On Thu, 2011-04-28 at 18:18 +0100, Mel Gorman wrote:

On Thu, Apr 28, 2011 at 11:56:17AM -0500, James Bottomley wrote:

quoted

On Thu, 2011-04-28 at 11:50 -0500, James Bottomley wrote:

quoted

This is the output of perf record -g -a -f sleep 5

(hopefully the list won't choke)

Um, this one actually shows kswapd

James

---

# Events: 6K cycles
#
# Overhead      Command        Shared Object                                   Symbol
# ........  ...........  ...................  .......................................
#
    20.41%      kswapd0  [kernel.kallsyms]    [k] shrink_slab
                |
                --- shrink_slab
                   |          
                   |--99.91%-- kswapd
                   |          kthread
                   |          kernel_thread_helper
                    --0.09%-- [...]

Ok. I can't see how the patch "mm: vmscan: reclaim order-0 and use
compaction instead of lumpy reclaim" is related unless we are seeing
two problems that happen to manifest in a similar manner.

That is a distinct possibility.

However, there were a number of changes made to dcache in particular
for 2.6.38. Specifically thinks like dentry_kill use trylock and is
happy to loop around if it fails to acquire anything. See things like
this for example;

[ text deleted ]

Way hey, cgroups are also in the mix. How jolly.

Is systemd a common element of the machines hitting this bug by any
chance?

Not in my case, using upstart on my machine.

Colin