Skip to contentSkip to conversation

O_SYNC error handling fix

PATCH In review Chris Mason 2 participants 4 messages Jan 16, 2006

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 O_SYNC error handling fix

If we hit errors during generic_osync_inode(), generic_file_buffered_write will still return written ? written : status; This patch ensures the -EIO gets back up to userland at all times.

mm/filemap.c +15 -2
diff -r 7e48cbfbccad mm/filemap.c
@@ -2026,7 +2026,6 @@ generic_file_buffered_write(struct kiocb
 		balance_dirty_pages_ratelimited(mapping);
 		cond_resched();
 	} while (count);
-	*ppos = pos;
 
 	if (cached_page)
 		page_cache_release(cached_page);
@@ -2047,10 +2046,24 @@ generic_file_buffered_write(struct kiocb
 	 * to buffered writes (block instantiation inside i_size).  So we sync
 	 * the file data here, to try to honour O_DIRECT expectations.
 	 */
-	if (unlikely(file->f_flags & O_DIRECT) && written)
+	if (unlikely(file->f_flags & O_DIRECT) && status >= 0 && written)
 		status = filemap_write_and_wait(mapping);
 
 	pagevec_lru_add(&lru_pvec);
+
+	/*
+	 * We must let know userspace if something hasn't been written
+	 * correctly. If we got an I/O error it means we got an hardware
+	 * failure, anything can be happening to the on-disk data,
+	 * letting know userspace that a bit of data might have been
+	 * written correctly on disk is a very low priority, compared
+	 * to letting know userspace that some data has _not_ been
+	 * written at all.
+	 */
+	if (unlikely(status == -EIO))
+		return status;
+	*ppos = pos;
+
 	return written ? written : status;
 }
 EXPORT_SYMBOL(generic_file_buffered_write);
#1 lore raw
Quoted 33 lines · Chris Mason

If we hit errors during generic_osync_inode(), generic_file_buffered_write will still return written ? written : status; This patch ensures the -EIO gets back up to userland at all times.

diff -r 7e48cbfbccad mm/filemap.c

mm/filemap.c +8 -2
@@ -2026,7 +2026,6 @@ generic_file_buffered_write(struct kiocb
 		balance_dirty_pages_ratelimited(mapping);
 		cond_resched();
 	} while (count);
-	*ppos = pos;
 
 	if (cached_page)
 		page_cache_release(cached_page);
@@ -2047,10 +2046,24 @@ generic_file_buffered_write(struct kiocb
 	 * to buffered writes (block instantiation inside i_size).  So we sync
 	 * the file data here, to try to honour O_DIRECT expectations.
 	 */
-	if (unlikely(file->f_flags & O_DIRECT) && written)
+	if (unlikely(file->f_flags & O_DIRECT) && status >= 0 && written)
 		status = filemap_write_and_wait(mapping);
 
 	pagevec_lru_add(&lru_pvec);
+
+	/*
+	 * We must let know userspace if something hasn't been written
+	 * correctly. If we got an I/O error it means we got an hardware
+	 * failure, anything can be happening to the on-disk data,
+	 * letting know userspace that a bit of data might have been
+	 * written correctly on disk is a very low priority, compared

mm/filemap.c generic_file_buffered_write(struct kiocb · line 2069

Quoted 10 lines
mm/filemap.c +7 -0
+	 * to letting know userspace that some data has _not_ been
+	 * written at all.
+	 */
+	if (unlikely(status == -EIO))
+		return status;
+	*ppos = pos;
+
 	return written ? written : status;
 }
 EXPORT_SYMBOL(generic_file_buffered_write);

The patch does more than this. It also changes the O_DIRECT logic and it also refuses to update the file position on I/O errors.

Probably reasonable changes, but I'd like to have your description of why these changes were made please.

The patch also makes write() non-linuxy. We normally return a short write on errors. In this case I'd say that returning 0 (and not updating f_pos) would be appropriate?

#2 lore raw
Quoted 4 lines · Andrew Morton
Quoted 3 lines · Chris Mason

If we hit errors during generic_osync_inode(), generic_file_buffered_write will still return written ? written : status;
This patch ensures the -EIO gets back up to userland at all times.

[ ... ]

Quoted 3 lines

The patch does more than this. It also changes the O_DIRECT logic and it also refuses to update the file position on I/O errors.

True, I can get rid of the O_DIRECT change, it was just to preserve the status variable for the -EIO test.

Quoted 4 lines

Probably reasonable changes, but I'd like to have your description of why these changes were made please.

Skipping the file position update was because we are not returning the short write. We want the next write to happen at the appropriate place given the number of bytes we've told userland went out.

Quoted 4 lines

The patch also makes write() non-linuxy. We normally return a short write on errors. In this case I'd say that returning 0 (and not updating f_pos) would be appropriate?

With O_SYNC we've no way of knowing which bytes didn't make it down to the disk. I can special case the -EIO check for just O_SYNC. I think that along with changing around the O_DIRECT bits will make us both happy?

-chris

#3 lore raw
Quoted 3 lines · Chris Mason

I can special case the -EIO check for just O_SYNC. I think that along
with changing around the O_DIRECT bits will make us both happy?

I guess so...

It seems I'm full of crap anyway. We already return -EFOO on failing O_SYNC writes, no?