sched: Add core wide task selection and scheduling
authorPeter Zijlstra <peterz@infradead.org>
Tue, 17 Nov 2020 23:19:37 +0000 (18:19 -0500)
committerPeter Zijlstra <peterz@infradead.org>
Wed, 12 May 2021 09:43:28 +0000 (11:43 +0200)
Instead of only selecting a local task, select a task for all SMT
siblings for every reschedule on the core (irrespective which logical
CPU does the reschedule).

Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Tested-by: Don Hiatt <dhiatt@digitalocean.com>
Tested-by: Hongyu Ning <hongyu.ning@linux.intel.com>
Tested-by: Vincent Guittot <vincent.guittot@linaro.org>
Link: https://lkml.kernel.org/r/20210422123308.557559654@infradead.org
kernel/sched/core.c
kernel/sched/sched.h

index c057d47..db763f4 100644 (file)
@@ -5282,7 +5282,7 @@ static void put_prev_task_balance(struct rq *rq, struct task_struct *prev,
  * Pick up the highest-prio task:
  */
 static inline struct task_struct *
-pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
+__pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
 {
        const struct sched_class *class;
        struct task_struct *p;
@@ -5323,6 +5323,294 @@ restart:
 }
 
 #ifdef CONFIG_SCHED_CORE
+static inline bool is_task_rq_idle(struct task_struct *t)
+{
+       return (task_rq(t)->idle == t);
+}
+
+static inline bool cookie_equals(struct task_struct *a, unsigned long cookie)
+{
+       return is_task_rq_idle(a) || (a->core_cookie == cookie);
+}
+
+static inline bool cookie_match(struct task_struct *a, struct task_struct *b)
+{
+       if (is_task_rq_idle(a) || is_task_rq_idle(b))
+               return true;
+
+       return a->core_cookie == b->core_cookie;
+}
+
+// XXX fairness/fwd progress conditions
+/*
+ * Returns
+ * - NULL if there is no runnable task for this class.
+ * - the highest priority task for this runqueue if it matches
+ *   rq->core->core_cookie or its priority is greater than max.
+ * - Else returns idle_task.
+ */
+static struct task_struct *
+pick_task(struct rq *rq, const struct sched_class *class, struct task_struct *max)
+{
+       struct task_struct *class_pick, *cookie_pick;
+       unsigned long cookie = rq->core->core_cookie;
+
+       class_pick = class->pick_task(rq);
+       if (!class_pick)
+               return NULL;
+
+       if (!cookie) {
+               /*
+                * If class_pick is tagged, return it only if it has
+                * higher priority than max.
+                */
+               if (max && class_pick->core_cookie &&
+                   prio_less(class_pick, max))
+                       return idle_sched_class.pick_task(rq);
+
+               return class_pick;
+       }
+
+       /*
+        * If class_pick is idle or matches cookie, return early.
+        */
+       if (cookie_equals(class_pick, cookie))
+               return class_pick;
+
+       cookie_pick = sched_core_find(rq, cookie);
+
+       /*
+        * If class > max && class > cookie, it is the highest priority task on
+        * the core (so far) and it must be selected, otherwise we must go with
+        * the cookie pick in order to satisfy the constraint.
+        */
+       if (prio_less(cookie_pick, class_pick) &&
+           (!max || prio_less(max, class_pick)))
+               return class_pick;
+
+       return cookie_pick;
+}
+
+static struct task_struct *
+pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
+{
+       struct task_struct *next, *max = NULL;
+       const struct sched_class *class;
+       const struct cpumask *smt_mask;
+       bool need_sync;
+       int i, j, cpu;
+
+       if (!sched_core_enabled(rq))
+               return __pick_next_task(rq, prev, rf);
+
+       cpu = cpu_of(rq);
+
+       /* Stopper task is switching into idle, no need core-wide selection. */
+       if (cpu_is_offline(cpu)) {
+               /*
+                * Reset core_pick so that we don't enter the fastpath when
+                * coming online. core_pick would already be migrated to
+                * another cpu during offline.
+                */
+               rq->core_pick = NULL;
+               return __pick_next_task(rq, prev, rf);
+       }
+
+       /*
+        * If there were no {en,de}queues since we picked (IOW, the task
+        * pointers are all still valid), and we haven't scheduled the last
+        * pick yet, do so now.
+        *
+        * rq->core_pick can be NULL if no selection was made for a CPU because
+        * it was either offline or went offline during a sibling's core-wide
+        * selection. In this case, do a core-wide selection.
+        */
+       if (rq->core->core_pick_seq == rq->core->core_task_seq &&
+           rq->core->core_pick_seq != rq->core_sched_seq &&
+           rq->core_pick) {
+               WRITE_ONCE(rq->core_sched_seq, rq->core->core_pick_seq);
+
+               next = rq->core_pick;
+               if (next != prev) {
+                       put_prev_task(rq, prev);
+                       set_next_task(rq, next);
+               }
+
+               rq->core_pick = NULL;
+               return next;
+       }
+
+       put_prev_task_balance(rq, prev, rf);
+
+       smt_mask = cpu_smt_mask(cpu);
+
+       /*
+        * core->core_task_seq, core->core_pick_seq, rq->core_sched_seq
+        *
+        * @task_seq guards the task state ({en,de}queues)
+        * @pick_seq is the @task_seq we did a selection on
+        * @sched_seq is the @pick_seq we scheduled
+        *
+        * However, preemptions can cause multiple picks on the same task set.
+        * 'Fix' this by also increasing @task_seq for every pick.
+        */
+       rq->core->core_task_seq++;
+       need_sync = !!rq->core->core_cookie;
+
+       /* reset state */
+       rq->core->core_cookie = 0UL;
+       for_each_cpu(i, smt_mask) {
+               struct rq *rq_i = cpu_rq(i);
+
+               rq_i->core_pick = NULL;
+
+               if (rq_i->core_forceidle) {
+                       need_sync = true;
