@@ -35,6 +35,7 @@ void nilfs_init_btnc_inode(struct inode *btnc_inode)
ii->i_flags = 0;
memset(&ii->i_bmap_data, 0, sizeof(struct nilfs_bmap));
mapping_set_gfp_mask(btnc_inode->i_mapping, GFP_NOFS);
+ btnc_inode->i_mapping->a_ops = &nilfs_buffer_cache_aops;
}
void nilfs_btnode_cache_clear(struct address_space *btnc)
@@ -163,7 +163,7 @@ int nilfs_init_gcinode(struct inode *inode)
inode->i_mode = S_IFREG;
mapping_set_gfp_mask(inode->i_mapping, GFP_NOFS);
- inode->i_mapping->a_ops = &empty_aops;
+ inode->i_mapping->a_ops = &nilfs_buffer_cache_aops;
ii->i_flags = 0;
nilfs_bmap_init_gc(ii->i_bmap);
@@ -276,6 +276,10 @@ const struct address_space_operations nilfs_aops = {
.is_partially_uptodate = block_is_partially_uptodate,
};
+const struct address_space_operations nilfs_buffer_cache_aops = {
+ .invalidate_folio = block_invalidate_folio,
+};
+
static int nilfs_insert_inode_locked(struct inode *inode,
struct nilfs_root *root,
unsigned long ino)
@@ -681,6 +685,7 @@ struct inode *nilfs_iget_for_shadow(struct inode *inode)
NILFS_I(s_inode)->i_flags = 0;
memset(NILFS_I(s_inode)->i_bmap, 0, sizeof(struct nilfs_bmap));
mapping_set_gfp_mask(s_inode->i_mapping, GFP_NOFS);
+ s_inode->i_mapping->a_ops = &nilfs_buffer_cache_aops;
err = nilfs_attach_btree_node_cache(s_inode);
if (unlikely(err)) {
@@ -401,6 +401,7 @@ extern const struct file_operations nilfs_dir_operations;
extern const struct inode_operations nilfs_file_inode_operations;
extern const struct file_operations nilfs_file_operations;
extern const struct address_space_operations nilfs_aops;
+extern const struct address_space_operations nilfs_buffer_cache_aops;
extern const struct inode_operations nilfs_dir_inode_operations;
extern const struct inode_operations nilfs_special_inode_operations;
extern const struct inode_operations nilfs_symlink_inode_operations;
When block_invalidatepage was converted to block_invalidate_folio, the fallback to block_invalidatepage in folio_invalidate() if the address_space_operations method invalidatepage (currently invalidate_folio) was not set, was removed. Unfortunately, some pseudo-inodes in nilfs2 use empty_aops set by inode_init_always_gfp() as is, or explicitly set it to address_space_operations. Therefore, with this change, block_invalidatepage() is no longer called from folio_invalidate(), and as a result, the buffer_head structures attached to these pages/folios are no longer freed via try_to_free_buffers(). Thus, these buffer heads are now leaked by truncate_inode_pages(), which cleans up the page cache from inode evict(), etc. Three types of caches use empty_aops: gc inode caches and the DAT shadow inode used by GC, and b-tree node caches. Of these, b-tree node caches explicitly call invalidate_mapping_pages() during cleanup, which involves calling try_to_free_buffers(), so the leak was not visible during normal operation but worsened when GC was performed. Fix this issue by using address_space_operations with invalidate_folio set to block_invalidate_folio instead of empty_aops, which will ensure the same behavior as before. Signed-off-by: Ryusuke Konishi <konishi.ryusuke@gmail.com> Fixes: 7ba13abbd31e ("fs: Turn block_invalidatepage into block_invalidate_folio") Cc: stable@vger.kernel.org # v5.18+ --- Hi Andrew, please apply this as a bug fix. This fixes buffer head memory leaks (seen by slabinfo and runtime function call checks) that gets worse during GC. Thanks, Ryusuke Konishi fs/nilfs2/btnode.c | 1 + fs/nilfs2/gcinode.c | 2 +- fs/nilfs2/inode.c | 5 +++++ fs/nilfs2/nilfs.h | 1 + 4 files changed, 8 insertions(+), 1 deletion(-)