f2fs: Fix deadlock in f2fs_gc() context during atomic files handling

The FS got stuck in the below stack when the storage is almost
full/dirty condition (when FG_GC is being done).

schedule_timeout
io_schedule_timeout
congestion_wait
f2fs_drop_inmem_pages_all
f2fs_gc
f2fs_balance_fs
__write_node_page
f2fs_fsync_node_pages
f2fs_do_sync_file
f2fs_ioctl

The root cause for this issue is there is a potential infinite loop
in f2fs_drop_inmem_pages_all() for the case where gc_failure is true
and when there an inode whose i_gc_failures[GC_FAILURE_ATOMIC] is
not set. Fix this by keeping track of the total atomic files
currently opened and using that to exit from this condition.

Fix-suggested-by: Chao Yu <yuchao0@huawei.com>
Signed-off-by: Chao Yu <yuchao0@huawei.com>
Signed-off-by: Sahitya Tummala <stummala@codeaurora.org>
Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>

3 files changed, 17 insertions, 6 deletions

diff --git a/fs/f2fs/f2fs.h b/fs/f2fs/f2fs.h
index 3f6204202788..5a888a063c7f 100644
--- a/fs/f2fs/f2fs.h
+++ b/fs/f2fs/f2fs.h
@@ -1297,6 +1297,7 @@ struct f2fs_sb_info {
 	unsigned int gc_mode;			/* current GC state */
 	unsigned int next_victim_seg[2];	/* next segment in victim section */
 	/* for skip statistic */
+	unsigned int atomic_files;              /* # of opened atomic file */
 	unsigned long long skipped_atomic_files[2];	/* FG_GC and BG_GC */
 	unsigned long long skipped_gc_rwsem;		/* FG_GC only */
 
diff --git a/fs/f2fs/file.c b/fs/f2fs/file.c
index f9f3a417a0cd..c0560d62dbee 100644
--- a/fs/f2fs/file.c
+++ b/fs/f2fs/file.c
@@ -1922,6 +1922,7 @@ static int f2fs_ioc_start_atomic_write(struct file *filp)
 	spin_lock(&sbi->inode_lock[ATOMIC_FILE]);
 	if (list_empty(&fi->inmem_ilist))
 		list_add_tail(&fi->inmem_ilist, &sbi->inode_list[ATOMIC_FILE]);
+	sbi->atomic_files++;
 	spin_unlock(&sbi->inode_lock[ATOMIC_FILE]);
 
 	/* add inode in inmem_list first and set atomic_file */
diff --git a/fs/f2fs/segment.c b/fs/f2fs/segment.c
index 43daa9431160..fa32ce92ed65 100644
--- a/fs/f2fs/segment.c
+++ b/fs/f2fs/segment.c
@@ -288,6 +288,8 @@ void f2fs_drop_inmem_pages_all(struct f2fs_sb_info *sbi, bool gc_failure)
 	struct list_head *head = &sbi->inode_list[ATOMIC_FILE];
 	struct inode *inode;
 	struct f2fs_inode_info *fi;
+	unsigned int count = sbi->atomic_files;
+	unsigned int looped = 0;
 next:
 	spin_lock(&sbi->inode_lock[ATOMIC_FILE]);
 	if (list_empty(head)) {
@@ -296,22 +298,26 @@ next:
 	}
 	fi = list_first_entry(head, struct f2fs_inode_info, inmem_ilist);
 	inode = igrab(&fi->vfs_inode);
+	if (inode)
+		list_move_tail(&fi->inmem_ilist, head);
 	spin_unlock(&sbi->inode_lock[ATOMIC_FILE]);
 
 	if (inode) {
 		if (gc_failure) {
-			if (fi->i_gc_failures[GC_FAILURE_ATOMIC])
-				goto drop;
-			goto skip;
+			if (!fi->i_gc_failures[GC_FAILURE_ATOMIC])
+				goto skip;
 		}
-drop:
 		set_inode_flag(inode, FI_ATOMIC_REVOKE_REQUEST);
 		f2fs_drop_inmem_pages(inode);
+skip:
 		iput(inode);
 	}
-skip:
 	congestion_wait(BLK_RW_ASYNC, HZ/50);
 	cond_resched();
+	if (gc_failure) {
+		if (++looped >= count)
+			return;
+	}
 	goto next;
 }
 
@@ -327,13 +333,16 @@ void f2fs_drop_inmem_pages(struct inode *inode)
 		mutex_unlock(&fi->inmem_lock);
 	}
 
-	clear_inode_flag(inode, FI_ATOMIC_FILE);
 	fi->i_gc_failures[GC_FAILURE_ATOMIC] = 0;
 	stat_dec_atomic_write(inode);
 
 	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]);
 }