Re: [PATCH] btrfs: don't take dev_replace rwsem on task already holding it

[Johannes Thumshirn <Johannes.Thumshirn@wdc.com>]

On 26.08.24 13:11, Filipe Manana wrote:
> On Thu, Aug 15, 2024 at 1:57 PM Johannes Thumshirn <jth@kernel.org> wrote:
>>
>> From: Johannes Thumshirn <johannes.thumshirn@wdc.com>
>>
>> Running fstests btrfs/011 with MKFS_OPTIONS="-O rst" to force the usage of
>> the RAID stripe-tree, we get the following splat from lockdep:
>>
>>   BTRFS info (device sdd): dev_replace from /dev/sdd (devid 1) to /dev/sdb started
>>
>>   ============================================
>>   WARNING: possible recursive locking detected
>>   6.11.0-rc3-btrfs-for-next #599 Not tainted
>>   --------------------------------------------
>>   btrfs/2326 is trying to acquire lock:
>>   ffff88810f215c98 (&fs_info->dev_replace.rwsem){++++}-{3:3}, at: btrfs_map_block+0x39f/0x2250
>>
>>   but task is already holding lock:
>>   ffff88810f215c98 (&fs_info->dev_replace.rwsem){++++}-{3:3}, at: btrfs_map_block+0x39f/0x2250
>>
>>   other info that might help us debug this:
>>    Possible unsafe locking scenario:
>>
>>          CPU0
>>          ----
>>     lock(&fs_info->dev_replace.rwsem);
>>     lock(&fs_info->dev_replace.rwsem);
>>
>>    *** DEADLOCK ***
>>
>>    May be due to missing lock nesting notation
>>
>>   1 lock held by btrfs/2326:
>>    #0: ffff88810f215c98 (&fs_info->dev_replace.rwsem){++++}-{3:3}, at: btrfs_map_block+0x39f/0x2250
>>
>>   stack backtrace:
>>   CPU: 1 UID: 0 PID: 2326 Comm: btrfs Not tainted 6.11.0-rc3-btrfs-for-next #599
>>   Hardware name: Bochs Bochs, BIOS Bochs 01/01/2011
>>   Call Trace:
>>    <TASK>
>>    dump_stack_lvl+0x5b/0x80
>>    __lock_acquire+0x2798/0x69d0
>>    ? __pfx___lock_acquire+0x10/0x10
>>    ? __pfx___lock_acquire+0x10/0x10
>>    lock_acquire+0x19d/0x4a0
>>    ? btrfs_map_block+0x39f/0x2250
>>    ? __pfx_lock_acquire+0x10/0x10
>>    ? find_held_lock+0x2d/0x110
>>    ? lock_is_held_type+0x8f/0x100
>>    down_read+0x8e/0x440
>>    ? btrfs_map_block+0x39f/0x2250
>>    ? __pfx_down_read+0x10/0x10
>>    ? do_raw_read_unlock+0x44/0x70
>>    ? _raw_read_unlock+0x23/0x40
>>    btrfs_map_block+0x39f/0x2250
>>    ? btrfs_dev_replace_by_ioctl+0xd69/0x1d00
>>    ? btrfs_bio_counter_inc_blocked+0xd9/0x2e0
>>    ? __kasan_slab_alloc+0x6e/0x70
>>    ? __pfx_btrfs_map_block+0x10/0x10
>>    ? __pfx_btrfs_bio_counter_inc_blocked+0x10/0x10
>>    ? kmem_cache_alloc_noprof+0x1f2/0x300
>>    ? mempool_alloc_noprof+0xed/0x2b0
>>    btrfs_submit_chunk+0x28d/0x17e0
>>    ? __pfx_btrfs_submit_chunk+0x10/0x10
>>    ? bvec_alloc+0xd7/0x1b0
>>    ? bio_add_folio+0x171/0x270
>>    ? __pfx_bio_add_folio+0x10/0x10
>>    ? __kasan_check_read+0x20/0x20
>>    btrfs_submit_bio+0x37/0x80
>>    read_extent_buffer_pages+0x3df/0x6c0
>>    btrfs_read_extent_buffer+0x13e/0x5f0
>>    read_tree_block+0x81/0xe0
>>    read_block_for_search+0x4bd/0x7a0
