sys/kern/subr_gtaskqueue.c

   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 struct gtaskqueue_busy {
  56         struct gtask    *tb_running;
  57         TAILQ_ENTRY(gtaskqueue_busy) tb_link;
  58 };
  59
  60 static struct gtask * const TB_DRAIN_WAITER = (struct gtask *)0x1;
  61
  62 struct gtaskqueue {
  63         STAILQ_HEAD(, gtask)    tq_queue;
  64         gtaskqueue_enqueue_fn   tq_enqueue;
  65         void                    *tq_context;
  66         char                    *tq_name;
  67         TAILQ_HEAD(, gtaskqueue_busy) tq_active;
  68         struct mtx              tq_mutex;
  69         struct thread           **tq_threads;
  70         int                     tq_tcount;
  71         int                     tq_spin;
  72         int                     tq_flags;
  73         int                     tq_callouts;
  74         taskqueue_callback_fn   tq_callbacks[TASKQUEUE_NUM_CALLBACKS];
  75         void                    *tq_cb_contexts[TASKQUEUE_NUM_CALLBACKS];
  76 };
  77
  78 #define TQ_FLAGS_ACTIVE         (1 << 0)
  79 #define TQ_FLAGS_BLOCKED        (1 << 1)
  80 #define TQ_FLAGS_UNLOCKED_ENQUEUE       (1 << 2)
  81
  82 #define DT_CALLOUT_ARMED        (1 << 0)
  83
  84 #define TQ_LOCK(tq)                                                     \
  85         do {                                                            \
  86                 if ((tq)->tq_spin)                                      \
  87                         mtx_lock_spin(&(tq)->tq_mutex);                 \
  88                 else                                                    \
  89                         mtx_lock(&(tq)->tq_mutex);                      \
  90         } while (0)
  91 #define TQ_ASSERT_LOCKED(tq)    mtx_assert(&(tq)->tq_mutex, MA_OWNED)
  92
  93 #define TQ_UNLOCK(tq)                                                   \
  94         do {                                                            \
  95                 if ((tq)->tq_spin)                                      \
  96                         mtx_unlock_spin(&(tq)->tq_mutex);               \
  97                 else                                                    \
  98                         mtx_unlock(&(tq)->tq_mutex);                    \
  99         } while (0)
 100 #define TQ_ASSERT_UNLOCKED(tq)  mtx_assert(&(tq)->tq_mutex, MA_NOTOWNED)
 101
 102 static __inline int
 103 TQ_SLEEP(struct gtaskqueue *tq, void *p, struct mtx *m, int pri, const char *wm,
 104     int t)
 105 {
 106         if (tq->tq_spin)
 107                 return (msleep_spin(p, m, wm, t));
 108         return (msleep(p, m, pri, wm, t));
 109 }
 110
 111 static struct gtaskqueue *
 112 _gtaskqueue_create(const char *name, int mflags,
 113                  taskqueue_enqueue_fn enqueue, void *context,
 114                  int mtxflags, const char *mtxname __unused)
 115 {
 116         struct gtaskqueue *queue;
 117         char *tq_name;
 118
 119         tq_name = malloc(TASKQUEUE_NAMELEN, M_GTASKQUEUE, mflags | M_ZERO);
 120         if (!tq_name)
 121                 return (NULL);
 122
 123         snprintf(tq_name, TASKQUEUE_NAMELEN, "%s", (name) ? name : "taskqueue");
 124
 125         queue = malloc(sizeof(struct gtaskqueue), M_GTASKQUEUE, mflags | M_ZERO);
 126         if (!queue)
 127                 return (NULL);
 128
 129         STAILQ_INIT(&queue->tq_queue);
 130         TAILQ_INIT(&queue->tq_active);
 131         queue->tq_enqueue = enqueue;
 132         queue->tq_context = context;
 133         queue->tq_name = tq_name;
 134         queue->tq_spin = (mtxflags & MTX_SPIN) != 0;
 135         queue->tq_flags |= TQ_FLAGS_ACTIVE;
 136         if (enqueue == gtaskqueue_thread_enqueue)
 137                 queue->tq_flags |= TQ_FLAGS_UNLOCKED_ENQUEUE;
 138         mtx_init(&queue->tq_mutex, tq_name, NULL, mtxflags);
 139
 140         return (queue);
 141 }
 142
 143
 144 /*
 145  * Signal a taskqueue thread to terminate.
