sched_ext: Replace SCX_TASK_BAL_KEEP with SCX_RQ_BAL_KEEP
authorTejun Heo <tj@kernel.org>
Wed, 4 Sep 2024 07:54:28 +0000 (21:54 -1000)
committerTejun Heo <tj@kernel.org>
Wed, 4 Sep 2024 07:54:28 +0000 (21:54 -1000)
SCX_TASK_BAL_KEEP is used by balance_one() to tell pick_next_task_scx() to
keep running the current task. It's not really a task property. Replace it
with SCX_RQ_BAL_KEEP which resides in rq->scx.flags and is a better fit for
the usage. Also, the existing clearing rule is unnecessarily strict and
makes it difficult to use with core-sched. Just clear it on entry to
balance_one().

Signed-off-by: Tejun Heo <tj@kernel.org>
include/linux/sched/ext.h
kernel/sched/ext.c
kernel/sched/sched.h

index 69f68e2121a8f8586a8f8ac5d91be091a50f9f6c..db2a266113acbd5d8efefa008a9b01a656546b7d 100644 (file)
@@ -71,7 +71,6 @@ struct scx_dispatch_q {
 /* scx_entity.flags */
 enum scx_ent_flags {
        SCX_TASK_QUEUED         = 1 << 0, /* on ext runqueue */
-       SCX_TASK_BAL_KEEP       = 1 << 1, /* balance decided to keep current */
        SCX_TASK_RESET_RUNNABLE_AT = 1 << 2, /* runnable_at should be reset */
        SCX_TASK_DEQD_FOR_SLEEP = 1 << 3, /* last dequeue was for SLEEP */
 
index be86dbfa75a8ec50c90f7a0b20e2973b47dc8df1..fd979c69de1f803954e1ac23e666052f37e9c3c5 100644 (file)
@@ -2508,6 +2508,7 @@ static int balance_one(struct rq *rq, struct task_struct *prev, bool local)
 
        lockdep_assert_rq_held(rq);
        rq->scx.flags |= SCX_RQ_IN_BALANCE;
+       rq->scx.flags &= ~SCX_RQ_BAL_KEEP;
 
        if (static_branch_unlikely(&scx_ops_cpu_preempt) &&
            unlikely(rq->scx.cpu_released)) {
@@ -2523,7 +2524,6 @@ static int balance_one(struct rq *rq, struct task_struct *prev, bool local)
        }
 
        if (prev_on_scx) {
-               WARN_ON_ONCE(local && (prev->scx.flags & SCX_TASK_BAL_KEEP));
                update_curr_scx(rq);
 
                /*
@@ -2538,13 +2538,13 @@ static int balance_one(struct rq *rq, struct task_struct *prev, bool local)
                 *
                 * When balancing a remote CPU for core-sched, there won't be a
                 * following put_prev_task_scx() call and we don't own
-                * %SCX_TASK_BAL_KEEP. Instead, pick_task_scx() will test the
-                * same conditions later and pick @rq->curr accordingly.
+                * %SCX_RQ_BAL_KEEP. Instead, pick_task_scx() will test the same
+                * conditions later and pick @rq->curr accordingly.
                 */
                if ((prev->scx.flags & SCX_TASK_QUEUED) &&
                    prev->scx.slice && !scx_ops_bypassing()) {
                        if (local)
-                               prev->scx.flags |= SCX_TASK_BAL_KEEP;
+                               rq->scx.flags |= SCX_RQ_BAL_KEEP;
                        goto has_tasks;
                }
        }
@@ -2604,7 +2604,7 @@ no_tasks:
        if ((prev->scx.flags & SCX_TASK_QUEUED) &&
            (!static_branch_unlikely(&scx_ops_enq_last) || scx_ops_bypassing())) {
                if (local)
-                       prev->scx.flags |= SCX_TASK_BAL_KEEP;
+                       rq->scx.flags |= SCX_RQ_BAL_KEEP;
                goto has_tasks;
        }
        rq->scx.flags &= ~SCX_RQ_IN_BALANCE;
@@ -2726,8 +2726,6 @@ static void put_prev_task_scx(struct rq *rq, struct task_struct *p,
                SCX_CALL_OP_TASK(SCX_KF_REST, stopping, p, true);
 
        if (p->scx.flags & SCX_TASK_QUEUED) {
-               p->scx.flags &= ~SCX_TASK_BAL_KEEP;
-
                set_task_runnable(rq, p);
 
                /*
@@ -2772,8 +2770,8 @@ static struct task_struct *pick_next_task_scx(struct rq *rq,
         * if necessary and keep running @prev. Otherwise, pop the first one
         * from the local DSQ.
         */
-       if (prev->scx.flags & SCX_TASK_BAL_KEEP) {
-               prev->scx.flags &= ~SCX_TASK_BAL_KEEP;
+       if ((rq->scx.flags & SCX_RQ_BAL_KEEP) &&
+           !WARN_ON_ONCE(prev->sched_class != &ext_sched_class)) {
                p = prev;
                if (!p->scx.slice)
                        p->scx.slice = SCX_SLICE_DFL;
@@ -2841,7 +2839,7 @@ bool scx_prio_less(const struct task_struct *a, const struct task_struct *b,
  *
  * As put_prev_task_scx() hasn't been called on remote CPUs, we can't just look
  * at the first task in the local dsq. @rq->curr has to be considered explicitly
- * to mimic %SCX_TASK_BAL_KEEP.
+ * to mimic %SCX_RQ_BAL_KEEP.
  */
 static struct task_struct *pick_task_scx(struct rq *rq)
 {
@@ -3872,7 +3870,7 @@ bool task_should_scx(struct task_struct *p)
  *
  * b. ops.dispatch() is ignored.
  *
- * c. balance_scx() does not set %SCX_TASK_BAL_KEEP on non-zero slice as slice
+ * c. balance_scx() does not set %SCX_RQ_BAL_KEEP on non-zero slice as slice
  *    can't be trusted. Whenever a tick triggers, the running task is rotated to
  *    the tail of the queue with core_sched_at touched.
  *
index 1eda2ce3178766c362595420b7ceb42ae2c20bd8..477e7a861d10dc0a5b9e460eee0fb1777c9b7ce8 100644 (file)
@@ -737,6 +737,7 @@ enum scx_rq_flags {
         */
        SCX_RQ_ONLINE           = 1 << 0,
        SCX_RQ_CAN_STOP_TICK    = 1 << 1,
+       SCX_RQ_BAL_KEEP         = 1 << 2, /* balance decided to keep current */
 
        SCX_RQ_IN_WAKEUP        = 1 << 16,
        SCX_RQ_IN_BALANCE       = 1 << 17,