On Wed, Apr 29, 2020 at 06:40:02PM +0300, Kirill A. Shutemov wrote: > On Wed, Apr 29, 2020 at 06:36:32AM -0700, Matthew Wilcox wrote: > > From: "Matthew Wilcox (Oracle)" <willy@xxxxxxxxxxxxx> > > > > This patch set does not pass xfstests. Test at your own risk. It is > > based on the readahead rewrite which is in Andrew's tree. The large > > pages somehow manage to fall off the LRU, so the test VM quickly runs > > out of memory and freezes. To reproduce: > > > > # mkfs.xfs /dev/sdb && mount /dev/sdb /mnt && dd if=/dev/zero bs=1M count=2048 of=/mnt/bigfile && sync && sleep 2 && sync && echo 1 >/proc/sys/vm/drop_caches > > # /host/home/willy/kernel/xarray-2/tools/vm/page-types | grep thp > > 0x0000000000401800 511 1 ___________Ma_________t____________________ mmap,anonymous,thp > > 0x0000000000405868 1 0 ___U_lA____Ma_b_______t____________________ uptodate,lru,active,mmap,anonymous,swapbacked,thp > > # dd if=/mnt/bigfile of=/dev/null bs=2M count=5 > > # /host/home/willy/kernel/xarray-2/tools/vm/page-types | grep thp > > 0x0000000000400000 2516 9 ______________________t____________________ thp > > 0x0000000000400028 1 0 ___U_l________________t____________________ uptodate,lru,thp > > 0x000000000040006c 106 0 __RU_lA_______________t____________________ referenced,uptodate,lru,active,thp > > Note that you have 107 pages on LRU. It is only head pages. With order-5 > pages it is over 13MiB. > > Looks like everything is fine. /proc/kpageflags reads page's flag bit directly instead of relying on PageLRU: u |= kpf_copy_bit(k, KPF_LRU, PG_lru); Tail pages don't have this bit set. They rely on head page's flag for PageLRU(). It would be nice to get it fixed, but I guess it is too late. Somebody may rely on the current behaviour by now. -- Kirill A. Shutemov
