[dpdk-dev] [PATCH v9 04/10] eal: implement functions for thread affinity management

[Narcisa Ana Maria Vasile]

On Wed, Jun 09, 2021 at 02:03:57AM +0300, Dmitry Kozlyuk wrote:
> 2021-06-04 16:44 (UTC-0700), Narcisa Ana Maria Vasile:
> [...]
> > diff --git a/lib/eal/windows/rte_thread.c b/lib/eal/windows/rte_thread.c
> > index 6ea1dc2a05..9e74a538c2 100644
> > --- a/lib/eal/windows/rte_thread.c
> > +++ b/lib/eal/windows/rte_thread.c
> > @@ -7,7 +7,8 @@
> >  #include <rte_errno.h>
> >  #include <rte_debug.h>
> >  #include <rte_thread.h>
> > -#include <rte_windows.h>
> > +
> > +#include "eal_windows.h"
> >  
> >  struct eal_tls_key {
> >  	DWORD thread_index;
> > @@ -77,6 +78,130 @@ rte_thread_equal(rte_thread_t t1, rte_thread_t t2)
> >  	return t1.opaque_id == t2.opaque_id;
> >  }
> >  
> > +static int
> > +rte_convert_cpuset_to_affinity(const rte_cpuset_t *cpuset,
> > +			       PGROUP_AFFINITY affinity)
> > +{
> > +	int ret = 0;
> > +	PGROUP_AFFINITY cpu_affinity = NULL;
> > +
> > +	memset(affinity, 0, sizeof(GROUP_AFFINITY));
> > +	affinity->Group = (USHORT)-1;
> > +
> > +	/* Check that all cpus of the set belong to the same processor group and
> > +	 * accumulate thread affinity to be applied.
> > +	 */
> > +	for (unsigned int cpu_idx = 0; cpu_idx < CPU_SETSIZE; cpu_idx++) {
> > +		if (!CPU_ISSET(cpu_idx, cpuset))
> > +			continue;
> > +
> > +		cpu_affinity = eal_get_cpu_affinity(cpu_idx);
> > +
> > +		if (affinity->Group == (USHORT)-1) {
> > +			affinity->Group = cpu_affinity->Group;
> > +		} else if (affinity->Group != cpu_affinity->Group) {
> > +			ret = EINVAL;
> > +			goto cleanup;
> > +		}
> > +
> > +		affinity->Mask |= cpu_affinity->Mask;
> > +	}
> > +
> > +	if (affinity->Mask == 0) {
> > +		ret = EINVAL;
> > +		goto cleanup;
> > +	}
> > +
> > +cleanup:
> > +	return ret;
> > +}
> 
> For v5 I asked a question that possibly got lost among other comments.
> Repeating the question for convenience:
> 
> 	Just to be clear: is it a kernel limitation that a thread can only
> 	run on cores of one processor group, or do we impose it so that API
> 	is atomic (transactional), i.e. because one of multiple
> 	SetThreadGroupAffinity() calls may fail and leave thread partially
> 	affinitized?
The second reason (to ensure full affinitization). I am not aware of a kernel limitation,
but I'll double check with Dmitry as we co-engineered this patch.