 146  */
 147 static void
 148 gtaskqueue_terminate(struct thread **pp, struct gtaskqueue *tq)
 149 {
 150
 151         while (tq->tq_tcount > 0 || tq->tq_callouts > 0) {
 152                 wakeup(tq);
 153                 TQ_SLEEP(tq, pp, &tq->tq_mutex, PWAIT, "taskqueue_destroy", 0);
 154         }
 155 }
 156
 157 static void
 158 gtaskqueue_free(struct gtaskqueue *queue)
 159 {
 160
 161         TQ_LOCK(queue);
 162         queue->tq_flags &= ~TQ_FLAGS_ACTIVE;
 163         gtaskqueue_terminate(queue->tq_threads, queue);
 164         KASSERT(TAILQ_EMPTY(&queue->tq_active), ("Tasks still running?"));
 165         KASSERT(queue->tq_callouts == 0, ("Armed timeout tasks"));
 166         mtx_destroy(&queue->tq_mutex);
 167         free(queue->tq_threads, M_GTASKQUEUE);
 168         free(queue->tq_name, M_GTASKQUEUE);
 169         free(queue, M_GTASKQUEUE);
 170 }
 171
 172 int
 173 grouptaskqueue_enqueue(struct gtaskqueue *queue, struct gtask *gtask)
 174 {
 175         TQ_LOCK(queue);
 176         if (gtask->ta_flags & TASK_ENQUEUED) {
 177                 TQ_UNLOCK(queue);
 178                 return (0);
 179         }
 180         STAILQ_INSERT_TAIL(&queue->tq_queue, gtask, ta_link);
 181         gtask->ta_flags |= TASK_ENQUEUED;
 182         TQ_UNLOCK(queue);
 183         if ((queue->tq_flags & TQ_FLAGS_BLOCKED) == 0)
 184                 queue->tq_enqueue(queue->tq_context);
 185         return (0);
 186 }
 187
 188 static void
 189 gtaskqueue_task_nop_fn(void *context)
 190 {
 191 }
 192
 193 /*
 194  * Block until all currently queued tasks in this taskqueue
 195  * have begun execution.  Tasks queued during execution of
 196  * this function are ignored.
 197  */
 198 static void
 199 gtaskqueue_drain_tq_queue(struct gtaskqueue *queue)
 200 {
 201         struct gtask t_barrier;
 202
 203         if (STAILQ_EMPTY(&queue->tq_queue))
 204                 return;
 205
 206         /*
 207          * Enqueue our barrier after all current tasks, but with
 208          * the highest priority so that newly queued tasks cannot
 209          * pass it.  Because of the high priority, we can not use
 210          * taskqueue_enqueue_locked directly (which drops the lock
 211          * anyway) so just insert it at tail while we have the
 212          * queue lock.
 213          */
 214         GTASK_INIT(&t_barrier, 0, USHRT_MAX, gtaskqueue_task_nop_fn, &t_barrier);
 215         STAILQ_INSERT_TAIL(&queue->tq_queue, &t_barrier, ta_link);
 216         t_barrier.ta_flags |= TASK_ENQUEUED;
 217
 218         /*
 219          * Once the barrier has executed, all previously queued tasks
 220          * have completed or are currently executing.
 221          */
 222         while (t_barrier.ta_flags & TASK_ENQUEUED)
 223                 TQ_SLEEP(queue, &t_barrier, &queue->tq_mutex, PWAIT, "-", 0);
 224 }
 225
 226 /*
 227  * Block until all currently executing tasks for this taskqueue
 228  * complete.  Tasks that begin execution during the execution
 229  * of this function are ignored.
 230  */
 231 static void
 232 gtaskqueue_drain_tq_active(struct gtaskqueue *queue)
 233 {
 234         struct gtaskqueue_busy tb_marker, *tb_first;
 235
 236         if (TAILQ_EMPTY(&queue->tq_active))
 237                 return;
 238
 239         /* Block taskq_terminate().*/
 240         queue->tq_callouts++;
 241
 242         /*
 243          * Wait for all currently executing taskqueue threads
 244          * to go idle.
 245          */
 246         tb_marker.tb_running = TB_DRAIN_WAITER;
 247         TAILQ_INSERT_TAIL(&queue->tq_active, &tb_marker, tb_link);
 248         while (TAILQ_FIRST(&queue->tq_active) != &tb_marker)
 249                 TQ_SLEEP(queue, &tb_marker, &queue->tq_mutex, PWAIT, "-", 0);
 250         TAILQ_REMOVE(&queue->tq_active, &tb_marker, tb_link);
 251
 252         /*
 253          * Wakeup any other drain waiter that happened to queue up
 254          * without any intervening active thread.
