Skip to contentSkip to conversation

block: fix q->blkg_list corruption during disk rebind

PATCH Applied Ming Lei 4 participants 6 messages Apr 7, 2024 – Apr 8, 2024

Subsystems (2)
lore
Share & tools
lore

Ming Lei

author
#0 lore raw

PATCH 1/1 block: fix q->blkg_list corruption during disk rebind

Multiple gendisk instances can allocated/added for single request queue in case of disk rebind. blkg may still stay in q->blkg_list when calling blkcg_init_disk() for rebind, then q->blkg_list becomes corrupted.

Fix the list corruption issue by:

- add blkg_init_queue() to initialize q->blkg_list & q->blkcg_mutex only
- move calling blkg_init_queue() into blk_alloc_queue()

The list corruption should be started since commit f1c006f1c685 ("blk-cgroup:
synchronize pd_free_fn() from blkg_free_workfn() and blkcg_deactivate_policy()") which delays removing blkg from q->blkg_list into blkg_free_workfn().

Fixes f1c006f1c685 ("blk-cgroup: synchronize pd_free_fn() from blkg_free_workfn() and blkcg_deactivate_policy()")
Fixes 1059699f87eb ("block: move blkcg initialization/destroy into disk allocation/release handler")
Cc Yu Kuai <[email protected]>
Cc Tejun Heo <[email protected]>

---
blktests:

	https://lore.kernel.org/linux-block/[email protected]/
 block/blk-cgroup.c | 9 ++++++---
 block/blk-cgroup.h | 2 ++
 block/blk-core.c   | 2 ++
 3 files changed, 10 insertions(+), 3 deletions(-)
block/blk-cgroup.c +6 -3
index bdbb557feb5a..059467086b13 100644
@@ -1409,6 +1409,12 @@ static int blkcg_css_online(struct cgroup_subsys_state *css)
 	return 0;
 }
 
+void blkg_init_queue(struct request_queue *q)
+{
+	INIT_LIST_HEAD(&q->blkg_list);
+	mutex_init(&q->blkcg_mutex);
+}
+
 int blkcg_init_disk(struct gendisk *disk)
 {
 	struct request_queue *q = disk->queue;
@@ -1416,9 +1422,6 @@ int blkcg_init_disk(struct gendisk *disk)
 	bool preloaded;
 	int ret;
 
-	INIT_LIST_HEAD(&q->blkg_list);
-	mutex_init(&q->blkcg_mutex);
-
 	new_blkg = blkg_alloc(&blkcg_root, disk, GFP_KERNEL);
 	if (!new_blkg)
 		return -ENOMEM;
block/blk-cgroup.h +2 -0
index 78b74106bf10..90b3959d88cf 100644
@@ -189,6 +189,7 @@ struct blkcg_policy {
 extern struct blkcg blkcg_root;
 extern bool blkcg_debug_stats;
 
+void blkg_init_queue(struct request_queue *q);
 int blkcg_init_disk(struct gendisk *disk);
 void blkcg_exit_disk(struct gendisk *disk);
 
@@ -482,6 +483,7 @@ struct blkcg {
 };
 
 static inline struct blkcg_gq *blkg_lookup(struct blkcg *blkcg, void *key) { return NULL; }
+static inline void blkg_init_queue(struct request_queue *q) { }
 static inline int blkcg_init_disk(struct gendisk *disk) { return 0; }
 static inline void blkcg_exit_disk(struct gendisk *disk) { }
 static inline int blkcg_policy_register(struct blkcg_policy *pol) { return 0; }
block/blk-core.c +2 -0
index a16b5abdbbf5..3a6f5603fb44 100644
@@ -442,6 +442,8 @@ struct request_queue *blk_alloc_queue(struct queue_limits *lim, int node_id)
 	init_waitqueue_head(&q->mq_freeze_wq);
 	mutex_init(&q->mq_freeze_lock);
 
+	blkg_init_queue(q);
+
 	/*
 	 * Init percpu_ref in atomic mode so that it's faster to shutdown.
 	 * See blk_register_queue() for details.
Signature · 1 line
2.41.0
Reviewed-by on v1 1/1
#1 lore raw

Hi,

Quoted 4 lines · Ming Lei

Multiple gendisk instances can allocated/added for single request queue in case of disk rebind. blkg may still stay in q->blkg_list when calling blkcg_init_disk() for rebind, then q->blkg_list becomes corrupted.

