Hi Bart, Martin
I'm seeing a number of deadlocks, most of them due to ide__sti
enabling interrupts in a critical section which needs to be protected
against interrupts too.
Another dangerous scenario is the following, from here the usage of
ide__sti becomes questionable.
queue_commands() {
ide__sti();
start_request();
}
...
start_request() {
spin_unlock_irq();
frob_ide();
spin_lock_irq();
}
and also;
if (ch->unmask)
ide__sti(); /* local CPU only */
/* service this interrupt, may set handler for next interrupt */
startstop = handler(drive, drive->rq);
spin_lock_irq(ch->lock);
If someone can explain to me what ide__sti really is trying to achieve
i'd greatly appreciate it.
Regards,
Zwane Mwaikambo
--
function.linuxpower.ca
On Sun, 7 Jul 2002, Zwane Mwaikambo wrote:
> Hi Bart, Martin
> I'm seeing a number of deadlocks, most of them due to ide__sti
> enabling interrupts in a critical section which needs to be protected
> against interrupts too.
I'm seeing blue sky ;-)
Which IDE patch?
> Another dangerous scenario is the following, from here the usage of
> ide__sti becomes questionable.
>
> queue_commands() {
> ide__sti();
> start_request();
> }
> ...
> start_request() {
> spin_unlock_irq();
> frob_ide();
Whats that?
> spin_lock_irq();
> }
>
> and also;
>
> if (ch->unmask)
> ide__sti(); /* local CPU only */
>
> /* service this interrupt, may set handler for next interrupt */
> startstop = handler(drive, drive->rq);
> spin_lock_irq(ch->lock);
>
> If someone can explain to me what ide__sti really is trying to achieve
> i'd greatly appreciate it.
ide_sti() its just __sti() (except atari).
Note that ide_do_request() is called under spin_lock_irqsave(ch->lock,
flags). We have to unlock or we will get deadlock - imagine we are holding
lock and we get irq (shared irq, unexpected one) and we try to lock in
ata_irq_request() -> deadlock.
Also for most code in start_request() we dont need lock, we need it only
for calling block layer helpers, changing/reading IDE_BUSY bit and
ch->handler, timer and drive->rq.
Also we cannot disable interrupts, because we wont know when drive is
interrupting us, missed irqs.
Please also read my previous mails to Alex...
Regards
--
Bartlomiej
> Regards,
> Zwane Mwaikambo
>
> --
> function.linuxpower.ca
>
On Sun, 7 Jul 2002, Bartlomiej Zolnierkiewicz wrote:
> Also for most code in start_request() we dont need lock, we need it only
> for calling block layer helpers, changing/reading IDE_BUSY bit and
> ch->handler, timer and drive->rq.
>
> Also we cannot disable interrupts, because we wont know when drive is
> interrupting us, missed irqs.
Thanks that explains quite a bit.
Cheers,
Zwane Mwaikambo
--
function.linuxpower.ca