>>    ? __pfx_read_block_for_search+0x10/0x10
>>    btrfs_search_slot+0x78d/0x2720
>>    ? __pfx_btrfs_search_slot+0x10/0x10
>>    ? lock_is_held_type+0x8f/0x100
>>    ? kasan_save_track+0x14/0x30
>>    ? __kasan_slab_alloc+0x6e/0x70
>>    ? kmem_cache_alloc_noprof+0x1f2/0x300
>>    btrfs_get_raid_extent_offset+0x181/0x820
>>    ? __pfx_lock_acquire+0x10/0x10
>>    ? __pfx_btrfs_get_raid_extent_offset+0x10/0x10
>>    ? down_read+0x194/0x440
>>    ? __pfx_down_read+0x10/0x10
>>    ? do_raw_read_unlock+0x44/0x70
>>    ? _raw_read_unlock+0x23/0x40
>>    btrfs_map_block+0x5b5/0x2250
>>    ? __pfx_btrfs_map_block+0x10/0x10
>>    scrub_submit_initial_read+0x8fe/0x11b0
>>    ? __pfx_scrub_submit_initial_read+0x10/0x10
>>    submit_initial_group_read+0x161/0x3a0
>>    ? lock_release+0x20e/0x710
>>    ? __pfx_submit_initial_group_read+0x10/0x10
>>    ? __pfx_lock_release+0x10/0x10
>>    scrub_simple_mirror.isra.0+0x3eb/0x580
>>    scrub_stripe+0xe4d/0x1440
>>    ? lock_release+0x20e/0x710
>>    ? __pfx_scrub_stripe+0x10/0x10
>>    ? __pfx_lock_release+0x10/0x10
>>    ? do_raw_read_unlock+0x44/0x70
>>    ? _raw_read_unlock+0x23/0x40
>>    scrub_chunk+0x257/0x4a0
>>    scrub_enumerate_chunks+0x64c/0xf70
>>    ? __mutex_unlock_slowpath+0x147/0x5f0
>>    ? __pfx_scrub_enumerate_chunks+0x10/0x10
>>    ? bit_wait_timeout+0xb0/0x170
>>    ? __up_read+0x189/0x700
>>    ? scrub_workers_get+0x231/0x300
>>    ? up_write+0x490/0x4f0
>>    btrfs_scrub_dev+0x52e/0xcd0
>>    ? create_pending_snapshots+0x230/0x250
>>    ? __pfx_btrfs_scrub_dev+0x10/0x10
>>    btrfs_dev_replace_by_ioctl+0xd69/0x1d00
>>    ? lock_acquire+0x19d/0x4a0
>>    ? __pfx_btrfs_dev_replace_by_ioctl+0x10/0x10
>>    ? lock_release+0x20e/0x710
>>    ? btrfs_ioctl+0xa09/0x74f0
>>    ? __pfx_lock_release+0x10/0x10
>>    ? do_raw_spin_lock+0x11e/0x240
>>    ? __pfx_do_raw_spin_lock+0x10/0x10
>>    btrfs_ioctl+0xa14/0x74f0
>>    ? lock_acquire+0x19d/0x4a0
>>    ? find_held_lock+0x2d/0x110
>>    ? __pfx_btrfs_ioctl+0x10/0x10
>>    ? lock_release+0x20e/0x710
>>    ? do_sigaction+0x3f0/0x860
>>    ? __pfx_do_vfs_ioctl+0x10/0x10
>>    ? do_raw_spin_lock+0x11e/0x240
>>    ? lockdep_hardirqs_on_prepare+0x270/0x3e0
>>    ? _raw_spin_unlock_irq+0x28/0x50
>>    ? do_sigaction+0x3f0/0x860
>>    ? __pfx_do_sigaction+0x10/0x10
>>    ? __x64_sys_rt_sigaction+0x18e/0x1e0
>>    ? __pfx___x64_sys_rt_sigaction+0x10/0x10
>>    ? __x64_sys_close+0x7c/0xd0
>>    __x64_sys_ioctl+0x137/0x190
>>    do_syscall_64+0x71/0x140
>>    entry_SYSCALL_64_after_hwframe+0x76/0x7e
>>   RIP: 0033:0x7f0bd1114f9b
>>   Code: Unable to access opcode bytes at 0x7f0bd1114f71.
>>   RSP: 002b:00007ffc8a8c3130 EFLAGS: 00000246 ORIG_RAX: 0000000000000010
>>   RAX: ffffffffffffffda RBX: 0000000000000003 RCX: 00007f0bd1114f9b
>>   RDX: 00007ffc8a8c35e0 RSI: 00000000ca289435 RDI: 0000000000000003
