XFS allows CoW on non-shared extents to combat fragmentation[1].  The
old non-shared extent could be mwrited before, its dax entry is marked
dirty.  To be able to delete this entry, clear its dirty mark before
invalidate_inode_pages2_range().

[1] https://lore.kernel.org/linux-xfs/20230321151339.GA11376@frogsfrogsfrogs/

Signed-off-by: Shiyang Ruan <ruansy.f...@fujitsu.com>
---
 fs/dax.c | 37 +++++++++++++++++++++++++++++++++++++
 1 file changed, 37 insertions(+)

diff --git a/fs/dax.c b/fs/dax.c
index 5d2e9b10030e..2ababb89918d 100644
--- a/fs/dax.c
+++ b/fs/dax.c
@@ -781,6 +781,33 @@ static int __dax_invalidate_entry(struct address_space 
*mapping,
        return ret;
 }
 
+static int __dax_clear_dirty_range(struct address_space *mapping,
+               pgoff_t start, pgoff_t end)
+{
+       XA_STATE(xas, &mapping->i_pages, start);
+       unsigned int scanned = 0;
+       void *entry;
+
+       xas_lock_irq(&xas);
+       xas_for_each(&xas, entry, end) {
+               entry = get_unlocked_entry(&xas, 0);
+               xas_clear_mark(&xas, PAGECACHE_TAG_DIRTY);
+               xas_clear_mark(&xas, PAGECACHE_TAG_TOWRITE);
+               put_unlocked_entry(&xas, entry, WAKE_NEXT);
+
+               if (++scanned % XA_CHECK_SCHED)
+                       continue;
+
+               xas_pause(&xas);
+               xas_unlock_irq(&xas);
+               cond_resched();
+               xas_lock_irq(&xas);
+       }
+       xas_unlock_irq(&xas);
+
+       return 0;
+}
+
 /*
  * Delete DAX entry at @index from @mapping.  Wait for it
  * to be unlocked before deleting it.
@@ -1440,6 +1467,16 @@ static loff_t dax_iomap_iter(const struct iomap_iter 
*iomi,
         * written by write(2) is visible in mmap.
         */
        if (iomap->flags & IOMAP_F_NEW || cow) {
+               /*
+                * Filesystem allows CoW on non-shared extents. The src extents
+                * may have been mmapped with dirty mark before. To be able to
+                * invalidate its dax entries, we need to clear the dirty mark
+                * in advance.
+                */
+               if (cow)
+                       __dax_clear_dirty_range(iomi->inode->i_mapping,
+                                               pos >> PAGE_SHIFT,
+                                               (end - 1) >> PAGE_SHIFT);
                invalidate_inode_pages2_range(iomi->inode->i_mapping,
                                              pos >> PAGE_SHIFT,
                                              (end - 1) >> PAGE_SHIFT);
-- 
2.39.2


Reply via email to