Skip to contentSkip to conversation

f2fs: protect published gc_thread during teardown

PATCH Superseded Zhang Cen 1 participant 1 message May 18 F2FS file system

v2 of 5 Superseded Latest is v5 →

lore
Share & tools
lore

Zhang Cen

author
#0 lore raw

PATCH v2 1/1 f2fs: protect published gc_thread during teardown

f2fs stores the background GC worker in a heap-allocated sbi->gc_thread. Foreground GC merge waiters, GC-thread sysfs callbacks, and victim- selection readers can still touch waitqueues, policy knobs, or f2fs_gc_task through that published object, but the dead-device shutdown path can reach f2fs_stop_gc_thread() under shared s_umount coverage and free it.

The buggy scenario involves two paths, with each column showing the order within that path:

f2fs_balance_fs() reader:          dead-device shutdown:
1. sees low free space and         1. enters f2fs_shutdown()
   enters GC_MERGE                 2. calls f2fs_do_shutdown(..., false)
2. borrows sbi->gc_thread          3. calls f2fs_stop_gc_thread()
3. queues on gc_th->fggc_wq        4. stops gc_th->f2fs_gc_task

4. wakes gc_th->gc_wait_queue_head 5. frees gc_th
5. sleeps in io_schedule() 6. clears sbi->gc_thread
6. resumes in finish_wait()

Fix this by making gc_thread a properly published object. Initialize it fully before publishing it, remove it from sbi->gc_thread with xchg() before teardown, keep f2fs_gc_task alive until an SRCU grace period has drained the borrowed readers, and route the waitqueue, scalar, and sysfs users through shared get/put helpers. Recheck f2fs_gc_task after prepare_to_wait() in f2fs_balance_fs() so teardown cannot miss an already-queued foreground waiter.

Validation reproduced this kernel report:
KASAN slab-use-after-free in _raw_spin_lock_irqsave()

Read of size 1
Call trace:
  dump_stack_lvl() (?:?)
  print_address_description() (mm/kasan/report.c:373)
  _raw_spin_lock_irqsave() (?:?)
  print_report() (?:?)
  __virt_addr_valid() (?:?)
  srso_alias_return_thunk() (arch/x86/include/asm/nospec-branch.h:375)
  kasan_addr_to_slab() (mm/kasan/common.c:45)
  kasan_report() (?:?)
  __kasan_check_byte() (mm/kasan/common.c:571)
  lock_acquire() (?:?)
  rcu_is_watching() (?:?)
  prepare_to_wait() (kernel/sched/wait.c:248)
  f2fs_balance_fs() (fs/f2fs/segment.c:448)
  f2fs_write_single_data_page() (fs/f2fs/f2fs.h:4184)
  do_raw_spin_lock() (?:?)
  find_held_lock() (kernel/locking/lockdep.c:5340)
  folio_mkclean() (?:?)
  f2fs_write_cache_pages() (fs/f2fs/data.c:3238)
  __lock_acquire() (kernel/locking/lockdep.c:5077)
  unwind_next_frame() (?:?)
  arch_stack_walk() (?:?)
  __f2fs_write_data_pages() (fs/f2fs/data.c:3565)
  validate_chain() (kernel/locking/lockdep.c:3861)
  insert_work() (kernel/workqueue.c:2220)
  do_writepages() (mm/page-writeback.c:2560)
  filemap_writeback() (mm/filemap.c:371)
  __queue_work() (kernel/workqueue.c:2275)
  file_write_and_wait_range() (?:?)
  f2fs_do_sync_file() (fs/f2fs/file.c:285)
  __lock_release() (kernel/locking/lockdep.c:5511)
  f2fs_sync_file() (fs/f2fs/f2fs.h:3781)
  do_fsync() (fs/sync.c:204)
  __x64_sys_fsync() (?:?)
  do_syscall_64() (arch/x86/entry/syscall_64.c:87)
  entry_SYSCALL_64_after_hwframe() (?:?)
  kasan_save_stack() (mm/kasan/common.c:52)
  kasan_save_track() (mm/kasan/common.c:74)
  __kasan_kmalloc() (?:?)
  f2fs_start_gc_thread() (fs/f2fs/f2fs.h:4203)
  f2fs_fill_super() (fs/f2fs/super.c:4993)

