On 12/03/2014 09:49 PM, Linus Torvalds wrote:
> On Wed, Dec 3, 2014 at 7:15 PM, Chris Mason <c...@fb.com> wrote:
>>
>> One guess is that trinity is generating a huge number of tlb
>> invalidations over sparse and horrible ranges. Perhaps the old code was
>> falling back to full tlb flushes before Dave Hansen's string of fixes?
>
> Hmm. I agree that we've had some of the backtraces look like TLB
> flushing might be involved. Not all, though. And I'm not seeing where
> a loop over up to 33 pages should matter over doing a full TLB flush.
>
> What *might* matter is if we somehow get that number wrong, and the loops like
>
> addr = f->flush_start;
> while (addr < f->flush_end) {
> __flush_tlb_single(addr);
> addr += PAGE_SIZE;
> }
>
> ends up looping a *lot* due to some bug, and then the IPI itself would
> take so long that the watchdog could trigger.
>
> But I do not see how that could actually happen. As far as I can tell,
> either the number of pages is limited to less than 33, or we have that
> TLB_FLUSH_ALL case.
>
> Do you see something I don't?
The one thing I _do_ see now is a missed TLB flush is we're flushing one
page at the end of the address space. We'd overflow flush_end back so
flush_end=0:
if (!f->flush_end)
f->flush_end = f->flush_start + PAGE_SIZE; <-- overflow
and we'll never enter the while loop where we actually do the flush:
while (addr < f->flush_end) {
__flush_tlb_single(addr);
addr += PAGE_SIZE;
}
But we have a hole up there on x86_64, so this will never happen in
practice there. It might theoretically apply to 32-bit, but this still
doesn't help with the bug.
Oh, and the tracepoint is spitting out bogus numbers because we need
some parenthesis around the 'nr_pages' calculation.
--
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/
