On Mon, Apr 06, 2015 at 06:20:17AM +0000, Naoya Horiguchi wrote: > Hi, > > When doing hugepage migration test on mmotm-2015-04-01-14-54, I found 2 cases > of the VM_BUG_ONs detected by page flag sanitization for compound pages. > > 1) > > [ 3071.845424] page:ffffea0018f08000 count:1 mapcount:0 mapping:ffff88065dcc1a51 index:0x3 > [ 3071.854375] flags: 0x5ffc0000004009(locked|uptodate|head) > [ 3071.860532] page dumped because: VM_BUG_ON_PAGE(PageCompound(page)) > [ 3071.867560] ------------[ cut here ]------------ > [ 3071.872715] kernel BUG at include/linux/page-flags.h:333! > [ 3071.878742] invalid opcode: 0000 [#1] SMP > [ 3071.883336] Modules linked in: xt_CHECKSUM iptable_mangle ipt_MASQUERADE nf_nat_masquerade_ipv4 iptable_nat nf_nat_ipv4 nf_nat nf_conntrack_ipv4 nf_defrag_ipv4 xt_conntrack nf_conntrack bridge stp llc x86_pkg_temp_thermal coretemp kvm_intel kvm crct10dif_pclmul crc32_pclmul crc32c_intel iTCO_wdt i2c_i801 ghash_clmulni_intel iTCO_vendor_support sb_edac edac_core lpc_ich mfd_core pcspkr ipmi_ssif tpm_tis ipmi_si tpm wmi ipmi_msghandler acpi_power_meter mei_me ioatdma mei shpchp dca nfsd auth_rpcgss nfs_acl lockd grace sunrpc uas usb_storage mgag200 bnx2x i2c_algo_bit drm_kms_helper e1000e ttm mdio libcrc32c drm ptp megaraid_sas pps_core > [ 3071.946801] CPU: 0 PID: 2796 Comm: test_mbind_fuzz Not tainted 3.19.0 #1 > [ 3071.954282] Hardware name: NEC Express5800/B120e [N8400-221Y]/G7LYN, BIOS 4.6.2106 11/08/2013 > [ 3071.963802] task: ffff88065d068a50 ti: ffff8806621b4000 task.ti: ffff8806621b4000 > [ 3071.972157] RIP: 0010:[<ffffffff8120882a>] [<ffffffff8120882a>] migrate_page_copy+0x58a/0x7f0 > [ 3071.981781] RSP: 0018:ffff8806621b7c98 EFLAGS: 00010296 > [ 3071.987711] RAX: 0000000000000037 RBX: ffffea0018f08000 RCX: 0000000000000037 > [ 3071.995680] RDX: 0000000000000000 RSI: 0000000000000000 RDI: ffff880667c0e838 > [ 3072.003646] RBP: ffff8806621b7ce8 R08: 0000000000000096 R09: 0000000000000534 > [ 3072.011615] R10: 0000000000000000 R11: 0000000000000534 R12: ffffea0019010000 > [ 3072.019583] R13: 0000000000000200 R14: 0000000018f10000 R15: ffff880000000000 > [ 3072.027551] FS: 00007f674471d740(0000) GS:ffff880667c00000(0000) knlGS:0000000000000000 > [ 3072.036586] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 > [ 3072.043001] CR2: 00007fffccc549f8 CR3: 0000000662c05000 CR4: 00000000001407f0 > [ 3072.050968] Stack: > [ 3072.053224] 0000000000000000 800000063c2000e7 ffff880600000040 0000020000e00000 > [ 3072.061525] ffffea0001dfddf0 ffffea0018f08000 ffffea0019010000 0000000000000001 > [ 3072.069826] 0000000000000000 00000000fffffff5 ffff8806621b7d18 ffffffff81208acd > [ 3072.078124] Call Trace: > [ 3072.080858] [<ffffffff81208acd>] migrate_page+0x3d/0x60 > [ 3072.086790] [<ffffffff81208dbd>] move_to_new_page+0x2cd/0x340 > [ 3072.093304] [<ffffffff811e03e6>] ? try_to_unmap+0x66/0x130 > [ 