workqueue: Use active mask for new worker when pool is DISASSOCIATED
PATCH v3 1/1 workqueue: Use active mask for new worker when pool is DISASSOCIATED
When CPU-[un]hotplugs, all workers will be bound to active CPU via unbind_workers().
But the unbound worker still has a chance to create a new worker, which has bound the newly created task to pool->attrs->cpumask. But the CPU has been unplugged.
Please refer to the following scenarios.
CPU0 CPU1
------------------------------------------------------------------sched_cpu_deactivate(cpu_active_mask clear) workqueue_offline_cpu(work pool POOL_DISASSOCIATED)
-- all worker will migrate to another cpu --
worker_thread
-- will create new worker if
pool->worklist is not empty create_worker()
-- new kworker will bound to CPU0
(pool->attrs->cpumask will be mask of CPU0).
kworker/0:x will running on rq
sched_cpu_dying
if (rq->nr_running != 1 || rq_has_pinned_tasks(rq))
WARN(true, "Dying CPU not properly vacated!");
---------OOPS-------------And the crash log as fellowing.
The crash log is as follows:
[ 1622.829074] ------------[ cut here ]------------ [ 1622.829081] Dying CPU not properly vacated! [ 1622.829091] WARNING: CPU: 3 PID: 31 at kernel/sched/core.c:7756 sched_cpu_dying+0x74/0x204 [ 1622.829374] CPU: 3 PID: 31 Comm: migration/3 Tainted: P O 5.10.59-rt52 #2 [ 1622.829386] Stopper: multi_cpu_stop+0x0/0x160 <- 0x0 [ 1622.829400] pstate: 60c00009 (nZCv daif +PAN +UAO -TCO BTYPE=--) [ 1622.829408] pc : sched_cpu_dying+0x74/0x204 [ 1622.829415] lr : sched_cpu_dying+0x74/0x204 [ 1622.829421] sp : ffff800012933c70 [ 1622.829424] pmr_save: 00000060 [ 1622.829426] x29: ffff800012933c70 x28: 0000000000000000 [ 1622.829435] x27: 0000000000000000 x26: 0000000000000001 [ 1622.829444] x25: 0000000000000000 x24: ffff800018353c2c [ 1622.829452] x23: 0000000000000003 x22: 0000000000000003 [ 1622.829460] x21: 0000000000000059 x20: 0000000000000000 [ 1622.829468] x19: ffff00027ee68a80 x18: 0000000000000000 [ 1622.829477] x17: 0000000000000000 x16: 0000000000000000 [ 1622.829485] x15: ffffffffffffffff x14: ffff80001169ae30 [ 1622.829493] x13: ffffffffffc38c07 x12: ffffffffffffffff [ 1622.829501] x11: ffffffffffe00000 x10: ffff80001169ae58 [ 1622.829510] x9 : 000000000000001e x8 : ffff80001169ae30 [ 1622.829518] x7 : ffff800012933ab0 x6 : 00000000ffff0e20 [ 1622.829526] x5 : ffff00027ee62988 x4 : 00000000ffff0e20 [ 1622.829535] x3 : ffff800011e30180 x2 : 0000000100000002 [ 1622.829543] x1 : 0000000000000000 x0 : 0000000000000000 [ 1622.829552] Call trace: [ 1622.829555] sched_cpu_dying+0x74/0x204 [ 1622.829562] cpuhp_invoke_callback+0xc0/0x1b0 [ 1622.829571] take_cpu_down+0xbc/0xd4 [ 1622.829577] multi_cpu_stop+0x138/0x160 [ 1622.829584] cpu_stopper_thread+0x9c/0x118 [ 1622.829591] smpboot_thread_fn+0x1e8/0x1ec [ 1622.829598] kthread+0x114/0x124 [ 1622.829606] ret_from_fork+0x10/0x30 [ 1622.829615] ---[ end trace 0000000000000002 ]--- [ 1623.830273] CPU3 enqueued tasks (2 total): [ 1623.830291] pid: 31, name: migration/3 [ 1623.830440] pid: 25654, name: kworker/3:0 [ 1623.830444] task:kworker/3:0 state:R running task stack: 0 pid:25654 ppid: 2 flags:0x00000028 [ 1623.830458] Call trace: [ 1623.830460] __switch_to+0x164/0x17c [ 1623.830472] __schedule+0x4cc/0x5c0 [ 1623.830483] schedule+0x7c/0xcc [ 1623.830491] schedule_preempt_disabled+0x14/0x24 [ 1623.830500] kthread+0xd8/0x124 [ 1623.830509] ret_from_fork+0x10/0x30
In the crash log, the error enqueued kworker(pid: 25654, name: kworker/3:0) is a newly created thread, and have bind to a unpluged CPU 3.
