xref: /freebsd/sys/kern/subr_gtaskqueue.c (revision 3fc36ee018bb836bd1796067cf4ef8683f166ebc)
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 __FBSDID("$FreeBSD$");
31 
32 #include <sys/param.h>
33 #include <sys/systm.h>
34 #include <sys/bus.h>
35 #include <sys/cpuset.h>
36 #include <sys/interrupt.h>
37 #include <sys/kernel.h>
38 #include <sys/kthread.h>
39 #include <sys/libkern.h>
40 #include <sys/limits.h>
41 #include <sys/lock.h>
42 #include <sys/malloc.h>
43 #include <sys/mutex.h>
44 #include <sys/proc.h>
45 #include <sys/sched.h>
46 #include <sys/smp.h>
47 #include <sys/gtaskqueue.h>
48 #include <sys/unistd.h>
49 #include <machine/stdarg.h>
50 
51 static MALLOC_DEFINE(M_GTASKQUEUE, "taskqueue", "Task Queues");
52 static void	gtaskqueue_thread_enqueue(void *);
53 static void	gtaskqueue_thread_loop(void *arg);
54 
55 
56 struct gtaskqueue_busy {
57 	struct gtask	*tb_running;
58 	TAILQ_ENTRY(gtaskqueue_busy) tb_link;
59 };
60 
61 static struct gtask * const TB_DRAIN_WAITER = (struct gtask *)0x1;
62 
63 struct gtaskqueue {
64 	STAILQ_HEAD(, gtask)	tq_queue;
65 	gtaskqueue_enqueue_fn	tq_enqueue;
66 	void			*tq_context;
67 	char			*tq_name;
68 	TAILQ_HEAD(, gtaskqueue_busy) tq_active;
69 	struct mtx		tq_mutex;
70 	struct thread		**tq_threads;
71 	int			tq_tcount;
72 	int			tq_spin;
73 	int			tq_flags;
74 	int			tq_callouts;
75 	taskqueue_callback_fn	tq_callbacks[TASKQUEUE_NUM_CALLBACKS];
76 	void			*tq_cb_contexts[TASKQUEUE_NUM_CALLBACKS];
77 };
78 
79 #define	TQ_FLAGS_ACTIVE		(1 << 0)
80 #define	TQ_FLAGS_BLOCKED	(1 << 1)
81 #define	TQ_FLAGS_UNLOCKED_ENQUEUE	(1 << 2)
82 
83 #define	DT_CALLOUT_ARMED	(1 << 0)
84 
85 #define	TQ_LOCK(tq)							\
86 	do {								\
87 		if ((tq)->tq_spin)					\
88 			mtx_lock_spin(&(tq)->tq_mutex);			\
89 		else							\
90 			mtx_lock(&(tq)->tq_mutex);			\
91 	} while (0)
92 #define	TQ_ASSERT_LOCKED(tq)	mtx_assert(&(tq)->tq_mutex, MA_OWNED)
93 
94 #define	TQ_UNLOCK(tq)							\
95 	do {								\
96 		if ((tq)->tq_spin)					\
97 			mtx_unlock_spin(&(tq)->tq_mutex);		\
98 		else							\
99 			mtx_unlock(&(tq)->tq_mutex);			\
100 	} while (0)
101 #define	TQ_ASSERT_UNLOCKED(tq)	mtx_assert(&(tq)->tq_mutex, MA_NOTOWNED)
102 
103 static __inline int
104 TQ_SLEEP(struct gtaskqueue *tq, void *p, struct mtx *m, int pri, const char *wm,
105     int t)
106 {
107 	if (tq->tq_spin)
108 		return (msleep_spin(p, m, wm, t));
109 	return (msleep(p, m, pri, wm, t));
110 }
111 
112 static struct gtaskqueue *
113 _gtaskqueue_create(const char *name, int mflags,
114 		 taskqueue_enqueue_fn enqueue, void *context,
115 		 int mtxflags, const char *mtxname __unused)
116 {
117 	struct gtaskqueue *queue;
118 	char *tq_name;
119 
120 	tq_name = malloc(TASKQUEUE_NAMELEN, M_GTASKQUEUE, mflags | M_ZERO);
121 	if (!tq_name)
122 		return (NULL);
123 
124 	snprintf(tq_name, TASKQUEUE_NAMELEN, "%s", (name) ? name : "taskqueue");
125 
126 	queue = malloc(sizeof(struct gtaskqueue), M_GTASKQUEUE, mflags | M_ZERO);
127 	if (!queue)
128 		return (NULL);
129 
130 	STAILQ_INIT(&queue->tq_queue);
131 	TAILQ_INIT(&queue->tq_active);
132 	queue->tq_enqueue = enqueue;
133 	queue->tq_context = context;
134 	queue->tq_name = tq_name;
135 	queue->tq_spin = (mtxflags & MTX_SPIN) != 0;
136 	queue->tq_flags |= TQ_FLAGS_ACTIVE;
137 	if (enqueue == gtaskqueue_thread_enqueue)
138 		queue->tq_flags |= TQ_FLAGS_UNLOCKED_ENQUEUE;
139 	mtx_init(&queue->tq_mutex, tq_name, NULL, mtxflags);
140 
141 	return (queue);
142 }
143 
144 
145 /*
146  * Signal a taskqueue thread to terminate.
