On Wed, Mar 20, 2024 at 7:33 PM Valentin Schneider <vschneid@xxxxxxxxxx> wrote:
> On 20/03/24 19:17, Valentin Schneider wrote:
> > On 20/03/24 15:03, Daniel Vacek wrote:
> >> Hi Valentin,
> >>
> >> On Mon, Mar 18, 2024 at 6:34 PM Valentin Schneider <vschneid@xxxxxxxxxx> wrote:
> >>> On 18/03/24 12:17, Daniel Vacek wrote:
> >>> > Bill Peters reported CPU hangs while offlining/onlining CPUs on s390.
> >>> >
> >>> > Analyzing the vmcore data shows `stop_one_cpu_nowait()` in `affine_move_task()`
> >>> > can fail when racing with off-/on-lining resulting in a deadlock waiting for
> >>> > the pending migration stop work completion which is never done.
> >>> >
> >>> > Fix this by correctly handling such a condition.
> >>> >
> >>>
> >>> IIUC the problem is that the dest_cpu and its stopper thread can be taken
> >>> down by take_cpu_down(), and affine_move_task() currently isn't aware of
> >>> that. I thought we had tested this vs hotplug, but oh well...
> >>
> >> I'm sorry, I should have provided more context in the first place. The machine
> >> is an LPAR with 2 CPUs and CPU 0 was onlining (hotplugging?) CPU 1. The traces
> >> show this scenario:
> >>
> >> CPU 0 | CPU 1
> >> |
> >> cpuplugd task 1429 |
> >> holds the `cpu_hotplug_lock` |
> >> for writing in _cpu_up+0x16a |
> >> blocked on `cpuhp_state:1.done_up` |
> >> completion in __cpuhp_kick_ap+0x76 |
> >> |
> >> | cpuhp/1 task 17
> >> |supposed to complete bringup of the CPU
> >> | (`cpuhp_state:1.done_up`) in cpuhp_thread_fun+0x108
> >> |blocked on `wq_pool_attach_mutex`
> >> | in workqueue_online_cpu+0x9e
> >> |
> >> xfs-conv/dm-0 task 745 |
> >> holds the `wq_pool_attach_mutex` |
> >> in worker_attach_to_pool+0x66 \
> >> blocked on `task->migration_pending->done`|
> >> completion in affine_move_task+0x10a/
> >>
> >> ~~~
> >> crash> b 1429
> >> PID: 1429 TASK: 99398000 CPU: 0 COMMAND: "cpuplugd"
> >> #0 [997df970] __schedule+0x34c at 3089c424
> >> #1 [997df9e0] schedule+0x7e at 3089cafe
> >> #2 [997dfa20] schedule_timeout+0x26e at 308a1d8e
> >> [inlined] do_wait_for_common
> >> [inlined] __wait_for_common
> >> #3 [997dfad8] wait_for_common+0x14a at 3089d902
> >> [ret call] wait_for_completion+0x1a at 3089d96a
> >>
> >> [inlined] wait_for_ap_thread <<< blocked on `cpuhp_state:1.done_up` completion
> >> [ret call] __cpuhp_kick_ap+0x76 at 300c610e
> >> #4 [997dfb58] cpuhp_kick_ap+0xc4 at 300c61dc
> >> [inlined] bringup_wait_for_ap
> >> [ret call] bringup_cpu+0xea at 300c6402
> >> #5 [997dfba8] cpuhp_invoke_callback+0xcc at 300c4f14
> >> #6 [997dfc40] _cpu_up+0x16a at 300c798a <<< holds the `cpu_hotplug_lock` for writing
> >> #7 [997dfc98] do_cpu_up+0xc6 at 300c7b66
> >> #8 [997dfcd8] cpu_subsys_online+0x58 at 305a0a00
> >> #9 [997dfd28] device_online+0x9e at 30598e7e
> >> #10 [997dfd68] online_store+0x88 at 30598f28
> >> #11 [997dfda8] kernfs_fop_write+0xdc at 3040276c
> >> #12 [997dfdf8] vfs_write+0xa8 at 30354760
> >> #13 [997dfe58] ksys_write+0x62 at 30354a32
> >>
> >> crash> cpuhp_cpu_state.state cpuhp_state:a | paste - -
> >> [0]: 1aef424e0 state = CPUHP_ONLINE, # (195)
> >> [1]: 1aef654e0 state = CPUHP_AP_WORKQUEUE_ONLINE, # (159)
> >>
> >> crash> cpuhp_cpu_state.bringup,thread,done_up.done cpuhp_state:1 -d | paste - - - -
