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, "gtaskqueue", "Group Task Queues");
  52 static void     gtaskqueue_thread_enqueue(void *);
  53 static void     gtaskqueue_thread_loop(void *arg);
  54 static int      task_is_running(struct gtaskqueue *queue, struct gtask *gtask);
  55 static void     gtaskqueue_drain_locked(struct gtaskqueue *queue, struct gtask *gtask);
  56
  57 TASKQGROUP_DEFINE(softirq, mp_ncpus, 1);
  58 TASKQGROUP_DEFINE(config, 1, 1);
  59
  60 struct gtaskqueue_busy {
  61         struct gtask    *tb_running;
  62         TAILQ_ENTRY(gtaskqueue_busy) tb_link;
  63 };
  64
  65 static struct gtask * const TB_DRAIN_WAITER = (struct gtask *)0x1;
  66
  67 struct gtaskqueue {
  68         STAILQ_HEAD(, gtask)    tq_queue;
  69         gtaskqueue_enqueue_fn   tq_enqueue;
  70         void                    *tq_context;
  71         char                    *tq_name;
  72         TAILQ_HEAD(, gtaskqueue_busy) tq_active;
  73         struct mtx              tq_mutex;
  74         struct thread           **tq_threads;
  75         int                     tq_tcount;
  76         int                     tq_spin;
  77         int                     tq_flags;
  78         int                     tq_callouts;
  79         taskqueue_callback_fn   tq_callbacks[TASKQUEUE_NUM_CALLBACKS];
  80         void                    *tq_cb_contexts[TASKQUEUE_NUM_CALLBACKS];
  81 };
  82
  83 #define TQ_FLAGS_ACTIVE         (1 << 0)
  84 #define TQ_FLAGS_BLOCKED        (1 << 1)
  85 #define TQ_FLAGS_UNLOCKED_ENQUEUE       (1 << 2)
  86
  87 #define DT_CALLOUT_ARMED        (1 << 0)
  88
  89 #define TQ_LOCK(tq)                                                     \
  90         do {                                                            \
  91                 if ((tq)->tq_spin)                                      \
  92                         mtx_lock_spin(&(tq)->tq_mutex);                 \
  93                 else                                                    \
  94                         mtx_lock(&(tq)->tq_mutex);                      \
  95         } while (0)
  96 #define TQ_ASSERT_LOCKED(tq)    mtx_assert(&(tq)->tq_mutex, MA_OWNED)
  97
  98 #define TQ_UNLOCK(tq)                                                   \
  99         do {                                                            \
 100                 if ((tq)->tq_spin)                                      \
 101                         mtx_unlock_spin(&(tq)->tq_mutex);               \
 102                 else                                                    \
 103                         mtx_unlock(&(tq)->tq_mutex);                    \
 104         } while (0)
 105 #define TQ_ASSERT_UNLOCKED(tq)  mtx_assert(&(tq)->tq_mutex, MA_NOTOWNED)
 106
 107 #ifdef INVARIANTS
 108 static void
 109 gtask_dump(struct gtask *gtask)
 110 {
 111         printf("gtask: %p ta_flags=%x ta_priority=%d ta_func=%p ta_context=%p\n",
 112                gtask, gtask->ta_flags, gtask->ta_priority, gtask->ta_func, gtask->ta_context);
 113 }
 114 #endif
 115
 116 static __inline int
 117 TQ_SLEEP(struct gtaskqueue *tq, void *p, struct mtx *m, int pri, const char *wm,
 118     int t)
 119 {
 120         if (tq->tq_spin)
 121                 return (msleep_spin(p, m, wm, t));
 122         return (msleep(p, m, pri, wm, t));
 123 }
 124
 125 static struct gtaskqueue *
 126 _gtaskqueue_create(const char *name, int mflags,
 127                  taskqueue_enqueue_fn enqueue, void *context,
 128                  int mtxflags, const char *mtxname __unused)
 129 {
 130         struct gtaskqueue *queue;
 131         char *tq_name;
 132
 133         tq_name = malloc(TASKQUEUE_NAMELEN, M_GTASKQUEUE, mflags | M_ZERO);
 134         if (!tq_name)
 135                 return (NULL);
 136
 137         snprintf(tq_name, TASKQUEUE_NAMELEN, "%s", (name) ? name : "taskqueue");
 138
 139         queue = malloc(sizeof(struct gtaskqueue), M_GTASKQUEUE, mflags | M_ZERO);
 140         if (!queue) {
 141                 free(tq_name, M_GTASKQUEUE);
 142                 return (NULL);
 143         }
 144
 145         STAILQ_INIT(&queue->tq_queue);
 146         TAILQ_INIT(&queue->tq_active);
 147         queue->tq_enqueue = enqueue;
 148         queue->tq_context = context;
 149         queue->tq_name = tq_name;
 150         queue->tq_spin = (mtxflags & MTX_SPIN) != 0;
 151         queue->tq_flags |= TQ_FLAGS_ACTIVE;
 152         if (enqueue == gtaskqueue_thread_enqueue)
 153                 queue->tq_flags |= TQ_FLAGS_UNLOCKED_ENQUEUE;
 154         mtx_init(&queue->tq_mutex, tq_name, NULL, mtxflags);
 155
 156         return (queue);
 157 }
 158
 159
 160 /*
 161  * Signal a taskqueue thread to terminate.
