MFC r307566 (by sbruno):

Ensure that tasks with a specific cpu set prior to smp starting get
re-attached to a thread running on that cpu.

ref: https://github.com/NextBSD/NextBSD/commit/fcc20e306bc93ebbbe51f3775d1afb527970a2e9

1 files changed, 85 insertions, 43 deletions

diff --git a/sys/kern/subr_gtaskqueue.c b/sys/kern/subr_gtaskqueue.c
index 0518be9..115d834 100644
--- a/sys/kern/subr_gtaskqueue.c
+++ b/sys/kern/subr_gtaskqueue.c
@@ -555,7 +555,7 @@ struct taskq_bind_task {
 };
 
 static void
-taskqgroup_cpu_create(struct taskqgroup *qgroup, int idx)
+taskqgroup_cpu_create(struct taskqgroup *qgroup, int idx, int cpu)
 {
 	struct taskqgroup_cpu *qcpu;
 
@@ -565,7 +565,7 @@ taskqgroup_cpu_create(struct taskqgroup *qgroup, int idx)
 	    taskqueue_thread_enqueue, &qcpu->tgc_taskq);
 	gtaskqueue_start_threads(&qcpu->tgc_taskq, 1, PI_SOFT,
 	    "%s_%d", qgroup->tqg_name, idx);
-	qcpu->tgc_cpu = idx * qgroup->tqg_stride;
+	qcpu->tgc_cpu = cpu;
 }
 
 static void
@@ -634,8 +634,8 @@ taskqgroup_attach(struct taskqgroup *qgroup, struct grouptask *gtask,
 	qgroup->tqg_queue[qid].tgc_cnt++;
 	LIST_INSERT_HEAD(&qgroup->tqg_queue[qid].tgc_tasks, gtask, gt_list);
 	gtask->gt_taskqueue = qgroup->tqg_queue[qid].tgc_taskq;
-	gtask->gt_cpu = qgroup->tqg_queue[qid].tgc_cpu;
 	if (irq != -1 && smp_started) {
+		gtask->gt_cpu = qgroup->tqg_queue[qid].tgc_cpu;
 		CPU_ZERO(&mask);
 		CPU_SET(qgroup->tqg_queue[qid].tgc_cpu, &mask);
 		mtx_unlock(&qgroup->tqg_lock);
@@ -644,6 +644,32 @@ taskqgroup_attach(struct taskqgroup *qgroup, struct grouptask *gtask,
 		mtx_unlock(&qgroup->tqg_lock);
 }
 
+static void
+taskqgroup_attach_deferred(struct taskqgroup *qgroup, struct grouptask *gtask)
+{
+	cpuset_t mask;
+	int qid, cpu;
+
+	mtx_lock(&qgroup->tqg_lock);
+	qid = taskqgroup_find(qgroup, gtask->gt_uniq);
+	cpu = qgroup->tqg_queue[qid].tgc_cpu;
+	if (gtask->gt_irq != -1) {
+		mtx_unlock(&qgroup->tqg_lock);
+
+			CPU_ZERO(&mask);
+			CPU_SET(cpu, &mask);
+			intr_setaffinity(gtask->gt_irq, &mask);
+
+			mtx_lock(&qgroup->tqg_lock);
+	}
+	qgroup->tqg_queue[qid].tgc_cnt++;
+
+	LIST_INSERT_HEAD(&qgroup->tqg_queue[qid].tgc_tasks, gtask,
+			 gt_list);
+	gtask->gt_taskqueue = qgroup->tqg_queue[qid].tgc_taskq;
+	mtx_unlock(&qgroup->tqg_lock);
+}
+
 int
 taskqgroup_attach_cpu(struct taskqgroup *qgroup, struct grouptask *gtask,
 	void *uniq, int cpu, int irq, char *name)
@@ -672,13 +698,46 @@ taskqgroup_attach_cpu(struct taskqgroup *qgroup, struct grouptask *gtask,
 	qgroup->tqg_queue[qid].tgc_cnt++;
 	LIST_INSERT_HEAD(&qgroup->tqg_queue[qid].tgc_tasks, gtask, gt_list);
 	gtask->gt_taskqueue = qgroup->tqg_queue[qid].tgc_taskq;
-	if (irq != -1 && smp_started) {
-		CPU_ZERO(&mask);
-		CPU_SET(qgroup->tqg_queue[qid].tgc_cpu, &mask);
-		mtx_unlock(&qgroup->tqg_lock);
+	cpu = qgroup->tqg_queue[qid].tgc_cpu;
+	mtx_unlock(&qgroup->tqg_lock);
+
+	CPU_ZERO(&mask);
+	CPU_SET(cpu, &mask);
+	if (irq != -1 && smp_started)
 		intr_setaffinity(irq, &mask);
-	} else
+	return (0);
+}
+
+static int
+taskqgroup_attach_cpu_deferred(struct taskqgroup *qgroup, struct grouptask *gtask)
+{
+	cpuset_t mask;
+	int i, qid, irq, cpu;
+
+	qid = -1;
+	irq = gtask->gt_irq;
+	cpu = gtask->gt_cpu;
+	MPASS(smp_started);
+	mtx_lock(&qgroup->tqg_lock);
+	for (i = 0; i < qgroup->tqg_cnt; i++)
+		if (qgroup->tqg_queue[i].tgc_cpu == cpu) {
+			qid = i;
+			break;
+		}
+	if (qid == -1) {
 		mtx_unlock(&qgroup->tqg_lock);
+		return (EINVAL);
+	}
+	qgroup->tqg_queue[qid].tgc_cnt++;
+	LIST_INSERT_HEAD(&qgroup->tqg_queue[qid].tgc_tasks, gtask, gt_list);
+	gtask->gt_taskqueue = qgroup->tqg_queue[qid].tgc_taskq;
+	mtx_unlock(&qgroup->tqg_lock);
+
+	CPU_ZERO(&mask);
+	CPU_SET(cpu, &mask);
+
+	if (irq != -1)
+		intr_setaffinity(irq, &mask);
 	return (0);
 }
 
