diff mbox series

[5.12,144/363] f2fs: Fix a hungtask problem in atomic write

Message ID 20210517140307.482258508@linuxfoundation.org
State New
Headers show
Series None | expand

Commit Message

Greg Kroah-Hartman May 17, 2021, 2 p.m. UTC 
From: Yi Zhuang <zhuangyi1@huawei.com>

[ Upstream commit be1ee45d51384161681ecf21085a42d316ae25f7 ]

In the cache writing process, if it is an atomic file, increase the page
count of F2FS_WB_CP_DATA, otherwise increase the page count of
F2FS_WB_DATA.

When you step into the hook branch due to insufficient memory in
f2fs_write_begin, f2fs_drop_inmem_pages_all will be called to traverse
all atomic inodes and clear the FI_ATOMIC_FILE mark of all atomic files.

In f2fs_drop_inmem_pages，first acquire the inmem_lock , revoke all the
inmem_pages, and then clear the FI_ATOMIC_FILE mark. Before this mark is
cleared, other threads may hold inmem_lock to add inmem_pages to the inode
that has just been emptied inmem_pages, and increase the page count of
F2FS_WB_CP_DATA.

When the IO returns, it is found that the FI_ATOMIC_FILE flag is cleared
by f2fs_drop_inmem_pages_all, and f2fs_is_atomic_file returns false,which
causes the page count of F2FS_WB_DATA to be decremented. The page count of
F2FS_WB_CP_DATA cannot be cleared. Finally, hungtask is triggered in
f2fs_wait_on_all_pages because get_pages will never return zero.

process A:				process B:
f2fs_drop_inmem_pages_all
->f2fs_drop_inmem_pages of inode#1
    ->mutex_lock(&fi->inmem_lock)
    ->__revoke_inmem_pages of inode#1	f2fs_ioc_commit_atomic_write
    ->mutex_unlock(&fi->inmem_lock)	->f2fs_commit_inmem_pages of inode#1
					->mutex_lock(&fi->inmem_lock)
					->__f2fs_commit_inmem_pages
					    ->f2fs_do_write_data_page
					        ->f2fs_outplace_write_data
					            ->do_write_page
					                ->f2fs_submit_page_write
					                    ->inc_page_count(sbi, F2FS_WB_CP_DATA )
					->mutex_unlock(&fi->inmem_lock)
    ->spin_lock(&sbi->inode_lock[ATOMIC_FILE]);
    ->clear_inode_flag(inode, FI_ATOMIC_FILE)
    ->spin_unlock(&sbi->inode_lock[ATOMIC_FILE])
					f2fs_write_end_io
					->dec_page_count(sbi, F2FS_WB_DATA );

We can fix the problem by putting the action of clearing the FI_ATOMIC_FILE
mark into the inmem_lock lock. This operation can ensure that no one will
submit the inmem pages before the FI_ATOMIC_FILE mark is cleared, so that
there will be no atomic writes waiting for writeback.

Fixes: 57864ae5ce3a ("f2fs: limit # of inmemory pages")
Signed-off-by: Yi Zhuang <zhuangyi1@huawei.com>
Reviewed-by: Chao Yu <yuchao0@huawei.com>
Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
Signed-off-by: Sasha Levin <sashal@kernel.org>
---
 fs/f2fs/segment.c | 30 +++++++++++++++++-------------
 1 file changed, 17 insertions(+), 13 deletions(-)
diff mbox series

Patch

diff --git a/fs/f2fs/segment.c b/fs/f2fs/segment.c
index 4fba82294a73..d83cb1359e53 100644
--- a/fs/f2fs/segment.c
+++ b/fs/f2fs/segment.c
@@ -324,23 +324,27 @@  void f2fs_drop_inmem_pages(struct inode *inode)
 	struct f2fs_sb_info *sbi = F2FS_I_SB(inode);
 	struct f2fs_inode_info *fi = F2FS_I(inode);
 
-	while (!list_empty(&fi->inmem_pages)) {
+	do {
 		mutex_lock(&fi->inmem_lock);
+		if (list_empty(&fi->inmem_pages)) {
+			fi->i_gc_failures[GC_FAILURE_ATOMIC] = 0;
+
+			spin_lock(&sbi->inode_lock[ATOMIC_FILE]);
+			if (!list_empty(&fi->inmem_ilist))
+				list_del_init(&fi->inmem_ilist);
+			if (f2fs_is_atomic_file(inode)) {
+				clear_inode_flag(inode, FI_ATOMIC_FILE);
+				sbi->atomic_files--;
+			}
+			spin_unlock(&sbi->inode_lock[ATOMIC_FILE]);
+
+			mutex_unlock(&fi->inmem_lock);
+			break;
+		}
 		__revoke_inmem_pages(inode, &fi->inmem_pages,
 						true, false, true);
 		mutex_unlock(&fi->inmem_lock);
-	}
-
-	fi->i_gc_failures[GC_FAILURE_ATOMIC] = 0;
-
-	spin_lock(&sbi->inode_lock[ATOMIC_FILE]);
-	if (!list_empty(&fi->inmem_ilist))
-		list_del_init(&fi->inmem_ilist);
-	if (f2fs_is_atomic_file(inode)) {
-		clear_inode_flag(inode, FI_ATOMIC_FILE);
-		sbi->atomic_files--;
-	}
-	spin_unlock(&sbi->inode_lock[ATOMIC_FILE]);
+	} while (1);
 }
 
 void f2fs_drop_inmem_page(struct inode *inode, struct page *page)