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/kernel.h>
  37 #include <sys/kthread.h>
  38 #include <sys/libkern.h>
  39 #include <sys/limits.h>
  40 #include <sys/lock.h>
  41 #include <sys/malloc.h>
  42 #include <sys/mutex.h>
  43 #include <sys/proc.h>
  44 #include <sys/sched.h>
  45 #include <sys/smp.h>
  46 #include <sys/gtaskqueue.h>
  47 #include <sys/unistd.h>
  48 #include <machine/stdarg.h>
  49
  50 static MALLOC_DEFINE(M_GTASKQUEUE, "gtaskqueue", "Group Task Queues");
  51 static void     gtaskqueue_thread_enqueue(void *);
  52 static void     gtaskqueue_thread_loop(void *arg);
  53 static int      task_is_running(struct gtaskqueue *queue, struct gtask *gtask);
  54 static void     gtaskqueue_drain_locked(struct gtaskqueue *queue, struct gtask *gtask);
  55
  56 TASKQGROUP_DEFINE(softirq, mp_ncpus, 1);
  57 TASKQGROUP_DEFINE(config, 1, 1);
  58
  59 struct gtaskqueue_busy {
  60         struct gtask    *tb_running;
  61         TAILQ_ENTRY(gtaskqueue_busy) tb_link;
  62 };
  63
  64 static struct gtask * const TB_DRAIN_WAITER = (struct gtask *)0x1;
  65
  66 typedef void (*gtaskqueue_enqueue_fn)(void *context);
  67
  68 struct gtaskqueue {
  69         STAILQ_HEAD(, gtask)    tq_queue;
  70         gtaskqueue_enqueue_fn   tq_enqueue;
  71         void                    *tq_context;
  72         char                    *tq_name;
  73         TAILQ_HEAD(, gtaskqueue_busy) tq_active;
  74         struct mtx              tq_mutex;
  75         struct thread           **tq_threads;
  76         int                     tq_tcount;
  77         int                     tq_spin;
  78         int                     tq_flags;
  79         int                     tq_callouts;
  80         taskqueue_callback_fn   tq_callbacks[TASKQUEUE_NUM_CALLBACKS];
  81         void                    *tq_cb_contexts[TASKQUEUE_NUM_CALLBACKS];
  82 };
  83
  84 #define TQ_FLAGS_ACTIVE         (1 << 0)
  85 #define TQ_FLAGS_BLOCKED        (1 << 1)
  86 #define TQ_FLAGS_UNLOCKED_ENQUEUE       (1 << 2)
  87
  88 #define DT_CALLOUT_ARMED        (1 << 0)
  89
  90 #define TQ_LOCK(tq)                                                     \
  91         do {                                                            \
  92                 if ((tq)->tq_spin)                                      \
  93                         mtx_lock_spin(&(tq)->tq_mutex);                 \
  94                 else                                                    \
  95                         mtx_lock(&(tq)->tq_mutex);                      \
  96         } while (0)
  97 #define TQ_ASSERT_LOCKED(tq)    mtx_assert(&(tq)->tq_mutex, MA_OWNED)
  98
  99 #define TQ_UNLOCK(tq)                                                   \
 100         do {                                                            \
 101                 if ((tq)->tq_spin)                                      \
 102                         mtx_unlock_spin(&(tq)->tq_mutex);               \
 103                 else                                                    \
 104                         mtx_unlock(&(tq)->tq_mutex);                    \
 105         } while (0)
 106 #define TQ_ASSERT_UNLOCKED(tq)  mtx_assert(&(tq)->tq_mutex, MA_NOTOWNED)
 107
 108 #ifdef INVARIANTS
 109 static void
 110 gtask_dump(struct gtask *gtask)
 111 {
 112         printf("gtask: %p ta_flags=%x ta_priority=%d ta_func=%p ta_context=%p\n",
 113                gtask, gtask->ta_flags, gtask->ta_priority, gtask->ta_func, gtask->ta_context);
 114 }
 115 #endif
 116
 117 static __inline int
 118 TQ_SLEEP(struct gtaskqueue *tq, void *p, struct mtx *m, int pri, const char *wm,
 119     int t)
 120 {
 121         if (tq->tq_spin)
 122                 return (msleep_spin(p, m, wm, t));
 123         return (msleep(p, m, pri, wm, t));
 124 }
 125
 126 static struct gtaskqueue *
 127 _gtaskqueue_create(const char *name, int mflags,
 128                  taskqueue_enqueue_fn enqueue, void *context,
 129                  int mtxflags, const char *mtxname __unused)
 130 {
 131         struct gtaskqueue *queue;
 132         char *tq_name;
 133
 134         tq_name = malloc(TASKQUEUE_NAMELEN, M_GTASKQUEUE, mflags | M_ZERO);
 135         if (!tq_name)
 136                 return (NULL);
 137
 138         snprintf(tq_name, TASKQUEUE_NAMELEN, "%s", (name) ? name : "taskqueue");
 139
 140         queue = malloc(sizeof(struct gtaskqueue), M_GTASKQUEUE, mflags | M_ZERO);
 141         if (!queue) {
 142                 free(tq_name, M_GTASKQUEUE);
 143                 return (NULL);
 144         }
 145
 146         STAILQ_INIT(&queue->tq_queue);
 147         TAILQ_INIT(&queue->tq_active);
 148         queue->tq_enqueue = enqueue;
 149         queue->tq_context = context;
 150         queue->tq_name = tq_name;
 151         queue->tq_spin = (mtxflags & MTX_SPIN) != 0;
 152         queue->tq_flags |= TQ_FLAGS_ACTIVE;
 153         if (enqueue == gtaskqueue_thread_enqueue)
 154                 queue->tq_flags |= TQ_FLAGS_UNLOCKED_ENQUEUE;
 155         mtx_init(&queue->tq_mutex, tq_name, NULL, mtxflags);
 156
 157         return (queue);
 158 }
 159
 160
 161 /*
 162  * Signal a taskqueue thread to terminate.