+                       rq_i->core_forceidle = false;
+               }
+
+               if (i != cpu)
+                       update_rq_clock(rq_i);
+       }
+
+       /*
+        * Try and select tasks for each sibling in decending sched_class
+        * order.
+        */
+       for_each_class(class) {
+again:
+               for_each_cpu_wrap(i, smt_mask, cpu) {
+                       struct rq *rq_i = cpu_rq(i);
+                       struct task_struct *p;
+
+                       if (rq_i->core_pick)
+                               continue;
+
+                       /*
+                        * If this sibling doesn't yet have a suitable task to
+                        * run; ask for the most elegible task, given the
+                        * highest priority task already selected for this
+                        * core.
+                        */
+                       p = pick_task(rq_i, class, max);
+                       if (!p) {
+                               /*
+                                * If there weren't no cookies; we don't need to
+                                * bother with the other siblings.
+                                * If the rest of the core is not running a tagged
+                                * task, i.e.  need_sync == 0, and the current CPU
+                                * which called into the schedule() loop does not
+                                * have any tasks for this class, skip selecting for
+                                * other siblings since there's no point. We don't skip
+                                * for RT/DL because that could make CFS force-idle RT.
+                                */
+                               if (i == cpu && !need_sync && class == &fair_sched_class)
+                                       goto next_class;
+
+                               continue;
+                       }
+
+                       /*
+                        * Optimize the 'normal' case where there aren't any
+                        * cookies and we don't need to sync up.
+                        */
+                       if (i == cpu && !need_sync && !p->core_cookie) {
+                               next = p;
+                               goto done;
+                       }
+
+                       rq_i->core_pick = p;
+
+                       /*
+                        * If this new candidate is of higher priority than the
+                        * previous; and they're incompatible; we need to wipe
+                        * the slate and start over. pick_task makes sure that
+                        * p's priority is more than max if it doesn't match
+                        * max's cookie.
+                        *
+                        * NOTE: this is a linear max-filter and is thus bounded
+                        * in execution time.
+                        */
+                       if (!max || !cookie_match(max, p)) {
+                               struct task_struct *old_max = max;
+
+                               rq->core->core_cookie = p->core_cookie;
+                               max = p;
+
+                               if (old_max) {
+                                       for_each_cpu(j, smt_mask) {
+                                               if (j == i)
+                                                       continue;
+
+                                               cpu_rq(j)->core_pick = NULL;
+                                       }
+                                       goto again;
+                               } else {
+                                       /*
+                                        * Once we select a task for a cpu, we
+                                        * should not be doing an unconstrained
+                                        * pick because it might starve a task
+                                        * on a forced idle cpu.
+                                        */
+                                       need_sync = true;
+                               }
+
+                       }
+               }
+next_class:;
+       }
+
+       rq->core->core_pick_seq = rq->core->core_task_seq;
+       next = rq->core_pick;
+       rq->core_sched_seq = rq->core->core_pick_seq;
+
+       /* Something should have been selected for current CPU */
+       WARN_ON_ONCE(!next);
+
+       /*
+        * Reschedule siblings
+        *
+        * NOTE: L1TF -- at this point we're no longer running the old task and
+        * sending an IPI (below) ensures the sibling will no longer be running
+        * their task. This ensures there is no inter-sibling overlap between
+        * non-matching user state.
+        */
+       for_each_cpu(i, smt_mask) {
+               struct rq *rq_i = cpu_rq(i);
+
+               /*
+                * An online sibling might have gone offline before a task
+                * could be picked for it, or it might be offline but later
+                * happen to come online, but its too late and nothing was
+                * picked for it.  That's Ok - it will pick tasks for itself,
+                * so ignore it.
+                */
+               if (!rq_i->core_pick)
+                       continue;
+
+               if (is_task_rq_idle(rq_i->core_pick) && rq_i->nr_running)
+                       rq_i->core_forceidle = true;
+
+               if (i == cpu) {
+                       rq_i->core_pick = NULL;
+                       continue;
+               }
+
+               /* Did we break L1TF mitigation requirements? */
+               WARN_ON_ONCE(!cookie_match(next, rq_i->core_pick));
+
+               if (rq_i->curr == rq_i->core_pick) {
+                       rq_i->core_pick = NULL;
+                       continue;
+               }
+
+               resched_curr(rq_i);
+       }
+
+done:
+       set_next_task(rq, next);
+       return next;
+}
 
