sys/kern/subr_taskqueue.c

   1 /*-
   2  * Copyright (c) 2000 Doug Rabson
   3  * All rights reserved.
   4  *
   5  * Redistribution and use in source and binary forms, with or without
   6  * modification, are permitted provided that the following conditions
   7  * are met:
   8  * 1. Redistributions of source code must retain the above copyright
   9  *    notice, this list of conditions and the following disclaimer.
  10  * 2. Redistributions in binary form must reproduce the above copyright
  11  *    notice, this list of conditions and the following disclaimer in the
  12  *    documentation and/or other materials provided with the distribution.
  13  *
  14  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
  15  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  16  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  17  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
  18  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  19  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  20  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  21  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  22  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  23  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  24  * SUCH DAMAGE.
  25  */
  26
  27 #include <sys/cdefs.h>
  28 __FBSDID("$FreeBSD$");
  29
  30 #include <sys/param.h>
  31 #include <sys/systm.h>
  32 #include <sys/bus.h>
  33 #include <sys/cpuset.h>
  34 #include <sys/interrupt.h>
  35 #include <sys/kernel.h>
  36 #include <sys/kthread.h>
  37 #include <sys/libkern.h>
  38 #include <sys/limits.h>
  39 #include <sys/lock.h>
  40 #include <sys/malloc.h>
  41 #include <sys/mutex.h>
  42 #include <sys/proc.h>
  43 #include <sys/sched.h>
  44 #include <sys/smp.h>
  45 #include <sys/taskqueue.h>
  46 #include <sys/unistd.h>
  47 #include <machine/stdarg.h>
  48
  49 static MALLOC_DEFINE(M_TASKQUEUE, "taskqueue", "Task Queues");
  50 static void     *taskqueue_giant_ih;
  51 static void     *taskqueue_ih;
  52 static void      taskqueue_fast_enqueue(void *);
  53 static void      taskqueue_swi_enqueue(void *);
  54 static void      taskqueue_swi_giant_enqueue(void *);
  55
  56 struct taskqueue_busy {
  57         struct task     *tb_running;
  58         TAILQ_ENTRY(taskqueue_busy) tb_link;
  59 };
  60
  61 struct task * const TB_DRAIN_WAITER = (struct task *)0x1;
  62
  63 struct taskqueue {
  64         STAILQ_HEAD(, task)     tq_queue;
  65         taskqueue_enqueue_fn    tq_enqueue;
  66         void                    *tq_context;
  67         char                    *tq_name;
  68         TAILQ_HEAD(, taskqueue_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 void
 104 _timeout_task_init(struct taskqueue *queue, struct timeout_task *timeout_task,
 105     int priority, task_fn_t func, void *context)
 106 {
 107
 108         TASK_INIT(&timeout_task->t, priority, func, context);
 109         callout_init_mtx(&timeout_task->c, &queue->tq_mutex,
 110             CALLOUT_RETURNUNLOCKED);
 111         timeout_task->q = queue;
 112         timeout_task->f = 0;
 113 }
 114
 115 static __inline int
 116 TQ_SLEEP(struct taskqueue *tq, void *p, struct mtx *m, int pri, const char *wm,
 117     int t)
 118 {
 119         if (tq->tq_spin)
 120                 return (msleep_spin(p, m, wm, t));
 121         return (msleep(p, m, pri, wm, t));
 122 }
 123
 124 static struct taskqueue *
 125 _taskqueue_create(const char *name, int mflags,
 126                  taskqueue_enqueue_fn enqueue, void *context,
 127                  int mtxflags, const char *mtxname __unused)
 128 {
 129         struct taskqueue *queue;
 130         char *tq_name;
 131
 132         tq_name = malloc(TASKQUEUE_NAMELEN, M_TASKQUEUE, mflags | M_ZERO);
 133         if (!tq_name)
 134                 return (NULL);
 135
 136         snprintf(tq_name, TASKQUEUE_NAMELEN, "%s", (name) ? name : "taskqueue");
 137
 138         queue = malloc(sizeof(struct taskqueue), M_TASKQUEUE, mflags | M_ZERO);
 139         if (!queue)
 140                 return (NULL);
 141
 142         STAILQ_INIT(&queue->tq_queue);
 143         TAILQ_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 == taskqueue_fast_enqueue ||
 150             enqueue == taskqueue_swi_enqueue ||
 151             enqueue == taskqueue_swi_giant_enqueue ||
 152             enqueue == taskqueue_thread_enqueue)
 153                 queue->tq_flags |= TQ_FLAGS_UNLOCKED_ENQUEUE;
 154         mtx_init(&queue->tq_mutex, tq_name, NULL, mtxflags);
 155
 156         return (queue);
 157 }
 158
 159 struct taskqueue *
 160 taskqueue_create(const char *name, int mflags,
 161                  taskqueue_enqueue_fn enqueue, void *context)
 162 {
 163
 164         return _taskqueue_create(name, mflags, enqueue, context,
 165                         MTX_DEF, name);
 166 }
 167
 168 void
 169 taskqueue_set_callback(struct taskqueue *queue,
 170     enum taskqueue_callback_type cb_type, taskqueue_callback_fn callback,
 171     void *context)
 172 {
 173
 174         KASSERT(((cb_type >= TASKQUEUE_CALLBACK_TYPE_MIN) &&
 175             (cb_type <= TASKQUEUE_CALLBACK_TYPE_MAX)),
 176             ("Callback type %d not valid, must be %d-%d", cb_type,
 177             TASKQUEUE_CALLBACK_TYPE_MIN, TASKQUEUE_CALLBACK_TYPE_MAX));
 178         KASSERT((queue->tq_callbacks[cb_type] == NULL),
 179             ("Re-initialization of taskqueue callback?"));
 180
 181         queue->tq_callbacks[cb_type] = callback;
 182         queue->tq_cb_contexts[cb_type] = context;
 183 }
 184
 185 /*
 186  * Signal a taskqueue thread to terminate.
