On 11/14/2014 04:06 PM, Rik van Riel wrote:
On 11/14/2014 08:08 AM, Michal Hocko wrote:
Hi,
back in 2012 [1] there was a discussion about a forking load which
accumulates anon_vmas. There was a trivial test case which triggers this
and can potentially deplete the memory by local user.
We have a report for an older enterprise distribution where nsd is
suffering from this issue most probably (I haven't debugged it throughly
but accumulating anon_vma structs over time sounds like a good enough
fit) and has to be restarted after some time to release the accumulated
anon_vma objects.
There was a patch which tried to work around the issue [2] but I do not
see any follow ups nor any indication that the issue would be addressed
in other way.
The test program from [1] was running for around 39 mins on my laptop
and here is the result:
$ date +%s; grep anon_vma /proc/slabinfo
1415960225
anon_vma 11664 11900 160 25 1 : tunables 0 0 0 :
slabdata 476 476 0
$ ./a # The reproducer
$ date +%s; grep anon_vma /proc/slabinfo
1415962592
anon_vma 34875 34875 160 25 1 : tunables 0 0 0 :
slabdata 1395 1395 0
$ killall a
$ date +%s; grep anon_vma /proc/slabinfo
1415962607
anon_vma 11277 12175 160 25 1 : tunables 0 0 0 :
slabdata 487 487 0
So we have accumulated 23211 objects over that time period before the
offender was killed which released all of them.
The proposed workaround is kind of ugly but do people have a better idea
than reference counting? If not should we merge it?
I believe we should just merge that patch.
I have not seen any better ideas come by.
I have some very vague idea that if we could distinguish (with a flag?)
anon_vma_chain (avc) pointing to parent's anon_vma, from avc's created
for new anon_vma's in the child, we could maybe detect at "child-type"
avc removal time, that the only avc's left for a non-root anon_vma are
those of "parent-type" pointing from children. Then we could go through
all pages that map the anon_vma, and change their mapping to the root
anon_vma. The root would have to stay, orphaned or not, because of the
lock there.
That would remove the need for determining a magic constant and the
possibility that we still leave non-useful "orphaned" anon_vma's on the
top levels of the fork hierarchy, while all the bottom levels have to
share the last anon_vma's that were allowed to be created. I'm not sure
if that's the case of nsd - if besides the "orphaned parent" forks it
also forks some workers that would no longer benefit from having their
private anon_vma's.
Of course the downside is that the idea would be too complicated wrt
locking and incur overhead on some fast paths (process exit?). And I
admit I'm not very familiar with the code (which is perhaps euphemism :)
Still, what do you think, Rik?
Vlastimil
The comment should probably be fixed to reflect the
chain length of 5 though :)
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
