On Wed 13-07-16 10:35:01, Jerome Marchand wrote: > On 07/13/2016 01:44 AM, Mikulas Patocka wrote: > > The problem of swapping to dm-crypt is this. > > > > The free memory goes low, kswapd decides that some page should be swapped > > out. However, when you swap to an ecrypted device, writeback of each page > > requires another page to hold the encrypted data. dm-crypt uses mempools > > for all its structures and pages, so that it can make forward progress > > even if there is no memory free. However, the mempool code first allocates > > from general memory allocator and resorts to the mempool only if the > > memory is below limit. > > > > So every attempt to swap out some page allocates another page. > > > > As long as swapping is in progress, the free memory is below the limit > > (because the swapping activity itself consumes any memory over the limit). > > And that triggered the OOM killer prematurely. > > There is a quite recent sysctl vm knob that I believe can help in this > case: watermark_scale_factor. If you increase this value, kswapd will > start paging out earlier, when there might still be enough free memory. > > Ondrej, have you tried to increase /proc/sys/vm/watermark_scale_factor? I suspect this would just change the timing or the real problem gets hidden. -- Michal Hocko SUSE Labs -- dm-devel mailing list dm-devel@xxxxxxxxxx https://www.redhat.com/mailman/listinfo/dm-devel