 255          */
 256         tb_first = TAILQ_FIRST(&queue->tq_active);
 257         if (tb_first != NULL && tb_first->tb_running == TB_DRAIN_WAITER)
 258                 wakeup(tb_first);
 259
 260         /* Release taskqueue_terminate(). */
 261         queue->tq_callouts--;
 262         if ((queue->tq_flags & TQ_FLAGS_ACTIVE) == 0)
 263                 wakeup_one(queue->tq_threads);
 264 }
 265
 266 void
 267 gtaskqueue_block(struct gtaskqueue *queue)
 268 {
 269
 270         TQ_LOCK(queue);
 271         queue->tq_flags |= TQ_FLAGS_BLOCKED;
 272         TQ_UNLOCK(queue);
 273 }
 274
 275 void
 276 gtaskqueue_unblock(struct gtaskqueue *queue)
 277 {
 278
 279         TQ_LOCK(queue);
 280         queue->tq_flags &= ~TQ_FLAGS_BLOCKED;
 281         if (!STAILQ_EMPTY(&queue->tq_queue))
 282                 queue->tq_enqueue(queue->tq_context);
 283         TQ_UNLOCK(queue);
 284 }
 285
 286 static void
 287 gtaskqueue_run_locked(struct gtaskqueue *queue)
 288 {
 289         struct gtaskqueue_busy tb;
 290         struct gtaskqueue_busy *tb_first;
 291         struct gtask *gtask;
 292
 293         KASSERT(queue != NULL, ("tq is NULL"));
 294         TQ_ASSERT_LOCKED(queue);
 295         tb.tb_running = NULL;
 296
 297         while (STAILQ_FIRST(&queue->tq_queue)) {
 298                 TAILQ_INSERT_TAIL(&queue->tq_active, &tb, tb_link);
 299
 300                 /*
 301                  * Carefully remove the first task from the queue and
 302                  * clear its TASK_ENQUEUED flag
 303                  */
 304                 gtask = STAILQ_FIRST(&queue->tq_queue);
 305                 KASSERT(gtask != NULL, ("task is NULL"));
 306                 STAILQ_REMOVE_HEAD(&queue->tq_queue, ta_link);
 307                 gtask->ta_flags &= ~TASK_ENQUEUED;
 308                 tb.tb_running = gtask;
 309                 TQ_UNLOCK(queue);
 310
 311                 KASSERT(gtask->ta_func != NULL, ("task->ta_func is NULL"));
 312                 gtask->ta_func(gtask->ta_context);
 313
 314                 TQ_LOCK(queue);
 315                 tb.tb_running = NULL;
 316                 wakeup(gtask);
 317
 318                 TAILQ_REMOVE(&queue->tq_active, &tb, tb_link);
 319                 tb_first = TAILQ_FIRST(&queue->tq_active);
 320                 if (tb_first != NULL &&
 321                     tb_first->tb_running == TB_DRAIN_WAITER)
 322                         wakeup(tb_first);
 323         }
 324 }
 325
 326 static int
 327 task_is_running(struct gtaskqueue *queue, struct gtask *gtask)
 328 {
 329         struct gtaskqueue_busy *tb;
 330
 331         TQ_ASSERT_LOCKED(queue);
 332         TAILQ_FOREACH(tb, &queue->tq_active, tb_link) {
 333                 if (tb->tb_running == gtask)
 334                         return (1);
 335         }
 336         return (0);
 337 }
 338
 339 static int
 340 gtaskqueue_cancel_locked(struct gtaskqueue *queue, struct gtask *gtask)
 341 {
 342
 343         if (gtask->ta_flags & TASK_ENQUEUED)
 344                 STAILQ_REMOVE(&queue->tq_queue, gtask, gtask, ta_link);
 345         gtask->ta_flags &= ~TASK_ENQUEUED;
 346         return (task_is_running(queue, gtask) ? EBUSY : 0);
 347 }
 348
 349 int
 350 gtaskqueue_cancel(struct gtaskqueue *queue, struct gtask *gtask)
 351 {
 352         int error;
 353
 354         TQ_LOCK(queue);
 355         error = gtaskqueue_cancel_locked(queue, gtask);
 356         TQ_UNLOCK(queue);
 357
 358         return (error);
 359 }
 360
 361 void
 362 gtaskqueue_drain(struct gtaskqueue *queue, struct gtask *gtask)
 363 {
 364
 365         if (!queue->tq_spin)
 366                 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, __func__);
 367
 368         TQ_LOCK(queue);
 369         while ((gtask->ta_flags & TASK_ENQUEUED) || task_is_running(queue, gtask))
 370                 TQ_SLEEP(queue, gtask, &queue->tq_mutex, PWAIT, "-", 0);
 371         TQ_UNLOCK(queue);
 372 }
 373
 374 void
 375 gtaskqueue_drain_all(struct gtaskqueue *queue)
 376 {
 377
 378         if (!queue->tq_spin)
 379                 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, __func__);
 380
 381         TQ_LOCK(queue);
 382         gtaskqueue_drain_tq_queue(queue);
