blk-mq: fix race conditions on cpu hotplug handling
v4 cover blk-mq: fix race conditions on cpu hotplug handling
This patchset addresses several race conditions on cpu hotplug handling for blk-mq. All problems can be reproducible by the following script.
while true; do
echo 0 > /sys/devices/system/cpu/cpu1/online echo 1 > /sys/devices/system/cpu/cpu1/online
done &
while true; do
modprobe -r null_blk
modprobe null_blk queue_mode=2 irqmode=1
sleep 0.1
done
* Changes from v3
- Rebased to the latest kernel
- Add Reviewed-by tags
* Changes from v2
- Add regression fix for hctx->tags->cpumask
- Remove BLK_MQ_F_SYSFS_UP per-hctx flag and use mq_sysfs_init_done per-queue flag instead with appropriate locking in order to keep track of 'mq' sysfs entry's life
- Add comments on non-trivial stuffs, suggested by Ming
* Changes from v1
- Release q->mq_map in blk_mq_release()
- Fix deadlock when reading cpu_list
- Fix race freeze and unfreeze
Akinobu Mita (7):
blk-mq: avoid setting hctx->tags->cpumask before allocation
blk-mq: fix sysfs registration/unregistration race
blk-mq: Fix use after of free q->mq_map
blk-mq: fix q->mq_usage_counter access race
blk-mq: avoid inserting requests before establishing new mapping
blk-mq: fix freeze queue race
blk-mq: fix deadlock when reading cpu_list
block/blk-core.c | 1 + block/blk-mq-cpumap.c | 9 +++-- block/blk-mq-sysfs.c | 36 ++++++++++++------ block/blk-mq.c | 100 +++++++++++++++++++++++++++++++++++++------------ block/blk-mq.h | 3 +- include/linux/blk-mq.h | 1 - include/linux/blkdev.h | 8 ++++ 7 files changed, 116 insertions(+), 42 deletions(-)
Signature · 1 line
1.9.1
PATCH v4 1/7 blk-mq: avoid setting hctx->tags->cpumask before allocation
When unmapped hw queue is remapped after CPU topology is changed, hctx->tags->cpumask has to be set after hctx->tags is setup in blk_mq_map_swqueue(), otherwise it causes null pointer dereference.
--- block/blk-mq.c | 9 ++++++++- 1 file changed, 8 insertions(+), 1 deletion(-)
block/blk-mq.c +8 -1
@@ -1811,7 +1811,6 @@ static void blk_mq_map_swqueue(struct request_queue *q) hctx = q->mq_ops->map_queue (q, i);cpumask_set_cpu (i, hctx->cpumask); -cpumask_set_cpu (i, hctx->tags->cpumask); ctx->index_hw = hctx->nr_ctx; hctx->ctxs[hctx->nr_ctx++] = ctx; }
@@ -1851,6 +1850,14 @@ static void blk_mq_map_swqueue(struct request_queue *q) hctx->next_cpu =cpumask_first (hctx->cpumask); hctx->next_cpu_batch =BLK_MQ_CPU_WORK_BATCH ; } + +queue_for_each_ctx (q, ctx, i) { +if (!cpu_online (i)) +continue ; + + hctx = q->mq_ops->map_queue (q, i); +cpumask_set_cpu (i, hctx->tags->cpumask); + } }static void blk_mq_update_tag_set_depth (struct blk_mq_tag_set *set)
Signature · 1 line
1.9.1
Re: [PATCH v4 1/7] blk-mq: avoid setting hctx->tags->cpumask before allocation
Looks good. I actually just came up with exactly the same fix after Keith told me to test hotplug for my nvme changes.
PATCH v4 2/7 blk-mq: fix sysfs registration/unregistration race
In reply to Akinobu MitaThere is a race between cpu hotplug handling and adding/deleting gendisk for blk-mq, where both are trying to register and unregister the same sysfs entries.
null_add_dev
--> blk_mq_init_queue
--> blk_mq_init_allocated_queue
--> add to 'all_q_list' (*)
--> add_disk
--> blk_register_queue
--> blk_mq_register_disk (++)null_del_dev
--> del_gendisk
--> blk_unregister_queue
--> blk_mq_unregister_disk (--)
--> blk_cleanup_queue
--> blk_mq_free_queue
--> del from 'all_q_list' (*)blk_mq_queue_reinit
--> blk_mq_sysfs_unregister (-)
--> blk_mq_sysfs_register (+)While the request queue is added to 'all_q_list' (*), blk_mq_queue_reinit() can be called for the queue anytime by CPU hotplug callback. But blk_mq_sysfs_unregister (-) and blk_mq_sysfs_register (+) in blk_mq_queue_reinit must not be called before blk_mq_register_disk (++) and after blk_mq_unregister_disk (--) is finished. Because '/sys/block/*/mq/' is not exists.
