linux.git/drivers/scsi/libiscsi.c, branch v6.0

scsi: libiscsi: Improve conn_send_pdu API

2022-06-22T01:19:23+00:00

The conn_send_pdu API is evil in that it returns a pointer to an
iscsi_task, but that task might have been freed already so you can't touch
it. This patch splits the task allocation and transmission, so functions
like iscsi_send_nopout() can access the task before its sent and do
whatever bookkeeping is needed before it is sent.

Link: https://lore.kernel.org/r/20220616224557.115234-10-michael.christie@oracle.com
Reviewed-by: Lee Duncan 
Signed-off-by: Mike Christie 
Signed-off-by: Martin K. Petersen

scsi: iscsi: Try to avoid taking back_lock in xmit path

2022-06-22T01:19:23+00:00

We need the back lock when freeing a task, so we hold it when calling
__iscsi_put_task() from the completion path to make it easier and to avoid
having to retake it in that path. For iscsi_put_task() we just grabbed it
while also doing the decrement on the refcount but it's only really needed
if the refcount is zero and we free the task. This modifies
iscsi_put_task() to just take the lock when needed then has the xmit path
use it. Normally we will then not take the back lock from the xmit path. It
will only be rare cases where the network is so fast that we get a response
right after we send the header/data.

Link: https://lore.kernel.org/r/20220616224557.115234-9-michael.christie@oracle.com
Reviewed-by: Lee Duncan 
Signed-off-by: Mike Christie 
Signed-off-by: Martin K. Petersen

scsi: iscsi: Remove iscsi_get_task back_lock requirement

2022-06-22T01:19:23+00:00

We currently require that the back_lock is held when calling the functions
that manipulate the iscsi_task refcount. The only reason for this is to
handle races where we are handling SCSI-ml EH callbacks and the cmd is
completing at the same time the normal completion path is running, and we
can't return from the EH callback until the driver has stopped accessing
the cmd. Holding the back_lock while also accessing the task->state made it
simple to check that a cmd is completing and also get/put a refcount at the
same time, and at the time we were not as concerned about performance.

The problem is that we don't want to take the back_lock from the xmit path
for normal I/O since it causes contention with the completion path if the
user has chosen to try and split those paths on different CPUs (in this
case abusing the CPUs and ignoring caching improves perf for some uses).

Begins to remove the back_lock requirement for iscsi_get/put_task by
removing the requirement for the get path. Instead of always holding the
back_lock we detect if something has done the last put and is about to call
iscsi_free_task(). A subsequent commit will then allow iSCSI code to do the
last put on a task and only grab the back_lock if the refcount is now zero
and it's going to call iscsi_free_task().

Link: https://lore.kernel.org/r/20220616224557.115234-8-michael.christie@oracle.com
Reviewed-by: Lee Duncan 
Signed-off-by: Mike Christie 
Signed-off-by: Martin K. Petersen

scsi: iscsi: Remove unneeded task state check

2022-06-22T01:19:23+00:00

Commit 5923d64b7ab6 ("scsi: libiscsi: Drop taskqueuelock") added an extra
task->state because for commit 6f8830f5bbab ("scsi: libiscsi: add lock
around task lists to fix list corruption regression") we didn't know why we
ended up with cmds on the list and thought it might have been a bad target
sending a response while we were still sending the cmd. We were never able
to get a target to send us a response early, because it turns out the bug
was just a race in libiscsi/libiscsi_tcp where:

 1. iscsi_tcp_r2t_rsp() queues a r2t to tcp_task->r2tqueue.

 2. iscsi_tcp_task_xmit() runs iscsi_tcp_get_curr_r2t() and sees we have a
    r2t. It dequeues it and iscsi_tcp_task_xmit() starts to process it.

 3. iscsi_tcp_r2t_rsp() runs iscsi_requeue_task() and puts the task on the
    requeue list.

 4. iscsi_tcp_task_xmit() sends the data for r2t. This is the final chunk
    of data, so the cmd is done.

 5. target sends the response.

 6. On a different CPU from #3, iscsi_complete_task() processes the
    response.  Since there was no common lock for the list, the lists/tasks
    pointers are not fully in sync, so could end up with list corruption.