 383         gtaskqueue_drain_tq_active(queue);
 384         TQ_UNLOCK(queue);
 385 }
 386
 387 static int
 388 _gtaskqueue_start_threads(struct gtaskqueue **tqp, int count, int pri,
 389     cpuset_t *mask, const char *name, va_list ap)
 390 {
 391         char ktname[MAXCOMLEN + 1];
 392         struct thread *td;
 393         struct gtaskqueue *tq;
 394         int i, error;
 395
 396         if (count <= 0)
 397                 return (EINVAL);
 398
 399         vsnprintf(ktname, sizeof(ktname), name, ap);
 400         tq = *tqp;
 401
 402         tq->tq_threads = malloc(sizeof(struct thread *) * count, M_GTASKQUEUE,
 403             M_NOWAIT | M_ZERO);
 404         if (tq->tq_threads == NULL) {
 405                 printf("%s: no memory for %s threads\n", __func__, ktname);
 406                 return (ENOMEM);
 407         }
 408
 409         for (i = 0; i < count; i++) {
 410                 if (count == 1)
 411                         error = kthread_add(gtaskqueue_thread_loop, tqp, NULL,
 412                             &tq->tq_threads[i], RFSTOPPED, 0, "%s", ktname);
 413                 else
 414                         error = kthread_add(gtaskqueue_thread_loop, tqp, NULL,
 415                             &tq->tq_threads[i], RFSTOPPED, 0,
 416                             "%s_%d", ktname, i);
 417                 if (error) {
 418                         /* should be ok to continue, taskqueue_free will dtrt */
 419                         printf("%s: kthread_add(%s): error %d", __func__,
 420                             ktname, error);
 421                         tq->tq_threads[i] = NULL;               /* paranoid */
 422                 } else
 423                         tq->tq_tcount++;
 424         }
 425         for (i = 0; i < count; i++) {
 426                 if (tq->tq_threads[i] == NULL)
 427                         continue;
 428                 td = tq->tq_threads[i];
 429                 if (mask) {
 430                         error = cpuset_setthread(td->td_tid, mask);
 431                         /*
 432                          * Failing to pin is rarely an actual fatal error;
 433                          * it'll just affect performance.
 434                          */
 435                         if (error)
 436                                 printf("%s: curthread=%llu: can't pin; "
 437                                     "error=%d\n",
 438                                     __func__,
 439                                     (unsigned long long) td->td_tid,
 440                                     error);
 441                 }
 442                 thread_lock(td);
 443                 sched_prio(td, pri);
 444                 sched_add(td, SRQ_BORING);
 445                 thread_unlock(td);
 446         }
 447
 448         return (0);
 449 }
 450
 451 static int
 452 gtaskqueue_start_threads(struct gtaskqueue **tqp, int count, int pri,
 453     const char *name, ...)
 454 {
 455         va_list ap;
 456         int error;
 457
 458         va_start(ap, name);
 459         error = _gtaskqueue_start_threads(tqp, count, pri, NULL, name, ap);
 460         va_end(ap);
 461         return (error);
 462 }
 463
 464 static inline void
 465 gtaskqueue_run_callback(struct gtaskqueue *tq,
 466     enum taskqueue_callback_type cb_type)
 467 {
 468         taskqueue_callback_fn tq_callback;
 469
 470         TQ_ASSERT_UNLOCKED(tq);
 471         tq_callback = tq->tq_callbacks[cb_type];
 472         if (tq_callback != NULL)
 473                 tq_callback(tq->tq_cb_contexts[cb_type]);
 474 }
 475
 476 static void
 477 gtaskqueue_thread_loop(void *arg)
 478 {
 479         struct gtaskqueue **tqp, *tq;
 480
 481         tqp = arg;
 482         tq = *tqp;
 483         gtaskqueue_run_callback(tq, TASKQUEUE_CALLBACK_TYPE_INIT);
 484         TQ_LOCK(tq);
 485         while ((tq->tq_flags & TQ_FLAGS_ACTIVE) != 0) {
 486                 /* XXX ? */
 487                 gtaskqueue_run_locked(tq);
 488                 /*
 489                  * Because taskqueue_run() can drop tq_mutex, we need to
 490                  * check if the TQ_FLAGS_ACTIVE flag wasn't removed in the
 491                  * meantime, which means we missed a wakeup.
