>

Give it a name symmetric to nvme_rdma_free_queue. Also pass in the ctrl
sqsize+1 and not the opts queue_size.  And suppress a superflous
failure message.
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>
Signed-off-by: NChristoph Hellwig <hch@lst.de>