147  */
148 static void
149 gtaskqueue_terminate(struct thread **pp, struct gtaskqueue *tq)
150 {
151 
152 	while (tq->tq_tcount > 0 || tq->tq_callouts > 0) {
153 		wakeup(tq);
154 		TQ_SLEEP(tq, pp, &tq->tq_mutex, PWAIT, "taskqueue_destroy", 0);
155 	}
156 }
157 
158 static void
159 gtaskqueue_free(struct gtaskqueue *queue)
160 {
161 
162 	TQ_LOCK(queue);
163 	queue->tq_flags &= ~TQ_FLAGS_ACTIVE;
164 	gtaskqueue_terminate(queue->tq_threads, queue);
165 	KASSERT(TAILQ_EMPTY(&queue->tq_active), ("Tasks still running?"));
166 	KASSERT(queue->tq_callouts == 0, ("Armed timeout tasks"));
167 	mtx_destroy(&queue->tq_mutex);
168 	free(queue->tq_threads, M_GTASKQUEUE);
169 	free(queue->tq_name, M_GTASKQUEUE);
170 	free(queue, M_GTASKQUEUE);
171 }
172 
173 int
174 grouptaskqueue_enqueue(struct gtaskqueue *queue, struct gtask *gtask)
175 {
176 	TQ_LOCK(queue);
177 	if (gtask->ta_flags & TASK_ENQUEUED) {
178 		TQ_UNLOCK(queue);
179 		return (0);
180 	}
181 	STAILQ_INSERT_TAIL(&queue->tq_queue, gtask, ta_link);
182 	gtask->ta_flags |= TASK_ENQUEUED;
183 	TQ_UNLOCK(queue);
184 	if ((queue->tq_flags & TQ_FLAGS_BLOCKED) == 0)
185 		queue->tq_enqueue(queue->tq_context);
186 	return (0);
187 }
188 
189 static void
190 gtaskqueue_task_nop_fn(void *context)
191 {
192 }
193 
194 /*
195  * Block until all currently queued tasks in this taskqueue
196  * have begun execution.  Tasks queued during execution of
197  * this function are ignored.
198  */
199 static void
200 gtaskqueue_drain_tq_queue(struct gtaskqueue *queue)
201 {
202 	struct gtask t_barrier;
203 
204 	if (STAILQ_EMPTY(&queue->tq_queue))
205 		return;
206 
207 	/*
208 	 * Enqueue our barrier after all current tasks, but with
209 	 * the highest priority so that newly queued tasks cannot
210 	 * pass it.  Because of the high priority, we can not use
211 	 * taskqueue_enqueue_locked directly (which drops the lock
212 	 * anyway) so just insert it at tail while we have the
213 	 * queue lock.
214 	 */
215 	GTASK_INIT(&t_barrier, 0, USHRT_MAX, gtaskqueue_task_nop_fn, &t_barrier);
216 	STAILQ_INSERT_TAIL(&queue->tq_queue, &t_barrier, ta_link);
217 	t_barrier.ta_flags |= TASK_ENQUEUED;
218 
219 	/*
220 	 * Once the barrier has executed, all previously queued tasks
221 	 * have completed or are currently executing.
