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 u_int tb_seq; 62 LIST_ENTRY(gtaskqueue_busy) tb_link; 63 }; 64 65 typedef void (*gtaskqueue_enqueue_fn)(void *context); 66 67 struct gtaskqueue { 68 STAILQ_HEAD(, gtask) tq_queue; 69 LIST_HEAD(, gtaskqueue_busy) tq_active; 70 u_int tq_seq; 71 int tq_callouts; 72 struct mtx_padalign tq_mutex; 73 gtaskqueue_enqueue_fn tq_enqueue; 74 void *tq_context; 75 char *tq_name; 76 struct thread **tq_threads; 77 int tq_tcount; 78 int tq_spin; 79 int tq_flags; 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, const char *wm) 119 { 120 if (tq->tq_spin) 121 return (msleep_spin(p, (struct mtx *)&tq->tq_mutex, wm, 0)); 122 return (msleep(p, &tq->tq_mutex, 0, wm, 0)); 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 LIST_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, "gtq_destroy"); 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(LIST_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, "gtq_qdrain"); 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; 299 u_int seq; 300 301 if (LIST_EMPTY(&queue->tq_active)) 302 return; 303 304 /* Block taskq_terminate().*/ 305 queue->tq_callouts++; 306 307 /* Wait for any active task with sequence from the past. */ 308 seq = queue->tq_seq; 309 restart: 310 LIST_FOREACH(tb, &queue->tq_active, tb_link) { 311 if ((int)(tb->tb_seq - seq) <= 0) { 312 TQ_SLEEP(queue, tb->tb_running, "gtq_adrain"); 313 goto restart; 314 } 315 } 316 317 /* Release taskqueue_terminate(). */ 318 queue->tq_callouts--; 319 if ((queue->tq_flags & TQ_FLAGS_ACTIVE) == 0) 320 wakeup_one(queue->tq_threads); 321 } 322 323 void 324 gtaskqueue_block(struct gtaskqueue *queue) 325 { 326 327 TQ_LOCK(queue); 328 queue->tq_flags |= TQ_FLAGS_BLOCKED; 329 TQ_UNLOCK(queue); 330 } 331 332 void 333 gtaskqueue_unblock(struct gtaskqueue *queue) 334 { 335 336 TQ_LOCK(queue); 337 queue->tq_flags &= ~TQ_FLAGS_BLOCKED; 338 if (!STAILQ_EMPTY(&queue->tq_queue)) 339 queue->tq_enqueue(queue->tq_context); 340 TQ_UNLOCK(queue); 341 } 342 343 static void 344 gtaskqueue_run_locked(struct gtaskqueue *queue) 345 { 346 struct gtaskqueue_busy tb; 347 struct gtask *gtask; 348 349 KASSERT(queue != NULL, ("tq is NULL")); 350 TQ_ASSERT_LOCKED(queue); 351 tb.tb_running = NULL; 352 LIST_INSERT_HEAD(&queue->tq_active, &tb, tb_link); 353 354 while ((gtask = STAILQ_FIRST(&queue->tq_queue)) != NULL) { 355 STAILQ_REMOVE_HEAD(&queue->tq_queue, ta_link); 356 gtask->ta_flags &= ~TASK_ENQUEUED; 357 tb.tb_running = gtask; 358 tb.tb_seq = ++queue->tq_seq; 359 TQ_UNLOCK(queue); 360 361 KASSERT(gtask->ta_func != NULL, ("task->ta_func is NULL")); 362 gtask->ta_func(gtask->ta_context); 363 364 TQ_LOCK(queue); 365 wakeup(gtask); 366 } 367 LIST_REMOVE(&tb, tb_link); 368 } 369 370 static int 371 task_is_running(struct gtaskqueue *queue, struct gtask *gtask) 372 { 373 struct gtaskqueue_busy *tb; 374 375 TQ_ASSERT_LOCKED(queue); 376 LIST_FOREACH(tb, &queue->tq_active, tb_link) { 377 if (tb->tb_running == gtask) 378 return (1); 379 } 380 return (0); 381 } 382 383 static int 