On Wed, 2003-02-05 at 17:56, Ross Biro wrote:
> Benjamin Herrenschmidt wrote:
>
> >>Okay, I had to watch for it a bit longer and it turns out that the kernel PDC driver has a problem in this shared interrupt setup. When loads get high it seems to run into some timing problem which causes things like:
> >>
> >>Feb 4 01:02:22 admin kernel: hde: dma_intr: status=0xd0 { Busy }
> >>
> >>
> >>
> Since the busy bit is set, we know the drive must have received a
> command. Since dma_intr thought the drive was not busy, an interrupt
> must have snuck through between the command being issued and the dma
> being started. I think in my original patch, I had the dma start
> outside of the spinlock, that is a bug. The command to the controller
> to start the dma must be inside of the spinlock.
While I agree with you here, I don't think it's what's happening.
In ide-disk, do_rw_disk sets up the taskfile, then basically calls
hwif->ide_dma_read/write to start the command.
In ide-dma.c in 2.4.21-pre4, what happens is:
/* PRD table */
hwif->OUTL(hwif->dmatable_dma, hwif->dma_prdtable);
/* specify r/w */
hwif->OUTB(reading, hwif->dma_command);
/* read dma_status for INTR & ERROR flags */
dma_stat = hwif->INB(hwif->dma_status);
/* clear INTR & ERROR flags */
hwif->OUTB(dma_stat|6, hwif->dma_status);
drive->waiting_for_dma = 1;
if (drive->media != ide_disk)
return 0;
.../...
Then issue command byte.
Below we clear the DMA status _and_ set waiting_for_dma to 1.
That means that if an IRQ sneaks in, we will call
drive_is_ready(), which shouldn't return INTR 1 since we
just cleared it. I don't see how a race could happen here,
but I might have missed something.
Even if, on SMP, the code below executes _simultaneously_
with ide_intr, the later will check for handler beeing
non-NULL before checking waiting_for_dma (drive_is_ready),
and thus will not race since we set the handler after.
The only thing I see is a possible wraparound of
waiting_for_dma. It's an u8, so it wraps at 255. However,
it's incremented in each __ide_dma_test_irq call. So if you
get more than 255 shared (network in your case) interrupts
before the end of the command, you die.
Alan: you can remove safely the waiting_for_dma++, I beleive,
in drive_is_ready(). I don't know how that code sneaked in
ide-dma. I indeed do that in ppc/pmac.c for other reasons
(sort of timeout condition on the DMA controller that happens
when I get an initial error), but this is totally unrelated
HW on which I know I have no shared IRQ.
Stephan: Can you try editing ide-dma.c, function
__ide_dma_test_irq(), and remove that line:
- drive->waiting_for_dma++;
And tell us if it helps in any way.
Ben.
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
This archive was generated by hypermail 2b29 : Fri Feb 07 2003 - 22:00:17 EST