Quoted 8 lines

Fix the list corruption issue by:

- add blkg_init_queue() to initialize q->blkg_list & q->blkcg_mutex only
- move calling blkg_init_queue() into blk_alloc_queue()

The list corruption should be started since commit f1c006f1c685 ("blk-cgroup:
synchronize pd_free_fn() from blkg_free_workfn() and blkcg_deactivate_policy()") which delays removing blkg from q->blkg_list into blkg_free_workfn().

I'm not familiar with how bind/unbind works yet, however, the patch itself looks reasonable to me, the initialization of fields related to queue should not be delayed to disk allocation.

BTW, it looks like the whole blkcg_init_disk() can go away:
- init of ioprio and blk-throttle can be delayed to the first user
configuration;
- root_blkg allocation doesn't rely on disk at all;

Or is there any plan to move the blkcg related field or code path to gendisk instead of queue? I might missing some previous discussions.

Thanks,
Kuai

Quoted 76 lines
Fixes f1c006f1c685 ("blk-cgroup: synchronize pd_free_fn() from blkg_free_workfn() and blkcg_deactivate_policy()")
Fixes 1059699f87eb ("block: move blkcg initialization/destroy into disk allocation/release handler")
Cc Yu Kuai <[email protected]>
Cc Tejun Heo <[email protected]>

---
blktests:

	https://lore.kernel.org/linux-block/[email protected]/
  block/blk-cgroup.c | 9 ++++++---
  block/blk-cgroup.h | 2 ++
  block/blk-core.c   | 2 ++
  3 files changed, 10 insertions(+), 3 deletions(-)
block/blk-cgroup.c +6 -3
index bdbb557feb5a..059467086b13 100644
@@ -1409,6 +1409,12 @@ static int blkcg_css_online(struct cgroup_subsys_state *css)
  	return 0;
  }
 
+void blkg_init_queue(struct request_queue *q)
+{
+	INIT_LIST_HEAD(&q->blkg_list);
+	mutex_init(&q->blkcg_mutex);
+}
+
  int blkcg_init_disk(struct gendisk *disk)
  {
  	struct request_queue *q = disk->queue;
@@ -1416,9 +1422,6 @@ int blkcg_init_disk(struct gendisk *disk)
  	bool preloaded;
  	int ret;
 
-	INIT_LIST_HEAD(&q->blkg_list);
-	mutex_init(&q->blkcg_mutex);
-
  	new_blkg = blkg_alloc(&blkcg_root, disk, GFP_KERNEL);
  	if (!new_blkg)
  		return -ENOMEM;
block/blk-cgroup.h +2 -0
index 78b74106bf10..90b3959d88cf 100644
@@ -189,6 +189,7 @@ struct blkcg_policy {
  extern struct blkcg blkcg_root;
  extern bool blkcg_debug_stats;
 
+void blkg_init_queue(struct request_queue *q);
  int blkcg_init_disk(struct gendisk *disk);
  void blkcg_exit_disk(struct gendisk *disk);
 
@@ -482,6 +483,7 @@ struct blkcg {
  };
 
  static inline struct blkcg_gq *blkg_lookup(struct blkcg *blkcg, void *key) { return NULL; }
+static inline void blkg_init_queue(struct request_queue *q) { }
  static inline int blkcg_init_disk(struct gendisk *disk) { return 0; }
  static inline void blkcg_exit_disk(struct gendisk *disk) { }
  static inline int blkcg_policy_register(struct blkcg_policy *pol) { return 0; }
block/blk-core.c +2 -0
index a16b5abdbbf5..3a6f5603fb44 100644
@@ -442,6 +442,8 @@ struct request_queue *blk_alloc_queue(struct queue_limits *lim, int node_id)
  	init_waitqueue_head(&q->mq_freeze_wq);
  	mutex_init(&q->mq_freeze_lock);
 
