[PATCH v4] f2fs: protect published gc_thread during teardown

[Zhang Cen]

f2fs_stop_gc_thread() stops the background GC task, wakes foreground
GC_MERGE waiters, frees sbi->gc_thread, and then clears the published
pointer.  A foreground f2fs_balance_fs() caller can already have copied
that pointer and queued itself on gc_th->fggc_wq, so freeing gc_th at
stop time can leave finish_wait() operating on a freed waitqueue.

Keep the allocated GC-thread state until the superblock is destroyed and
use gc_th->f2fs_gc_task as the running-state marker.  The stop path now
withdraws the task pointer with xchg(), stops the task, and wakes any
foreground waiters, but leaves the waitqueue storage valid.  The start
path reuses a stopped gc_thread object instead of reinitializing its
waitqueues, and remount restart decisions check the task pointer rather
than only the object pointer.

f2fs_balance_fs() also snapshots sbi->gc_thread once and rechecks
f2fs_gc_task after prepare_to_wait().  If teardown wins the race after
the first check, the foreground caller removes its wait entry without
sleeping on a worker that has already been withdrawn.

Validation reproduced this kernel report:

BUG: KASAN: slab-use-after-free in finish_wait+0x276/0x290
Write of size 8 at addr ffff8881150819b8 by task dd/802
The buggy address belongs to the object at ffff888115081900 which
belongs to the cache kmalloc-256 of size 256
The buggy address is located 184 bytes inside of freed 256-byte region
Call trace:
  finish_wait()
  f2fs_balance_fs()
  f2fs_write_single_data_page()
  f2fs_write_cache_pages()
  __f2fs_write_data_pages()
  do_writepages()
  filemap_fdatawrite_wbc()
  __filemap_fdatawrite_range()
  file_write_and_wait_range()
  f2fs_do_sync_file()
  f2fs_sync_file()
  do_fsync()
Freed by task stack:
  kfree()
  f2fs_stop_gc_thread()
  f2fs_do_shutdown()
  f2fs_shutdown()
  fs_bdev_mark_dead()

Fixes: 5911d2d1d1a3 ("f2fs: introduce gc_merge mount option")
Signed-off-by: Zhang Cen <rollkingzzc@xxxxxxxxx>
---
v4:
- Replace the v3 SRCU/refcounted lifetime model with a smaller fix that
  keeps the existing heap-allocated gc_thread object alive until
  superblock teardown.
- Use f2fs_gc_task as the running-state marker and withdraw it with
  xchg() before waking GC_MERGE waiters.
- Reuse a stopped gc_thread object across remount restarts so the
  waitqueues are not reinitialized while old waiters can still finish.
- Recheck f2fs_gc_task after prepare_to_wait() so a waiter that races
  with teardown does not sleep after the worker has been withdrawn.

v3:
- Add the Fixes tag for the GC_MERGE foreground wait path.
- Fix checkpatch style issues in the broader lifetime variant.

v2:
- Sashiko.dev pointed out that GC_MERGE foreground waiters and
  GC-thread users needed lifetime-safe access after teardown.

 fs/f2fs/gc.c      | 48 +++++++++++++++++++++++++++++------------------
 fs/f2fs/segment.c | 19 ++++++++++++-------
 fs/f2fs/super.c   |  7 +++++--
 3 files changed, 47 insertions(+), 27 deletions(-)

diff --git a/fs/f2fs/gc.c b/fs/f2fs/gc.c
index ba93010924c06..20f8394482a09 100644
--- a/fs/f2fs/gc.c
+++ b/fs/f2fs/gc.c
@@ -193,12 +193,23 @@ static int gc_thread_func(void *data)
 
 int f2fs_start_gc_thread(struct f2fs_sb_info *sbi)
 {
-	struct f2fs_gc_kthread *gc_th;
+	struct f2fs_gc_kthread *gc_th = sbi->gc_thread;
+	struct task_struct *task;
+	bool allocated = false;
 	dev_t dev = sbi->sb->s_bdev->bd_dev;
 
-	gc_th = f2fs_kmalloc(sbi, sizeof(struct f2fs_gc_kthread), GFP_KERNEL);
-	if (!gc_th)
-		return -ENOMEM;
+	if (gc_th && READ_ONCE(gc_th->f2fs_gc_task))
+		return 0;
+
+	if (!gc_th) {
+		gc_th = f2fs_kmalloc(sbi, sizeof(*gc_th), GFP_KERNEL);
+		if (!gc_th)
+			return -ENOMEM;
+		init_waitqueue_head(&gc_th->gc_wait_queue_head);
+		init_waitqueue_head(&gc_th->fggc_wq);
+		sbi->gc_thread = gc_th;
+		allocated = true;
+	}
 
 	gc_th->urgent_sleep_time = DEF_GC_THREAD_URGENT_SLEEP_TIME;
 	gc_th->valid_thresh_ratio = DEF_GC_THREAD_VALID_THRESH_RATIO;
@@ -221,34 +232,35 @@ int f2fs_start_gc_thread(struct f2fs_sb_info *sbi)
 
 	gc_th->gc_wake = false;
 