384 gtaskqueue_cancel_locked(struct gtaskqueue *queue, struct gtask *gtask) 385 { 386 387 if (gtask->ta_flags & TASK_ENQUEUED) 388 STAILQ_REMOVE(&queue->tq_queue, gtask, gtask, ta_link); 389 gtask->ta_flags &= ~TASK_ENQUEUED; 390 return (task_is_running(queue, gtask) ? EBUSY : 0); 391 } 392 393 int 394 gtaskqueue_cancel(struct gtaskqueue *queue, struct gtask *gtask) 395 { 396 int error; 397 398 TQ_LOCK(queue); 399 error = gtaskqueue_cancel_locked(queue, gtask); 400 TQ_UNLOCK(queue); 401 402 return (error); 403 } 404 405 static void 406 gtaskqueue_drain_locked(struct gtaskqueue *queue, struct gtask *gtask) 407 { 408 while ((gtask->ta_flags & TASK_ENQUEUED) || task_is_running(queue, gtask)) 409 TQ_SLEEP(queue, gtask, "gtq_drain"); 410 } 411 412 void 413 gtaskqueue_drain(struct gtaskqueue *queue, struct gtask *gtask) 414 { 415 416 if (!queue->tq_spin) 417 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, __func__); 418 419 TQ_LOCK(queue); 420 gtaskqueue_drain_locked(queue, gtask); 421 TQ_UNLOCK(queue); 422 } 423 424 void 425 gtaskqueue_drain_all(struct gtaskqueue *queue) 426 { 427 428 if (!queue->tq_spin) 429 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, __func__); 430 431 TQ_LOCK(queue); 432 gtaskqueue_drain_tq_queue(queue); 433 gtaskqueue_drain_tq_active(queue); 434 TQ_UNLOCK(queue); 435 } 436 437 static int 438 _gtaskqueue_start_threads(struct gtaskqueue **tqp, int count, int pri, 439 cpuset_t *mask, const char *name, va_list ap) 440 { 441 char ktname[MAXCOMLEN + 1]; 442 struct thread *td; 443 struct gtaskqueue *tq; 444 int i, error; 445 446 if (count <= 0) 447 return (EINVAL); 448 449 vsnprintf(ktname, sizeof(ktname), name, ap); 450 tq = *tqp; 451 452 tq->tq_threads = malloc(sizeof(struct thread *) * count, M_GTASKQUEUE, 453 M_NOWAIT | M_ZERO); 454 if (tq->tq_threads == NULL) { 455 printf("%s: no memory for %s threads\n", __func__, ktname); 456 return (ENOMEM); 457 } 458 459 for (i = 0; i < count; i++) { 460 if (count == 1) 461 error = kthread_add(gtaskqueue_thread_loop, tqp, NULL, 462 &tq->tq_threads[i], RFSTOPPED, 0, "%s", ktname); 463 else 464 error = kthread_add(gtaskqueue_thread_loop, tqp, NULL, 465 &tq->tq_threads[i], RFSTOPPED, 0, 466 "%s_%d", ktname, i); 467 if (error) { 468 /* should be ok to continue, taskqueue_free will dtrt */ 469 printf("%s: kthread_add(%s): error %d", __func__, 470 ktname, error); 471 tq->tq_threads[i] = NULL; /* paranoid */ 472 } else 473 tq->tq_tcount++; 474 } 475 for (i = 0; i < count; i++) { 476 if (tq->tq_threads[i] == NULL) 477 continue; 478 td = tq->tq_threads[i]; 479 if (mask) { 480 error = cpuset_setthread(td->td_tid, mask); 481 /* 482 * Failing to pin is rarely an actual fatal error; 483 * it'll just affect performance. 484 */ 485 if (error) 486 printf("%s: curthread=%llu: can't pin; " 487 "error=%d\n", 488 __func__, 489 (unsigned long long) td->td_tid, 490 error); 491 } 492 thread_lock(td); 493 sched_prio(td, pri); 494 sched_add(td, SRQ_BORING); 495 thread_unlock(td); 496 } 497 498 return (0); 499 } 500 501 static int 502 gtaskqueue_start_threads(struct gtaskqueue **tqp, int count, int pri, 503 const char *name, ...) 