 162  */
 163 static void
 164 gtaskqueue_terminate(struct thread **pp, struct gtaskqueue *tq)
 165 {
 166
 167         while (tq->tq_tcount > 0 || tq->tq_callouts > 0) {
 168                 wakeup(tq);
 169                 TQ_SLEEP(tq, pp, &tq->tq_mutex, PWAIT, "taskqueue_destroy", 0);
 170         }
 171 }
 172
 173 static void
 174 gtaskqueue_free(struct gtaskqueue *queue)
 175 {
 176
 177         TQ_LOCK(queue);
 178         queue->tq_flags &= ~TQ_FLAGS_ACTIVE;
 179         gtaskqueue_terminate(queue->tq_threads, queue);
 180         KASSERT(TAILQ_EMPTY(&queue->tq_active), ("Tasks still running?"));
 181         KASSERT(queue->tq_callouts == 0, ("Armed timeout tasks"));
 182         mtx_destroy(&queue->tq_mutex);
 183         free(queue->tq_threads, M_GTASKQUEUE);
 184         free(queue->tq_name, M_GTASKQUEUE);
 185         free(queue, M_GTASKQUEUE);
 186 }
 187
 188 /*
 189  * Wait for all to complete, then prevent it from being enqueued
 190  */
 191 void
 192 grouptask_block(struct grouptask *grouptask)
 193 {
 194         struct gtaskqueue *queue = grouptask->gt_taskqueue;
 195         struct gtask *gtask = &grouptask->gt_task;
 196
 197 #ifdef INVARIANTS
 198         if (queue == NULL) {
 199                 gtask_dump(gtask);
 200                 panic("queue == NULL");
 201         }
 202 #endif
 203         TQ_LOCK(queue);
 204         gtask->ta_flags |= TASK_NOENQUEUE;
 205         gtaskqueue_drain_locked(queue, gtask);
 206         TQ_UNLOCK(queue);
 207 }
 208
 209 void
 210 grouptask_unblock(struct grouptask *grouptask)
 211 {
 212         struct gtaskqueue *queue = grouptask->gt_taskqueue;
 213         struct gtask *gtask = &grouptask->gt_task;
 214
 215 #ifdef INVARIANTS
 216         if (queue == NULL) {
 217                 gtask_dump(gtask);
 218                 panic("queue == NULL");
 219         }
 220 #endif
 221         TQ_LOCK(queue);
 222         gtask->ta_flags &= ~TASK_NOENQUEUE;
 223         TQ_UNLOCK(queue);
 224 }
 225
 226 int
 227 grouptaskqueue_enqueue(struct gtaskqueue *queue, struct gtask *gtask)
 228 {
 229 #ifdef INVARIANTS
 230         if (queue == NULL) {
 231                 gtask_dump(gtask);
 232                 panic("queue == NULL");
 233         }
 234 #endif
 235         TQ_LOCK(queue);
 236         if (gtask->ta_flags & TASK_ENQUEUED) {
 237                 TQ_UNLOCK(queue);
 238                 return (0);
 239         }
 240         if (gtask->ta_flags & TASK_NOENQUEUE) {
 241                 TQ_UNLOCK(queue);
 242                 return (EAGAIN);
 243         }
 244         STAILQ_INSERT_TAIL(&queue->tq_queue, gtask, ta_link);
 245         gtask->ta_flags |= TASK_ENQUEUED;
 246         TQ_UNLOCK(queue);
 247         if ((queue->tq_flags & TQ_FLAGS_BLOCKED) == 0)
 248                 queue->tq_enqueue(queue->tq_context);
 249         return (0);
 250 }
 251
 252 static void
 253 gtaskqueue_task_nop_fn(void *context)
 254 {
 255 }
 256
 257 /*
 258  * Block until all currently queued tasks in this taskqueue
 259  * have begun execution.  Tasks queued during execution of
 260  * this function are ignored.
 261  */
 262 static void
 263 gtaskqueue_drain_tq_queue(struct gtaskqueue *queue)
 264 {
 265         struct gtask t_barrier;
 266
 267         if (STAILQ_EMPTY(&queue->tq_queue))
 268                 return;
 269
 270         /*
 271          * Enqueue our barrier after all current tasks, but with
 272          * the highest priority so that newly queued tasks cannot
 273          * pass it.  Because of the high priority, we can not use
 274          * taskqueue_enqueue_locked directly (which drops the lock
 275          * anyway) so just insert it at tail while we have the
 276          * queue lock.
 277          */
 278         GTASK_INIT(&t_barrier, 0, USHRT_MAX, gtaskqueue_task_nop_fn, &t_barrier);
 279         STAILQ_INSERT_TAIL(&queue->tq_queue, &t_barrier, ta_link);
 280         t_barrier.ta_flags |= TASK_ENQUEUED;
 281
 282         /*
 283          * Once the barrier has executed, all previously queued tasks
 284          * have completed or are currently executing.
 285          */
 286         while (t_barrier.ta_flags & TASK_ENQUEUED)
 287                 TQ_SLEEP(queue, &t_barrier, &queue->tq_mutex, PWAIT, "-", 0);
 288 }
 289
 290 /*
 291  * Block until all currently executing tasks for this taskqueue
 292  * complete.  Tasks that begin execution during the execution
 293  * of this function are ignored.
