[PATCH v2] blk-cgroup: Use cond_resched() when destroy blkgs

[PATCH v2] blk-cgroup: Use cond_resched() when destroy blkgs

[Baolin Wang]

On !PREEMPT kernel, we can get below softlockup when doing stress
testing with creating and destroying block cgroup repeatly. The
reason is it may take a long time to acquire the queue's lock in
the loop of blkcg_destroy_blkgs(), or the system can accumulate a
huge number of blkgs in pathological cases. We can add a need_resched()
check on each loop and release locks and do cond_resched() if true
to avoid this issue, since the blkcg_destroy_blkgs() is not called
from atomic contexts.

[ 4757.010308] watchdog: BUG: soft lockup - CPU#11 stuck for 94s!
[ 4757.010698] Call trace:
[ 4757.010700]  blkcg_destroy_blkgs+0x68/0x150
[ 4757.010701]  cgwb_release_workfn+0x104/0x158
[ 4757.010702]  process_one_work+0x1bc/0x3f0
[ 4757.010704]  worker_thread+0x164/0x468
[ 4757.010705]  kthread+0x108/0x138

Suggested-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Baolin Wang <baolin.wang@linux.alibaba.com>
---
Changes from v1:
 - Add might_sleep() in blkcg_destroy_blkgs().
 - Add an explicitly need_resched() check before releasing lock.
 - Add some comments.
---
 block/blk-cgroup.c | 13 +++++++++++++
 1 file changed, 13 insertions(+)

diff --git a/block/blk-cgroup.c b/block/blk-cgroup.c
index 3465d6e..94eeed7 100644
--- a/block/blk-cgroup.c
+++ b/block/blk-cgroup.c
@@ -1016,6 +1016,8 @@ static void blkcg_css_offline(struct cgroup_subsys_state *css)
  */
 void blkcg_destroy_blkgs(struct blkcg *blkcg)
 {
+	might_sleep();
+
 	spin_lock_irq(&blkcg->lock);
 
 	while (!hlist_empty(&blkcg->blkg_list)) {
@@ -1031,6 +1033,17 @@ void blkcg_destroy_blkgs(struct blkcg *blkcg)
 			cpu_relax();
 			spin_lock_irq(&blkcg->lock);
 		}
+
+		/*
+		 * Given that the system can accumulate a huge number
+		 * of blkgs in pathological cases, check to see if we
+		 * need to rescheduling to avoid softlockup.
+		 */
+		if (need_resched()) {
+			spin_unlock_irq(&blkcg->lock);
+			cond_resched();
+			spin_lock_irq(&blkcg->lock);
+		}
 	}
 
 	spin_unlock_irq(&blkcg->lock);
-- 
1.8.3.1

Re: [PATCH v2] blk-cgroup: Use cond_resched() when destroy blkgs

[Tejun Heo]

On Thu, Jan 28, 2021 at 11:22:00AM +0800, Baolin Wang wrote:
> On !PREEMPT kernel, we can get below softlockup when doing stress
> testing with creating and destroying block cgroup repeatly. The
> reason is it may take a long time to acquire the queue's lock in
> the loop of blkcg_destroy_blkgs(), or the system can accumulate a
> huge number of blkgs in pathological cases. We can add a need_resched()
> check on each loop and release locks and do cond_resched() if true
> to avoid this issue, since the blkcg_destroy_blkgs() is not called
> from atomic contexts.
> 
> [ 4757.010308] watchdog: BUG: soft lockup - CPU#11 stuck for 94s!
> [ 4757.010698] Call trace:
> [ 4757.010700]  blkcg_destroy_blkgs+0x68/0x150
> [ 4757.010701]  cgwb_release_workfn+0x104/0x158
> [ 4757.010702]  process_one_work+0x1bc/0x3f0
> [ 4757.010704]  worker_thread+0x164/0x468
> [ 4757.010705]  kthread+0x108/0x138
> 
> Suggested-by: Tejun Heo <tj@kernel.org>
> Signed-off-by: Baolin Wang <baolin.wang@linux.alibaba.com>

Acked-by: Tejun Heo <tj@kernel.org>

Thanks.

