On Mon, Apr 18, 2016 at 03:44:59PM -0400, Sasha Levin wrote: > On 09/24/2015 10:51 AM, Kirill A. Shutemov wrote: > > Transparent huge pages can be mlocked -- whole compund page at once. > > Something went wrong if we're trying to mlock() tail page. > > Let's use PF_NO_TAIL. > > Kirill, Hugh, > > I seem to be hitting this with trinity: > > [ 242.257552] page:ffffea0001517fc0 count:0 mapcount:0 > mapping:dead000000000400 index:0x0 compound_mapcount: 0 > [ 242.258680] flags: 0x1fffff80000000() > [ 242.261853] page dumped because: VM_BUG_ON_PAGE(1 && PageTail(page)) > [ 242.262712] ------------[ cut here ]------------ > [ 242.263182] kernel BUG at include/linux/page-flags.h:332! > [ 242.263731] invalid opcode: 0000 [#1] PREEMPT SMP KASAN > [ 242.264246] Modules linked in: > [ 242.264603] CPU: 2 PID: 9302 Comm: trinity-c67 Not tainted > 4.6.0-rc3-next-20160412-sasha-00024-geaec67e #3001 > [ 242.265574] task: ffff88013270c000 ti: ffff88012d0d0000 task.ti: > ffff88012d0d0000 > [ 242.266306] RIP: clear_pages_mlock (include/linux/page-flags.h:332 > mm/mlock.c:82) > [ 242.267325] RSP: 0018:ffff88012d0d70e0 EFLAGS: 00010286 > [ 242.267894] RAX: 0000000000000000 RBX: ffffea0001517fc0 RCX: > 0000000000000000 > [ 242.268773] RDX: 1ffffd40002a2fff RSI: 0000000000000282 RDI: > ffffea0001517ff8 > [ 242.269495] RBP: ffff88012d0d7120 R08: ffffffffb66a1566 R09: > fffffbfff6cd42ad > [ 242.270227] R10: ffffffffb61dc140 R11: 00000000e27286ff R12: > ffffea0001517fe0 > [ 242.270961] R13: ffffea0001518000 R14: dffffc0000000000 R15: > 0000000000000000 > [ 242.271686] FS: 00007f594abad700(0000) GS:ffff8801b5a00000(0000) > knlGS:0000000000000000 > [ 242.272494] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 > [ 242.273073] CR2: 0000000001f27000 CR3: 0000000132418000 CR4: > 00000000000006a0 > [ 242.273815] Stack: > [ 242.274036] 0000000000000008 ffff880083fd3000 0000000100000000 > ffffea0001517fc0 > [ 242.274847] ffffea0001517fe0 ffffea0001510000 ffffea0001510001 > 0000000000000000 > [ 242.275656] ffff88012d0d7170 ffffffff9d712322 ffffffff9d7221ab > 0000000001400000 > [ 242.276458] Call Trace: > [ 242.276746] page_remove_rmap (include/linux/page-flags.h:157 > include/linux/page-flags.h:522 mm/rmap.c:1383) > [ 242.277989] __split_huge_pmd_locked (include/linux/compiler.h:222 > (discriminator 3) include/linux/page-flags.h:143 (discriminator 3) > include/linux/mm.h:736 (discriminator 3) mm/huge_memory.c:3075 (discriminator > 3)) > [ 242.279909] __split_huge_pmd (include/linux/spinlock.h:347 > mm/huge_memory.c:3102) > [ 242.280515] split_huge_pmd_address (mm/huge_memory.c:3137) > [ 242.281181] try_to_unmap_one (include/linux/compiler.h:222 > include/linux/page-flags.h:143 include/linux/page-flags.h:268 > include/linux/mm.h:495 mm/rmap.c:1425) > [ 242.285008] rmap_walk_anon (mm/rmap.c:1762) > [ 242.286167] rmap_walk_locked (mm/rmap.c:1845) > [ 242.286753] try_to_unmap (mm/rmap.c:1643) > [ 242.291004] split_huge_page_to_list (mm/huge_memory.c:3191 > mm/huge_memory.c:3380) > [ 242.293499] queue_pages_pte_range (mm/mempolicy.c:505) > [ 242.294728] __walk_page_range (mm/pagewalk.c:51 mm/pagewalk.c:90 > mm/pagewalk.c:116 mm/pagewalk.c:204) > [ 242.295331] walk_page_range (mm/pagewalk.c:282) > [ 242.295910] queue_pages_range (mm/mempolicy.c:667) > [ 242.299552] migrate_to_node (include/linux/compiler.h:222 > include/linux/list.h:189 mm/mempolicy.c:1002) > [ 242.301994] do_migrate_pages (mm/mempolicy.c:1105) > [ 242.303811] SYSC_migrate_pages (mm/mempolicy.c:1451) > [ 242.307712] SyS_migrate_pages (mm/mempolicy.c:1369) > [ 242.308290] ? SyS_set_mempolicy (mm/mempolicy.c:1369) > [ 242.308892] do_syscall_64 (arch/x86/entry/common.c:350) > [ 242.310743] entry_SYSCALL64_slow_path (arch/x86/entry/entry_64.S:251) > [ 242.311395] Code: 42 80 3c 30 00 74 08 4c 89 e7 e8 c7 f8 08 00 48 8b 43 20 > a8 01 74 22 e8 da e2 ea ff 48 c7 c6 e0 9b 31 a7 48 89 df e8 0b 01 fe ff <0f> > 0b 48 c7 c7 e0 3b 52 ab e8 5f 3b 9d 01 e8 b8 e2 ea ff 48 8b
That's very strange. It's the same bug you've reported here: http://lkml.kernel.org/r/571638cf.5090...@oracle.com Looks like the page was re-mlocked() under us while we're splitting pmd. In both cases it's last tail page of the THP, which probably some hint, but I don't get it. Do you still see the issue? I have never seen anything like this myself. -- Kirill A. Shutemov
