xref: /freebsd/sys/kern/subr_gtaskqueue.c (revision 4d65a7c6951cea0333f1a0c1b32c38489cdfa6c5)
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/param.h>
30 #include <sys/systm.h>
31 #include <sys/bus.h>
32 #include <sys/cpuset.h>
33 #include <sys/kernel.h>
34 #include <sys/kthread.h>
35 #include <sys/libkern.h>
36 #include <sys/limits.h>
37 #include <sys/lock.h>
38 #include <sys/malloc.h>
39 #include <sys/mutex.h>
40 #include <sys/proc.h>
41 #include <sys/epoch.h>
42 #include <sys/sched.h>
43 #include <sys/smp.h>
44 #include <sys/gtaskqueue.h>
45 #include <sys/unistd.h>
46 #include <machine/stdarg.h>
47 
48 static MALLOC_DEFINE(M_GTASKQUEUE, "gtaskqueue", "Group Task Queues");
49 static void	gtaskqueue_thread_enqueue(void *);
50 static void	gtaskqueue_thread_loop(void *arg);
51 static int	task_is_running(struct gtaskqueue *queue, struct gtask *gtask);
52 static void	gtaskqueue_drain_locked(struct gtaskqueue *queue, struct gtask *gtask);
53 
54 TASKQGROUP_DEFINE(softirq, mp_ncpus, 1);
55 
56 struct gtaskqueue_busy {
57 	struct gtask		*tb_running;
58 	u_int			 tb_seq;
59 	LIST_ENTRY(gtaskqueue_busy) tb_link;
60 };
61 
62 typedef void (*gtaskqueue_enqueue_fn)(void *context);
63 
64 struct gtaskqueue {
65 	STAILQ_HEAD(, gtask)	tq_queue;
66 	LIST_HEAD(, gtaskqueue_busy) tq_active;
67 	u_int			tq_seq;
68 	int			tq_callouts;
69 	struct mtx_padalign	tq_mutex;
70 	gtaskqueue_enqueue_fn	tq_enqueue;
71 	void			*tq_context;
72 	char			*tq_name;
73 	struct thread		**tq_threads;
74 	int			tq_tcount;
75 	int			tq_spin;
76 	int			tq_flags;
77 	taskqueue_callback_fn	tq_callbacks[TASKQUEUE_NUM_CALLBACKS];
78 	void			*tq_cb_contexts[TASKQUEUE_NUM_CALLBACKS];
79 };
80 
81 #define	TQ_FLAGS_ACTIVE		(1 << 0)
82 #define	TQ_FLAGS_BLOCKED	(1 << 1)
83 #define	TQ_FLAGS_UNLOCKED_ENQUEUE	(1 << 2)
84 
85 #define	DT_CALLOUT_ARMED	(1 << 0)
86 
87 #define	TQ_LOCK(tq)							\
88 	do {								\
89 		if ((tq)->tq_spin)					\
90 			mtx_lock_spin(&(tq)->tq_mutex);			\
91 		else							\
92 			mtx_lock(&(tq)->tq_mutex);			\
93 	} while (0)
94 #define	TQ_ASSERT_LOCKED(tq)	mtx_assert(&(tq)->tq_mutex, MA_OWNED)
95 
96 #define	TQ_UNLOCK(tq)							\
97 	do {								\
98 		if ((tq)->tq_spin)					\
99 			mtx_unlock_spin(&(tq)->tq_mutex);		\
100 		else							\
101 			mtx_unlock(&(tq)->tq_mutex);			\
102 	} while (0)
103 #define	TQ_ASSERT_UNLOCKED(tq)	mtx_assert(&(tq)->tq_mutex, MA_NOTOWNED)
104 
105 #ifdef INVARIANTS
106 static void
107 gtask_dump(struct gtask *gtask)
108 {
109 	printf("gtask: %p ta_flags=%x ta_priority=%d ta_func=%p ta_context=%p\n",
