btrfs: fix race in read_extent_buffer_pages()

In the Linux kernel, the following vulnerability has been resolved:

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.

[ minor update of changelog ]