Btrfs: fix deadlock between direct IO reads and buffered writes
authorFilipe Manana <fdmanana@suse.com>
Thu, 18 Feb 2016 14:28:55 +0000 (14:28 +0000)
committerChris Mason <clm@fb.com>
Tue, 1 Mar 2016 16:23:37 +0000 (08:23 -0800)
commitade770294df29e08f913e5d733a756893128f45e
treec73de7a8809a7cbaaacb44430bb1e9d46cc45612
parentf4dfe6871006c62abdccc77b2818b11f376e98e2
Btrfs: fix deadlock between direct IO reads and buffered writes

While running a test with a mix of buffered IO and direct IO against
the same files I hit a deadlock reported by the following trace:

[11642.140352] INFO: task kworker/u32:3:15282 blocked for more than 120 seconds.
[11642.142452]       Not tainted 4.4.0-rc6-btrfs-next-21+ #1
[11642.143982] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
[11642.146332] kworker/u32:3   D ffff880230ef7988 [11642.147737] systemd-journald[571]: Sent WATCHDOG=1 notification.
[11642.149771]     0 15282      2 0x00000000
[11642.151205] Workqueue: btrfs-flush_delalloc btrfs_flush_delalloc_helper [btrfs]
[11642.154074]  ffff880230ef7988 0000000000000246 0000000000014ec0 ffff88023ec94ec0
[11642.156722]  ffff880233fe8f80 ffff880230ef8000 ffff88023ec94ec0 7fffffffffffffff
[11642.159205]  0000000000000002 ffffffff8147b7f9 ffff880230ef79a0 ffffffff8147b541
[11642.161403] Call Trace:
[11642.162129]  [<ffffffff8147b7f9>] ? bit_wait+0x2f/0x2f
[11642.163396]  [<ffffffff8147b541>] schedule+0x82/0x9a
[11642.164871]  [<ffffffff8147e7fe>] schedule_timeout+0x43/0x109
[11642.167020]  [<ffffffff8147b7f9>] ? bit_wait+0x2f/0x2f
[11642.167931]  [<ffffffff8108afd1>] ? trace_hardirqs_on_caller+0x17b/0x197
[11642.182320]  [<ffffffff8108affa>] ? trace_hardirqs_on+0xd/0xf
[11642.183762]  [<ffffffff810b079b>] ? timekeeping_get_ns+0xe/0x33
[11642.185308]  [<ffffffff810b0f61>] ? ktime_get+0x41/0x52
[11642.186782]  [<ffffffff8147ac08>] io_schedule_timeout+0xa0/0x102
[11642.188217]  [<ffffffff8147ac08>] ? io_schedule_timeout+0xa0/0x102
[11642.189626]  [<ffffffff8147b814>] bit_wait_io+0x1b/0x39
[11642.190803]  [<ffffffff8147bb21>] __wait_on_bit_lock+0x4c/0x90
[11642.192158]  [<ffffffff8111829f>] __lock_page+0x66/0x68
[11642.193379]  [<ffffffff81082f29>] ? autoremove_wake_function+0x3a/0x3a
[11642.194831]  [<ffffffffa0450ddd>] lock_page+0x31/0x34 [btrfs]
[11642.197068]  [<ffffffffa0454e3b>] extent_write_cache_pages.isra.19.constprop.35+0x1af/0x2f4 [btrfs]
[11642.199188]  [<ffffffffa0455373>] extent_writepages+0x4b/0x5c [btrfs]
[11642.200723]  [<ffffffffa043c913>] ? btrfs_writepage_start_hook+0xce/0xce [btrfs]
[11642.202465]  [<ffffffffa043aa82>] btrfs_writepages+0x28/0x2a [btrfs]
[11642.203836]  [<ffffffff811236bc>] do_writepages+0x23/0x2c
[11642.205624]  [<ffffffff811198c9>] __filemap_fdatawrite_range+0x5a/0x61
[11642.207057]  [<ffffffff81119946>] filemap_fdatawrite_range+0x13/0x15
[11642.208529]  [<ffffffffa044f87e>] btrfs_start_ordered_extent+0xd0/0x1a1 [btrfs]