110 	       gtask, gtask->ta_flags, gtask->ta_priority, gtask->ta_func, gtask->ta_context);
111 }
112 #endif
113 
114 static __inline int
115 TQ_SLEEP(struct gtaskqueue *tq, void *p, const char *wm)
116 {
117 	if (tq->tq_spin)
118 		return (msleep_spin(p, (struct mtx *)&tq->tq_mutex, wm, 0));
119 	return (msleep(p, &tq->tq_mutex, 0, wm, 0));
120 }
121 
122 static struct gtaskqueue *
123 _gtaskqueue_create(const char *name, int mflags,
124 		 taskqueue_enqueue_fn enqueue, void *context,
125 		 int mtxflags, const char *mtxname __unused)
126 {
127 	struct gtaskqueue *queue;
128 	char *tq_name;
129 
130 	tq_name = malloc(TASKQUEUE_NAMELEN, M_GTASKQUEUE, mflags | M_ZERO);
131 	if (!tq_name)
132 		return (NULL);
133 
134 	snprintf(tq_name, TASKQUEUE_NAMELEN, "%s", (name) ? name : "taskqueue");
135 
136 	queue = malloc(sizeof(struct gtaskqueue), M_GTASKQUEUE, mflags | M_ZERO);
137 	if (!queue) {
138 		free(tq_name, M_GTASKQUEUE);
139 		return (NULL);
140 	}
141 
142 	STAILQ_INIT(&queue->tq_queue);
143 	LIST_INIT(&queue->tq_active);
144 	queue->tq_enqueue = enqueue;
145 	queue->tq_context = context;
146 	queue->tq_name = tq_name;
147 	queue->tq_spin = (mtxflags & MTX_SPIN) != 0;
148 	queue->tq_flags |= TQ_FLAGS_ACTIVE;
149 	if (enqueue == gtaskqueue_thread_enqueue)
150 		queue->tq_flags |= TQ_FLAGS_UNLOCKED_ENQUEUE;
151 	mtx_init(&queue->tq_mutex, tq_name, NULL, mtxflags);
152 
153 	return (queue);
154 }
155 
156 /*
157  * Signal a taskqueue thread to terminate.
158  */
159 static void
160 gtaskqueue_terminate(struct thread **pp, struct gtaskqueue *tq)
161 {
162 
163 	while (tq->tq_tcount > 0 || tq->tq_callouts > 0) {
164 		wakeup(tq);
165 		TQ_SLEEP(tq, pp, "gtq_destroy");
166 	}
167 }
168 
169 static void __unused
170 gtaskqueue_free(struct gtaskqueue *queue)
171 {
172 
173 	TQ_LOCK(queue);
174 	queue->tq_flags &= ~TQ_FLAGS_ACTIVE;
175 	gtaskqueue_terminate(queue->tq_threads, queue);
176 	KASSERT(LIST_EMPTY(&queue->tq_active), ("Tasks still running?"));
177 	KASSERT(queue->tq_callouts == 0, ("Armed timeout tasks"));
178 	mtx_destroy(&queue->tq_mutex);
179 	free(queue->tq_threads, M_GTASKQUEUE);
180 	free(queue->tq_name, M_GTASKQUEUE);
181 	free(queue, M_GTASKQUEUE);
182 }
183 
184 /*
185  * Wait for all to complete, then prevent it from being enqueued
186  */
187 void
188 grouptask_block(struct grouptask *grouptask)
189 {
190 	struct gtaskqueue *queue = grouptask->gt_taskqueue;
191 	struct gtask *gtask = &grouptask->gt_task;
192 
193 #ifdef INVARIANTS
194 	if (queue == NULL) {
195 		gtask_dump(gtask);
196 		panic("queue == NULL");
197 	}
198 #endif
