Mel Gorman wrote: > On (13/09/07 19:07), KAMEZAWA Hiroyuki didst pronounce: >> Hi, >> >> While I'm playing with memory controller of 2.6.23-rc4-mm1, I met following. >> >> == >> [root@drpq test-2.6.23-rc4-mm1]# echo $$ > /opt/mem_control/group_1/tasks >> [root@drpq test-2.6.23-rc4-mm1]# cat /opt/mem_control/group_1/memory.limit >> 32768 >> [root@drpq test-2.6.23-rc4-mm1]# cat /opt/mem_control/group_1/memory.usage >> 286 >> // Memory is limited to 512 GiB. try "dd" 1GiB (page size is 16KB) >> >> [root@drpq test-2.6.23-rc4-mm1]# dd if=/dev/zero of=/tmp/tmpfile bs=1024 count=1048576 >> Killed >> [root@drpq test-2.6.23-rc4-mm1]# ls >> Killed >> //above are caused by OOM. >> [root@drpq test-2.6.23-rc4-mm1]# cat /opt/mem_control/group_1/memory.usage >> 32763 >> [root@drpq test-2.6.23-rc4-mm1]# cat /opt/mem_control/group_1/memory.limit >> 32768 >> // fully filled by page cache. no reclaim run. >> == >> >> The reason this happens is because I used kernelcore= boot option, i.e >> ZONE_MOVABLE. Seems try_to_free_mem_container_pages() ignores ZONE_MOVABLE. >> >> Quick fix is attached, but Mel's one-zonelist-pernode patch may change this. >> I'll continue to watch. >> > > You are right on both counts. This is a valid fix but > one-zonelist-pernode overwrites it. Specifically the code in question > with one-zonelist will look like; > > for_each_online_node(node) { > zonelist = &NODE_DATA(node)->node_zonelist; > if (do_try_to_free_pages(zonelist, sc.gfp_mask, &sc)) > return 1; > } > > We should be careful that this problem does not get forgotten about if > one-zonelist gets delayed for a long period of time. Have the fix at the > end of the container patchset where it can be easily dropped if > one-zonelist is merged. > > Thanks Yes, I second that. So, we should get KAMEZAWA's fix in. -- Warm Regards, Balbir Singh Linux Technology Center IBM, ISTL _______________________________________________ Containers mailing list Containers@xxxxxxxxxxxxxxxxxxxxxxxxxx https://lists.linux-foundation.org/mailman/listinfo/containers