(In reply to Christian Wittmer from comment #3) > (In reply to Michal Hocko from comment #2) > > These all seem to be temporary allocation stalls 3 hitting one window Mar 8 > > and one mar 13. One hitting xfs, 2 skb and one vfs allocation paths. So > > nothing really systematic. Checking the free memory proves that the memory > > was mostly used. > > > > I am strongly suspecting that some part of the memory reclaim (shrinkers > > probably) took excessive amount of time to make a forward progress. We would > > need vmsacan tracepoints data to know better though. > > How can I provide such info about 'vmsacan tracepoints data' ? There are two ways. You can either mount tracefs and enable $TRACEFS_MNT/events/vmscan/{mm_vmscan_direct_reclaim_begin,mm_vmscan_direct_reclaim_end,mm_shrink_slab_start,mm_shrink_slab_end} and read the output from $TRACE_MNT/trace_pipe or use trace-cmd to do the same. Btw. is this reproducible?