Btrfs: fix hang during inode eviction due to concurrent readahead

Btrfs: fix hang during inode eviction due to concurrent readahead

Zygo Blaxell and other users have reported occasional hangs while an
inode is being evicted, leading to traces like the following:

[ 5281.972322] INFO: task rm:20488 blocked for more than 120 seconds.
[ 5281.973836]       Not tainted 4.0.0-rc5-btrfs-next-9+ #2
[ 5281.974818] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
[ 5281.976364] rm              D ffff8800724cfc38     0 20488   7747 0x00000000
[ 5281.977506]  ffff8800724cfc38 ffff8800724cfc38 ffff880065da5c50 0000000000000001
[ 5281.978461]  ffff8800724cffd8 ffff8801540a5f50 0000000000000008 ffff8801540a5f78
[ 5281.979541]  ffff8801540a5f50 ffff8800724cfc58 ffffffff8143107e 0000000000000123
[ 5281.981396] Call Trace:
[ 5281.982066]  [<ffffffff8143107e>] schedule+0x74/0x83
[ 5281.983341]  [<ffffffffa03b33cf>] wait_on_state+0xac/0xcd [btrfs]
[ 5281.985127]  [<ffffffff81075cd6>] ? signal_pending_state+0x31/0x31
[ 5281.986715]  [<ffffffffa03b4b71>] wait_extent_bit.constprop.32+0x7c/0xde [btrfs]
[ 5281.988680]  [<ffffffffa03b540b>] lock_extent_bits+0x5d/0x88 [btrfs]
[ 5281.990200]  [<ffffffffa03a621d>] btrfs_evict_inode+0x24e/0x5be [btrfs]
[ 5281.991781]  [<ffffffff8116964d>] evict+0xa0/0x148
[ 5281.992735]  [<ffffffff8116a43d>] iput+0x18f/0x1e5
[ 5281.993796]  [<ffffffff81160d4a>] do_unlinkat+0x15b/0x1fa
[ 5281.994806]  [<ffffffff81435b54>] ? ret_from_sys_call+0x1d/0x58
[ 5281.996120]  [<ffffffff8107d314>] ? trace_hardirqs_on_caller+0x18f/0x1ab
[ 5281.997562]  [<ffffffff8123960b>] ? trace_hardirqs_on_thunk+0x3a/0x3f
[ 5281.998815]  [<ffffffff81161a16>] SyS_unlinkat+0x29/0x2b
[ 5281.999920]  [<ffffffff81435b32>] system_call_fastpath+0x12/0x17
[ 5282.001299] 1 lock held by rm/20488:
[ 5282.002066]  #0:  (sb_writers#12){.+.+.+}, at: [<ffffffff8116dd81>] mnt_want_write+0x24/0x4b

This happens when we have readahead, which calls readpages(), happening
right before the inode eviction handler is invoked. So the reason is
essentially:

1) readpages() is called while a reference on the inode is held, so
   eviction can not be triggered before readpages() returns. It also
   locks one or more ranges in the inode's io_tree (which is done at
   extent_io.c:__do_contiguous_readpages());

2) readpages() submits several read bios, all with an end io callback
   that runs extent_io.c:end_bio_extent_readpage() and that is executed
   by other task when a bio finishes, corresponding to a work queue
   (fs_info->end_io_workers) worker kthread. This callback unlocks
   the ranges in the inode's io_tree that were previously locked in
   step 1;

3) readpages() returns, the reference on the inode is dropped;

4) One or more of the read bios previously submitted are still not
   complete (their end io callback was not yet invoked or has not
   yet finished execution);

5) Inode eviction is triggered (through an unlink call for example).
   The inode reference count was not incremented before submitting
   the read bios, therefore this is possible;

6) The eviction handler starts executing and enters the loop that
   iterates over all extent states in the inode's io_tree;

7) The loop picks one extent state record and uses its ->start and
   ->end fields, after releasing the inode's io_tree spinlock, to
   call lock_extent_bits() and clear_extent_bit(). The call to lock
   the range [state->start, state->end] blocks because the whole
   range or a part of it was locked by the previous call to
   readpages() and the corresponding end io callback, which unlocks
   the range was not yet executed;

8) The end io callback for the read bio is executed and unlocks the
   range [state->start, state->end] (or a superset of that range).
   And at clear_extent_bit() the extent_state record state is used
   as a second argument to split_state(), which sets state->start to
   a larger value;

9) The task executing the eviction handler is woken up by the task
   executing the bio's end io callback (through clear_state_bit) and
   the eviction handler locks the range
   [old value for state->start, state->end]. Shortly after, when
   calling clear_extent_bit(), it unlocks the range
   [new value for state->start, state->end], so it ends up unlocking
   only part of the range that it locked, leaving an extent state
   record in the io_tree that represents the unlocked subrange;

10) The eviction handler loop, in its next iteration, gets the
    extent_state record for the subrange that it did not unlock in the
    previous step and then tries to lock it, resulting in an hang.

So fix this by not using the ->start and ->end fields of an existing
extent_state record. This is a simple solution, and an alternative
could be to bump the inode's reference count before submitting each
read bio and having it dropped in the bio's end io callback. But that
would be a more invasive/complex change and would not protect against
other possible places that are not holding a reference on the inode
as well. Something to consider in the future.

Many thanks to Zygo Blaxell for reporting, in the mailing list, the
issue, a set of scripts to trigger it and testing this fix.

Reported-by: Zygo Blaxell <ce3g8jdj@umail.furryterror.org>
Tested-by: Zygo Blaxell <ce3g8jdj@umail.furryterror.org>
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>