 163  */
 164 static void
 165 gtaskqueue_terminate(struct thread **pp, struct gtaskqueue *tq)
 166 {
 167
 168         while (tq->tq_tcount > 0 || tq->tq_callouts > 0) {
 169                 wakeup(tq);
 170                 TQ_SLEEP(tq, pp, &tq->tq_mutex, PWAIT, "taskqueue_destroy", 0);
 171         }
 172 }
 173
 174 static void
 175 gtaskqueue_free(struct gtaskqueue *queue)
 176 {
 177
 178         TQ_LOCK(queue);
 179         queue->tq_flags &= ~TQ_FLAGS_ACTIVE;
 180         gtaskqueue_terminate(queue->tq_threads, queue);
 181         KASSERT(TAILQ_EMPTY(&queue->tq_active), ("Tasks still running?"));
 182         KASSERT(queue->tq_callouts == 0, ("Armed timeout tasks"));
 183         mtx_destroy(&queue->tq_mutex);
 184         free(queue->tq_threads, M_GTASKQUEUE);
 185         free(queue->tq_name, M_GTASKQUEUE);
 186         free(queue, M_GTASKQUEUE);
 187 }
 188
 189 /*
 190  * Wait for all to complete, then prevent it from being enqueued
 191  */
 192 void
 193 grouptask_block(struct grouptask *grouptask)
 194 {
 195         struct gtaskqueue *queue = grouptask->gt_taskqueue;
 196         struct gtask *gtask = &grouptask->gt_task;
 197
 198 #ifdef INVARIANTS
 199         if (queue == NULL) {
 200                 gtask_dump(gtask);
 201                 panic("queue == NULL");
 202         }
 203 #endif
 204         TQ_LOCK(queue);
 205         gtask->ta_flags |= TASK_NOENQUEUE;
 206         gtaskqueue_drain_locked(queue, gtask);
 207         TQ_UNLOCK(queue);
 208 }
 209
 210 void
 211 grouptask_unblock(struct grouptask *grouptask)
 212 {
 213         struct gtaskqueue *queue = grouptask->gt_taskqueue;
 214         struct gtask *gtask = &grouptask->gt_task;
 215
 216 #ifdef INVARIANTS
 217         if (queue == NULL) {
 218                 gtask_dump(gtask);
 219                 panic("queue == NULL");
 220         }
 221 #endif
 222         TQ_LOCK(queue);
 223         gtask->ta_flags &= ~TASK_NOENQUEUE;
 224         TQ_UNLOCK(queue);
 225 }
 226
 227 int
 228 grouptaskqueue_enqueue(struct gtaskqueue *queue, struct gtask *gtask)
 229 {
 230 #ifdef INVARIANTS
 231         if (queue == NULL) {
 232                 gtask_dump(gtask);
 233                 panic("queue == NULL");
 234         }
 235 #endif
 236         TQ_LOCK(queue);
 237         if (gtask->ta_flags & TASK_ENQUEUED) {
 238                 TQ_UNLOCK(queue);
 239                 return (0);
 240         }
 241         if (gtask->ta_flags & TASK_NOENQUEUE) {
 242                 TQ_UNLOCK(queue);
 243                 return (EAGAIN);
 244         }
 245         STAILQ_INSERT_TAIL(&queue->tq_queue, gtask, ta_link);
 246         gtask->ta_flags |= TASK_ENQUEUED;
 247         TQ_UNLOCK(queue);
 248         if ((queue->tq_flags & TQ_FLAGS_BLOCKED) == 0)
 249                 queue->tq_enqueue(queue->tq_context);
 250         return (0);
 251 }
 252
 253 static void
 254 gtaskqueue_task_nop_fn(void *context)
 255 {
 256 }
 257
 258 /*
 259  * Block until all currently queued tasks in this taskqueue
 260  * have begun execution.  Tasks queued during execution of
 261  * this function are ignored.
 262  */
 263 static void
 264 gtaskqueue_drain_tq_queue(struct gtaskqueue *queue)
 265 {
 266         struct gtask t_barrier;
 267
 268         if (STAILQ_EMPTY(&queue->tq_queue))
 269                 return;
 270
 271         /*
 272          * Enqueue our barrier after all current tasks, but with
 273          * the highest priority so that newly queued tasks cannot
 274          * pass it.  Because of the high priority, we can not use
 275          * taskqueue_enqueue_locked directly (which drops the lock
 276          * anyway) so just insert it at tail while we have the
 277          * queue lock.
 278          */
 279         GTASK_INIT(&t_barrier, 0, USHRT_MAX, gtaskqueue_task_nop_fn, &t_barrier);
 280         STAILQ_INSERT_TAIL(&queue->tq_queue, &t_barrier, ta_link);
 281         t_barrier.ta_flags |= TASK_ENQUEUED;
 282
 283         /*
 284          * Once the barrier has executed, all previously queued tasks
 285          * have completed or are currently executing.