 187  */
 188 static void
 189 taskqueue_terminate(struct thread **pp, struct taskqueue *tq)
 190 {
 191
 192         while (tq->tq_tcount > 0 || tq->tq_callouts > 0) {
 193                 wakeup(tq);
 194                 TQ_SLEEP(tq, pp, &tq->tq_mutex, PWAIT, "taskqueue_destroy", 0);
 195         }
 196 }
 197
 198 void
 199 taskqueue_free(struct taskqueue *queue)
 200 {
 201
 202         TQ_LOCK(queue);
 203         queue->tq_flags &= ~TQ_FLAGS_ACTIVE;
 204         taskqueue_terminate(queue->tq_threads, queue);
 205         KASSERT(TAILQ_EMPTY(&queue->tq_active), ("Tasks still running?"));
 206         KASSERT(queue->tq_callouts == 0, ("Armed timeout tasks"));
 207         mtx_destroy(&queue->tq_mutex);
 208         free(queue->tq_threads, M_TASKQUEUE);
 209         free(queue->tq_name, M_TASKQUEUE);
 210         free(queue, M_TASKQUEUE);
 211 }
 212
 213 static int
 214 taskqueue_enqueue_locked(struct taskqueue *queue, struct task *task)
 215 {
 216         struct task *ins;
 217         struct task *prev;
 218
 219         KASSERT(task->ta_func != NULL, ("enqueueing task with NULL func"));
 220         /*
 221          * Count multiple enqueues.
 222          */
 223         if (task->ta_pending) {
 224                 if (task->ta_pending < USHRT_MAX)
 225                         task->ta_pending++;
 226                 TQ_UNLOCK(queue);
 227                 return (0);
 228         }
 229
 230         /*
 231          * Optimise the case when all tasks have the same priority.
 232          */
 233         prev = STAILQ_LAST(&queue->tq_queue, task, ta_link);
 234         if (!prev || prev->ta_priority >= task->ta_priority) {
 235                 STAILQ_INSERT_TAIL(&queue->tq_queue, task, ta_link);
 236         } else {
 237                 prev = NULL;
 238                 for (ins = STAILQ_FIRST(&queue->tq_queue); ins;
 239                      prev = ins, ins = STAILQ_NEXT(ins, ta_link))
 240                         if (ins->ta_priority < task->ta_priority)
 241                                 break;
 242
 243                 if (prev)
 244                         STAILQ_INSERT_AFTER(&queue->tq_queue, prev, task, ta_link);
 245                 else
 246                         STAILQ_INSERT_HEAD(&queue->tq_queue, task, ta_link);
 247         }
 248
 249         task->ta_pending = 1;
 250         if ((queue->tq_flags & TQ_FLAGS_UNLOCKED_ENQUEUE) != 0)
 251                 TQ_UNLOCK(queue);
 252         if ((queue->tq_flags & TQ_FLAGS_BLOCKED) == 0)
 253                 queue->tq_enqueue(queue->tq_context);
 254         if ((queue->tq_flags & TQ_FLAGS_UNLOCKED_ENQUEUE) == 0)
 255                 TQ_UNLOCK(queue);
 256
 257         /* Return with lock released. */
 258         return (0);
 259 }
 260
 261 int
 262 grouptaskqueue_enqueue(struct taskqueue *queue, struct task *task)
 263 {
 264         TQ_LOCK(queue);
 265         if (task->ta_pending) {
 266                 TQ_UNLOCK(queue);
 267                 return (0);
 268         }
 269         STAILQ_INSERT_TAIL(&queue->tq_queue, task, ta_link);
 270         task->ta_pending = 1;
 271         TQ_UNLOCK(queue);
 272         if ((queue->tq_flags & TQ_FLAGS_BLOCKED) == 0)
 273                 queue->tq_enqueue(queue->tq_context);
 274         return (0);
 275 }
 276
 277 int
 278 taskqueue_enqueue(struct taskqueue *queue, struct task *task)
 279 {
 280         int res;
 281
 282         TQ_LOCK(queue);
 283         res = taskqueue_enqueue_locked(queue, task);
 284         /* The lock is released inside. */
 285
 286         return (res);
 287 }
 288
 289 static void
 290 taskqueue_timeout_func(void *arg)
 291 {
 292         struct taskqueue *queue;
 293         struct timeout_task *timeout_task;
 294
 295         timeout_task = arg;
 296         queue = timeout_task->q;
 297         KASSERT((timeout_task->f & DT_CALLOUT_ARMED) != 0, ("Stray timeout"));
 298         timeout_task->f &= ~DT_CALLOUT_ARMED;
 299         queue->tq_callouts--;
 300         taskqueue_enqueue_locked(timeout_task->q, &timeout_task->t);
 301         /* The lock is released inside. */
 302 }
 303
 304 int
 305 taskqueue_enqueue_timeout(struct taskqueue *queue,
 306     struct timeout_task *timeout_task, int ticks)
 307 {
 308         int res;
 309
 310         TQ_LOCK(queue);
 311         KASSERT(timeout_task->q == NULL || timeout_task->q == queue,
 312             ("Migrated queue"));
 313         KASSERT(!queue->tq_spin, ("Timeout for spin-queue"));
 314         timeout_task->q = queue;
 315         res = timeout_task->t.ta_pending;
 316         if (ticks == 0) {
 317                 taskqueue_enqueue_locked(queue, &timeout_task->t);
 318                 /* The lock is released inside. */
 319         } else {
 320                 if ((timeout_task->f & DT_CALLOUT_ARMED) != 0) {
 321                         res++;
 322                 } else {
 323                         queue->tq_callouts++;
 324                         timeout_task->f |= DT_CALLOUT_ARMED;
 325                         if (ticks < 0)
 326                                 ticks = -ticks; /* Ignore overflow. */
 327                 }
 328                 if (ticks > 0) {
 329                         callout_reset(&timeout_task->c, ticks,
 330                             taskqueue_timeout_func, timeout_task);
 331                 }
 332                 TQ_UNLOCK(queue);
 333         }
 334         return (res);
 335 }
 336
 337 static void
 338 taskqueue_task_nop_fn(void *context, int pending)
 339 {
 340 }
 341
 342 /*
 343  * Block until all currently queued tasks in this taskqueue
 344  * have begun execution.  Tasks queued during execution of
 345  * this function are ignored.
