btrfs: subpage: only allow compression if the range is fully page aligned

For compressed write, we use a mechanism called async COW, which unlike
regular run_delalloc_cow() or cow_file_range() will also unlock the
first page.

This mechanism allows us to continue handling next ranges, without
waiting for the time consuming compression.

But this has a problem for subpage case, as we could have the following
delalloc range for a page:

0		32K		64K
|	|///////|	|///////|
		\- A		\- B

In the above case, if we pass both ranges to cow_file_range_async(),
both range A and range B will try to unlock the full page [0, 64K).

And which one finishes later than the other one will try to do other
page operations like end_page_writeback() on a unlocked page, triggering
VM layer BUG_ON().

To make subpage compression work at least partially, here we add another
restriction for it, only allow compression if the delalloc range is
fully page aligned.

By that, async extent is always ensured to unlock the first page
exclusively, just like it used to be for regular sectorsize.

In theory, we only need to make sure the delalloc range fully covers its
first page, but the tail page will be locked anyway, blocking later
writeback until the compression finishes.

Thus here we choose to make sure the range is fully page aligned before
doing the compression.

In the future, we could optimize the situation by properly increasing
subpage::writers number for the locked page, but that also means we need
to change how we run delalloc range of page.
(Instead of running each delalloc range we hit, we need to find and lock
all delalloc ranges covering the page, then run each of them).

Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>

1 files changed, 44 insertions, 4 deletions

diff --git a/fs/btrfs/inode.c b/fs/btrfs/inode.c
index 0276fa951641..28e0ad1478c3 100644
--- a/fs/btrfs/inode.c
+++ b/fs/btrfs/inode.c
@@ -489,9 +489,6 @@ static noinline int add_async_extent(struct async_chunk *cow,
  */
 static inline bool inode_can_compress(struct btrfs_inode *inode)
 {
-	/* Subpage doesn't support compression yet */
-	if (inode->root->fs_info->sectorsize < PAGE_SIZE)
-		return false;
 	if (inode->flags & BTRFS_INODE_NODATACOW ||
 	    inode->flags & BTRFS_INODE_NODATASUM)
 		return false;
@@ -513,6 +510,38 @@ static inline int inode_need_compress(struct btrfs_inode *inode, u64 start,
 			btrfs_ino(inode));
 		return 0;
 	}
+	/*
+	 * Special check for subpage.
+	 *
+	 * We lock the full page then run each delalloc range in the page, thus
+	 * for the following case, we will hit some subpage specific corner case:
+	 *
+	 * 0		32K		64K
+	 * |	|///////|	|///////|
+	 *		\- A		\- B
+	 *
+	 * In above case, both range A and range B will try to unlock the full
+	 * page [0, 64K), causing the one finished later will have page
+	 * unlocked already, triggering various page lock requirement BUG_ON()s.
+	 *
+	 * So here we add an artificial limit that subpage compression can only
+	 * if the range is fully page aligned.
+	 *
+	 * In theory we only need to ensure the first page is fully covered, but
+	 * the tailing partial page will be locked until the full compression
+	 * finishes, delaying the write of other range.
+	 *
+	 * TODO: Make btrfs_run_delalloc_range() to lock all delalloc range
+	 * first to prevent any submitted async extent to unlock the full page.
+	 * By this, we can ensure for subpage case that only the last async_cow
+	 * will unlock the full page.
+	 */
+	if (fs_info->sectorsize < PAGE_SIZE) {
+		if (!IS_ALIGNED(start, PAGE_SIZE) ||
+		    !IS_ALIGNED(end + 1, PAGE_SIZE))
+			return 0;
+	}
+
 	/* force compress */
 	if (btrfs_test_opt(fs_info, FORCE_COMPRESS))
 		return 1;
@@ -614,13 +643,24 @@ again:
 	total_compressed = actual_end - start;
 
 	/*
-	 * skip compression for a small file range(<=blocksize) that
+	 * Skip compression for a small file range(<=blocksize) that
 	 * isn't an inline extent, since it doesn't save disk space at all.
 	 */
 	if (total_compressed <= blocksize &&
 	   (start > 0 || end + 1 < BTRFS_I(inode)->disk_i_size))
 		goto cleanup_and_bail_uncompressed;
 
+	/*
+	 * For subpage case, we require full page alignment for the sector
+	 * aligned range.
+	 * Thus we must also check against @actual_end, not just @end.
+	 */
+	if (blocksize < PAGE_SIZE) {
+		if (!IS_ALIGNED(start, PAGE_SIZE) ||
+		    !IS_ALIGNED(round_up(actual_end, blocksize), PAGE_SIZE))
+			goto cleanup_and_bail_uncompressed;
+	}
+
 	total_compressed = min_t(unsigned long, total_compressed,
 			BTRFS_MAX_UNCOMPRESSED);
 	total_in = 0;