222 	 */
223 	while (t_barrier.ta_flags & TASK_ENQUEUED)
224 		TQ_SLEEP(queue, &t_barrier, &queue->tq_mutex, PWAIT, "-", 0);
225 }
226 
227 /*
228  * Block until all currently executing tasks for this taskqueue
229  * complete.  Tasks that begin execution during the execution
230  * of this function are ignored.
231  */
232 static void
233 gtaskqueue_drain_tq_active(struct gtaskqueue *queue)
234 {
235 	struct gtaskqueue_busy tb_marker, *tb_first;
236 
237 	if (TAILQ_EMPTY(&queue->tq_active))
238 		return;
239 
240 	/* Block taskq_terminate().*/
241 	queue->tq_callouts++;
242 
243 	/*
244 	 * Wait for all currently executing taskqueue threads
245 	 * to go idle.
246 	 */
247 	tb_marker.tb_running = TB_DRAIN_WAITER;
248 	TAILQ_INSERT_TAIL(&queue->tq_active, &tb_marker, tb_link);
249 	while (TAILQ_FIRST(&queue->tq_active) != &tb_marker)
250 		TQ_SLEEP(queue, &tb_marker, &queue->tq_mutex, PWAIT, "-", 0);
251 	TAILQ_REMOVE(&queue->tq_active, &tb_marker, tb_link);
252 
253 	/*
254 	 * Wakeup any other drain waiter that happened to queue up
255 	 * without any intervening active thread.
256 	 */
257 	tb_first = TAILQ_FIRST(&queue->tq_active);
258 	if (tb_first != NULL && tb_first->tb_running == TB_DRAIN_WAITER)
259 		wakeup(tb_first);
260 
261 	/* Release taskqueue_terminate(). */
262 	queue->tq_callouts--;
263 	if ((queue->tq_flags & TQ_FLAGS_ACTIVE) == 0)
264 		wakeup_one(queue->tq_threads);
265 }
266 
267 void
268 gtaskqueue_block(struct gtaskqueue *queue)
269 {
270 
271 	TQ_LOCK(queue);
272 	queue->tq_flags |= TQ_FLAGS_BLOCKED;
273 	TQ_UNLOCK(queue);
274 }
275 
276 void
277 gtaskqueue_unblock(struct gtaskqueue *queue)
278 {
279 
280 	TQ_LOCK(queue);
281 	queue->tq_flags &= ~TQ_FLAGS_BLOCKED;
282 	if (!STAILQ_EMPTY(&queue->tq_queue))
283 		queue->tq_enqueue(queue->tq_context);
284 	TQ_UNLOCK(queue);
285 }
286 
287 static void
288 gtaskqueue_run_locked(struct gtaskqueue *queue)
289 {
290 	struct gtaskqueue_busy tb;
291 	struct gtaskqueue_busy *tb_first;
292 	struct gtask *gtask;
293 
294 	KASSERT(queue != NULL, ("tq is NULL"));
295 	TQ_ASSERT_LOCKED(queue);
296 	tb.tb_running = NULL;
297 
298 	while (STAILQ_FIRST(&queue->tq_queue)) {
299 		TAILQ_INSERT_TAIL(&queue->tq_active, &tb, tb_link);
300 
301 		/*
302 		 * Carefully remove the first task from the queue and
303 		 * clear its TASK_ENQUEUED flag
304 		 */
305 		gtask = STAILQ_FIRST(&queue->tq_queue);
306 		KASSERT(gtask != NULL, ("task is NULL"));
307 		STAILQ_REMOVE_HEAD(&queue->tq_queue, ta_link);
308 		gtask->ta_flags &= ~TASK_ENQUEUED;
309 		tb.tb_running = gtask;
310 		TQ_UNLOCK(queue);
311 
312 		KASSERT(gtask->ta_func != NULL, ("task->ta_func is NULL"));
313 		gtask->ta_func(gtask->ta_context);
314 
315 		TQ_LOCK(queue);
316 		tb.tb_running = NULL;
317 		wakeup(gtask);
318 
319 		TAILQ_REMOVE(&queue->tq_active, &tb, tb_link);
320 		tb_first = TAILQ_FIRST(&queue->tq_active);
321 		if (tb_first != NULL &&
322 		    tb_first->tb_running == TB_DRAIN_WAITER)
323 			wakeup(tb_first);
324 	}
325 }
326 
327 static int
