Re: [PATCH v2] f2fs: quiesce background threads during system suspend using PM notifier

Re: [PATCH v2] f2fs: quiesce background threads during system suspend using PM notifier

[Christoph Hellwig]

On Thu, Aug 06, 2026 at 10:01:08AM -0700, Daeho Jeong wrote:
> From: Daeho Jeong <daehojeong@google.com>
> 
> During system suspend, a race condition can cause f2fs_gc and f2fs_discard
> threads to call submit_bio() while the underlying block device (e.g., UFS)
> is in Runtime PM suspend. Because Runtime PM worker threads are already
> frozen during task freezing, the threads become trapped in
> __bio_queue_enter() waiting on mq_freeze_wq, leading to a PM freezer
> timeout.

That does sound like a general issue with our block device / threading
handling.

> To prevent this deadlock, register a PM notifier to set SBI_IS_SUSPENDING
> during PM_SUSPEND_PREPARE. Background GC and discard threads check this
> flag and immediately stop issuing new bios, allowing them to enter a
> freezable sleep state cleanly before process freezing begins.
> 
> In addition, check freezing() as a fast path to stop issuing new I/O
> when non-PM freezing (e.g. dm-snapshot or cgroup freezer) is requested.

.. which means that we really sould have all the relevant parties
invited into figuring out whast is happening here, rather than
band-aiding something that looks like a horrible hack inside a
file system.

Unfortunately I see this a lot with f2fs.  Please reach out to all
relevant maintainers for something that does not look strictly local
to f2fs.


