Change nvme-rdma to use the IB Client API to detect device removal.
This has the wonderful benefit of being able to blow away all the
ib/rdma_cm resources for the device being removed.  No craziness about
not destroying the cm_id handling the event.  No deadlocks due to broken
iw_cm/rdma_cm/iwarp dependencies.  And no need to have a bound cm_id
around during controller recovery/reconnect to catch device removal
events.

We don't use the device_add aspect of the ib_client service since we only
want to create resources for an IB device if we have a target utilizing
that device.
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NSteve Wise <swise@opengridcomputing.com>
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>

When we get a surprise disconnect from the target we queue a periodic
reconnect (which is the sane thing to do...).

We only move the queues out of CONNECTED when we retry to reconnect (after
10 seconds in the default case) but we stop the blk queues immediately
so we are not bothered with traffic from now on. If delete() is kicking
off in this period the queues are still in CONNECTED state.

Part of the delete sequence is trying to issue ctrl shutdown if the
admin queue is CONNECTED (which it is!). This request is issued but
stuck in blk-mq waiting for the queues to start again. This might be
the one preventing us from forward progress...

The patch separates the queue flags to CONNECTED and DELETING. Now we
will move out of CONNECTED as soon as error recovery kicks in (before
stopping the queues) and DELETING is on when we start the queue deletion.
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>

After address resolution, the nvme_rdma_queue rdma resources are
allocated.  If rdma route resolution or the connect fails, or the
controller reconnect times out and gives up, then the rdma resources
need to be freed.  Otherwise, rdma resources are leaked.
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NSagi Grimberg <sagi@grimbrg.me>
Signed-off-by: NSteve Wise <swise@opengridcomputing.com>
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>

Reviewed-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NSagi Grimberg <sagi@grimbrg.me>
Signed-off-by: NSteve Wise <swise@opengridcomputing.com>
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>

Signed-off-by: NSagi Grimberg <sagi@grimberg.me>
Reviewed-by: NChristoph Hellwig <hch@lst.de>

We already have need_inval in ib_mr, lets use
that instead.
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>
Reviewed-by: NChristoph Hellwig <hch@lst.de>

Per NVMe-over-Fabrics 1.0 spec, sqsize is represented as
a 0-based value.

Also per spec, the RDMA binding values shall be set
to sqsize, which makes hsqsize 0-based values.

Thus, the sqsize during NVMf connect() is now:

[root@fedora23-fabrics-host1 for-48]# dmesg
[  318.720645] nvme_fabrics: nvmf_connect_admin_queue(): sqsize for
admin queue: 31
[  318.720884] nvme nvme0: creating 16 I/O queues.
[  318.810114] nvme_fabrics: nvmf_connect_io_queue(): sqsize for i/o
queue: 127

Finally, current interpretation implies hrqsize is 1's based
so set it appropriately.
Reported-by: NDaniel Verkamp <daniel.verkamp@intel.com>
Signed-off-by: NJay Freyensee <james_p_freyensee@linux.intel.com>
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>

Upon admin queue connect(), the rdma qp was being
set based on NVMF_AQ_DEPTH.  However, the fabrics layer was
using the sqsize field value set for I/O queues for the admin
queue, which threw the nvme layer and rdma layer off-whack:

root@fedora23-fabrics-host1 nvmf]# dmesg
[ 3507.798642] nvme_fabrics: nvmf_connect_admin_queue():admin sqsize
being sent is: 128
[ 3507.798858] nvme nvme0: creating 16 I/O queues.
[ 3507.896407] nvme nvme0: new ctrl: NQN "nullside-nqn", addr
192.168.1.3:4420

Thus, to have a different admin queue value, we use
NVMF_AQ_DEPTH for connect() and RDMA private data
as the minimum depth specified in the NVMe-over-Fabrics 1.0 spec
(and in that RDMA private data we treat hrqsize as 1's-based
value, per current understanding of the fabrics spec).
Reported-by: NDaniel Verkamp <daniel.verkamp@intel.com>
Signed-off-by: NJay Freyensee <james_p_freyensee@linux.intel.com>
Reviewed-by: NDaniel Verkamp <daniel.verkamp@intel.com>
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>

ret is not initialized so it contains garbage.  Ensure garbage
is not returned by initializing rc to 0.
Signed-off-by: NColin Ian King <colin.king@canonical.com>
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>

Signed-off-by: NSagi Grimberg <sagi@grimberg.me>
Reviewed-by: NSteve Wise <swise@opengridcomputing.com>

When we reset or reconnect to a controller, we are cancelling the
async event handler so we can safely re-establish resources, but we
need to remember to start it again when we successfully reconnect.
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>
Reviewed-by: NChristoph Hellwig <hch@lst.de>

Relying on ctrl state in nvme_rdma_shutdown_ctrl is wrong because
it will never be NVME_CTRL_LIVE (delete_ctrl or reset_ctrl invoked it).