 static inline void sched_core_cpu_starting(unsigned int cpu)
 {
@@ -5354,6 +5642,12 @@ static inline void sched_core_cpu_starting(unsigned int cpu)
 
 static inline void sched_core_cpu_starting(unsigned int cpu) {}
 
+static struct task_struct *
+pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
+{
+       return __pick_next_task(rq, prev, rf);
+}
+
 #endif /* CONFIG_SCHED_CORE */
 
 /*
@@ -8609,7 +8903,12 @@ void __init sched_init(void)
 
 #ifdef CONFIG_SCHED_CORE
                rq->core = NULL;
+               rq->core_pick = NULL;
                rq->core_enabled = 0;
+               rq->core_tree = RB_ROOT;
+               rq->core_forceidle = false;
+
+               rq->core_cookie = 0UL;
 #endif
        }
 
index e43a217..dd44a31 100644 (file)
@@ -1079,11 +1079,16 @@ struct rq {
 #ifdef CONFIG_SCHED_CORE
        /* per rq */
        struct rq               *core;
+       struct task_struct      *core_pick;
        unsigned int            core_enabled;
+       unsigned int            core_sched_seq;
        struct rb_root          core_tree;
+       unsigned char           core_forceidle;
 
        /* shared state */
        unsigned int            core_task_seq;
+       unsigned int            core_pick_seq;
+       unsigned long           core_cookie;
 #endif
 };
 
@@ -2060,7 +2065,6 @@ static inline void put_prev_task(struct rq *rq, struct task_struct *prev)
 
 static inline void set_next_task(struct rq *rq, struct task_struct *next)
 {
-       WARN_ON_ONCE(rq->curr != next);
        next->sched_class->set_next_task(rq, next, false);
 }