Re: possible deadlock in start_this_handle (2)

On Sat 13-02-21 23:26:37, Tetsuo Handa wrote:

On 2021/02/11 19:49, Jan Kara wrote:

quoted

This stacktrace should never happen. ext4_xattr_set() starts a transaction.
That internally goes through start_this_handle() which calls:

	handle->saved_alloc_context = memalloc_nofs_save();

and we restore the allocation context only in stop_this_handle() when
stopping the handle. And with this fs_reclaim_acquire() should remove
__GFP_FS from the mask and not call __fs_reclaim_acquire().

Excuse me, but it seems to me that nothing prevents
ext4_xattr_set_handle() from reaching ext4_xattr_inode_lookup_create()
without memalloc_nofs_save() when hitting ext4_get_nojournal() path.
Will you explain when ext4_get_nojournal() path is executed?

That's a good question but sadly I don't think that's it.
ext4_get_nojournal() is called when the filesystem is created without a
journal. In that case we also don't acquire jbd2_handle lockdep map. In the
syzbot report we can see:

kswapd0/2246 is trying to acquire lock:
ffff888041a988e0 (jbd2_handle){++++}-{0:0}, at: start_this_handle+0xf81/0x1380 fs/jbd2/transaction.c:444

but task is already holding lock:
ffffffff8be892c0 (fs_reclaim){+.+.}-{0:0}, at: __fs_reclaim_acquire+0x0/0x30 mm/page_alloc.c:5195

So this filesystem has very clearly been created with a journal. Also the
journal lockdep tracking machinery uses:

rwsem_acquire_read(&journal->j_trans_commit_map, 0, 0, _THIS_IP_);

so a lockdep key is per-filesystem. Thus it is not possible that lockdep
would combine lock dependencies from two different filesystems.

But I guess we could narrow the search for this problem by adding WARN_ONs
to ext4_xattr_set_handle() and ext4_xattr_inode_lookup_create() like:

WARN_ON(ext4_handle_valid(handle) && !(current->flags & PF_MEMALLOC_NOFS));

It would narrow down a place in which PF_MEMALLOC_NOFS flag isn't set
properly... At least that seems like the most plausible way forward to me.

								Honza

ext4_xattr_set() {
  handle = ext4_journal_start(inode, EXT4_HT_XATTR, credits) == __ext4_journal_start() {
      return __ext4_journal_start_sb() {
        journal = EXT4_SB(sb)->s_journal;
        if (!journal || (EXT4_SB(sb)->s_mount_state & EXT4_FC_REPLAY))
          return ext4_get_nojournal(); // Never calls memalloc_nofs_save() despite returning !IS_ERR() value.
        return jbd2__journal_start(journal, blocks, rsv_blocks, revoke_creds, GFP_NOFS, type, line); // Calls memalloc_nofs_save() when start_this_handle() returns 0.
      }
    }
  }
  error = ext4_xattr_set_handle(handle, inode, name_index, name, value, value_len, flags); {
    ext4_write_lock_xattr(inode, &no_expand); // Grabs &ei->xattr_sem
    error = ext4_xattr_ibody_set(handle, inode, &i, &is) {
      error = ext4_xattr_set_entry(i, s, handle, inode, false /* is_block */) {
        ret = ext4_xattr_inode_lookup_create(handle, inode, i->value, i->value_len, &new_ea_inode); // Using GFP_KERNEL based on assumption that ext4_journal_start() called memalloc_nofs_save().
      }
    }
  }
}

-- 
Jan Kara [off-list ref]
SUSE Labs, CR