On 03/03/2014 09:28 AM, Joonsoo Kim wrote:
On Fri, Feb 28, 2014 at 03:15:04PM +0100, Vlastimil Babka wrote:set_pageblock_flags_group() is used to set either migratetype or skip bit of a pageblock. Setting migratetype is done under zone->lock (except from __init code), however changing the skip bits is not protected and the pageblock flags bitmap packs migratetype and skip bits together and uses non-atomic bit ops. Therefore, races between setting migratetype and skip bit are possible and the non-atomic read-modify-update of the skip bit may cause lost updates to migratetype bits, resulting in invalid migratetype values, which are in turn used to e.g. index free_list array. The race has been observed to happen and cause panics, albeit during development of series that increases frequency of migratetype changes through {start,undo}_isolate_page_range() calls. Two possible solutions were investigated: 1) using zone->lock for changing pageblock_skip bit and 2) changing the bitmap operations to be atomic. The problem of 1) is that zone->lock is already contended and almost never held in the compaction code that updates pageblock_skip bits. Solution 2) should scale better, but adds atomic operations also to migratype changes which are already protected by zone->lock.How about 3) introduce new bitmap for pageblock_skip? I guess that migratetype bitmap is read-intensive and set/clear pageblock_skip could make performance degradation.
Yes that would be also possible, but was deemed too ugly and maybe even uglier in case some new pageblock bits are introduced. But it seems no performance degradation was observed for 1) and 2).
I guess if we left the whole idea of packed bitmap we could also make atomic the update of the whole migratetype instead of processing each bit separately. But that would mean at least 8 bits per pageblock for migratetype (and I have no idea about specifics for other archs than x86 here). Maybe 4 bits if it's even more ugly and distinguishes odd and even pageblocks...
Using mmtests' stress-highalloc benchmark, little difference was found between the two solutions. The base is 3.13 with recent compaction series by myself and Joonsoo Kim applied. 3.13 3.13 3.13 base 2)atomic 1)lock User 6103.92 6072.09 6178.79 System 1039.68 1033.96 1042.92 Elapsed 2114.27 2090.20 2110.23I really wonder how 2) is better than base although there is a little difference. Is it the avg result of 10 runs? Do you have any idea what happens?
It is avg of 10 runs but I guess this just means 10 runs are not enough to get results precise enough. One difference is that atomic version does not clear/set bits that don't need it, but the profiles show the whole operation is pretty negligible. And if at least one bit is changed (I guess it is, unless migratetypes are somewhere set to the same value as they already are), cache line becomes dirty anyway. And again, profiles suggest that very little cache is dirtied here.
Vlastimil
Thanks.
-- 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>