commit 9dfa4b90867dedf4b1db0523a76c7007cbb9bd40 Author: George Dunlap Commit: George Dunlap xen: credit2: implement true SMT support In fact, right now, we recommend keepeing runqueues arranged per-core, so that it is the inter-runqueue load balancing code that automatically spreads the work in an SMT friendly way. This means that any other runq arrangement one may want to use falls short of SMT scheduling optimizations. This commit implements SMT awareness --similar to the one we have in Credit1-- for any possible runq arrangement. This turned out to be pretty easy to do, as the logic can live entirely in runq_tickle() (although, in order to avoid for_each_cpu loops in that function, we use a new cpumask which indeed needs to be updated in other places). In addition to disentangling SMT awareness from load balancing, this also allows us to support the sched_smt_power_savings parametar in Credit2 as well. Signed-off-by: Dario Faggioli Signed-off-by: George Dunlap Reviewed-by: Anshul Makkar diff --git a/xen/common/sched_credit2.c b/xen/common/sched_credit2.c index b33ba7a..3e1720c 100644 --- a/xen/common/sched_credit2.c +++ b/xen/common/sched_credit2.c @@ -353,8 +353,9 @@ struct csched2_runqueue_data { struct list_head svc; /* List of all vcpus assigned to this runqueue */ unsigned int max_weight; - cpumask_t idle, /* Currently idle */ - tickled; /* Another cpu in the queue is already targeted for this one */ + cpumask_t idle, /* Currently idle pcpus */ + smt_idle, /* Fully idle-and-untickled cores (see below) */ + tickled; /* Have been asked to go through schedule */ int load; /* Instantaneous load: Length of queue + num non-idle threads */ s_time_t load_last_update; /* Last time average was updated */ s_time_t avgload; /* Decaying queue load */ @@ -415,6 +416,79 @@ struct csched2_dom { }; /* + * Hyperthreading (SMT) support. + * + * We use a special per-runq mask (smt_idle) and update it according to the + * following logic: + * - when _all_ the SMT sibling in a core are idle, all their corresponding + * bits are set in the smt_idle mask; + * - when even _just_one_ of the SMT siblings in a core is not idle, all the + * bits correspondings to it and to all its siblings are clear in the + * smt_idle mask. + * + * Once we have such a mask, it is easy to implement a policy that, either: + * - uses fully idle cores first: it is enough to try to schedule the vcpus + * on pcpus from smt_idle mask first. This is what happens if + * sched_smt_power_savings was not set at boot (default), and it maximizes + * true parallelism, and hence performance; + * - uses already busy cores first: it is enough to try to schedule the vcpus + * on pcpus that are idle, but are not in smt_idle. This is what happens if + * sched_smt_power_savings is set at boot, and it allows as more cores as + * possible to stay in low power states, minimizing power consumption. + * + * This logic is entirely implemented in runq_tickle(), and that is enough. + * In fact, in this scheduler, placement of a vcpu on one of the pcpus of a + * runq, _always_ happens by means of tickling: + * - when a vcpu wakes up, it calls csched2_vcpu_wake(), which calls + * runq_tickle(); + * - when a migration is initiated in schedule.c, we call csched2_cpu_pick(), + * csched2_vcpu_migrate() (which calls migrate()) and csched2_vcpu_wake(). + * csched2_cpu_pick() looks for the least loaded runq and