Re: [RFC 1/6] Optimize wake-up task for Task Packing heuristic

[Parth Shah <parth015@linux.vnet.ibm.com>]

On 3/27/19 3:01 PM, Dietmar Eggemann wrote:
> Hi Parth,
> 
> On 3/22/19 7:06 AM, Parth Shah wrote:
>> TurboSched feature requires the minimal number of cores to be active
>> inorder to sustain higher Turbo-frequency in SMP sytems. The jitter
>> tasks if packed on the idle CPUs of already active cores will result in
>> better performance for throughput intensive workloads.
>>
>> Signed-off-by: Parth Shah <parth015@linux.vnet.ibm.com>
>> ---
>>   kernel/sched/fair.c | 91 ++++++++++++++++++++++++++++++++++++++++++++-
>>   1 file changed, 89 insertions(+), 2 deletions(-)
>>
>> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
>> index 51003e1c794d..dcf48f37e0fa 100644
>> --- a/kernel/sched/fair.c
>> +++ b/kernel/sched/fair.c
>> @@ -6202,6 +6202,82 @@ static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, int t
>>       return cpu;
>>   }
>>   +/*
>> + * Core is defined as under-utilized in case if the aggregated utilization of a
>> + * all the CPUs in a core is less than 12.5%
>> + */
>> +static inline bool core_underutilized(long unsigned core_util,
>> +           long unsigned core_capacity)
>> +{
>> +    return core_util < core_capacity>>3;
>> +}
>> +
>> +/*
>> + * Core Capacity Mulitplication Factor
>> + * The capacity of a core is defined to be 1.6x the capacity of any
>> + * CPU(or SM thread), since the architecture is symmetric
>> + */
>> +static const int core_cap_mf = 16;
>> +
>> +/*
>> + * Try to find non idle core in the system, but with spare capacity available
>> + * for task packing, thereby keeping minimal cores active.
>> + */
>> +static int select_non_idle_core(struct task_struct *p, int prev_cpu)
>> +{
>> +    struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
>> +    int core, smt;
>> +
>> +    cpumask_and(cpus, cpu_online_mask, &p->cpus_allowed);
>> +
>> +    for_each_cpu_wrap(core, cpus, prev_cpu)
>> +    {
>> +        long unsigned int core_util = 0;
>> +        long unsigned int core_cap = core_cap_mf*capacity_of(core)/10;
>> +        long unsigned int cache_cpu_util = (unsigned)-1;
>> +        long unsigned est_util = 0, est_util_enqueued = 0;
>> +        int cache_cpu = core;
>> +        struct cfs_rq *cfs_rq;
>> +
>> +        for_each_cpu(smt, cpu_smt_mask(core)) {
> 
> This one doesn't build for me on arm64 (make defconfig) since it uses cpu_smt_mask() outside the CONFIG_SCHED_SMT guard.
> 
> kernel/sched/fair.c: In function ‘select_non_idle_core’:
> kernel/sched/fair.c:6243:21: error: implicit declaration of function ‘cpu_smt_mask’; did you mean ‘cpu_cpu_mask’? [-Werror=implicit-function-declaration]
>    for_each_cpu(smt, cpu_smt_mask(core)) {
>                      ^
> ./include/linux/cpumask.h:242:32: note: in definition of macro ‘for_each_cpu’
> 
> [...]
> 

Thanks for pointing out. It will not build for individual patches for the current version of RFC.

Please try to build it with full patch set. I assure, the following iterations of RFC will resolve this issue.