On Tue 28-02-17 09:59:35, Robert Kudyba wrote: > > > On Feb 28, 2017, at 9:40 AM, Michal Hocko <[email protected]> wrote: > > > > On Tue 28-02-17 09:33:49, Robert Kudyba wrote: > >> > >>> On Feb 28, 2017, at 9:15 AM, Michal Hocko <[email protected]> wrote: > >>> and this one is hitting the min watermark while there is not really > >>> much to reclaim. Only the page cache which might be pinned and not > >>> reclaimable from this context because this is GFP_NOFS request. It is > >>> not all that surprising the reclaim context fights to get some memory. > >>> There is a huge amount of the reclaimable slab which probably just makes > >>> a slow progress. > >>> > >>> That is not something completely surprsing on 32b system I am afraid. > >>> > >>> Btw. is the stall repeating with the increased time or it gets resolved > >>> eventually? > >> > >> Yes and if you mean by repeating it’s not only affecting rsync but > >> you can see just now automount and NetworkManager get these page > >> allocation stalls and kswapd0 is getting heavy CPU load, are there any > >> other settings I can adjust? > > > > None that I am aware of. You might want to talk to FS guys, maybe they > > can figure out who is pinning file pages so that they cannot be > > reclaimed. They do not seem to be dirty or under writeback. It would be > > also interesting to see whether that is a regression. The warning is > > relatively new so you might have had this problem before just haven't > > noticed it. > > We have been getting out of memory errors for a while but those seem > to have gone away.
this sounds suspicious. Are you really sure that this is a new problem? Btw. is there any reason to use 32b kernel at all? It will always suffer from a really small lowmem... > We did just replace the controller in the VessRAID > as there were some timeouts observed and multiple login/logout > attempts. > > By FS guys do you mean the linux-fsdevel or linux-fsf list? yeah linux-fsdevel. No idea what linux-fsf is. It would be great if you could collect some tracepoints before reporting the issue. At least those in events/vmscan/*. -- Michal Hocko SUSE Labs