199 	TQ_LOCK(queue);
200 	gtask->ta_flags |= TASK_NOENQUEUE;
201   	gtaskqueue_drain_locked(queue, gtask);
202 	TQ_UNLOCK(queue);
203 }
204 
205 void
206 grouptask_unblock(struct grouptask *grouptask)
207 {
208 	struct gtaskqueue *queue = grouptask->gt_taskqueue;
209 	struct gtask *gtask = &grouptask->gt_task;
210 
211 #ifdef INVARIANTS
212 	if (queue == NULL) {
213 		gtask_dump(gtask);
214 		panic("queue == NULL");
215 	}
216 #endif
217 	TQ_LOCK(queue);
218 	gtask->ta_flags &= ~TASK_NOENQUEUE;
219 	TQ_UNLOCK(queue);
220 }
221 
222 int
223 grouptaskqueue_enqueue(struct gtaskqueue *queue, struct gtask *gtask)
224 {
225 #ifdef INVARIANTS
226 	if (queue == NULL) {
227 		gtask_dump(gtask);
228 		panic("queue == NULL");
229 	}
230 #endif
231 	TQ_LOCK(queue);
232 	if (gtask->ta_flags & TASK_ENQUEUED) {
233 		TQ_UNLOCK(queue);
234 		return (0);
235 	}
236 	if (gtask->ta_flags & TASK_NOENQUEUE) {
237 		TQ_UNLOCK(queue);
238 		return (EAGAIN);
239 	}
240 	STAILQ_INSERT_TAIL(&queue->tq_queue, gtask, ta_link);
241 	gtask->ta_flags |= TASK_ENQUEUED;
242 	TQ_UNLOCK(queue);
243 	if ((queue->tq_flags & TQ_FLAGS_BLOCKED) == 0)
244 		queue->tq_enqueue(queue->tq_context);
245 	return (0);
246 }
247 
248 static void
249 gtaskqueue_task_nop_fn(void *context)
250 {
251 }
252 
253 /*
254  * Block until all currently queued tasks in this taskqueue
255  * have begun execution.  Tasks queued during execution of
256  * this function are ignored.
257  */
258 static void
259 gtaskqueue_drain_tq_queue(struct gtaskqueue *queue)
260 {
261 	struct gtask t_barrier;
262 
263 	if (STAILQ_EMPTY(&queue->tq_queue))
264 		return;
265 
266 	/*
267 	 * Enqueue our barrier after all current tasks, but with
268 	 * the highest priority so that newly queued tasks cannot
269 	 * pass it.  Because of the high priority, we can not use
270 	 * taskqueue_enqueue_locked directly (which drops the lock
271 	 * anyway) so just insert it at tail while we have the
272 	 * queue lock.
273 	 */
274 	GTASK_INIT(&t_barrier, 0, USHRT_MAX, gtaskqueue_task_nop_fn, &t_barrier);
275 	STAILQ_INSERT_TAIL(&queue->tq_queue, &t_barrier, ta_link);
276 	t_barrier.ta_flags |= TASK_ENQUEUED;
277 
278 	/*
279 	 * Once the barrier has executed, all previously queued tasks
280 	 * have completed or are currently executing.
281 	 */
282 	while (t_barrier.ta_flags & TASK_ENQUEUED)
283 		TQ_SLEEP(queue, &t_barrier, "gtq_qdrain");
284 }
285 
286 /*
287  * Block until all currently executing tasks for this taskqueue
288  * complete.  Tasks that begin execution during the execution
289  * of this function are ignored.
