Re: [PATCH v2] sched: Store restrict_cpus_allowed_ptr() call state

[Waiman Long <longman@redhat.com>]

On 1/24/23 14:48, Will Deacon wrote:
> Hi Waiman,
>
> [+Thorsten given where we are in the release cycle]
>
> On Fri, Jan 20, 2023 at 09:17:49PM -0500, Waiman Long wrote:
>> The user_cpus_ptr field was originally added by commit b90ca8badbd1
>> ("sched: Introduce task_struct::user_cpus_ptr to track requested
>> affinity"). It was used only by arm64 arch due to possible asymmetric
>> CPU setup.
>>
>> Since commit 8f9ea86fdf99 ("sched: Always preserve the user requested
>> cpumask"), task_struct::user_cpus_ptr is repurposed to store user
>> requested cpu affinity specified in the sched_setaffinity().
>>
>> This results in a performance regression in an arm64 system when booted
>> with "allow_mismatched_32bit_el0" on the command-line. The arch code will
>> (amongst other things) calls force_compatible_cpus_allowed_ptr() and
>> relax_compatible_cpus_allowed_ptr() when exec()'ing a 32-bit or a 64-bit
>> task respectively. Now a call to relax_compatible_cpus_allowed_ptr()
>> will always result in a __sched_setaffinity() call whether there is a
>> previous force_compatible_cpus_allowed_ptr() call or not.
> I'd argue it's more than just a performance regression -- the affinity
> masks are set incorrectly, which is a user visible thing
> (i.e. sched_getaffinity() gives unexpected values).
>
>> In order to fix this regression, a new scheduler flag
>> task_struct::cpus_allowed_restricted is now added to track if
>> force_compatible_cpus_allowed_ptr() has been called before or not. This
>> patch also updates the comments in force_compatible_cpus_allowed_ptr()
>> and relax_compatible_cpus_allowed_ptr() and handles their interaction
>> with sched_setaffinity().
>>
>> Fixes: 8f9ea86fdf99 ("sched: Always preserve the user requested cpumask")
>> Reported-by: Will Deacon <will@kernel.org>
>> Signed-off-by: Waiman Long <longman@redhat.com>
>> ---
>>   include/linux/sched.h |  3 +++
>>   kernel/sched/core.c   | 46 ++++++++++++++++++++++++++++++++++---------
>>   kernel/sched/sched.h  |  2 ++
>>   3 files changed, 42 insertions(+), 9 deletions(-)
> I find this pretty invasive, but I guess it's up to Peter and Ingo.
> It also doesn't the whole problem for me; see below.
>
>> diff --git a/include/linux/sched.h b/include/linux/sched.h
>> index 853d08f7562b..f93f62a1f858 100644
>> --- a/include/linux/sched.h
>> +++ b/include/linux/sched.h
>> @@ -886,6 +886,9 @@ struct task_struct {
>>   	unsigned			sched_contributes_to_load:1;
>>   	unsigned			sched_migrated:1;
>>   
>> +	/* restrict_cpus_allowed_ptr() bit, serialized by scheduler locks */
>> +	unsigned			cpus_allowed_restricted:1;
>> +
>>   	/* Force alignment to the next boundary: */
>>   	unsigned			:0;
>>   
>> diff --git a/kernel/sched/core.c b/kernel/sched/core.c
>> index bb1ee6d7bdde..48234dc9005b 100644
>> --- a/kernel/sched/core.c
>> +++ b/kernel/sched/core.c
>> @@ -2999,15 +2999,40 @@ static int __set_cpus_allowed_ptr(struct task_struct *p,
>>   	struct rq *rq;
>>   
>>   	rq = task_rq_lock(p, &rf);
>> +
>> +	if (ctx->flags & (SCA_CLR_RESTRICT | SCA_SET_RESTRICT)) {
>> +		p->cpus_allowed_restricted = 0;
> I don't think this is ever called on the SCA_SET_RESTRICT path, as
> restrict_cpus_allowed_ptr() calls __set_cpus_allowed_ptr_locked() directly.
> In my testing, we see a failure in the following sequence:
>
>    1. A 64-bit task has an affinity of 0xf
>    2. It exec()s a 32-bit program and is forcefully restricted to the set
>       of 32-bit-capable cores. Let's say that new mask is 0xc
>    3. The 32-bit task now exec()s a 64-bit program again
>
> And now we're still stuck with 0xc after step 3 whereas we should restore
> 0xf.
I am sorry that missed it. You are right. For setting the 
cpus_allowed_restricted bit, it should be done directly in 
restrict_cpus_allowed_ptr().
>> +	} else if (p->cpus_allowed_restricted) {
>> +		/*
>> +		 * If force_compatible_cpus_allowed_ptr() has been called,
>> +		 * we can't extend cpumask to beyond what is in cpus_mask.
>> +		 */
>> +		if (!cpumask_and(rq->scratch_mask, ctx->new_mask,
>> +				 &p->cpus_mask)) {
>> +			task_rq_unlock(rq, p, &rf);
>> +			return -EINVAL;
>> +		}
> Why is this masking actually needed? __sched_setaffinity() already
> takes into account the task_cpu_possible_mask(), which is why I asked you
> before [1] about cases where the saved affinity is not simply a superset
> of the effective affinity.

I kind of overlook the use of task_cpu_possible_mask() in 
__set_cpus_allowed_ptr_locked. So we don't really need that masking. 
That make the patch even simpler then. I will send out a v3.

Cheers,
Longman