On Wed, Sep 07, 2011 at 01:33:55AM +0900, Tejun Heo wrote:

Hello,

On Tue, Sep 06, 2011 at 01:19:44PM +0100, Bruce Stenning wrote:

quoted

ata4: EH complete
Waking error handler thread
scsi_eh_wakeup: succeeded
scsi_schedule_eh: succeeded
scsi_restart_operations: waking up host to restart
Error handler scsi_eh_3 sleeping

I think the following should fix the problem.  The code there is from
the time when libata shared scsi_host->host_lock.  libata no longer
does that so, in the current state, host_eh_scheduled can be cleared
with actual pending EH condition.

Hmmm... maybe not.  Such race condition exists iff host_eh_scheduled
is incremented outside of ap->lock, which I can't see how.  Weird.
Can you please instrument the followings?

* print the caller of scsi_eh_wakeup().  "%pF" w/ (void *)_RET_IP_
  should do it.

* print why scsi_eh is going back to sleep immediately.
  ie. shost->host_failed, host_eh_scheduled, host_busy in
  scsi_error_handler().  It would also be nice to add some printks
  around schedule() and enable PRINTK_TIME.

Thanks.

-- 
tejun