> Signed-off-by: Daeho Jeong <daehojeong@google.com>
> ---
> v2: check freezing() together for non-PM freezing.
> ---
>  fs/f2fs/f2fs.h    |  9 +++++++++
>  fs/f2fs/gc.c      | 13 ++++++++-----
>  fs/f2fs/segment.c | 12 ++++++++----
>  fs/f2fs/super.c   | 25 +++++++++++++++++++++++++
>  4 files changed, 50 insertions(+), 9 deletions(-)
> 
> diff --git a/fs/f2fs/f2fs.h b/fs/f2fs/f2fs.h
> index f1774d4e18d2..2a7b0fe8419b 100644
> --- a/fs/f2fs/f2fs.h
> +++ b/fs/f2fs/f2fs.h
> @@ -25,6 +25,7 @@
>  #include <linux/quotaops.h>
>  #include <linux/part_stat.h>
>  #include <linux/rw_hint.h>
> +#include <linux/suspend.h>
>  
>  #include <linux/fscrypt.h>
>  #include <linux/fsverity.h>
> @@ -1492,6 +1493,7 @@ enum {
>  	SBI_IS_FREEZING,			/* freezefs is in process */
>  	SBI_IS_WRITABLE,			/* remove ro mountoption transiently */
>  	SBI_ENABLE_CHECKPOINT,			/* indicate it's during f2fs_enable_checkpoint() */
> +	SBI_IS_SUSPENDING,			/* system suspend is in progress */
>  	MAX_SBI_FLAG,
>  };
>  
> @@ -1755,6 +1757,7 @@ struct f2fs_sb_info {
>  	struct f2fs_rwsem sb_lock;		/* lock for raw super block */
>  	int valid_super_block;			/* valid super block no */
>  	unsigned long s_flag;				/* flags for sbi */
> +	struct notifier_block pm_nb;		/* for PM notifier */
>  	struct mutex writepages;		/* mutex for writepages() */
>  
>  #ifdef CONFIG_BLK_DEV_ZONED
> @@ -2309,6 +2312,12 @@ static inline void clear_sbi_flag(struct f2fs_sb_info *sbi, unsigned int type)
>  	clear_bit(type, &sbi->s_flag);
>  }
>  
> +static inline bool f2fs_is_suspending(struct f2fs_sb_info *sbi)
> +{
> +	return is_sbi_flag_set(sbi, SBI_IS_SUSPENDING) ||
> +				unlikely(freezing(current));
> +}
> +
>  static inline unsigned long long cur_cp_version(struct f2fs_checkpoint *cp)
>  {
>  	return le64_to_cpu(cp->checkpoint_ver);
> diff --git a/fs/f2fs/gc.c b/fs/f2fs/gc.c
> index ffaa7ba76a1b..6c7ea38eb70d 100644
> --- a/fs/f2fs/gc.c
> +++ b/fs/f2fs/gc.c
> @@ -71,7 +71,8 @@ static int gc_thread_func(void *data)
>  		if (kthread_should_stop())
>  			break;
>  
> -		if (sbi->sb->s_writers.frozen >= SB_FREEZE_WRITE) {
> +		if (sbi->sb->s_writers.frozen >= SB_FREEZE_WRITE ||
> +			f2fs_is_suspending(sbi)) {
>  			increase_sleep_time(gc_th, &wait_ms);
>  			stat_other_skip_bggc_count(sbi);
>  			continue;
> @@ -1064,8 +1065,9 @@ static int gc_node_segment(struct f2fs_sb_info *sbi,
>  		struct node_info ni;
>  		int err;
>  
> -		/* stop BG_GC if there is not enough free sections. */
> -		if (gc_type == BG_GC && has_not_enough_free_secs(sbi, 0, 0))
> +		/* stop BG_GC if there is not enough free sections or suspending/freezing. */
> +		if (gc_type == BG_GC && (has_not_enough_free_secs(sbi, 0, 0) ||
> +					f2fs_is_suspending(sbi)))
>  			return submitted;
>  
>  		if (check_valid_map(sbi, segno, off) == 0)
> @@ -1611,7 +1613,8 @@ static int gc_data_segment(struct f2fs_sb_info *sbi, struct f2fs_summary *sum,
>  		 * Or, stop GC if the segment becomes fully valid caused by
>  		 * race condition along with SSR block allocation.
>  		 */
> -		if ((gc_type == BG_GC && has_not_enough_free_secs(sbi, 0, 0)) ||
> +		if ((gc_type == BG_GC && (has_not_enough_free_secs(sbi, 0, 0) ||
> +					f2fs_is_suspending(sbi))) ||
>  			(!force_migrate && get_valid_blocks(sbi, segno, true) ==
>  							CAP_BLKS_PER_SEC(sbi)))
>  			return submitted;
> @@ -2015,7 +2018,7 @@ int f2fs_gc(struct f2fs_sb_info *sbi, struct f2fs_gc_control *gc_control)