>>   RBP: 0000000000000000 R08: 0000000000000000 R09: 0000000000000007
>>   R10: 0000000000000008 R11: 0000000000000246 R12: 00007ffc8a8c6c85
>>   R13: 00000000398e72a0 R14: 0000000000004361 R15: 0000000000000004
>>    </TASK>
>>
>> This happens because on RAID stripe-tree filesystems we recurse back into
>> btrfs_map_block() on scrub to perform the logical to device physical
>> mapping.
>>
>> But as the device replace task is already holding the dev_replace::rwsem
>> we deadlock.
>>
>> So don't take the dev_replace::rwsem in case our task is the task performing
>> the device replace.
>>
>> Suggested-by: Filipe Manana <fdmanana@suse.com>
>> Signed-off-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
>> ---
>>   fs/btrfs/dev-replace.c | 2 ++
>>   fs/btrfs/fs.h          | 2 ++
>>   fs/btrfs/volumes.c     | 4 +++-
>>   3 files changed, 7 insertions(+), 1 deletion(-)
>>
>> diff --git a/fs/btrfs/dev-replace.c b/fs/btrfs/dev-replace.c
>> index 83d5cdd77f29..604399e59a3d 100644
>> --- a/fs/btrfs/dev-replace.c
>> +++ b/fs/btrfs/dev-replace.c
>> @@ -641,6 +641,7 @@ static int btrfs_dev_replace_start(struct btrfs_fs_info *fs_info,
>>                  return ret;
>>
>>          down_write(&dev_replace->rwsem);
>> +       dev_replace->replace_task = current;
>>          switch (dev_replace->replace_state) {
>>          case BTRFS_IOCTL_DEV_REPLACE_STATE_NEVER_STARTED:
>>          case BTRFS_IOCTL_DEV_REPLACE_STATE_FINISHED:
>> @@ -994,6 +995,7 @@ static int btrfs_dev_replace_finishing(struct btrfs_fs_info *fs_info,
>>          list_add(&tgt_device->dev_alloc_list, &fs_devices->alloc_list);
>>          fs_devices->rw_devices++;
>>
>> +       dev_replace->replace_task = NULL;
>>          up_write(&dev_replace->rwsem);
>>          btrfs_rm_dev_replace_blocked(fs_info);
>>
>> diff --git a/fs/btrfs/fs.h b/fs/btrfs/fs.h
>> index 3d6d4b503220..53824da92cc3 100644
>> --- a/fs/btrfs/fs.h
>> +++ b/fs/btrfs/fs.h
>> @@ -317,6 +317,8 @@ struct btrfs_dev_replace {
>>
>>          struct percpu_counter bio_counter;
>>          wait_queue_head_t replace_wait;
>> +
>> +       struct task_struct *replace_task;
>>   };
>>
>>   /*
>> diff --git a/fs/btrfs/volumes.c b/fs/btrfs/volumes.c
>> index 4b9b647a7e29..d2e80a1f258d 100644
>> --- a/fs/btrfs/volumes.c
>> +++ b/fs/btrfs/volumes.c
>> @@ -6481,7 +6481,9 @@ int btrfs_map_block(struct btrfs_fs_info *fs_info, enum btrfs_map_op op,
>>          *length = min_t(u64, map->chunk_len - map_offset, max_len);
>>
>>          down_read(&dev_replace->rwsem);
>> -       dev_replace_is_ongoing = btrfs_dev_replace_is_ongoing(dev_replace);
>> +       if (btrfs_dev_replace_is_ongoing(dev_replace) && dev_replace->replace_task != current)
>> +               dev_replace_is_ongoing = 1;
> 
> So the idea I suggested was not like this.
> The idea was just to not ever down_read() the semaphore if the current
> task is the replace task.

