xref: /freebsd/sys/kern/subr_gtaskqueue.c (revision 1d386b48a555f61cb7325543adbbb5c3f3407a66)
1 /*-
2  * Copyright (c) 2000 Doug Rabson
3  * Copyright (c) 2014 Jeff Roberson
4  * Copyright (c) 2016 Matthew Macy
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  *
16  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
17  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
20  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
22  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
23  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
26  * SUCH DAMAGE.
27  */
28 
29 #include <sys/cdefs.h>
30 #include <sys/param.h>
31 #include <sys/systm.h>
32 #include <sys/bus.h>
33 #include <sys/cpuset.h>
34 #include <sys/kernel.h>
35 #include <sys/kthread.h>
36 #include <sys/libkern.h>
37 #include <sys/limits.h>
38 #include <sys/lock.h>
39 #include <sys/malloc.h>
40 #include <sys/mutex.h>
41 #include <sys/proc.h>
42 #include <sys/epoch.h>
43 #include <sys/sched.h>
44 #include <sys/smp.h>
45 #include <sys/gtaskqueue.h>
46 #include <sys/unistd.h>
47 #include <machine/stdarg.h>
48 
49 static MALLOC_DEFINE(M_GTASKQUEUE, "gtaskqueue", "Group Task Queues");
50 static void	gtaskqueue_thread_enqueue(void *);
51 static void	gtaskqueue_thread_loop(void *arg);
52 static int	task_is_running(struct gtaskqueue *queue, struct gtask *gtask);
53 static void	gtaskqueue_drain_locked(struct gtaskqueue *queue, struct gtask *gtask);
54 
55 TASKQGROUP_DEFINE(softirq, mp_ncpus, 1);
56 
57 struct gtaskqueue_busy {
58 	struct gtask		*tb_running;
59 	u_int			 tb_seq;
60 	LIST_ENTRY(gtaskqueue_busy) tb_link;
61 };
62 
63 typedef void (*gtaskqueue_enqueue_fn)(void *context);
64 
65 struct gtaskqueue {
66 	STAILQ_HEAD(, gtask)	tq_queue;
67 	LIST_HEAD(, gtaskqueue_busy) tq_active;
68 	u_int			tq_seq;
69 	int			tq_callouts;
70 	struct mtx_padalign	tq_mutex;
71 	gtaskqueue_enqueue_fn	tq_enqueue;
72 	void			*tq_context;
73 	char			*tq_name;
74 	struct thread		**tq_threads;
75 	int			tq_tcount;
76 	int			tq_spin;
77 	int			tq_flags;
78 	taskqueue_callback_fn	tq_callbacks[TASKQUEUE_NUM_CALLBACKS];
79 	void			*tq_cb_contexts[TASKQUEUE_NUM_CALLBACKS];
80 };
81 
82 #define	TQ_FLAGS_ACTIVE		(1 << 0)
83 #define	TQ_FLAGS_BLOCKED	(1 << 1)
84 #define	TQ_FLAGS_UNLOCKED_ENQUEUE	(1 << 2)
85 
86 #define	DT_CALLOUT_ARMED	(1 << 0)
87 
88 #define	TQ_LOCK(tq)							\
89 	do {								\
90 		if ((tq)->tq_spin)					\
91 			mtx_lock_spin(&(tq)->tq_mutex);			\
92 		else							\
93 			mtx_lock(&(tq)->tq_mutex);			\
94 	} while (0)
95 #define	TQ_ASSERT_LOCKED(tq)	mtx_assert(&(tq)->tq_mutex, MA_OWNED)
96 
97 #define	TQ_UNLOCK(tq)							\
98 	do {								\
99 		if ((tq)->tq_spin)					\
100 			mtx_unlock_spin(&(tq)->tq_mutex);		\
101 		else							\