 294  */
 295 static void
 296 gtaskqueue_drain_tq_active(struct gtaskqueue *queue)
 297 {
 298         struct gtaskqueue_busy tb_marker, *tb_first;
 299
 300         if (TAILQ_EMPTY(&queue->tq_active))
 301                 return;
 302
 303         /* Block taskq_terminate().*/
 304         queue->tq_callouts++;
 305
 306         /*
 307          * Wait for all currently executing taskqueue threads
 308          * to go idle.
 309          */
 310         tb_marker.tb_running = TB_DRAIN_WAITER;
 311         TAILQ_INSERT_TAIL(&queue->tq_active, &tb_marker, tb_link);
 312         while (TAILQ_FIRST(&queue->tq_active) != &tb_marker)
 313                 TQ_SLEEP(queue, &tb_marker, &queue->tq_mutex, PWAIT, "-", 0);
 314         TAILQ_REMOVE(&queue->tq_active, &tb_marker, tb_link);
 315
 316         /*
 317          * Wakeup any other drain waiter that happened to queue up
 318          * without any intervening active thread.
 319          */
 320         tb_first = TAILQ_FIRST(&queue->tq_active);
 321         if (tb_first != NULL && tb_first->tb_running == TB_DRAIN_WAITER)
 322                 wakeup(tb_first);
 323
 324         /* Release taskqueue_terminate(). */
 325         queue->tq_callouts--;
 326         if ((queue->tq_flags & TQ_FLAGS_ACTIVE) == 0)
 327                 wakeup_one(queue->tq_threads);
 328 }
 329
 330 void
 331 gtaskqueue_block(struct gtaskqueue *queue)
 332 {
 333
 334         TQ_LOCK(queue);
 335         queue->tq_flags |= TQ_FLAGS_BLOCKED;
 336         TQ_UNLOCK(queue);
 337 }
 338
 339 void
 340 gtaskqueue_unblock(struct gtaskqueue *queue)
 341 {
 342
 343         TQ_LOCK(queue);
 344         queue->tq_flags &= ~TQ_FLAGS_BLOCKED;
 345         if (!STAILQ_EMPTY(&queue->tq_queue))
 346                 queue->tq_enqueue(queue->tq_context);
 347         TQ_UNLOCK(queue);
 348 }
 349
 350 static void
 351 gtaskqueue_run_locked(struct gtaskqueue *queue)
 352 {
 353         struct gtaskqueue_busy tb;
 354         struct gtaskqueue_busy *tb_first;
 355         struct gtask *gtask;
 356
 357         KASSERT(queue != NULL, ("tq is NULL"));
 358         TQ_ASSERT_LOCKED(queue);
 359         tb.tb_running = NULL;
 360
 361         while (STAILQ_FIRST(&queue->tq_queue)) {
 362                 TAILQ_INSERT_TAIL(&queue->tq_active, &tb, tb_link);
 363
 364                 /*
 365                  * Carefully remove the first task from the queue and
 366                  * clear its TASK_ENQUEUED flag
 367                  */
 368                 gtask = STAILQ_FIRST(&queue->tq_queue);
 369                 KASSERT(gtask != NULL, ("task is NULL"));
 370                 STAILQ_REMOVE_HEAD(&queue->tq_queue, ta_link);
 371                 gtask->ta_flags &= ~TASK_ENQUEUED;
 372                 tb.tb_running = gtask;
 373                 TQ_UNLOCK(queue);
 374
 375                 KASSERT(gtask->ta_func != NULL, ("task->ta_func is NULL"));
 376                 gtask->ta_func(gtask->ta_context);
 377
 378                 TQ_LOCK(queue);
 379                 tb.tb_running = NULL;
 380                 wakeup(gtask);
 381
 382                 TAILQ_REMOVE(&queue->tq_active, &tb, tb_link);
 383                 tb_first = TAILQ_FIRST(&queue->tq_active);
 384                 if (tb_first != NULL &&
 385                     tb_first->tb_running == TB_DRAIN_WAITER)
 386                         wakeup(tb_first);
 387         }
 388 }
 389
 390 static int
 391 task_is_running(struct gtaskqueue *queue, struct gtask *gtask)
 392 {
 393         struct gtaskqueue_busy *tb;
 394
 395         TQ_ASSERT_LOCKED(queue);
 396         TAILQ_FOREACH(tb, &queue->tq_active, tb_link) {
 397                 if (tb->tb_running == gtask)
 398                         return (1);
 399         }
 400         return (0);
 401 }
 402
 403 static int
 404 gtaskqueue_cancel_locked(struct gtaskqueue *queue, struct gtask *gtask)
 405 {
 406
 407         if (gtask->ta_flags & TASK_ENQUEUED)
 408                 STAILQ_REMOVE(&queue->tq_queue, gtask, gtask, ta_link);
 409         gtask->ta_flags &= ~TASK_ENQUEUED;
 410         return (task_is_running(queue, gtask) ? EBUSY : 0);
 411 }
 412
 413 int
 414 gtaskqueue_cancel(struct gtaskqueue *queue, struct gtask *gtask)
 415 {
 416         int error;
 417
 418         TQ_LOCK(queue);
 419         error = gtaskqueue_cancel_locked(queue, gtask);
 420         TQ_UNLOCK(queue);
 421
 422         return (error);
 423 }
 424
 425 static void
 426 gtaskqueue_drain_locked(struct gtaskqueue *queue, struct gtask *gtask)
 427 {
 428         while ((gtask->ta_flags & TASK_ENQUEUED) || task_is_running(queue, gtask))
 429                 TQ_SLEEP(queue, gtask, &queue->tq_mutex, PWAIT, "-", 0);
 430 }
 431
 432 void
 433 gtaskqueue_drain(struct gtaskqueue *queue, struct gtask *gtask)
 434 {
 435
 436         if (!queue->tq_spin)
 437                 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, __func__);
 438
 439         TQ_LOCK(queue);
 440         gtaskqueue_drain_locked(queue, gtask);
 441         TQ_UNLOCK(queue);