504 { 505 va_list ap; 506 int error; 507 508 va_start(ap, name); 509 error = _gtaskqueue_start_threads(tqp, count, pri, NULL, name, ap); 510 va_end(ap); 511 return (error); 512 } 513 514 static inline void 515 gtaskqueue_run_callback(struct gtaskqueue *tq, 516 enum taskqueue_callback_type cb_type) 517 { 518 taskqueue_callback_fn tq_callback; 519 520 TQ_ASSERT_UNLOCKED(tq); 521 tq_callback = tq->tq_callbacks[cb_type]; 522 if (tq_callback != NULL) 523 tq_callback(tq->tq_cb_contexts[cb_type]); 524 } 525 526 static void 527 gtaskqueue_thread_loop(void *arg) 528 { 529 struct gtaskqueue **tqp, *tq; 530 531 tqp = arg; 532 tq = *tqp; 533 gtaskqueue_run_callback(tq, TASKQUEUE_CALLBACK_TYPE_INIT); 534 TQ_LOCK(tq); 535 while ((tq->tq_flags & TQ_FLAGS_ACTIVE) != 0) { 536 /* XXX ? */ 537 gtaskqueue_run_locked(tq); 538 /* 539 * Because taskqueue_run() can drop tq_mutex, we need to 540 * check if the TQ_FLAGS_ACTIVE flag wasn't removed in the 541 * meantime, which means we missed a wakeup. 542 */ 543 if ((tq->tq_flags & TQ_FLAGS_ACTIVE) == 0) 544 break; 545 TQ_SLEEP(tq, tq, "-"); 546 } 547 gtaskqueue_run_locked(tq); 548 /* 549 * This thread is on its way out, so just drop the lock temporarily 550 * in order to call the shutdown callback. This allows the callback 551 * to look at the taskqueue, even just before it dies. 552 */ 553 TQ_UNLOCK(tq); 554 gtaskqueue_run_callback(tq, TASKQUEUE_CALLBACK_TYPE_SHUTDOWN); 555 TQ_LOCK(tq); 556 557 /* rendezvous with thread that asked us to terminate */ 558 tq->tq_tcount--; 559 wakeup_one(tq->tq_threads); 560 TQ_UNLOCK(tq); 561 kthread_exit(); 562 } 563 564 static void 565 gtaskqueue_thread_enqueue(void *context) 566 { 567 struct gtaskqueue **tqp, *tq; 568 569 tqp = context; 570 tq = *tqp; 571 wakeup_any(tq); 572 } 573 574 575 static struct gtaskqueue * 576 gtaskqueue_create_fast(const char *name, int mflags, 577 taskqueue_enqueue_fn enqueue, void *context) 578 { 579 return _gtaskqueue_create(name, mflags, enqueue, context, 580 MTX_SPIN, "fast_taskqueue"); 581 } 582 583 584 struct taskqgroup_cpu { 585 LIST_HEAD(, grouptask) tgc_tasks; 586 struct gtaskqueue *tgc_taskq; 587 int tgc_cnt; 588 int tgc_cpu; 589 }; 590 591 struct taskqgroup { 592 struct taskqgroup_cpu tqg_queue[MAXCPU]; 593 struct mtx tqg_lock; 594 const char * tqg_name; 595 int tqg_adjusting; 596 int tqg_stride; 597 int tqg_cnt; 598 }; 599 600 struct taskq_bind_task { 601 struct gtask bt_task; 602 int bt_cpuid; 603 }; 604 605 static void 606 taskqgroup_cpu_create(struct taskqgroup *qgroup, int idx, int cpu) 607 { 608 struct taskqgroup_cpu *qcpu; 609 610 qcpu = &qgroup->tqg_queue[idx]; 611 LIST_INIT(&qcpu->tgc_tasks); 612 qcpu->tgc_taskq = gtaskqueue_create_fast(NULL, M_WAITOK, 613 taskqueue_thread_enqueue, &qcpu->tgc_taskq); 614 gtaskqueue_start_threads(&qcpu->tgc_taskq, 1, PI_SOFT, 615 "%s_%d", qgroup->tqg_name, idx); 616 qcpu->tgc_cpu = cpu; 617 } 618 619 static void 620 taskqgroup_cpu_remove(struct taskqgroup *qgroup, int idx) 621 { 622 623 gtaskqueue_free(qgroup->tqg_queue[idx].tgc_taskq); 624 } 625 626 /* 627 * Find the taskq with least # of tasks that doesn't currently have any 628 * other queues from the uniq identifier. 