Can crash dump can also verify this.
crash> task -R nr_cpus_allowed,cpus_mask 25654
PID: 25654 TASK: ffff000181ff0000 CPU: 3 COMMAND: "kworker/3:0"
nr_cpus_allowed = 1,
cpus_mask = {
bits = {8, 0}},
crash> struct worker_pool.cpu,nr_workers,attrs 0xffff00027ee68380
cpu = 3,
nr_workers = 3,
attrs = 0xffff000180004480,
crash> struct workqueue_attrs 0xffff000180004480
struct workqueue_attrs {
nice = 0,
cpumask = {{
bits = {8, 0}
}}, no_numa = false
}
The stack trace of the bad running task was dumped via the following patch:
And I think this debug patch needs to be added to the mainline, it can help us to debug this kind of problem
To fix it, we can use cpu_active_mask when work pool is DISASSOCIATED.
--
Changelog:
v1 -> v2:
- Move worker task bind to worker_attach_to_pool, remove extra wq_pool_attach_mutex added.
- Add a timing diagram to make this question clearer.
v2 -> v3:
- Add missing PF_NO_SETAFFINITY, use cpumask_intersects to avoid setting bad mask for unbound work pool as Lai Jiangshan advised.
- Call kthread_set_pre_cpu correctly for unbound worker.
--- kernel/workqueue.c | 16 +++++++++++----- 1 file changed, 11 insertions(+), 5 deletions(-)
kernel/workqueue.c +11 -5
@@ -1860,8 +1860,16 @@ static struct worker *alloc_worker(int node)static void worker_attach_to_pool (struct worker *worker,struct worker_pool *pool) { +const struct cpumask *cpu_mask; +mutex_lock (&wq_pool_attach_mutex); +if (cpumask_intersects (pool->attrs->cpumask, cpu_active_mask)) + cpu_mask = pool->attrs->cpumask; +else + cpu_mask = wq_unbound_cpumask; + +set_cpus_allowed_ptr (worker->task, cpu_mask);/* * The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains * stable across this function. See the comments above the flag
@@ -1870,10 +1878,8 @@ static void worker_attach_to_pool(struct worker *worker,if (pool->flags &POOL_DISASSOCIATED ) worker->flags |=WORKER_UNBOUND ;else -kthread_set_per_cpu (worker->task, pool->cpu); - -if (worker->rescue_wq) -set_cpus_allowed_ptr (worker->task, pool->attrs->cpumask); +kthread_set_per_cpu (worker->task, + cpu_mask == wq_unbound_cpumask ?-1 : pool->cpu);list_add_tail (&worker->node, &pool->workers); worker->pool = pool;
@@ -1952,8 +1958,8 @@ static struct worker *create_worker(struct worker_pool *pool) goto fail;set_user_nice (worker->task, pool->attrs->nice); -kthread_bind_mask (worker->task, pool->attrs->cpumask); + worker->task->flags |=PF_NO_SETAFFINITY ;/* successful, attach the worker to the pool */ worker_attach_to_pool (worker, pool);
Signature · 1 line
2.29.0
Lai Jiangshan
reviewerQuoted 10 lines · Schspa Shi
When CPU-[un]hotplugs, all workers will be bound to active CPU via
unbind_workers().But the unbound worker still has a chance to create a new worker, which has bound the newly created task to
pool->attrs->cpumask. But the CPU has been unplugged.Please refer to the following scenarios.
Quoted 10 lines
CPU0 CPU1
------------------------------------------------------------------sched_cpu_deactivate(cpu_active_mask clear)workqueue_offline_cpu(work pool POOL_DISASSOCIATED)
-- all worker will migrate to another cpu --
worker_thread
-- will create new worker if
pool->worklistis not emptycreate_worker()
-- new kworker will bound to CPU0
How will the new kworker bound to CPU0? Could you give more details?
Since the pool is POOL_DISASSOCIATED and kthread_is_per_cpu() will be false for the new worker. ttwu() will put it on a fallback CPU IIUC (see select_task_rq()).
