Put the ctrl reference count at nvme_uninit_ctrl as opposed to
nvme_init_ctrl which takes it. This decrease the reference count at the
core layer instead of decreasing it on each transport separately.
Also move the call of nvme_uninit_ctrl at PCI driver after calling to
nvme_release_prp_pools and nvme_dev_unmap, in order to put the reference
count after using the dev. This is safe because those functions use
nvme_dev which is freed only later at nvme_pci_free_ctrl.
Signed-off-by: NIsrael Rukshin <israelr@mellanox.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NKeith Busch <kbusch@kernel.org>

In case nvme_sysfs_delete() is called by the user before taking the ctrl
reference count, the ctrl may be freed during the creation and cause the
bug. Take the reference as soon as the controller is externally visible,
which is done by cdev_device_add() in nvme_init_ctrl(). Also take the
reference count at the core layer instead of taking it on each transport
separately.
Signed-off-by: NIsrael Rukshin <israelr@mellanox.com>
Reviewed-by: NMax Gurtovoy <maxg@mellanox.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NKeith Busch <kbusch@kernel.org>

If we failed to receive data from the socket, don't try
to further process it, we will for sure be handling a queue
error at this point. While no issue was seen with the
current behavior thus far, its safer to cease socket processing
if we detected an error.
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>
Signed-off-by: NKeith Busch <kbusch@kernel.org>

Consolidate the request failure handling code to where
it is being fetched (nvme_tcp_try_send).
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>
Signed-off-by: NKeith Busch <kbusch@kernel.org>

Currently, queue io_cpu assignment is done sequentially for default,
read and poll queues based on queue id. This causes miss-alignment between
context of CPU initiating I/O and the I/O worker thread processing
queued requests or completions.

Change to modify queue io_cpu assignment to take into account queue
maps offset. Each queue io_cpu will start at zero for each queue map.
This essentially aligns read/poll queues to start over the same range as
default queues.

Testing performed by Mark with:
- ram device (nvmet)
- single CPU core (pinned)
- 100% 4k reads
- engine io_uring (not using sq_thread option)
- hipri flag set

Micro-benchmark results show a net gain of:
- increase of 18%-29% in IOPs
- reduction of 16%-22% in average latency
- reduction of 7%-23% in 99.99% latency

Baseline:
========
QDepth/Batch	| IOPs [k]	| Avg. Lat [us]	| 99.99% Lat [us]
-----------------------------------------------------------------
1/1 		| 32.4		| 30.11		| 50.94
32/8		| 179		| 168.20	| 371

CPU alignment:
=============
QDepth/Batch	| IOPs [k]	| Avg. Lat [us]	| 99.99% Lat [us]
-----------------------------------------------------------------
1/1 		| 38.5		|   25.18	| 39.16
32/8		| 231		|   130.75	| 343
Reported-by: NMark Wunderlich <mark.wunderlich@intel.com>
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>
Signed-off-by: NKeith Busch <kbusch@kernel.org>

Enable ability to associate all sockets related to NVMf TCP traffic
to a priority group that will perform optimized network processing for
this traffic class. Maintain initial default behavior of using priority
of zero.
Signed-off-by: NKiran Patil <kiran.patil@intel.com>
Signed-off-by: NMark Wunderlich <mark.wunderlich@intel.com>
Reviewed-by: NSagi Grimberg <sagi@grimberg.me>
Signed-off-by: NKeith Busch <kbusch@kernel.org>

Delayed keep alive work is queued on system workqueue and may be cancelled
via nvme_stop_keep_alive from nvme_reset_wq, nvme_fc_wq or nvme_wq.

Check_flush_dependency detects mismatched attributes between the work-queue
context used to cancel the keep alive work and system-wq. Specifically
system-wq does not have the WQ_MEM_RECLAIM flag, whereas the contexts used
to cancel keep alive work have WQ_MEM_RECLAIM flag.

Example warning:

  workqueue: WQ_MEM_RECLAIM nvme-reset-wq:nvme_fc_reset_ctrl_work [nvme_fc]
	is flushing !WQ_MEM_RECLAIM events:nvme_keep_alive_work [nvme_core]

To avoid the flags mismatch, delayed keep alive work is queued on nvme_wq.