There has already been BLK_MQ_F_SYSFS_UP flag in hctx->flags which can be used to track these sysfs stuff, but it is only fixing this issue partially.
In order to fix it completely, we just need per-queue flag instead of per-hctx flag with appropriate locking. So this introduces q->mq_sysfs_init_done which is properly protected with all_q_mutex.
Also, we need to ensure that blk_mq_map_swqueue() is called with all_q_mutex is held. Since hctx->nr_ctx is reset temporarily and updated in blk_mq_map_swqueue(), so we should avoid blk_mq_register_hctx() seeing the temporary hctx->nr_ctx value in CPU hotplug handling or adding/deleting gendisk .
--- block/blk-mq-sysfs.c | 30 ++++++++++++++++++++++-------- block/blk-mq.c | 6 +++--- include/linux/blk-mq.h | 1 - include/linux/blkdev.h | 2 ++ 4 files changed, 27 insertions(+), 12 deletions(-)
block/blk-mq-sysfs.c +22 -8
@@ -343,7 +343,7 @@ static void blk_mq_unregister_hctx(struct blk_mq_hw_ctx *hctx)struct blk_mq_ctx *ctx;int i; -if (!hctx->nr_ctx || !(hctx->flags &BLK_MQ_F_SYSFS_UP )) +if (!hctx->nr_ctx)return ;hctx_for_each_ctx (hctx, ctx, i)
@@ -358,7 +358,7 @@ static int blk_mq_register_hctx(struct blk_mq_hw_ctx *hctx)struct blk_mq_ctx *ctx;int i, ret; -if (!hctx->nr_ctx || !(hctx->flags &BLK_MQ_F_SYSFS_UP )) +if (!hctx->nr_ctx)return 0 ; ret =kobject_add (&hctx->kobj, &q->mq_kobj,"%u" , hctx->queue_num);
@@ -381,6 +381,8 @@ void blk_mq_unregister_disk(struct gendisk *disk)struct blk_mq_ctx *ctx;int i, j; +blk_mq_disable_hotplug (); +queue_for_each_hw_ctx (q, hctx, i) {blk_mq_unregister_hctx (hctx);
@@ -395,6 +397,9 @@ void blk_mq_unregister_disk(struct gendisk *disk)kobject_put (&q->mq_kobj);kobject_put (&disk_to_dev (disk)->kobj); + + q->mq_sysfs_init_done = false; +blk_mq_enable_hotplug (); }static void blk_mq_sysfs_init (struct request_queue *q)
@@ -425,27 +430,30 @@ int blk_mq_register_disk(struct gendisk *disk)struct blk_mq_hw_ctx *hctx;int ret, i; +blk_mq_disable_hotplug (); +blk_mq_sysfs_init (q); ret =kobject_add (&q->mq_kobj,kobject_get (&dev->kobj),"%s" ,"mq" );if (ret <0 ) -return ret; + goto out;kobject_uevent (&q->mq_kobj,KOBJ_ADD );queue_for_each_hw_ctx (q, hctx, i) { - hctx->flags |=BLK_MQ_F_SYSFS_UP ; ret =blk_mq_register_hctx (hctx);if (ret)break ; } -if (ret) { +if (ret)blk_mq_unregister_disk (disk); -return ret; - } +else + q->mq_sysfs_init_done = true; +out: +blk_mq_enable_hotplug (); -return 0 ; +return ret; }EXPORT_SYMBOL_GPL (blk_mq_register_disk);
@@ -454,6 +462,9 @@ void blk_mq_sysfs_unregister(struct request_queue *q)struct blk_mq_hw_ctx *hctx;int i; +if (!q->mq_sysfs_init_done) +return ; +queue_for_each_hw_ctx (q, hctx, i)blk_mq_unregister_hctx (hctx); }
@@ -463,6 +474,9 @@ int blk_mq_sysfs_register(struct request_queue *q)struct blk_mq_hw_ctx *hctx;int i, ret =0 ; +if (!q->mq_sysfs_init_done) +return ret; +queue_for_each_hw_ctx (q, hctx, i) { ret =blk_mq_register_hctx (hctx);if (ret)
block/blk-mq.c +3 -3
@@ -2035,13 +2035,13 @@ struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set, goto err_hctxs;mutex_lock (&all_q_mutex); -list_add_tail (&q->all_q_node, &all_q_list); -mutex_unlock (&all_q_mutex); +list_add_tail (&q->all_q_node, &all_q_list);blk_mq_add_queue_tag_set (set, q); -blk_mq_map_swqueue (q); +mutex_unlock (&all_q_mutex); +return q; err_hctxs:
include/linux/blk-mq.h +0 -1
@@ -145,7 +145,6 @@ enum {BLK_MQ_F_SHOULD_MERGE =1 <<0 ,BLK_MQ_F_TAG_SHARED =1 <<1 ,BLK_MQ_F_SG_MERGE =1 <<2 , -BLK_MQ_F_SYSFS_UP =1 <<3 ,BLK_MQ_F_DEFER_ISSUE =1 <<4 ,BLK_MQ_F_ALLOC_POLICY_START_BIT =8 , BLK_MQ_F_ALLOC_POLICY_BITS = 1,