[11642.210375]  [<ffffffffa0462613>] ? btrfs_scrubparity_helper+0x140/0x33a [btrfs]
[11642.212132]  [<ffffffffa044f974>] btrfs_run_ordered_extent_work+0x25/0x34 [btrfs]
[11642.213837]  [<ffffffffa046262f>] btrfs_scrubparity_helper+0x15c/0x33a [btrfs]
[11642.215457]  [<ffffffffa046293b>] btrfs_flush_delalloc_helper+0xe/0x10 [btrfs]
[11642.217095]  [<ffffffff8106483e>] process_one_work+0x256/0x48b
[11642.218324]  [<ffffffff81064f20>] worker_thread+0x1f5/0x2a7
[11642.219466]  [<ffffffff81064d2b>] ? rescuer_thread+0x289/0x289
[11642.220801]  [<ffffffff8106a500>] kthread+0xd4/0xdc
[11642.222032]  [<ffffffff8106a42c>] ? kthread_parkme+0x24/0x24
[11642.223190]  [<ffffffff8147fdef>] ret_from_fork+0x3f/0x70
[11642.224394]  [<ffffffff8106a42c>] ? kthread_parkme+0x24/0x24
[11642.226295] 2 locks held by kworker/u32:3/15282:
[11642.227273]  #0:  ("%s-%s""btrfs", name){++++.+}, at: [<ffffffff8106474d>] process_one_work+0x165/0x48b
[11642.229412]  #1:  ((&work->normal_work)){+.+.+.}, at: [<ffffffff8106474d>] process_one_work+0x165/0x48b
[11642.231414] INFO: task kworker/u32:8:15289 blocked for more than 120 seconds.
[11642.232872]       Not tainted 4.4.0-rc6-btrfs-next-21+ #1
[11642.234109] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
[11642.235776] kworker/u32:8   D ffff88020de5f848     0 15289      2 0x00000000
[11642.237412] Workqueue: writeback wb_workfn (flush-btrfs-481)
[11642.238670]  ffff88020de5f848 0000000000000246 0000000000014ec0 ffff88023ed54ec0
[11642.240475]  ffff88021b1ece40 ffff88020de60000 ffff88023ed54ec0 7fffffffffffffff
[11642.242154]  0000000000000002 ffffffff8147b7f9 ffff88020de5f860 ffffffff8147b541
[11642.243715] Call Trace:
[11642.244390]  [<ffffffff8147b7f9>] ? bit_wait+0x2f/0x2f
[11642.245432]  [<ffffffff8147b541>] schedule+0x82/0x9a
[11642.246392]  [<ffffffff8147e7fe>] schedule_timeout+0x43/0x109
[11642.247479]  [<ffffffff8147b7f9>] ? bit_wait+0x2f/0x2f
[11642.248551]  [<ffffffff8108afd1>] ? trace_hardirqs_on_caller+0x17b/0x197
[11642.249968]  [<ffffffff8108affa>] ? trace_hardirqs_on+0xd/0xf
[11642.251043]  [<ffffffff810b079b>] ? timekeeping_get_ns+0xe/0x33
[11642.252202]  [<ffffffff810b0f61>] ? ktime_get+0x41/0x52
[11642.253210]  [<ffffffff8147ac08>] io_schedule_timeout+0xa0/0x102
[11642.254307]  [<ffffffff8147ac08>] ? io_schedule_timeout+0xa0/0x102
[11642.256118]  [<ffffffff8147b814>] bit_wait_io+0x1b/0x39
[11642.257131]  [<ffffffff8147bb21>] __wait_on_bit_lock+0x4c/0x90
[11642.258200]  [<ffffffff8111829f>] __lock_page+0x66/0x68
[11642.259168]  [<ffffffff81082f29>] ? autoremove_wake_function+0x3a/0x3a
[11642.260516]  [<ffffffffa0450ddd>] lock_page+0x31/0x34 [btrfs]
[11642.261841]  [<ffffffffa0454e3b>] extent_write_cache_pages.isra.19.constprop.35+0x1af/0x2f4 [btrfs]
[11642.263531]  [<ffffffffa0455373>] extent_writepages+0x4b/0x5c [btrfs]
[11642.264747]  [<ffffffffa043c913>] ? btrfs_writepage_start_hook+0xce/0xce [btrfs]