102 			mtx_unlock(&(tq)->tq_mutex);			\
103 	} while (0)
104 #define	TQ_ASSERT_UNLOCKED(tq)	mtx_assert(&(tq)->tq_mutex, MA_NOTOWNED)
105 
106 #ifdef INVARIANTS
107 static void
108 gtask_dump(struct gtask *gtask)
109 {
110 	printf("gtask: %p ta_flags=%x ta_priority=%d ta_func=%p ta_context=%p\n",
111 	       gtask, gtask->ta_flags, gtask->ta_priority, gtask->ta_func, gtask->ta_context);
112 }
113 #endif
114 
115 static __inline int
116 TQ_SLEEP(struct gtaskqueue *tq, void *p, const char *wm)
117 {
118 	if (tq->tq_spin)
119 		return (msleep_spin(p, (struct mtx *)&tq->tq_mutex, wm, 0));
120 	return (msleep(p, &tq->tq_mutex, 0, wm, 0));
121 }
122 
123 static struct gtaskqueue *
124 _gtaskqueue_create(const char *name, int mflags,
125 		 taskqueue_enqueue_fn enqueue, void *context,
126 		 int mtxflags, const char *mtxname __unused)
127 {
128 	struct gtaskqueue *queue;
129 	char *tq_name;
130 
131 	tq_name = malloc(TASKQUEUE_NAMELEN, M_GTASKQUEUE, mflags | M_ZERO);
132 	if (!tq_name)
133 		return (NULL);
134 
135 	snprintf(tq_name, TASKQUEUE_NAMELEN, "%s", (name) ? name : "taskqueue");
136 
137 	queue = malloc(sizeof(struct gtaskqueue), M_GTASKQUEUE, mflags | M_ZERO);
138 	if (!queue) {
139 		free(tq_name, M_GTASKQUEUE);
140 		return (NULL);
141 	}
142 
143 	STAILQ_INIT(&queue->tq_queue);
144 	LIST_INIT(&queue->tq_active);
145 	queue->tq_enqueue = enqueue;
146 	queue->tq_context = context;
147 	queue->tq_name = tq_name;
148 	queue->tq_spin = (mtxflags & MTX_SPIN) != 0;
149 	queue->tq_flags |= TQ_FLAGS_ACTIVE;
150 	if (enqueue == gtaskqueue_thread_enqueue)
151 		queue->tq_flags |= TQ_FLAGS_UNLOCKED_ENQUEUE;
152 	mtx_init(&queue->tq_mutex, tq_name, NULL, mtxflags);
153 
154 	return (queue);
155 }
156 
157 /*
158  * Signal a taskqueue thread to terminate.
159  */
160 static void
161 gtaskqueue_terminate(struct thread **pp, struct gtaskqueue *tq)
162 {
163 
164 	while (tq->tq_tcount > 0 || tq->tq_callouts > 0) {
165 		wakeup(tq);
166 		TQ_SLEEP(tq, pp, "gtq_destroy");
167 	}
168 }
169 
170 static void __unused
171 gtaskqueue_free(struct gtaskqueue *queue)
172 {
173 
174 	TQ_LOCK(queue);
175 	queue->tq_flags &= ~TQ_FLAGS_ACTIVE;
176 	gtaskqueue_terminate(queue->tq_threads, queue);
177 	KASSERT(LIST_EMPTY(&queue->tq_active), ("Tasks still running?"));
178 	KASSERT(queue->tq_callouts == 0, ("Armed timeout tasks"));
179 	mtx_destroy(&queue->tq_mutex);
180 	free(queue->tq_threads, M_GTASKQUEUE);
181 	free(queue->tq_name, M_GTASKQUEUE);
182 	free(queue, M_GTASKQUEUE);
183 }
184 
185 /*
186  * Wait for all to complete, then prevent it from being enqueued
187  */
188 void
189 grouptask_block(struct grouptask *grouptask)
190 {
191 	struct gtaskqueue *queue = grouptask->gt_taskqueue;
