Re: [PATCH v3]sched/rt: Stop for_each_process_thread() iterations in tg_has_rt_tasks()

[Kirill Tkhai]

On 24.01.2020 00:56, Phil Auld wrote:
> Hi,
> 
> On Thu, Apr 26, 2018 at 12:54:41PM +0300 Kirill Tkhai wrote:
>> From: Kirill Tkhai <ktkhai@xxxxxxxxxxxxx>
>>
>> tg_rt_schedulable() iterates over all child task groups,
>> while tg_has_rt_tasks() iterates over all linked tasks.
>> In case of systems with big number of tasks, this may
>> take a lot of time.
>>
>> I observed hard LOCKUP on machine with 20000+ processes
>> after write to "cpu.rt_period_us" of cpu cgroup with
>> 39 children. The problem occurred because of tasklist_lock
>> is held for a long time and other processes can't do fork().
>>
>> PID: 1036268  TASK: ffff88766c310000  CPU: 36  COMMAND: "criu"
>>  #0 [ffff887f7f408e48] crash_nmi_callback at ffffffff81050601
>>  #1 [ffff887f7f408e58] nmi_handle at ffffffff816e0cc7
>>  #2 [ffff887f7f408eb0] do_nmi at ffffffff816e0fb0
>>  #3 [ffff887f7f408ef0] end_repeat_nmi at ffffffff816e00b9
>>     [exception RIP: tg_rt_schedulable+463]
>>     RIP: ffffffff810bf49f  RSP: ffff886537ad7d50  RFLAGS: 00000202
>>     RAX: 0000000000000000  RBX: 000000003b9aca00  RCX: ffff883e9cb4b1b0
>>     RDX: ffff887d0be43608  RSI: ffff886537ad7dd8  RDI: ffff8840a6ad0000
>>     RBP: ffff886537ad7d68   R8: ffff887d0be431b0   R9: 00000000000e7ef0
>>     R10: ffff88164fc39400  R11: 0000000000023380  R12: ffffffff81ef8d00
>>     R13: ffffffff810bea40  R14: 0000000000000000  R15: ffff8840a6ad0000
>>     ORIG_RAX: ffffffffffffffff  CS: 0010  SS: 0018
>> --- <NMI exception stack> ---
>>  #4 [ffff886537ad7d50] tg_rt_schedulable at ffffffff810bf49f
>>  #5 [ffff886537ad7d70] walk_tg_tree_from at ffffffff810c6c91
>>  #6 [ffff886537ad7dc0] tg_set_rt_bandwidth at ffffffff810c6dd0
>>  #7 [ffff886537ad7e28] cpu_rt_period_write_uint at ffffffff810c6eea
>>  #8 [ffff886537ad7e38] cgroup_file_write at ffffffff8111cfd3
>>  #9 [ffff886537ad7ec8] vfs_write at ffffffff8121eced
>> #10 [ffff886537ad7f08] sys_write at ffffffff8121faff
>> #11 [ffff886537ad7f50] system_call_fastpath at ffffffff816e8a7d
>>
>> The patch reworks tg_has_rt_tasks() and makes it to iterate over
>> task group process list instead of iteration over all tasks list.
>> This makes the function to scale well, and reduces its execution
>> time.
>>
>> Note, that since tasklist_lock doesn't protect a task against
>> sched_class changing, we don't introduce new races in comparison
>> to that we had before.
>>
>> Signed-off-by: Kirill Tkhai <ktkhai@xxxxxxxxxxxxx>
>> ---
>>  kernel/sched/rt.c |   21 +++++++++++----------
>>  1 file changed, 11 insertions(+), 10 deletions(-)
>>
>> diff --git a/kernel/sched/rt.c b/kernel/sched/rt.c
>> index 7aef6b4e885a..a40535c604b8 100644
>> --- a/kernel/sched/rt.c
>> +++ b/kernel/sched/rt.c
>> @@ -2395,10 +2395,11 @@ const struct sched_class rt_sched_class = {
>>   */
>>  static DEFINE_MUTEX(rt_constraints_mutex);
>>  
>> -/* Must be called with tasklist_lock held */
>>  static inline int tg_has_rt_tasks(struct task_group *tg)
>>  {
>> -	struct task_struct *g, *p;
>> +	struct task_struct *task;
>> +	struct css_task_iter it;
>> +	int ret = 0;
>>  
>>  	/*
>>  	 * Autogroups do not have RT tasks; see autogroup_create().
>> @@ -2406,12 +2407,16 @@ static inline int tg_has_rt_tasks(struct task_group *tg)
>>  	if (task_group_is_autogroup(tg))
>>  		return 0;
>>  
>> -	for_each_process_thread(g, p) {
>> -		if (rt_task(p) && task_group(p) == tg)
>> -			return 1;
>> +	css_task_iter_start(&tg->css, 0, &it);
>> +	while ((task = css_task_iter_next(&it))) {
>> +		if (rt_task(task)) {
>> +			ret = 1;
>> +			break;
>> +		}
>>  	}
>> +	css_task_iter_end(&it);
>>  
>> -	return 0;
>> +	return ret;
>>  }
>>  
>>  struct rt_schedulable_data {
>> @@ -2510,7 +2515,6 @@ static int tg_set_rt_bandwidth(struct task_group *tg,
>>  		return -EINVAL;
>>  
>>  	mutex_lock(&rt_constraints_mutex);
>> -	read_lock(&tasklist_lock);
>>  	err = __rt_schedulable(tg, rt_period, rt_runtime);
>>  	if (err)
>>  		goto unlock;
>> @@ -2528,7 +2532,6 @@ static int tg_set_rt_bandwidth(struct task_group *tg,
>>  	}
>>  	raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
>>  unlock:
>> -	read_unlock(&tasklist_lock);
>>  	mutex_unlock(&rt_constraints_mutex);
>>  
>>  	return err;
>> @@ -2582,9 +2585,7 @@ static int sched_rt_global_constraints(void)
>>  	int ret = 0;
>>  
>>  	mutex_lock(&rt_constraints_mutex);
>> -	read_lock(&tasklist_lock);
>>  	ret = __rt_schedulable(NULL, 0, 0);
>> -	read_unlock(&tasklist_lock);
>>  	mutex_unlock(&rt_constraints_mutex);
>>  
>>  	return ret;
> 
> Sorry to necro this...  
> 
> I have a customer case that has hit the issue here. It looks like this
> fix didn't end up going in. 
> 
> The only way I could reproduce it myself was to add a udelay(2) to 
> the tg_has_rt_tasks loop. With this patch applied, even with the 
> udelay nothing bad happens.
> 
> Kirill, did you have a good way to reproduce it without adding 
> delays? 

I have no a reproducer, it just used to fire on some of our customer nodes.
It depends on many parameters, and main is workload. Also, I think, interrupts
affinity may be involved (whether they bound to small subset of cpus, or they
distributed over all cores). If this is possible theoretically, it occurs
practically with some probability.

We fixed this with out-of-tree patch in our vzkernel tree, so we haven't seen
a reproduction anymore.

> Peter, is there any chance of taking something like this?
> 
> Is there a better fix?