Since it was just a race on our side, remove the extra check and fix up the
comments.

Link: https://lore.kernel.org/r/20220616224557.115234-7-michael.christie@oracle.com
Reviewed-by: Wu Bo 
Reviewed-by: Lee Duncan 
Signed-off-by: Mike Christie 
Signed-off-by: Martin K. Petersen

scsi: iscsi: Add recv workqueue helpers

2022-06-22T01:19:22+00:00

Add helpers to allow the drivers to run their recv paths from libiscsi's
workqueue.

Link: https://lore.kernel.org/r/20220616224557.115234-3-michael.christie@oracle.com
Reviewed-by: Lee Duncan 
Signed-off-by: Mike Christie 
Signed-off-by: Martin K. Petersen

scsi: iscsi: Rename iscsi_conn_queue_work()

2022-06-22T01:19:22+00:00

Rename iscsi_conn_queue_work() to iscsi_conn_queue_xmit() to reflect that
it handles queueing of xmits only.

Link: https://lore.kernel.org/r/20220616224557.115234-2-michael.christie@oracle.com
Reviewed-by: Lee Duncan 
Reviewed-by: Wu Bo 
Signed-off-by: Mike Christie 
Signed-off-by: Martin K. Petersen

scsi: iscsi: Fix session removal on shutdown

2022-06-22T01:14:54+00:00

When the system is shutting down, iscsid is not running so we will not get
a response to the ISCSI_ERR_INVALID_HOST error event. The system shutdown
will then hang waiting on userspace to remove the session.

This has libiscsi force the destruction of the session from the kernel when
iscsi_host_remove() is called from a driver's shutdown callout.

This fixes a regression added in qedi boot with commit d1f2ce77638d ("scsi:
qedi: Fix host removal with running sessions") which made qedi use the
common session removal function that waits on userspace instead of rolling
its own kernel based removal.

Link: https://lore.kernel.org/r/20220616222738.5722-7-michael.christie@oracle.com
Fixes: d1f2ce77638d ("scsi: qedi: Fix host removal with running sessions")
Tested-by: Nilesh Javali 
Reviewed-by: Lee Duncan 
Reviewed-by: Nilesh Javali 
Signed-off-by: Mike Christie 
Signed-off-by: Martin K. Petersen

scsi: iscsi: Prefer xmit of DataOut over new commands

2022-06-17T01:53:39+00:00

iscsi_data_xmit() (TX worker) is iterating over the queue of new SCSI
commands concurrently with the queue being replenished. Only after the
queue is emptied will we start sending pending DataOut PDUs. That leads to
DataOut timeout on the target side and to connection reinstatement.

Give priority to pending DataOut commands over new commands.

Link: https://lore.kernel.org/r/20220607131953.11584-1-d.bogdanov@yadro.com
Reviewed-by: Konstantin Shelekhin 
Reviewed-by: Mike Christie 
Signed-off-by: Dmitry Bogdanov 
Signed-off-by: Martin K. Petersen

scsi: iscsi: Fix NOP handling during conn recovery

2022-04-12T02:09:35+00:00

If a offload driver doesn't use the xmit workqueue, then when we are doing
ep_disconnect libiscsi can still inject PDUs to the driver. This adds a
check for if the connection is bound before trying to inject PDUs.

Link: https://lore.kernel.org/r/20220408001314.5014-9-michael.christie@oracle.com
Tested-by: Manish Rangankar 
Reviewed-by: Lee Duncan 
Reviewed-by: Chris Leech 
Signed-off-by: Mike Christie 
Signed-off-by: Martin K. Petersen

scsi: iscsi: Merge suspend fields

2022-04-12T02:09:35+00:00

Move the tx and rx suspend fields into one flags field.

Link: https://lore.kernel.org/r/20220408001314.5014-8-michael.christie@oracle.com
Tested-by: Manish Rangankar 
Reviewed-by: Lee Duncan 
Reviewed-by: Chris Leech 
Signed-off-by: Mike Christie 
Signed-off-by: Martin K. Petersen