@@ -742,9 +801,8 @@ static int
 _taskqgroup_adjust(struct taskqgroup *qgroup, int cnt, int stride)
 {
 	LIST_HEAD(, grouptask) gtask_head = LIST_HEAD_INITIALIZER(NULL);
-	cpuset_t mask;
 	struct grouptask *gtask;
-	int i, k, old_cnt, qid, cpu;
+	int i, k, old_cnt, old_cpu, cpu;
 
 	mtx_assert(&qgroup->tqg_lock, MA_OWNED);
 
@@ -759,6 +817,9 @@ _taskqgroup_adjust(struct taskqgroup *qgroup, int cnt, int stride)
 	}
 	qgroup->tqg_adjusting = 1;
 	old_cnt = qgroup->tqg_cnt;
+	old_cpu = 0;
+	if (old_cnt < cnt)
+		old_cpu = qgroup->tqg_queue[old_cnt].tgc_cpu;
 	mtx_unlock(&qgroup->tqg_lock);
 	/*
 	 * Set up queue for tasks added before boot.
@@ -772,8 +833,13 @@ _taskqgroup_adjust(struct taskqgroup *qgroup, int cnt, int stride)
 	/*
 	 * If new taskq threads have been added.
 	 */
-	for (i = old_cnt; i < cnt; i++)
-		taskqgroup_cpu_create(qgroup, i);
+	cpu = old_cpu;
+	for (i = old_cnt; i < cnt; i++) {
+		for (k = 0; k < qgroup->tqg_stride; k++)
+			cpu = CPU_NEXT(cpu);
+
+		taskqgroup_cpu_create(qgroup, i, cpu);
+	}
 	mtx_lock(&qgroup->tqg_lock);
 	qgroup->tqg_cnt = cnt;
 	qgroup->tqg_stride = stride;
@@ -789,39 +855,15 @@ _taskqgroup_adjust(struct taskqgroup *qgroup, int cnt, int stride)
 		}
 	}
 
+	mtx_unlock(&qgroup->tqg_lock);
+
 	while ((gtask = LIST_FIRST(&gtask_head))) {
 		LIST_REMOVE(gtask, gt_list);
 		if (gtask->gt_cpu == -1)
-			qid = taskqgroup_find(qgroup, gtask->gt_uniq);
-		else {
-			for (i = 0; i < qgroup->tqg_cnt; i++)
-				if (qgroup->tqg_queue[i].tgc_cpu == gtask->gt_cpu) {
-					qid = i;
-					break;
-				}
-		}
-		qgroup->tqg_queue[qid].tgc_cnt++;
-		LIST_INSERT_HEAD(&qgroup->tqg_queue[qid].tgc_tasks, gtask,
-		    gt_list);
-		gtask->gt_taskqueue = qgroup->tqg_queue[qid].tgc_taskq;
-	}
-	/*
-	 * Set new CPU and IRQ affinity
-	 */
-	cpu = CPU_FIRST();
-	for (i = 0; i < cnt; i++) {
-		qgroup->tqg_queue[i].tgc_cpu = cpu;
-		for (k = 0; k < qgroup->tqg_stride; k++)
-			cpu = CPU_NEXT(cpu);
-		CPU_ZERO(&mask);
-		CPU_SET(qgroup->tqg_queue[i].tgc_cpu, &mask);
-		LIST_FOREACH(gtask, &qgroup->tqg_queue[i].tgc_tasks, gt_list) {
-			if (gtask->gt_irq == -1)
-				continue;
-			intr_setaffinity(gtask->gt_irq, &mask);
-		}
+			taskqgroup_attach_deferred(qgroup, gtask);
+		else if (taskqgroup_attach_cpu_deferred(qgroup, gtask))
+			taskqgroup_attach_deferred(qgroup, gtask);
 	}
-	mtx_unlock(&qgroup->tqg_lock);
 
 	/*
 	 * If taskq thread count has been reduced.
@@ -838,12 +880,12 @@ _taskqgroup_adjust(struct taskqgroup *qgroup, int cnt, int stride)
 }
 
 int
-taskqgroup_adjust(struct taskqgroup *qgroup, int cpu, int stride)
+taskqgroup_adjust(struct taskqgroup *qgroup, int cnt, int stride)
 {
 	int error;
 
 	mtx_lock(&qgroup->tqg_lock);
-	error = _taskqgroup_adjust(qgroup, cpu, stride);
+	error = _taskqgroup_adjust(qgroup, cnt, stride);
 	mtx_unlock(&qgroup->tqg_lock);
 
 	return (error);