 286          */
 287         while (t_barrier.ta_flags & TASK_ENQUEUED)
 288                 TQ_SLEEP(queue, &t_barrier, &queue->tq_mutex, PWAIT, "-", 0);
 289 }
 290
 291 /*
 292  * Block until all currently executing tasks for this taskqueue
 293  * complete.  Tasks that begin execution during the execution
 294  * of this function are ignored.
 295  */
 296 static void
 297 gtaskqueue_drain_tq_active(struct gtaskqueue *queue)
 298 {
 299         struct gtaskqueue_busy tb_marker, *tb_first;
 300
 301         if (TAILQ_EMPTY(&queue->tq_active))
 302                 return;
 303
 304         /* Block taskq_terminate().*/
 305         queue->tq_callouts++;
 306
 307         /*
 308          * Wait for all currently executing taskqueue threads
 309          * to go idle.
 310          */
 311         tb_marker.tb_running = TB_DRAIN_WAITER;
 312         TAILQ_INSERT_TAIL(&queue->tq_active, &tb_marker, tb_link);
 313         while (TAILQ_FIRST(&queue->tq_active) != &tb_marker)
 314                 TQ_SLEEP(queue, &tb_marker, &queue->tq_mutex, PWAIT, "-", 0);
 315         TAILQ_REMOVE(&queue->tq_active, &tb_marker, tb_link);
 316
 317         /*
 318          * Wakeup any other drain waiter that happened to queue up
 319          * without any intervening active thread.
 320          */
 321         tb_first = TAILQ_FIRST(&queue->tq_active);
 322         if (tb_first != NULL && tb_first->tb_running == TB_DRAIN_WAITER)
 323                 wakeup(tb_first);
 324
 325         /* Release taskqueue_terminate(). */
 326         queue->tq_callouts--;
 327         if ((queue->tq_flags & TQ_FLAGS_ACTIVE) == 0)
 328                 wakeup_one(queue->tq_threads);
 329 }
 330
 331 void
 332 gtaskqueue_block(struct gtaskqueue *queue)
 333 {
 334
 335         TQ_LOCK(queue);
 336         queue->tq_flags |= TQ_FLAGS_BLOCKED;
 337         TQ_UNLOCK(queue);
 338 }
 339
 340 void
 341 gtaskqueue_unblock(struct gtaskqueue *queue)
 342 {
 343
 344         TQ_LOCK(queue);
 345         queue->tq_flags &= ~TQ_FLAGS_BLOCKED;
 346         if (!STAILQ_EMPTY(&queue->tq_queue))
 347                 queue->tq_enqueue(queue->tq_context);
 348         TQ_UNLOCK(queue);
 349 }
 350
 351 static void
 352 gtaskqueue_run_locked(struct gtaskqueue *queue)
 353 {
 354         struct gtaskqueue_busy tb;
 355         struct gtaskqueue_busy *tb_first;
 356         struct gtask *gtask;
 357
 358         KASSERT(queue != NULL, ("tq is NULL"));
 359         TQ_ASSERT_LOCKED(queue);
 360         tb.tb_running = NULL;
 361
 362         while (STAILQ_FIRST(&queue->tq_queue)) {
 363                 TAILQ_INSERT_TAIL(&queue->tq_active, &tb, tb_link);
 364
 365                 /*
 366                  * Carefully remove the first task from the queue and
 367                  * clear its TASK_ENQUEUED flag
 368                  */
 369                 gtask = STAILQ_FIRST(&queue->tq_queue);
 370                 KASSERT(gtask != NULL, ("task is NULL"));
 371                 STAILQ_REMOVE_HEAD(&queue->tq_queue, ta_link);
 372                 gtask->ta_flags &= ~TASK_ENQUEUED;
 373                 tb.tb_running = gtask;
 374                 TQ_UNLOCK(queue);
 375
 376                 KASSERT(gtask->ta_func != NULL, ("task->ta_func is NULL"));
 377                 gtask->ta_func(gtask->ta_context);
 378
 379                 TQ_LOCK(queue);
 380                 tb.tb_running = NULL;
 381                 wakeup(gtask);
 382
 383                 TAILQ_REMOVE(&queue->tq_active, &tb, tb_link);
 384                 tb_first = TAILQ_FIRST(&queue->tq_active);
 385                 if (tb_first != NULL &&
 386                     tb_first->tb_running == TB_DRAIN_WAITER)
 387                         wakeup(tb_first);
 388         }
 389 }
 390
 391 static int
 392 task_is_running(struct gtaskqueue *queue, struct gtask *gtask)
 393 {
 394         struct gtaskqueue_busy *tb;
 395
 396         TQ_ASSERT_LOCKED(queue);
 397         TAILQ_FOREACH(tb, &queue->tq_active, tb_link) {
 398                 if (tb->tb_running == gtask)
 399                         return (1);
 400         }
 401         return (0);
 402 }
 403
 404 static int
 405 gtaskqueue_cancel_locked(struct gtaskqueue *queue, struct gtask *gtask)
 406 {
 407
 408         if (gtask->ta_flags & TASK_ENQUEUED)
 409                 STAILQ_REMOVE(&queue->tq_queue, gtask, gtask, ta_link);
 410         gtask->ta_flags &= ~TASK_ENQUEUED;
 411         return (task_is_running(queue, gtask) ? EBUSY : 0);
 412 }
 413
 414 int
 415 gtaskqueue_cancel(struct gtaskqueue *queue, struct gtask *gtask)
 416 {
 417         int error;
 418
 419         TQ_LOCK(queue);
 420         error = gtaskqueue_cancel_locked(queue, gtask);
 421         TQ_UNLOCK(queue);
 422
 423         return (error);
 424 }
 425
 426 static void
 427 gtaskqueue_drain_locked(struct gtaskqueue *queue, struct gtask *gtask)
 428 {
 429         while ((gtask->ta_flags & TASK_ENQUEUED) || task_is_running(queue, gtask))
