On 5/28/21 3:32 AM, Ming Lei wrote:

On Thu, May 27, 2021 at 05:43:07PM +0200, Hannes Reinecke wrote:

quoted

On 5/25/21 6:03 PM, Douglas Gilbert wrote:

quoted

On 2021-05-21 5:56 p.m., Douglas Gilbert wrote:

quoted

The REQ_HIPRI flag on requests is associated with blk_poll() (aka iopoll)
and assumes the user space (or some higher level) will be calling
blk_poll() on requests marked with REQ_HIPRI and that will lead to their
completion.

In lk 5.13-rc1 the megaraid and scsi_debug LLDs support blk_poll() [seen
by searching for 'mq_poll'] with more to follow, I assume. I have tested
blk_poll() on the scsi_debug driver using both fio and the new sg driver.
It works well with one caveat: as long as there isn't an error.
After fighting with that error processing from the ULD side (i.e. the
new sg driver) and the LLD side I am concluding that the glue that
holds them together, that is, the mid-level is not as REQ_HIPRI aware
as it should be.

Yes REQ_HIPRI is there in scsi_lib.c but it is missing from scsi_error.c
How can scsi_error.c re-issue requests _without_ taking into account
that the original was issued with REQ_HIPRI ? Well I don't know but I'm
pretty sure that is close to the area that I see causing problems
(mainly lockups).

As an example the scsi_debug driver has an in-use bitmap that when a new
request arrives the code looks for an empty slot. Due to (incorrect)
parameter setup that may fail. If the driver returns:
     device_qfull_result = (DID_OK << 16) | SAM_STAT_TASK_SET_FULL;
then I see lock-ups if the request in question has REQ_HIPRI set.

If that is changed to:
     device_qfull_result = (DID_ABORT << 16) | SAM_STAT_TASK_SET_FULL;
then my user space test program sees that error and aborts showing the
TASK SET FULL SCSI status. That is much better than a lockup ...

That's because with the first result the command is requeued (due to
DID_OK), whereas in the latter result the command is aborted (due to
DID_ABORT).

So the question really is whether we should retry the commands which have
REQ_HIPRI set, or whether we shouldn't rather complete them with appropriate
error code.
A bit like enhanced BLOCK_PC requests, if you will.

quoted

quoted

Having played around with variants of the above for a few weeks, I'd
like to throw this problem into the open :-)


Suggestion: perhaps the eh could give up immediately on any request
with REQ_HIPRI set (i.e. make it a higher level layer's problem).

Like I said above: it's not only scsi EH which would need to be modified,
but possibly also the result evaluation in scsi_decide_disposition(); it's
questionable whether a HIPRI command should be requeued at all.

Why can't HIPRI req be requeued?

Oh, it can.
As I said: it's questionable; HIPRI / polled completions are just that,
polling for completions. And a completion indicating a requeue is
_still_ a completion.
So one could argue that we should return here (as it's a completion, and
we're polling for completion).

quoted

But this might even affect the NVMe folks; they do return commands with
BLK_STS_RESOURCE, too.

Block layer will be responsible for re-queueing BLK_STS_RESOURCE requests,
so still not understand why it is one issue for HIPRI req. Also
rq->mq_hctx won't be changed since its allocation, blk_poll()
will keep polling on the correct hw queue for reaping the IO.

As mentioned above, I was talking about completions indicating a requeue.
Requeues due to resource shortage on the initiator side would of course
be requeued.

Cheers,

Hannes
-- 
Dr. Hannes Reinecke		        Kernel Storage Architect
hare@suse.de			               +49 911 74053 688
SUSE Software Solutions Germany GmbH, 90409 Nürnberg
GF: F. Imendörffer, HRB 36809 (AG Nürnberg)