192 	struct gtask *gtask = &grouptask->gt_task;
193 
194 #ifdef INVARIANTS
195 	if (queue == NULL) {
196 		gtask_dump(gtask);
197 		panic("queue == NULL");
198 	}
199 #endif
200 	TQ_LOCK(queue);
201 	gtask->ta_flags |= TASK_NOENQUEUE;
202   	gtaskqueue_drain_locked(queue, gtask);
203 	TQ_UNLOCK(queue);
204 }
205 
206 void
207 grouptask_unblock(struct grouptask *grouptask)
208 {
209 	struct gtaskqueue *queue = grouptask->gt_taskqueue;
210 	struct gtask *gtask = &grouptask->gt_task;
211 
212 #ifdef INVARIANTS
213 	if (queue == NULL) {
214 		gtask_dump(gtask);
215 		panic("queue == NULL");
216 	}
217 #endif
218 	TQ_LOCK(queue);
219 	gtask->ta_flags &= ~TASK_NOENQUEUE;
220 	TQ_UNLOCK(queue);
221 }
222 
223 int
224 grouptaskqueue_enqueue(struct gtaskqueue *queue, struct gtask *gtask)
225 {
226 #ifdef INVARIANTS
227 	if (queue == NULL) {
228 		gtask_dump(gtask);
229 		panic("queue == NULL");
230 	}
231 #endif
232 	TQ_LOCK(queue);
233 	if (gtask->ta_flags & TASK_ENQUEUED) {
234 		TQ_UNLOCK(queue);
235 		return (0);
236 	}
237 	if (gtask->ta_flags & TASK_NOENQUEUE) {
238 		TQ_UNLOCK(queue);
239 		return (EAGAIN);
240 	}
241 	STAILQ_INSERT_TAIL(&queue->tq_queue, gtask, ta_link);
242 	gtask->ta_flags |= TASK_ENQUEUED;
243 	TQ_UNLOCK(queue);
244 	if ((queue->tq_flags & TQ_FLAGS_BLOCKED) == 0)
245 		queue->tq_enqueue(queue->tq_context);
246 	return (0);
247 }
248 
249 static void
250 gtaskqueue_task_nop_fn(void *context)
251 {
252 }
253 
254 /*
255  * Block until all currently queued tasks in this taskqueue
256  * have begun execution.  Tasks queued during execution of
257  * this function are ignored.
258  */
259 static void
260 gtaskqueue_drain_tq_queue(struct gtaskqueue *queue)
261 {
262 	struct gtask t_barrier;
263 
264 	if (STAILQ_EMPTY(&queue->tq_queue))
265 		return;
266 
267 	/*
268 	 * Enqueue our barrier after all current tasks, but with
269 	 * the highest priority so that newly queued tasks cannot
270 	 * pass it.  Because of the high priority, we can not use
271 	 * taskqueue_enqueue_locked directly (which drops the lock
272 	 * anyway) so just insert it at tail while we have the
273 	 * queue lock.
274 	 */
275 	GTASK_INIT(&t_barrier, 0, USHRT_MAX, gtaskqueue_task_nop_fn, &t_barrier);
276 	STAILQ_INSERT_TAIL(&queue->tq_queue, &t_barrier, ta_link);
277 	t_barrier.ta_flags |= TASK_ENQUEUED;
278 
279 	/*
280 	 * Once the barrier has executed, all previously queued tasks
281 	 * have completed or are currently executing.
282 	 */
283 	while (t_barrier.ta_flags & TASK_ENQUEUED)
284 		TQ_SLEEP(queue, &t_barrier, "gtq_qdrain");
285 }
286 
287 /*
288  * Block until all currently executing tasks for this taskqueue
289  * complete.  Tasks that begin execution during the execution
290  * of this function are ignored.