 346  */
 347 static void
 348 taskqueue_drain_tq_queue(struct taskqueue *queue)
 349 {
 350         struct task t_barrier;
 351
 352         if (STAILQ_EMPTY(&queue->tq_queue))
 353                 return;
 354
 355         /*
 356          * Enqueue our barrier after all current tasks, but with
 357          * the highest priority so that newly queued tasks cannot
 358          * pass it.  Because of the high priority, we can not use
 359          * taskqueue_enqueue_locked directly (which drops the lock
 360          * anyway) so just insert it at tail while we have the
 361          * queue lock.
 362          */
 363         TASK_INIT(&t_barrier, USHRT_MAX, taskqueue_task_nop_fn, &t_barrier);
 364         STAILQ_INSERT_TAIL(&queue->tq_queue, &t_barrier, ta_link);
 365         t_barrier.ta_pending = 1;
 366
 367         /*
 368          * Once the barrier has executed, all previously queued tasks
 369          * have completed or are currently executing.
 370          */
 371         while (t_barrier.ta_pending != 0)
 372                 TQ_SLEEP(queue, &t_barrier, &queue->tq_mutex, PWAIT, "-", 0);
 373 }
 374
 375 /*
 376  * Block until all currently executing tasks for this taskqueue
 377  * complete.  Tasks that begin execution during the execution
 378  * of this function are ignored.
 379  */
 380 static void
 381 taskqueue_drain_tq_active(struct taskqueue *queue)
 382 {
 383         struct taskqueue_busy tb_marker, *tb_first;
 384
 385         if (TAILQ_EMPTY(&queue->tq_active))
 386                 return;
 387
 388         /* Block taskq_terminate().*/
 389         queue->tq_callouts++;
 390
 391         /*
 392          * Wait for all currently executing taskqueue threads
 393          * to go idle.
 394          */
 395         tb_marker.tb_running = TB_DRAIN_WAITER;
 396         TAILQ_INSERT_TAIL(&queue->tq_active, &tb_marker, tb_link);
 397         while (TAILQ_FIRST(&queue->tq_active) != &tb_marker)
 398                 TQ_SLEEP(queue, &tb_marker, &queue->tq_mutex, PWAIT, "-", 0);
 399         TAILQ_REMOVE(&queue->tq_active, &tb_marker, tb_link);
 400
 401         /*
 402          * Wakeup any other drain waiter that happened to queue up
 403          * without any intervening active thread.
 404          */
 405         tb_first = TAILQ_FIRST(&queue->tq_active);
 406         if (tb_first != NULL && tb_first->tb_running == TB_DRAIN_WAITER)
 407                 wakeup(tb_first);
 408
 409         /* Release taskqueue_terminate(). */
 410         queue->tq_callouts--;
 411         if ((queue->tq_flags & TQ_FLAGS_ACTIVE) == 0)
 412                 wakeup_one(queue->tq_threads);
 413 }
 414
 415 void
 416 taskqueue_block(struct taskqueue *queue)
 417 {
 418
 419         TQ_LOCK(queue);
 420         queue->tq_flags |= TQ_FLAGS_BLOCKED;
 421         TQ_UNLOCK(queue);
 422 }
 423
 424 void
 425 taskqueue_unblock(struct taskqueue *queue)
 426 {
 427
 428         TQ_LOCK(queue);
 429         queue->tq_flags &= ~TQ_FLAGS_BLOCKED;
 430         if (!STAILQ_EMPTY(&queue->tq_queue))
 431                 queue->tq_enqueue(queue->tq_context);
 432         TQ_UNLOCK(queue);
 433 }
 434
 435 static void
 436 taskqueue_run_locked(struct taskqueue *queue)
 437 {
 438         struct taskqueue_busy tb;
 439         struct taskqueue_busy *tb_first;
 440         struct task *task;
 441         int pending;
 442
 443         KASSERT(queue != NULL, ("tq is NULL"));
 444         TQ_ASSERT_LOCKED(queue);
 445         tb.tb_running = NULL;
 446
 447         while (STAILQ_FIRST(&queue->tq_queue)) {
 448                 TAILQ_INSERT_TAIL(&queue->tq_active, &tb, tb_link);
 449
 450                 /*
 451                  * Carefully remove the first task from the queue and
 452                  * zero its pending count.
