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
  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, old_cnt, qid;
 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         for (i = 0; i < cnt; i++) {
 810                 qgroup->tqg_queue[i].tgc_cpu = i * qgroup->tqg_stride;
 811                 CPU_ZERO(&mask);
 812                 CPU_SET(qgroup->tqg_queue[i].tgc_cpu, &mask);
 813                 LIST_FOREACH(gtask, &qgroup->tqg_queue[i].tgc_tasks, gt_list) {
 814                         if (gtask->gt_irq == -1)
 815                                 continue;
 816                         intr_setaffinity(gtask->gt_irq, &mask);
 817                 }
 818         }
 819         mtx_unlock(&qgroup->tqg_lock);
 820
 821         /*
 822          * If taskq thread count has been reduced.
 823          */
 824         for (i = cnt; i < old_cnt; i++)
 825                 taskqgroup_cpu_remove(qgroup, i);
 826
 827         mtx_lock(&qgroup->tqg_lock);
 828         qgroup->tqg_adjusting = 0;
 829
 830         taskqgroup_bind(qgroup);
 831
 832         return (0);
 833 }
 834
 835 int
 836 taskqgroup_adjust(struct taskqgroup *qgroup, int cpu, int stride)
 837 {
 838         int error;
 839
 840         mtx_lock(&qgroup->tqg_lock);
 841         error = _taskqgroup_adjust(qgroup, cpu, stride);
 842         mtx_unlock(&qgroup->tqg_lock);
 843
 844         return (error);
 845 }
 846
 847 struct taskqgroup *
 848 taskqgroup_create(char *name)
 849 {
 850         struct taskqgroup *qgroup;
 851
 852         qgroup = malloc(sizeof(*qgroup), M_GTASKQUEUE, M_WAITOK | M_ZERO);
 853         mtx_init(&qgroup->tqg_lock, "taskqgroup", NULL, MTX_DEF);
 854         qgroup->tqg_name = name;
 855         LIST_INIT(&qgroup->tqg_queue[0].tgc_tasks);
 856
 857         return (qgroup);
 858 }
 859
 860 void
 861 taskqgroup_destroy(struct taskqgroup *qgroup)
 862 {
 863
 864 }