include/linux/blkdev.h +2 -0
@@ -456,6 +456,8 @@ struct request_queue {struct blk_mq_tag_set *tag_set;struct list_head tag_set_list;struct bio_set *bio_split; + +bool mq_sysfs_init_done; };#define QUEUE_FLAG_QUEUED 1/* uses generic tag queueing */
Signature · 1 line
1.9.1
Re: [PATCH v4 2/7] blk-mq: fix sysfs registration/unregistration race
Looks good,
PATCH v4 3/7 blk-mq: Fix use after of free q->mq_map
In reply to Akinobu MitaCPU hotplug handling for blk-mq (blk_mq_queue_reinit) updates q->mq_map by blk_mq_update_queue_map() for all request queues in all_q_list. On the other hand, q->mq_map is released before deleting the queue from all_q_list.
So if CPU hotplug event occurs in the window, invalid memory access can happen. Fix it by releasing q->mq_map in blk_mq_release() to make it happen latter than removal from all_q_list.
--- block/blk-mq.c | 8 +++----- 1 file changed, 3 insertions(+), 5 deletions(-)
block/blk-mq.c +3 -5
@@ -1925,6 +1925,9 @@ void blk_mq_release(struct request_queue *q)kfree (hctx); } +kfree (q->mq_map); + q->mq_map =NULL ; +kfree (q->queue_hw_ctx);/* ctx kobj stays in queue_ctx */
@@ -2070,11 +2073,6 @@ void blk_mq_free_queue(struct request_queue *q)blk_mq_free_hw_queues (q, set);percpu_ref_exit (&q->mq_usage_counter); - -kfree (q->mq_map); - - q->mq_map =NULL ; -mutex_lock (&all_q_mutex);list_del_init (&q->all_q_node);mutex_unlock (&all_q_mutex);
Signature · 1 line
1.9.1
Re: [PATCH v4 3/7] blk-mq: Fix use after of free q->mq_map
Looks good,
Can you also add a patch that renames blk_mq_free_queue to blk_mq_cleaup_queue and adds a comment that we should not free any memory here? We had way too many bugs of this kinds unfortunately.
Re: [PATCH v4 3/7] blk-mq: Fix use after of free q->mq_map
2015-09-29 15:52 GMT+09:00 Christoph Hellwig <[email protected]>:
Quoted 7 lines
Looks good,
Reviewed-by Christoph Hellwig <[email protected]>Can you also add a patch that renames
blk_mq_free_queuetoblk_mq_cleaup_queueand adds a comment that we should not free any memory here? We had way too many bugs of this kinds unfortunately.
Renaming blk_mq_free_queue to blk_mq_cleaup_queue sounds good because it is called from blk_cleanup_queue().
How about adding comment like below?
/*
* The resources allocated by blk_mq_init_allocated_queue() are released
* by blk_mq_cleanup_queue() and blk_mq_release().
*
* blk_mq_cleanup_queue() is called from blk_cleanup_queue(), so
* the resources which may be used after blk_cleanup_queue() shouldn't
* be released here. Instead, those should be released by blk_mq_release()
* which is called from blk_release_queue().
*
* For example, loop and md drivers call del_gendisk() after
* blk_cleanup_queue(), so the resources used when accessing sysfs entries
* for blk-mq shouldn't be released by blk_mq_cleanup_queue() as these sysfs
* entries can be accessible before del_gendisk() is called.
*/
Re: [PATCH v4 3/7] blk-mq: Fix use after of free q->mq_map
Quoted 5 lines · Akinobu Mita
2015-09-29 15:52 GMT+09:00 Christoph Hellwig <[email protected]>:
Quoted 4 lines
Looks good,
Reviewed-by Christoph Hellwig <[email protected]>
Quoted 8 lines
Quoted 3 lines
Can you also add a patch that renames
blk_mq_free_queuetoblk_mq_cleaup_queueand adds a comment that we should not free any memory here? We had way too many bugs of this kinds unfortunately.Renaming
blk_mq_free_queuetoblk_mq_cleaup_queuesounds good because it is called fromblk_cleanup_queue().How about adding comment like below?