328 task_is_running(struct gtaskqueue *queue, struct gtask *gtask)
329 {
330 	struct gtaskqueue_busy *tb;
331 
332 	TQ_ASSERT_LOCKED(queue);
333 	TAILQ_FOREACH(tb, &queue->tq_active, tb_link) {
334 		if (tb->tb_running == gtask)
335 			return (1);
336 	}
337 	return (0);
338 }
339 
340 static int
341 gtaskqueue_cancel_locked(struct gtaskqueue *queue, struct gtask *gtask)
342 {
343 
344 	if (gtask->ta_flags & TASK_ENQUEUED)
345 		STAILQ_REMOVE(&queue->tq_queue, gtask, gtask, ta_link);
346 	gtask->ta_flags &= ~TASK_ENQUEUED;
347 	return (task_is_running(queue, gtask) ? EBUSY : 0);
348 }
349 
350 int
351 gtaskqueue_cancel(struct gtaskqueue *queue, struct gtask *gtask)
352 {
353 	int error;
354 
355 	TQ_LOCK(queue);
356 	error = gtaskqueue_cancel_locked(queue, gtask);
357 	TQ_UNLOCK(queue);
358 
359 	return (error);
360 }
361 
362 void
363 gtaskqueue_drain(struct gtaskqueue *queue, struct gtask *gtask)
364 {
365 
366 	if (!queue->tq_spin)
367 		WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, __func__);
368 
369 	TQ_LOCK(queue);
370 	while ((gtask->ta_flags & TASK_ENQUEUED) || task_is_running(queue, gtask))
371 		TQ_SLEEP(queue, gtask, &queue->tq_mutex, PWAIT, "-", 0);
372 	TQ_UNLOCK(queue);
373 }
374 
375 void
376 gtaskqueue_drain_all(struct gtaskqueue *queue)
377 {
378 
379 	if (!queue->tq_spin)
380 		WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, __func__);
381 
382 	TQ_LOCK(queue);
383 	gtaskqueue_drain_tq_queue(queue);
384 	gtaskqueue_drain_tq_active(queue);
385 	TQ_UNLOCK(queue);
386 }
387 
388 static int
389 _gtaskqueue_start_threads(struct gtaskqueue **tqp, int count, int pri,
390     cpuset_t *mask, const char *name, va_list ap)
391 {
392 	char ktname[MAXCOMLEN + 1];
393 	struct thread *td;
394 	struct gtaskqueue *tq;
395 	int i, error;
396 
397 	if (count <= 0)
398 		return (EINVAL);
399 
400 	vsnprintf(ktname, sizeof(ktname), name, ap);
401 	tq = *tqp;
402 
403 	tq->tq_threads = malloc(sizeof(struct thread *) * count, M_GTASKQUEUE,
404 	    M_NOWAIT | M_ZERO);
405 	if (tq->tq_threads == NULL) {
406 		printf("%s: no memory for %s threads\n", __func__, ktname);
407 		return (ENOMEM);
408 	}
409 
410 	for (i = 0; i < count; i++) {
411 		if (count == 1)
412 			error = kthread_add(gtaskqueue_thread_loop, tqp, NULL,
413 			    &tq->tq_threads[i], RFSTOPPED, 0, "%s", ktname);
414 		else
415 			error = kthread_add(gtaskqueue_thread_loop, tqp, NULL,
416 			    &tq->tq_threads[i], RFSTOPPED, 0,
417 			    "%s_%d", ktname, i);
418 		if (error) {
419 			/* should be ok to continue, taskqueue_free will dtrt */
420 			printf("%s: kthread_add(%s): error %d", __func__,
421 			    ktname, error);
422 			tq->tq_threads[i] = NULL;		/* paranoid */
423 		} else
424 			tq->tq_tcount++;
425 	}
426 	for (i = 0; i < count; i++) {
427 		if (tq->tq_threads[i] == NULL)
428 			continue;
429 		td = tq->tq_threads[i];
430 		if (mask) {
431 			error = cpuset_setthread(td->td_tid, mask);
432 			/*
433 			 * Failing to pin is rarely an actual fatal error;
434 			 * it'll just affect performance.