 453                  */
 454                 task = STAILQ_FIRST(&queue->tq_queue);
 455                 KASSERT(task != NULL, ("task is NULL"));
 456                 STAILQ_REMOVE_HEAD(&queue->tq_queue, ta_link);
 457                 pending = task->ta_pending;
 458                 task->ta_pending = 0;
 459                 tb.tb_running = task;
 460                 TQ_UNLOCK(queue);
 461
 462                 KASSERT(task->ta_func != NULL, ("task->ta_func is NULL"));
 463                 task->ta_func(task->ta_context, pending);
 464
 465                 TQ_LOCK(queue);
 466                 tb.tb_running = NULL;
 467                 wakeup(task);
 468
 469                 TAILQ_REMOVE(&queue->tq_active, &tb, tb_link);
 470                 tb_first = TAILQ_FIRST(&queue->tq_active);
 471                 if (tb_first != NULL &&
 472                     tb_first->tb_running == TB_DRAIN_WAITER)
 473                         wakeup(tb_first);
 474         }
 475 }
 476
 477 void
 478 taskqueue_run(struct taskqueue *queue)
 479 {
 480
 481         TQ_LOCK(queue);
 482         taskqueue_run_locked(queue);
 483         TQ_UNLOCK(queue);
 484 }
 485
 486 static int
 487 task_is_running(struct taskqueue *queue, struct task *task)
 488 {
 489         struct taskqueue_busy *tb;
 490
 491         TQ_ASSERT_LOCKED(queue);
 492         TAILQ_FOREACH(tb, &queue->tq_active, tb_link) {
 493                 if (tb->tb_running == task)
 494                         return (1);
 495         }
 496         return (0);
 497 }
 498
 499 static int
 500 taskqueue_cancel_locked(struct taskqueue *queue, struct task *task,
 501     u_int *pendp)
 502 {
 503
 504         if (task->ta_pending > 0)
 505                 STAILQ_REMOVE(&queue->tq_queue, task, task, ta_link);
 506         if (pendp != NULL)
 507                 *pendp = task->ta_pending;
 508         task->ta_pending = 0;
 509         return (task_is_running(queue, task) ? EBUSY : 0);
 510 }
 511
 512 int
 513 taskqueue_cancel(struct taskqueue *queue, struct task *task, u_int *pendp)
 514 {
 515         int error;
 516
 517         TQ_LOCK(queue);
 518         error = taskqueue_cancel_locked(queue, task, pendp);
 519         TQ_UNLOCK(queue);
 520
 521         return (error);
 522 }
 523
 524 int
 525 taskqueue_cancel_timeout(struct taskqueue *queue,
 526     struct timeout_task *timeout_task, u_int *pendp)
 527 {
 528         u_int pending, pending1;
 529         int error;
 530
 531         TQ_LOCK(queue);
 532         pending = !!(callout_stop(&timeout_task->c) > 0);
 533         error = taskqueue_cancel_locked(queue, &timeout_task->t, &pending1);
 534         if ((timeout_task->f & DT_CALLOUT_ARMED) != 0) {
 535                 timeout_task->f &= ~DT_CALLOUT_ARMED;
 536                 queue->tq_callouts--;
 537         }
 538         TQ_UNLOCK(queue);
 539
 540         if (pendp != NULL)
 541                 *pendp = pending + pending1;
 542         return (error);
 543 }
 544
 545 void
 546 taskqueue_drain(struct taskqueue *queue, struct task *task)
 547 {
 548
 549         if (!queue->tq_spin)
 550                 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, __func__);
 551
 552         TQ_LOCK(queue);
 553         while (task->ta_pending != 0 || task_is_running(queue, task))
 554                 TQ_SLEEP(queue, task, &queue->tq_mutex, PWAIT, "-", 0);
 555         TQ_UNLOCK(queue);
 556 }
 557
 558 void
 559 taskqueue_drain_all(struct taskqueue *queue)
 560 {
 561
 562         if (!queue->tq_spin)
 563                 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, __func__);
 564
 565         TQ_LOCK(queue);
 566         taskqueue_drain_tq_queue(queue);
 567         taskqueue_drain_tq_active(queue);
 568         TQ_UNLOCK(queue);
 569 }
 570
 571 void
 572 taskqueue_drain_timeout(struct taskqueue *queue,
 573     struct timeout_task *timeout_task)
 574 {
 575
 576         callout_drain(&timeout_task->c);
 577         taskqueue_drain(queue, &timeout_task->t);
 578 }
 579
 580 static void
 581 taskqueue_swi_enqueue(void *context)
 582 {
 583         swi_sched(taskqueue_ih, 0);
 584 }
 585
 586 static void
 587 taskqueue_swi_run(void *dummy)
 588 {
 589         taskqueue_run(taskqueue_swi);
 590 }
 591
 592 static void
 593 taskqueue_swi_giant_enqueue(void *context)
 594 {
 595         swi_sched(taskqueue_giant_ih, 0);
 596 }
 597
 598 static void
 599 taskqueue_swi_giant_run(void *dummy)
 600 {
 601         taskqueue_run(taskqueue_swi_giant);
 602 }
 603
 604 static int
 605 _taskqueue_start_threads(struct taskqueue **tqp, int count, int pri,
 606     cpuset_t *mask, const char *name, va_list ap)
 607 {
 608         char ktname[MAXCOMLEN + 1];
 609         struct thread *td;
 610         struct taskqueue *tq;