This looks great. It's a lot more than what I though but it looks correct and useful!
PATCH v4 4/7 blk-mq: fix q->mq_usage_counter access race
In reply to Akinobu MitaCPU hotplug handling for blk-mq (blk_mq_queue_reinit) accesses q->mq_usage_counter while freezing all request queues in all_q_list. On the other hand, q->mq_usage_counter is deinitialized in blk_mq_free_queue() before deleting the queue from all_q_list.
So if CPU hotplug event occurs in the window, percpu_ref_kill() is called with q->mq_usage_counter which has already been marked dead, and it triggers warning. Fix it by deleting the queue from all_q_list earlier than destroying q->mq_usage_counter.
--- block/blk-mq.c | 7 ++++--- 1 file changed, 4 insertions(+), 3 deletions(-)
block/blk-mq.c +4 -3
@@ -2067,15 +2067,16 @@ void blk_mq_free_queue(struct request_queue *q) {struct blk_mq_tag_set *set = q->tag_set; +mutex_lock (&all_q_mutex); +list_del_init (&q->all_q_node); +mutex_unlock (&all_q_mutex); +blk_mq_del_queue_tag_set (q);blk_mq_exit_hw_queues (q, set, set->nr_hw_queues);blk_mq_free_hw_queues (q, set);percpu_ref_exit (&q->mq_usage_counter); -mutex_lock (&all_q_mutex); -list_del_init (&q->all_q_node); -mutex_unlock (&all_q_mutex); }/* Basically redo blk_mq_init_queue with queue frozen */
Signature · 1 line
1.9.1
Re: [PATCH v4 4/7] blk-mq: fix q->mq_usage_counter access race
Looks good,
PATCH v4 5/7 blk-mq: avoid inserting requests before establishing new mapping
In reply to Akinobu MitaNotifier callbacks for CPU_ONLINE action can be run on the other CPU than the CPU which was just onlined. So it is possible for the process running on the just onlined CPU to insert request and run hw queue before establishing new mapping which is done by blk_mq_queue_reinit_notify().
This can cause a problem when the CPU has just been onlined first time since the request queue was initialized. At this time ctx->index_hw for the CPU, which is the index in hctx->ctxs[] for this ctx, is still zero before blk_mq_queue_reinit_notify() is called by notifier callbacks for CPU_ONLINE action.
For example, there is a single hw queue (hctx) and two CPU queues (ctx0 for CPU0, and ctx1 for CPU1). Now CPU1 is just onlined and a request is inserted into ctx1->rq_list and set bit0 in pending bitmap as ctx1->index_hw is still zero.
And then while running hw queue, flush_busy_ctxs() finds bit0 is set in pending bitmap and tries to retrieve requests in hctx->ctxs[0]->rq_list. But htx->ctxs[0] is a pointer to ctx0, so the request in ctx1->rq_list is ignored.
Fix it by ensuring that new mapping is established before onlined cpu starts running.
--- block/blk-mq-cpumap.c | 9 ++++---- block/blk-mq.c | 59 +++++++++++++++++++++++++++++++++++++++------------ block/blk-mq.h | 3 ++- 3 files changed, 52 insertions(+), 19 deletions(-)
block/blk-mq-cpumap.c +5 -4
@@ -31,7 +31,8 @@ static int get_first_sibling(unsigned int cpu)return cpu; } -int blk_mq_update_queue_map (unsigned int *map,unsigned int nr_queues) +int blk_mq_update_queue_map (unsigned int *map,unsigned int nr_queues, +const struct cpumask *online_mask) {unsigned int i, nr_cpus, nr_uniq_cpus, queue, first_sibling;cpumask_var_t cpus;
@@ -41,7 +42,7 @@ int blk_mq_update_queue_map(unsigned int *map, unsigned int nr_queues)cpumask_clear (cpus); nr_cpus = nr_uniq_cpus =0 ; -for_each_online_cpu (i) { +for_each_cpu (i, online_mask) { nr_cpus++; first_sibling =get_first_sibling (i);if (!cpumask_test_cpu (first_sibling, cpus))
@@ -51,7 +52,7 @@ int blk_mq_update_queue_map(unsigned int *map, unsigned int nr_queues) queue =0 ;for_each_possible_cpu (i) { -if (!cpu_online (i)) { +if (!cpumask_test_cpu (i, online_mask)) { map[i] =0 ;continue ; }
@@ -95,7 +96,7 @@ unsigned int *blk_mq_make_queue_map(struct blk_mq_tag_set *set)if (!map)return NULL ; -if (!blk_mq_update_queue_map (map, set->nr_hw_queues)) +if (!blk_mq_update_queue_map (map, set->nr_hw_queues, cpu_online_mask))return map;kfree (map);
block/blk-mq.c +45 -14
122 lines not on this page · raw message
block/blk-mq.h +2 -1
@@ -51,7 +51,8 @@ void blk_mq_disable_hotplug(void);* CPU -> queue mappings */ extern unsigned int *blk_mq_make_queue_map (struct blk_mq_tag_set *set); -extern int blk_mq_update_queue_map (unsigned int *map,unsigned int nr_queues); +extern int blk_mq_update_queue_map (unsigned int *map,unsigned int nr_queues, +const struct cpumask *online_mask);extern int blk_mq_hw_queue_to_node (unsigned int *map,unsigned int );/*
Signature · 1 line
1.9.1
Re: [PATCH v4 5/7] blk-mq: avoid inserting requests before establishing new mapping
Looks good,
PATCH v4 6/7 blk-mq: fix freeze queue race
In reply to Akinobu MitaThere are several race conditions while freezing queue.