435 			 */
436 			if (error)
437 				printf("%s: curthread=%llu: can't pin; "
438 				    "error=%d\n",
439 				    __func__,
440 				    (unsigned long long) td->td_tid,
441 				    error);
442 		}
443 		thread_lock(td);
444 		sched_prio(td, pri);
445 		sched_add(td, SRQ_BORING);
446 		thread_unlock(td);
447 	}
448 
449 	return (0);
450 }
451 
452 static int
453 gtaskqueue_start_threads(struct gtaskqueue **tqp, int count, int pri,
454     const char *name, ...)
455 {
456 	va_list ap;
457 	int error;
458 
459 	va_start(ap, name);
460 	error = _gtaskqueue_start_threads(tqp, count, pri, NULL, name, ap);
461 	va_end(ap);
462 	return (error);
463 }
464 
465 static inline void
466 gtaskqueue_run_callback(struct gtaskqueue *tq,
467     enum taskqueue_callback_type cb_type)
468 {
469 	taskqueue_callback_fn tq_callback;
470 
471 	TQ_ASSERT_UNLOCKED(tq);
472 	tq_callback = tq->tq_callbacks[cb_type];
473 	if (tq_callback != NULL)
474 		tq_callback(tq->tq_cb_contexts[cb_type]);
475 }
476 
477 static void
478 gtaskqueue_thread_loop(void *arg)
479 {
480 	struct gtaskqueue **tqp, *tq;
481 
482 	tqp = arg;
483 	tq = *tqp;
484 	gtaskqueue_run_callback(tq, TASKQUEUE_CALLBACK_TYPE_INIT);
485 	TQ_LOCK(tq);
486 	while ((tq->tq_flags & TQ_FLAGS_ACTIVE) != 0) {
487 		/* XXX ? */
488 		gtaskqueue_run_locked(tq);
489 		/*
490 		 * Because taskqueue_run() can drop tq_mutex, we need to
491 		 * check if the TQ_FLAGS_ACTIVE flag wasn't removed in the
492 		 * meantime, which means we missed a wakeup.
493 		 */
494 		if ((tq->tq_flags & TQ_FLAGS_ACTIVE) == 0)
495 			break;
496 		TQ_SLEEP(tq, tq, &tq->tq_mutex, 0, "-", 0);
497 	}
498 	gtaskqueue_run_locked(tq);
499 	/*
500 	 * This thread is on its way out, so just drop the lock temporarily
501 	 * in order to call the shutdown callback.  This allows the callback
502 	 * to look at the taskqueue, even just before it dies.
503 	 */
504 	TQ_UNLOCK(tq);
505 	gtaskqueue_run_callback(tq, TASKQUEUE_CALLBACK_TYPE_SHUTDOWN);
506 	TQ_LOCK(tq);
507 
508 	/* rendezvous with thread that asked us to terminate */
509 	tq->tq_tcount--;
510 	wakeup_one(tq->tq_threads);
511 	TQ_UNLOCK(tq);
512 	kthread_exit();
513 }
514 
515 static void
516 gtaskqueue_thread_enqueue(void *context)
517 {
518 	struct gtaskqueue **tqp, *tq;
519 
520 	tqp = context;
521 	tq = *tqp;
522 	wakeup_one(tq);
523 }
524 
525 
526 static struct gtaskqueue *
527 gtaskqueue_create_fast(const char *name, int mflags,
528 		 taskqueue_enqueue_fn enqueue, void *context)
529 {
530 	return _gtaskqueue_create(name, mflags, enqueue, context,
531 			MTX_SPIN, "fast_taskqueue");
532 }
533 
534 
535 struct taskqgroup_cpu {
536 	LIST_HEAD(, grouptask)	tgc_tasks;
537 	struct gtaskqueue	*tgc_taskq;
538 	int	tgc_cnt;
539 	int	tgc_cpu;
540 };
541 
542 struct taskqgroup {
543 	struct taskqgroup_cpu tqg_queue[MAXCPU];
544 	struct mtx	tqg_lock;
545 	char *		tqg_name;
546 	int		tqg_adjusting;
547 	int		tqg_stride;
548 	int		tqg_cnt;
549 };
550 
551 struct taskq_bind_task {
552 	struct gtask bt_task;
553 	int	bt_cpuid;
554 };
555 
556 static void
557 taskqgroup_cpu_create(struct taskqgroup *qgroup, int idx)
558 {
559 	struct taskqgroup_cpu *qcpu;
560 
561 	qcpu = &qgroup->tqg_queue[idx];
562 	LIST_INIT(&qcpu->tgc_tasks);
563 	qcpu->tgc_taskq = gtaskqueue_create_fast(NULL, M_WAITOK,
564 	    taskqueue_thread_enqueue, &qcpu->tgc_taskq);
565 	gtaskqueue_start_threads(&qcpu->tgc_taskq, 1, PI_SOFT,
566 	    "%s_%d", qgroup->tqg_name, idx);
567 	qcpu->tgc_cpu = idx * qgroup->tqg_stride;
568 }
569 
570 static void
571 taskqgroup_cpu_remove(struct taskqgroup *qgroup, int idx)
572 {
573 
574 	gtaskqueue_free(qgroup->tqg_queue[idx].tgc_taskq);
575 }
576 
577 /*
578  * Find the taskq with least # of tasks that doesn't currently have any
579  * other queues from the uniq identifier.
