Skip to contentSkip to conversation

FOP_DONTCACHE when btrfs Direct IO fallback to buffered IO

DISCUSSION Wang Yugui 2 participants 3 messages Dec 21, 2025 – Dec 22, 2025 Btrfs file system

lore
Share & tools
lore
author
#1 lore raw

Hi,

Quoted 4 lines

Hi,

Could we add FOP_DONTCACHE support when btrfs Direct IO fallback to buffered IO?

Quoted 8 lines

I noticed similar logic in zfs-2.4.0 too.

https://github.com/openzfs/zfs/releases/tag/zfs-2.4.0
Uncached IO: Direct IO fallback to a light-weight uncached IO when unaligned

Best Regards
Wang Yugui ([email protected])
2025/12/21

The following dirty patch seems work here.

Best Regards
Wang Yugui ([email protected])
2025/12/22

fs/btrfs/direct-io.c +1 -0
index 802d4dbe5b38..2642dceb6911 100644
@@ -881,6 +881,7 @@ ssize_t btrfs_direct_write(struct kiocb *iocb, struct iov_iter *from)
 	 */
 	if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)) {
 		btrfs_inode_unlock(BTRFS_I(inode), ilock_flags);
+		iocb->ki_flags |= IOCB_DONTCACHE;
 		goto buffered;
 	}
 
fs/btrfs/file.c +1 -1
index fa82def46e39..64eae7417242 100644
@@ -3843,7 +3843,7 @@ const struct file_operations btrfs_file_operations = {
 #endif
 	.remap_file_range = btrfs_remap_file_range,
 	.uring_cmd	= btrfs_uring_cmd,
-	.fop_flags	= FOP_BUFFER_RASYNC | FOP_BUFFER_WASYNC,
+	.fop_flags	= FOP_BUFFER_RASYNC | FOP_BUFFER_WASYNC | FOP_DONTCACHE,
 };
 
 int btrfs_fdatawrite_range(struct btrfs_inode *inode, loff_t start, loff_t end)
#2 lore raw
Quoted 27 lines · Wang Yugui

Hi,

Quoted 4 lines

Hi,

Could we add FOP_DONTCACHE support when btrfs Direct IO fallback to buffered IO?

Quoted 8 lines

I noticed similar logic in zfs-2.4.0 too.

https://github.com/openzfs/zfs/releases/tag/zfs-2.4.0
Uncached IO: Direct IO fallback to a light-weight uncached IO when unaligned

Best Regards
Wang Yugui ([email protected])
2025/12/21

The following dirty patch seems work here.

Best Regards
Wang Yugui ([email protected])
2025/12/22

fs/btrfs/direct-io.c +0 -0
index 802d4dbe5b38..2642dceb6911 100644

fs/btrfs/direct-io.c ssize_t btrfs_direct_write(struct kiocb *iocb, struct iov_iter *from) · line 886

Quoted 6 lines
fs/btrfs/direct-io.c +1 -0
@@ -881,6 +881,7 @@ ssize_t btrfs_direct_write(struct kiocb *iocb, struct iov_iter *from)
  	 */
  	if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)) {
  		btrfs_inode_unlock(BTRFS_I(inode), ilock_flags);
+		iocb->ki_flags |= IOCB_DONTCACHE;
  		goto buffered;
  	}

Nope, run the full fstest and you will crash.

The problem is how btrfs handling compression, the writeback range for compression will stay locked (thus async), and can be marked writeback at almost any time.

That's why we have all the extra per-block handling for writeback other than other fses' regular mark the full folio writeback.

Quoted 18 lines
fs/btrfs/file.c +1 -1
index fa82def46e39..64eae7417242 100644
@@ -3843,7 +3843,7 @@ const struct file_operations btrfs_file_operations = {
  #endif
  	.remap_file_range = btrfs_remap_file_range,
  	.uring_cmd	= btrfs_uring_cmd,
-	.fop_flags	= FOP_BUFFER_RASYNC | FOP_BUFFER_WASYNC,
+	.fop_flags	= FOP_BUFFER_RASYNC | FOP_BUFFER_WASYNC | FOP_DONTCACHE,
  };
 
  int btrfs_fdatawrite_range(struct btrfs_inode *inode, loff_t start, loff_t end)