291  */
292 static void
293 gtaskqueue_drain_tq_active(struct gtaskqueue *queue)
294 {
295 	struct gtaskqueue_busy *tb;
296 	u_int seq;
297 
298 	if (LIST_EMPTY(&queue->tq_active))
299 		return;
300 
301 	/* Block taskq_terminate().*/
302 	queue->tq_callouts++;
303 
304 	/* Wait for any active task with sequence from the past. */
305 	seq = queue->tq_seq;
306 restart:
307 	LIST_FOREACH(tb, &queue->tq_active, tb_link) {
308 		if ((int)(tb->tb_seq - seq) <= 0) {
309 			TQ_SLEEP(queue, tb->tb_running, "gtq_adrain");
310 			goto restart;
311 		}
312 	}
313 
314 	/* Release taskqueue_terminate(). */
315 	queue->tq_callouts--;
316 	if ((queue->tq_flags & TQ_FLAGS_ACTIVE) == 0)
317 		wakeup_one(queue->tq_threads);
318 }
319 
320 void
321 gtaskqueue_block(struct gtaskqueue *queue)
322 {
323 
324 	TQ_LOCK(queue);
325 	queue->tq_flags |= TQ_FLAGS_BLOCKED;
326 	TQ_UNLOCK(queue);
327 }
328 
329 void
330 gtaskqueue_unblock(struct gtaskqueue *queue)
331 {
332 
333 	TQ_LOCK(queue);
334 	queue->tq_flags &= ~TQ_FLAGS_BLOCKED;
335 	if (!STAILQ_EMPTY(&queue->tq_queue))
336 		queue->tq_enqueue(queue->tq_context);
337 	TQ_UNLOCK(queue);
338 }
339 
340 static void
341 gtaskqueue_run_locked(struct gtaskqueue *queue)
342 {
343 	struct epoch_tracker et;
344 	struct gtaskqueue_busy tb;
345 	struct gtask *gtask;
346 	bool in_net_epoch;
347 
348 	KASSERT(queue != NULL, ("tq is NULL"));
349 	TQ_ASSERT_LOCKED(queue);
350 	tb.tb_running = NULL;
351 	LIST_INSERT_HEAD(&queue->tq_active, &tb, tb_link);
352 	in_net_epoch = false;
353 
354 	while ((gtask = STAILQ_FIRST(&queue->tq_queue)) != NULL) {
355 		STAILQ_REMOVE_HEAD(&queue->tq_queue, ta_link);
356 		gtask->ta_flags &= ~TASK_ENQUEUED;
357 		tb.tb_running = gtask;
358 		tb.tb_seq = ++queue->tq_seq;
359 		TQ_UNLOCK(queue);
360 
361 		KASSERT(gtask->ta_func != NULL, ("task->ta_func is NULL"));
362 		if (!in_net_epoch && TASK_IS_NET(gtask)) {
363 			in_net_epoch = true;
364 			NET_EPOCH_ENTER(et);
365 		} else if (in_net_epoch && !TASK_IS_NET(gtask)) {
366 			NET_EPOCH_EXIT(et);
367 			in_net_epoch = false;
368 		}
369 		gtask->ta_func(gtask->ta_context);
370 
371 		TQ_LOCK(queue);
372 		wakeup(gtask);
373 	}
374 	if (in_net_epoch)
375 		NET_EPOCH_EXIT(et);
376 	LIST_REMOVE(&tb, tb_link);
377 }
378 
379 static int
380 task_is_running(struct gtaskqueue *queue, struct gtask *gtask)
381 {
382 	struct gtaskqueue_busy *tb;
383 
384 	TQ_ASSERT_LOCKED(queue);
385 	LIST_FOREACH(tb, &queue->tq_active, tb_link) {
386 		if (tb->tb_running == gtask)
387 			return (1);
388 	}
389 	return (0);
390 }
391 
392 static int
393 gtaskqueue_cancel_locked(struct gtaskqueue *queue, struct gtask *gtask)
394 {
395 
396 	if (gtask->ta_flags & TASK_ENQUEUED)
397 		STAILQ_REMOVE(&queue->tq_queue, gtask, gtask, ta_link);
398 	gtask->ta_flags &= ~TASK_ENQUEUED;
399 	return (task_is_running(queue, gtask) ? EBUSY : 0);
400 }
401 
402 int
403 gtaskqueue_cancel(struct gtaskqueue *queue, struct gtask *gtask)
404 {
405 	int error;
406 
407 	TQ_LOCK(queue);
408 	error = gtaskqueue_cancel_locked(queue, gtask);
409 	TQ_UNLOCK(queue);
410 
411 	return (error);
412 }
413 
414 static void
415 gtaskqueue_drain_locked(struct gtaskqueue *queue, struct gtask *gtask)
416 {
417 	while ((gtask->ta_flags & TASK_ENQUEUED) || task_is_running(queue, gtask))
418 		TQ_SLEEP(queue, gtask, "gtq_drain");
419 }
420 
421 void
422 gtaskqueue_drain(struct gtaskqueue *queue, struct gtask *gtask)
423 {
424 
425 	if (!queue->tq_spin)
