Hello, On kernel 4.4 I observe that the memory gets really fragmented fairly quickly. E.g. there are no order > 4 pages even after 2 days of uptime. This leads to certain data structures on XFS (in my case order 4/order 5 allocations) not being allocated and causes the server to stall. When this happens either someone has to log on the server and manually invoke the memory compaction or plain reboot the server. Before that the server was running with the exact same workload but with 3.12.52 kernel and no such issue were observed. That is - memory was fragmented but allocation didn't fail, maybe alloc_pages_direct_compact was doing a better job? FYI the allocation is performed with GFP_KERNEL | GFP_NOFS Manual compaction usually does the job, however I'm wondering why isn't invoking __alloc_pages_direct_compact from within __alloc_pages_nodemask satisfying the request if manual compaction would do the job. Is there a difference in the efficiency of manually invoking memory compaction and the one invoked from the page allocator path? Another question for my own satisfaction - I created a kernel module which allocate pages of very high order - 8/9) then later when those pages are returned I see the number of unmovable pages increase by the amount of pages returned. So should freed pages go to the unmovable category? -- 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>