290  */
291 static void
292 gtaskqueue_drain_tq_active(struct gtaskqueue *queue)
293 {
294 	struct gtaskqueue_busy *tb;
295 	u_int seq;
296 
297 	if (LIST_EMPTY(&queue->tq_active))
298 		return;
299 
300 	/* Block taskq_terminate().*/
301 	queue->tq_callouts++;
302 
303 	/* Wait for any active task with sequence from the past. */
304 	seq = queue->tq_seq;
305 restart:
306 	LIST_FOREACH(tb, &queue->tq_active, tb_link) {
307 		if ((int)(tb->tb_seq - seq) <= 0) {
308 			TQ_SLEEP(queue, tb->tb_running, "gtq_adrain");
309 			goto restart;
310 		}
311 	}
312 
313 	/* Release taskqueue_terminate(). */
314 	queue->tq_callouts--;
315 	if ((queue->tq_flags & TQ_FLAGS_ACTIVE) == 0)
316 		wakeup_one(queue->tq_threads);
317 }
318 
319 void
320 gtaskqueue_block(struct gtaskqueue *queue)
321 {
322 
323 	TQ_LOCK(queue);
324 	queue->tq_flags |= TQ_FLAGS_BLOCKED;
325 	TQ_UNLOCK(queue);
326 }
327 
328 void
329 gtaskqueue_unblock(struct gtaskqueue *queue)
330 {
331 
332 	TQ_LOCK(queue);
333 	queue->tq_flags &= ~TQ_FLAGS_BLOCKED;
334 	if (!STAILQ_EMPTY(&queue->tq_queue))
335 		queue->tq_enqueue(queue->tq_context);
336 	TQ_UNLOCK(queue);
337 }
338 
339 static void
340 gtaskqueue_run_locked(struct gtaskqueue *queue)
341 {
342 	struct epoch_tracker et;
343 	struct gtaskqueue_busy tb;
344 	struct gtask *gtask;
345 	bool in_net_epoch;
346 
347 	KASSERT(queue != NULL, ("tq is NULL"));
348 	TQ_ASSERT_LOCKED(queue);
349 	tb.tb_running = NULL;
350 	LIST_INSERT_HEAD(&queue->tq_active, &tb, tb_link);
351 	in_net_epoch = false;
352 
353 	while ((gtask = STAILQ_FIRST(&queue->tq_queue)) != NULL) {
354 		STAILQ_REMOVE_HEAD(&queue->tq_queue, ta_link);
355 		gtask->ta_flags &= ~TASK_ENQUEUED;
356 		tb.tb_running = gtask;
357 		tb.tb_seq = ++queue->tq_seq;
358 		TQ_UNLOCK(queue);
359 
360 		KASSERT(gtask->ta_func != NULL, ("task->ta_func is NULL"));
361 		if (!in_net_epoch && TASK_IS_NET(gtask)) {
362 			in_net_epoch = true;
363 			NET_EPOCH_ENTER(et);
364 		} else if (in_net_epoch && !TASK_IS_NET(gtask)) {
365 			NET_EPOCH_EXIT(et);
366 			in_net_epoch = false;
367 		}
368 		gtask->ta_func(gtask->ta_context);
369 
370 		TQ_LOCK(queue);
371 		wakeup(gtask);
372 	}
373 	if (in_net_epoch)
374 		NET_EPOCH_EXIT(et);
375 	LIST_REMOVE(&tb, tb_link);
376 }
377 
378 static int
379 task_is_running(struct gtaskqueue *queue, struct gtask *gtask)
380 {
381 	struct gtaskqueue_busy *tb;
382 
383 	TQ_ASSERT_LOCKED(queue);
384 	LIST_FOREACH(tb, &queue->tq_active, tb_link) {
385 		if (tb->tb_running == gtask)
386 			return (1);
387 	}
388 	return (0);
389 }
390 
391 static int
392 gtaskqueue_cancel_locked(struct gtaskqueue *queue, struct gtask *gtask)
393 {
394 
395 	if (gtask->ta_flags & TASK_ENQUEUED)
396 		STAILQ_REMOVE(&queue->tq_queue, gtask, gtask, ta_link);
397 	gtask->ta_flags &= ~TASK_ENQUEUED;
398 	return (task_is_running(queue, gtask) ? EBUSY : 0);
399 }
400 
401 int
402 gtaskqueue_cancel(struct gtaskqueue *queue, struct gtask *gtask)
403 {
404 	int error;
405 
406 	TQ_LOCK(queue);
407 	error = gtaskqueue_cancel_locked(queue, gtask);
408 	TQ_UNLOCK(queue);
409 
410 	return (error);
411 }
412 
413 static void
414 gtaskqueue_drain_locked(struct gtaskqueue *queue, struct gtask *gtask)
415 {
416 	while ((gtask->ta_flags & TASK_ENQUEUED) || task_is_running(queue, gtask))
417 		TQ_SLEEP(queue, gtask, "gtq_drain");
418 }
419 
420 void
421 gtaskqueue_drain(struct gtaskqueue *queue, struct gtask *gtask)
422 {
423 
424 	if (!queue->tq_spin)