629 */ 630 static int 631 taskqgroup_find(struct taskqgroup *qgroup, void *uniq) 632 { 633 struct grouptask *n; 634 int i, idx, mincnt; 635 int strict; 636 637 mtx_assert(&qgroup->tqg_lock, MA_OWNED); 638 if (qgroup->tqg_cnt == 0) 639 return (0); 640 idx = -1; 641 mincnt = INT_MAX; 642 /* 643 * Two passes; First scan for a queue with the least tasks that 644 * does not already service this uniq id. If that fails simply find 645 * the queue with the least total tasks; 646 */ 647 for (strict = 1; mincnt == INT_MAX; strict = 0) { 648 for (i = 0; i < qgroup->tqg_cnt; i++) { 649 if (qgroup->tqg_queue[i].tgc_cnt > mincnt) 650 continue; 651 if (strict) { 652 LIST_FOREACH(n, 653 &qgroup->tqg_queue[i].tgc_tasks, gt_list) 654 if (n->gt_uniq == uniq) 655 break; 656 if (n != NULL) 657 continue; 658 } 659 mincnt = qgroup->tqg_queue[i].tgc_cnt; 660 idx = i; 661 } 662 } 663 if (idx == -1) 664 panic("%s: failed to pick a qid.", __func__); 665 666 return (idx); 667 } 668 669 /* 670 * smp_started is unusable since it is not set for UP kernels or even for 671 * SMP kernels when there is 1 CPU. This is usually handled by adding a 672 * (mp_ncpus == 1) test, but that would be broken here since we need to 673 * to synchronize with the SI_SUB_SMP ordering. Even in the pure SMP case 674 * smp_started only gives a fuzzy ordering relative to SI_SUB_SMP. 675 * 676 * So maintain our own flag. It must be set after all CPUs are started 677 * and before SI_SUB_SMP:SI_ORDER_ANY so that the SYSINIT for delayed 678 * adjustment is properly delayed. SI_ORDER_FOURTH is clearly before 679 * SI_ORDER_ANY and unclearly after the CPUs are started. It would be 680 * simpler for adjustment to pass a flag indicating if it is delayed. 681 */ 682 683 static int tqg_smp_started; 684 685 static void 686 tqg_record_smp_started(void *arg) 687 { 688 tqg_smp_started = 1; 689 } 690 691 SYSINIT(tqg_record_smp_started, SI_SUB_SMP, SI_ORDER_FOURTH, 692 tqg_record_smp_started, NULL); 693 694 void 695 taskqgroup_attach(struct taskqgroup *qgroup, struct grouptask *gtask, 696 void *uniq, device_t dev, struct resource *irq, const char *name) 697 { 698 int cpu, qid, error; 699 700 gtask->gt_uniq = uniq; 701 snprintf(gtask->gt_name, GROUPTASK_NAMELEN, "%s", name ? name : "grouptask"); 702 gtask->gt_dev = dev; 703 gtask->gt_irq = irq; 704 gtask->gt_cpu = -1; 705 mtx_lock(&qgroup->tqg_lock); 706 qid = taskqgroup_find(qgroup, uniq); 707 qgroup->tqg_queue[qid].tgc_cnt++; 708 LIST_INSERT_HEAD(&qgroup->tqg_queue[qid].tgc_tasks, gtask, gt_list); 709 gtask->gt_taskqueue = qgroup->tqg_queue[qid].tgc_taskq; 710 if (dev != NULL && irq != NULL && tqg_smp_started) { 711 cpu = qgroup->tqg_queue[qid].tgc_cpu; 712 gtask->gt_cpu = cpu; 713 mtx_unlock(&qgroup->tqg_lock); 714 error = bus_bind_intr(dev, irq, cpu); 715 if (error) 716 printf("%s: binding interrupt failed for %s: %d\n", 717 __func__, gtask->gt_name, error); 718 } else 719 mtx_unlock(&qgroup->tqg_lock); 720 } 721 722 static void 723 taskqgroup_attach_deferred(struct taskqgroup *qgroup, struct grouptask *gtask) 724 { 725 int qid, cpu, error; 726 727 mtx_lock(&qgroup->tqg_lock); 728 qid = taskqgroup_find(qgroup, gtask->gt_uniq); 729 cpu = qgroup->tqg_queue[qid].tgc_cpu; 730 if (gtask->gt_dev != NULL && gtask->gt_irq != NULL) { 731 mtx_unlock(&qgroup->tqg_lock); 732 error = bus_bind_intr(gtask->gt_dev, gtask->gt_irq, cpu); 733 mtx_lock(&qgroup->tqg_lock); 734 if (error) 735 printf("%s: binding interrupt failed for %s: %d\n", 736 __func__, gtask->gt_name, error); 737 738 } 739 qgroup->tqg_queue[qid].tgc_cnt++; 740 LIST_INSERT_HEAD(&qgroup->tqg_queue[qid].tgc_tasks, gtask, gt_list); 741 MPASS(qgroup->tqg_queue[qid].tgc_taskq != NULL); 742 gtask->gt_taskqueue = qgroup->tqg_queue[qid].tgc_taskq; 743 mtx_unlock(&qgroup->tqg_lock); 744 } 745 746 int 747 taskqgroup_attach_cpu(struct taskqgroup *qgroup, struct grouptask *gtask, 748 void *uniq, int cpu, device_t dev, struct resource *irq, const char *name) 749 { 750 int i, qid, error; 751 752 qid = -1; 753 gtask->gt_uniq = uniq; 754 snprintf(gtask->gt_name, GROUPTASK_NAMELEN, "%s", name ? name : "grouptask"); 755 gtask->gt_dev = dev; 756 gtask->gt_irq = irq; 757 gtask->gt_cpu = cpu; 758 mtx_lock(&qgroup->tqg_lock); 759 if (tqg_smp_started) { 760 for (i = 0; i < qgroup->tqg_cnt; i++) 761 if (qgroup->tqg_queue[i].tgc_cpu == cpu) { 762 qid = i; 763 break; 764 } 765 if (qid == -1) { 766 mtx_unlock(&qgroup->tqg_lock); 767 printf("%s: qid not found for %s cpu=%d\n", __func__, gtask->gt_name, cpu); 768 return (EINVAL); 769 } 770 } else 771 qid = 0; 772 qgroup->tqg_queue[qid].tgc_cnt++; 773 LIST_INSERT_HEAD(&qgroup->tqg_queue[qid].tgc_tasks, gtask, gt_list); 774 gtask->gt_taskqueue = qgroup->tqg_queue[qid].tgc_taskq; 775 cpu = qgroup->tqg_queue[qid].tgc_cpu; 776 mtx_unlock(&qgroup->tqg_lock); 777 778 if (dev != NULL && irq != NULL && tqg_smp_started) { 779 error = bus_bind_intr(dev, irq, cpu); 780 if (error) 781 printf("%s: binding interrupt failed for %s: %d\n", 782 __func__, gtask->gt_name, error); 783 } 784 return (0); 785 } 786 787 static int 788 taskqgroup_attach_cpu_deferred(struct taskqgroup *qgroup, struct grouptask *gtask) 789 { 790 device_t dev; 791 struct resource *irq; 792 int cpu, error, i, qid; 793 794 qid = -1; 795 dev = gtask->gt_dev; 796 irq = gtask->gt_irq; 797 cpu = gtask->gt_cpu; 798 MPASS(tqg_smp_started); 799 mtx_lock(&qgroup->tqg_lock); 800 for (i = 0; i < qgroup->tqg_cnt; i++) 801 if (qgroup->tqg_queue[i].tgc_cpu == cpu) { 802 qid = i; 803 break; 804 } 805 if (qid == -1) { 806 mtx_unlock(&qgroup->tqg_lock); 807 printf("%s: qid not found for %s cpu=%d\n", __func__, gtask->gt_name, cpu); 808 return (EINVAL); 809 } 810 qgroup->tqg_queue[qid].tgc_cnt++; 811 LIST_INSERT_HEAD(&qgroup->tqg_queue[qid].tgc_tasks, gtask, gt_list); 812 MPASS(qgroup->tqg_queue[qid].tgc_taskq != NULL); 813 gtask->gt_taskqueue = qgroup->tqg_queue[qid].tgc_taskq; 814 mtx_unlock(&qgroup->tqg_lock); 815 816 if (dev != NULL && irq != NULL) { 817 error = bus_bind_intr(dev, irq, cpu); 818 if (error) 819 printf("%s: binding interrupt failed for %s: %d\n", 820 __func__, gtask->gt_name, error); 821 } 822 return (0); 823 } 824 825 void 826 taskqgroup_detach(struct taskqgroup *qgroup, struct grouptask *gtask) 827 { 828 int i; 829 830 grouptask_block(gtask); 831 mtx_lock(&qgroup->tqg_lock); 832 for (i = 0; i < qgroup->tqg_cnt; i++) 833 if (qgroup->tqg_queue[i].tgc_taskq == gtask->gt_taskqueue) 834 break; 835 if (i == qgroup->tqg_cnt) 836 panic("%s: task %s not in group", __func__, gtask->gt_name); 837 qgroup->tqg_queue[i].tgc_cnt--; 838 LIST_REMOVE(gtask, gt_list); 839 mtx_unlock(&qgroup->tqg_lock); 840 gtask->gt_taskqueue = NULL; 841 gtask->gt_task.ta_flags &= ~TASK_NOENQUEUE; 842 } 843 844 static void 845 taskqgroup_binder(void *ctx) 846 { 847 struct taskq_bind_task *gtask = (struct taskq_bind_task *)ctx; 848 cpuset_t mask; 849 int error; 850 851 CPU_ZERO(&mask); 852 CPU_SET(gtask->bt_cpuid, &mask); 853 error = cpuset_setthread(curthread->td_tid, &mask); 854 thread_lock(curthread); 855 sched_bind(curthread, gtask->bt_cpuid); 856 thread_unlock(curthread); 857 858 if (error) 859 printf("%s: binding curthread failed: %d\n", __func__, error); 860 free(gtask, M_DEVBUF); 861 } 862 863 static void 864 taskqgroup_bind(struct taskqgroup *qgroup) 865 { 866 struct taskq_bind_task *gtask; 867 int i; 868 869 /* 870 * Bind taskqueue threads to specific CPUs, if they have been assigned 871 * one. 872 */ 873 if (qgroup->tqg_cnt == 1) 874 return; 875 876 for (i = 0; i < qgroup->tqg_cnt; i++) { 877 gtask = malloc(sizeof (*gtask), M_DEVBUF, M_WAITOK); 878 GTASK_INIT(>ask->bt_task, 0, 0, taskqgroup_binder, gtask); 879 gtask->bt_cpuid = qgroup->tqg_queue[i].tgc_cpu; 880 grouptaskqueue_enqueue(qgroup->tqg_queue[i].tgc_taskq, 881 >ask->bt_task); 882 } 883 } 884 885 static void 886 taskqgroup_config_init(void *arg) 887 { 888 struct taskqgroup *qgroup = qgroup_config; 889 LIST_HEAD(, grouptask) gtask_head = LIST_HEAD_INITIALIZER(NULL); 890 891 LIST_SWAP(>ask_head, &qgroup->tqg_queue[0].tgc_tasks, 892 grouptask, gt_list); 893 qgroup->tqg_queue[0].tgc_cnt = 0; 894 taskqgroup_cpu_create(qgroup, 0, 0); 895 896 qgroup->tqg_cnt = 1; 897 qgroup->tqg_stride = 1; 898 } 899 900 SYSINIT(taskqgroup_config_init, SI_SUB_TASKQ, SI_ORDER_SECOND, 901 taskqgroup_config_init, NULL); 902 903 static int 904 _taskqgroup_adjust(struct taskqgroup *qgroup, int cnt, int stride) 905 { 906 LIST_HEAD(, grouptask) gtask_head = LIST_HEAD_INITIALIZER(NULL); 907 struct grouptask *gtask; 908 int i, k, old_cnt, old_cpu, cpu; 909 910 mtx_assert(&qgroup->tqg_lock, MA_OWNED); 911 912 if (cnt < 1 || cnt * stride > mp_ncpus || !tqg_smp_started) { 913 printf("%s: failed cnt: %d stride: %d " 914 "mp_ncpus: %d tqg_smp_started: %d\n", 915 __func__, cnt, stride, mp_ncpus, tqg_smp_started); 916 return (EINVAL); 917 } 918 if (qgroup->tqg_adjusting) { 919 printf("%s failed: adjusting\n", __func__); 