Message ID | 20210210165407.9770-3-dwagner@suse.de |
---|---|
State | New |
Headers | show |
Series | oslat fixes | expand |
On Wed, Feb 10, 2021 at 05:54:07PM +0100, Daniel Wagner wrote: > To assign the threads to the correct CPU we need to use the cpuset > size as upper bound for the loop and not the number of threads. > > Fixes: 85b0763dacd9 ("oslat: Use parse_cpumask() from rt-numa.h") > Reported-by: Peter Xu <peterx@redhat.com> Please instead add: Reported-by: Pradipta Kumar Sahoo <psahoo@redhat.com> Pradipta, would you verify and see whether this patch fixes your issue? Thanks,
On Wed, 10 Feb 2021, Daniel Wagner wrote: > To assign the threads to the correct CPU we need to use the cpuset > size as upper bound for the loop and not the number of threads. > > Fixes: 85b0763dacd9 ("oslat: Use parse_cpumask() from rt-numa.h") > Reported-by: Peter Xu <peterx@redhat.com> > Signed-off-by: Daniel Wagner <dwagner@suse.de> > --- > src/oslat/oslat.c | 7 +++++-- > 1 file changed, 5 insertions(+), 2 deletions(-) > > diff --git a/src/oslat/oslat.c b/src/oslat/oslat.c > index d3f659b218b0..62b82098419a 100644 > --- a/src/oslat/oslat.c > +++ b/src/oslat/oslat.c > @@ -742,9 +742,12 @@ int main(int argc, char *argv[]) > n_cores = numa_bitmask_weight(cpu_set); > > TEST(threads = calloc(1, n_cores * sizeof(threads[0]))); > - for (i = 0; i < n_cores; ++i) > - if (numa_bitmask_isbitset(cpu_set, i) && move_to_core(i) == 0) > + for (i = 0; n_cores && i < cpu_set->size; i++) { > + if (numa_bitmask_isbitset(cpu_set, i) && move_to_core(i) == 0) { > threads[g.n_threads_total++].core_i = i; > + n_cores--; > + } > + } > > if (numa_bitmask_isbitset(cpu_set, 0) && g.rtprio) > printf("WARNING: Running SCHED_FIFO workload on CPU 0 may hang the thread\n"); > -- > 2.30.0 > > This looks okay, but we are waiting for Pradipta's testing to confirm before I push it. Thanks John
On Wed, Feb 10, 2021 at 11:28:16PM -0500, John Kacur wrote: > > > On Wed, 10 Feb 2021, Daniel Wagner wrote: > > > To assign the threads to the correct CPU we need to use the cpuset > > size as upper bound for the loop and not the number of threads. > > > > Fixes: 85b0763dacd9 ("oslat: Use parse_cpumask() from rt-numa.h") > > Reported-by: Peter Xu <peterx@redhat.com> > > Signed-off-by: Daniel Wagner <dwagner@suse.de> > > --- > > src/oslat/oslat.c | 7 +++++-- > > 1 file changed, 5 insertions(+), 2 deletions(-) > > > > diff --git a/src/oslat/oslat.c b/src/oslat/oslat.c > > index d3f659b218b0..62b82098419a 100644 > > --- a/src/oslat/oslat.c > > +++ b/src/oslat/oslat.c > > @@ -742,9 +742,12 @@ int main(int argc, char *argv[]) > > n_cores = numa_bitmask_weight(cpu_set); > > > > TEST(threads = calloc(1, n_cores * sizeof(threads[0]))); > > - for (i = 0; i < n_cores; ++i) > > - if (numa_bitmask_isbitset(cpu_set, i) && move_to_core(i) == 0) > > + for (i = 0; n_cores && i < cpu_set->size; i++) { > > + if (numa_bitmask_isbitset(cpu_set, i) && move_to_core(i) == 0) { > > threads[g.n_threads_total++].core_i = i; > > + n_cores--; > > + } > > + } > > > > if (numa_bitmask_isbitset(cpu_set, 0) && g.rtprio) > > printf("WARNING: Running SCHED_FIFO workload on CPU 0 may hang the thread\n"); > > -- > > 2.30.0 > > > > > > This looks okay, but we are waiting for Pradipta's testing to confirm > before I push it. After the reporter's confirmation there seems to have two issues. Daniel's patch should be a valid fix of at least one of the problem. The other problem is still under investigation and I think it's not affecting the upstream branch. So I think we don't need to wait for that anymore: Reviewed-by: Peter Xu <peterx@redhat.com> Tested-by: Peter Xu <peterx@redhat.com> Thanks, -- Peter Xu
diff --git a/src/oslat/oslat.c b/src/oslat/oslat.c index d3f659b218b0..62b82098419a 100644 --- a/src/oslat/oslat.c +++ b/src/oslat/oslat.c @@ -742,9 +742,12 @@ int main(int argc, char *argv[]) n_cores = numa_bitmask_weight(cpu_set); TEST(threads = calloc(1, n_cores * sizeof(threads[0]))); - for (i = 0; i < n_cores; ++i) - if (numa_bitmask_isbitset(cpu_set, i) && move_to_core(i) == 0) + for (i = 0; n_cores && i < cpu_set->size; i++) { + if (numa_bitmask_isbitset(cpu_set, i) && move_to_core(i) == 0) { threads[g.n_threads_total++].core_i = i; + n_cores--; + } + } if (numa_bitmask_isbitset(cpu_set, 0) && g.rtprio) printf("WARNING: Running SCHED_FIFO workload on CPU 0 may hang the thread\n");
To assign the threads to the correct CPU we need to use the cpuset size as upper bound for the loop and not the number of threads. Fixes: 85b0763dacd9 ("oslat: Use parse_cpumask() from rt-numa.h") Reported-by: Peter Xu <peterx@redhat.com> Signed-off-by: Daniel Wagner <dwagner@suse.de> --- src/oslat/oslat.c | 7 +++++-- 1 file changed, 5 insertions(+), 2 deletions(-)