 442 }
 443
 444 void
 445 gtaskqueue_drain_all(struct gtaskqueue *queue)
 446 {
 447
 448         if (!queue->tq_spin)
 449                 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, __func__);
 450
 451         TQ_LOCK(queue);
 452         gtaskqueue_drain_tq_queue(queue);
 453         gtaskqueue_drain_tq_active(queue);
 454         TQ_UNLOCK(queue);
 455 }
 456
 457 static int
 458 _gtaskqueue_start_threads(struct gtaskqueue **tqp, int count, int pri,
 459     cpuset_t *mask, const char *name, va_list ap)
 460 {
 461         char ktname[MAXCOMLEN + 1];
 462         struct thread *td;
 463         struct gtaskqueue *tq;
 464         int i, error;
 465
 466         if (count <= 0)
 467                 return (EINVAL);
 468
 469         vsnprintf(ktname, sizeof(ktname), name, ap);
 470         tq = *tqp;
 471
 472         tq->tq_threads = malloc(sizeof(struct thread *) * count, M_GTASKQUEUE,
 473             M_NOWAIT | M_ZERO);
 474         if (tq->tq_threads == NULL) {
 475                 printf("%s: no memory for %s threads\n", __func__, ktname);
 476                 return (ENOMEM);
 477         }
 478
 479         for (i = 0; i < count; i++) {
 480                 if (count == 1)
 481                         error = kthread_add(gtaskqueue_thread_loop, tqp, NULL,
 482                             &tq->tq_threads[i], RFSTOPPED, 0, "%s", ktname);
 483                 else
 484                         error = kthread_add(gtaskqueue_thread_loop, tqp, NULL,
 485                             &tq->tq_threads[i], RFSTOPPED, 0,
 486                             "%s_%d", ktname, i);
 487                 if (error) {
 488                         /* should be ok to continue, taskqueue_free will dtrt */
 489                         printf("%s: kthread_add(%s): error %d", __func__,
 490                             ktname, error);
 491                         tq->tq_threads[i] = NULL;               /* paranoid */
 492                 } else
 493                         tq->tq_tcount++;
 494         }
 495         for (i = 0; i < count; i++) {
 496                 if (tq->tq_threads[i] == NULL)
 497                         continue;
 498                 td = tq->tq_threads[i];
 499                 if (mask) {
 500                         error = cpuset_setthread(td->td_tid, mask);
 501                         /*
 502                          * Failing to pin is rarely an actual fatal error;
 503                          * it'll just affect performance.
 504                          */
 505                         if (error)
 506                                 printf("%s: curthread=%llu: can't pin; "
 507                                     "error=%d\n",
 508                                     __func__,
 509                                     (unsigned long long) td->td_tid,
 510                                     error);
 511                 }
 512                 thread_lock(td);
 513                 sched_prio(td, pri);
 514                 sched_add(td, SRQ_BORING);
 515                 thread_unlock(td);
 516         }
 517
 518         return (0);
 519 }
 520
 521 static int
 522 gtaskqueue_start_threads(struct gtaskqueue **tqp, int count, int pri,
 523     const char *name, ...)
 524 {
 525         va_list ap;
 526         int error;
 527
 528         va_start(ap, name);
 529         error = _gtaskqueue_start_threads(tqp, count, pri, NULL, name, ap);
 530         va_end(ap);
 531         return (error);
 532 }
 533
 534 static inline void
 535 gtaskqueue_run_callback(struct gtaskqueue *tq,
 536     enum taskqueue_callback_type cb_type)
 537 {
 538         taskqueue_callback_fn tq_callback;
 539
 540         TQ_ASSERT_UNLOCKED(tq);
 541         tq_callback = tq->tq_callbacks[cb_type];
 542         if (tq_callback != NULL)
 543                 tq_callback(tq->tq_cb_contexts[cb_type]);
 544 }
 545
 546 static void
 547 gtaskqueue_thread_loop(void *arg)
 548 {
 549         struct gtaskqueue **tqp, *tq;
 550
 551         tqp = arg;
 552         tq = *tqp;
 553         gtaskqueue_run_callback(tq, TASKQUEUE_CALLBACK_TYPE_INIT);
 554         TQ_LOCK(tq);
 555         while ((tq->tq_flags & TQ_FLAGS_ACTIVE) != 0) {
 556                 /* XXX ? */
 557                 gtaskqueue_run_locked(tq);
 558                 /*
 559                  * Because taskqueue_run() can drop tq_mutex, we need to
 560                  * check if the TQ_FLAGS_ACTIVE flag wasn't removed in the
 561                  * meantime, which means we missed a wakeup.
 562                  */
 563                 if ((tq->tq_flags & TQ_FLAGS_ACTIVE) == 0)
 564                         break;
 565                 TQ_SLEEP(tq, tq, &tq->tq_mutex, 0, "-", 0);
 566         }
 567         gtaskqueue_run_locked(tq);
 568         /*
 569          * This thread is on its way out, so just drop the lock temporarily
 570          * in order to call the shutdown callback.  This allows the callback
 571          * to look at the taskqueue, even just before it dies.