426 		WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, __func__);
427 
428 	TQ_LOCK(queue);
429 	gtaskqueue_drain_locked(queue, gtask);
430 	TQ_UNLOCK(queue);
431 }
432 
433 void
434 gtaskqueue_drain_all(struct gtaskqueue *queue)
435 {
436 
437 	if (!queue->tq_spin)
438 		WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, __func__);
439 
440 	TQ_LOCK(queue);
441 	gtaskqueue_drain_tq_queue(queue);
442 	gtaskqueue_drain_tq_active(queue);
443 	TQ_UNLOCK(queue);
444 }
445 
446 static int
447 _gtaskqueue_start_threads(struct gtaskqueue **tqp, int count, int pri,
448     cpuset_t *mask, const char *name, va_list ap)
449 {
450 	char ktname[MAXCOMLEN + 1];
451 	struct thread *td;
452 	struct gtaskqueue *tq;
453 	int i, error;
454 
455 	if (count <= 0)
456 		return (EINVAL);
457 
458 	vsnprintf(ktname, sizeof(ktname), name, ap);
459 	tq = *tqp;
460 
461 	tq->tq_threads = malloc(sizeof(struct thread *) * count, M_GTASKQUEUE,
462 	    M_NOWAIT | M_ZERO);
463 	if (tq->tq_threads == NULL) {
464 		printf("%s: no memory for %s threads\n", __func__, ktname);
465 		return (ENOMEM);
466 	}
467 
468 	for (i = 0; i < count; i++) {
469 		if (count == 1)
470 			error = kthread_add(gtaskqueue_thread_loop, tqp, NULL,
471 			    &tq->tq_threads[i], RFSTOPPED, 0, "%s", ktname);
472 		else
473 			error = kthread_add(gtaskqueue_thread_loop, tqp, NULL,
474 			    &tq->tq_threads[i], RFSTOPPED, 0,
475 			    "%s_%d", ktname, i);
476 		if (error) {
477 			/* should be ok to continue, taskqueue_free will dtrt */
478 			printf("%s: kthread_add(%s): error %d", __func__,
479 			    ktname, error);
480 			tq->tq_threads[i] = NULL;		/* paranoid */
481 		} else
482 			tq->tq_tcount++;
483 	}
484 	for (i = 0; i < count; i++) {
485 		if (tq->tq_threads[i] == NULL)
486 			continue;
487 		td = tq->tq_threads[i];
488 		if (mask) {
489 			error = cpuset_setthread(td->td_tid, mask);
490 			/*
491 			 * Failing to pin is rarely an actual fatal error;
492 			 * it'll just affect performance.
493 			 */
494 			if (error)
495 				printf("%s: curthread=%llu: can't pin; "
496 				    "error=%d\n",
497 				    __func__,
498 				    (unsigned long long) td->td_tid,
499 				    error);
500 		}
501 		thread_lock(td);
502 		sched_prio(td, pri);
503 		sched_add(td, SRQ_BORING);
504 	}
505 
506 	return (0);
507 }
508 
509 static int
510 gtaskqueue_start_threads(struct gtaskqueue **tqp, int count, int pri,
511     const char *name, ...)
512 {
513 	va_list ap;
514 	int error;
515 
516 	va_start(ap, name);
517 	error = _gtaskqueue_start_threads(tqp, count, pri, NULL, name, ap);
518 	va_end(ap);
519 	return (error);
520 }
521 
522 static inline void
523 gtaskqueue_run_callback(struct gtaskqueue *tq,
524     enum taskqueue_callback_type cb_type)
525 {
526 	taskqueue_callback_fn tq_callback;
527 
528 	TQ_ASSERT_UNLOCKED(tq);
529 	tq_callback = tq->tq_callbacks[cb_type];
530 	if (tq_callback != NULL)
531 		tq_callback(tq->tq_cb_contexts[cb_type]);
532 }
533 
534 static void
535 gtaskqueue_thread_loop(void *arg)
536 {
537 	struct gtaskqueue **tqp, *tq;
538 
539 	tqp = arg;
540 	tq = *tqp;
541 	gtaskqueue_run_callback(tq, TASKQUEUE_CALLBACK_TYPE_INIT);
542 	TQ_LOCK(tq);
543 	while ((tq->tq_flags & TQ_FLAGS_ACTIVE) != 0) {
544 		/* XXX ? */
545 		gtaskqueue_run_locked(tq);
546 		/*
547 		 * Because taskqueue_run() can drop tq_mutex, we need to
548 		 * check if the TQ_FLAGS_ACTIVE flag wasn't removed in the
549 		 * meantime, which means we missed a wakeup.