 430                 TQ_SLEEP(queue, gtask, &queue->tq_mutex, PWAIT, "-", 0);
 431 }
 432
 433 void
 434 gtaskqueue_drain(struct gtaskqueue *queue, struct gtask *gtask)
 435 {
 436
 437         if (!queue->tq_spin)
 438                 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, __func__);
 439
 440         TQ_LOCK(queue);
 441         gtaskqueue_drain_locked(queue, gtask);
 442         TQ_UNLOCK(queue);
 443 }
 444
 445 void
 446 gtaskqueue_drain_all(struct gtaskqueue *queue)
 447 {
 448
 449         if (!queue->tq_spin)
 450                 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, __func__);
 451
 452         TQ_LOCK(queue);
 453         gtaskqueue_drain_tq_queue(queue);
 454         gtaskqueue_drain_tq_active(queue);
 455         TQ_UNLOCK(queue);
 456 }
 457
 458 static int
 459 _gtaskqueue_start_threads(struct gtaskqueue **tqp, int count, int pri,
 460     cpuset_t *mask, const char *name, va_list ap)
 461 {
 462         char ktname[MAXCOMLEN + 1];
 463         struct thread *td;
 464         struct gtaskqueue *tq;
 465         int i, error;
 466
 467         if (count <= 0)
 468                 return (EINVAL);
 469
 470         vsnprintf(ktname, sizeof(ktname), name, ap);
 471         tq = *tqp;
 472
 473         tq->tq_threads = malloc(sizeof(struct thread *) * count, M_GTASKQUEUE,
 474             M_NOWAIT | M_ZERO);
 475         if (tq->tq_threads == NULL) {
 476                 printf("%s: no memory for %s threads\n", __func__, ktname);
 477                 return (ENOMEM);
 478         }
 479
 480         for (i = 0; i < count; i++) {
 481                 if (count == 1)
 482                         error = kthread_add(gtaskqueue_thread_loop, tqp, NULL,
 483                             &tq->tq_threads[i], RFSTOPPED, 0, "%s", ktname);
 484                 else
 485                         error = kthread_add(gtaskqueue_thread_loop, tqp, NULL,
 486                             &tq->tq_threads[i], RFSTOPPED, 0,
 487                             "%s_%d", ktname, i);
 488                 if (error) {
 489                         /* should be ok to continue, taskqueue_free will dtrt */
 490                         printf("%s: kthread_add(%s): error %d", __func__,
 491                             ktname, error);
 492                         tq->tq_threads[i] = NULL;               /* paranoid */
 493                 } else
 494                         tq->tq_tcount++;
 495         }
 496         for (i = 0; i < count; i++) {
 497                 if (tq->tq_threads[i] == NULL)
 498                         continue;
 499                 td = tq->tq_threads[i];
 500                 if (mask) {
 501                         error = cpuset_setthread(td->td_tid, mask);
 502                         /*
 503                          * Failing to pin is rarely an actual fatal error;
 504                          * it'll just affect performance.
 505                          */
 506                         if (error)
 507                                 printf("%s: curthread=%llu: can't pin; "
 508                                     "error=%d\n",
 509                                     __func__,
 510                                     (unsigned long long) td->td_tid,
 511                                     error);
 512                 }
 513                 thread_lock(td);
 514                 sched_prio(td, pri);
 515                 sched_add(td, SRQ_BORING);
 516                 thread_unlock(td);
 517         }
 518
 519         return (0);
 520 }
 521
 522 static int
 523 gtaskqueue_start_threads(struct gtaskqueue **tqp, int count, int pri,
 524     const char *name, ...)
 525 {
 526         va_list ap;
 527         int error;
 528
 529         va_start(ap, name);
 530         error = _gtaskqueue_start_threads(tqp, count, pri, NULL, name, ap);
 531         va_end(ap);
 532         return (error);
 533 }
 534
 535 static inline void
 536 gtaskqueue_run_callback(struct gtaskqueue *tq,
 537     enum taskqueue_callback_type cb_type)
 538 {
 539         taskqueue_callback_fn tq_callback;
 540
 541         TQ_ASSERT_UNLOCKED(tq);
 542         tq_callback = tq->tq_callbacks[cb_type];
 543         if (tq_callback != NULL)
 544                 tq_callback(tq->tq_cb_contexts[cb_type]);
 545 }
 546
 547 static void
 548 gtaskqueue_thread_loop(void *arg)
 549 {
 550         struct gtaskqueue **tqp, *tq;
 551
 552         tqp = arg;
 553         tq = *tqp;
 554         gtaskqueue_run_callback(tq, TASKQUEUE_CALLBACK_TYPE_INIT);
 555         TQ_LOCK(tq);
 556         while ((tq->tq_flags & TQ_FLAGS_ACTIVE) != 0) {
 557                 /* XXX ? */
 558                 gtaskqueue_run_locked(tq);
 559                 /*
 560                  * Because taskqueue_run() can drop tq_mutex, we need to
 561                  * check if the TQ_FLAGS_ACTIVE flag wasn't removed in the
 562                  * meantime, which means we missed a wakeup.