425 		WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, __func__);
426 
427 	TQ_LOCK(queue);
428 	gtaskqueue_drain_locked(queue, gtask);
429 	TQ_UNLOCK(queue);
430 }
431 
432 void
433 gtaskqueue_drain_all(struct gtaskqueue *queue)
434 {
435 
436 	if (!queue->tq_spin)
437 		WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, __func__);
438 
439 	TQ_LOCK(queue);
440 	gtaskqueue_drain_tq_queue(queue);
441 	gtaskqueue_drain_tq_active(queue);
442 	TQ_UNLOCK(queue);
443 }
444 
445 static int
446 _gtaskqueue_start_threads(struct gtaskqueue **tqp, int count, int pri,
447     cpuset_t *mask, const char *name, va_list ap)
448 {
449 	char ktname[MAXCOMLEN + 1];
450 	struct thread *td;
451 	struct gtaskqueue *tq;
452 	int i, error;
453 
454 	if (count <= 0)
455 		return (EINVAL);
456 
457 	vsnprintf(ktname, sizeof(ktname), name, ap);
458 	tq = *tqp;
459 
460 	tq->tq_threads = malloc(sizeof(struct thread *) * count, M_GTASKQUEUE,
461 	    M_NOWAIT | M_ZERO);
462 	if (tq->tq_threads == NULL) {
463 		printf("%s: no memory for %s threads\n", __func__, ktname);
464 		return (ENOMEM);
465 	}
466 
467 	for (i = 0; i < count; i++) {
468 		if (count == 1)
469 			error = kthread_add(gtaskqueue_thread_loop, tqp, NULL,
470 			    &tq->tq_threads[i], RFSTOPPED, 0, "%s", ktname);
471 		else
472 			error = kthread_add(gtaskqueue_thread_loop, tqp, NULL,
473 			    &tq->tq_threads[i], RFSTOPPED, 0,
474 			    "%s_%d", ktname, i);
475 		if (error) {
476 			/* should be ok to continue, taskqueue_free will dtrt */
477 			printf("%s: kthread_add(%s): error %d", __func__,
478 			    ktname, error);
479 			tq->tq_threads[i] = NULL;		/* paranoid */
480 		} else
481 			tq->tq_tcount++;
482 	}
483 	for (i = 0; i < count; i++) {
484 		if (tq->tq_threads[i] == NULL)
485 			continue;
486 		td = tq->tq_threads[i];
487 		if (mask) {
488 			error = cpuset_setthread(td->td_tid, mask);
489 			/*
490 			 * Failing to pin is rarely an actual fatal error;
491 			 * it'll just affect performance.
492 			 */
493 			if (error)
494 				printf("%s: curthread=%llu: can't pin; "
495 				    "error=%d\n",
496 				    __func__,
497 				    (unsigned long long) td->td_tid,
498 				    error);
499 		}
500 		thread_lock(td);
501 		sched_prio(td, pri);
502 		sched_add(td, SRQ_BORING);
503 	}
504 
505 	return (0);
506 }
507 
508 static int
509 gtaskqueue_start_threads(struct gtaskqueue **tqp, int count, int pri,
510     const char *name, ...)
511 {
512 	va_list ap;
513 	int error;
514 
515 	va_start(ap, name);
516 	error = _gtaskqueue_start_threads(tqp, count, pri, NULL, name, ap);
517 	va_end(ap);
518 	return (error);
519 }
520 
521 static inline void
522 gtaskqueue_run_callback(struct gtaskqueue *tq,
523     enum taskqueue_callback_type cb_type)
524 {
525 	taskqueue_callback_fn tq_callback;
526 
527 	TQ_ASSERT_UNLOCKED(tq);
528 	tq_callback = tq->tq_callbacks[cb_type];
529 	if (tq_callback != NULL)
530 		tq_callback(tq->tq_cb_contexts[cb_type]);
531 }
532 
533 static void
534 gtaskqueue_thread_loop(void *arg)
535 {
536 	struct gtaskqueue **tqp, *tq;
537 
538 	tqp = arg;
539 	tq = *tqp;
540 	gtaskqueue_run_callback(tq, TASKQUEUE_CALLBACK_TYPE_INIT);
541 	TQ_LOCK(tq);
542 	while ((tq->tq_flags & TQ_FLAGS_ACTIVE) != 0) {
543 		/* XXX ? */
544 		gtaskqueue_run_locked(tq);
545 		/*
546 		 * Because taskqueue_run() can drop tq_mutex, we need to
547 		 * check if the TQ_FLAGS_ACTIVE flag wasn't removed in the
548 		 * meantime, which means we missed a wakeup.