[11642.266148]  [<ffffffffa043aa82>] btrfs_writepages+0x28/0x2a [btrfs]
[11642.267264]  [<ffffffff811236bc>] do_writepages+0x23/0x2c
[11642.268280]  [<ffffffff81192a2b>] __writeback_single_inode+0xda/0x5ba
[11642.269407]  [<ffffffff811939f0>] writeback_sb_inodes+0x27b/0x43d
[11642.270476]  [<ffffffff81193c28>] __writeback_inodes_wb+0x76/0xae
[11642.271547]  [<ffffffff81193ea6>] wb_writeback+0x19e/0x41c
[11642.272588]  [<ffffffff81194821>] wb_workfn+0x201/0x341
[11642.273523]  [<ffffffff81194821>] ? wb_workfn+0x201/0x341
[11642.274479]  [<ffffffff8106483e>] process_one_work+0x256/0x48b
[11642.275497]  [<ffffffff81064f20>] worker_thread+0x1f5/0x2a7
[11642.276518]  [<ffffffff81064d2b>] ? rescuer_thread+0x289/0x289
[11642.277520]  [<ffffffff81064d2b>] ? rescuer_thread+0x289/0x289
[11642.278517]  [<ffffffff8106a500>] kthread+0xd4/0xdc
[11642.279371]  [<ffffffff8106a42c>] ? kthread_parkme+0x24/0x24
[11642.280468]  [<ffffffff8147fdef>] ret_from_fork+0x3f/0x70
[11642.281607]  [<ffffffff8106a42c>] ? kthread_parkme+0x24/0x24
[11642.282604] 3 locks held by kworker/u32:8/15289:
[11642.283423]  #0:  ("writeback"){++++.+}, at: [<ffffffff8106474d>] process_one_work+0x165/0x48b
[11642.285629]  #1:  ((&(&wb->dwork)->work)){+.+.+.}, at: [<ffffffff8106474d>] process_one_work+0x165/0x48b
[11642.287538]  #2:  (&type->s_umount_key#37){+++++.}, at: [<ffffffff81171217>] trylock_super+0x1b/0x4b
[11642.289423] INFO: task fdm-stress:26848 blocked for more than 120 seconds.
[11642.290547]       Not tainted 4.4.0-rc6-btrfs-next-21+ #1
[11642.291453] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
[11642.292864] fdm-stress      D ffff88022c107c20     0 26848  26591 0x00000000
[11642.294118]  ffff88022c107c20 000000038108affa 0000000000014ec0 ffff88023ed54ec0
[11642.295602]  ffff88013ab1ca40 ffff88022c108000 ffff8800b2fc19d0 00000000000e0fff
[11642.297098]  ffff8800b2fc19b0 ffff88022c107c88 ffff88022c107c38 ffffffff8147b541
[11642.298433] Call Trace:
[11642.298896]  [<ffffffff8147b541>] schedule+0x82/0x9a
[11642.299738]  [<ffffffffa045225d>] lock_extent_bits+0xfe/0x1a3 [btrfs]
[11642.300833]  [<ffffffff81082eef>] ? add_wait_queue_exclusive+0x44/0x44
[11642.301943]  [<ffffffffa0447516>] lock_and_cleanup_extent_if_need+0x68/0x18e [btrfs]
[11642.303270]  [<ffffffffa04485ba>] __btrfs_buffered_write+0x238/0x4c1 [btrfs]
[11642.304552]  [<ffffffffa044b50a>] ? btrfs_file_write_iter+0x17c/0x408 [btrfs]
[11642.305782]  [<ffffffffa044b682>] btrfs_file_write_iter+0x2f4/0x408 [btrfs]
[11642.306878]  [<ffffffff8116e298>] __vfs_write+0x7c/0xa5
[11642.307729]  [<ffffffff8116e7d1>] vfs_write+0x9d/0xe8
[11642.308602]  [<ffffffff8116efbb>] SyS_write+0x50/0x7e
[11642.309410]  [<ffffffff8147fa97>] entry_SYSCALL_64_fastpath+0x12/0x6b
[11642.310403] 3 locks held by fdm-stress/26848:
[11642.311108]  #0:  (&f->f_pos_lock){+.+.+.}, at: [<ffffffff811877e8>] __fdget_pos+0x3a/0x40