 563                  */
 564                 if ((tq->tq_flags & TQ_FLAGS_ACTIVE) == 0)
 565                         break;
 566                 TQ_SLEEP(tq, tq, &tq->tq_mutex, 0, "-", 0);
 567         }
 568         gtaskqueue_run_locked(tq);
 569         /*
 570          * This thread is on its way out, so just drop the lock temporarily
 571          * in order to call the shutdown callback.  This allows the callback
 572          * to look at the taskqueue, even just before it dies.
 573          */
 574         TQ_UNLOCK(tq);
 575         gtaskqueue_run_callback(tq, TASKQUEUE_CALLBACK_TYPE_SHUTDOWN);
 576         TQ_LOCK(tq);
 577
 578         /* rendezvous with thread that asked us to terminate */
 579         tq->tq_tcount--;
 580         wakeup_one(tq->tq_threads);
 581         TQ_UNLOCK(tq);
 582         kthread_exit();
 583 }
 584
 585 static void
 586 gtaskqueue_thread_enqueue(void *context)
 587 {
 588         struct gtaskqueue **tqp, *tq;
 589
 590         tqp = context;
 591         tq = *tqp;
 592         wakeup_one(tq);
 593 }
 594
 595
 596 static struct gtaskqueue *
 597 gtaskqueue_create_fast(const char *name, int mflags,
 598                  taskqueue_enqueue_fn enqueue, void *context)
 599 {
 600         return _gtaskqueue_create(name, mflags, enqueue, context,
 601                         MTX_SPIN, "fast_taskqueue");
 602 }
 603
 604
 605 struct taskqgroup_cpu {
 606         LIST_HEAD(, grouptask)  tgc_tasks;
 607         struct gtaskqueue       *tgc_taskq;
 608         int     tgc_cnt;
 609         int     tgc_cpu;
 610 };
 611
 612 struct taskqgroup {
 613         struct taskqgroup_cpu tqg_queue[MAXCPU];
 614         struct mtx      tqg_lock;
 615         const char *    tqg_name;
 616         int             tqg_adjusting;
 617         int             tqg_stride;
 618         int             tqg_cnt;
 619 };
 620
 621 struct taskq_bind_task {
 622         struct gtask bt_task;
 623         int     bt_cpuid;
 624 };
 625
 626 static void
 627 taskqgroup_cpu_create(struct taskqgroup *qgroup, int idx, int cpu)
 628 {
 629         struct taskqgroup_cpu *qcpu;
 630
 631         qcpu = &qgroup->tqg_queue[idx];
 632         LIST_INIT(&qcpu->tgc_tasks);
 633         qcpu->tgc_taskq = gtaskqueue_create_fast(NULL, M_WAITOK,
 634             taskqueue_thread_enqueue, &qcpu->tgc_taskq);
 635         gtaskqueue_start_threads(&qcpu->tgc_taskq, 1, PI_SOFT,
 636             "%s_%d", qgroup->tqg_name, idx);
 637         qcpu->tgc_cpu = cpu;
 638 }
 639
 640 static void
 641 taskqgroup_cpu_remove(struct taskqgroup *qgroup, int idx)
 642 {
 643
 644         gtaskqueue_free(qgroup->tqg_queue[idx].tgc_taskq);
 645 }
 646
 647 /*
 648  * Find the taskq with least # of tasks that doesn't currently have any
 649  * other queues from the uniq identifier.
 650  */
 651 static int
 652 taskqgroup_find(struct taskqgroup *qgroup, void *uniq)
 653 {
 654         struct grouptask *n;
 655         int i, idx, mincnt;
 656         int strict;
 657
 658         mtx_assert(&qgroup->tqg_lock, MA_OWNED);
 659         if (qgroup->tqg_cnt == 0)
 660                 return (0);
 661         idx = -1;
 662         mincnt = INT_MAX;
 663         /*
 664          * Two passes;  First scan for a queue with the least tasks that
 665          * does not already service this uniq id.  If that fails simply find
 666          * the queue with the least total tasks;
 667          */
 668         for (strict = 1; mincnt == INT_MAX; strict = 0) {
 669                 for (i = 0; i < qgroup->tqg_cnt; i++) {
 670                         if (qgroup->tqg_queue[i].tgc_cnt > mincnt)
 671                                 continue;
 672                         if (strict) {
 673                                 LIST_FOREACH(n,
 674                                     &qgroup->tqg_queue[i].tgc_tasks, gt_list)
 675                                         if (n->gt_uniq == uniq)
 676                                                 break;
 677                                 if (n != NULL)
 678                                         continue;
 679                         }
 680                         mincnt = qgroup->tqg_queue[i].tgc_cnt;
 681                         idx = i;
 682                 }
 683         }
 684         if (idx == -1)
 685                 panic("%s: failed to pick a qid.", __func__);
 686
 687         return (idx);
 688 }
 689
 690 /*
 691  * smp_started is unusable since it is not set for UP kernels or even for
 692  * SMP kernels when there is 1 CPU.  This is usually handled by adding a
 693  * (mp_ncpus == 1) test, but that would be broken here since we need to
 694  * to synchronize with the SI_SUB_SMP ordering.  Even in the pure SMP case
 695  * smp_started only gives a fuzzy ordering relative to SI_SUB_SMP.
