Skip to contentSkip to conversation

sbp2.c on SMP

DISCUSSION Douglas Gilbert 1 participant 1 message Nov 12, 2001

lore
Share & tools
lore
#0 lore raw
Quoted 21 lines · Andrew Morton

drivers/ieee1394/sbp2.c deadlocks immediately on SMP, because io_request_lock is not held over its call to scsi_old_done().

I don't know why scsi_old_done() actually requires io_request_lock; perhaps Jens could comment on whether I've taken the lock in the appropriate place?

With the appended patch I can confirm that the driver happily runs
`dbench 40' for half an hour on dual x86. Even when you kick the
disk onto the floor (sorry, HJ).

The games which scsi_old_done() plays with spinlocks and interrupt enabling are really foul. If someone calls it with interrupts disabled (sbp2 does this) then scsi_old_done() will enable interrupts for you. If, for example, you call scsi_old_done() under spin_lock_irqsave(), the reenabling of interrupts will expose you to deadlocks. Perhaps scsi_old_done() should just use spin_unlock()/spin_lock()?

I tried enabling SBP2_USE_REAL_SPINLOCKS in sbp2.c and that appears to work just fine, although I haven't left that change in place here.

Quoted 7 lines

You don't actually _need_ SMP hardware to test SMP code, BTW. You can just build an SMP kernel and run that on a uniprocessor box. This will still catch a wide range of bugs - it certainly detects the lockup which was occurring because of the scsi_old_done() thing.

Incidentally, it would be nice to be able to get this driver working properly when linked into the kernel - it makes debugging much easier :)

It would also be useful to replace the old style scsi error handling with the improved "eh" error handling. The old style was deprecated in the lk 2.2 series but lives on.

drivers/ieee1394/sbp2.c Sun Nov 11 17:22:47 2001 static void sbp2scsi_complete_command(st · line 2775

Quoted 13 lines
drivers/ieee1394/sbp2.c Sun Nov 11 17:22:47 2001 renamed from drivers/ieee1394/sbp2.c Wed Oct 17 14:19:20 2001 +2 -0
@@ -2767,7 +2767,9 @@ static void sbp2scsi_complete_command(st
        /*
         * Tell scsi stack that we're done with this command
         */
+       spin_lock_irq(&io_request_lock);
        done (SCpnt);
+       spin_unlock_irq(&io_request_lock);
 
        return;
 }

In the meantime, this patch is obviously required.

Doug Gilbert