 572          */
 573         TQ_UNLOCK(tq);
 574         gtaskqueue_run_callback(tq, TASKQUEUE_CALLBACK_TYPE_SHUTDOWN);
 575         TQ_LOCK(tq);
 576
 577         /* rendezvous with thread that asked us to terminate */
 578         tq->tq_tcount--;
 579         wakeup_one(tq->tq_threads);
 580         TQ_UNLOCK(tq);
 581         kthread_exit();
 582 }
 583
 584 static void
 585 gtaskqueue_thread_enqueue(void *context)
 586 {
 587         struct gtaskqueue **tqp, *tq;
 588
 589         tqp = context;
 590         tq = *tqp;
 591         wakeup_one(tq);
 592 }
 593
 594
 595 static struct gtaskqueue *
 596 gtaskqueue_create_fast(const char *name, int mflags,
 597                  taskqueue_enqueue_fn enqueue, void *context)
 598 {
 599         return _gtaskqueue_create(name, mflags, enqueue, context,
 600                         MTX_SPIN, "fast_taskqueue");
 601 }
 602
 603
 604 struct taskqgroup_cpu {
 605         LIST_HEAD(, grouptask)  tgc_tasks;
 606         struct gtaskqueue       *tgc_taskq;
 607         int     tgc_cnt;
 608         int     tgc_cpu;
 609 };
 610
 611 struct taskqgroup {
 612         struct taskqgroup_cpu tqg_queue[MAXCPU];
 613         struct mtx      tqg_lock;
 614         const char *    tqg_name;
 615         int             tqg_adjusting;
 616         int             tqg_stride;
 617         int             tqg_cnt;
 618 };
 619
 620 struct taskq_bind_task {
 621         struct gtask bt_task;
 622         int     bt_cpuid;
 623 };
 624
 625 static void
 626 taskqgroup_cpu_create(struct taskqgroup *qgroup, int idx, int cpu)
 627 {
 628         struct taskqgroup_cpu *qcpu;
 629
 630         qcpu = &qgroup->tqg_queue[idx];
 631         LIST_INIT(&qcpu->tgc_tasks);
 632         qcpu->tgc_taskq = gtaskqueue_create_fast(NULL, M_WAITOK,
 633             taskqueue_thread_enqueue, &qcpu->tgc_taskq);
 634         gtaskqueue_start_threads(&qcpu->tgc_taskq, 1, PI_SOFT,
 635             "%s_%d", qgroup->tqg_name, idx);
 636         qcpu->tgc_cpu = cpu;
 637 }
 638
 639 static void
 640 taskqgroup_cpu_remove(struct taskqgroup *qgroup, int idx)
 641 {
 642
 643         gtaskqueue_free(qgroup->tqg_queue[idx].tgc_taskq);
 644 }
 645
 646 /*
 647  * Find the taskq with least # of tasks that doesn't currently have any
 648  * other queues from the uniq identifier.
 649  */
 650 static int
 651 taskqgroup_find(struct taskqgroup *qgroup, void *uniq)
 652 {
 653         struct grouptask *n;
 654         int i, idx, mincnt;
 655         int strict;
 656
 657         mtx_assert(&qgroup->tqg_lock, MA_OWNED);
 658         if (qgroup->tqg_cnt == 0)
 659                 return (0);
 660         idx = -1;
 661         mincnt = INT_MAX;
 662         /*
 663          * Two passes;  First scan for a queue with the least tasks that
 664          * does not already service this uniq id.  If that fails simply find
 665          * the queue with the least total tasks;
 666          */
 667         for (strict = 1; mincnt == INT_MAX; strict = 0) {
 668                 for (i = 0; i < qgroup->tqg_cnt; i++) {
 669                         if (qgroup->tqg_queue[i].tgc_cnt > mincnt)
 670                                 continue;
 671                         if (strict) {
 672                                 LIST_FOREACH(n,
 673                                     &qgroup->tqg_queue[i].tgc_tasks, gt_list)
 674                                         if (n->gt_uniq == uniq)
 675                                                 break;
 676                                 if (n != NULL)
 677                                         continue;
 678                         }
 679                         mincnt = qgroup->tqg_queue[i].tgc_cnt;
 680                         idx = i;
 681                 }
 682         }
 683         if (idx == -1)
 684                 panic("taskqgroup_find: Failed to pick a qid.");
 685
 686         return (idx);
 687 }
 688
 689 /*
 690  * smp_started is unusable since it is not set for UP kernels or even for
 691  * SMP kernels when there is 1 CPU.  This is usually handled by adding a
 692  * (mp_ncpus == 1) test, but that would be broken here since we need to
 693  * to synchronize with the SI_SUB_SMP ordering.  Even in the pure SMP case
 694  * smp_started only gives a fuzzy ordering relative to SI_SUB_SMP.
 695  *
 696  * So maintain our own flag.  It must be set after all CPUs are started
 697  * and before SI_SUB_SMP:SI_ORDER_ANY so that the SYSINIT for delayed
 698  * adjustment is properly delayed.  SI_ORDER_FOURTH is clearly before
 699  * SI_ORDER_ANY and unclearly after the CPUs are started.  It would be
 700  * simpler for adjustment to pass a flag indicating if it is delayed.