[11642.312578]  #1:  (sb_writers#11){.+.+.+}, at: [<ffffffff811706ee>] __sb_start_write+0x5f/0xb0
[11642.314170]  #2:  (&sb->s_type->i_mutex_key#15){+.+.+.}, at: [<ffffffffa044b401>] btrfs_file_write_iter+0x73/0x408 [btrfs]
[11642.316796] INFO: task fdm-stress:26849 blocked for more than 120 seconds.
[11642.317842]       Not tainted 4.4.0-rc6-btrfs-next-21+ #1
[11642.318691] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
[11642.319959] fdm-stress      D ffff8801964ffa68     0 26849  26591 0x00000000
[11642.321312]  ffff8801964ffa68 00ff8801e9975f80 0000000000014ec0 ffff88023ed94ec0
[11642.322555]  ffff8800b00b4840 ffff880196500000 ffff8801e9975f20 0000000000000002
[11642.323715]  ffff8801e9975f18 ffff8800b00b4840 ffff8801964ffa80 ffffffff8147b541
[11642.325096] Call Trace:
[11642.325532]  [<ffffffff8147b541>] schedule+0x82/0x9a
[11642.326303]  [<ffffffff8147e7fe>] schedule_timeout+0x43/0x109
[11642.327180]  [<ffffffff8108ae40>] ? mark_held_locks+0x5e/0x74
[11642.328114]  [<ffffffff8147f30e>] ? _raw_spin_unlock_irq+0x2c/0x4a
[11642.329051]  [<ffffffff8108afd1>] ? trace_hardirqs_on_caller+0x17b/0x197
[11642.330053]  [<ffffffff8147bceb>] __wait_for_common+0x109/0x147
[11642.330952]  [<ffffffff8147bceb>] ? __wait_for_common+0x109/0x147
[11642.331869]  [<ffffffff8147e7bb>] ? usleep_range+0x4a/0x4a
[11642.332925]  [<ffffffff81074075>] ? wake_up_q+0x47/0x47
[11642.333736]  [<ffffffff8147bd4d>] wait_for_completion+0x24/0x26
[11642.334672]  [<ffffffffa044f5ce>] btrfs_wait_ordered_extents+0x1c8/0x217 [btrfs]
[11642.335858]  [<ffffffffa0465b5a>] btrfs_mksubvol+0x224/0x45d [btrfs]
[11642.336854]  [<ffffffff81082eef>] ? add_wait_queue_exclusive+0x44/0x44
[11642.337820]  [<ffffffffa0465edb>] btrfs_ioctl_snap_create_transid+0x148/0x17a [btrfs]
[11642.339026]  [<ffffffffa046603b>] btrfs_ioctl_snap_create_v2+0xc7/0x110 [btrfs]
[11642.340214]  [<ffffffffa0468582>] btrfs_ioctl+0x590/0x27bd [btrfs]
[11642.341123]  [<ffffffff8147dc00>] ? mutex_unlock+0xe/0x10
[11642.341934]  [<ffffffffa00fa6e9>] ? ext4_file_write_iter+0x2a3/0x36f [ext4]
[11642.342936]  [<ffffffff8108895d>] ? __lock_is_held+0x3c/0x57
[11642.343772]  [<ffffffff81186a1d>] ? rcu_read_unlock+0x3e/0x5d
[11642.344673]  [<ffffffff8117dc95>] do_vfs_ioctl+0x458/0x4dc
[11642.346024]  [<ffffffff81186bbe>] ? __fget_light+0x62/0x71
[11642.346873]  [<ffffffff8117dd70>] SyS_ioctl+0x57/0x79
[11642.347720]  [<ffffffff8147fa97>] entry_SYSCALL_64_fastpath+0x12/0x6b
[11642.350222] 4 locks held by fdm-stress/26849:
[11642.350898]  #0:  (sb_writers#11){.+.+.+}, at: [<ffffffff811706ee>] __sb_start_write+0x5f/0xb0
[11642.352375]  #1:  (&type->i_mutex_dir_key#4/1){+.+.+.}, at: [<ffffffffa0465981>] btrfs_mksubvol+0x4b/0x45d [btrfs]
[11642.354072]  #2:  (&fs_info->subvol_sem){++++..}, at: [<ffffffffa0465a2a>] btrfs_mksubvol+0xf4/0x45d [btrfs]
[11642.355647]  #3:  (&root->ordered_extent_mutex){+.+...}, at: [<ffffffffa044f456>] btrfs_wait_ordered_extents+0x50/0x217 [btrfs]
