Re: [PATCH] f2fs: Fix deadlock under storage almost full/dirty condition

[Chao Yu]

On 2019/11/8 19:03, Sahitya Tummala wrote:
> There could be a potential deadlock when the storage capacity
> is almost full and theren't enough free segments available, due
> to which FG_GC is needed in the atomic commit ioctl as shown in
> the below callstack -
> 
> 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
> 
> If this inode doesn't have i_gc_failures[GC_FAILURE_ATOMIC] set,
> then it waits forever in f2fs_drop_inmem_pages_all(), for this
> atomic inode to be dropped. And the rest of the system is stuck
> waiting for sbi->gc_mutex lock, which is acquired by f2fs_balance_fs()
> in the stack above.

I think the root cause of this issue is there is potential infinite loop in
f2fs_drop_inmem_pages_all() for the case of gc_failure is true, because once the
first inode in inode_list[ATOMIC_FILE] list didn't suffer gc failure, we will
skip dropping its in-memory cache and calling iput(), and traverse the list
again, most possibly there is the same inode in the head of that list.

Could you please check below fix:

diff --git a/fs/f2fs/f2fs.h b/fs/f2fs/f2fs.h
index 7bf7b0194944..8a3a35b42a37 100644
--- a/fs/f2fs/f2fs.h
+++ b/fs/f2fs/f2fs.h
@@ -1395,6 +1395,7 @@ 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 ecd063239642..79f4b348951a 100644
--- a/fs/f2fs/file.c
+++ b/fs/f2fs/file.c
@@ -2047,6 +2047,7 @@ 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 8b977bbd6822..6aa0bb693697 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 +298,29 @@ void f2fs_drop_inmem_pages_all(struct f2fs_sb_info *sbi,
bool gc_failure)
 	}
 	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;
 }

@@ -334,6 +343,7 @@ void f2fs_drop_inmem_pages(struct inode *inode)
 	spin_lock(&sbi->inode_lock[ATOMIC_FILE]);
 	if (!list_empty(&fi->inmem_ilist))
 		list_del_init(&fi->inmem_ilist);
+	sbi->atomic_files--;
 	spin_unlock(&sbi->inode_lock[ATOMIC_FILE]);
 }

Thanks,