920 return (EBUSY); 921 } 922 qgroup->tqg_adjusting = 1; 923 old_cnt = qgroup->tqg_cnt; 924 old_cpu = 0; 925 if (old_cnt < cnt) 926 old_cpu = qgroup->tqg_queue[old_cnt].tgc_cpu; 927 mtx_unlock(&qgroup->tqg_lock); 928 /* 929 * Set up queue for tasks added before boot. 930 */ 931 if (old_cnt == 0) { 932 LIST_SWAP(>ask_head, &qgroup->tqg_queue[0].tgc_tasks, 933 grouptask, gt_list); 934 qgroup->tqg_queue[0].tgc_cnt = 0; 935 } 936 937 /* 938 * If new taskq threads have been added. 939 */ 940 cpu = old_cpu; 941 for (i = old_cnt; i < cnt; i++) { 942 taskqgroup_cpu_create(qgroup, i, cpu); 943 944 for (k = 0; k < stride; k++) 945 cpu = CPU_NEXT(cpu); 946 } 947 mtx_lock(&qgroup->tqg_lock); 948 qgroup->tqg_cnt = cnt; 949 qgroup->tqg_stride = stride; 950 951 /* 952 * Adjust drivers to use new taskqs. 953 */ 954 for (i = 0; i < old_cnt; i++) { 955 while ((gtask = LIST_FIRST(&qgroup->tqg_queue[i].tgc_tasks))) { 956 LIST_REMOVE(gtask, gt_list); 957 qgroup->tqg_queue[i].tgc_cnt--; 958 LIST_INSERT_HEAD(>ask_head, gtask, gt_list); 959 } 960 } 961 mtx_unlock(&qgroup->tqg_lock); 962 963 while ((gtask = LIST_FIRST(>ask_head))) { 964 LIST_REMOVE(gtask, gt_list); 965 if (gtask->gt_cpu == -1) 966 taskqgroup_attach_deferred(qgroup, gtask); 967 else if (taskqgroup_attach_cpu_deferred(qgroup, gtask)) 968 taskqgroup_attach_deferred(qgroup, gtask); 969 } 970 971 #ifdef INVARIANTS 972 mtx_lock(&qgroup->tqg_lock); 973 for (i = 0; i < qgroup->tqg_cnt; i++) { 974 MPASS(qgroup->tqg_queue[i].tgc_taskq != NULL); 975 LIST_FOREACH(gtask, &qgroup->tqg_queue[i].tgc_tasks, gt_list) 976 MPASS(gtask->gt_taskqueue != NULL); 977 } 978 mtx_unlock(&qgroup->tqg_lock); 979 #endif 980 /* 981 * If taskq thread count has been reduced. 982 */ 983 for (i = cnt; i < old_cnt; i++) 984 taskqgroup_cpu_remove(qgroup, i); 985 986 taskqgroup_bind(qgroup); 987 988 mtx_lock(&qgroup->tqg_lock); 989 qgroup->tqg_adjusting = 0; 990 991 return (0); 992 } 993 994 int 995 taskqgroup_adjust(struct taskqgroup *qgroup, int cnt, int stride) 996 { 997 int error; 998 999 mtx_lock(&qgroup->tqg_lock); 1000 error = _taskqgroup_adjust(qgroup, cnt, stride); 1001 mtx_unlock(&qgroup->tqg_lock); 1002 1003 return (error); 1004 } 1005 1006 struct taskqgroup * 1007 taskqgroup_create(const char *name) 1008 { 1009 struct taskqgroup *qgroup; 1010 1011 qgroup = malloc(sizeof(*qgroup), M_GTASKQUEUE, M_WAITOK | M_ZERO); 1012 mtx_init(&qgroup->tqg_lock, "taskqgroup", NULL, MTX_DEF); 1013 qgroup->tqg_name = name; 1014 LIST_INIT(&qgroup->tqg_queue[0].tgc_tasks); 1015 1016 return (qgroup); 1017 } 1018 1019 void 1020 taskqgroup_destroy(struct taskqgroup *qgroup) 1021 { 1022 1023 } 1024 1025 void 1026 taskqgroup_config_gtask_init(void *ctx, struct grouptask *gtask, gtask_fn_t *fn, 1027 const char *name) 1028 { 1029 1030 GROUPTASK_INIT(gtask, 0, fn, ctx); 1031 taskqgroup_attach(qgroup_config, gtask, gtask, NULL, NULL, name); 1032 } 1033 1034 void 1035 taskqgroup_config_gtask_deinit(struct grouptask *gtask) 1036 { 1037 1038 taskqgroup_detach(qgroup_config, gtask); 1039 } 1040