create_discard_cmd_control() is called from f2fs_build_segment_manager(),
before f2fs_build_node_manager() and mount recovery have completed.  It
currently starts the discard thread immediately, so issue_discard_thread()
can run while f2fs_fill_super() is still initializing mount-time state.

This is not the failure-unwind case where free_nm stops the discard
thread before f2fs_destroy_node_manager() frees nm_info.  The window is
earlier: the thread may run while f2fs_build_node_manager() has published
sbi->nm_info but init_node_manager() is still initializing it.  After
commit d6d2b491a82e, issue_discard_thread() may call
f2fs_available_free_memory() and read fields such as nm_i->ram_thresh.

The same early-start window also lets the discard thread observe the
superblock read-only state while mount recovery is still making temporary
SB_RDONLY transitions.

Keep the discard command control available early, but start the discard
thread later in f2fs_fill_super(), after node-manager initialization and
mount recovery have completed.


Fixes: d6d2b491a82e1e411a6766fbfb87c697d8701554 ("f2fs: allow to change discard policy based on cached discard cmds")
Signed-off-by: Cen Zhang <zzzccc427@gmail.com>
---
 fs/f2fs/segment.c | 17 ++++-------------
 fs/f2fs/super.c   | 12 ++++++++++++
 2 files changed, 16 insertions(+), 13 deletions(-)

diff --git a/fs/f2fs/segment.c b/fs/f2fs/segment.c
index 8390994a8826..deb98f564165 100644
--- a/fs/f2fs/segment.c
+++ b/fs/f2fs/segment.c
@@ -2302,12 +2302,10 @@ int f2fs_start_discard_thread(struct f2fs_sb_info *sbi)
 static int create_discard_cmd_control(struct f2fs_sb_info *sbi)
 {
 	struct discard_cmd_control *dcc;
-	int err = 0, i;
+	int i;
 
-	if (SM_I(sbi)->dcc_info) {
-		dcc = SM_I(sbi)->dcc_info;
-		goto init_thread;
-	}
+	if (SM_I(sbi)->dcc_info)
+		return 0;
 
 	dcc = f2fs_kzalloc(sbi, sizeof(struct discard_cmd_control), GFP_KERNEL);
 	if (!dcc)
@@ -2344,14 +2342,7 @@ static int create_discard_cmd_control(struct f2fs_sb_info *sbi)
 
 	init_waitqueue_head(&dcc->discard_wait_queue);
 	SM_I(sbi)->dcc_info = dcc;
-init_thread:
-	err = f2fs_start_discard_thread(sbi);
-	if (err) {
-		kfree(dcc);
-		SM_I(sbi)->dcc_info = NULL;
-	}
-
-	return err;
+	return 0;
 }
 
 static void destroy_discard_cmd_control(struct f2fs_sb_info *sbi)
diff --git a/fs/f2fs/super.c b/fs/f2fs/super.c
index 40079fd7886b..8228be53d036 100644
--- a/fs/f2fs/super.c
+++ b/fs/f2fs/super.c
@@ -5340,6 +5340,15 @@ static int f2fs_fill_super(struct super_block *sb, struct fs_context *fc)
 
 	f2fs_tuning_parameters(sbi);
 
+	/*
+	 * After POR and mount-time recovery, we can run the discard thread. It
+	 * reads node-manager memory thresholds and the superblock read-only
+	 * state, so keep it out of the fill_super() initialization window.
+	 */
+	err = f2fs_start_discard_thread(sbi);
+	if (err)
+		goto leave_shrinker;
+
 	f2fs_notice(sbi, "Mounted with checkpoint version = %llx",
 		    cur_cp_version(F2FS_CKPT(sbi)));
 	f2fs_update_time(sbi, CP_TIME);
@@ -5349,6 +5358,9 @@ static int f2fs_fill_super(struct super_block *sb, struct fs_context *fc)
 	sbi->umount_lock_holder = NULL;
 	return 0;
 
+leave_shrinker:
+	f2fs_leave_shrinker(sbi);
+	f2fs_stop_gc_thread(sbi);
 sync_free_meta:
 	/* safe to flush all the data */
 	sync_filesystem(sbi->sb);
-- 
2.43.0