 492                  */
 493                 if ((tq->tq_flags & TQ_FLAGS_ACTIVE) == 0)
 494                         break;
 495                 TQ_SLEEP(tq, tq, &tq->tq_mutex, 0, "-", 0);
 496         }
 497         gtaskqueue_run_locked(tq);
 498         /*
 499          * This thread is on its way out, so just drop the lock temporarily
 500          * in order to call the shutdown callback.  This allows the callback
 501          * to look at the taskqueue, even just before it dies.
 502          */
 503         TQ_UNLOCK(tq);
 504         gtaskqueue_run_callback(tq, TASKQUEUE_CALLBACK_TYPE_SHUTDOWN);
 505         TQ_LOCK(tq);
 506
 507         /* rendezvous with thread that asked us to terminate */
 508         tq->tq_tcount--;
 509         wakeup_one(tq->tq_threads);
 510         TQ_UNLOCK(tq);
 511         kthread_exit();
 512 }
 513
 514 static void
 515 gtaskqueue_thread_enqueue(void *context)
 516 {
 517         struct gtaskqueue **tqp, *tq;
 518
 519         tqp = context;
 520         tq = *tqp;
 521         wakeup_one(tq);
 522 }
 523
 524
 525 static struct gtaskqueue *
 526 gtaskqueue_create_fast(const char *name, int mflags,
 527                  taskqueue_enqueue_fn enqueue, void *context)
 528 {
 529         return _gtaskqueue_create(name, mflags, enqueue, context,
 530                         MTX_SPIN, "fast_taskqueue");
 531 }
 532
 533
 534 struct taskqgroup_cpu {
 535         LIST_HEAD(, grouptask)  tgc_tasks;
 536         struct gtaskqueue       *tgc_taskq;
 537         int     tgc_cnt;
 538         int     tgc_cpu;
 539 };
 540
 541 struct taskqgroup {
 542         struct taskqgroup_cpu tqg_queue[MAXCPU];
 543         struct mtx      tqg_lock;
 544         char *          tqg_name;
 545         int             tqg_adjusting;
 546         int             tqg_stride;
 547         int             tqg_cnt;
 548 };
 549
 550 struct taskq_bind_task {
 551         struct gtask bt_task;
 552         int     bt_cpuid;
 553 };
 554
 555 static void
 556 taskqgroup_cpu_create(struct taskqgroup *qgroup, int idx, int cpu)
 557 {
 558         struct taskqgroup_cpu *qcpu;
 559
 560         qcpu = &qgroup->tqg_queue[idx];
 561         LIST_INIT(&qcpu->tgc_tasks);
 562         qcpu->tgc_taskq = gtaskqueue_create_fast(NULL, M_WAITOK,
 563             taskqueue_thread_enqueue, &qcpu->tgc_taskq);
 564         gtaskqueue_start_threads(&qcpu->tgc_taskq, 1, PI_SOFT,
 565             "%s_%d", qgroup->tqg_name, idx);
 566         qcpu->tgc_cpu = cpu;
 567 }
 568
 569 static void
 570 taskqgroup_cpu_remove(struct taskqgroup *qgroup, int idx)
 571 {
 572
 573         gtaskqueue_free(qgroup->tqg_queue[idx].tgc_taskq);
 574 }
 575
 576 /*
 577  * Find the taskq with least # of tasks that doesn't currently have any
 578  * other queues from the uniq identifier.
