Re: [RFC PATCH-cgroup 5/6] cgroup: Skip dying css in cgroup_apply_control_{enable,disable}

[Waiman Long <longman@redhat.com>]

On 06/21/2017 05:42 PM, Tejun Heo wrote:
> Hello,
>
> On Wed, Jun 14, 2017 at 11:05:36AM -0400, Waiman Long wrote:
>> While constantly turning on and off controllers, it is possible to
>> trigger the dying CSS warnings in cgroup_apply_control_enable() and
>> cgroup_apply_control_disable(). The current code, however, proceeds
>> after the warning leading to other secondary warnings and maybe even
>> data corruption, like
>>
>>   cgroup: cgroup_addrm_files: failed to add current, err=-17
>>
>> To avoid the secondary errors, the dying CSS is now ignored or skipped
>> so as not to cause other problem.
>>
>> Signed-off-by: Waiman Long <longman@redhat.com>
>> ---
>>  kernel/cgroup/cgroup.c | 20 +++++++++++++++-----
>>  1 file changed, 15 insertions(+), 5 deletions(-)
>>
>> diff --git a/kernel/cgroup/cgroup.c b/kernel/cgroup/cgroup.c
>> index f0bea32..2a5bd49 100644
>> --- a/kernel/cgroup/cgroup.c
>> +++ b/kernel/cgroup/cgroup.c
>> @@ -2846,12 +2846,24 @@ static int cgroup_apply_control_enable(struct cgroup *cgrp)
>>  		for_each_subsys(ss, ssid) {
>>  			struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
>>  
>> -			WARN_ON_ONCE(css && percpu_ref_is_dying(&css->refcnt));
>> -
>>  			if (!(cgroup_ss_mask(dsct, false) & (1 << ss->id)) ||
>>  			    (dsct->bypass_ss_mask & (1 << ss->id)))
>>  				continue;
>>  
>> +			/*
>> +			 * If the css is dying, we will just skip it after
>> +			 * warning.
>> +			 */
>> +			if (css && (css->flags & CSS_DYING)) {
>> +				char name[NAME_MAX+1];
>> +
>> +				cgroup_name(cgrp, name, NAME_MAX);
>> +				pr_warn("%s: %s css of cgroup %s is dying!\n",
>> +					__func__, ss->name, name);
>> +				WARN_ON_ONCE(1);
>> +				continue;
>> +			}
> Can you trigger this without your patches because this triggering
> means that the code screwed up before it reached this point.  We
> should be fixing that bug rather than masking it up here.
>
> Thanks.
>
I will try to reproduce it without my patch.

I do think that it can happen with existing code because CSS killing is
asynchronous, I think. So the command can complete before the CSS is
actually gone. If the next command to reactivate it happens fast enough,
we can trigger that. When I added more checking to my test script
essentially increasing the latency between successive tests, I couldn't
trigger it anymore.

Cheers,
Longman