Instead, check that the admin queue is connected. Note that it is safe
because we can never see a copmeting thread trying to destroy the admin
queue (reset or delete controller).
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>
Reviewed-by: NChristoph Hellwig <hch@lst.de>

If we wait until we free the controller (free_ctrl) we might
lose our rdma device without any notification while we still
have open resources (tags mrs and dma mappings).

Instead, destroy the tags with their rdma resources once we
delete the device and not when freeing it.

Note that we don't do that in nvme_rdma_shutdown_ctrl because
controller reset uses it as well and we want to give active I/O
a chance to complete successfully.
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>
Reviewed-by: NChristoph Hellwig <hch@lst.de>

nvme_uninit_ctrl already does that for us. Note that we reordered
nvme_rdma_shutdown_ctrl and nvme_uninit_ctrl, this is perfectly
fine because we actually want ctrl uninit (aen, scan cancellation
and namespaces removal) to happen before we shutdown the rdma
resources.

Also, centralize the deletion work and the dead controller removal
work code duplication into __nvme_rdma_shutdown_ctrl that accepts
a shutdown boolean.
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>
Reviewed-by: NChristoph Hellwig <hch@lst.de>

Device removal sequence may have crashed because the
controller (and admin queue space) was freed before
we destroyed the admin queue resources. Thus we
want to destroy the admin queue and only then queue
controller deletion and wait for it to complete.

More specifically we:
1. own the controller deletion (make sure we are not
   competing with another deletion).
2. get rid of inflight reconnects if exists (which
   also destroy and create queues).
3. destroy the queue.
4. safely queue controller deletion (and wait for it
   to complete).
Reported-by: NSteve Wise <swise@opengridcomputing.com>
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>
Reviewed-by: NChristoph Hellwig <hch@lst.de>

On an ordered target shutdown, the target can send a AEN on a namespace
removal, this will trigger the host to queue ns-list query. The shutdown
will trigger error recovery which will attepmt periodic reconnect.

We can hit a race where the ns rescanning fails (error recovery kicked
in and we're not connected) causing removing all the namespaces and when
we reconnect we won't see any namespaces for this controller.

So, queue a namespace rescan after we successfully reconnected to the target.
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>
Reviewed-by: NChristoph Hellwig <hch@lst.de>

Zero out the full nvme_rdma_cm_req structure before sending it.
Otherwise we end up leaking kernel memory in the reserved field, which
might break forward compatibility in the future.
Signed-off-by: NRoland Dreier <roland@purestorage.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>

Always use the maximum qp retry count as the
error recovery timeout is dictated from the nvme
keep-alive.
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>
Signed-off-by: NJens Axboe <axboe@fb.com>

PTR_ERR should be applied before its argument is reassigned, otherwise the
return value will be set to 0, not error code.
Signed-off-by: NWei Yongjun <yongjun_wei@trendmicro.com.cn>
Reviewed-by: NJay Freyensee <james_p_freyensee@linux.intel.com>
Signed-off-by: NJens Axboe <axboe@fb.com>

This patch implements the RDMA host (initiator in SCSI speak) driver.  It
can be used to connect to remote NVMe over Fabrics controllers over
Infiniband, RoCE or iWarp, and uses the existing NVMe core driver as well
a the new fabrics library.

To connect to all NVMe over Fabrics controller reachable on a given taget
port using RDMA/CM use the following command:

	nvme connect-all -t rdma -a $IPADDR

This requires the latest version of nvme-cli with Fabrics support.
Signed-off-by: NJay Freyensee <james.p.freyensee@intel.com>
Signed-off-by: NMing Lin <ming.l@ssi.samsung.com>
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NSteve Wise <swise@opengridcomputing.com>
Tested-by: NSteve Wise <swise@opengridcomputing.com>
Signed-off-by: NJens Axboe <axboe@fb.com>