However this creates a secondary concern where work and a request to cancel
that work may be in the same work queue - namely err_work in the rdma and
tcp transports, which will want to flush/cancel the keep alive work which
will now be on nvme_wq.

After reviewing the transports, it looks like err_work can be moved to
nvme_reset_wq. In fact that aligns them better with transition into
RESETTING and performing related reset work in nvme_reset_wq.

Change nvme-rdma and nvme-tcp to perform err_work in nvme_reset_wq.
Signed-off-by: NNigel Kirkland <nigel.kirkland@broadcom.com>
Signed-off-by: NJames Smart <jsmart2021@gmail.com>
Reviewed-by: NSagi Grimberg <sagi@grimberg.me>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NKeith Busch <kbusch@kernel.org>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

When nvme_tcp_io_work() fails to send to socket due to
connection close/reset, error_recovery work is triggered
from nvme_tcp_state_change() socket callback.
This cancels all the active requests in the tagset,
which requeues them.

The failed request, however, was ended and thus requeued
individually as well unless send returned -EPIPE.
Another return code to be treated the same way is -ECONNRESET.

Double requeue caused BUG_ON(blk_queued_rq(rq))
in blk_mq_requeue_request() from either the individual requeue
of the failed request or the bulk requeue from
blk_mq_tagset_busy_iter(, nvme_cancel_request, );
Signed-off-by: NAnton Eidelman <anton@lightbitslabs.com>
Reviewed-by: NSagi Grimberg <sagi@grimberg.me>
Signed-off-by: NKeith Busch <kbusch@kernel.org>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

This function improves code readability and reduces code duplication.
Signed-off-by: NIsrael Rukshin <israelr@mellanox.com>
Signed-off-by: NMax Gurtovoy <maxg@mellanox.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NKeith Busch <kbusch@kernel.org>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

Busy polling usually runs without locks.
Let's use skb_queue_empty_lockless() instead of skb_queue_empty()

Also uses READ_ONCE() in __skb_try_recv_datagram() to address
a similar potential problem.
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

During nvme_tcp_setup_cmd_pdu error flow, one must call nvme_cleanup_cmd
since it's symmetric to nvme_setup_cmd.
Signed-off-by: NMax Gurtovoy <maxg@mellanox.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NKeith Busch <kbusch@kernel.org>

The access to sk->sk_ll_usec should be hidden behind
CONFIG_NET_RX_BUSY_POLL like the definition of sk_ll_usec.

Put access to ->sk_ll_usec behind CONFIG_NET_RX_BUSY_POLL.

Fixes: 1a9460ce ("nvme-tcp: support simple polling")
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NSebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: NKeith Busch <kbusch@kernel.org>

A controller in the resetting state has not yet completed its recovery
actions. The pci and fc transports were already handling this, so update
the remaining transports to not attempt additional recovery in this
state. Instead, just restart the request timer.
Tested-by: NEdmund Nadolski <edmund.nadolski@intel.com>
Reviewed-by: NJames Smart <james.smart@broadcom.com>
Reviewed-by: NSagi Grimberg <sagi@grimberg.me>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NKeith Busch <kbusch@kernel.org>

Allow the do/while statement to continue if current time
is not after the proposed time 'deadline'. Intent is to
allow loop to proceed for a specific time period. Currently
the loop, as coded, will exit after first pass.
Signed-off-by: NMark Wunderlich <mark.wunderlich@intel.com>
Reviewed-by: NSagi Grimberg <sagi@grimberg.me>
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>

The variable ret is being initialized with a value that is never read
and is being re-assigned immediately afterwards. The assignment is
redundant and hence can be removed.

Addresses-Coverity: ("Unused value")
Signed-off-by: NColin Ian King <colin.king@canonical.com>
Reviewed-by: NMax Gurtovoy <maxg@mellanox.com>
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>

This is a more appropriate error status for a transport error
detected by us (the host).
Reviewed-by: NHannes Reinecke <hare@suse.com>
Reviewed-by: NJames Smart <james.smart@broadcom.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NChaitanya Kulkarni <chaitanya.kulkarni@wdc.com>
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>

We have a fundamental issue that fabric commands use the admin_q.
The reason is, that admin-connect, register reads and writes and
admin commands cannot be guaranteed ordering while we are running
controller resets.