When unfreezing queue, there is a small window between decrementing q->mq_freeze_depth to zero and percpu_ref_reinit() call with q->mq_usage_counter. If the other calls blk_mq_freeze_queue_start() in the window, q->mq_freeze_depth is increased from zero to one and percpu_ref_kill() is called with q->mq_usage_counter which is already killed. percpu refcount should be re-initialized before killed again.
Also, there is a race condition while switching to percpu mode. percpu_ref_switch_to_percpu() and percpu_ref_kill() must not be executed at the same time as the following scenario is possible:
1. q->mq_usage_counter is initialized in atomic mode. (atomic counter: 1)
2. After the disk registration, a process like systemd-udev starts accessing the disk, and successfully increases refcount successfully by percpu_ref_tryget_live() in blk_mq_queue_enter(). (atomic counter: 2)
3. In the final stage of initialization, q->mq_usage_counter is being switched to percpu mode by percpu_ref_switch_to_percpu() in blk_mq_finish_init(). But if CONFIG_PREEMPT_VOLUNTARY is enabled, the process is rescheduled in the middle of switching when calling wait_event() in __percpu_ref_switch_to_percpu(). (atomic counter: 2)
4. CPU hotplug handling for blk-mq calls percpu_ref_kill() to freeze request queue. q->mq_usage_counter is decreased and marked as DEAD. Wait until all requests have finished.
(atomic counter: 1)
5. The process rescheduled in the step 3. is resumed and finishes all remaining work in __percpu_ref_switch_to_percpu(). A bias value is added to atomic counter of q->mq_usage_counter. (atomic counter: PERCPU_COUNT_BIAS + 1)
6. A request issed in the step 2. is finished and q->mq_usage_counter is decreased by blk_mq_queue_exit(). q->mq_usage_counter is DEAD, so atomic counter is decreased and no release handler is called. (atomic counter: PERCPU_COUNT_BIAS)
7. CPU hotplug handling in the step 4. will wait forever as q->mq_usage_counter will never be zero.
Also, percpu_ref_reinit() and percpu_ref_kill() must not be executed at the same time. Because both functions could call __percpu_ref_switch_to_percpu() which adds the bias value and initialize percpu counter.
Fix those races by serializing with per-queue mutex.
--- block/blk-core.c | 1 + block/blk-mq-sysfs.c | 2 ++ block/blk-mq.c | 8 ++++++++ include/linux/blkdev.h | 6 ++++++ 4 files changed, 17 insertions(+)
block/blk-core.c +1 -0
@@ -689,6 +689,7 @@ struct request_queue *blk_alloc_queue_node(gfp_t gfp_mask, int node_id)__set_bit (QUEUE_FLAG_BYPASS , &q->queue_flags);init_waitqueue_head (&q->mq_freeze_wq); +mutex_init (&q->mq_freeze_lock);if (blkcg_init_queue (q)) goto fail_bdi;
block/blk-mq-sysfs.c +2 -0
@@ -420,7 +420,9 @@ static void blk_mq_sysfs_init(struct request_queue *q)/* see blk_register_queue() */ void blk_mq_finish_init (struct request_queue *q) { +mutex_lock (&q->mq_freeze_lock);percpu_ref_switch_to_percpu (&q->mq_usage_counter); +mutex_unlock (&q->mq_freeze_lock); }int blk_mq_register_disk (struct gendisk *disk)
block/blk-mq.c +8 -0
@@ -115,11 +115,15 @@ void blk_mq_freeze_queue_start(struct request_queue *q) {int freeze_depth; +mutex_lock (&q->mq_freeze_lock); + freeze_depth =atomic_inc_return (&q->mq_freeze_depth);if (freeze_depth ==1 ) {percpu_ref_kill (&q->mq_usage_counter);blk_mq_run_hw_queues (q, false); } + +mutex_unlock (&q->mq_freeze_lock); }EXPORT_SYMBOL_GPL (blk_mq_freeze_queue_start);
