Re: frequent lockups in 3.18rc4

[Chris Mason]

On Thu, Dec 4, 2014 at 10:22 AM, Dave Hansen <dave.hansen@xxxxxxxxx> 
wrote:
> On 12/03/2014 09:49 PM, Linus Torvalds wrote:
> >  On Wed, Dec 3, 2014 at 7:15 PM, Chris Mason <clm@xxxxxx> 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.

Yeah, I didn't see any problems with your changes, but I was hoping 
that even a small change like doing 33 flushes at a time was pushing 
Dave's box just over the line.

-chris



--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/