Re: difficult to pinpoint exhaustion of swap between 4.2.0-rc6 and 4.2.0-rc7

From: Vlastimil Babka
Date: Fri Aug 21 2015 - 07:38:04 EST


On 08/21/2015 11:17 AM, Arthur Marsh wrote:


Vlastimil Babka wrote on 20/08/15 17:46:
On 08/19/2015 05:44 PM, Arthur Marsh wrote:
Hi, I've found that the Linus' git head kernel has had some unwelcome
behaviour where chromium browser would exhaust all swap space in the
course of a few hours. The behaviour appeared before the release of
4.2.0-rc7.

Do you have any more details about the memory/swap usage? Is it really
that chromium process(es) itself eats more memory and starts swapping,
or that something else (a graphics driver?) eats kernel memory, and
chromium as one of the biggest processes is driven to swap by that? Can
you provide e.g. top output with good/bad kernels?

Also what does /proc/meminfo and /proc/zoneinfo look like when it's
swapping?

To see which processes use swap, you can try [1] :
for file in /proc/*/status ; do awk '/VmSwap|Name/{printf $2 " " $3}END{
print ""}' $file; done | sort -k 2 -n -r | less

Thanks

[1] http://www.cyberciti.biz/faq/linux-which-process-is-using-swap/

This does not happen with kernel 4.2.0-rc6.

Sorry for the delay in replying. I had to give an extended run under
kernel 4.2.0-rc6 to obtain comparative results. Both kernels' config
files are attached.

The applications running are the same both times, mainly iceweasel
38.1.0esr-3 and chromium 44.0.2403.107-1.

With the rc7+ kernel but not the rc6 kernel, chromium eventually gets
into a state of consuming lots of swap.

I was able to capture the output requested when running a 4.2.0-rc7+
kernel (Linus' git head as of around 05:00 UTC 19 August 2015) just
before swap was exhausted, forcing me to do a control-alt-delete
shutdown and waiting ages. The kernel config for the rc7+ is attached

The comparison good kernel is from Debian:
Linux am64 4.2.0-rc6-amd64 #1 SMP Debian 4.2~rc6-1~exp1 (2015-08-12)
x86_64 GNU/Linux

Hm I didn't how similar are the configs, was the debian one used as a base for the self-compiled one? Just to rule out config differences... during the bisection you did use the same for compiling a "good" rc6 kernel and "bad" rc7 kernel, right?

That, said, looking at the memory values:

rc6: Free+Buffers+A/I(Anon)+A/I(File)+Slab = 6769MB
rc7: ... = 4714MB

That's 2GB unaccounted for. Which is bad, and yet not enough to explain a full 4GB swap. Another noticeable difference is rc7 using 1560MB ShMem vs 476MB. The rest must be due to more anonymous memory used by the processes. Iceweasel looks unchanged, so I'm guessing the chromiums... the top output probably doesn't give us the whole picture here. I'm still suspecting a graphics driver, which one do you use?

The shmem could be inspected by listing ipcs -m and ipcs -mp and grep grep SYSV /proc/*/maps and figuring out what processes are behind the pids. Doing that for rc6 and rc7 could tell us which processes use the extra 1GB of shmem in rc7.


--
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/