 696  *
 697  * So maintain our own flag.  It must be set after all CPUs are started
 698  * and before SI_SUB_SMP:SI_ORDER_ANY so that the SYSINIT for delayed
 699  * adjustment is properly delayed.  SI_ORDER_FOURTH is clearly before
 700  * SI_ORDER_ANY and unclearly after the CPUs are started.  It would be
 701  * simpler for adjustment to pass a flag indicating if it is delayed.
 702  */
 703
 704 static int tqg_smp_started;
 705
 706 static void
 707 tqg_record_smp_started(void *arg)
 708 {
 709         tqg_smp_started = 1;
 710 }
 711
 712 SYSINIT(tqg_record_smp_started, SI_SUB_SMP, SI_ORDER_FOURTH,
 713         tqg_record_smp_started, NULL);
 714
 715 void
 716 taskqgroup_attach(struct taskqgroup *qgroup, struct grouptask *gtask,
 717     void *uniq, device_t dev, struct resource *irq, const char *name)
 718 {
 719         int cpu, qid, error;
 720
 721         gtask->gt_uniq = uniq;
 722         snprintf(gtask->gt_name, GROUPTASK_NAMELEN, "%s", name ? name : "grouptask");
 723         gtask->gt_dev = dev;
 724         gtask->gt_irq = irq;
 725         gtask->gt_cpu = -1;
 726         mtx_lock(&qgroup->tqg_lock);
 727         qid = taskqgroup_find(qgroup, uniq);
 728         qgroup->tqg_queue[qid].tgc_cnt++;
 729         LIST_INSERT_HEAD(&qgroup->tqg_queue[qid].tgc_tasks, gtask, gt_list);
 730         gtask->gt_taskqueue = qgroup->tqg_queue[qid].tgc_taskq;
 731         if (dev != NULL && irq != NULL && tqg_smp_started) {
 732                 cpu = qgroup->tqg_queue[qid].tgc_cpu;
 733                 gtask->gt_cpu = cpu;
 734                 mtx_unlock(&qgroup->tqg_lock);
 735                 error = bus_bind_intr(dev, irq, cpu);
 736                 if (error)
 737                         printf("%s: binding interrupt failed for %s: %d\n",
 738                             __func__, gtask->gt_name, error);
 739         } else
 740                 mtx_unlock(&qgroup->tqg_lock);
 741 }
 742
 743 static void
 744 taskqgroup_attach_deferred(struct taskqgroup *qgroup, struct grouptask *gtask)
 745 {
 746         int qid, cpu, error;
 747
 748         mtx_lock(&qgroup->tqg_lock);
 749         qid = taskqgroup_find(qgroup, gtask->gt_uniq);
 750         cpu = qgroup->tqg_queue[qid].tgc_cpu;
 751         if (gtask->gt_dev != NULL && gtask->gt_irq != NULL) {
 752                 mtx_unlock(&qgroup->tqg_lock);
 753                 error = bus_bind_intr(gtask->gt_dev, gtask->gt_irq, cpu);
 754                 mtx_lock(&qgroup->tqg_lock);
 755                 if (error)
 756                         printf("%s: binding interrupt failed for %s: %d\n",
 757                             __func__, gtask->gt_name, error);
 758
 759         }
 760         qgroup->tqg_queue[qid].tgc_cnt++;
 761         LIST_INSERT_HEAD(&qgroup->tqg_queue[qid].tgc_tasks, gtask, gt_list);
 762         MPASS(qgroup->tqg_queue[qid].tgc_taskq != NULL);
 763         gtask->gt_taskqueue = qgroup->tqg_queue[qid].tgc_taskq;
 764         mtx_unlock(&qgroup->tqg_lock);
 765 }
 766
 767 int
 768 taskqgroup_attach_cpu(struct taskqgroup *qgroup, struct grouptask *gtask,
 769     void *uniq, int cpu, device_t dev, struct resource *irq, const char *name)
 770 {
 771         int i, qid, error;
 772
 773         qid = -1;
 774         gtask->gt_uniq = uniq;
 775         snprintf(gtask->gt_name, GROUPTASK_NAMELEN, "%s", name ? name : "grouptask");
 776         gtask->gt_dev = dev;
 777         gtask->gt_irq = irq;
 778         gtask->gt_cpu = cpu;
 779         mtx_lock(&qgroup->tqg_lock);
 780         if (tqg_smp_started) {
 781                 for (i = 0; i < qgroup->tqg_cnt; i++)
 782                         if (qgroup->tqg_queue[i].tgc_cpu == cpu) {
 783                                 qid = i;
 784                                 break;
 785                         }
 786                 if (qid == -1) {
 787                         mtx_unlock(&qgroup->tqg_lock);
 788                         printf("%s: qid not found for %s cpu=%d\n", __func__, gtask->gt_name, cpu);
 789                         return (EINVAL);
 790                 }
 791         } else
 792                 qid = 0;
 793         qgroup->tqg_queue[qid].tgc_cnt++;
 794         LIST_INSERT_HEAD(&qgroup->tqg_queue[qid].tgc_tasks, gtask, gt_list);
 795         gtask->gt_taskqueue = qgroup->tqg_queue[qid].tgc_taskq;
 796         cpu = qgroup->tqg_queue[qid].tgc_cpu;
 797         mtx_unlock(&qgroup->tqg_lock);
 798
 799         if (dev != NULL && irq != NULL && tqg_smp_started) {
 800                 error = bus_bind_intr(dev, irq, cpu);
 801                 if (error)
 802                         printf("%s: binding interrupt failed for %s: %d\n",
 803                             __func__, gtask->gt_name, error);
 804         }
 805         return (0);
 806 }
 807
 808 static int
 809 taskqgroup_attach_cpu_deferred(struct taskqgroup *qgroup, struct grouptask *gtask)
 810 {
 811         device_t dev;
 812         struct resource *irq;
 813         int cpu, error, i, qid;
 814
 815         qid = -1;
 816         dev = gtask->gt_dev;
 817         irq = gtask->gt_irq;