-- 
tejun

Re: [PATCH v2] blk-cgroup: Use cond_resched() when destroy blkgs

[Jens Axboe]

On 1/27/21 8:22 PM, Baolin Wang wrote:
> On !PREEMPT kernel, we can get below softlockup when doing stress
> testing with creating and destroying block cgroup repeatly. The
> reason is it may take a long time to acquire the queue's lock in
> the loop of blkcg_destroy_blkgs(), or the system can accumulate a
> huge number of blkgs in pathological cases. We can add a need_resched()
> check on each loop and release locks and do cond_resched() if true
> to avoid this issue, since the blkcg_destroy_blkgs() is not called
> from atomic contexts.
> 
> [ 4757.010308] watchdog: BUG: soft lockup - CPU#11 stuck for 94s!
> [ 4757.010698] Call trace:
> [ 4757.010700]  blkcg_destroy_blkgs+0x68/0x150
> [ 4757.010701]  cgwb_release_workfn+0x104/0x158
> [ 4757.010702]  process_one_work+0x1bc/0x3f0
> [ 4757.010704]  worker_thread+0x164/0x468
> [ 4757.010705]  kthread+0x108/0x138

Kind of ugly with the two clauses for dropping the blkcg lock, one
being a cpu_relax() and the other a resched. How about something
like this:


diff --git a/block/blk-cgroup.c b/block/blk-cgroup.c
index 031114d454a6..4221a1539391 100644
--- a/block/blk-cgroup.c
+++ b/block/blk-cgroup.c
@@ -1016,6 +1016,8 @@ static void blkcg_css_offline(struct cgroup_subsys_state *css)
  */
 void blkcg_destroy_blkgs(struct blkcg *blkcg)
 {
+	might_sleep();
+
 	spin_lock_irq(&blkcg->lock);
 
 	while (!hlist_empty(&blkcg->blkg_list)) {
@@ -1023,14 +1025,20 @@ void blkcg_destroy_blkgs(struct blkcg *blkcg)
 						struct blkcg_gq, blkcg_node);
 		struct request_queue *q = blkg->q;
 
-		if (spin_trylock(&q->queue_lock)) {
-			blkg_destroy(blkg);
-			spin_unlock(&q->queue_lock);
-		} else {
+		if (need_resched() || !spin_trylock(&q->queue_lock)) {
+			/*
+			 * Given that the system can accumulate a huge number
+			 * of blkgs in pathological cases, check to see if we
+			 * need to rescheduling to avoid softlockup.
+			 */
 			spin_unlock_irq(&blkcg->lock);
-			cpu_relax();
+			cond_resched();
 			spin_lock_irq(&blkcg->lock);
+			continue;
 		}
+
+		blkg_destroy(blkg);
+		spin_unlock(&q->queue_lock);
 	}
 
 	spin_unlock_irq(&blkcg->lock);

-- 
Jens Axboe

Re: [PATCH v2] blk-cgroup: Use cond_resched() when destroy blkgs

[Baolin Wang]

在 2021/1/28 11:41, Jens Axboe 写道:
> On 1/27/21 8:22 PM, Baolin Wang wrote:
>> On !PREEMPT kernel, we can get below softlockup when doing stress
>> testing with creating and destroying block cgroup repeatly. The
>> reason is it may take a long time to acquire the queue's lock in
>> the loop of blkcg_destroy_blkgs(), or the system can accumulate a
>> huge number of blkgs in pathological cases. We can add a need_resched()
>> check on each loop and release locks and do cond_resched() if true
>> to avoid this issue, since the blkcg_destroy_blkgs() is not called
>> from atomic contexts.
>>
>> [ 4757.010308] watchdog: BUG: soft lockup - CPU#11 stuck for 94s!
>> [ 4757.010698] Call trace:
>> [ 4757.010700]  blkcg_destroy_blkgs+0x68/0x150
>> [ 4757.010701]  cgwb_release_workfn+0x104/0x158
>> [ 4757.010702]  process_one_work+0x1bc/0x3f0
>> [ 4757.010704]  worker_thread+0x164/0x468
>> [ 4757.010705]  kthread+0x108/0x138
> 
> Kind of ugly with the two clauses for dropping the blkcg lock, one
> being a cpu_relax() and the other a resched. How about something
> like this:
> 
> 
> diff --git a/block/blk-cgroup.c b/block/blk-cgroup.c
> index 031114d454a6..4221a1539391 100644
> --- a/block/blk-cgroup.c
> +++ b/block/blk-cgroup.c
> @@ -1016,6 +1016,8 @@ static void blkcg_css_offline(struct cgroup_subsys_state *css)
>    */
>   void blkcg_destroy_blkgs(struct blkcg *blkcg)
>   {
> +	might_sleep();
> +
>   	spin_lock_irq(&blkcg->lock);
>   
>   	while (!hlist_empty(&blkcg->blkg_list)) {
> @@ -1023,14 +1025,20 @@ void blkcg_destroy_blkgs(struct blkcg *blkcg)
>   						struct blkcg_gq, blkcg_node);
>   		struct request_queue *q = blkg->q;
>   
> -		if (spin_trylock(&q->queue_lock)) {
> -			blkg_destroy(blkg);
> -			spin_unlock(&q->queue_lock);
> -		} else {
> +		if (need_resched() || !spin_trylock(&q->queue_lock)) {
> +			/*
> +			 * Given that the system can accumulate a huge number
> +			 * of blkgs in pathological cases, check to see if we
> +			 * need to rescheduling to avoid softlockup.
> +			 */
>   			spin_unlock_irq(&blkcg->lock);
> -			cpu_relax();
> +			cond_resched();
>   			spin_lock_irq(&blkcg->lock);
> +			continue;
>   		}
> +
> +		blkg_destroy(blkg);
> +		spin_unlock(&q->queue_lock);
>   	}
>   
>   	spin_unlock_irq(&blkcg->lock);
> 