 611         int i, error;
 612
 613         if (count <= 0)
 614                 return (EINVAL);
 615
 616         vsnprintf(ktname, sizeof(ktname), name, ap);
 617         tq = *tqp;
 618
 619         tq->tq_threads = malloc(sizeof(struct thread *) * count, M_TASKQUEUE,
 620             M_NOWAIT | M_ZERO);
 621         if (tq->tq_threads == NULL) {
 622                 printf("%s: no memory for %s threads\n", __func__, ktname);
 623                 return (ENOMEM);
 624         }
 625
 626         for (i = 0; i < count; i++) {
 627                 if (count == 1)
 628                         error = kthread_add(taskqueue_thread_loop, tqp, NULL,
 629                             &tq->tq_threads[i], RFSTOPPED, 0, "%s", ktname);
 630                 else
 631                         error = kthread_add(taskqueue_thread_loop, tqp, NULL,
 632                             &tq->tq_threads[i], RFSTOPPED, 0,
 633                             "%s_%d", ktname, i);
 634                 if (error) {
 635                         /* should be ok to continue, taskqueue_free will dtrt */
 636                         printf("%s: kthread_add(%s): error %d", __func__,
 637                             ktname, error);
 638                         tq->tq_threads[i] = NULL;               /* paranoid */
 639                 } else
 640                         tq->tq_tcount++;
 641         }
 642         for (i = 0; i < count; i++) {
 643                 if (tq->tq_threads[i] == NULL)
 644                         continue;
 645                 td = tq->tq_threads[i];
 646                 if (mask) {
 647                         error = cpuset_setthread(td->td_tid, mask);
 648                         /*
 649                          * Failing to pin is rarely an actual fatal error;
 650                          * it'll just affect performance.
 651                          */
 652                         if (error)
 653                                 printf("%s: curthread=%llu: can't pin; "
 654                                     "error=%d\n",
 655                                     __func__,
 656                                     (unsigned long long) td->td_tid,
 657                                     error);
 658                 }
 659                 thread_lock(td);
 660                 sched_prio(td, pri);
 661                 sched_add(td, SRQ_BORING);
 662                 thread_unlock(td);
 663         }
 664
 665         return (0);
 666 }
 667
 668 int
 669 taskqueue_start_threads(struct taskqueue **tqp, int count, int pri,
 670     const char *name, ...)
 671 {
 672         va_list ap;
 673         int error;
 674
 675         va_start(ap, name);
 676         error = _taskqueue_start_threads(tqp, count, pri, NULL, name, ap);
 677         va_end(ap);
 678         return (error);
 679 }
 680
 681 int
 682 taskqueue_start_threads_cpuset(struct taskqueue **tqp, int count, int pri,
 683     cpuset_t *mask, const char *name, ...)
 684 {
 685         va_list ap;
 686         int error;
 687
 688         va_start(ap, name);
 689         error = _taskqueue_start_threads(tqp, count, pri, mask, name, ap);
 690         va_end(ap);
 691         return (error);
 692 }
 693
 694 static inline void
 695 taskqueue_run_callback(struct taskqueue *tq,
 696     enum taskqueue_callback_type cb_type)
 697 {
 698         taskqueue_callback_fn tq_callback;
 699
 700         TQ_ASSERT_UNLOCKED(tq);
 701         tq_callback = tq->tq_callbacks[cb_type];
 702         if (tq_callback != NULL)
 703                 tq_callback(tq->tq_cb_contexts[cb_type]);
 704 }
 705
 706 void
 707 taskqueue_thread_loop(void *arg)
 708 {
 709         struct taskqueue **tqp, *tq;
 710
 711         tqp = arg;
 712         tq = *tqp;
 713         taskqueue_run_callback(tq, TASKQUEUE_CALLBACK_TYPE_INIT);
 714         TQ_LOCK(tq);
 715         while ((tq->tq_flags & TQ_FLAGS_ACTIVE) != 0) {
 716                 /* XXX ? */
 717                 taskqueue_run_locked(tq);
 718                 /*
 719                  * Because taskqueue_run() can drop tq_mutex, we need to
 720                  * check if the TQ_FLAGS_ACTIVE flag wasn't removed in the
 721                  * meantime, which means we missed a wakeup.
 722                  */
 723                 if ((tq->tq_flags & TQ_FLAGS_ACTIVE) == 0)
 724                         break;
 725                 TQ_SLEEP(tq, tq, &tq->tq_mutex, 0, "-", 0);
 726         }
 727         taskqueue_run_locked(tq);
 728         /*
 729          * This thread is on its way out, so just drop the lock temporarily
 730          * in order to call the shutdown callback.  This allows the callback
 731          * to look at the taskqueue, even just before it dies.