 579  */
 580 static int
 581 taskqgroup_find(struct taskqgroup *qgroup, void *uniq)
 582 {
 583         struct grouptask *n;
 584         int i, idx, mincnt;
 585         int strict;
 586
 587         mtx_assert(&qgroup->tqg_lock, MA_OWNED);
 588         if (qgroup->tqg_cnt == 0)
 589                 return (0);
 590         idx = -1;
 591         mincnt = INT_MAX;
 592         /*
 593          * Two passes;  First scan for a queue with the least tasks that
 594          * does not already service this uniq id.  If that fails simply find
 595          * the queue with the least total tasks;
 596          */
 597         for (strict = 1; mincnt == INT_MAX; strict = 0) {
 598                 for (i = 0; i < qgroup->tqg_cnt; i++) {
 599                         if (qgroup->tqg_queue[i].tgc_cnt > mincnt)
 600                                 continue;
 601                         if (strict) {
 602                                 LIST_FOREACH(n,
 603                                     &qgroup->tqg_queue[i].tgc_tasks, gt_list)
 604                                         if (n->gt_uniq == uniq)
 605                                                 break;
 606                                 if (n != NULL)
 607                                         continue;
 608                         }
 609                         mincnt = qgroup->tqg_queue[i].tgc_cnt;
 610                         idx = i;
 611                 }
 612         }
 613         if (idx == -1)
 614                 panic("taskqgroup_find: Failed to pick a qid.");
 615
 616         return (idx);
 617 }
 618
 619 void
 620 taskqgroup_attach(struct taskqgroup *qgroup, struct grouptask *gtask,
 621     void *uniq, int irq, char *name)
 622 {
 623         cpuset_t mask;
 624         int qid;
 625
 626         gtask->gt_uniq = uniq;
 627         gtask->gt_name = name;
 628         gtask->gt_irq = irq;
 629         gtask->gt_cpu = -1;
 630         mtx_lock(&qgroup->tqg_lock);
 631         qid = taskqgroup_find(qgroup, uniq);
 632         qgroup->tqg_queue[qid].tgc_cnt++;
 633         LIST_INSERT_HEAD(&qgroup->tqg_queue[qid].tgc_tasks, gtask, gt_list);
 634         gtask->gt_taskqueue = qgroup->tqg_queue[qid].tgc_taskq;
 635         if (irq != -1 && smp_started) {
 636                 gtask->gt_cpu = qgroup->tqg_queue[qid].tgc_cpu;
 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 static void
 646 taskqgroup_attach_deferred(struct taskqgroup *qgroup, struct grouptask *gtask)
 647 {
 648         cpuset_t mask;
 649         int qid, cpu;
 650
 651         mtx_lock(&qgroup->tqg_lock);
 652         qid = taskqgroup_find(qgroup, gtask->gt_uniq);
 653         cpu = qgroup->tqg_queue[qid].tgc_cpu;
 654         if (gtask->gt_irq != -1) {
 655                 mtx_unlock(&qgroup->tqg_lock);
 656
 657                 CPU_ZERO(&mask);
 658                 CPU_SET(cpu, &mask);
 659                 intr_setaffinity(gtask->gt_irq, &mask);
 660
 661                 mtx_lock(&qgroup->tqg_lock);
 662         }
 663         qgroup->tqg_queue[qid].tgc_cnt++;
 664
 665         LIST_INSERT_HEAD(&qgroup->tqg_queue[qid].tgc_tasks, gtask,
 666                          gt_list);
 667         MPASS(qgroup->tqg_queue[qid].tgc_taskq != NULL);
 668         gtask->gt_taskqueue = qgroup->tqg_queue[qid].tgc_taskq;
 669         mtx_unlock(&qgroup->tqg_lock);
 670 }
 671
 672 int
 673 taskqgroup_attach_cpu(struct taskqgroup *qgroup, struct grouptask *gtask,
 674         void *uniq, int cpu, int irq, char *name)
 675 {
 676         cpuset_t mask;
 677         int i, qid;
 678
 679         qid = -1;
 680         gtask->gt_uniq = uniq;
 681         gtask->gt_name = name;
 682         gtask->gt_irq = irq;
 683         gtask->gt_cpu = cpu;
 684         mtx_lock(&qgroup->tqg_lock);
 685         if (smp_started) {
 686                 for (i = 0; i < qgroup->tqg_cnt; i++)
 687                         if (qgroup->tqg_queue[i].tgc_cpu == cpu) {
 688                                 qid = i;
 689                                 break;
 690                         }
 691                 if (qid == -1) {
 692                         mtx_unlock(&qgroup->tqg_lock);
 693                         return (EINVAL);
 694                 }
 695         } else
 696                 qid = 0;
 697         qgroup->tqg_queue[qid].tgc_cnt++;
 698         LIST_INSERT_HEAD(&qgroup->tqg_queue[qid].tgc_tasks, gtask, gt_list);
 699         gtask->gt_taskqueue = qgroup->tqg_queue[qid].tgc_taskq;
 700         cpu = qgroup->tqg_queue[qid].tgc_cpu;
 701         mtx_unlock(&qgroup->tqg_lock);
