btrfs: fix lock inversion problem when doing qgroup extent tracing
authorFilipe Manana <fdmanana@suse.com>
Wed, 21 Jul 2021 16:31:48 +0000 (17:31 +0100)
committerDavid Sterba <dsterba@suse.com>
Thu, 22 Jul 2021 13:50:07 +0000 (15:50 +0200)
commit8949b9a114019b03fbd0d03d65b8647cba4feef3
tree18f002084ce4a3d35f21c44160b42e89c0169069
parent16a200f66ede3f9afa2e51d90ade017aaa18d213
btrfs: fix lock inversion problem when doing qgroup extent tracing

At btrfs_qgroup_trace_extent_post() we call btrfs_find_all_roots() with a
NULL value as the transaction handle argument, which makes that function
take the commit_root_sem semaphore, which is necessary when we don't hold
a transaction handle or any other mechanism to prevent a transaction
commit from wiping out commit roots.

However btrfs_qgroup_trace_extent_post() can be called in a context where
we are holding a write lock on an extent buffer from a subvolume tree,
namely from btrfs_truncate_inode_items(), called either during truncate
or unlink operations. In this case we end up with a lock inversion problem
because the commit_root_sem is a higher level lock, always supposed to be
acquired before locking any extent buffer.

Lockdep detects this lock inversion problem since we switched the extent
buffer locks from custom locks to semaphores, and when running btrfs/158
from fstests, it reported the following trace:

[ 9057.626435] ======================================================
[ 9057.627541] WARNING: possible circular locking dependency detected
[ 9057.628334] 5.14.0-rc2-btrfs-next-93 #1 Not tainted
[ 9057.628961] ------------------------------------------------------
[ 9057.629867] kworker/u16:4/30781 is trying to acquire lock:
[ 9057.630824] ffff8e2590f58760 (btrfs-tree-00){++++}-{3:3}, at: __btrfs_tree_read_lock+0x24/0x110 [btrfs]
[ 9057.632542]
               but task is already holding lock:
[ 9057.633551] ffff8e25582d4b70 (&fs_info->commit_root_sem){++++}-{3:3}, at: iterate_extent_inodes+0x10b/0x280 [btrfs]
[ 9057.635255]
               which lock already depends on the new lock.

[ 9057.636292]
               the existing dependency chain (in reverse order) is:
[ 9057.637240]
               -> #1 (&fs_info->commit_root_sem){++++}-{3:3}:
[ 9057.638138]        down_read+0x46/0x140
[ 9057.638648]        btrfs_find_all_roots+0x41/0x80 [btrfs]
[ 9057.639398]        btrfs_qgroup_trace_extent_post+0x37/0x70 [btrfs]
[ 9057.640283]        btrfs_add_delayed_data_ref+0x418/0x490 [btrfs]
[ 9057.641114]        btrfs_free_extent+0x35/0xb0 [btrfs]
[ 9057.641819]        btrfs_truncate_inode_items+0x424/0xf70 [btrfs]
[ 9057.642643]        btrfs_evict_inode+0x454/0x4f0 [btrfs]
[ 9057.643418]        evict+0xcf/0x1d0
[ 9057.643895]        do_unlinkat+0x1e9/0x300
[ 9057.644525]        do_syscall_64+0x3b/0xc0
[ 9057.645110]        entry_SYSCALL_64_after_hwframe+0x44/0xae
[ 9057.645835]
               -> #0 (btrfs-tree-00){++++}-{3:3}:
[ 9057.646600]        __lock_acquire+0x130e/0x2210
[ 9057.647248]        lock_acquire+0xd7/0x310
[ 9057.647773]        down_read_nested+0x4b/0x140
[ 9057.648350]        __btrfs_tree_read_lock+0x24/0x110 [btrfs]
[ 9057.649175]        btrfs_read_lock_root_node+0x31/0x40 [btrfs]
[ 9057.650010]        btrfs_search_slot+0x537/0xc00 [btrfs]
[ 9057.650849]        scrub_print_warning_inode+0x89/0x370 [btrfs]
[ 9057.651733]        iterate_extent_inodes+0x1e3/0x280 [btrfs]
[ 9057.652501]        scrub_print_warning+0x15d/0x2f0 [btrfs]
[ 9057.653264]        scrub_handle_errored_block.isra.0+0x135f/0x1640 [btrfs]
[ 9057.654295]        scrub_bio_end_io_worker+0x101/0x2e0 [btrfs]
[ 9057.655111]        btrfs_work_helper+0xf8/0x400 [btrfs]
[ 9057.655831]        process_one_work+0x247/0x5a0
[ 9057.656425]        worker_thread+0x55/0x3c0
[ 9057.656993]        kthread+0x155/0x180
[ 9057.657494]        ret_from_fork+0x22/0x30
[ 9057.658030]
               other info that might help us debug this:

[ 9057.659064]  Possible unsafe locking scenario:

[ 9057.659824]        CPU0                    CPU1
[ 9057.660402]        ----                    ----
[ 9057.660988]   lock(&fs_info->commit_root_sem);
[ 9057.661581]                                lock(btrfs-tree-00);
[ 9057.662348]                                lock(&fs_info->commit_root_sem);
[ 9057.663254]   lock(btrfs-tree-00);
[ 9057.663690]
                *** DEADLOCK ***

[ 9057.664437] 4 locks held by kworker/u16:4/30781:
[ 9057.665023]  #0: ffff8e25922a1148 ((wq_completion)btrfs-scrub){+.+.}-{0:0}, at: process_one_work+0x1c7/0x5a0
[ 9057.666260]  #1: ffffabb3451ffe70 ((work_completion)(&work->normal_work)){+.+.}-{0:0}, at: process_one_work+0x1c7/0x5a0
[ 9057.667639]  #2: ffff8e25922da198 (&ret->mutex){+.+.}-{3:3}, at: scrub_handle_errored_block.isra.0+0x5d2/0x1640 [btrfs]
[ 9057.669017]  #3: ffff8e25582d4b70 (&fs_info->commit_root_sem){++++}-{3:3}, at: iterate_extent_inodes+0x10b/0x280 [btrfs]
[ 9057.670408]
               stack backtrace:
[ 9057.670976] CPU: 7 PID: 30781 Comm: kworker/u16:4 Not tainted 5.14.0-rc2-btrfs-next-93 #1
[ 9057.672030] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.14.0-0-g155821a1990b-prebuilt.qemu.org 04/01/2014
[ 9057.673492] Workqueue: btrfs-scrub btrfs_work_helper [btrfs]
[ 9057.674258] Call Trace:
[ 9057.674588]  dump_stack_lvl+0x57/0x72
[ 9057.675083]  check_noncircular+0xf3/0x110
[ 9057.675611]  __lock_acquire+0x130e/0x2210
[ 9057.676132]  lock_acquire+0xd7/0x310
[ 9057.676605]  ? __btrfs_tree_read_lock+0x24/0x110 [btrfs]
[ 9057.677313]  ? lock_is_held_type+0xe8/0x140
[ 9057.677849]  down_read_nested+0x4b/0x140
[ 9057.678349]  ? __btrfs_tree_read_lock+0x24/0x110 [btrfs]
[ 9057.679068]  __btrfs_tree_read_lock+0x24/0x110 [btrfs]
[ 9057.679760]  btrfs_read_lock_root_node+0x31/0x40 [btrfs]
[ 9057.680458]  btrfs_search_slot+0x537/0xc00 [btrfs]
[ 9057.681083]  ? _raw_spin_unlock+0x29/0x40
[ 9057.681594]  ? btrfs_find_all_roots_safe+0x11f/0x140 [btrfs]
[ 9057.682336]  scrub_print_warning_inode+0x89/0x370 [btrfs]
[ 9057.683058]  ? btrfs_find_all_roots_safe+0x11f/0x140 [btrfs]
[ 9057.683834]  ? scrub_write_block_to_dev_replace+0xb0/0xb0 [btrfs]
[ 9057.684632]  iterate_extent_inodes+0x1e3/0x280 [btrfs]
[ 9057.685316]  scrub_print_warning+0x15d/0x2f0 [btrfs]
[ 9057.685977]  ? ___ratelimit+0xa4/0x110
[ 9057.686460]  scrub_handle_errored_block.isra.0+0x135f/0x1640 [btrfs]
[ 9057.687316]  scrub_bio_end_io_worker+0x101/0x2e0 [btrfs]
[ 9057.688021]  btrfs_work_helper+0xf8/0x400 [btrfs]
[ 9057.688649]  ? lock_is_held_type+0xe8/0x140
[ 9057.689180]  process_one_work+0x247/0x5a0
[ 9057.689696]  worker_thread+0x55/0x3c0
[ 9057.690175]  ? process_one_work+0x5a0/0x5a0
[ 9057.690731]  kthread+0x155/0x180
[ 9057.691158]  ? set_kthread_struct+0x40/0x40
[ 9057.691697]  ret_from_fork+0x22/0x30

Fix this by making btrfs_find_all_roots() never attempt to lock the
commit_root_sem when it is called from btrfs_qgroup_trace_extent_post().

We can't just pass a non-NULL transaction handle to btrfs_find_all_roots()
from btrfs_qgroup_trace_extent_post(), because that would make backref
lookup not use commit roots and acquire read locks on extent buffers, and
therefore could deadlock when btrfs_qgroup_trace_extent_post() is called
from the btrfs_truncate_inode_items() code path which has acquired a write
lock on an extent buffer of the subvolume btree.

CC: stable@vger.kernel.org # 4.19+
Reviewed-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
fs/btrfs/backref.c
fs/btrfs/backref.h
fs/btrfs/delayed-ref.c
fs/btrfs/qgroup.c
fs/btrfs/qgroup.h
fs/btrfs/tests/qgroup-tests.c