btrfs: fix race in read_extent_buffer_pages()

There are reports from tree-checker that detects corrupted nodes,
without any obvious pattern so possibly an overwrite in memory.
After some debugging it turns out there's a race when reading an extent
buffer the uptodate status can be missed.

To prevent concurrent reads for the same extent buffer,
read_extent_buffer_pages() performs these checks:

    /* (1) */
    if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
        return 0;

    /* (2) */
    if (test_and_set_bit(EXTENT_BUFFER_READING, &eb->bflags))
        goto done;

At this point, it seems safe to start the actual read operation. Once
that completes, end_bbio_meta_read() does

    /* (3) */
    set_extent_buffer_uptodate(eb);

    /* (4) */
    clear_bit(EXTENT_BUFFER_READING, &eb->bflags);

Normally, this is enough to ensure only one read happens, and all other
callers wait for it to finish before returning.  Unfortunately, there is
a racey interleaving:

    Thread A | Thread B | Thread C
    ---------+----------+---------
       (1)   |          |
             |    (1)   |
       (2)   |          |
       (3)   |          |
       (4)   |          |
             |    (2)   |
             |          |    (1)

When this happens, thread B kicks of an unnecessary read. Worse, thread
C will see UPTODATE set and return immediately, while the read from
thread B is still in progress.  This race could result in tree-checker
errors like this as the extent buffer is concurrently modified:

    BTRFS critical (device dm-0): corrupted node, root=256
    block=8550954455682405139 owner mismatch, have 11858205567642294356
    expect [256, 18446744073709551360]

Fix it by testing UPTODATE again after setting the READING bit, and if
it's been set, skip the unnecessary read.

Fixes: d7172f52 ("btrfs: use per-buffer locking for extent_buffer reading")
Link: https://lore.kernel.org/linux-btrfs/CAHk-=whNdMaN9ntZ47XRKP6DBes2E5w7fi-0U3H2+PS18p+Pzw@mail.gmail.com/
Link: https://lore.kernel.org/linux-btrfs/f51a6d5d7432455a6a858d51b49ecac183e0bbc9.1706312914.git.wqu@suse.com/
Link: https://lore.kernel.org/linux-btrfs/c7241ea4-fcc6-48d2-98c8-b5ea790d6c89@gmx.com/
CC: stable@vger.kernel.org # 6.5+
Reviewed-by: Qu Wenruo <wqu@suse.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Tavian Barnes <tavianator@tavianator.com>
Reviewed-by: David Sterba <dsterba@suse.com>
[ minor update of changelog ]
Signed-off-by: David Sterba <dsterba@suse.com>

fs/btrfs/extent_io.c

@@ -4333,6 +4333,19 @@ int read_extent_buffer_pages(struct extent_buffer *eb, int wait, int mirror_num,
 	if (test_and_set_bit(EXTENT_BUFFER_READING, &eb->bflags))
 		goto done;
 
+	/*
+	 * Between the initial test_bit(EXTENT_BUFFER_UPTODATE) and the above
+	 * test_and_set_bit(EXTENT_BUFFER_READING), someone else could have
+	 * started and finished reading the same eb.  In this case, UPTODATE
+	 * will now be set, and we shouldn't read it in again.
+	 */
+	if (unlikely(test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))) {
+		clear_bit(EXTENT_BUFFER_READING, &eb->bflags);
+		smp_mb__after_atomic();
+		wake_up_bit(&eb->bflags, EXTENT_BUFFER_READING);
+		return 0;
+	}
+
 	clear_bit(EXTENT_BUFFER_READ_ERR, &eb->bflags);
 	eb->read_mirror = 0;
 	check_buffer_tree_ref(eb);