So your idea is sth like the following (not yet tested):

diff --git a/fs/btrfs/dev-replace.c b/fs/btrfs/dev-replace.c
index 83d5cdd77f29..604399e59a3d 100644
--- a/fs/btrfs/dev-replace.c
+++ b/fs/btrfs/dev-replace.c
@@ -641,6 +641,7 @@ static int btrfs_dev_replace_start(struct 
btrfs_fs_info *fs_info,
                 return ret;

         down_write(&dev_replace->rwsem);
+       dev_replace->replace_task = current;
         switch (dev_replace->replace_state) {
         case BTRFS_IOCTL_DEV_REPLACE_STATE_NEVER_STARTED:
         case BTRFS_IOCTL_DEV_REPLACE_STATE_FINISHED:
@@ -994,6 +995,7 @@ static int btrfs_dev_replace_finishing(struct 
btrfs_fs_info *fs_info,
         list_add(&tgt_device->dev_alloc_list, &fs_devices->alloc_list);
         fs_devices->rw_devices++;

+       dev_replace->replace_task = NULL;
         up_write(&dev_replace->rwsem);
         btrfs_rm_dev_replace_blocked(fs_info);

diff --git a/fs/btrfs/fs.h b/fs/btrfs/fs.h
index 79f64e383edd..cbfb225858a5 100644
--- a/fs/btrfs/fs.h
+++ b/fs/btrfs/fs.h
@@ -317,6 +317,8 @@ struct btrfs_dev_replace {

         struct percpu_counter bio_counter;
         wait_queue_head_t replace_wait;
+
+       struct task_struct *replace_task;
  };

  /*
diff --git a/fs/btrfs/volumes.c b/fs/btrfs/volumes.c
index 8f340ad1d938..995b0647f538 100644
--- a/fs/btrfs/volumes.c
+++ b/fs/btrfs/volumes.c
@@ -6480,13 +6480,15 @@ int btrfs_map_block(struct btrfs_fs_info 
*fs_info, enum btrfs_map_op op,
         max_len = btrfs_max_io_len(map, map_offset, &io_geom);
         *length = min_t(u64, map->chunk_len - map_offset, max_len);

-       down_read(&dev_replace->rwsem);
+       if (dev_replace->replace_task != current)
+               down_read(&dev_replace->rwsem);
+
         dev_replace_is_ongoing = btrfs_dev_replace_is_ongoing(dev_replace);
         /*
          * Hold the semaphore for read during the whole operation, write is
          * requested at commit time but must wait.
          */
-       if (!dev_replace_is_ongoing)
+       if (!dev_replace_is_ongoing && dev_replace->replace_task != current)
                 up_read(&dev_replace->rwsem);

         switch (map->type & BTRFS_BLOCK_GROUP_PROFILE_MASK) {
@@ -6626,7 +6628,7 @@ int btrfs_map_block(struct btrfs_fs_info *fs_info, 
enum btrfs_map_op op,
         bioc->mirror_num = io_geom.mirror_num;

  out:
-       if (dev_replace_is_ongoing) {
+       if (dev_replace_is_ongoing && dev_replace->replace_task != 
current) {
                 lockdep_assert_held(&dev_replace->rwsem);
                 /* Unlock and let waiting writers proceed */
                 up_read(&dev_replace->rwsem);


> 
> What this patch is doing is fishy, because it's setting
> dev_replace_is_ongoing to false when there's an actual dev replace in
> progress.
> By setting it to false we are preventing the redirection of a write to
> the new device.
> 
> Was that intentional? If so, why (the changelog doesn't mention it,
> and that's a very important detail)?

Nope that was unintentional. Thanks for spotting.

> For my suggestion, since we will skip the locking of the semaphore,
> we'll have to compare "current" with "dev_replace->replace_task"
> without any locking,
> but that's ok and we can use data_race() to silence KCSAN.

I'll check if KCSAN actually complains about this.

As for David's remark, do you prefere a pid or a task_struct to be 
stored in struct dev_replace?

Thanks,
	Johannes