On 5/4/26 21:35, Cen Zhang wrote:
> create_discard_cmd_control() is called from f2fs_build_segment_manager(),
> before f2fs_build_node_manager() and mount recovery have completed.  It
> currently starts the discard thread immediately, so issue_discard_thread()
> can run while f2fs_fill_super() is still initializing mount-time state.
> 
> This is not the failure-unwind case where free_nm stops the discard
> thread before f2fs_destroy_node_manager() frees nm_info.  The window is
> earlier: the thread may run while f2fs_build_node_manager() has published
> sbi->nm_info but init_node_manager() is still initializing it.  After
> commit d6d2b491a82e, issue_discard_thread() may call
> f2fs_available_free_memory() and read fields such as nm_i->ram_thresh.
> 
> The same early-start window also lets the discard thread observe the
> superblock read-only state while mount recovery is still making temporary
> SB_RDONLY transitions.
> 
> Keep the discard command control available early, but start the discard
> thread later in f2fs_fill_super(), after node-manager initialization and
> mount recovery have completed.
> 
> 

Cc: stable@kernel.org

> Fixes: d6d2b491a82e1e411a6766fbfb87c697d8701554 ("f2fs: allow to change discard policy based on cached discard cmds")

Fixes: d6d2b491a82e ("f2fs: allow to change discard policy based on cached discard cmds")

otherwise, ./scripts/checkpatch.pl will complain as below:

ERROR: Please use git commit description style 'commit <12+ chars of sha1> ("<title line>")' - ie: 'commit d6d2b491a82e ("f2fs: allow to change discard policy based on cached discard cmds")'

Thanks,

> Signed-off-by: Cen Zhang <zzzccc427@gmail.com>
> ---
>   fs/f2fs/segment.c | 17 ++++-------------
>   fs/f2fs/super.c   | 12 ++++++++++++
>   2 files changed, 16 insertions(+), 13 deletions(-)
> 
> diff --git a/fs/f2fs/segment.c b/fs/f2fs/segment.c
> index 8390994a8826..deb98f564165 100644
> --- a/fs/f2fs/segment.c
> +++ b/fs/f2fs/segment.c
> @@ -2302,12 +2302,10 @@ int f2fs_start_discard_thread(struct f2fs_sb_info *sbi)
>   static int create_discard_cmd_control(struct f2fs_sb_info *sbi)
>   {
>   	struct discard_cmd_control *dcc;
> -	int err = 0, i;
> +	int i;
>   
> -	if (SM_I(sbi)->dcc_info) {
> -		dcc = SM_I(sbi)->dcc_info;
> -		goto init_thread;
> -	}
> +	if (SM_I(sbi)->dcc_info)
> +		return 0;
>   
>   	dcc = f2fs_kzalloc(sbi, sizeof(struct discard_cmd_control), GFP_KERNEL);
>   	if (!dcc)
> @@ -2344,14 +2342,7 @@ static int create_discard_cmd_control(struct f2fs_sb_info *sbi)
>   
>   	init_waitqueue_head(&dcc->discard_wait_queue);
>   	SM_I(sbi)->dcc_info = dcc;
> -init_thread:
> -	err = f2fs_start_discard_thread(sbi);
> -	if (err) {
> -		kfree(dcc);
> -		SM_I(sbi)->dcc_info = NULL;
> -	}
> -
> -	return err;
> +	return 0;
>   }
>   
>   static void destroy_discard_cmd_control(struct f2fs_sb_info *sbi)
> diff --git a/fs/f2fs/super.c b/fs/f2fs/super.c
> index 40079fd7886b..8228be53d036 100644
> --- a/fs/f2fs/super.c
> +++ b/fs/f2fs/super.c
> @@ -5340,6 +5340,15 @@ static int f2fs_fill_super(struct super_block *sb, struct fs_context *fc)
>   
>   	f2fs_tuning_parameters(sbi);
>   
> +	/*
> +	 * After POR and mount-time recovery, we can run the discard thread. It
> +	 * reads node-manager memory thresholds and the superblock read-only
> +	 * state, so keep it out of the fill_super() initialization window.
> +	 */
> +	err = f2fs_start_discard_thread(sbi);
> +	if (err)
> +		goto leave_shrinker;
> +
>   	f2fs_notice(sbi, "Mounted with checkpoint version = %llx",
>   		    cur_cp_version(F2FS_CKPT(sbi)));
>   	f2fs_update_time(sbi, CP_TIME);
> @@ -5349,6 +5358,9 @@ static int f2fs_fill_super(struct super_block *sb, struct fs_context *fc)
>   	sbi->umount_lock_holder = NULL;
>   	return 0;
>   
> +leave_shrinker:
> +	f2fs_leave_shrinker(sbi);
> +	f2fs_stop_gc_thread(sbi);
>   sync_free_meta:
>   	/* safe to flush all the data */
>   	sync_filesystem(sbi->sb);