 732          */
 733         TQ_UNLOCK(tq);
 734         taskqueue_run_callback(tq, TASKQUEUE_CALLBACK_TYPE_SHUTDOWN);
 735         TQ_LOCK(tq);
 736
 737         /* rendezvous with thread that asked us to terminate */
 738         tq->tq_tcount--;
 739         wakeup_one(tq->tq_threads);
 740         TQ_UNLOCK(tq);
 741         kthread_exit();
 742 }
 743
 744 void
 745 taskqueue_thread_enqueue(void *context)
 746 {
 747         struct taskqueue **tqp, *tq;
 748
 749         tqp = context;
 750         tq = *tqp;
 751         wakeup_one(tq);
 752 }
 753
 754 TASKQUEUE_DEFINE(swi, taskqueue_swi_enqueue, NULL,
 755                  swi_add(NULL, "task queue", taskqueue_swi_run, NULL, SWI_TQ,
 756                      INTR_MPSAFE, &taskqueue_ih));
 757
 758 TASKQUEUE_DEFINE(swi_giant, taskqueue_swi_giant_enqueue, NULL,
 759                  swi_add(NULL, "Giant taskq", taskqueue_swi_giant_run,
 760                      NULL, SWI_TQ_GIANT, 0, &taskqueue_giant_ih));
 761
 762 TASKQUEUE_DEFINE_THREAD(thread);
 763
 764 struct taskqueue *
 765 taskqueue_create_fast(const char *name, int mflags,
 766                  taskqueue_enqueue_fn enqueue, void *context)
 767 {
 768         return _taskqueue_create(name, mflags, enqueue, context,
 769                         MTX_SPIN, "fast_taskqueue");
 770 }
 771
 772 static void     *taskqueue_fast_ih;
 773
 774 static void
 775 taskqueue_fast_enqueue(void *context)
 776 {
 777         swi_sched(taskqueue_fast_ih, 0);
 778 }
 779
 780 static void
 781 taskqueue_fast_run(void *dummy)
 782 {
 783         taskqueue_run(taskqueue_fast);
 784 }
 785
 786 TASKQUEUE_FAST_DEFINE(fast, taskqueue_fast_enqueue, NULL,
 787         swi_add(NULL, "fast taskq", taskqueue_fast_run, NULL,
 788         SWI_TQ_FAST, INTR_MPSAFE, &taskqueue_fast_ih));
 789
 790 int
 791 taskqueue_member(struct taskqueue *queue, struct thread *td)
 792 {
 793         int i, j, ret = 0;
 794
 795         for (i = 0, j = 0; ; i++) {
 796                 if (queue->tq_threads[i] == NULL)
 797                         continue;
 798                 if (queue->tq_threads[i] == td) {
 799                         ret = 1;
 800                         break;
 801                 }
 802                 if (++j >= queue->tq_tcount)
 803                         break;
 804         }
 805         return (ret);
 806 }
 807
 808 struct taskqgroup_cpu {
 809         LIST_HEAD(, grouptask)  tgc_tasks;
 810         struct taskqueue        *tgc_taskq;
 811         int     tgc_cnt;
 812         int     tgc_cpu;
 813 };
 814
 815 struct taskqgroup {
 816         struct taskqgroup_cpu tqg_queue[MAXCPU];
 817         struct mtx      tqg_lock;
 818         char *          tqg_name;
 819         int             tqg_adjusting;
 820         int             tqg_stride;
 821         int             tqg_cnt;
 822 };
 823
 824 struct taskq_bind_task {
 825         struct task bt_task;
 826         int     bt_cpuid;
 827 };
 828
 829 static void
 830 taskqgroup_cpu_create(struct taskqgroup *qgroup, int idx)
 831 {
 832         struct taskqgroup_cpu *qcpu;
 833
 834         qcpu = &qgroup->tqg_queue[idx];
 835         LIST_INIT(&qcpu->tgc_tasks);
 836         qcpu->tgc_taskq = taskqueue_create_fast(NULL, M_WAITOK,
 837             taskqueue_thread_enqueue, &qcpu->tgc_taskq);
 838         taskqueue_start_threads(&qcpu->tgc_taskq, 1, PI_SOFT,
 839             "%s_%d", qgroup->tqg_name, idx);
 840         qcpu->tgc_cpu = idx * qgroup->tqg_stride;
 841 }
 842
 843 static void
 844 taskqgroup_cpu_remove(struct taskqgroup *qgroup, int idx)
 845 {
 846
 847         taskqueue_free(qgroup->tqg_queue[idx].tgc_taskq);
 848 }
 849
 850 /*
 851  * Find the taskq with least # of tasks that doesn't currently have any
 852  * other queues from the uniq identifier.