> >> [1]: 1aef654e0 bringup = true, thread = 0x81134400, done_up.done = 0, <<<
> >>
> >> crash> b 17
> >> PID: 17 TASK: 81134400 CPU: 1 COMMAND: "cpuhp/1"
> >> #0 [81143b68] __schedule+0x34c at 3089c424
> >> #1 [81143bd8] schedule+0x7e at 3089cafe
> >> #2 [81143c18] schedule_preempt_disabled+0x2a at 3089cfba
> >> #3 [81143c30] __mutex_lock+0x320 at 3089df60
> >>
> >> #4 [81143cb0] workqueue_online_cpu+0x9e at 300e847e <<< blocked on `wq_pool_attach_mutex`
> >> #5 [81143d20] cpuhp_invoke_callback+0xcc at 300c4f14
> >> #6 [81143db8] cpuhp_thread_fun+0x108 at 300c6848 <<< supposed to complete the bring-up of the CPU (`cpuhp_state:1.done_up`)
> >>
> >> crash> b 745
> >> PID: 745 TASK: 82359100 CPU: 0 COMMAND: "xfs-conv/dm-0"
> >> #0 [8b4bfa20] __schedule+0x34c at 3089c424
> >> #1 [8b4bfa90] schedule+0x7e at 3089cafe
> >> #2 [8b4bfad0] schedule_timeout+0x26e at 308a1d8e
> >> [inlined] do_wait_for_common
> >> [inlined] __wait_for_common
> >> #3 [8b4bfb88] wait_for_common+0x14a at 3089d902
> >> [ret call] wait_for_completion+0x1a at 3089d96a
> >>
> >> #4 [8b4bfc08] affine_move_task+0x10a at 300fb51a <<< blocked on `task->migration_pending->done` completion
> >> #5 [8b4bfd08] __set_cpus_allowed_ptr+0x12e at 300fb926
> >> [ret call] set_cpus_allowed_ptr+0xa at 300fba32
> >> #6 [8b4bfd78] worker_attach_to_pool+0x66 at 300e1dae <<< holds the `wq_pool_attach_mutex`
> >> #7 [8b4bfdc8] rescuer_thread+0x12c at 300e5bac
> >>
> >> crash> rx 8b4bfea0
> >> 8b4bfea0: [863373c0:kmalloc-192]
> >>
> >> crash> worker.task,rescue_wq 863373c0
> >> task = 0x82359100,
> >> rescue_wq = 0x8aa44400,
> >>
> >> crash> list -s pool_workqueue.pool pool_workqueue.mayday_node -hO workqueue_struct.maydays 0x8aa44400 | paste - -
> >> 1fffff7f751900 pool = 0x1aef56a00,
> >>
> >> crash> worker_pool.attrs 0x1aef56a00
> >> attrs = 0x80088180,
> >>
> >> crash> workqueue_attrs.cpumask[0].bits 0x80088180
> >> cpumask[0].bits = {0x1, 0x0, ...
> >>
> >> crash> cpumask.bits __cpu_active_mask
> >> bits = {0x1, 0x0, ...
> >>
> >> crash> cpumask.bits __cpu_online_mask
> >> bits = {0x3, 0x0, ...
> >>
> >> crash> task_struct.migration_pending,flags 0x82359100
> >> migration_pending = 0x8b4bfce8,
> >> flags = 0x4208060,
> >> ^ PF_KTHREAD
> >>
> >> crash> pd distribute_cpu_mask_prev:0
> >> per_cpu(distribute_cpu_mask_prev, 0) = 0
> >>
> >> crash> set_affinity_pending.refs.refs.counter,arg,stop_pending,done.done 0x8b4bfce8 -d
> >> refs.refs.counter = 1
> >> arg = {
> >> task = 0x82359100,
> >> dest_cpu = 0,
> >> pending = 0x8b4bfce8
> >> }
> >> stop_pending = 1,
> >> done.done = 0,
> >> ~~~
> >>
> >> In other words the `set_cpus_allowed_ptr()` is called from a worker thread which
> >> tries to migrate. The worker pool is only allowed on CPU 0 and that was supposed
> >> to be the destination as per the stack structure. In this case I thought it's OK
> >> to leave the task on the old CPU
> >
> > AFAICT if a call to set_cpus_allowed_ptr() ends up in affine_move_task()
> > and down to the stopper call, that means the task isn't allowed on its
> > current CPU and needs to be moved.
> >
> >> and the Bill's testing scenario was successful
> >> with the proposed patch. IIUC, it's exercising the hotplug due to load-balancing.