550 		 */
551 		if ((tq->tq_flags & TQ_FLAGS_ACTIVE) == 0)
552 			break;
553 		TQ_SLEEP(tq, tq, "-");
554 	}
555 	gtaskqueue_run_locked(tq);
556 	/*
557 	 * This thread is on its way out, so just drop the lock temporarily
558 	 * in order to call the shutdown callback.  This allows the callback
559 	 * to look at the taskqueue, even just before it dies.
560 	 */
561 	TQ_UNLOCK(tq);
562 	gtaskqueue_run_callback(tq, TASKQUEUE_CALLBACK_TYPE_SHUTDOWN);
563 	TQ_LOCK(tq);
564 
565 	/* rendezvous with thread that asked us to terminate */
566 	tq->tq_tcount--;
567 	wakeup_one(tq->tq_threads);
568 	TQ_UNLOCK(tq);
569 	kthread_exit();
570 }
571 
572 static void
573 gtaskqueue_thread_enqueue(void *context)
574 {
575 	struct gtaskqueue **tqp, *tq;
576 
577 	tqp = context;
578 	tq = *tqp;
579 	wakeup_any(tq);
580 }
581 
582 static struct gtaskqueue *
583 gtaskqueue_create_fast(const char *name, int mflags,
584 		 taskqueue_enqueue_fn enqueue, void *context)
585 {
586 	return _gtaskqueue_create(name, mflags, enqueue, context,
587 			MTX_SPIN, "fast_taskqueue");
588 }
589 
590 struct taskqgroup_cpu {
591 	LIST_HEAD(, grouptask) tgc_tasks;
592 	struct gtaskqueue *tgc_taskq;
593 	int		tgc_cnt;
594 	int		tgc_cpu;
595 };
596 
597 struct taskqgroup {
598 	struct taskqgroup_cpu tqg_queue[MAXCPU];
599 	struct mtx	tqg_lock;
600 	const char *	tqg_name;
601 	int		tqg_cnt;
602 };
603 
604 struct taskq_bind_task {
605 	struct gtask bt_task;
606 	int	bt_cpuid;
607 };
608 
609 static void
610 taskqgroup_cpu_create(struct taskqgroup *qgroup, int idx, int cpu)
611 {
612 	struct taskqgroup_cpu *qcpu;
613 
614 	qcpu = &qgroup->tqg_queue[idx];
615 	LIST_INIT(&qcpu->tgc_tasks);
616 	qcpu->tgc_taskq = gtaskqueue_create_fast(NULL, M_WAITOK,
617 	    taskqueue_thread_enqueue, &qcpu->tgc_taskq);
618 	gtaskqueue_start_threads(&qcpu->tgc_taskq, 1, PI_SOFT,
619 	    "%s_%d", qgroup->tqg_name, idx);
620 	qcpu->tgc_cpu = cpu;
621 }
622 
623 /*
624  * Find the taskq with least # of tasks that doesn't currently have any
625  * other queues from the uniq identifier.
626  */
627 static int
628 taskqgroup_find(struct taskqgroup *qgroup, void *uniq)
629 {
630 	struct grouptask *n;
631 	int i, idx, mincnt;
632 	int strict;
633 
634 	mtx_assert(&qgroup->tqg_lock, MA_OWNED);
635 	KASSERT(qgroup->tqg_cnt != 0,
636 	    ("qgroup %s has no queues", qgroup->tqg_name));
637 
638 	/*
639 	 * Two passes: first scan for a queue with the least tasks that
640 	 * does not already service this uniq id.  If that fails simply find
641 	 * the queue with the least total tasks.
