On Thu, May 21, 2015 at 10:08 AM, Borislav Petkov <b...@suse.de> wrote:
> On Thu, May 21, 2015 at 09:45:10AM -0700, Andy Lutomirski wrote:
>> I must be missing something. In this sequence, you're sleeping with
>> IF=0 and ECX[0] = 0, so an IRQ won't get handled. Don't we want
>> ECX[0] = 1?
>
> Hmm, so actually we don't want to sleep with interrupts disabled. If
> ECX[0]=1b, then an interrupt will wake MWAIT. So then you have to do the
> loop thing as tglx suggested.
>
>> > The NMI argument is a problem though - if and NMI gets you out of
>> > MWAITX, a simple perf tool workload would kill all MWAITX executions.
>> > Which is bad. :-\
>>
>> I'm not sure it's a show-stopper. NMI handlers are meant to be fast.
>> If an NMI comes in between rdtsc and mwaitx, then we oversleep, but by
>> at most the time it takes to handle an NMI, and nothing would have
>> stopped us from oversleeping that long if an NMI came in right after
>> mwaitx returned.
>
> Actually, I'm thinking about an NMI happening after we've issued MWAIT.
> NMIs wake it up. So you have the same problem as above:
>
> NMIs will wake MWAIT so you'd need to check how long you've slept and
> sleep for the remaining time. I.e., something like that thing from a
> couple of mails ago:
>
> delay = usec_to_tsc(delay_usec);
>
> if (delay > ((1 << 32) - 1)) {
> mdelay(delay_usec);
> return;
> }
>
> end = rdtsc() + delay;
> while (1) {
>
> monitorx( ...); /* Do we need that here? */
> mwaitx(delay);
>
> /* possible wakeups */
>
> now = rdtsc();
> if (end <= now)
> break;
> delay = end - now;
> }
>
>
> Yes, no?
Yes, but there should already be an adequate outer loop around the
whole thing. After all, even regular mwait can have spurious wakeups
due to monitor monitoring the entire cache line.
--Andy
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