>  		goto stop;
>  	}
>  retry:
> -	if (unlikely(freezing(current))) {
> +	if (f2fs_is_suspending(sbi)) {
>  		ret = 0;
>  		goto stop;
>  	}
> diff --git a/fs/f2fs/segment.c b/fs/f2fs/segment.c
> index d71ddb3ee918..7b03b3d06161 100644
> --- a/fs/f2fs/segment.c
> +++ b/fs/f2fs/segment.c
> @@ -1300,7 +1300,7 @@ static int __submit_discard_cmd(struct f2fs_sb_info *sbi,
>  	if (dc->state != D_PREP)
>  		return 0;
>  
> -	if (is_sbi_flag_set(sbi, SBI_NEED_FSCK))
> +	if (is_sbi_flag_set(sbi, SBI_NEED_FSCK) || f2fs_is_suspending(sbi))
>  		return 0;
>  
>  #ifdef CONFIG_BLK_DEV_ZONED
> @@ -1341,6 +1341,9 @@ static int __submit_discard_cmd(struct f2fs_sb_info *sbi,
>  		unsigned long flags;
>  		bool last = true;
>  
> +		if (f2fs_is_suspending(sbi))
> +			break;
> +
>  		if (len > max_discard_blocks) {
>  			len = max_discard_blocks;
>  			last = false;
> @@ -1615,7 +1618,7 @@ static void __issue_discard_cmd_orderly(struct f2fs_sb_info *sbi,
>  		if (dc->state != D_PREP)
>  			goto next;
>  
> -		if (*issued > 0 && unlikely(freezing(current)))
> +		if (f2fs_is_suspending(sbi))
>  			break;
>  
>  		if (dpolicy->io_aware && !is_idle(sbi, DISCARD_TIME)) {
> @@ -1688,7 +1691,7 @@ static int __issue_discard_cmd(struct f2fs_sb_info *sbi,
>  		list_for_each_entry_safe(dc, tmp, pend_list, list) {
>  			f2fs_bug_on(sbi, dc->state != D_PREP);
>  
> -			if (issued > 0 && unlikely(freezing(current))) {
> +			if (f2fs_is_suspending(sbi)) {
>  				suspended = true;
>  				break;
>  			}
> @@ -1955,7 +1958,8 @@ static int issue_discard_thread(void *data)
>  			continue;
>  		if (kthread_should_stop())
>  			return 0;
> -		if (is_sbi_flag_set(sbi, SBI_NEED_FSCK) ||
> +		if (f2fs_is_suspending(sbi) ||
> +			is_sbi_flag_set(sbi, SBI_NEED_FSCK) ||
>  			!atomic_read(&dcc->discard_cmd_cnt)) {
>  			wait_ms = dpolicy.max_interval;
>  			continue;
> diff --git a/fs/f2fs/super.c b/fs/f2fs/super.c
> index c448d992ff2a..8c97b5d1ee68 100644
> --- a/fs/f2fs/super.c
> +++ b/fs/f2fs/super.c
> @@ -1979,6 +1979,26 @@ static void destroy_device_list(struct f2fs_sb_info *sbi)
>  	kvfree(sbi->devs);
>  }
>  
> +static int f2fs_pm_notifier(struct notifier_block *nb,
> +			    unsigned long action, void *ptr)
> +{
> +	struct f2fs_sb_info *sbi = container_of(nb, struct f2fs_sb_info, pm_nb);
> +
> +	switch (action) {
> +	case PM_HIBERNATION_PREPARE:
> +	case PM_SUSPEND_PREPARE:
> +	case PM_RESTORE_PREPARE:
> +		set_sbi_flag(sbi, SBI_IS_SUSPENDING);
> +		break;
> +	case PM_POST_SUSPEND:
> +	case PM_POST_HIBERNATION:
> +	case PM_POST_RESTORE:
> +		clear_sbi_flag(sbi, SBI_IS_SUSPENDING);
> +		break;
> +	}
> +	return NOTIFY_OK;
> +}
> +
>  static void f2fs_put_super(struct super_block *sb)
>  {
>  	struct f2fs_sb_info *sbi = F2FS_SB(sb);
> @@ -1986,6 +2006,8 @@ static void f2fs_put_super(struct super_block *sb)
>  	int err = 0;
>  	bool done;
>  
> +	unregister_pm_notifier(&sbi->pm_nb);
> +
>  	/* unregister procfs/sysfs entries in advance to avoid race case */
>  	f2fs_unregister_sysfs(sbi);
>  
> @@ -5436,6 +5458,9 @@ static int f2fs_fill_super(struct super_block *sb, struct fs_context *fc)
>  	f2fs_update_time(sbi, REQ_TIME);
>  	clear_sbi_flag(sbi, SBI_CP_DISABLED_QUICK);
>  
> +	sbi->pm_nb.notifier_call = f2fs_pm_notifier;
> +	register_pm_notifier(&sbi->pm_nb);
> +
>  	sbi->umount_lock_holder = NULL;
>  	return 0;
>  
> -- 
> 2.55.0.654.g21b8a5bc05-goog
> 
> 
---end quoted text---