 818         cpu = gtask->gt_cpu;
 819         MPASS(tqg_smp_started);
 820         mtx_lock(&qgroup->tqg_lock);
 821         for (i = 0; i < qgroup->tqg_cnt; i++)
 822                 if (qgroup->tqg_queue[i].tgc_cpu == cpu) {
 823                         qid = i;
 824                         break;
 825                 }
 826         if (qid == -1) {
 827                 mtx_unlock(&qgroup->tqg_lock);
 828                 printf("%s: qid not found for %s cpu=%d\n", __func__, gtask->gt_name, cpu);
 829                 return (EINVAL);
 830         }
 831         qgroup->tqg_queue[qid].tgc_cnt++;
 832         LIST_INSERT_HEAD(&qgroup->tqg_queue[qid].tgc_tasks, gtask, gt_list);
 833         MPASS(qgroup->tqg_queue[qid].tgc_taskq != NULL);
 834         gtask->gt_taskqueue = qgroup->tqg_queue[qid].tgc_taskq;
 835         mtx_unlock(&qgroup->tqg_lock);
 836
 837         if (dev != NULL && irq != NULL) {
 838                 error = bus_bind_intr(dev, irq, cpu);
 839                 if (error)
 840                         printf("%s: binding interrupt failed for %s: %d\n",
 841                             __func__, gtask->gt_name, error);
 842         }
 843         return (0);
 844 }
 845
 846 void
 847 taskqgroup_detach(struct taskqgroup *qgroup, struct grouptask *gtask)
 848 {
 849         int i;
 850
 851         grouptask_block(gtask);
 852         mtx_lock(&qgroup->tqg_lock);
 853         for (i = 0; i < qgroup->tqg_cnt; i++)
 854                 if (qgroup->tqg_queue[i].tgc_taskq == gtask->gt_taskqueue)
 855                         break;
 856         if (i == qgroup->tqg_cnt)
 857                 panic("%s: task %s not in group", __func__, gtask->gt_name);
 858         qgroup->tqg_queue[i].tgc_cnt--;
 859         LIST_REMOVE(gtask, gt_list);
 860         mtx_unlock(&qgroup->tqg_lock);
 861         gtask->gt_taskqueue = NULL;
 862         gtask->gt_task.ta_flags &= ~TASK_NOENQUEUE;
 863 }
 864
 865 static void
 866 taskqgroup_binder(void *ctx)
 867 {
 868         struct taskq_bind_task *gtask = (struct taskq_bind_task *)ctx;
 869         cpuset_t mask;
 870         int error;
 871
 872         CPU_ZERO(&mask);
 873         CPU_SET(gtask->bt_cpuid, &mask);
 874         error = cpuset_setthread(curthread->td_tid, &mask);
 875         thread_lock(curthread);
 876         sched_bind(curthread, gtask->bt_cpuid);
 877         thread_unlock(curthread);
 878
 879         if (error)
 880                 printf("%s: binding curthread failed: %d\n", __func__, error);
 881         free(gtask, M_DEVBUF);
 882 }
 883
 884 static void
 885 taskqgroup_bind(struct taskqgroup *qgroup)
 886 {
 887         struct taskq_bind_task *gtask;
 888         int i;
 889
 890         /*
 891          * Bind taskqueue threads to specific CPUs, if they have been assigned
 892          * one.
 893          */
 894         if (qgroup->tqg_cnt == 1)
 895                 return;
 896
 897         for (i = 0; i < qgroup->tqg_cnt; i++) {
 898                 gtask = malloc(sizeof (*gtask), M_DEVBUF, M_WAITOK);
 899                 GTASK_INIT(&gtask->bt_task, 0, 0, taskqgroup_binder, gtask);
 900                 gtask->bt_cpuid = qgroup->tqg_queue[i].tgc_cpu;
 901                 grouptaskqueue_enqueue(qgroup->tqg_queue[i].tgc_taskq,
 902                     &gtask->bt_task);
 903         }
 904 }
 905
 906 static void
 907 taskqgroup_config_init(void *arg)
 908 {
 909         struct taskqgroup *qgroup = qgroup_config;
 910         LIST_HEAD(, grouptask) gtask_head = LIST_HEAD_INITIALIZER(NULL);
 911
 912         LIST_SWAP(&gtask_head, &qgroup->tqg_queue[0].tgc_tasks,
 913             grouptask, gt_list);
 914         qgroup->tqg_queue[0].tgc_cnt = 0;
 915         taskqgroup_cpu_create(qgroup, 0, 0);
 916
 917         qgroup->tqg_cnt = 1;
 918         qgroup->tqg_stride = 1;
 919 }
 920
 921 SYSINIT(taskqgroup_config_init, SI_SUB_TASKQ, SI_ORDER_SECOND,
 922         taskqgroup_config_init, NULL);
 923
 924 static int
 925 _taskqgroup_adjust(struct taskqgroup *qgroup, int cnt, int stride)
 926 {
 927         LIST_HEAD(, grouptask) gtask_head = LIST_HEAD_INITIALIZER(NULL);
 928         struct grouptask *gtask;
 929         int i, k, old_cnt, old_cpu, cpu;
 930
 931         mtx_assert(&qgroup->tqg_lock, MA_OWNED);
 932
 933         if (cnt < 1 || cnt * stride > mp_ncpus || !tqg_smp_started) {
 934                 printf("%s: failed cnt: %d stride: %d "
 935                     "mp_ncpus: %d tqg_smp_started: %d\n",
 936                     __func__, cnt, stride, mp_ncpus, tqg_smp_started);
 937                 return (EINVAL);
 938         }
 939         if (qgroup->tqg_adjusting) {
 940                 printf("%s failed: adjusting\n", __func__);
 941                 return (EBUSY);
 942         }
 943         qgroup->tqg_adjusting = 1;
 944         old_cnt = qgroup->tqg_cnt;
 945         old_cpu = 0;
 946         if (old_cnt < cnt)
 947                 old_cpu = qgroup->tqg_queue[old_cnt].tgc_cpu;
 948         mtx_unlock(&qgroup->tqg_lock);
 949         /*
 950          * Set up queue for tasks added before boot.