[11642.357516] INFO: task fdm-stress:26850 blocked for more than 120 seconds.
[11642.358508]       Not tainted 4.4.0-rc6-btrfs-next-21+ #1
[11642.359376] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
[11642.368625] fdm-stress      D ffff88021f167688     0 26850  26591 0x00000000
[11642.369716]  ffff88021f167688 0000000000000001 0000000000014ec0 ffff88023edd4ec0
[11642.370950]  ffff880128a98680 ffff88021f168000 ffff88023edd4ec0 7fffffffffffffff
[11642.372210]  0000000000000002 ffffffff8147b7f9 ffff88021f1676a0 ffffffff8147b541
[11642.373430] Call Trace:
[11642.373853]  [<ffffffff8147b7f9>] ? bit_wait+0x2f/0x2f
[11642.374623]  [<ffffffff8147b541>] schedule+0x82/0x9a
[11642.375948]  [<ffffffff8147e7fe>] schedule_timeout+0x43/0x109
[11642.376862]  [<ffffffff8147b7f9>] ? bit_wait+0x2f/0x2f
[11642.377637]  [<ffffffff8108afd1>] ? trace_hardirqs_on_caller+0x17b/0x197
[11642.378610]  [<ffffffff8108affa>] ? trace_hardirqs_on+0xd/0xf
[11642.379457]  [<ffffffff810b079b>] ? timekeeping_get_ns+0xe/0x33
[11642.380366]  [<ffffffff810b0f61>] ? ktime_get+0x41/0x52
[11642.381353]  [<ffffffff8147ac08>] io_schedule_timeout+0xa0/0x102
[11642.382255]  [<ffffffff8147ac08>] ? io_schedule_timeout+0xa0/0x102
[11642.383162]  [<ffffffff8147b814>] bit_wait_io+0x1b/0x39
[11642.383945]  [<ffffffff8147bb21>] __wait_on_bit_lock+0x4c/0x90
[11642.384875]  [<ffffffff8111829f>] __lock_page+0x66/0x68
[11642.385749]  [<ffffffff81082f29>] ? autoremove_wake_function+0x3a/0x3a
[11642.386721]  [<ffffffffa0450ddd>] lock_page+0x31/0x34 [btrfs]
[11642.387596]  [<ffffffffa0454e3b>] extent_write_cache_pages.isra.19.constprop.35+0x1af/0x2f4 [btrfs]
[11642.389030]  [<ffffffffa0455373>] extent_writepages+0x4b/0x5c [btrfs]
[11642.389973]  [<ffffffff810a25ad>] ? rcu_read_lock_sched_held+0x61/0x69
[11642.390939]  [<ffffffffa043c913>] ? btrfs_writepage_start_hook+0xce/0xce [btrfs]
[11642.392271]  [<ffffffffa0451c32>] ? __clear_extent_bit+0x26e/0x2c0 [btrfs]
[11642.393305]  [<ffffffffa043aa82>] btrfs_writepages+0x28/0x2a [btrfs]
[11642.394239]  [<ffffffff811236bc>] do_writepages+0x23/0x2c
[11642.395045]  [<ffffffff811198c9>] __filemap_fdatawrite_range+0x5a/0x61
[11642.395991]  [<ffffffff81119946>] filemap_fdatawrite_range+0x13/0x15
[11642.397144]  [<ffffffffa044f87e>] btrfs_start_ordered_extent+0xd0/0x1a1 [btrfs]
[11642.398392]  [<ffffffffa0452094>] ? clear_extent_bit+0x17/0x19 [btrfs]
[11642.399363]  [<ffffffffa0445945>] btrfs_get_blocks_direct+0x12b/0x61c [btrfs]
[11642.400445]  [<ffffffff8119f7a1>] ? dio_bio_add_page+0x3d/0x54
[11642.401309]  [<ffffffff8119fa93>] ? submit_page_section+0x7b/0x111
[11642.402213]  [<ffffffff811a0258>] do_blockdev_direct_IO+0x685/0xc24
[11642.403139]  [<ffffffffa044581a>] ? btrfs_page_exists_in_range+0x1a1/0x1a1 [btrfs]
[11642.404360]  [<ffffffffa043d267>] ? btrfs_get_extent_fiemap+0x1c0/0x1c0 [btrfs]