580  */
581 static int
582 taskqgroup_find(struct taskqgroup *qgroup, void *uniq)
583 {
584 	struct grouptask *n;
585 	int i, idx, mincnt;
586 	int strict;
587 
588 	mtx_assert(&qgroup->tqg_lock, MA_OWNED);
589 	if (qgroup->tqg_cnt == 0)
590 		return (0);
591 	idx = -1;
592 	mincnt = INT_MAX;
593 	/*
594 	 * Two passes;  First scan for a queue with the least tasks that
595 	 * does not already service this uniq id.  If that fails simply find
596 	 * the queue with the least total tasks;
597 	 */
598 	for (strict = 1; mincnt == INT_MAX; strict = 0) {
599 		for (i = 0; i < qgroup->tqg_cnt; i++) {
600 			if (qgroup->tqg_queue[i].tgc_cnt > mincnt)
601 				continue;
602 			if (strict) {
603 				LIST_FOREACH(n,
604 				    &qgroup->tqg_queue[i].tgc_tasks, gt_list)
605 					if (n->gt_uniq == uniq)
606 						break;
607 				if (n != NULL)
608 					continue;
609 			}
610 			mincnt = qgroup->tqg_queue[i].tgc_cnt;
611 			idx = i;
612 		}
613 	}
614 	if (idx == -1)
615 		panic("taskqgroup_find: Failed to pick a qid.");
616 
617 	return (idx);
618 }
619 
620 void
621 taskqgroup_attach(struct taskqgroup *qgroup, struct grouptask *gtask,
622     void *uniq, int irq, char *name)
623 {
624 	cpuset_t mask;
625 	int qid;
626 
627 	gtask->gt_uniq = uniq;
628 	gtask->gt_name = name;
629 	gtask->gt_irq = irq;
630 	gtask->gt_cpu = -1;
631 	mtx_lock(&qgroup->tqg_lock);
632 	qid = taskqgroup_find(qgroup, uniq);
633 	qgroup->tqg_queue[qid].tgc_cnt++;
634 	LIST_INSERT_HEAD(&qgroup->tqg_queue[qid].tgc_tasks, gtask, gt_list);
635 	gtask->gt_taskqueue = qgroup->tqg_queue[qid].tgc_taskq;
636 	if (irq != -1 && smp_started) {
637 		CPU_ZERO(&mask);
638 		CPU_SET(qgroup->tqg_queue[qid].tgc_cpu, &mask);
639 		mtx_unlock(&qgroup->tqg_lock);
640 		intr_setaffinity(irq, &mask);
641 	} else
642 		mtx_unlock(&qgroup->tqg_lock);
643 }
644 
645 int
646 taskqgroup_attach_cpu(struct taskqgroup *qgroup, struct grouptask *gtask,
647 	void *uniq, int cpu, int irq, char *name)
648 {
649 	cpuset_t mask;
650 	int i, qid;
651 
652 	qid = -1;
653 	gtask->gt_uniq = uniq;
654 	gtask->gt_name = name;
655 	gtask->gt_irq = irq;
656 	gtask->gt_cpu = cpu;
657 	mtx_lock(&qgroup->tqg_lock);
658 	if (smp_started) {
659 		for (i = 0; i < qgroup->tqg_cnt; i++)
660 			if (qgroup->tqg_queue[i].tgc_cpu == cpu) {
661 				qid = i;
662 				break;
663 			}
664 		if (qid == -1) {
665 			mtx_unlock(&qgroup->tqg_lock);
666 			return (EINVAL);
667 		}
668 	} else
669 		qid = 0;
670 	qgroup->tqg_queue[qid].tgc_cnt++;
671 	LIST_INSERT_HEAD(&qgroup->tqg_queue[qid].tgc_tasks, gtask, gt_list);
672 	gtask->gt_taskqueue = qgroup->tqg_queue[qid].tgc_taskq;
