On Thu 03-12-15 21:59:50, Minchan Kim wrote: > On Thu, Dec 03, 2015 at 09:54:52AM +0100, Michal Hocko wrote: > > On Thu 03-12-15 11:10:06, Minchan Kim wrote: > > > On Thu, Dec 03, 2015 at 10:34:04AM +0900, Minchan Kim wrote: > > > > On Wed, Dec 02, 2015 at 11:16:43AM +0100, Michal Hocko wrote: [...] > > > > > Also, how big is the underflow? > > [...] > > > > nr_pages 293 new -324 > > > > nr_pages 16 new -340 > > > > nr_pages 342 new -91 > > > > nr_pages 246 new -337 > > > > nr_pages 15 new -352 > > > > nr_pages 15 new -367 > > > > They are quite large but that is not that surprising if we consider that > > we are batching many uncharges at once. > > > > > My guess is that it's related to new feature of Kirill's THP 'PageDoubleMap' > > > so a THP page could be mapped a pte but !pmd_trans_huge(*pmd) so memcg > > > precharge in move_charge should handle it? > > > > I am not familiar with the current state of THP after the rework > > unfortunately. So if I got you right then you are saying that > > pmd_trans_huge_lock fails to notice a THP so we will not charge it as > > THP and only charge one head page and then the tear down path will > > correctly recognize it as a THP and uncharge the full size, right? > > Exactly. Hmm, but are pages represented by those ptes on the LRU list? __split_huge_pmd_locked doesn't seem to do any lru care. If they are not on any LRU then mem_cgroup_move_charge_pte_range should ignore such a pte and the THP (which the pte is part of) should stay in the original memcg. How do you trigger this issue btw? -- Michal Hocko SUSE Labs -- 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>