[11642.406187]  [<ffffffff811a0828>] __blockdev_direct_IO+0x31/0x33
[11642.407070]  [<ffffffff811a0828>] ? __blockdev_direct_IO+0x31/0x33
[11642.407990]  [<ffffffffa043d267>] ? btrfs_get_extent_fiemap+0x1c0/0x1c0 [btrfs]
[11642.409192]  [<ffffffffa043b4ca>] btrfs_direct_IO+0x1c7/0x27e [btrfs]
[11642.410146]  [<ffffffffa043d267>] ? btrfs_get_extent_fiemap+0x1c0/0x1c0 [btrfs]
[11642.411291]  [<ffffffff81119a2c>] generic_file_read_iter+0x89/0x4e1
[11642.412263]  [<ffffffff8108ac05>] ? mark_lock+0x24/0x201
[11642.413057]  [<ffffffff8116e1f8>] __vfs_read+0x79/0x9d
[11642.413897]  [<ffffffff8116e6f1>] vfs_read+0x8f/0xd2
[11642.414708]  [<ffffffff8116ef3d>] SyS_read+0x50/0x7e
[11642.415573]  [<ffffffff8147fa97>] entry_SYSCALL_64_fastpath+0x12/0x6b
[11642.416572] 1 lock held by fdm-stress/26850:
[11642.417345]  #0:  (&f->f_pos_lock){+.+.+.}, at: [<ffffffff811877e8>] __fdget_pos+0x3a/0x40
[11642.418703] INFO: task fdm-stress:26851 blocked for more than 120 seconds.
[11642.419698]       Not tainted 4.4.0-rc6-btrfs-next-21+ #1
[11642.420612] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
[11642.421807] fdm-stress      D ffff880196483d28     0 26851  26591 0x00000000
[11642.422878]  ffff880196483d28 00ff8801c8f60740 0000000000014ec0 ffff88023ed94ec0
[11642.424149]  ffff8801c8f60740 ffff880196484000 0000000000000246 ffff8801c8f60740
[11642.425374]  ffff8801bb711840 ffff8801bb711878 ffff880196483d40 ffffffff8147b541
[11642.426591] Call Trace:
[11642.427013]  [<ffffffff8147b541>] schedule+0x82/0x9a
[11642.427856]  [<ffffffff8147b6d5>] schedule_preempt_disabled+0x18/0x24
[11642.428852]  [<ffffffff8147c23a>] mutex_lock_nested+0x1d7/0x3b4
[11642.429743]  [<ffffffffa044f456>] ? btrfs_wait_ordered_extents+0x50/0x217 [btrfs]
[11642.430911]  [<ffffffffa044f456>] btrfs_wait_ordered_extents+0x50/0x217 [btrfs]
[11642.432102]  [<ffffffffa044f674>] ? btrfs_wait_ordered_roots+0x57/0x191 [btrfs]
[11642.433259]  [<ffffffffa044f456>] ? btrfs_wait_ordered_extents+0x50/0x217 [btrfs]
[11642.434431]  [<ffffffffa044f6ea>] btrfs_wait_ordered_roots+0xcd/0x191 [btrfs]
[11642.436079]  [<ffffffffa0410cab>] btrfs_sync_fs+0xe0/0x1ad [btrfs]
[11642.437009]  [<ffffffff81197900>] ? SyS_tee+0x23c/0x23c
[11642.437860]  [<ffffffff81197920>] sync_fs_one_sb+0x20/0x22
[11642.438723]  [<ffffffff81171435>] iterate_supers+0x75/0xc2
[11642.439597]  [<ffffffff81197d00>] sys_sync+0x52/0x80
[11642.440454]  [<ffffffff8147fa97>] entry_SYSCALL_64_fastpath+0x12/0x6b
[11642.441533] 3 locks held by fdm-stress/26851:
[11642.442370]  #0:  (&type->s_umount_key#37){+++++.}, at: [<ffffffff8117141f>] iterate_supers+0x5f/0xc2
[11642.444043]  #1:  (&fs_info->ordered_operations_mutex){+.+...}, at: [<ffffffffa044f661>] btrfs_wait_ordered_roots+0x44/0x191 [btrfs]
[11642.446010]  #2:  (&root->ordered_extent_mutex){+.+...}, at: [<ffffffffa044f456>] btrfs_wait_ordered_extents+0x50/0x217 [btrfs]