---
v2:
Sashiko.dev pointed out that the sysfs critical_task_priority path could use a stale f2fs_gc_task.
Sashiko.dev pointed out that the GC_MERGE foreground wait could miss teardown wakeups and sleep forever.
Sashiko.dev pointed out that holding gc_thread_lock across kthread_run() could deadlock with reclaim.
Sashiko.dev pointed out that the gc_urgent and gc_idle sysfs paths still touched GC state without safe lifetime protection.
While rechecking the mount retry path, I found that gc_thread_srcu_inited must be reset per f2fs_fill_super() attempt so cleanup only covers the current SRCU state.

 fs/f2fs/f2fs.h    |  26 ++++++++++-
 fs/f2fs/gc.c      |  75 ++++++++++++++++++++----------
 fs/f2fs/gc.h      |   8 +++-
 fs/f2fs/segment.c |  34 ++++++++++----
 fs/f2fs/super.c   |  11 +++++
 fs/f2fs/sysfs.c   | 116 ++++++++++++++++++++++++++++++++++++----------
 6 files changed, 207 insertions(+), 63 deletions(-)
fs/f2fs/f2fs.h +25 -1
index 91f506e7c9cfb..8bf4c2873ac03 100644
@@ -25,6 +25,7 @@
 #include <linux/quotaops.h>
 #include <linux/part_stat.h>
 #include <linux/rw_hint.h>
+#include <linux/srcu.h>
 
 #include <linux/fscrypt.h>
 #include <linux/fsverity.h>
@@ -156,6 +157,8 @@ typedef u32 block_t;	/*
 			 */
 typedef u32 nid_t;
 
+struct f2fs_gc_kthread;
+
 #define COMPRESS_EXT_NUM		16
 
 enum blkzone_allocation_policy {
@@ -1873,7 +1876,8 @@ struct f2fs_sb_info {
 						 * semaphore for GC, avoid
 						 * race between GC and GC or CP
 						 */
-	struct f2fs_gc_kthread	*gc_thread;	/* GC thread */
+	struct f2fs_gc_kthread	*gc_thread;	/* published GC thread */
+	struct srcu_struct gc_thread_srcu;	/* protects gc_thread readers */
 	struct atgc_management am;		/* atgc management */
 	unsigned int cur_victim_sec;		/* current victim section num */
 	unsigned int gc_mode;			/* current GC state */
@@ -4200,6 +4204,26 @@ extern const struct iomap_ops f2fs_iomap_ops;
 /*
  * gc.c
  */
+static inline struct f2fs_gc_kthread *f2fs_get_gc_thread(
+					struct f2fs_sb_info *sbi, int *srcu_idx)
+{
+	struct f2fs_gc_kthread *gc_th;
+
+	*srcu_idx = srcu_read_lock(&sbi->gc_thread_srcu);
+	gc_th = smp_load_acquire(&sbi->gc_thread);
+	if (!gc_th) {
+		srcu_read_unlock(&sbi->gc_thread_srcu, *srcu_idx);
+		*srcu_idx = -1;
+	}
+
+	return gc_th;
+}
+
+static inline void f2fs_put_gc_thread(struct f2fs_sb_info *sbi, int srcu_idx)
+{
+	srcu_read_unlock(&sbi->gc_thread_srcu, srcu_idx);
+}
+
 int f2fs_start_gc_thread(struct f2fs_sb_info *sbi);
 void f2fs_stop_gc_thread(struct f2fs_sb_info *sbi);
 block_t f2fs_start_bidx_of_node(unsigned int node_ofs, struct inode *inode);
fs/f2fs/gc.c +50 -25

128 lines not on this page · raw message

fs/f2fs/gc.h +6 -2
index 6c4d4567571eb..b32d57a70f760 100644
@@ -45,7 +45,10 @@
 
 #define NR_GC_CHECKPOINT_SECS (3)	/* data/node/dentry sections */
 
+struct f2fs_sb_info;
+
 struct f2fs_gc_kthread {
+	struct f2fs_sb_info *sbi;
 	struct task_struct *f2fs_gc_task;
 	wait_queue_head_t gc_wait_queue_head;
 
