On 09/12/2021 13:52, Kashyap Desai wrote:

quoted

+ scsi mailing list

quoted

On Mon, 18 Oct 2021 17:41:23 +0800, John Garry wrote:

quoted

Since it is now possible for a tagset to share a single set of tags,
the iter function should not re-iter the tags for the count of #hw
queues in that case. Rather it should just iter once.

John - Recently we found issue of error hander thread never kicked off
and this patch fix the issue.
Without this patch, scsi error hander will not find correct host_busy
counter.

Take one simple case. There is one IO outstanding and that is getting
timedout.
Now SML wants to wake up EH thread only if, below condition met
"scsi_host_busy(shost) == shost->host_failed"

Without this patch, shared host tag enabled meagaraid_sas driver will
find host_busy = actual outstanding * nr_hw_queues.
Error handler thread will never be kicked-off.

This patch is mandatory for fixing shared host tag feature and require
to be part of stable kernel.

Do you need more data for posting to stable kernel ?

To be clear, are you saying that you see the issue which patch "blk-mq:
Fix blk_mq_tagset_busy_iter() for shared tags" fixes before v5.16-rc?

This patch (now commit 0994c64eb415) and the commit which it is supposed
to fix, e155b0c238b2, will only be in v5.16, so I don't see anything which
is
needed in stable.

Hi John

Yes. No need of posting this to stable.  There is still an issue which we
are tracking. It is not always reproducible. I am injecting artificial Task
abort on my setup to reproduce it.
It happens on rhel8.5 most of the time. It is a timing issue so thinking of
reproducing on other kernel as well.
I am suspecting issue might be due to  missing commit -
67f3b2f822b7e71cfc9b42dbd9f3144fa2933e0b of  [PATCH] blk-mq: avoid to
iterate over stale request

Whenever I notice the issue, there was a symptoms that host_busy is getting
counted for each hctx individually. Let me collect more data and I will
start another thread.

Kashyap

Thanks,
John