642 	 */
643 	for (idx = -1, mincnt = INT_MAX, strict = 1; mincnt == INT_MAX;
644 	    strict = 0) {
645 		for (i = 0; i < qgroup->tqg_cnt; i++) {
646 			if (qgroup->tqg_queue[i].tgc_cnt > mincnt)
647 				continue;
648 			if (strict) {
649 				LIST_FOREACH(n, &qgroup->tqg_queue[i].tgc_tasks,
650 				    gt_list)
651 					if (n->gt_uniq == uniq)
652 						break;
653 				if (n != NULL)
654 					continue;
655 			}
656 			mincnt = qgroup->tqg_queue[i].tgc_cnt;
657 			idx = i;
658 		}
659 	}
660 	if (idx == -1)
661 		panic("%s: failed to pick a qid.", __func__);
662 
663 	return (idx);
664 }
665 
666 void
667 taskqgroup_attach(struct taskqgroup *qgroup, struct grouptask *gtask,
668     void *uniq, device_t dev, struct resource *irq, const char *name)
669 {
670 	int cpu, qid, error;
671 
672 	KASSERT(qgroup->tqg_cnt > 0,
673 	    ("qgroup %s has no queues", qgroup->tqg_name));
674 
675 	gtask->gt_uniq = uniq;
676 	snprintf(gtask->gt_name, GROUPTASK_NAMELEN, "%s", name ? name : "grouptask");
677 	gtask->gt_dev = dev;
678 	gtask->gt_irq = irq;
679 	gtask->gt_cpu = -1;
680 	mtx_lock(&qgroup->tqg_lock);
681 	qid = taskqgroup_find(qgroup, uniq);
682 	qgroup->tqg_queue[qid].tgc_cnt++;
683 	LIST_INSERT_HEAD(&qgroup->tqg_queue[qid].tgc_tasks, gtask, gt_list);
684 	gtask->gt_taskqueue = qgroup->tqg_queue[qid].tgc_taskq;
685 	if (dev != NULL && irq != NULL) {
686 		cpu = qgroup->tqg_queue[qid].tgc_cpu;
687 		gtask->gt_cpu = cpu;
688 		mtx_unlock(&qgroup->tqg_lock);
689 		error = bus_bind_intr(dev, irq, cpu);
690 		if (error)
691 			printf("%s: binding interrupt failed for %s: %d\n",
692 			    __func__, gtask->gt_name, error);
693 	} else
694 		mtx_unlock(&qgroup->tqg_lock);
695 }
696 
697 int
698 taskqgroup_attach_cpu(struct taskqgroup *qgroup, struct grouptask *gtask,
699     void *uniq, int cpu, device_t dev, struct resource *irq, const char *name)
700 {
701 	int i, qid, error;
702 
703 	gtask->gt_uniq = uniq;
704 	snprintf(gtask->gt_name, GROUPTASK_NAMELEN, "%s", name ? name : "grouptask");
705 	gtask->gt_dev = dev;
706 	gtask->gt_irq = irq;
707 	gtask->gt_cpu = cpu;
708 	mtx_lock(&qgroup->tqg_lock);
709 	for (i = 0, qid = -1; i < qgroup->tqg_cnt; i++)
710 		if (qgroup->tqg_queue[i].tgc_cpu == cpu) {
711 			qid = i;
712 			break;
713 		}
714 	if (qid == -1) {
715 		mtx_unlock(&qgroup->tqg_lock);
716 		printf("%s: qid not found for %s cpu=%d\n", __func__, gtask->gt_name, cpu);
717 		return (EINVAL);
718 	}
719 	qgroup->tqg_queue[qid].tgc_cnt++;
720 	LIST_INSERT_HEAD(&qgroup->tqg_queue[qid].tgc_tasks, gtask, gt_list);
721 	gtask->gt_taskqueue = qgroup->tqg_queue[qid].tgc_taskq;
722 	cpu = qgroup->tqg_queue[qid].tgc_cpu;
723 	mtx_unlock(&qgroup->tqg_lock);
724 
725 	if (dev != NULL && irq != NULL) {