@@ -143,12 +147,16 @@ void blk_mq_unfreeze_queue(struct request_queue *q) {int freeze_depth; +mutex_lock (&q->mq_freeze_lock); + freeze_depth =atomic_dec_return (&q->mq_freeze_depth);WARN_ON_ONCE (freeze_depth <0 );if (!freeze_depth) {percpu_ref_reinit (&q->mq_usage_counter);wake_up_all (&q->mq_freeze_wq); } + +mutex_unlock (&q->mq_freeze_lock); }EXPORT_SYMBOL_GPL (blk_mq_unfreeze_queue);
include/linux/blkdev.h +6 -0
@@ -450,6 +450,12 @@ struct request_queue {#endif struct rcu_head rcu_head;wait_queue_head_t mq_freeze_wq; +/* +* Protect concurrent access to mq_usage_counter by +* percpu_ref_switch_to_percpu(), percpu_ref_kill(), and +* percpu_ref_reinit(). +*/ +struct mutex mq_freeze_lock;struct percpu_ref mq_usage_counter;struct list_head all_q_node;
Signature · 1 line
1.9.1
Re: [PATCH v4 6/7] blk-mq: fix freeze queue race
Hello,
block/blk-mq-sysfs.c static void blk_mq_sysfs_init(struct request_queue *q) · line 425
Quoted 7 lines · Akinobu Mita
block/blk-mq-sysfs.c +2
-0@@ -420,7 +420,9 @@ static void blk_mq_sysfs_init(struct request_queue *q)/* see blk_register_queue() */ void blk_mq_finish_init (struct request_queue *q) { +mutex_lock (&q->mq_freeze_lock);percpu_ref_switch_to_percpu (&q->mq_usage_counter); +mutex_unlock (&q->mq_freeze_lock);
This looks weird to me. What can it race against at this point?
block/blk-mq.c void blk_mq_freeze_queue_start(struct request_queue *q) · line 120
Quoted 7 lines
block/blk-mq.c +2
-0@@ -115,11 +115,15 @@ void blk_mq_freeze_queue_start(struct request_queue *q) {int freeze_depth; +mutex_lock (&q->mq_freeze_lock); + freeze_depth =atomic_inc_return (&q->mq_freeze_depth);
It doesn't have to be an atomic anymore, right?
Thanks.
Signature · 1 line
tejun
Re: [PATCH v4 6/7] blk-mq: fix freeze queue race
Hi Tejun,
2015-09-27 2:32 GMT+09:00 Tejun Heo <[email protected]>:
Quoted 12 lines
Hello,
block/blk-mq-sysfs.c static void blk_mq_sysfs_init(struct request_queue *q) · line 425
Quoted 7 lines · Akinobu Mita
block/blk-mq-sysfs.c +2
-0@@ -420,7 +420,9 @@ static void blk_mq_sysfs_init(struct request_queue *q)/* see blk_register_queue() */ void blk_mq_finish_init (struct request_queue *q) { +mutex_lock (&q->mq_freeze_lock);percpu_ref_switch_to_percpu (&q->mq_usage_counter); +mutex_unlock (&q->mq_freeze_lock);This looks weird to me. What can it race against at this point?
The possible scenario is described in commit log (1. ~ 7.). In summary, blk_mq_finish_init() and blk_mq_freeze_queue_start() can be executed at the same time, so this is required to serialize the execution of percpu_ref_switch_to_percpu() by blk_mq_finish_init() and percpu_ref_kill() by blk_mq_freeze_queue_start().
Quoted 9 lines
block/blk-mq.c void blk_mq_freeze_queue_start(struct request_queue *q) · line 120
Quoted 7 lines
block/blk-mq.c +2
-0@@ -115,11 +115,15 @@ void blk_mq_freeze_queue_start(struct request_queue *q) {int freeze_depth; +mutex_lock (&q->mq_freeze_lock); + freeze_depth =atomic_inc_return (&q->mq_freeze_depth);It doesn't have to be an atomic anymore, right?
Yes, you are right. I would like to make it in another patch in order to simplify each change.
Re: [PATCH v4 6/7] blk-mq: fix freeze queue race
Hello,
Quoted 6 lines · Akinobu Mita
Quoted 6 lines
Quoted 5 lines
void
blk_mq_finish_init(struct request_queue *q)
{block/blk-mq-sysfs.c +2
-0+mutex_lock (&q->mq_freeze_lock);percpu_ref_switch_to_percpu (&q->mq_usage_counter); +mutex_unlock (&q->mq_freeze_lock);
Quoted 7 lines
Quoted 1 line
This looks weird to me. What can it race against at this point?