3072.099527] [<ffffffff811ddde0>] ? invalid_migration_vma+0x30/0x30 > [ 3072.106527] [<ffffffff81209817>] migrate_pages+0x8a7/0x9f0 > [ 3072.112752] [<ffffffff811f7640>] ? alloc_pages_vma+0x220/0x220 > [ 3072.119364] [<ffffffff811f7e14>] do_mbind+0x504/0x5d0 > [ 3072.125101] [<ffffffff811f8138>] SyS_mbind+0xa8/0xb0 > [ 3072.130743] [<ffffffff8177d769>] system_call_fastpath+0x12/0x17 > [ 3072.137448] Code: 0f 0b 0f 1f 80 00 00 00 00 48 8b 43 30 48 8b 13 80 e6 80 48 0f 44 c3 e9 36 fc ff ff 48 c7 c6 48 71 a3 81 48 89 df e8 f6 4c fc ff <0f> 0b 48 c7 c6 48 71 a3 81 4c 89 e7 e8 e5 4c fc ff 0f 0b 48 c7 > [ 3072.159196] RIP [<ffffffff8120882a>] migrate_page_copy+0x58a/0x7f0 > [ 3072.166217] RSP <ffff8806621b7c98> > [ 3073.015139] ---[ end trace f52f8daca3fe12f2 ]--- > > migrate_page_copy() always calls ClearPageSwapCache(), but recently calling > it for compound pages is banned by Kirill's page flag sanitization patch. > I think that the sanitization made this bug visible, and we need fix the ill > usage of ClearPageSwapCache. > I think thp migration should also hit this BUG_ON. > > I'm thinking of one simplest fix like this: > > diff --git a/mm/migrate.c b/mm/migrate.c > index 25fd7f6291de..5fa399d20435 100644 > --- a/mm/migrate.c > +++ b/mm/migrate.c > @@ -537,7 +537,8 @@ void migrate_page_copy(struct page *newpage, struct page *page) > * Please do not reorder this without considering how mm/ksm.c's > * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache(). > */ > - ClearPageSwapCache(page); > + if (PageSwapCache(page)) > + ClearPageSwapCache(page); > ClearPagePrivate(page); > set_page_private(page, 0); > > but, this looks a dirty workaround. I think this change is reasonable for general case: we would avoid atoimic opration if the bit is not set. > Creating a migrate_copy_page() variant for compound page is better? > Or any better idea? > > > 2) > > # This is triggered easily by libhugetlbfs functional test. > > [ 2662.212167] page:ffffea0002b38000 count:2 mapcount:1 mapping:ffff8800db8ec380 index:0x0 > [ 2662.213480] flags: 0xbffc0000004008(uptodate|head) > [ 2662.214252] page dumped because: VM_BUG_ON_PAGE(PageCompound(page)) > [ 2662.215448] ------------[ cut here ]------------ > [ 2662.216135] kernel BUG at /src/linux-dev/include/linux/page-flags.h:317! > [ 2662.216427] invalid opcode: 0000 [#1] SMP > [ 2662.216427] Modules linked in: fuse btrfs xor raid6_pq ufs hfsplus hfs minix vfat msdos fat jfs xfs libcrc32c reiserfs cfg80211 rfkill ppdev crc32c_intel virtio_console pcspkr virtio_balloon serio_raw pa > rport_pc parport pvpanic i2c_piix4 virtio_blk virtio_net ata_generic pata_acpi virtio_pci virtio_ring virtio floppy > [ 2662.216427] CPU: 1 PID: 12413 Comm: test_mbind_fuzz Not tainted 3.19.0-mmotm-2015-04-01-14-54-150406-1312-00001-gfe6d51b5f154 #36 > [ 2662.216427] Hardware name: Bochs Bochs, BIOS Bochs 01/01/2011 > [ 2662.216427] task: ffff88007b39f170 ti: ffff88007b4a4000 task.ti: ffff88007b4a4000 > [ 2662.216427] RIP: 0010:[<ffffffff811837ff>] [<ffffffff811837ff>] set_page_dirty+0x8f/0xc0 > [ 2662.216427] RSP: 0000:ffff88007b4a7a08 EFLAGS: 00010292 > [ 2662.216427] RAX: 0000000000000037 RBX: ffffea0002b38000 RCX: 0000000000000037 > [ 2662.216427] RDX: 0000000000000000 RSI: 0000000000000000 RDI: ffff88011fc0e838 > [ 2662.216427] RBP: ffff88007b4a7a18 R08: 0000000000000092 R09: 000000000000020c > [ 2662.216427] R10: 0000000000000000 R11: 000000000000020c R12: ffff8800da599f80 > [ 2662.216427] R13: ffffffff81fbc6c0 R14: ffff8800da065000 R15: ffffea0003681970 > [ 2662.216427] FS: 0000000000000000(0000) GS:ffff88011fc00000(0000) knlGS:0000000000000000 > [ 2662.216427] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 > [ 2662.216427] CR2: 00007fffe8a99ff8 CR3: 00000000db1be000 CR4: 00000000000006e0 > [ 2662.216427] Stack: > [ 2662.216427] ffffffff81fbc6c0 0000700000200000 ffff88007b4a7aa8 ffffffff811cdb39 > [ 2662.216427] 80000000ace000e7 ffffea0002b38000 ffff88007c74f730 ffff88007b4a7c40 > [ 2662.216427] 0000000000000000 ffff8800da599fe4 0000700000000000 0000000000200000 > [ 2662.216427] Call Trace: > [ 2662.216427] [<ffffffff811cdb39>] __unmap_hugepage_range+0x329/0x340 > [ 2662.216427] [<ffffffff811cdb66>] __unmap_hugepage_range_final+0x16/0x30 > [ 2662.216427] [<ffffffff811acace>] unmap_single_vma+0x51e/0x910 > [ 2662.216427] [<ffffffff811ad7c4>] unmap_vmas+0x54/0xb0 > [ 2662.216427] [<ffffffff811b735c>] exit_mmap+0xac/0x180 > [ 2662.216427] [<ffffffff81074d33>] mmput+0x63/0xf0 > [ 2662.216427] [<ffffffff8107a41d>] do_exit+0x2ad/0xac0 > [ 2662.216427] [<ffffffff8107acc7>] do_group_exit+0x47/0xc0 > [ 2662.216427] [<ffffffff81086c04>] get_signal+0x2a4/0x6b0 > [ 2662.216427] [<ffffffff810145b7>] do_signal+0x37/0x800 > [ 2662.216427] [<ffffffff81202365>] ? __sb_end_write+0x35/0x70 > [ 2662.216427] [<ffffffff811ffc22>] ? vfs_write+0x1b2/0x1f0 > [ 2662.216427] [<ffffffff81014de9>] do_notify_resume+0x69/0xb0 > [ 2662.216427] [<ffffffff816dd4e2>] retint_signal+0x48/0x86 > [ 2662.216427] Code: df 48 8b 03 f6 c4 80 75 3f f0 0f ba 2b 04 72 24 b8 01 00 00 00 eb c9 0f 1f 44 00 00 48 c7 c6 e0 d2 a0 81 48 89 df e8 71 59 02 00 <0f> 0b 0f 1f 80 00 00 00 00 31 c0 eb a8 48 8b 43 30 48 > 8b 13 80 > [ 2662.216427] RIP [<ffffffff811837ff>] set_page_dirty+0x8f/0xc0 > [ 2662.216427] RSP <ffff88007b4a7a08> > > set_page_dirty() calls ClearPageReclaim(), which is also banned to call on > compound pages (but it is called now.) > thp is not affected by this because page_mapping() is NULL for thp (maybe > this will not be true when thp supports page cache?). > As a short term fix, hugetlb had better have its own set_page_dirty? The same check-before-set should work fine, right? -- Kirill A. Shutemov -- 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>