726 		error = bus_bind_intr(dev, irq, cpu);
727 		if (error)
728 			printf("%s: binding interrupt failed for %s: %d\n",
729 			    __func__, gtask->gt_name, error);
730 	}
731 	return (0);
732 }
733 
734 void
735 taskqgroup_detach(struct taskqgroup *qgroup, struct grouptask *gtask)
736 {
737 	int i;
738 
739 	grouptask_block(gtask);
740 	mtx_lock(&qgroup->tqg_lock);
741 	for (i = 0; i < qgroup->tqg_cnt; i++)
742 		if (qgroup->tqg_queue[i].tgc_taskq == gtask->gt_taskqueue)
743 			break;
744 	if (i == qgroup->tqg_cnt)
745 		panic("%s: task %s not in group", __func__, gtask->gt_name);
746 	qgroup->tqg_queue[i].tgc_cnt--;
747 	LIST_REMOVE(gtask, gt_list);
748 	mtx_unlock(&qgroup->tqg_lock);
749 	gtask->gt_taskqueue = NULL;
750 	gtask->gt_task.ta_flags &= ~TASK_NOENQUEUE;
751 }
752 
753 static void
754 taskqgroup_binder(void *ctx)
755 {
756 	struct taskq_bind_task *gtask;
757 	cpuset_t mask;
758 	int error;
759 
760 	gtask = ctx;
761 	CPU_ZERO(&mask);
762 	CPU_SET(gtask->bt_cpuid, &mask);
763 	error = cpuset_setthread(curthread->td_tid, &mask);
764 	thread_lock(curthread);
765 	sched_bind(curthread, gtask->bt_cpuid);
766 	thread_unlock(curthread);
767 
768 	if (error)
769 		printf("%s: binding curthread failed: %d\n", __func__, error);
770 	free(gtask, M_DEVBUF);
771 }
772 
773 void
774 taskqgroup_bind(struct taskqgroup *qgroup)
775 {
776 	struct taskq_bind_task *gtask;
777 	int i;
778 
779 	/*
780 	 * Bind taskqueue threads to specific CPUs, if they have been assigned
781 	 * one.
782 	 */
783 	if (qgroup->tqg_cnt == 1)
784 		return;
785 
786 	for (i = 0; i < qgroup->tqg_cnt; i++) {
787 		gtask = malloc(sizeof(*gtask), M_DEVBUF, M_WAITOK);
788 		GTASK_INIT(&gtask->bt_task, 0, 0, taskqgroup_binder, gtask);
789 		gtask->bt_cpuid = qgroup->tqg_queue[i].tgc_cpu;
790 		grouptaskqueue_enqueue(qgroup->tqg_queue[i].tgc_taskq,
791 		    &gtask->bt_task);
792 	}
793 }
794 
795 struct taskqgroup *
796 taskqgroup_create(const char *name, int cnt, int stride)
797 {
798 	struct taskqgroup *qgroup;
799 	int cpu, i, j;
800 
801 	qgroup = malloc(sizeof(*qgroup), M_GTASKQUEUE, M_WAITOK | M_ZERO);
802 	mtx_init(&qgroup->tqg_lock, "taskqgroup", NULL, MTX_DEF);
803 	qgroup->tqg_name = name;
804 	qgroup->tqg_cnt = cnt;
805 
806 	for (cpu = i = 0; i < cnt; i++) {
807 		taskqgroup_cpu_create(qgroup, i, cpu);
808 		for (j = 0; j < stride; j++)
809 			cpu = CPU_NEXT(cpu);
810 	}
811 	return (qgroup);
812 }
813 
814 void
815 taskqgroup_destroy(struct taskqgroup *qgroup)
816 {
817 }
818 
819 void
820 taskqgroup_drain_all(struct taskqgroup *tqg)
821 {
822 	struct gtaskqueue *q;
823 
824 	for (int i = 0; i < mp_ncpus; i++) {
825 		q = tqg->tqg_queue[i].tgc_taskq;
826 		if (q == NULL)
827 			continue;
828 		gtaskqueue_drain_all(q);
829 	}
830 }
831