-	sbi->gc_thread = gc_th;
-	init_waitqueue_head(&sbi->gc_thread->gc_wait_queue_head);
-	init_waitqueue_head(&sbi->gc_thread->fggc_wq);
-	sbi->gc_thread->f2fs_gc_task = kthread_run(gc_thread_func, sbi,
-			"f2fs_gc-%u:%u", MAJOR(dev), MINOR(dev));
-	if (IS_ERR(gc_th->f2fs_gc_task)) {
-		int err = PTR_ERR(gc_th->f2fs_gc_task);
+	task = kthread_run(gc_thread_func, sbi, "f2fs_gc-%u:%u",
+			   MAJOR(dev), MINOR(dev));
+	if (IS_ERR(task)) {
+		int err = PTR_ERR(task);
 
-		kfree(gc_th);
-		sbi->gc_thread = NULL;
+		if (allocated) {
+			kfree(gc_th);
+			sbi->gc_thread = NULL;
+		}
 		return err;
 	}
 
-	set_user_nice(gc_th->f2fs_gc_task,
-			PRIO_TO_NICE(sbi->critical_task_priority));
+	WRITE_ONCE(gc_th->f2fs_gc_task, task);
+	set_user_nice(task, PRIO_TO_NICE(sbi->critical_task_priority));
 	return 0;
 }
 
 void f2fs_stop_gc_thread(struct f2fs_sb_info *sbi)
 {
 	struct f2fs_gc_kthread *gc_th = sbi->gc_thread;
+	struct task_struct *task;
 
 	if (!gc_th)
 		return;
-	kthread_stop(gc_th->f2fs_gc_task);
+	task = xchg(&gc_th->f2fs_gc_task, NULL);
+	if (!task)
+		return;
+	kthread_stop(task);
 	wake_up_all(&gc_th->fggc_wq);
-	kfree(gc_th);
-	sbi->gc_thread = NULL;
 }
 
 static int select_gc_type(struct f2fs_sb_info *sbi, int gc_type)
diff --git a/fs/f2fs/segment.c b/fs/f2fs/segment.c
index 788f8b0502492..84307525edd27 100644
--- a/fs/f2fs/segment.c
+++ b/fs/f2fs/segment.c
@@ -424,6 +424,8 @@ int f2fs_commit_atomic_write(struct inode *inode)
  */
 void f2fs_balance_fs(struct f2fs_sb_info *sbi, bool need)
 {
+	struct f2fs_gc_kthread *gc_th;
+
 	if (f2fs_cp_error(sbi))
 		return;
 
@@ -444,15 +446,18 @@ void f2fs_balance_fs(struct f2fs_sb_info *sbi, bool need)
 	if (has_enough_free_secs(sbi, 0, 0))
 		return;
 
-	if (test_opt(sbi, GC_MERGE) && sbi->gc_thread &&
-				sbi->gc_thread->f2fs_gc_task) {
+	gc_th = sbi->gc_thread;
+	if (test_opt(sbi, GC_MERGE) && gc_th &&
+	    READ_ONCE(gc_th->f2fs_gc_task)) {
 		DEFINE_WAIT(wait);
 
-		prepare_to_wait(&sbi->gc_thread->fggc_wq, &wait,
-					TASK_UNINTERRUPTIBLE);
-		wake_up(&sbi->gc_thread->gc_wait_queue_head);
-		io_schedule();
-		finish_wait(&sbi->gc_thread->fggc_wq, &wait);
+		prepare_to_wait(&gc_th->fggc_wq, &wait,
+				TASK_UNINTERRUPTIBLE);
+		if (READ_ONCE(gc_th->f2fs_gc_task)) {
+			wake_up(&gc_th->gc_wait_queue_head);
+			io_schedule();
+		}
+		finish_wait(&gc_th->fggc_wq, &wait);
 	} else {
 		struct f2fs_gc_control gc_control = {
 			.victim_segno = NULL_SEGNO,
diff --git a/fs/f2fs/super.c b/fs/f2fs/super.c
index ccf806b676f53..d6863da05a7c2 100644
--- a/fs/f2fs/super.c
+++ b/fs/f2fs/super.c
@@ -2925,11 +2925,12 @@ static int __f2fs_remount(struct fs_context *fc, struct super_block *sb)
 	if ((flags & SB_RDONLY) ||
 			(F2FS_OPTION(sbi).bggc_mode == BGGC_MODE_OFF &&
 			!test_opt(sbi, GC_MERGE))) {
-		if (sbi->gc_thread) {
+		if (sbi->gc_thread && READ_ONCE(sbi->gc_thread->f2fs_gc_task)) {
 			f2fs_stop_gc_thread(sbi);
 			need_restart_gc = true;
 		}
-	} else if (!sbi->gc_thread) {
+	} else if (!sbi->gc_thread ||
+			!READ_ONCE(sbi->gc_thread->f2fs_gc_task)) {
 		err = f2fs_start_gc_thread(sbi);
 		if (err)
 			goto restore_opts;
@@ -5451,6 +5452,7 @@ static int f2fs_fill_super(struct super_block *sb, struct fs_context *fc)
 free_sb_buf:
 	kfree(raw_super);
 free_sbi:
+	kfree(sbi->gc_thread);
 #ifdef CONFIG_DEBUG_LOCK_ALLOC
 	lockdep_unregister_key(&sbi->cp_global_sem_key);
 #endif
@@ -5535,6 +5537,7 @@ static void kill_f2fs_super(struct super_block *sb)
 	/* Release block devices last, after fscrypt_destroy_keyring(). */
 	if (sbi) {
 		destroy_device_list(sbi);
+		kfree(sbi->gc_thread);
 #ifdef CONFIG_DEBUG_LOCK_ALLOC
 		lockdep_unregister_key(&sbi->cp_global_sem_key);
 #endif
-- 
2.43.0