 702
 703         CPU_ZERO(&mask);
 704         CPU_SET(cpu, &mask);
 705         if (irq != -1 && smp_started)
 706                 intr_setaffinity(irq, &mask);
 707         return (0);
 708 }
 709
 710 static int
 711 taskqgroup_attach_cpu_deferred(struct taskqgroup *qgroup, struct grouptask *gtask)
 712 {
 713         cpuset_t mask;
 714         int i, qid, irq, cpu;
 715
 716         qid = -1;
 717         irq = gtask->gt_irq;
 718         cpu = gtask->gt_cpu;
 719         MPASS(smp_started);
 720         mtx_lock(&qgroup->tqg_lock);
 721         for (i = 0; i < qgroup->tqg_cnt; i++)
 722                 if (qgroup->tqg_queue[i].tgc_cpu == cpu) {
 723                         qid = i;
 724                         break;
 725                 }
 726         if (qid == -1) {
 727                 mtx_unlock(&qgroup->tqg_lock);
 728                 return (EINVAL);
 729         }
 730         qgroup->tqg_queue[qid].tgc_cnt++;
 731         LIST_INSERT_HEAD(&qgroup->tqg_queue[qid].tgc_tasks, gtask, gt_list);
 732         MPASS(qgroup->tqg_queue[qid].tgc_taskq != NULL);
 733         gtask->gt_taskqueue = qgroup->tqg_queue[qid].tgc_taskq;
 734         mtx_unlock(&qgroup->tqg_lock);
 735
 736         CPU_ZERO(&mask);
 737         CPU_SET(cpu, &mask);
 738
 739         if (irq != -1)
 740                 intr_setaffinity(irq, &mask);
 741         return (0);
 742 }
 743
 744 void
 745 taskqgroup_detach(struct taskqgroup *qgroup, struct grouptask *gtask)
 746 {
 747         int i;
 748
 749         mtx_lock(&qgroup->tqg_lock);
 750         for (i = 0; i < qgroup->tqg_cnt; i++)
 751                 if (qgroup->tqg_queue[i].tgc_taskq == gtask->gt_taskqueue)
 752                         break;
 753         if (i == qgroup->tqg_cnt)
 754                 panic("taskqgroup_detach: task not in group\n");
 755         qgroup->tqg_queue[i].tgc_cnt--;
 756         LIST_REMOVE(gtask, gt_list);
 757         mtx_unlock(&qgroup->tqg_lock);
 758         gtask->gt_taskqueue = NULL;
 759 }
 760
 761 static void
 762 taskqgroup_binder(void *ctx)
 763 {
 764         struct taskq_bind_task *gtask = (struct taskq_bind_task *)ctx;
 765         cpuset_t mask;
 766         int error;
 767
 768         CPU_ZERO(&mask);
 769         CPU_SET(gtask->bt_cpuid, &mask);
 770         error = cpuset_setthread(curthread->td_tid, &mask);
 771         thread_lock(curthread);
 772         sched_bind(curthread, gtask->bt_cpuid);
 773         thread_unlock(curthread);
 774
 775         if (error)
 776                 printf("taskqgroup_binder: setaffinity failed: %d\n",
 777                     error);
 778         free(gtask, M_DEVBUF);
 779 }
 780
 781 static void
 782 taskqgroup_bind(struct taskqgroup *qgroup)
 783 {
 784         struct taskq_bind_task *gtask;
 785         int i;
 786
 787         /*
 788          * Bind taskqueue threads to specific CPUs, if they have been assigned
 789          * one.
 790          */
 791         if (qgroup->tqg_cnt == 1)
 792                 return;
 793
 794         for (i = 0; i < qgroup->tqg_cnt; i++) {
 795                 gtask = malloc(sizeof (*gtask), M_DEVBUF, M_WAITOK);
 796                 GTASK_INIT(&gtask->bt_task, 0, 0, taskqgroup_binder, gtask);
 797                 gtask->bt_cpuid = qgroup->tqg_queue[i].tgc_cpu;
 798                 grouptaskqueue_enqueue(qgroup->tqg_queue[i].tgc_taskq,
 799                     &gtask->bt_task);
 800         }
 801 }
 802
 803 static int
 804 _taskqgroup_adjust(struct taskqgroup *qgroup, int cnt, int stride)
 805 {
 806         LIST_HEAD(, grouptask) gtask_head = LIST_HEAD_INITIALIZER(NULL);
 807         struct grouptask *gtask;
 808         int i, k, old_cnt, old_cpu, cpu;
 809
 810         mtx_assert(&qgroup->tqg_lock, MA_OWNED);
 811
 812         if (cnt < 1 || cnt * stride > mp_ncpus || !smp_started) {
 813                 printf("taskqgroup_adjust failed cnt: %d stride: %d mp_ncpus: %d smp_started: %d\n",
 814                            cnt, stride, mp_ncpus, smp_started);
 815                 return (EINVAL);
 816         }
 817         if (qgroup->tqg_adjusting) {
 818                 printf("taskqgroup_adjust failed: adjusting\n");
 819                 return (EBUSY);
 820         }
 821         qgroup->tqg_adjusting = 1;
 822         old_cnt = qgroup->tqg_cnt;
 823         old_cpu = 0;
 824         if (old_cnt < cnt)
 825                 old_cpu = qgroup->tqg_queue[old_cnt].tgc_cpu;
 826         mtx_unlock(&qgroup->tqg_lock);
 827         /*
 828          * Set up queue for tasks added before boot.