return just any + * of its processors. Then, csched2_vcpu_migrate() just moves the vcpu to + * the chosen runq, and it is again runq_tickle(), called by + * csched2_vcpu_wake() that actually decides what pcpu to use within the + * chosen runq; + * - when a migration is initiated in sched_credit2.c, by calling migrate() + * directly, that again temporarily use a random pcpu from the new runq, + * and then calls runq_tickle(), by itself. + */ + +/* + * If all the siblings of cpu (including cpu itself) are both idle and + * untickled, set all their bits in mask. + * + * NB that rqd->smt_idle is different than rqd->idle. rqd->idle + * records pcpus that at are merely idle (i.e., at the moment do not + * have a vcpu running on them). But you have to manually filter out + * which pcpus have been tickled in order to find cores that are not + * going to be busy soon. Filtering out tickled cpus pairwise is a + * lot of extra pain; so for rqd->smt_idle, we explicitly make so that + * the bits of a pcpu are set only if all the threads on its core are + * both idle *and* untickled. + * + * This means changing the mask when either rqd->idle or rqd->tickled + * changes. + */ +static inline +void smt_idle_mask_set(unsigned int cpu, const cpumask_t *idlers, + cpumask_t *mask) +{ + const cpumask_t *cpu_siblings = per_cpu(cpu_sibling_mask, cpu); + + if ( cpumask_subset(cpu_siblings, idlers) ) + cpumask_or(mask, mask, cpu_siblings); +} + +/* + * Clear the bits of all the siblings of cpu from mask. + */ +static inline +void smt_idle_mask_clear(unsigned int cpu, cpumask_t *mask) +{ + cpumask_andnot(mask, mask, per_cpu(cpu_sibling_mask, cpu)); +} + +/* * When a hard affinity change occurs, we may not be able to check some * (any!) of the other runqueues, when looking for the best new processor * for svc (as trylock-s in csched2_cpu_pick() can fail). If that happens, we @@ -853,9 +927,30 @@ runq_tickle(const struct scheduler *ops, struct csched2_vcpu *new, s_time_t now) } /* - * Get a mask of idle, but not tickled, processors that new is - * allowed to run on. If that's not empty, choose someone from there - * (preferrably, the one were new was running on already). + * First of all, consider idle cpus, checking if we can just + * re-use the pcpu where we were running before. + * + * If there are cores where all the siblings are idle, consider + * them first, honoring whatever the spreading-vs-consolidation + * SMT policy wants us to do. + */ + if ( unlikely(sched_smt_power_savings) ) + cpumask_andnot(&mask, &rqd->idle, &rqd->smt_idle); + else + cpumask_copy(&mask, &rqd->smt_idle); + cpumask_and(&mask, &mask, new->vcpu->cpu_hard_affinity); + i = cpumask_test_or_cycle(cpu, &mask); + if ( i < nr_cpu_ids ) + { + SCHED_STAT_CRANK(tickled_idle_cpu); + ipid = i; + goto tickle; + } + + /* + * If there are no fully idle cores, check all idlers, after + * having filtered out pcpus that have been tickled but haven't + * gone through the scheduler yet. */ cpumask_andnot(&mask, &rqd->idle, &rqd->tickled); cpumask_and(&mask, &mask, new->vcpu->cpu_hard_affinity); @@ -947,6 +1042,7 @@ runq_tickle(const struct scheduler *ops, struct csched2_vcpu *new, s_time_t now) (unsigned char *)&d); } __cpumask_set_cpu(ipid, &rqd->tickled); + smt_idle_mask_clear(ipid, &rqd->smt_idle); cpu_raise_softirq(ipid, SCHEDULE_SOFTIRQ); } @@ -1442,8 +1538,10 @@ csched2_cpu_pick(const struct scheduler *ops, struct vcpu *vc) return get_fallback_cpu(svc); } - /* First check to see if we're here because someone else suggested a place - * for us to move. */ + /* + * First check to see if we're here because someone else suggested a place + * for us to move. + */ if ( __test_and_clear_bit(__CSFLAG_runq_migrate_request, &svc->flags) ) { if ( unlikely(svc->migrate_rqd->id < 0) ) @@ -1464,7 +1562,7 @@ csched2_cpu_pick(const struct scheduler *ops, struct vcpu *vc) min_avgload = MAX_LOAD; - /* Find the runqueue with the lowest instantaneous load */ + /* Find the runqueue with the lowest average load. */ for_each_cpu(i, &prv->active_queues) { struct csched2_runqueue_data *rqd; @@ -1507,16 +1605,17 @@ csched2_cpu_pick(const struct scheduler *ops, struct vcpu *vc) /* We didn't find anyone (most likely because of spinlock contention). */ if ( min_rqi == -1 ) - new_cpu = get_fallback_cpu(svc); - else { - cpumask_and(cpumask_scratch, vc->cpu_hard_affinity, - &prv->rqd[min_rqi].active); - new_cpu = cpumask_any(cpumask_scratch); - BUG_ON(new_cpu >= nr_cpu_ids); + new_cpu = get_fallback_cpu(svc); + goto out_up; } -out_up: + cpumask_and(cpumask_scratch, vc->cpu_hard_affinity, + &prv->rqd[min_rqi].active); + new_cpu = cpumask_any(cpumask_scratch); + BUG_ON(new_cpu >= nr_cpu_ids); + + out_up: read_unlock(&prv->lock); if ( unlikely(tb_init_done) ) @@ -2140,7 +2239,11 @@ csched2_schedule( /* Clear "tickled" bit now that we've been scheduled */ if ( cpumask_test_cpu(cpu, &rqd->tickled) ) + { __cpumask_clear_cpu(cpu, &rqd->tickled); + cpumask_andnot(cpumask_scratch, &rqd->idle, &rqd->tickled); + smt_idle_mask_set(cpu, cpumask_scratch, &rqd->smt_idle); + } /* Update credits */ burn_credits(rqd, scurr, now); @@ -2202,7 +2305,10 @@ csched2_schedule( /* Clear the idle mask if necessary */ if ( cpumask_test_cpu(cpu, &rqd->idle) ) + { __cpumask_clear_cpu(cpu, &rqd->idle); + smt_idle_mask_clear(cpu, &rqd->smt_idle); + } snext->start_time = now; @@ -2224,10 +2330,17 @@ csched2_schedule( if ( tasklet_work_scheduled ) { if ( cpumask_test_cpu(cpu, &rqd->idle) ) + { __cpumask_clear_cpu(cpu, &rqd->idle); + smt_idle_mask_clear(cpu, &rqd->smt_idle); + } } else if ( !cpumask_test_cpu(cpu, &rqd->idle) ) + { __cpumask_set_cpu(cpu, &rqd->idle); + cpumask_andnot(cpumask_scratch, &rqd->idle, &rqd->tickled); + smt_idle_mask_set(cpu, cpumask_scratch, &rqd->smt_idle); + } /* Make sure avgload gets updated periodically even * if there's no activity */ update_load(ops, rqd, NULL, 0, now); @@ -2357,6 +2470,8 @@ csched2_dump(const struct scheduler *ops) printk("\tidlers: %s\n", cpustr); cpumask_scnprintf(cpustr, sizeof(cpustr), &prv->rqd[i].tickled); printk("\ttickled: %s\n", cpustr); + cpumask_scnprintf(cpustr, sizeof(cpustr), &prv->rqd[i].smt_idle); + printk("\tfully idle cores: %s\n", cpustr); } printk("Domain info:\n"); @@ -2510,6 +2625,7 @@ init_pdata(struct csched2_private *prv, unsigned int cpu) __cpumask_set_cpu(cpu, &rqd->idle); __cpumask_set_cpu(cpu, &rqd->active); __cpumask_set_cpu(cpu, &prv->initialized); + __cpumask_set_cpu(cpu, &rqd->smt_idle); return rqi; } @@ -2615,6 +2731,7 @@ csched2_deinit_pdata(const struct scheduler *ops, void *pcpu, int cpu) printk(XENLOG_INFO "Removing cpu %d from runqueue %d\n", cpu, rqi); __cpumask_clear_cpu(cpu, &rqd->idle); + __cpumask_clear_cpu(cpu, &rqd->smt_idle); __cpumask_clear_cpu(cpu, &rqd->active); if ( cpumask_empty(&rqd->active) )