I mentioned in the thread "blk-mq: improvement on handling IO during CPU
hotplug" that I was using this series to test that patchset.
So just with this patchset (and without yours), I get what looks like some
IO errors in the LLDD. The error is an underflow error. I can't figure out
what is the cause.
Hi Ming,
Can you post the error log? Or interpret the 'underflow error' from hisi
sas or scsi viewpoint?
The check here fails:
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/drivers/scsi/libsas/sas_scsi_host.c?h=v5.4-rc7#n57
Indeed, no data is received.
I'm wondering if the SCSI command is getting corrupted someway.
Why do you think the command is corrupted?
I considered that the underflow may occur if we were to clobber a SCSI
command/request from another hctx and zero some fields, which is
detected as an underflow. But that's just guessing.
However do I find if I set shost->can_queue = HISI_SAS_MAX_COMMANDS /
#queues, then no issue. But maybe that's a coincidence. For this, total
queue depth = HISI_SAS_MAX_COMMANDS. I don't see the impact of that.
I need to test that more.
+ if (expose_mq_experimental) {
+ shost->can_queue = HISI_SAS_MAX_COMMANDS;
+ shost->cmd_per_lun = HISI_SAS_MAX_COMMANDS;
The above is contradictory with current 'nr_hw_queues''s meaning,
see commit on Scsi_Host.nr_hw_queues.
Right, so I am generating the hostwide tag in the LLDD. And the Scsi
host-wide host_busy counter should ensure that we don't pump too much IO to
the HBA.
Even without the host-wide host_busy, your approach should work if you
build the hisi sas tag correctly(uniquely), just not efficiently.
Yes, I do that.
I'd
suggest you to collect trace and observe if request with expected hisi sas
tag is sent to hardware.
I can add some debug for that. What trace do you mean?
BTW, the patch of 'scsi: core: avoid host-wide host_busy counter for scsi_mq'
will be merged to v5.5 if everything is fine.
https://git.kernel.org/pub/scm/linux/kernel/git/mkp/scsi.git/commit/?h=5.5/scsi-queue&id=6eb045e092efefafc6687409a6fa6d1dabf0fb69
Yeah, it seems a good change.
Thanks,
John