 853  */
 854 static int
 855 taskqgroup_find(struct taskqgroup *qgroup, void *uniq)
 856 {
 857         struct grouptask *n;
 858         int i, idx, mincnt;
 859         int strict;
 860
 861         mtx_assert(&qgroup->tqg_lock, MA_OWNED);
 862         if (qgroup->tqg_cnt == 0)
 863                 return (0);
 864         idx = -1;
 865         mincnt = INT_MAX;
 866         /*
 867          * Two passes;  First scan for a queue with the least tasks that
 868          * does not already service this uniq id.  If that fails simply find
 869          * the queue with the least total tasks;
 870          */
 871         for (strict = 1; mincnt == INT_MAX; strict = 0) {
 872                 for (i = 0; i < qgroup->tqg_cnt; i++) {
 873                         if (qgroup->tqg_queue[i].tgc_cnt > mincnt)
 874                                 continue;
 875                         if (strict) {
 876                                 LIST_FOREACH(n,
 877                                     &qgroup->tqg_queue[i].tgc_tasks, gt_list)
 878                                         if (n->gt_uniq == uniq)
 879                                                 break;
 880                                 if (n != NULL)
 881                                         continue;
 882                         }
 883                         mincnt = qgroup->tqg_queue[i].tgc_cnt;
 884                         idx = i;
 885                 }
 886         }
 887         if (idx == -1)
 888                 panic("taskqgroup_find: Failed to pick a qid.");
 889
 890         return (idx);
 891 }
 892
 893 void
 894 taskqgroup_attach(struct taskqgroup *qgroup, struct grouptask *gtask,
 895     void *uniq, int irq, char *name)
 896 {
 897         cpuset_t mask;
 898         int qid;
 899
 900         gtask->gt_uniq = uniq;
 901         gtask->gt_name = name;
 902         gtask->gt_irq = irq;
 903         gtask->gt_cpu = -1;
 904         mtx_lock(&qgroup->tqg_lock);
 905         qid = taskqgroup_find(qgroup, uniq);
 906         qgroup->tqg_queue[qid].tgc_cnt++;
 907         LIST_INSERT_HEAD(&qgroup->tqg_queue[qid].tgc_tasks, gtask, gt_list);
 908         gtask->gt_taskqueue = qgroup->tqg_queue[qid].tgc_taskq;
 909         if (irq != -1 && smp_started) {
 910                 CPU_ZERO(&mask);
 911                 CPU_SET(qgroup->tqg_queue[qid].tgc_cpu, &mask);
 912                 mtx_unlock(&qgroup->tqg_lock);
 913                 intr_setaffinity(irq, &mask);
 914         } else
 915                 mtx_unlock(&qgroup->tqg_lock);
 916 }
 917
 918 int
 919 taskqgroup_attach_cpu(struct taskqgroup *qgroup, struct grouptask *gtask,
 920         void *uniq, int cpu, int irq, char *name)
 921 {
 922         cpuset_t mask;
 923         int i, qid;
 924
 925         qid = -1;
 926         gtask->gt_uniq = uniq;
 927         gtask->gt_name = name;
 928         gtask->gt_irq = irq;
 929         gtask->gt_cpu = cpu;
 930         mtx_lock(&qgroup->tqg_lock);
 931         if (smp_started) {
 932                 for (i = 0; i < qgroup->tqg_cnt; i++)
 933                         if (qgroup->tqg_queue[i].tgc_cpu == cpu) {
 934                                 qid = i;
 935                                 break;
 936                         }
 937                 if (qid == -1) {
 938                         mtx_unlock(&qgroup->tqg_lock);
 939                         return (EINVAL);
 940                 }
 941         } else
 942                 qid = 0;
 943         qgroup->tqg_queue[qid].tgc_cnt++;
 944         LIST_INSERT_HEAD(&qgroup->tqg_queue[qid].tgc_tasks, gtask, gt_list);
 945         gtask->gt_taskqueue = qgroup->tqg_queue[qid].tgc_taskq;
 946         if (irq != -1 && smp_started) {
 947                 CPU_ZERO(&mask);
 948                 CPU_SET(qgroup->tqg_queue[qid].tgc_cpu, &mask);
 949                 mtx_unlock(&qgroup->tqg_lock);
 950                 intr_setaffinity(irq, &mask);
 951         } else
 952                 mtx_unlock(&qgroup->tqg_lock);
 953         return (0);
 954 }
 955
 956 void
 957 taskqgroup_detach(struct taskqgroup *qgroup, struct grouptask *gtask)
 958 {
 959         int i;
 960
 961         mtx_lock(&qgroup->tqg_lock);
 962         for (i = 0; i < qgroup->tqg_cnt; i++)
 963                 if (qgroup->tqg_queue[i].tgc_taskq == gtask->gt_taskqueue)
 964                         break;
 965         if (i == qgroup->tqg_cnt)
 966                 panic("taskqgroup_detach: task not in group\n");
 967         qgroup->tqg_queue[i].tgc_cnt--;
 968         LIST_REMOVE(gtask, gt_list);
 969         mtx_unlock(&qgroup->tqg_lock);
 970         gtask->gt_taskqueue = NULL;
 971 }
 972
 973 static void
 974 taskqgroup_binder(void *ctx, int pending)
 975 {
 976         struct taskq_bind_task *task = (struct taskq_bind_task *)ctx;
 977         cpuset_t mask;
 978         int error;
 979
 980         CPU_ZERO(&mask);
 981         CPU_SET(task->bt_cpuid, &mask);
 982         error = cpuset_setthread(curthread->td_tid, &mask);
 983         thread_lock(curthread);
 984         sched_bind(curthread, task->bt_cpuid);
 985         thread_unlock(curthread);
 986
 987         if (error)
 988                 printf("taskqgroup_binder: setaffinity failed: %d\n",
 989                     error);
 990         free(task, M_DEVBUF);
 991 }
 992
 993 static void
 994 taskqgroup_bind(struct taskqgroup *qgroup)
 995 {
 996         struct taskq_bind_task *task;
 997         int i;
 998
 999         /*
1000          * Bind taskqueue threads to specific CPUs, if they have been assigned
1001          * one.