673 	if (irq != -1 && smp_started) {
674 		CPU_ZERO(&mask);
675 		CPU_SET(qgroup->tqg_queue[qid].tgc_cpu, &mask);
676 		mtx_unlock(&qgroup->tqg_lock);
677 		intr_setaffinity(irq, &mask);
678 	} else
679 		mtx_unlock(&qgroup->tqg_lock);
680 	return (0);
681 }
682 
683 void
684 taskqgroup_detach(struct taskqgroup *qgroup, struct grouptask *gtask)
685 {
686 	int i;
687 
688 	mtx_lock(&qgroup->tqg_lock);
689 	for (i = 0; i < qgroup->tqg_cnt; i++)
690 		if (qgroup->tqg_queue[i].tgc_taskq == gtask->gt_taskqueue)
691 			break;
692 	if (i == qgroup->tqg_cnt)
693 		panic("taskqgroup_detach: task not in group\n");
694 	qgroup->tqg_queue[i].tgc_cnt--;
695 	LIST_REMOVE(gtask, gt_list);
696 	mtx_unlock(&qgroup->tqg_lock);
697 	gtask->gt_taskqueue = NULL;
698 }
699 
700 static void
701 taskqgroup_binder(void *ctx)
702 {
703 	struct taskq_bind_task *gtask = (struct taskq_bind_task *)ctx;
704 	cpuset_t mask;
705 	int error;
706 
707 	CPU_ZERO(&mask);
708 	CPU_SET(gtask->bt_cpuid, &mask);
709 	error = cpuset_setthread(curthread->td_tid, &mask);
710 	thread_lock(curthread);
711 	sched_bind(curthread, gtask->bt_cpuid);
712 	thread_unlock(curthread);
713 
714 	if (error)
715 		printf("taskqgroup_binder: setaffinity failed: %d\n",
716 		    error);
717 	free(gtask, M_DEVBUF);
718 }
719 
720 static void
721 taskqgroup_bind(struct taskqgroup *qgroup)
722 {
723 	struct taskq_bind_task *gtask;
724 	int i;
725 
726 	/*
727 	 * Bind taskqueue threads to specific CPUs, if they have been assigned
728 	 * one.
729 	 */
730 	for (i = 0; i < qgroup->tqg_cnt; i++) {
731 		gtask = malloc(sizeof (*gtask), M_DEVBUF, M_NOWAIT);
732 		GTASK_INIT(&gtask->bt_task, 0, 0, taskqgroup_binder, gtask);
733 		gtask->bt_cpuid = qgroup->tqg_queue[i].tgc_cpu;
734 		grouptaskqueue_enqueue(qgroup->tqg_queue[i].tgc_taskq,
735 		    &gtask->bt_task);
736 	}
737 }
738 
739 static int
740 _taskqgroup_adjust(struct taskqgroup *qgroup, int cnt, int stride)
741 {
742 	LIST_HEAD(, grouptask) gtask_head = LIST_HEAD_INITIALIZER(NULL);
743 	cpuset_t mask;
744 	struct grouptask *gtask;
745 	int i, k, old_cnt, qid, cpu;
746 
747 	mtx_assert(&qgroup->tqg_lock, MA_OWNED);
748 
749 	if (cnt < 1 || cnt * stride > mp_ncpus || !smp_started) {
750 		printf("taskqgroup_adjust failed cnt: %d stride: %d mp_ncpus: %d smp_started: %d\n",
751 			   cnt, stride, mp_ncpus, smp_started);
752 		return (EINVAL);
753 	}
754 	if (qgroup->tqg_adjusting) {
755 		printf("taskqgroup_adjust failed: adjusting\n");
756 		return (EBUSY);
757 	}
758 	qgroup->tqg_adjusting = 1;
759 	old_cnt = qgroup->tqg_cnt;
760 	mtx_unlock(&qgroup->tqg_lock);
761 	/*
762 	 * Set up queue for tasks added before boot.
