Hello,

On Wed, Sep 07, 2011 at 01:09:10PM +0100, Bruce Stenning wrote:

Sorry for sending so many emails yesterday; I blame the dental anaesthetic
I received in the morning for being so jumpy on the send button ;-)

Oh the fun. :)

I can certainly try this.  Could you confirm whether my thoughts about a race
between the scsi_eh thread and the wake-up are plausible?  I backtracked
yesterday because I thought the scsi_eh thread would get rescheduled naturally,
not realising that when the task state is TASK_INTERRUPTIBLE schedule() takes
the task off the run queue (so it needs to be explicitly woken.)

Here is my thinking again:

shost->host_eh_scheduled is read here in scsi_error_handler:

        set_current_state(TASK_INTERRUPTIBLE);
        while (!kthread_should_stop()) {
                if ((shost->host_failed == 0 && shost->host_eh_scheduled == 0) ||

There's no locking in scsi_error_handler (though functions it calls probably
claim locks.)

When scheduling an EH, scsi_schedule_eh takes the shost->host_lock, increments
shost->host_eh_scheduled, and then wakes the EH thread.  If this happens
between the scsi_eh thread reading host_eh_scheduled and sending itself back
to sleep (when the scsi_eh thread's state is TASK_INTERRUPTIBLE) nothing will
wake up the thread again and host_eh_scheduled will not get inspected.
host_eh_scheduled is stuck at 1 with the scsi_eh thread asleep, and it won't
get woken again because the ata port has been frozen and irqs are masked off.

I don't think there's a race condition there.  set_current_state()
implies memory barrier and wake_up_process() implies wmb().  host_eh
either sees the inrecremented eh_scheduled count or TASK_RUNNING set
by wake_up_process(), so it can't miss an event.

Thanks.

-- 
tejun