lkml.org 
[lkml]   [2005]   [Nov]   [19]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: shrinker->nr = LONG_MAX means deadlock for icache
On Sat, Nov 19, 2005 at 03:03:06AM -0800, Andrew Morton wrote:
> It would be nice to understand exactly what's gone wrong.

I found something more, see below.

> I guess so, although I worry that this way we'll obscure the real bug,
> whatever it is.

Now that I understand better the math around scanned and lru_pages I
believe their caller could be the reason they have this huge number in
"nr" is because they pass 0 to shrink all slabs entries. As said in the
previous email they lockup when invoking the slab shrinking with the
toss-cache feature. They should have passed "tossed" as third parameter
too, not 0.

int tossed = atomic_read(&npgs_tossed);
shrink_slab(tossed, GFP_KERNEL, 0 /* shrink max */);
atomic_set(&npgs_tossed, 0);
The zero as thrid parameter means nr will be "max_pass * scanned", so if
both the page-lru is huge and the icache is huge, that can lead to an
huge value.

They should also add a WARN_ON to be sure that "tossed" is never
negative just in case: when the "tossed" gets sign zero extended during
the int2unsigned-long conversion, that could generate the huge number if
tossed was negative.

So the caller has to be fixed too, even if now it would be ok to pass 0
without risking huge nr values (after fixing the unrelated __GFP_IO bug).

So hopefully the "0" as third parameter is good enough to explain the
(other) real bug and we won't be hiding more bugs with this fix.

> Sure. You've limited the number of scanned objects in one pass to twice
> the number of objects - there's no point in doing more work than that.

Agreed.

> A return value of 3 is very odd. I'd be suspecting a mismeasurement.
> Unless someone had altered vfs_cache_pressure.

Exactly.

> OK. Well If Edward&co could do a bit more investigation it'd be great -
> meanwhile I'll hang onto this (and might add some mm-only debugging,
> depending on how Edward gets on):

Looks good to me, thanks!
-
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/

\
 
 \ /
  Last update: 2005-11-19 12:41    [from the cache]
©2003-2011 Jasper Spaans