Michal Hocko wrote: > On Wed 21-10-15 09:49:07, Christoph Lameter wrote: > > On Wed, 21 Oct 2015, Michal Hocko wrote: > > > > > Because all the WQ workers are stuck somewhere, maybe in the memory > > > allocation which cannot make any progress and the vmstat update work is > > > queued behind them. After invoking the OOM killer, we can easily observe that vmstat_update cannot be processed due to memory allocation by disk_events_workfn stalls. http://lkml.kernel.org/r/201509120019.BJI48986.OOSVMJtOLFQHFF@xxxxxxxxxxxxxxxxxxx I worried that blocking forever from workqueue is an exclusive occupation of workqueue. In fact, changing to GFP_ATOMIC avoids this problem. http://lkml.kernel.org/r/201503012017.EAD00571.HOOJVOStMFLFQF@xxxxxxxxxxxxxxxxxxx Now we realized that we are hitting this problem before invoking the OOM killer. The situation is similar to the case after the OOM killer is invoked; there are no reclaimable pages but vmstat_update cannot be processed. We are caught by a small difference of vmstat counter values. > > > > > > At least this is my current understanding. > > > > Eww. Maybe need a queue that does not do such evil things as memory > > allocation? > > I am not sure how to achieve that. Requiring non-sleeping worker would > work out but do we have enough users to add such an API? If a queue does not need to sleep, can't that queue be processed from timer context (e.g. mod_timer()) ? -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxx. For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>