[PATCH] f2fs: wake up foreground GC waiters when the GC thread skips a round

[Wenjie Qi]

With GC_MERGE, f2fs_balance_fs() queues a foreground caller on fggc_wq,
wakes the GC thread, and sleeps without a timeout or condition recheck.

The root cause is that a foreground writer, the GC thread, and fsfreeze
can race as follows:

foreground writer                  GC thread                  fsfreeze
- f2fs_create
 - f2fs_balance_fs
  - prepare_to_wait(&fggc_wq)
  - wake_up(&gc_wait_queue_head)
  - io_schedule()
    waits on fggc_wq                                    - freeze_super()
                                                         - frozen =
                                                           SB_FREEZE_WRITE
                                                         - sb_wait_write()
                                                           waits for writer
                                  - gc_thread_func()
                                   - waitqueue_active(fggc_wq)
                                     returns true
                                   - foreground = true
                                   - frozen >= SB_FREEZE_WRITE
                                   - continue
                                     no wake_up_all(&fggc_wq)

The writer keeps its superblock write reference while sleeping. fsfreeze
waits for that writer, while the GC thread loops because the queued waiter
keeps waitqueue_active(fggc_wq) true.

Wake foreground waiters on each early skip. A skipped foreground-GC round
must not leave its caller blocked.

Fixes: 5911d2d1d1a3 ("f2fs: introduce gc_merge mount option")
Signed-off-by: Wenjie Qi <qiwenjie@xxxxxxxxxx>
---
Reproduced on a 512 MiB F2FS image mounted with gc_merge, filled and
fragmented under concurrent inode creation. The kernel reported:

  INFO: task sh:1381 blocked in I/O wait for more than 20 seconds
  Call Trace:
    io_schedule
    f2fs_balance_fs
    f2fs_create

The SysRq task dump showed fsfreeze in freeze_super.part.0. An NMI
backtrace reported:

  CPU: 2 PID: 229 Comm: f2fs_gc-254:0
  Call Trace:
    gc_thread_func

The f2fs_gc kthread also consumed 301 ticks in three seconds at
USER_HZ=100.
---
 fs/f2fs/gc.c | 6 ++++++
 1 file changed, 6 insertions(+)

diff --git a/fs/f2fs/gc.c b/fs/f2fs/gc.c
index 926e6976bc2..086670f9df2 100644
--- a/fs/f2fs/gc.c
+++ b/fs/f2fs/gc.c
@@ -66,6 +66,8 @@ static int gc_thread_func(void *data)
 
 		if (f2fs_readonly(sbi->sb)) {
 			stat_other_skip_bggc_count(sbi);
+			if (foreground)
+				wake_up_all(&gc_th->fggc_wq);
 			continue;
 		}
 		if (kthread_should_stop())
@@ -74,6 +76,8 @@ static int gc_thread_func(void *data)
 		if (sbi->sb->s_writers.frozen >= SB_FREEZE_WRITE) {
 			increase_sleep_time(gc_th, &wait_ms);
 			stat_other_skip_bggc_count(sbi);
+			if (foreground)
+				wake_up_all(&gc_th->fggc_wq);
 			continue;
 		}
 
@@ -83,6 +87,8 @@ static int gc_thread_func(void *data)
 
 		if (!sb_start_write_trylock(sbi->sb)) {
 			stat_other_skip_bggc_count(sbi);
+			if (foreground)
+				wake_up_all(&gc_th->fggc_wq);
 			continue;
 		}
 
-- 
2.43.0