 951          */
 952         if (old_cnt == 0) {
 953                 LIST_SWAP(&gtask_head, &qgroup->tqg_queue[0].tgc_tasks,
 954                     grouptask, gt_list);
 955                 qgroup->tqg_queue[0].tgc_cnt = 0;
 956         }
 957
 958         /*
 959          * If new taskq threads have been added.
 960          */
 961         cpu = old_cpu;
 962         for (i = old_cnt; i < cnt; i++) {
 963                 taskqgroup_cpu_create(qgroup, i, cpu);
 964
 965                 for (k = 0; k < stride; k++)
 966                         cpu = CPU_NEXT(cpu);
 967         }
 968         mtx_lock(&qgroup->tqg_lock);
 969         qgroup->tqg_cnt = cnt;
 970         qgroup->tqg_stride = stride;
 971
 972         /*
 973          * Adjust drivers to use new taskqs.
 974          */
 975         for (i = 0; i < old_cnt; i++) {
 976                 while ((gtask = LIST_FIRST(&qgroup->tqg_queue[i].tgc_tasks))) {
 977                         LIST_REMOVE(gtask, gt_list);
 978                         qgroup->tqg_queue[i].tgc_cnt--;
 979                         LIST_INSERT_HEAD(&gtask_head, gtask, gt_list);
 980                 }
 981         }
 982         mtx_unlock(&qgroup->tqg_lock);
 983
 984         while ((gtask = LIST_FIRST(&gtask_head))) {
 985                 LIST_REMOVE(gtask, gt_list);
 986                 if (gtask->gt_cpu == -1)
 987                         taskqgroup_attach_deferred(qgroup, gtask);
 988                 else if (taskqgroup_attach_cpu_deferred(qgroup, gtask))
 989                         taskqgroup_attach_deferred(qgroup, gtask);
 990         }
 991
 992 #ifdef INVARIANTS
 993         mtx_lock(&qgroup->tqg_lock);
 994         for (i = 0; i < qgroup->tqg_cnt; i++) {
 995                 MPASS(qgroup->tqg_queue[i].tgc_taskq != NULL);
 996                 LIST_FOREACH(gtask, &qgroup->tqg_queue[i].tgc_tasks, gt_list)
 997                         MPASS(gtask->gt_taskqueue != NULL);
 998         }
 999         mtx_unlock(&qgroup->tqg_lock);
1000 #endif
1001         /*
1002          * If taskq thread count has been reduced.
1003          */
1004         for (i = cnt; i < old_cnt; i++)
1005                 taskqgroup_cpu_remove(qgroup, i);
1006
1007         taskqgroup_bind(qgroup);
1008
1009         mtx_lock(&qgroup->tqg_lock);
1010         qgroup->tqg_adjusting = 0;
1011
1012         return (0);
1013 }
1014
1015 int
1016 taskqgroup_adjust(struct taskqgroup *qgroup, int cnt, int stride)
1017 {
1018         int error;
1019
1020         mtx_lock(&qgroup->tqg_lock);
1021         error = _taskqgroup_adjust(qgroup, cnt, stride);
1022         mtx_unlock(&qgroup->tqg_lock);
1023
1024         return (error);
1025 }
1026
1027 struct taskqgroup *
1028 taskqgroup_create(const char *name)
1029 {
1030         struct taskqgroup *qgroup;
1031
1032         qgroup = malloc(sizeof(*qgroup), M_GTASKQUEUE, M_WAITOK | M_ZERO);
1033         mtx_init(&qgroup->tqg_lock, "taskqgroup", NULL, MTX_DEF);
1034         qgroup->tqg_name = name;
1035         LIST_INIT(&qgroup->tqg_queue[0].tgc_tasks);
1036
1037         return (qgroup);
1038 }
1039
1040 void
1041 taskqgroup_destroy(struct taskqgroup *qgroup)
1042 {
1043
1044 }
1045
1046 void
1047 taskqgroup_config_gtask_init(void *ctx, struct grouptask *gtask, gtask_fn_t *fn,
1048     const char *name)
1049 {
1050
1051         GROUPTASK_INIT(gtask, 0, fn, ctx);
1052         taskqgroup_attach(qgroup_config, gtask, gtask, NULL, NULL, name);
1053 }
1054
1055 void
1056 taskqgroup_config_gtask_deinit(struct grouptask *gtask)
1057 {
1058
1059         taskqgroup_detach(qgroup_config, gtask);
1060 }