For example, when we reset a controller we perform:
1. disable the controller
2. teardown the admin queue
3. re-establish the admin queue
4. enable the controller

In order to perform (3), we need to unquiesce the admin queue, however
we may have some admin commands that are already pending on the
quiesced admin_q and will immediate execute when we unquiesce it before
we execute (4). The host must not send admin commands to the controller
before enabling the controller.

To fix this, we have the fabric commands (admin connect and property
get/set, but not I/O queue connect) use a separate fabrics_q and make
sure to quiesce the admin_q before we disable the controller, and
unquiesce it only after we enable the controller.

This fixes the error prints from nvmet in a controller reset storm test:
kernel: nvmet: got cmd 6 while CC.EN == 0 on qid = 0
Which indicate that the host is sending an admin command when the
controller is not enabled.
Reviewed-by: NJames Smart <james.smart@broadcom.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>

TOS provide clients the ability to segregate traffic flows for
different type of data.
One of the TOS usage is bandwidth management which allows setting bandwidth
limits for QoS classes, e.g. 80% bandwidth to controllers at QoS class A
and 20% to controllers at QoS class B.

usage examples:
nvme connect --tos=0 --transport=tcp --traddr=10.0.1.1 --nqn=test-nvme
Signed-off-by: NIsrael Rukshin <israelr@mellanox.com>
Reviewed-by: NMax Gurtovoy <maxg@mellanox.com>
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>

This patch doesn't change any functionality.
Signed-off-by: NIsrael Rukshin <israelr@mellanox.com>
Reviewed-by: NMax Gurtovoy <maxg@mellanox.com>
Reviewed-by: NSagi Grimberg <sagi@grimberg.me>
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>

Simple polling support via socket busy_poll interface.
Although we do not shutdown interrupts but simply hammer
the socket poll, we can sometimes find completions faster
than the normal interrupt driven RX path.

We add per queue nr_cqe counter that resets every time
RX path is invoked such that .poll callback can return it
to stay consistent with the semantics.
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>

All seem to call it with ctrl->cap so no need to pass it
at all.
Reviewed-by: NMinwoo Im <minwoo.im.dev@gmail.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>

nvme_enable_ctrl reads the cap register right after, so
no need to do that locally in the transport driver. Have
sqsize setting in nvme_init_identify.
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>

Using socket specific read_sock() calls instead of directly calling
tcp_read_sock() helps lld module registered handlers if any, to be called
from nvme-tcp host.
This patch therefore replaces the tcp_read_sock() with socket specific
prot_ops.
Signed-off-by: NPotnuri Bharat Teja <bharat@chelsio.com>
Acked-by: NSagi Grimberg <sagi@grimberg.me>
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>

Can return directly in the switch statement
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>

When aborting in-flight request for recovering controller, we have
to make sure that queue's complete function is called on completed
request before moving on. Otherwise, for example, the warning of
WARN_ON_ONCE(qp->mrs_used > 0) in ib_destroy_qp_user() may be
triggered on nvme-rdma.

Fix this issue by using blk_mq_tagset_wait_completed_request.

Cc: Max Gurtovoy <maxg@mellanox.com>
Cc: Sagi Grimberg <sagi@grimberg.me>
Cc: Keith Busch <keith.busch@intel.com>
Cc: Christoph Hellwig <hch@lst.de>
Reviewed-by: NSagi Grimberg <sagi@grimberg.me>
Signed-off-by: NMing Lei <ming.lei@redhat.com>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

According to commit a10674bf ("tcp: detecting the misuse of
.sendpage for Slab objects") and previous discussion, tcp_sendpage
should not be used for pages that is managed by SLAB, as SLAB is not
taking page reference counters into consideration.
Signed-off-by: NMikhail Skorzhinskii <mskorzhinskiy@solarflare.com>
Reviewed-by: NSagi Grimberg <sagi@grimberg.me>
Signed-off-by: NChristoph Hellwig <hch@lst.de>

When the controller supports less queues than requested, we
should make sure that queue mapping does the right thing and
not assume that all queues are available. This fixes a crash
when the controller supports less queues than requested.