> >>
> >> This was on RHEL 8.8.z kernel. I see upstream changed a bit so I'm not sure it's
> >> still reproducible. Also, I'm not sure why this only happens on s390 and not on
> >> x86. I imagine the CPU hotplug slightly differs? Anyways this seems to be timing
> >> sensitive and the timing will differ greatly for sure.
> >>
> >
> > Thanks for the extra context!
> >
> > Double checking what I wrote before, I forgot RCU considers preempt-off
> > sections as read-side critical sections. __set_cpus_allowed_ptr() already
> > has preemption disabled all the way from reading the cpu_active_mask to the
> > stop_one_cpu_nowait() call via task_rq_lock() + preempt_disable().
> >
> > IOW we have:
> >
> > __set_cpus_allowed_ptr()
> > task_rq_lock() <-- PREEMPT OFF
> > __set_cpus_allowed_ptr_locked()
> > cpu_valid_mask = cpu_active_mask;
> > dest_cpu = cpumask_any_and_distribute(cpu_valid_mask, ctx->new_mask);
> > affine_move_task()
> > preempt_disable();
> > task_rq_unlock();
> > stop_one_cpu_nowait(); <-- preemption still OFF
> >
> > And, considering:
> >
> > sched_cpu_deactivate()
> > set_cpu_active(cpu, false);
> > synchronize_rcu();
> >
> > Then, if __set_cpus_allowed_ptr() observes a CPU as being in the
> > cpu_active_mask and uses that one as a destination CPU, said CPU cannot
> > reach CPUHP_TEARDOWN_CPU:take_cpu_down() and park the stopper thread
> > because its hotplug machinery will wait on the synchronize_rcu() in
> > CPUHP_AP_ACTIVE:sched_cpu_deactivate().
> >
> > So "in theory", this shouldn't happen upstream.
>
> Eh nevermind, in your stacktrace the relevant task is a rescuer thread
> which is a kthread, so the cpu_valid_mask in use there is
> cpu_online_mask, not cpu_valid_mask... Back to reading code :-)
I just wanted to correct you on this point but you figured it out for
yourself after all.
On the other hand the `new_mask` (which comes from the worker pool allowed CPUs)
~~~
2102 static void worker_attach_to_pool(struct worker *worker,
2103 struct worker_pool *pool)
2104 {
2105 mutex_lock(&wq_pool_attach_mutex);
--
2117 if (worker->rescue_wq)
2118 set_cpus_allowed_ptr(worker->task,
pool_allowed_cpus(pool));
2827 static int rescuer_thread(void *__rescuer)
2828 {
2829 struct worker *rescuer = __rescuer;
2830 struct workqueue_struct *wq = rescuer->rescue_wq;
--
2856 while (!list_empty(&wq->maydays)) {
2857 struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2858 struct pool_workqueue,
mayday_node);
2859 struct worker_pool *pool = pwq->pool;
--
2863 list_del_init(&pwq->mayday_node);
2864
2865 raw_spin_unlock_irq(&wq_mayday_lock);
2866
2867 worker_attach_to_pool(rescuer, pool);
2868
2869 raw_spin_lock_irq(&pool->lock);
~~~
matches the `cpu_valid_mask` and is ANDed to the `cpu_online_mask` so
the final result is the same...
Actually double checking now, I may have pulled the wrong pool as the
right `pwq` is removed from the `maydays` list on line 2863 (the code
is similar in RHEL 8.8) and I have used whatever `pwq` was remaining
on that list, but possibly unrelated.
~~~
crash> rx 8b4bfe30
8b4bfe30: 00000001aef56a00
crash> worker_pool.attrs 1aef56a00
attrs = 0x80088180,
crash> workqueue_attrs.cpumask[0].bits 0x80088180
cpumask[0].bits = {0x1, 0x0, ...
crash> set_affinity_pending.refs.refs.counter,arg,stop_pending,done.done
0x8b4bfce8
refs.refs.counter = 0x1
arg = {
task = 0x82359100,
dest_cpu = 0x0,
pending = 0x8b4bfce8
}
stop_pending = 0x1,
done.done = 0x0,
~~~
Nah, so not unrelated. It's the right one. The rescuer was deadlocked
for almost 3 days at the time the vmcore was collected, so it seems
the `pwq` returned to the `maydays` list in the meantime. And I guess
the pool cpumask could have changed as well. Nevertheless, the local
`affine_move_task::my_pending.arg.dest_cpu` on stack still shows CPU 0
was selected that time.
So far I don't see why upstream would not be prone to the same issue.
But I may be missing something.
--nX
