On Tue, Jan 31, 2023 at 01:14:23PM +0100, Jan Kara wrote:
> This is the warning Willy has added as part of "mpage: convert
> __mpage_writepage() to use a folio more fully" and that warning can indeed
> easily trigger. There's nothing that serializes writeback against racing
> truncate setting new i_size so it is perfectly normal to see pages beyond
> EOF in this place. And the traditional response to such pages is "silently
> do nothing" since they will be soon discarded by truncate_inode_pages().
Absolutely right. Not sure what I was thinking; I may have been
confused by the label being called "confused". How about this for
Andrew to squash into that commit?
diff --git a/fs/mpage.c b/fs/mpage.c
index 2efa393f0db7..89bcefb4553a 100644
--- a/fs/mpage.c
+++ b/fs/mpage.c
@@ -559,6 +559,9 @@ static int __mpage_writepage(struct folio *folio, struct writeback_control *wbc,
first_unmapped = page_block;
page_is_mapped:
+ /* Don't bother writing beyond EOF, truncate will discard the folio */
+ if (folio_pos(folio) >= i_size)
+ goto confused;
length = folio_size(folio);
if (folio_pos(folio) + length > i_size) {
/*
@@ -570,8 +573,6 @@ static int __mpage_writepage(struct folio *folio, struct writeback_control *wbc,
* written out to the file."
*/
length = i_size - folio_pos(folio);
- if (WARN_ON_ONCE(folio_pos(folio) >= i_size))
- goto confused;
folio_zero_segment(folio, length, folio_size(folio));
}