 701  */
 702
 703 static int tqg_smp_started;
 704
 705 static void
 706 tqg_record_smp_started(void *arg)
 707 {
 708         tqg_smp_started = 1;
 709 }
 710
 711 SYSINIT(tqg_record_smp_started, SI_SUB_SMP, SI_ORDER_FOURTH,
 712         tqg_record_smp_started, NULL);
 713
 714 void
 715 taskqgroup_attach(struct taskqgroup *qgroup, struct grouptask *gtask,
 716     void *uniq, int irq, const char *name)
 717 {
 718         cpuset_t mask;
 719         int qid, error;
 720
 721         gtask->gt_uniq = uniq;
 722         snprintf(gtask->gt_name, GROUPTASK_NAMELEN, "%s", name ? name : "grouptask");
 723         gtask->gt_irq = irq;
 724         gtask->gt_cpu = -1;
 725         mtx_lock(&qgroup->tqg_lock);
 726         qid = taskqgroup_find(qgroup, uniq);
 727         qgroup->tqg_queue[qid].tgc_cnt++;
 728         LIST_INSERT_HEAD(&qgroup->tqg_queue[qid].tgc_tasks, gtask, gt_list);
 729         gtask->gt_taskqueue = qgroup->tqg_queue[qid].tgc_taskq;
 730         if (irq != -1 && tqg_smp_started) {
 731                 gtask->gt_cpu = qgroup->tqg_queue[qid].tgc_cpu;
 732                 CPU_ZERO(&mask);
 733                 CPU_SET(qgroup->tqg_queue[qid].tgc_cpu, &mask);
 734                 mtx_unlock(&qgroup->tqg_lock);
 735                 error = intr_setaffinity(irq, CPU_WHICH_IRQ, &mask);
 736                 if (error)
 737                         printf("%s: setaffinity failed for %s: %d\n", __func__, gtask->gt_name, error);
 738         } else
 739                 mtx_unlock(&qgroup->tqg_lock);
 740 }
 741
 742 static void
 743 taskqgroup_attach_deferred(struct taskqgroup *qgroup, struct grouptask *gtask)
 744 {
 745         cpuset_t mask;
 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_irq != -1) {
 752                 mtx_unlock(&qgroup->tqg_lock);
 753
 754                 CPU_ZERO(&mask);
 755                 CPU_SET(cpu, &mask);
 756                 error = intr_setaffinity(gtask->gt_irq, CPU_WHICH_IRQ, &mask);
 757                 mtx_lock(&qgroup->tqg_lock);
 758                 if (error)
 759                         printf("%s: %s setaffinity failed: %d\n", __func__, gtask->gt_name, error);
 760
 761         }
 762         qgroup->tqg_queue[qid].tgc_cnt++;
 763
 764         LIST_INSERT_HEAD(&qgroup->tqg_queue[qid].tgc_tasks, gtask,
 765                          gt_list);
 766         MPASS(qgroup->tqg_queue[qid].tgc_taskq != NULL);
 767         gtask->gt_taskqueue = qgroup->tqg_queue[qid].tgc_taskq;
 768         mtx_unlock(&qgroup->tqg_lock);
 769 }
 770
 771 int
 772 taskqgroup_attach_cpu(struct taskqgroup *qgroup, struct grouptask *gtask,
 773         void *uniq, int cpu, int irq, const char *name)
 774 {
 775         cpuset_t mask;
 776         int i, qid, error;
 777
 778         qid = -1;
 779         gtask->gt_uniq = uniq;
 780         snprintf(gtask->gt_name, GROUPTASK_NAMELEN, "%s", name ? name : "grouptask");
 781         gtask->gt_irq = irq;
 782         gtask->gt_cpu = cpu;
 783         mtx_lock(&qgroup->tqg_lock);
 784         if (tqg_smp_started) {
 785                 for (i = 0; i < qgroup->tqg_cnt; i++)
 786                         if (qgroup->tqg_queue[i].tgc_cpu == cpu) {
 787                                 qid = i;
 788                                 break;
 789                         }
 790                 if (qid == -1) {
 791                         mtx_unlock(&qgroup->tqg_lock);
 792                         printf("%s: qid not found for %s cpu=%d\n", __func__, gtask->gt_name, cpu);
 793                         return (EINVAL);
 794                 }
 795         } else
 796                 qid = 0;
 797         qgroup->tqg_queue[qid].tgc_cnt++;
 798         LIST_INSERT_HEAD(&qgroup->tqg_queue[qid].tgc_tasks, gtask, gt_list);
 799         gtask->gt_taskqueue = qgroup->tqg_queue[qid].tgc_taskq;
 800         cpu = qgroup->tqg_queue[qid].tgc_cpu;
 801         mtx_unlock(&qgroup->tqg_lock);
 802
 803         CPU_ZERO(&mask);
 804         CPU_SET(cpu, &mask);
 805         if (irq != -1 && tqg_smp_started) {
 806                 error = intr_setaffinity(irq, CPU_WHICH_IRQ, &mask);
 807                 if (error)
 808                         printf("%s: setaffinity failed: %d\n", __func__, error);
 809         }
 810         return (0);
 811 }
 812
 813 static int
 814 taskqgroup_attach_cpu_deferred(struct taskqgroup *qgroup, struct grouptask *gtask)
 815 {
 816         cpuset_t mask;
 817         int i, qid, irq, cpu, error;
 818
 819         qid = -1;
 820         irq = gtask->gt_irq;
 821         cpu = gtask->gt_cpu;
 822         MPASS(tqg_smp_started);
 823         mtx_lock(&qgroup->tqg_lock);
 824         for (i = 0; i < qgroup->tqg_cnt; i++)
 825                 if (qgroup->tqg_queue[i].tgc_cpu == cpu) {
 826                         qid = i;
 827                         break;