+	blkg_init_queue(q);
+
  	/*
  	 * Init percpu_ref in atomic mode so that it's faster to shutdown.
  	 * See blk_register_queue() for details.
author
#2 lore raw
Quoted 16 lines · Yu Kuai

Hi,

Quoted 4 lines · Ming Lei

Multiple gendisk instances can allocated/added for single request queue in case of disk rebind. blkg may still stay in q->blkg_list when calling blkcg_init_disk() for rebind, then q->blkg_list becomes corrupted.

Quoted 8 lines

Fix the list corruption issue by:

- add blkg_init_queue() to initialize q->blkg_list & q->blkcg_mutex only
- move calling blkg_init_queue() into blk_alloc_queue()

The list corruption should be started since commit f1c006f1c685 ("blk-cgroup:
synchronize pd_free_fn() from blkg_free_workfn() and blkcg_deactivate_policy()") which delays removing blkg from q->blkg_list into blkg_free_workfn().

Quoted 5 lines

I'm not familiar with how bind/unbind works yet, however, the patch itself looks reasonable to me, the initialization of fields related to queue should not be delayed to disk allocation.

Thanks!

Quoted 8 lines

BTW, it looks like the whole blkcg_init_disk() can go away:
- init of ioprio and blk-throttle can be delayed to the first user
configuration;
- root_blkg allocation doesn't rely on disk at all;

Or is there any plan to move the blkcg related field or code path to gendisk instead of queue? I might missing some previous discussions.

Christoph worked towards this direction, but not succeed:

a06377c5d01e Revert "blk-cgroup: pin the gendisk in struct blkcg_gq" 9a9c261e6b55 Revert "blk-cgroup: pass a gendisk to blkg_lookup" 1231039db31c Revert "blk-cgroup: move the cgroup information to struct gendisk" dcb522014351 Revert "blk-cgroup: simplify blkg freeing from initialization failure paths"

Thanks,
Ming

maintainer Applied
#3 lore raw
In reply to Ming Lei
Quoted 10 lines · Ming Lei

Multiple gendisk instances can allocated/added for single request queue in case of disk rebind. blkg may still stay in q->blkg_list when calling blkcg_init_disk() for rebind, then q->blkg_list becomes corrupted.

Fix the list corruption issue by:

- add blkg_init_queue() to initialize q->blkg_list & q->blkcg_mutex only
- move calling blkg_init_queue() into blk_alloc_queue()

[...]

Applied, thanks!

[1/1] block: fix q->blkg_list corruption during disk rebind

      commit: 8b8ace080319a866f5dfe9da8e665ae51d971c54

Best regards,

Signature · 1 line
Jens Axboe
#4 lore raw
In reply to Ming Lei
Quoted 5 lines · Ming Lei

Multiple gendisk instances can allocated/added for single request queue in case of disk rebind. blkg may still stay in q->blkg_list when calling blkcg_init_disk() for rebind, then q->blkg_list becomes corrupted.

Fix the list corruption issue by:

The right fix is to move the blkgs to the gendisk as there is no use for them on a request_queue without a gendisk.

author
#5 lore raw
Quoted 9 lines · Christoph Hellwig
Quoted 5 lines · Ming Lei

Multiple gendisk instances can allocated/added for single request queue in case of disk rebind. blkg may still stay in q->blkg_list when calling blkcg_init_disk() for rebind, then q->blkg_list becomes corrupted.

Fix the list corruption issue by:

The right fix is to move the blkgs to the gendisk as there is no use for them on a request_queue without a gendisk.

The fix needs to be backported to stable, so it isn't good to fix in that aggressive way, especially last time your attempt failed, please see the revert commits in my last reply to Yu Kuai.

Thanks,
Ming