549 		 */
550 		if ((tq->tq_flags & TQ_FLAGS_ACTIVE) == 0)
551 			break;
552 		TQ_SLEEP(tq, tq, "-");
553 	}
554 	gtaskqueue_run_locked(tq);
555 	/*
556 	 * This thread is on its way out, so just drop the lock temporarily
557 	 * in order to call the shutdown callback.  This allows the callback
558 	 * to look at the taskqueue, even just before it dies.
559 	 */
560 	TQ_UNLOCK(tq);
561 	gtaskqueue_run_callback(tq, TASKQUEUE_CALLBACK_TYPE_SHUTDOWN);
562 	TQ_LOCK(tq);
563 
564 	/* rendezvous with thread that asked us to terminate */
565 	tq->tq_tcount--;
566 	wakeup_one(tq->tq_threads);
567 	TQ_UNLOCK(tq);
568 	kthread_exit();
569 }
570 
571 static void
572 gtaskqueue_thread_enqueue(void *context)
573 {
574 	struct gtaskqueue **tqp, *tq;
575 
576 	tqp = context;
577 	tq = *tqp;
578 	wakeup_any(tq);
579 }
580 
581 static struct gtaskqueue *
582 gtaskqueue_create_fast(const char *name, int mflags,
583 		 taskqueue_enqueue_fn enqueue, void *context)
584 {
585 	return _gtaskqueue_create(name, mflags, enqueue, context,
586 			MTX_SPIN, "fast_taskqueue");
587 }
588 
589 struct taskqgroup_cpu {
590 	LIST_HEAD(, grouptask) tgc_tasks;
591 	struct gtaskqueue *tgc_taskq;
592 	int		tgc_cnt;
593 	int		tgc_cpu;
594 };
595 
596 struct taskqgroup {
597 	struct taskqgroup_cpu tqg_queue[MAXCPU];
598 	struct mtx	tqg_lock;
599 	const char *	tqg_name;
600 	int		tqg_cnt;
601 };
602 
603 struct taskq_bind_task {
604 	struct gtask bt_task;
605 	int	bt_cpuid;
606 };
607 
608 static void
609 taskqgroup_cpu_create(struct taskqgroup *qgroup, int idx, int cpu)
610 {
611 	struct taskqgroup_cpu *qcpu;
612 
613 	qcpu = &qgroup->tqg_queue[idx];
614 	LIST_INIT(&qcpu->tgc_tasks);
615 	qcpu->tgc_taskq = gtaskqueue_create_fast(NULL, M_WAITOK,
616 	    taskqueue_thread_enqueue, &qcpu->tgc_taskq);
617 	gtaskqueue_start_threads(&qcpu->tgc_taskq, 1, PI_SOFT,
618 	    "%s_%d", qgroup->tqg_name, idx);
619 	qcpu->tgc_cpu = cpu;
620 }
621 
622 /*
623  * Find the taskq with least # of tasks that doesn't currently have any
624  * other queues from the uniq identifier.
625  */
626 static int
627 taskqgroup_find(struct taskqgroup *qgroup, void *uniq)
628 {
629 	struct grouptask *n;
630 	int i, idx, mincnt;
631 	int strict;
632 
633 	mtx_assert(&qgroup->tqg_lock, MA_OWNED);
634 	KASSERT(qgroup->tqg_cnt != 0,
635 	    ("qgroup %s has no queues", qgroup->tqg_name));
636 
637 	/*
638 	 * Two passes: first scan for a queue with the least tasks that
639 	 * does not already service this uniq id.  If that fails simply find
640 	 * the queue with the least total tasks.