 828                 }
 829         if (qid == -1) {
 830                 mtx_unlock(&qgroup->tqg_lock);
 831                 printf("%s: qid not found for %s cpu=%d\n", __func__, gtask->gt_name, cpu);
 832                 return (EINVAL);
 833         }
 834         qgroup->tqg_queue[qid].tgc_cnt++;
 835         LIST_INSERT_HEAD(&qgroup->tqg_queue[qid].tgc_tasks, gtask, gt_list);
 836         MPASS(qgroup->tqg_queue[qid].tgc_taskq != NULL);
 837         gtask->gt_taskqueue = qgroup->tqg_queue[qid].tgc_taskq;
 838         mtx_unlock(&qgroup->tqg_lock);
 839
 840         CPU_ZERO(&mask);
 841         CPU_SET(cpu, &mask);
 842
 843         if (irq != -1) {
 844                 error = intr_setaffinity(irq, CPU_WHICH_IRQ, &mask);
 845                 if (error)
 846                         printf("%s: setaffinity failed: %d\n", __func__, error);
 847         }
 848         return (0);
 849 }
 850
 851 void
 852 taskqgroup_detach(struct taskqgroup *qgroup, struct grouptask *gtask)
 853 {
 854         int i;
 855
 856         grouptask_block(gtask);
 857         mtx_lock(&qgroup->tqg_lock);
 858         for (i = 0; i < qgroup->tqg_cnt; i++)
 859                 if (qgroup->tqg_queue[i].tgc_taskq == gtask->gt_taskqueue)
 860                         break;
 861         if (i == qgroup->tqg_cnt)
 862                 panic("taskqgroup_detach: task %s not in group\n", gtask->gt_name);
 863         qgroup->tqg_queue[i].tgc_cnt--;
 864         LIST_REMOVE(gtask, gt_list);
 865         mtx_unlock(&qgroup->tqg_lock);
 866         gtask->gt_taskqueue = NULL;
 867         gtask->gt_task.ta_flags &= ~TASK_NOENQUEUE;
 868 }
 869
 870 static void
 871 taskqgroup_binder(void *ctx)
 872 {
 873         struct taskq_bind_task *gtask = (struct taskq_bind_task *)ctx;
 874         cpuset_t mask;
 875         int error;
 876
 877         CPU_ZERO(&mask);
 878         CPU_SET(gtask->bt_cpuid, &mask);
 879         error = cpuset_setthread(curthread->td_tid, &mask);
 880         thread_lock(curthread);
 881         sched_bind(curthread, gtask->bt_cpuid);
 882         thread_unlock(curthread);
 883
 884         if (error)
 885                 printf("%s: setaffinity failed: %d\n", __func__,
 886                     error);
 887         free(gtask, M_DEVBUF);
 888 }
 889
 890 static void
 891 taskqgroup_bind(struct taskqgroup *qgroup)
 892 {
 893         struct taskq_bind_task *gtask;
 894         int i;
 895
 896         /*
 897          * Bind taskqueue threads to specific CPUs, if they have been assigned
 898          * one.
 899          */
 900         if (qgroup->tqg_cnt == 1)
 901                 return;
 902
 903         for (i = 0; i < qgroup->tqg_cnt; i++) {
 904                 gtask = malloc(sizeof (*gtask), M_DEVBUF, M_WAITOK);
 905                 GTASK_INIT(&gtask->bt_task, 0, 0, taskqgroup_binder, gtask);
 906                 gtask->bt_cpuid = qgroup->tqg_queue[i].tgc_cpu;
 907                 grouptaskqueue_enqueue(qgroup->tqg_queue[i].tgc_taskq,
 908                     &gtask->bt_task);
 909         }
 910 }
 911
 912 static void
 913 taskqgroup_config_init(void *arg)
 914 {
 915         struct taskqgroup *qgroup = qgroup_config;
 916         LIST_HEAD(, grouptask) gtask_head = LIST_HEAD_INITIALIZER(NULL);
 917
 918         LIST_SWAP(&gtask_head, &qgroup->tqg_queue[0].tgc_tasks,
 919             grouptask, gt_list);
 920         qgroup->tqg_queue[0].tgc_cnt = 0;
 921         taskqgroup_cpu_create(qgroup, 0, 0);
 922
 923         qgroup->tqg_cnt = 1;
 924         qgroup->tqg_stride = 1;
 925 }
 926
 927 SYSINIT(taskqgroup_config_init, SI_SUB_TASKQ, SI_ORDER_SECOND,
 928         taskqgroup_config_init, NULL);
 929
 930 static int
 931 _taskqgroup_adjust(struct taskqgroup *qgroup, int cnt, int stride)
 932 {
 933         LIST_HEAD(, grouptask) gtask_head = LIST_HEAD_INITIALIZER(NULL);
 934         struct grouptask *gtask;
 935         int i, k, old_cnt, old_cpu, cpu;
 936
 937         mtx_assert(&qgroup->tqg_lock, MA_OWNED);
 938
 939         if (cnt < 1 || cnt * stride > mp_ncpus || !tqg_smp_started) {
 940                 printf("%s: failed cnt: %d stride: %d "
 941                     "mp_ncpus: %d tqg_smp_started: %d\n",
 942                     __func__, cnt, stride, mp_ncpus, tqg_smp_started);
 943                 return (EINVAL);
 944         }
 945         if (qgroup->tqg_adjusting) {
 946                 printf("%s failed: adjusting\n", __func__);
 947                 return (EBUSY);
 948         }
 949         qgroup->tqg_adjusting = 1;
 950         old_cnt = qgroup->tqg_cnt;
 951         old_cpu = 0;
 952         if (old_cnt < cnt)
 953                 old_cpu = qgroup->tqg_queue[old_cnt].tgc_cpu;
 954         mtx_unlock(&qgroup->tqg_lock);
 955         /*
 956          * Set up queue for tasks added before boot.