 829          */
 830         if (old_cnt == 0) {
 831                 LIST_SWAP(&gtask_head, &qgroup->tqg_queue[0].tgc_tasks,
 832                     grouptask, gt_list);
 833                 qgroup->tqg_queue[0].tgc_cnt = 0;
 834         }
 835
 836         /*
 837          * If new taskq threads have been added.
 838          */
 839         cpu = old_cpu;
 840         for (i = old_cnt; i < cnt; i++) {
 841                 taskqgroup_cpu_create(qgroup, i, cpu);
 842
 843                 for (k = 0; k < stride; k++)
 844                         cpu = CPU_NEXT(cpu);
 845         }
 846         mtx_lock(&qgroup->tqg_lock);
 847         qgroup->tqg_cnt = cnt;
 848         qgroup->tqg_stride = stride;
 849
 850         /*
 851          * Adjust drivers to use new taskqs.
 852          */
 853         for (i = 0; i < old_cnt; i++) {
 854                 while ((gtask = LIST_FIRST(&qgroup->tqg_queue[i].tgc_tasks))) {
 855                         LIST_REMOVE(gtask, gt_list);
 856                         qgroup->tqg_queue[i].tgc_cnt--;
 857                         LIST_INSERT_HEAD(&gtask_head, gtask, gt_list);
 858                 }
 859         }
 860         mtx_unlock(&qgroup->tqg_lock);
 861
 862         while ((gtask = LIST_FIRST(&gtask_head))) {
 863                 LIST_REMOVE(gtask, gt_list);
 864                 if (gtask->gt_cpu == -1)
 865                         taskqgroup_attach_deferred(qgroup, gtask);
 866                 else if (taskqgroup_attach_cpu_deferred(qgroup, gtask))
 867                         taskqgroup_attach_deferred(qgroup, gtask);
 868         }
 869
 870 #ifdef INVARIANTS
 871         mtx_lock(&qgroup->tqg_lock);
 872         for (i = 0; i < qgroup->tqg_cnt; i++) {
 873                 MPASS(qgroup->tqg_queue[i].tgc_taskq != NULL);
 874                 LIST_FOREACH(gtask, &qgroup->tqg_queue[i].tgc_tasks, gt_list)
 875                         MPASS(gtask->gt_taskqueue != NULL);
 876         }
 877         mtx_unlock(&qgroup->tqg_lock);
 878 #endif
 879         /*
 880          * If taskq thread count has been reduced.
 881          */
 882         for (i = cnt; i < old_cnt; i++)
 883                 taskqgroup_cpu_remove(qgroup, i);
 884
 885         taskqgroup_bind(qgroup);
 886
 887         mtx_lock(&qgroup->tqg_lock);
 888         qgroup->tqg_adjusting = 0;
 889
 890         return (0);
 891 }
 892
 893 int
 894 taskqgroup_adjust(struct taskqgroup *qgroup, int cnt, int stride)
 895 {
 896         int error;
 897
 898         mtx_lock(&qgroup->tqg_lock);
 899         error = _taskqgroup_adjust(qgroup, cnt, stride);
 900         mtx_unlock(&qgroup->tqg_lock);
 901
 902         return (error);
 903 }
 904
 905 struct taskqgroup *
 906 taskqgroup_create(char *name)
 907 {
 908         struct taskqgroup *qgroup;
 909
 910         qgroup = malloc(sizeof(*qgroup), M_GTASKQUEUE, M_WAITOK | M_ZERO);
 911         mtx_init(&qgroup->tqg_lock, "taskqgroup", NULL, MTX_DEF);
 912         qgroup->tqg_name = name;
 913         LIST_INIT(&qgroup->tqg_queue[0].tgc_tasks);
 914
 915         return (qgroup);
 916 }
 917
 918 void
 919 taskqgroup_destroy(struct taskqgroup *qgroup)
 920 {
 921
 922 }