@@ -193,10 +196,11 @@ static inline bool has_enough_invalid_blocks(struct f2fs_sb_info *sbi)
 			limit_free_user_blocks(invalid_user_blocks));
 }
 
-static inline bool need_to_boost_gc(struct f2fs_sb_info *sbi)
+static inline bool need_to_boost_gc(struct f2fs_sb_info *sbi,
+				    struct f2fs_gc_kthread *gc_th)
 {
 	if (f2fs_sb_has_blkzoned(sbi))
 		return !has_enough_free_blocks(sbi,
-				sbi->gc_thread->boost_zoned_gc_percent);
+				gc_th->boost_zoned_gc_percent);
 	return has_enough_invalid_blocks(sbi);
 }
fs/f2fs/segment.c +24 -10
index 788f8b0502492..cd8d5084c3f31 100644
@@ -444,16 +444,30 @@ 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) {
-		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);
-	} else {
+	if (test_opt(sbi, GC_MERGE)) {
+		struct f2fs_gc_kthread *gc_th;
+		int srcu_idx;
+
+		gc_th = f2fs_get_gc_thread(sbi, &srcu_idx);
+		if (gc_th) {
+			if (READ_ONCE(gc_th->f2fs_gc_task)) {
+				DEFINE_WAIT(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);
+				f2fs_put_gc_thread(sbi, srcu_idx);
+				return;
+			}
+			f2fs_put_gc_thread(sbi, srcu_idx);
+		}
+	}
+
+	{
 		struct f2fs_gc_control gc_control = {
 			.victim_segno = NULL_SEGNO,
 			.init_gc_type = f2fs_sb_has_blkzoned(sbi) ?
fs/f2fs/super.c +11 -0
index ccf806b676f53..a1c3784656fb2 100644
@@ -4954,11 +4954,13 @@ static int f2fs_fill_super(struct super_block *sb, struct fs_context *fc)
 	int recovery, i, valid_super_block;
 	struct curseg_info *seg_i;
 	int retry_cnt = 1;
+	bool gc_thread_srcu_inited = false;
 #ifdef CONFIG_QUOTA
 	bool quota_enabled = false;
 #endif
 
 try_onemore:
+	gc_thread_srcu_inited = false;
 	err = -EINVAL;
 	raw_super = NULL;
 	valid_super_block = -1;
@@ -4993,6 +4995,10 @@ static int f2fs_fill_super(struct super_block *sb, struct fs_context *fc)
 		spin_lock_init(&sbi->inode_lock[i]);
 	}
 	mutex_init(&sbi->flush_lock);
+	err = init_srcu_struct(&sbi->gc_thread_srcu);
+	if (err)
+		goto free_sbi;
+	gc_thread_srcu_inited = true;
 
 	/* set a block size */
 	if (unlikely(!sb_set_blocksize(sb, F2FS_BLKSIZE))) {
@@ -5454,6 +5460,10 @@ static int f2fs_fill_super(struct super_block *sb, struct fs_context *fc)
 #ifdef CONFIG_DEBUG_LOCK_ALLOC
 	lockdep_unregister_key(&sbi->cp_global_sem_key);
 #endif
+	if (gc_thread_srcu_inited) {
+		cleanup_srcu_struct(&sbi->gc_thread_srcu);
+		gc_thread_srcu_inited = false;
+	}
 	kfree(sbi);
 	sb->s_fs_info = NULL;
 
@@ -5538,6 +5548,7 @@ static void kill_f2fs_super(struct super_block *sb)
 #ifdef CONFIG_DEBUG_LOCK_ALLOC
 		lockdep_unregister_key(&sbi->cp_global_sem_key);
 #endif
+		cleanup_srcu_struct(&sbi->gc_thread_srcu);
 		kfree(sbi);
 		sb->s_fs_info = NULL;
 	}
fs/f2fs/sysfs.c +91 -25

195 lines not on this page · raw message

Signature · 1 line
2.43.0