btrfs: avoid access-beyond-folio for bs > ps encoded writes

[POTENTIAL BUG]
If the system page size is 4K and fs block size is 8K, and max_inline
mount option is set to 6K, we can inline a 6K sized data extent.

Then a encoded write submitted a compressed extent which is at file
offset 0, and the compressed length is 6K, which is allowed to be inlined.

Now a read beyond page boundary is triggered inside write_extent_buffer()
from insert_inline_extent().

[CAUSE]
Currently the function __cow_file_range_inline() can only accept a
single folio.

For regular compressed write path, we always allocate the compressed
folios using the minimal order matching the block size, thus the
@compressed_folio should always cover a full fs block thus it is fine.

But for encoded writes, they allocate page size folios, this means we
can hit a case where the compressed data is smaller than block size but
still larger than page size, in that case __cow_file_range_inline() will
be called with @compressed_size larger than a page.

In that case we will trigger a read beyond the folio inside
insert_inline_extent().

Thankfully this is not that common, as the default max_inline is only
2048 bytes, smaller than PAGE_SIZE, and bs > ps support is still
experimental.

[FIX]
We need to either allow insert_inline_extent() to accept a page array to
properly support such case, or reject such inline extent.

The latter is a much simpler solution, and considering bs > ps will stay
as a corner case and non-default max_inline will be even rarer, I don't
think we really need to fulfill such niche.

So just reject any inline extent that's larger than PAGE_SIZE, and add
an extra ASSERT() to insert_inline_extent() to catch such beyond-boundary
access.

Fixes: ec20799064c8 ("btrfs: enable encoded read/write/send for bs > ps cases")
Signed-off-by: Qu Wenruo <wqu@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>

1 files changed, 18 insertions, 4 deletions

diff --git a/fs/btrfs/inode.c b/fs/btrfs/inode.c
index 03337fa7a61c..bfb8e98d8c02 100644
--- a/fs/btrfs/inode.c
+++ b/fs/btrfs/inode.c
@@ -480,13 +480,15 @@ static int insert_inline_extent(struct btrfs_trans_handle *trans,
 	ASSERT(size <= sectorsize);
 
 	/*
-	 * The compressed size also needs to be no larger than a sector.
-	 * That's also why we only need one page as the parameter.
+	 * The compressed size also needs to be no larger than a page.
+	 * That's also why we only need one folio as the parameter.
 	 */
-	if (compressed_folio)
+	if (compressed_folio) {
 		ASSERT(compressed_size <= sectorsize);
-	else
+		ASSERT(compressed_size <= PAGE_SIZE);
+	} else {
 		ASSERT(compressed_size == 0);
+	}
 
 	if (compressed_size && compressed_folio)
 		cur_size = compressed_size;
@@ -573,6 +575,18 @@ static bool can_cow_file_range_inline(struct btrfs_inode *inode,
 	if (offset != 0)
 		return false;
 
+	/*
+	 * Even for bs > ps cases, cow_file_range_inline() can only accept a
+	 * single folio.
+	 *
+	 * This can be problematic and cause access beyond page boundary if a
+	 * page sized folio is passed into that function.
+	 * And encoded write is doing exactly that.
+	 * So here limits the inlined extent size to PAGE_SIZE.
+	 */
+	if (size > PAGE_SIZE || compressed_size > PAGE_SIZE)
+		return false;
+
 	/* Inline extents are limited to sectorsize. */
 	if (size > fs_info->sectorsize)
 		return false;