The possible scenario is described in commit log (1. ~ 7.). In summary,
blk_mq_finish_init()andblk_mq_freeze_queue_start()can be executed at the same time, so this is required to serialize the execution ofpercpu_ref_switch_to_percpu()byblk_mq_finish_init()andpercpu_ref_kill()byblk_mq_freeze_queue_start().
Ah, you're right. I was thinking that percpu_ref_switch_to_percpu() being called after blk_mq_freeze_queue_start() would be buggy and thus the above can't be enough but that is safe as long as the calls are properly synchronized. Hmmm... maybe we should add synchronization to those operations from percpu_ref side.
Thanks.
Signature · 1 line
tejun
Re: [PATCH v4 6/7] blk-mq: fix freeze queue race
Quoted 5 lines · Tejun Heo
Ah, you're right. I was thinking that
percpu_ref_switch_to_percpu()being called afterblk_mq_freeze_queue_start()would be buggy and thus the above can't be enough but that is safe as long as the calls are properly synchronized. Hmmm... maybe we should add synchronization to those operations frompercpu_refside.
I think that would be very useful, as most users of the code will have similar issues.
Jens Axboe
maintainerRe: [PATCH v4 6/7] blk-mq: fix freeze queue race
In reply to Tejun HeoQuoted 17 lines · Tejun Heo
Hello,
Quoted 6 lines · Akinobu Mita
Quoted 6 lines
Quoted 5 lines
void blk_mq_finish_init(struct request_queue *q) {block/blk-mq-sysfs.c +2
-0+mutex_lock (&q->mq_freeze_lock);percpu_ref_switch_to_percpu (&q->mq_usage_counter); +mutex_unlock (&q->mq_freeze_lock);Quoted 7 lines
Quoted 1 line
This looks weird to me. What can it race against at this point?
The possible scenario is described in commit log (1. ~ 7.). In summary,
blk_mq_finish_init()andblk_mq_freeze_queue_start()can be executed at the same time, so this is required to serialize the execution ofpercpu_ref_switch_to_percpu()byblk_mq_finish_init()andpercpu_ref_kill()byblk_mq_freeze_queue_start().
Quoted 5 lines
Ah, you're right. I was thinking that
percpu_ref_switch_to_percpu()being called afterblk_mq_freeze_queue_start()would be buggy and thus the above can't be enough but that is safe as long as the calls are properly synchronized. Hmmm... maybe we should add synchronization to those operations frompercpu_refside.
I think that would be very useful, it seems sort of half-assed if the caller side has to provide serialization for that.
Signature · 1 line
Jens Axboe
Re: [PATCH v4 6/7] blk-mq: fix freeze queue race
Hello,
Quoted 2 lines · Jens Axboe
I think that would be very useful, it seems sort of half-assed if the caller side has to provide serialization for that.
Yeah, the thing is init/exit are usually caller synchronized but percpu_rwsem's kill/reinit are more of mode-switching operations which can be performed concurrently during operation so I think the right thing to do here is making it synchronize itself. Will spin a patch.
Thanks.
Signature · 1 line
tejun
Re: [PATCH v4 6/7] blk-mq: fix freeze queue race
Hello,
Quoted 8 lines · Tejun Heo
Quoted 2 lines · Jens Axboe
I think that would be very useful, it seems sort of half-assed if the caller side has to provide serialization for that.
Yeah, the thing is init/exit are usually caller synchronized but percpu_rwsem's kill/reinit are more of mode-switching operations which can be performed concurrently during operation so I think the right thing to do here is making it synchronize itself. Will spin a patch.
Patchset posted
http://lkml.kernel.org/g/[email protected]
Unfortunately, the required changes on percpu_ref side ended up somewhat invasive and don't seem fit for -stable backport. I think it'd be a good idea to apply Akinobu's patch for now and then revert it later when the percpu_ref changes hit mainline.
Thanks.
Signature · 1 line
tejun
Re: [PATCH v4 6/7] blk-mq: fix freeze queue race
2015-09-30 6:50 GMT+09:00 Tejun Heo <[email protected]>:
Quoted 7 lines
Hello,
Quoted 4 lines · Tejun Heo
Quoted 2 lines · Jens Axboe
I think that would be very useful, it seems sort of half-assed if the caller side has to provide serialization for that.
Quoted 8 lines
Quoted 4 lines
Yeah, the thing is init/exit are usually caller synchronized but percpu_rwsem's kill/reinit are more of mode-switching operations which can be performed concurrently during operation so I think the right thing to do here is making it synchronize itself. Will spin a patch.
Patchset posted
Thanks. So far it's working well without any problems.
Note that we still need a part of patch 6/7.