1002          */
1003         for (i = 0; i < qgroup->tqg_cnt; i++) {
1004                 task = malloc(sizeof (*task), M_DEVBUF, M_NOWAIT);
1005                 TASK_INIT(&task->bt_task, 0, taskqgroup_binder, task);
1006                 task->bt_cpuid = qgroup->tqg_queue[i].tgc_cpu;
1007                 taskqueue_enqueue(qgroup->tqg_queue[i].tgc_taskq,
1008                     &task->bt_task);
1009         }
1010 }
1011
1012 static int
1013 _taskqgroup_adjust(struct taskqgroup *qgroup, int cnt, int stride)
1014 {
1015         LIST_HEAD(, grouptask) gtask_head = LIST_HEAD_INITIALIZER(NULL);
1016         cpuset_t mask;
1017         struct grouptask *gtask;
1018         int i, old_cnt, qid;
1019
1020         mtx_assert(&qgroup->tqg_lock, MA_OWNED);
1021
1022         if (cnt < 1 || cnt * stride > mp_ncpus || !smp_started) {
1023                 printf("taskqgroup_adjust failed cnt: %d stride: %d mp_ncpus: %d smp_started: %d\n",
1024                            cnt, stride, mp_ncpus, smp_started);
1025                 return (EINVAL);
1026         }
1027         if (qgroup->tqg_adjusting) {
1028                 printf("taskqgroup_adjust failed: adjusting\n");
1029                 return (EBUSY);
1030         }
1031         qgroup->tqg_adjusting = 1;
1032         old_cnt = qgroup->tqg_cnt;
1033         mtx_unlock(&qgroup->tqg_lock);
1034         /*
1035          * Set up queue for tasks added before boot.
1036          */
1037         if (old_cnt == 0) {
1038                 LIST_SWAP(&gtask_head, &qgroup->tqg_queue[0].tgc_tasks,
1039                     grouptask, gt_list);
1040                 qgroup->tqg_queue[0].tgc_cnt = 0;
1041         }
1042
1043         /*
1044          * If new taskq threads have been added.
1045          */
1046         for (i = old_cnt; i < cnt; i++)
1047                 taskqgroup_cpu_create(qgroup, i);
1048         mtx_lock(&qgroup->tqg_lock);
1049         qgroup->tqg_cnt = cnt;
1050         qgroup->tqg_stride = stride;
1051
1052         /*
1053          * Adjust drivers to use new taskqs.
1054          */
1055         for (i = 0; i < old_cnt; i++) {
1056                 while ((gtask = LIST_FIRST(&qgroup->tqg_queue[i].tgc_tasks))) {
1057                         LIST_REMOVE(gtask, gt_list);
1058                         qgroup->tqg_queue[i].tgc_cnt--;
1059                         LIST_INSERT_HEAD(&gtask_head, gtask, gt_list);
1060                 }
1061         }
1062
1063         while ((gtask = LIST_FIRST(&gtask_head))) {
1064                 LIST_REMOVE(gtask, gt_list);
1065                 if (gtask->gt_cpu == -1)
1066                         qid = taskqgroup_find(qgroup, gtask->gt_uniq);
1067                 else {
1068                         for (i = 0; i < qgroup->tqg_cnt; i++)
1069                                 if (qgroup->tqg_queue[i].tgc_cpu == gtask->gt_cpu) {
1070                                         qid = i;
1071                                         break;
1072                                 }
1073                 }
1074                 qgroup->tqg_queue[qid].tgc_cnt++;
1075                 LIST_INSERT_HEAD(&qgroup->tqg_queue[qid].tgc_tasks, gtask,
1076                     gt_list);
1077                 gtask->gt_taskqueue = qgroup->tqg_queue[qid].tgc_taskq;
1078         }
1079         /*
1080          * Set new CPU and IRQ affinity
1081          */
1082         for (i = 0; i < cnt; i++) {
1083                 qgroup->tqg_queue[i].tgc_cpu = i * qgroup->tqg_stride;
1084                 CPU_ZERO(&mask);
1085                 CPU_SET(qgroup->tqg_queue[i].tgc_cpu, &mask);
1086                 LIST_FOREACH(gtask, &qgroup->tqg_queue[i].tgc_tasks, gt_list) {
1087                         if (gtask->gt_irq == -1)
1088                                 continue;
1089                         intr_setaffinity(gtask->gt_irq, &mask);
1090                 }
1091         }
1092         mtx_unlock(&qgroup->tqg_lock);
1093
1094         /*
1095          * If taskq thread count has been reduced.
1096          */
1097         for (i = cnt; i < old_cnt; i++)
1098                 taskqgroup_cpu_remove(qgroup, i);
1099
1100         mtx_lock(&qgroup->tqg_lock);
1101         qgroup->tqg_adjusting = 0;
1102
1103         taskqgroup_bind(qgroup);
1104
1105         return (0);
1106 }
1107
1108 int
1109 taskqgroup_adjust(struct taskqgroup *qgroup, int cpu, int stride)
1110 {
1111         int error;
1112
1113         mtx_lock(&qgroup->tqg_lock);
1114         error = _taskqgroup_adjust(qgroup, cpu, stride);
1115         mtx_unlock(&qgroup->tqg_lock);
1116
1117         return (error);
1118 }
1119
1120 struct taskqgroup *
1121 taskqgroup_create(char *name)
1122 {
1123         struct taskqgroup *qgroup;
1124
1125         qgroup = malloc(sizeof(*qgroup), M_TASKQUEUE, M_WAITOK | M_ZERO);
1126         mtx_init(&qgroup->tqg_lock, "taskqgroup", NULL, MTX_DEF);
1127         qgroup->tqg_name = name;
1128         LIST_INIT(&qgroup->tqg_queue[0].tgc_tasks);
1129
1130         return (qgroup);
1131 }
1132
1133 void
1134 taskqgroup_destroy(struct taskqgroup *qgroup)
1135 {
1136
1137 }