Hi! On Mon, Jun 17, 2013 at 02:12:44PM -0400, Rik van Riel wrote: > On 06/17/2013 05:30 AM, Mel Gorman wrote: > > On Fri, Jun 14, 2013 at 12:16:47PM -0400, Rik van Riel wrote: > >> On 06/06/2013 01:37 PM, Rik van Riel wrote: > >>> On 06/06/2013 05:04 AM, Mel Gorman wrote: > >>>> On Wed, Jun 05, 2013 at 05:10:31PM +0200, Andrea Arcangeli wrote: > >>>>> Zone reclaim locked breaks zone_reclaim_mode=1. If more than one > >>>>> thread allocates memory at the same time, it forces a premature > >>>>> allocation into remote NUMA nodes even when there's plenty of clean > >>>>> cache to reclaim in the local nodes. > >>>>> > >>>>> Signed-off-by: Andrea Arcangeli <aarcange@xxxxxxxxxx> > >>>> > >>>> Be aware that after this patch is applied that it is possible to have a > >>>> situation like this > >>>> > >>>> 1. 4 processes running on node 1 > >>>> 2. Each process tries to allocate 30% of memory > >>>> 3. Each process reads the full buffer in a loop (stupid, just an example) > >>>> > >>>> In this situation the processes will continually interfere with each > >>>> other until one of them gets migrated to another zone by the scheduler. > >>> > >>> This is a very good point. > >>> > >>> Andrea, I suspect we will need some kind of safeguard against > >>> this problem. > >> > >> Never mind me. > >> > >> In __zone_reclaim we set the flags in swap_control so > >> we never unmap pages or swap pages out at all by > >> default, so this should not be an issue at all. > >> > >> In order to get the problem illustrated above, the > >> user will have to enable RECLAIM_SWAP through sysfs > >> manually. > >> > > > > For the mapped case and the default tuning for zone_reclaim_mode then > > yes. If instead of allocating 30% of memory the processes are using using > > buffered reads/writes then they'll reach each others page cache pages and > > it's a very similar problem. > > Could we fix that problem by simply allowing page cache > allocations (__GFP_WRITE) to fall back to other zones, > regardless of the zone_reclaim setting? > > The ZONE_RECLAIM_LOCKED function seems to break as many > things as it fixes, so replacing it with something else > seems like a worthwhile pursuit... I actually don't see a connection between the various scenarios described above with ZONE_RECLAIM_LOCKED. I mean whatever problem you are having with swapping or excessive reclaim in a single zone/node despite the other zones/nodes are completely free, could materialize the same way with the current ZONE_RECLAIM_LOCKED code if you just use a mutex in userland to serialize the memory allocations. Or if they just happen to run serially for other reasons. If it was a problem to keep insisting calling zone_reclaim in any given zone, the problem would eventually materialize anyway, by just running a single thread in the whole system pinned to a single node. ZONE_RECLAIM_LOCKED isn't about swapping or memory pressure, it is only about preventing running more than one zone_reclaim function at once in any given zone. But that shall be ok. If all zone_reclaim() running in parallel are doing a .nr_to_reclaim = SWAP_CLUSTER_MAX shrinkage attempt with may_unmap/may_writepage unset (zone_reclaim_mode is <=1), there shall be no problem. And zone_reclaim won't be called anymore as soon as the watermark is above "low". -- 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>