641 	 */
642 	for (idx = -1, mincnt = INT_MAX, strict = 1; mincnt == INT_MAX;
643 	    strict = 0) {
644 		for (i = 0; i < qgroup->tqg_cnt; i++) {
645 			if (qgroup->tqg_queue[i].tgc_cnt > mincnt)
646 				continue;
647 			if (strict) {
648 				LIST_FOREACH(n, &qgroup->tqg_queue[i].tgc_tasks,
649 				    gt_list)
650 					if (n->gt_uniq == uniq)
651 						break;
652 				if (n != NULL)
653 					continue;
654 			}
655 			mincnt = qgroup->tqg_queue[i].tgc_cnt;
656 			idx = i;
657 		}
658 	}
659 	if (idx == -1)
660 		panic("%s: failed to pick a qid.", __func__);
661 
662 	return (idx);
663 }
664 
665 void
666 taskqgroup_attach(struct taskqgroup *qgroup, struct grouptask *gtask,
667     void *uniq, device_t dev, struct resource *irq, const char *name)
668 {
669 	int cpu, qid, error;
670 
671 	KASSERT(qgroup->tqg_cnt > 0,
672 	    ("qgroup %s has no queues", qgroup->tqg_name));
673 
674 	gtask->gt_uniq = uniq;
675 	snprintf(gtask->gt_name, GROUPTASK_NAMELEN, "%s", name ? name : "grouptask");
676 	gtask->gt_dev = dev;
677 	gtask->gt_irq = irq;
678 	gtask->gt_cpu = -1;
679 	mtx_lock(&qgroup->tqg_lock);
680 	qid = taskqgroup_find(qgroup, uniq);
681 	qgroup->tqg_queue[qid].tgc_cnt++;
682 	LIST_INSERT_HEAD(&qgroup->tqg_queue[qid].tgc_tasks, gtask, gt_list);
683 	gtask->gt_taskqueue = qgroup->tqg_queue[qid].tgc_taskq;
684 	if (dev != NULL && irq != NULL) {
685 		cpu = qgroup->tqg_queue[qid].tgc_cpu;
686 		gtask->gt_cpu = cpu;
687 		mtx_unlock(&qgroup->tqg_lock);
688 		error = bus_bind_intr(dev, irq, cpu);
689 		if (error)
690 			printf("%s: binding interrupt failed for %s: %d\n",
691 			    __func__, gtask->gt_name, error);
692 	} else
693 		mtx_unlock(&qgroup->tqg_lock);
694 }
695 
696 int
697 taskqgroup_attach_cpu(struct taskqgroup *qgroup, struct grouptask *gtask,
698     void *uniq, int cpu, device_t dev, struct resource *irq, const char *name)
699 {
700 	int i, qid, error;
701 
702 	gtask->gt_uniq = uniq;
703 	snprintf(gtask->gt_name, GROUPTASK_NAMELEN, "%s", name ? name : "grouptask");
704 	gtask->gt_dev = dev;
705 	gtask->gt_irq = irq;
706 	gtask->gt_cpu = cpu;
707 	mtx_lock(&qgroup->tqg_lock);
708 	for (i = 0, qid = -1; i < qgroup->tqg_cnt; i++)
709 		if (qgroup->tqg_queue[i].tgc_cpu == cpu) {
710 			qid = i;
711 			break;
712 		}
713 	if (qid == -1) {
714 		mtx_unlock(&qgroup->tqg_lock);
715 		printf("%s: qid not found for %s cpu=%d\n", __func__, gtask->gt_name, cpu);
716 		return (EINVAL);
717 	}
718 	qgroup->tqg_queue[qid].tgc_cnt++;
719 	LIST_INSERT_HEAD(&qgroup->tqg_queue[qid].tgc_tasks, gtask, gt_list);
720 	gtask->gt_taskqueue = qgroup->tqg_queue[qid].tgc_taskq;
721 	cpu = qgroup->tqg_queue[qid].tgc_cpu;
722 	mtx_unlock(&qgroup->tqg_lock);
723 
724 	if (dev != NULL && irq != NULL) {