The rules are:
1. if no write queues are requested, we assign the available queues
   to the default queue map. The default and read queue maps share the
   existing queues.
2. if write queues are requested:
  - first make sure that read queue map gets the requested
    nr_io_queues count
  - then grant the default queue map the minimum between the requested
    nr_write_queues and the remaining queues. If there are no available
    queues to dedicate to the default queue map, fallback to (1) and
    share all the queues in the existing queue map.

Also, provide a log indication on how we constructed the different
queue maps.
Reported-by: NHarris, James R <james.r.harris@intel.com>
Tested-by: NJim Harris <james.r.harris@intel.com>
Cc: <stable@vger.kernel.org> # v5.0+
Suggested-by: NRoy Shterman <roys@lightbitslabs.com>
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>

If I/O queue connect times out, we might have freed the queue socket
already, so check for that on the error path in nvme_tcp_start_queue.
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>
Signed-off-by: NChristoph Hellwig <hch@lst.de>

usually nvme_ prefix is for core functions.
While we're cleaning up, remove redundant empty lines
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>
Reviewed-by: NMinwoo Im <minwoo.im@samsung.com>
Signed-off-by: NChristoph Hellwig <hch@lst.de>

If we timeout the admin startup sequence we might not yet have
an I/O tagset allocated which causes the teardown sequence to crash.
Make nvme_tcp_teardown_io_queues safe by not iterating inflight tags
if the tagset wasn't allocated.

Fixes: 39d57757 ("nvme-tcp: fix timeout handler")
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>
Signed-off-by: NChristoph Hellwig <hch@lst.de>

nvme_tcp_end_request just takes the status value and the converts
it to little endian as well as shifting for the phase bit.

Fixes: 43ce38a6d823 ("nvme-tcp: support C2HData with SUCCESS flag")
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NSagi Grimberg <sagi@grimberg.me>

A C2HData PDU with the SUCCESS flag set indicates that the I/O was
completed by the controller successfully and means that a subsequent
completion response capsule PDU will be ommitted.

If we see this flag, fisrt we check that LAST_PDU flag is set as well,
and then we complete the request when the data transfer (and data digest
verification if its on) is done.

While we're at it, reuse a bit of code with nvme_fail_request.
Reported-by: NSteve Blightman <steve.blightman@oracle.com>
Suggested-by: NOliver Smith-Denny <osmithde@cisco.com>
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>
Reviewed-by: NOliver Smith-Denny <osmithde@cisco.com>
Tested-by: NOliver Smith-Denny <osmithde@cisco.com>
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

It is used now just to flush error recovery and reconnect work items in
the RDMA and TCP transports, which can simply be moved to the
corresponding teardown routines.
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>
Signed-off-by: NChristoph Hellwig <hch@lst.de>

Currently, we have several problems with the timeout
handler:
1. If we timeout on the controller establishment flow, we will hang
because we don't execute the error recovery (and we shouldn't because
the create_ctrl flow needs to fail and cleanup on its own)
2. We might also hang if we get a disconnet on a queue while the
controller is already deleting. This racy flow can cause the controller
disable/shutdown admin command to hang.

We cannot complete a timed out request from the timeout handler without
mutual exclusion from the teardown flow (e.g. nvme_rdma_error_recovery_work).
So we serialize it in the timeout handler and teardown io and admin
queues to guarantee that no one races with us from completing the
request.
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

For sure we are a fabric driver.
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>
Signed-off-by: NChristoph Hellwig <hch@lst.de>

We should never touch the opal device from the transport driver.
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>
Signed-off-by: NChristoph Hellwig <hch@lst.de>

Preparation for polling support for fabrics. Polling support
means that our completion queues are not generating any interrupts
which means we need to poll for the nvmf io queue connect as well.

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

There is a spelling mistake in a dev_info message, fix it.
Signed-off-by: NColin Ian King <colin.king@canonical.com>
Reviewed-by: NSagi Grimberg <sagi@grimberg.me>
Signed-off-by: NChristoph Hellwig <hch@lst.de>

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

Allow NVMF_OPT_NR_WRITE_QUEUES to describe additional write queues.  In
addition, implement .map_queues that will apply 2 queue maps for read
and write queue sets.

Note that with the separate queue map, HCTX_TYPE_READ will always use
nr_io_queues and HCTX_TYPE_DEFAULT will use nr_write_queues.
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>
Signed-off-by: NChristoph Hellwig <hch@lst.de>