Quoted from this patch description:
When unfreezing queue, there is a small window between decrementing
q->mq_freeze_depth to zero and percpu_ref_reinit() call with
q->mq_usage_counter. If the other calls blk_mq_freeze_queue_start()
in the window, q->mq_freeze_depth is increased from zero to one and
percpu_ref_kill() is called with q->mq_usage_counter which is already
killed. percpu refcount should be re-initialized before killed again.So we don't need to protect percpu_ref_switch_to_percpu() in blk_mq_finish_init() anymore by your percpu_ref patchset, but we still need to serialize blk_mq_freeze_queue_start() and blk_mq_unfreeze_queue(). (As you suggested earlier in this thread, q->mq_freeze_depth don't have to be atomic_t anymore)
PATCH v4 7/7 blk-mq: fix deadlock when reading cpu_list
In reply to Akinobu MitaCPU hotplug handling for blk-mq (blk_mq_queue_reinit) acquires all_q_mutex in blk_mq_queue_reinit_notify() and then removes sysfs entries by blk_mq_sysfs_unregister(). Removing sysfs entry needs to be blocked until the active reference of the kernfs_node to be zero.
On the other hand, reading blk_mq_hw_sysfs_cpu sysfs entry (e.g. /sys/block/nullb0/mq/0/cpu_list) acquires all_q_mutex in blk_mq_hw_sysfs_cpus_show().
If these happen at the same time, a deadlock can happen. Because one can wait for the active reference to be zero with holding all_q_mutex, and the other tries to acquire all_q_mutex with holding the active reference.
The reason that all_q_mutex is acquired in blk_mq_hw_sysfs_cpus_show() is to avoid reading an imcomplete hctx->cpumask. Since reading sysfs entry for blk-mq needs to acquire q->sysfs_lock, we can avoid deadlock and reading an imcomplete hctx->cpumask by protecting q->sysfs_lock while hctx->cpumask is being updated.
--- block/blk-mq-sysfs.c | 4 ---- block/blk-mq.c | 7 +++++++ 2 files changed, 7 insertions(+), 4 deletions(-)
block/blk-mq-sysfs.c +0 -4
@@ -229,8 +229,6 @@ static ssize_t blk_mq_hw_sysfs_cpus_show(struct blk_mq_hw_ctx *hctx, char *page)unsigned int i, first =1 ;ssize_t ret =0 ; -blk_mq_disable_hotplug (); -for_each_cpu (i, hctx->cpumask) {if (first) ret +=sprintf (ret + page,"%u" , i);
@@ -240,8 +238,6 @@ static ssize_t blk_mq_hw_sysfs_cpus_show(struct blk_mq_hw_ctx *hctx, char *page) first =0 ; } -blk_mq_enable_hotplug (); - ret +=sprintf (ret + page,"\n" );return ret; }
block/blk-mq.c +7 -0
@@ -1805,6 +1805,11 @@ static void blk_mq_map_swqueue(struct request_queue *q,struct blk_mq_ctx *ctx;struct blk_mq_tag_set *set = q->tag_set; +/* +* Avoid others reading imcomplete hctx->cpumask through sysfs +*/ +mutex_lock (&q->sysfs_lock); +queue_for_each_hw_ctx (q, hctx, i) {cpumask_clear (hctx->cpumask); hctx->nr_ctx =0 ;
@@ -1824,6 +1829,8 @@ static void blk_mq_map_swqueue(struct request_queue *q, hctx->ctxs[hctx->nr_ctx++] = ctx; } +mutex_unlock (&q->sysfs_lock); +queue_for_each_hw_ctx (q, hctx, i) {struct blk_mq_ctxmap *map = &hctx->ctx_map;
Signature · 1 line
1.9.1
Re: [PATCH v4 7/7] blk-mq: fix deadlock when reading cpu_list
Looks good,
Jens Axboe
maintainerQuoted 14 lines · Akinobu Mita
This patchset addresses several race conditions on cpu hotplug handling for blk-mq. All problems can be reproducible by the following script.
while true; do
echo 0 >/sys/devices/system/cpu/cpu1/onlineecho 1 >/sys/devices/system/cpu/cpu1/online
done &while true; do
modprobe -rnull_blk
modprobenull_blkqueue_mode=2 irqmode=1
sleep 0.1
done
Quoted 3 lines
* Changes from v3
- Rebased to the latest kernel
- Add Reviewed-by tags
I've reviewed the patches, looks good to me. I've skipped 6/7, as I think we should fix that as part of the percpu ref counting, and not handle it specifically in blk-mq. If we can't get that fix before 4.3 wraps up, then we can revisit and put that in as well.
Thanks!
Signature · 1 line
Jens Axboe