725 		error = bus_bind_intr(dev, irq, cpu);
726 		if (error)
727 			printf("%s: binding interrupt failed for %s: %d\n",
728 			    __func__, gtask->gt_name, error);
729 	}
730 	return (0);
731 }
732 
733 void
734 taskqgroup_detach(struct taskqgroup *qgroup, struct grouptask *gtask)
735 {
736 	int i;
737 
738 	grouptask_block(gtask);
739 	mtx_lock(&qgroup->tqg_lock);
740 	for (i = 0; i < qgroup->tqg_cnt; i++)
741 		if (qgroup->tqg_queue[i].tgc_taskq == gtask->gt_taskqueue)
742 			break;
743 	if (i == qgroup->tqg_cnt)
744 		panic("%s: task %s not in group", __func__, gtask->gt_name);
745 	qgroup->tqg_queue[i].tgc_cnt--;
746 	LIST_REMOVE(gtask, gt_list);
747 	mtx_unlock(&qgroup->tqg_lock);
748 	gtask->gt_taskqueue = NULL;
749 	gtask->gt_task.ta_flags &= ~TASK_NOENQUEUE;
750 }
751 
752 static void
753 taskqgroup_binder(void *ctx)
754 {
755 	struct taskq_bind_task *gtask;
756 	cpuset_t mask;
757 	int error;
758 
759 	gtask = ctx;
760 	CPU_ZERO(&mask);
761 	CPU_SET(gtask->bt_cpuid, &mask);
762 	error = cpuset_setthread(curthread->td_tid, &mask);
763 	thread_lock(curthread);
764 	sched_bind(curthread, gtask->bt_cpuid);
765 	thread_unlock(curthread);
766 
767 	if (error)
768 		printf("%s: binding curthread failed: %d\n", __func__, error);
769 	free(gtask, M_DEVBUF);
770 }
771 
772 void
773 taskqgroup_bind(struct taskqgroup *qgroup)
774 {
775 	struct taskq_bind_task *gtask;
776 	int i;
777 
778 	/*
779 	 * Bind taskqueue threads to specific CPUs, if they have been assigned
780 	 * one.
781 	 */
782 	if (qgroup->tqg_cnt == 1)
783 		return;
784 
785 	for (i = 0; i < qgroup->tqg_cnt; i++) {
786 		gtask = malloc(sizeof(*gtask), M_DEVBUF, M_WAITOK);
787 		GTASK_INIT(&gtask->bt_task, 0, 0, taskqgroup_binder, gtask);
788 		gtask->bt_cpuid = qgroup->tqg_queue[i].tgc_cpu;
789 		grouptaskqueue_enqueue(qgroup->tqg_queue[i].tgc_taskq,
790 		    &gtask->bt_task);
791 	}
792 }
793 
794 struct taskqgroup *
795 taskqgroup_create(const char *name, int cnt, int stride)
796 {
797 	struct taskqgroup *qgroup;
798 	int cpu, i, j;
799 
800 	qgroup = malloc(sizeof(*qgroup), M_GTASKQUEUE, M_WAITOK | M_ZERO);
801 	mtx_init(&qgroup->tqg_lock, "taskqgroup", NULL, MTX_DEF);
802 	qgroup->tqg_name = name;
803 	qgroup->tqg_cnt = cnt;
804 
805 	for (cpu = i = 0; i < cnt; i++) {
806 		taskqgroup_cpu_create(qgroup, i, cpu);
807 		for (j = 0; j < stride; j++)
808 			cpu = CPU_NEXT(cpu);
809 	}
810 	return (qgroup);
811 }
812 
813 void
814 taskqgroup_destroy(struct taskqgroup *qgroup)
815 {
816 }
817 
818 void
819 taskqgroup_drain_all(struct taskqgroup *tqg)
820 {
821 	struct gtaskqueue *q;
822 
823 	for (int i = 0; i < mp_ncpus; i++) {
824 		q = tqg->tqg_queue[i].tgc_taskq;
825 		if (q == NULL)
826 			continue;
827 		gtaskqueue_drain_all(q);
828 	}
829 }
830