Re: [PATCH v2] f2fs: quiesce background threads during system suspend using PM notifier

[Bart Van Assche]

On 8/10/26 8:54 AM, Christoph Hellwig wrote:
> On Thu, Aug 06, 2026 at 10:01:08AM -0700, Daeho Jeong wrote:
>> From: Daeho Jeong <daehojeong@google.com>
>>
>> During system suspend, a race condition can cause f2fs_gc and f2fs_discard
>> threads to call submit_bio() while the underlying block device (e.g., UFS)
>> is in Runtime PM suspend. Because Runtime PM worker threads are already
>> frozen during task freezing, the threads become trapped in
>> __bio_queue_enter() waiting on mq_freeze_wq, leading to a PM freezer
>> timeout.
> 
> That does sound like a general issue with our block device / threading
> handling.

Daeho's description above mixes up unrelated topics. A runtime suspended 
UFS device is resumed automatically by the block layer if necessary. The
issue Daeho is trying to solve is unrelated to runtime suspend according
to my understanding.

>> To prevent this deadlock, register a PM notifier to set SBI_IS_SUSPENDING
>> during PM_SUSPEND_PREPARE. Background GC and discard threads check this
>> flag and immediately stop issuing new bios, allowing them to enter a
>> freezable sleep state cleanly before process freezing begins.
>>
>> In addition, check freezing() as a fast path to stop issuing new I/O
>> when non-PM freezing (e.g. dm-snapshot or cgroup freezer) is requested.
> 
> .. which means that we really sould have all the relevant parties
> invited into figuring out whast is happening here, rather than
> band-aiding something that looks like a horrible hack inside a
> file system.
> 
> Unfortunately I see this a lot with f2fs.  Please reach out to all
> relevant maintainers for something that does not look strictly local
> to f2fs.

This information was shared earlier with Daeho 
(https://b.corp.google.com/issues/515470309#comment45): [ ... ]
Register a Freezable Kernel Thread (Recommended) [ ... ]
Why this works: During suspend, freezer stops freezable kernel threads 
before devices and block queues enter their PM suspend phases. When the 
queue freezes later, your thread is already safely sleeping in 
try_to_freeze() and will not attempt submit_bio(). [ ... ]

PS: I'm no longer subscribed to the f2fs-devel mailing list.

Bart.

Re: [PATCH v2] f2fs: quiesce background threads during system suspend using PM notifier

[Daeho Jeong]

On Mon, Aug 10, 2026 at 9:34 AM Bart Van Assche <bvanassche@acm.org> wrote:
>
> On 8/10/26 8:54 AM, Christoph Hellwig wrote:
> > On Thu, Aug 06, 2026 at 10:01:08AM -0700, Daeho Jeong wrote:
> >> From: Daeho Jeong <daehojeong@google.com>
> >>
> >> During system suspend, a race condition can cause f2fs_gc and f2fs_discard
> >> threads to call submit_bio() while the underlying block device (e.g., UFS)
> >> is in Runtime PM suspend. Because Runtime PM worker threads are already
> >> frozen during task freezing, the threads become trapped in
> >> __bio_queue_enter() waiting on mq_freeze_wq, leading to a PM freezer
> >> timeout.
> >
> > That does sound like a general issue with our block device / threading
> > handling.
>
> Daeho's description above mixes up unrelated topics. A runtime suspended
> UFS device is resumed automatically by the block layer if necessary. The
> issue Daeho is trying to solve is unrelated to runtime suspend according
> to my understanding.
>
> >> To prevent this deadlock, register a PM notifier to set SBI_IS_SUSPENDING
> >> during PM_SUSPEND_PREPARE. Background GC and discard threads check this
> >> flag and immediately stop issuing new bios, allowing them to enter a
> >> freezable sleep state cleanly before process freezing begins.
> >>
> >> In addition, check freezing() as a fast path to stop issuing new I/O
> >> when non-PM freezing (e.g. dm-snapshot or cgroup freezer) is requested.
> >
> > .. which means that we really sould have all the relevant parties
> > invited into figuring out whast is happening here, rather than
> > band-aiding something that looks like a horrible hack inside a
> > file system.
> >
> > Unfortunately I see this a lot with f2fs.  Please reach out to all
> > relevant maintainers for something that does not look strictly local
> > to f2fs.
>
> This information was shared earlier with Daeho
> (https://b.corp.google.com/issues/515470309#comment45): [ ... ]
> Register a Freezable Kernel Thread (Recommended) [ ... ]
> Why this works: During suspend, freezer stops freezable kernel threads
> before devices and block queues enter their PM suspend phases. When the
> queue freezes later, your thread is already safely sleeping in
> try_to_freeze() and will not attempt submit_bio(). [ ... ]

Hi Bart,

To clarify it:

f2fs threads are already freezable: f2fs_gc and f2fs_discard are
already registered with set_freezable() and call try_to_freeze().
However, a race window exists: a thread checks freezing() (false),
calls submit_bio(), and gets trapped inside __bio_queue_enter().
Because it gets blocked before reaching try_to_freeze(), it triggers a
PM freezer timeout.

Hi Christoph,

Point taken. Aside from this f2fs patch, I agree that addressing this
at the block layer or PM subsystem level would be a much cleaner,
system-wide solution.
I’ll give more thought to how we can properly solve this race
condition for the entire system, and I'll loop in the relevant
PM/block maintainers if a viable generic approach emerges.

Thank you,

>
> PS: I'm no longer subscribed to the f2fs-devel mailing list.
>
> Bart.