Quoted 10 lines
(
pool->attrs->cpumaskwill be mask of CPU0).
kworker/0:x will running on rq
sched_cpu_dying
if (rq->nr_running!= 1 ||rq_has_pinned_tasks(rq))WARN(true, "Dying CPU not properly vacated!"); ---------OOPS-------------
Quoted 6 lines
The stack trace of the bad running task was dumped via the following patch:
And I think this debug patch needs to be added to the mainline, it can help us to debug this kind of problem
To fix it, we can use
cpu_active_maskwhen work pool is DISASSOCIATED.
use wq_unbound_cpumask.
Quoted 2 lines
Signed-off-by Schspa Shi <[email protected]>
Please solo CC Peter, as:
Quoted 39 lines
--
Changelog:
v1 -> v2:
- Move worker task bind toworker_attach_to_pool, remove extrawq_pool_attach_mutexadded.
- Add a timing diagram to make this question clearer.
v2 -> v3:
- Add missingPF_NO_SETAFFINITY, usecpumask_intersectsto avoid setting bad mask for unbound work pool as Lai Jiangshan advised.
- Callkthread_set_pre_cpucorrectly for unbound worker.--- kernel/workqueue.c | 16 +++++++++++----- 1 file changed, 11 insertions(+), 5 deletions(-)kernel/workqueue.c +8
-0@@ -1860,8 +1860,16 @@ static struct worker *alloc_worker(int node)static void worker_attach_to_pool (struct worker *worker,struct worker_pool *pool) { +const struct cpumask *cpu_mask; +mutex_lock (&wq_pool_attach_mutex); +if (cpumask_intersects (pool->attrs->cpumask, cpu_active_mask)) + cpu_mask = pool->attrs->cpumask; +else + cpu_mask = wq_unbound_cpumask; + +set_cpus_allowed_ptr (worker->task, cpu_mask);/* * The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains * stable across this function. See the comments above the flag
kernel/workqueue.c static void worker_attach_to_pool(struct worker *worker, · line 1882
Quoted 9 lines
kernel/workqueue.c +2
-4@@ -1870,10 +1878,8 @@ static void worker_attach_to_pool(struct worker *worker,if (pool->flags &POOL_DISASSOCIATED ) worker->flags |=WORKER_UNBOUND ;else -kthread_set_per_cpu (worker->task, pool->cpu); - -if (worker->rescue_wq) -set_cpus_allowed_ptr (worker->task, pool->attrs->cpumask); +kthread_set_per_cpu (worker->task, + cpu_mask == wq_unbound_cpumask ?-1 : pool->cpu);
Only workers for percpu pool need to set kthread_set_per_cpu(). So it is already handled in the above code, the branch is unneeded.
Quoted 16 lines
kernel/workqueue.c +1
-2list_add_tail (&worker->node, &pool->workers); worker->pool = pool;@@ -1952,8 +1958,8 @@ static struct worker *create_worker(struct worker_pool *pool) goto fail;set_user_nice (worker->task, pool->attrs->nice); -kthread_bind_mask (worker->task, pool->attrs->cpumask); + worker->task->flags |=PF_NO_SETAFFINITY ;/* successful, attach the worker to the pool */ worker_attach_to_pool (worker, pool); --2.29.0
Quoted 22 lines · Lai Jiangshan
Quoted 10 lines · Schspa Shi
When CPU-[un]hotplugs, all workers will be bound to active CPU via
unbind_workers().But the unbound worker still has a chance to create a new worker, which has bound the newly created task to
pool->attrs->cpumask. But the CPU has been unplugged.Please refer to the following scenarios.
Quoted 10 lines
CPU0 CPU1
------------------------------------------------------------------sched_cpu_deactivate(cpu_active_mask clear)workqueue_offline_cpu(work pool POOL_DISASSOCIATED)
-- all worker will migrate to another cpu --
worker_thread
-- will create new worker if
pool->worklistis not emptycreate_worker()
-- new kworker will bound to CPU0
Quoted 2 lines
How will the new kworker bound to CPU0? Could you give more details?
It's because we use pool->attrs->cpumask for the newly created worker to bind it to CPU0.
Please note that the CPU0 marked here does not correspond to the fault log below, maybe I should change this CPU0 to CPU3, then it can be correspond to the exception log below.
Quoted 4 lines
Since the pool is
POOL_DISASSOCIATEDandkthread_is_per_cpu()will be false for the new worker.ttwu()will put it on a fallback CPU IIUC (seeselect_task_rq()).