Looks better to me. Do I need resend with your suggestion? Thanks.

Re: [PATCH v2] blk-cgroup: Use cond_resched() when destroy blkgs

[Jens Axboe]

On 1/27/21 8:49 PM, Baolin Wang wrote:
> 
> 
> 在 2021/1/28 11:41, Jens Axboe 写道:
>> On 1/27/21 8:22 PM, Baolin Wang wrote:
>>> On !PREEMPT kernel, we can get below softlockup when doing stress
>>> testing with creating and destroying block cgroup repeatly. The
>>> reason is it may take a long time to acquire the queue's lock in
>>> the loop of blkcg_destroy_blkgs(), or the system can accumulate a
>>> huge number of blkgs in pathological cases. We can add a need_resched()
>>> check on each loop and release locks and do cond_resched() if true
>>> to avoid this issue, since the blkcg_destroy_blkgs() is not called
>>> from atomic contexts.
>>>
>>> [ 4757.010308] watchdog: BUG: soft lockup - CPU#11 stuck for 94s!
>>> [ 4757.010698] Call trace:
>>> [ 4757.010700]  blkcg_destroy_blkgs+0x68/0x150
>>> [ 4757.010701]  cgwb_release_workfn+0x104/0x158
>>> [ 4757.010702]  process_one_work+0x1bc/0x3f0
>>> [ 4757.010704]  worker_thread+0x164/0x468
>>> [ 4757.010705]  kthread+0x108/0x138
>>
>> Kind of ugly with the two clauses for dropping the blkcg lock, one
>> being a cpu_relax() and the other a resched. How about something
>> like this:
>>
>>
>> diff --git a/block/blk-cgroup.c b/block/blk-cgroup.c
>> index 031114d454a6..4221a1539391 100644
>> --- a/block/blk-cgroup.c
>> +++ b/block/blk-cgroup.c
>> @@ -1016,6 +1016,8 @@ static void blkcg_css_offline(struct cgroup_subsys_state *css)
>>    */
>>   void blkcg_destroy_blkgs(struct blkcg *blkcg)
>>   {
>> +	might_sleep();
>> +
>>   	spin_lock_irq(&blkcg->lock);
>>   
>>   	while (!hlist_empty(&blkcg->blkg_list)) {
>> @@ -1023,14 +1025,20 @@ void blkcg_destroy_blkgs(struct blkcg *blkcg)
>>   						struct blkcg_gq, blkcg_node);
>>   		struct request_queue *q = blkg->q;
>>   
>> -		if (spin_trylock(&q->queue_lock)) {
>> -			blkg_destroy(blkg);
>> -			spin_unlock(&q->queue_lock);
>> -		} else {
>> +		if (need_resched() || !spin_trylock(&q->queue_lock)) {
>> +			/*
>> +			 * Given that the system can accumulate a huge number
>> +			 * of blkgs in pathological cases, check to see if we
>> +			 * need to rescheduling to avoid softlockup.
>> +			 */
>>   			spin_unlock_irq(&blkcg->lock);
>> -			cpu_relax();
>> +			cond_resched();
>>   			spin_lock_irq(&blkcg->lock);
>> +			continue;
>>   		}
>> +
>> +		blkg_destroy(blkg);
>> +		spin_unlock(&q->queue_lock);
>>   	}
>>   
>>   	spin_unlock_irq(&blkcg->lock);
>>
> 
> Looks better to me. Do I need resend with your suggestion? Thanks.

Probably best, gives Tejun another chance to sign off on it :-)


-- 
Jens Axboe