 957          */
 958         if (old_cnt == 0) {
 959                 LIST_SWAP(&gtask_head, &qgroup->tqg_queue[0].tgc_tasks,
 960                     grouptask, gt_list);
 961                 qgroup->tqg_queue[0].tgc_cnt = 0;
 962         }
 963
 964         /*
 965          * If new taskq threads have been added.
 966          */
 967         cpu = old_cpu;
 968         for (i = old_cnt; i < cnt; i++) {
 969                 taskqgroup_cpu_create(qgroup, i, cpu);
 970
 971                 for (k = 0; k < stride; k++)
 972                         cpu = CPU_NEXT(cpu);
 973         }
 974         mtx_lock(&qgroup->tqg_lock);
 975         qgroup->tqg_cnt = cnt;
 976         qgroup->tqg_stride = stride;
 977
 978         /*
 979          * Adjust drivers to use new taskqs.
 980          */
 981         for (i = 0; i < old_cnt; i++) {
 982                 while ((gtask = LIST_FIRST(&qgroup->tqg_queue[i].tgc_tasks))) {
 983                         LIST_REMOVE(gtask, gt_list);
 984                         qgroup->tqg_queue[i].tgc_cnt--;
 985                         LIST_INSERT_HEAD(&gtask_head, gtask, gt_list);
 986                 }
 987         }
 988         mtx_unlock(&qgroup->tqg_lock);
 989
 990         while ((gtask = LIST_FIRST(&gtask_head))) {
 991                 LIST_REMOVE(gtask, gt_list);
 992                 if (gtask->gt_cpu == -1)
 993                         taskqgroup_attach_deferred(qgroup, gtask);
 994                 else if (taskqgroup_attach_cpu_deferred(qgroup, gtask))
 995                         taskqgroup_attach_deferred(qgroup, gtask);
 996         }
 997
 998 #ifdef INVARIANTS
 999         mtx_lock(&qgroup->tqg_lock);
1000         for (i = 0; i < qgroup->tqg_cnt; i++) {
1001                 MPASS(qgroup->tqg_queue[i].tgc_taskq != NULL);
1002                 LIST_FOREACH(gtask, &qgroup->tqg_queue[i].tgc_tasks, gt_list)
1003                         MPASS(gtask->gt_taskqueue != NULL);
1004         }
1005         mtx_unlock(&qgroup->tqg_lock);
1006 #endif
1007         /*
1008          * If taskq thread count has been reduced.
1009          */
1010         for (i = cnt; i < old_cnt; i++)
1011                 taskqgroup_cpu_remove(qgroup, i);
1012
1013         taskqgroup_bind(qgroup);
1014
1015         mtx_lock(&qgroup->tqg_lock);
1016         qgroup->tqg_adjusting = 0;
1017
1018         return (0);
1019 }
1020
1021 int
1022 taskqgroup_adjust(struct taskqgroup *qgroup, int cnt, int stride)
1023 {
1024         int error;
1025
1026         mtx_lock(&qgroup->tqg_lock);
1027         error = _taskqgroup_adjust(qgroup, cnt, stride);
1028         mtx_unlock(&qgroup->tqg_lock);
1029
1030         return (error);
1031 }
1032
1033 struct taskqgroup *
1034 taskqgroup_create(const char *name)
1035 {
1036         struct taskqgroup *qgroup;
1037
1038         qgroup = malloc(sizeof(*qgroup), M_GTASKQUEUE, M_WAITOK | M_ZERO);
1039         mtx_init(&qgroup->tqg_lock, "taskqgroup", NULL, MTX_DEF);
1040         qgroup->tqg_name = name;
1041         LIST_INIT(&qgroup->tqg_queue[0].tgc_tasks);
1042
1043         return (qgroup);
1044 }
1045
1046 void
1047 taskqgroup_destroy(struct taskqgroup *qgroup)
1048 {
1049
1050 }
1051
1052 void
1053 taskqgroup_config_gtask_init(void *ctx, struct grouptask *gtask, gtask_fn_t *fn,
1054         const char *name)
1055 {
1056
1057         GROUPTASK_INIT(gtask, 0, fn, ctx);
1058         taskqgroup_attach(qgroup_config, gtask, gtask, -1, name);
1059 }
1060
1061 void
1062 taskqgroup_config_gtask_deinit(struct grouptask *gtask)
1063 {
1064         taskqgroup_detach(qgroup_config, gtask);
1065 }