On 2016.08.25 at 17:40 +0200, David Madore wrote: > TL;DR: Why is Firefox getting OOM-killed while I have 24GB free swap? > > A few days ago I upgraded the kernel on my desktop PC from 4.5.5 to > 4.7.2 and, since then, I've witnessed a huge number of cases where > various processes (typically Firefox) got OOM-killed by the kernel. > Before this kernel upgrade, I had never seen a single OOM event in > normal use; now I've had dozens in couple of days. Nothing has > changed in my config apart from the kernel. Clearly, something has > changed for the worse! > > In fact, this morning, the problem was so bad that I was simply unable > to start Firefox (any attempt to do so would result in it getting > killed immediately), even though I had about 24GB of free swap > available (and the system was idle). I had to kill a few unrelated > processes to be able to launch Firefox; and even then, at some point, > running "sync" in a different window caused the Firefox to be > OOM-killed. > > (Note: the problem IS NOT the choice of the process being killed: > Firefox is a reasonable target. The problem is that the OOM-killer is > being invoked at all, whereas there is plenty of free swap, and, > before this kernel upgrade, all seemed to work perfectly.) > > Unfortunately, all of this is very unreproducible: now it seems to > have disappeared completely, so I can't really test anything any more. > All I can offer is some sample output (below) from the kernel's logs. > Any suggestions as to what I should do if and when the problem returns > is welcome, either to debug or to work around these OOMs. > > The computer in question is an x86_64 (Intel Core 2 Quad Q6600) box > with 8GB RAM and 24GB swap (4*6GB swap across four different disks): I > can of course offer any additional details as to its hardware, kernel > or userland config. > > Sample log output from OOM-killer: > > ### cut after ### > Aug 25 11:46:12 vega kernel: [115461.357412] firefox invoked oom-killer: > gfp_mask=0x27000c0(GFP_KERNEL_ACCOUNT|__GFP_NOTRACK), order=2, oom_score_adj=0
Welcome to the club. See https://lkml.org/lkml/2016/8/22/184 for a possible fix. -- Markus