It won't put it on a fallback CPU, is_cpu_allowed() will use cpu_online_mask to check if a kthread can running on this CPU, we need cpu_active_mask in this case.
see:
static inline bool is_cpu_allowed(struct task_struct *p, int cpu)
{
...
/* KTHREAD_IS_PER_CPU is always allowed. */
if (kthread_is_per_cpu(p))
return cpu_online(cpu);
...
}Quoted 10 lines
Quoted 8 lines
(
pool->attrs->cpumaskwill be mask of CPU0).
kworker/0:x will running on rq
sched_cpu_dying
if (rq->nr_running!= 1 ||rq_has_pinned_tasks(rq))WARN(true, "Dying CPU not properly vacated!"); ---------OOPS-------------
Quoted 9 lines
Quoted 6 lines
The stack trace of the bad running task was dumped via the following patch:
And I think this debug patch needs to be added to the mainline, it can help us to debug this kind of problem
To fix it, we can use
cpu_active_maskwhen work pool is DISASSOCIATED.use
wq_unbound_cpumask.
Yes, I forgot to change this.
Quoted 7 lines
Quoted 2 lines
Signed-off-by Schspa Shi <[email protected]>Please solo CC Peter, as:
CC Peter Zijlstra <[email protected]>
OK, thanks for reminding me.
Quoted 43 lines
Quoted 43 lines
--
Changelog:
v1 -> v2:
- Move worker task bind toworker_attach_to_pool, remove extrawq_pool_attach_mutexadded.
- Add a timing diagram to make this question clearer.
v2 -> v3:
- Add missingPF_NO_SETAFFINITY, usecpumask_intersectsto avoid setting bad mask for unbound work pool as Lai Jiangshan advised.
- Callkthread_set_pre_cpucorrectly for unbound worker.--- kernel/workqueue.c | 16 +++++++++++----- 1 file changed, 11 insertions(+), 5 deletions(-)kernel/workqueue.c +8
-1@@ -1860,8 +1860,16 @@ static struct worker *alloc_worker(int node)static void worker_attach_to_pool (struct worker *worker,struct worker_pool *pool) { +const struct cpumask *cpu_mask; +mutex_lock (&wq_pool_attach_mutex); +if (cpumask_intersects (pool->attrs->cpumask, cpu_active_mask)) + cpu_mask = pool->attrs->cpumask; +else + cpu_mask = wq_unbound_cpumask; + +set_cpus_allowed_ptr (worker->task, cpu_mask);/* * The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains * stable across this function. See the comments above the flag @@ -1870,10 +1878,8 @@ static void worker_attach_to_pool(struct worker *worker,if (pool->flags &POOL_DISASSOCIATED ) worker->flags |=WORKER_UNBOUND ;else -kthread_set_per_cpu (worker->task, pool->cpu);
Quoted 9 lines
kernel/workqueue.c static void worker_attach_to_pool(struct worker *worker, · line 1882
Quoted 5 lines
kernel/workqueue.c +2
-3- -if (worker->rescue_wq) -set_cpus_allowed_ptr (worker->task, pool->attrs->cpumask); +kthread_set_per_cpu (worker->task, + cpu_mask == wq_unbound_cpumask ?-1 : pool->cpu);Only workers for percpu pool need to set
kthread_set_per_cpu(). So it is already handled in the above code, the branch is unneeded.
Change this to something like fellowing.
if (pool->flags & POOL_DISASSOCIATED)
worker->flags |= WORKER_UNBOUND;
else if (cpu_mask == pool->attrs->cpumask)
kthread_set_per_cpu(worker->task, pool->cpu);I add this because we will set it = -1 on unbind_workers too. static void unbind_workers(int cpu)
{
...
for_each_pool_worker(worker, pool) {
kthread_set_per_cpu(worker->task, -1);
WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, wq_unbound_cpumask) < 0);
}
...
}Quoted 16 lines
Quoted 16 lines
kernel/workqueue.c +1
-2list_add_tail (&worker->node, &pool->workers); worker->pool = pool;@@ -1952,8 +1958,8 @@ static struct worker *create_worker(struct worker_pool *pool) goto fail;set_user_nice (worker->task, pool->attrs->nice); -kthread_bind_mask (worker->task, pool->attrs->cpumask); + worker->task->flags |=PF_NO_SETAFFINITY ;/* successful, attach the worker to the pool */ worker_attach_to_pool (worker, pool); --2.29.0
Signature · 2 lines
BRs Zhaohui Shi