This happened because under specific timings the path for direct IO reads
can deadlock with concurrent buffered writes. The diagram below shows how
this happens for an example file that has the following layout:

     [  extent A  ]  [  extent B  ]  [ ....
     0K              4K              8K

     CPU 1                                               CPU 2                             CPU 3

DIO read against range
 [0K, 8K[ starts

btrfs_direct_IO()
  --> calls btrfs_get_blocks_direct()
      which finds the extent map for the
      extent A and leaves the range
      [0K, 4K[ locked in the inode's
      io tree

                                                   buffered write against
                                                   range [4K, 8K[ starts

                                                   __btrfs_buffered_write()
                                                     --> dirties page at 4K

                                                                                     a user space
                                                                                     task calls sync
                                                                                     for e.g or
                                                                                     writepages() is
                                                                                     invoked by mm

                                                                                     writepages()
                                                                                       run_delalloc_range()
                                                                                         cow_file_range()
                                                                                           --> ordered extent X
                                                                                               for the buffered
                                                                                               write is created
                                                                                               and
                                                                                               writeback starts

  --> calls btrfs_get_blocks_direct()
      again, without submitting first
      a bio for reading extent A, and
      finds the extent map for extent B

  --> calls lock_extent_direct()

      --> locks range [4K, 8K[
      --> finds ordered extent X
          covering range [4K, 8K[
      --> unlocks range [4K, 8K[

                                                  buffered write against
                                                  range [0K, 8K[ starts

                                                  __btrfs_buffered_write()
                                                    prepare_pages()
                                                      --> locks pages with
                                                          offsets 0 and 4K
                                                    lock_and_cleanup_extent_if_need()
                                                      --> blocks attempting to
                                                          lock range [0K, 8K[ in
                                                          the inode's io tree,
                                                          because the range [0, 4K[
                                                          is already locked by the
                                                          direct IO task at CPU 1

      --> calls
          btrfs_start_ordered_extent(oe X)

          btrfs_start_ordered_extent(oe X)

            --> At this point writeback for ordered
                extent X has not finished yet

            filemap_fdatawrite_range()
              btrfs_writepages()
                extent_writepages()
                  extent_write_cache_pages()
                    --> finds page with offset 0
                        with the writeback tag
                        (and not dirty)
                    --> tries to lock it
                         --> deadlock, task at CPU 2
                             has the page locked and
                             is blocked on the io range
                             [0, 4K[ that was locked
                             earlier by this task

So fix this by falling back to a buffered read in the direct IO read path
when an ordered extent for a buffered write is found.

Signed-off-by: Filipe Manana <fdmanana@suse.com>
Reviewed-by: Liu Bo <bo.li.liu@oracle.com>
Signed-off-by: Chris Mason <clm@fb.com>
fs/btrfs/inode.c