[PATCH] drm/amdkfd: Try to schedule bottom half on same core

[Felix Kuehling]

Am 2022-08-10 um 19:41 schrieb Felix Kuehling:
> On systems that support SMT (hyperthreading) schedule the bottom half of
> the KFD interrupt handler on the same core. This makes it possible to
> reserve a core for interrupt handling and have the bottom half run on
> that same core.
>
> On systems without SMT, pick another core in the same NUMA node, as
> before.
>
> Use for_each_cpu_wrap instead of open-coding it.
>
> Signed-off-by: Felix Kuehling <Felix.Kuehling at amd.com>
> ---
>   drivers/gpu/drm/amd/amdkfd/kfd_device.c | 20 ++++++++++++++++----
>   1 file changed, 16 insertions(+), 4 deletions(-)
>
> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device.c b/drivers/gpu/drm/amd/amdkfd/kfd_device.c
> index f5853835f03a..5b12ae69a301 100644
> --- a/drivers/gpu/drm/amd/amdkfd/kfd_device.c
> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device.c
> @@ -24,6 +24,7 @@
>   #include <linux/bsearch.h>
>   #include <linux/pci.h>
>   #include <linux/slab.h>
> +#include <linux/topology.h>
>   #include "kfd_priv.h"
>   #include "kfd_device_queue_manager.h"
>   #include "kfd_pm4_headers_vi.h"
> @@ -801,13 +802,24 @@ static inline void kfd_queue_work(struct workqueue_struct *wq,
>   				  struct work_struct *work)
>   {
>   	int cpu, new_cpu;
> +	const struct cpumask *mask = NULL;
>   
>   	cpu = new_cpu = smp_processor_id();
> -	do {
> -		new_cpu = cpumask_next(new_cpu, cpu_online_mask) % nr_cpu_ids;
> -		if (cpu_to_node(new_cpu) == numa_node_id())
> +
> +#if defined(CONFIG_SCHED_SMT)
> +	/* CPU threads in the same core */
> +	mask = cpu_smt_mask(cpu);
> +#endif
> +	if (!mask || cpumask_weight(mask) <= 1) {

There is a stray { here, left over from when I removed some 
instrumentation and switched branches. I'll fix that before submitting.

Regards,
   Felix


> +		/* CPU threads in the same NUMA node */
> +		mask = cpu_cpu_mask(cpu);
> +	/* Pick the next online CPU thread in the same core or NUMA node */
> +	for_each_cpu_wrap(cpu, mask, cpu+1) {
> +		if (cpu != new_cpu && cpu_online(cpu)) {
> +			new_cpu = cpu;
>   			break;
> -	} while (cpu != new_cpu);
> +		}
> +	}
>   
>   	queue_work_on(new_cpu, wq, work);
>   }