763 	 */
764 	if (old_cnt == 0) {
765 		LIST_SWAP(&gtask_head, &qgroup->tqg_queue[0].tgc_tasks,
766 		    grouptask, gt_list);
767 		qgroup->tqg_queue[0].tgc_cnt = 0;
768 	}
769 
770 	/*
771 	 * If new taskq threads have been added.
772 	 */
773 	for (i = old_cnt; i < cnt; i++)
774 		taskqgroup_cpu_create(qgroup, i);
775 	mtx_lock(&qgroup->tqg_lock);
776 	qgroup->tqg_cnt = cnt;
777 	qgroup->tqg_stride = stride;
778 
779 	/*
780 	 * Adjust drivers to use new taskqs.
781 	 */
782 	for (i = 0; i < old_cnt; i++) {
783 		while ((gtask = LIST_FIRST(&qgroup->tqg_queue[i].tgc_tasks))) {
784 			LIST_REMOVE(gtask, gt_list);
785 			qgroup->tqg_queue[i].tgc_cnt--;
786 			LIST_INSERT_HEAD(&gtask_head, gtask, gt_list);
787 		}
788 	}
789 
790 	while ((gtask = LIST_FIRST(&gtask_head))) {
791 		LIST_REMOVE(gtask, gt_list);
792 		if (gtask->gt_cpu == -1)
793 			qid = taskqgroup_find(qgroup, gtask->gt_uniq);
794 		else {
795 			for (i = 0; i < qgroup->tqg_cnt; i++)
796 				if (qgroup->tqg_queue[i].tgc_cpu == gtask->gt_cpu) {
797 					qid = i;
798 					break;
799 				}
800 		}
801 		qgroup->tqg_queue[qid].tgc_cnt++;
802 		LIST_INSERT_HEAD(&qgroup->tqg_queue[qid].tgc_tasks, gtask,
803 		    gt_list);
804 		gtask->gt_taskqueue = qgroup->tqg_queue[qid].tgc_taskq;
805 	}
806 	/*
807 	 * Set new CPU and IRQ affinity
808 	 */
809 	cpu = CPU_FIRST();
810 	for (i = 0; i < cnt; i++) {
811 		qgroup->tqg_queue[i].tgc_cpu = cpu;
812 		for (k = 0; k < qgroup->tqg_stride; k++)
813 			cpu = CPU_NEXT(cpu);
814 		CPU_ZERO(&mask);
815 		CPU_SET(qgroup->tqg_queue[i].tgc_cpu, &mask);
816 		LIST_FOREACH(gtask, &qgroup->tqg_queue[i].tgc_tasks, gt_list) {
817 			if (gtask->gt_irq == -1)
818 				continue;
819 			intr_setaffinity(gtask->gt_irq, &mask);
820 		}
821 	}
822 	mtx_unlock(&qgroup->tqg_lock);
823 
824 	/*
825 	 * If taskq thread count has been reduced.
826 	 */
827 	for (i = cnt; i < old_cnt; i++)
828 		taskqgroup_cpu_remove(qgroup, i);
829 
830 	mtx_lock(&qgroup->tqg_lock);
831 	qgroup->tqg_adjusting = 0;
832 
833 	taskqgroup_bind(qgroup);
834 
835 	return (0);
836 }
837 
838 int
839 taskqgroup_adjust(struct taskqgroup *qgroup, int cpu, int stride)
840 {
841 	int error;
842 
843 	mtx_lock(&qgroup->tqg_lock);
844 	error = _taskqgroup_adjust(qgroup, cpu, stride);
845 	mtx_unlock(&qgroup->tqg_lock);
846 
847 	return (error);
848 }
849 
850 struct taskqgroup *
851 taskqgroup_create(char *name)
852 {
853 	struct taskqgroup *qgroup;
854 
855 	qgroup = malloc(sizeof(*qgroup), M_GTASKQUEUE, M_WAITOK | M_ZERO);
856 	mtx_init(&qgroup->tqg_lock, "taskqgroup", NULL, MTX_DEF);
857 	qgroup->tqg_name = name;
858 	LIST_INIT(&qgroup->tqg_queue[0].tgc_tasks);
859 
860 	return (qgroup);
861 }
862 
863 void
864 taskqgroup_destroy(struct taskqgroup *qgroup)
865 {
866 
867 }
868