Allocation of hugetlb pages via sysctl or procfs can stall for minutes or hours. A simple example on a two node system with 8GB of memory is as follows: echo 4096 > /sys/devices/system/node/node1/hugepages/hugepages-2048kB/nr_hugepages echo 4096 > /proc/sys/vm/nr_hugepages Obviously, both allocation attempts will fall short of their 8GB goal. However, one or both of these commands may stall and not be interruptible. The issues were discussed in this thread [1]. This series attempts to address the issues causing the stalls. There are two distinct issues, and an optimization. For the reclaim and compaction issues, suggestions were made to simply remove some existing code. However, the impact of such changes would be hard to address. This series takes a more conservative approach in an attempt to minimally impact existing workloads. The question of which approach is better is debatable, hence the RFC designation. Patches in the series address these issues: 1) Should_continue_reclaim returns true too often. Michal Hocko suggested removing the special casing for __GFP_RETRY_MAYFAIL in should_continue_reclaim. This does indeed address the hugetlb allocations, but may impact other users. Hillf Danton restructured the code in such a way to preserve much of the original semantics. Hillf's patch also addresses hugetlb allocation issues and is included here. 2) With 1) addressed, should_compact_retry returns true too often. Mel Gorman suggested the removal of the compaction_zonelist_suitable() call. This routine/call was introduced by Michal Hocko for a specific use case. Therefore, removal would likely break that use case. While examining the reasons for compaction_withdrawn() as in [2], it appears that there are several places where we should be using MIN_COMPACT_COSTLY_PRIORITY instead of MIN_COMPACT_PRIORITY for costly allocations. This patch makes those changes which also causes more appropriate should_compact_retry behavior for hugetlb allocations. 3) This is simply an optimization of the allocation code for hugetlb pool pages. After first __GFP_RETRY_MAYFAIL allocation failure on a node, it drops the __GFP_RETRY_MAYFAIL flag. [1] http://lkml.kernel.org/r/d38a095e-dc39-7e82-bb76-2c9247929f07@xxxxxxxxxx [2] http://lkml.kernel.org/r/6377c199-2b9e-e30d-a068-c304d8a3f706@xxxxxxxxxx Hillf Danton (1): mm, reclaim: make should_continue_reclaim perform dryrun detection Mike Kravetz (2): mm, compaction: use MIN_COMPACT_COSTLY_PRIORITY everywhere for costly orders hugetlbfs: don't retry when pool page allocations start to fail mm/compaction.c | 18 +++++++--- mm/hugetlb.c | 87 +++++++++++++++++++++++++++++++++++++++++++------ mm/vmscan.c | 28 ++++++++-------- 3 files changed, 105 insertions(+), 28 deletions(-) -- 2.20.1
