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 #ifndef __HAIKU__ 36 #include <sys/cpuset.h> 37 #include <sys/interrupt.h> 38 #endif 39 #include <sys/kernel.h> 40 #include <sys/kthread.h> 41 #include <sys/libkern.h> 42 #include <sys/limits.h> 43 #include <sys/lock.h> 44 #include <sys/malloc.h> 45 #include <sys/mutex.h> 46 #include <sys/proc.h> 47 #ifndef __HAIKU__ 48 #include <sys/sched.h> 49 #endif 50 #include <sys/smp.h> 51 #include <sys/gtaskqueue.h> 52 #ifndef __HAIKU__ 53 #include <sys/unistd.h> 54 #endif 55 #include <machine/stdarg.h> 56 57 static MALLOC_DEFINE(M_GTASKQUEUE, "gtaskqueue", "Group Task Queues"); 58 static void gtaskqueue_thread_enqueue(void *); 59 static void gtaskqueue_thread_loop(void *arg); 60 static int task_is_running(struct gtaskqueue *queue, struct gtask *gtask); 61 static void gtaskqueue_drain_locked(struct gtaskqueue *queue, struct gtask *gtask); 62 63 TASKQGROUP_DEFINE(softirq, mp_ncpus, 1); 64 TASKQGROUP_DEFINE(config, 1, 1); 65 66 struct gtaskqueue_busy { 67 struct gtask *tb_running; 68 TAILQ_ENTRY(gtaskqueue_busy) tb_link; 69 }; 70 71 static struct gtask * const TB_DRAIN_WAITER = (struct gtask *)0x1; 72 73 typedef void (*gtaskqueue_enqueue_fn)(void *context); 74 75 struct gtaskqueue { 76 STAILQ_HEAD(, gtask) tq_queue; 77 gtaskqueue_enqueue_fn tq_enqueue; 78 void *tq_context; 79 char *tq_name; 80 TAILQ_HEAD(, gtaskqueue_busy) tq_active; 81 struct mtx tq_mutex; 82 struct thread **tq_threads; 83 int tq_tcount; 84 int tq_spin; 85 int tq_flags; 86 int tq_callouts; 87 taskqueue_callback_fn tq_callbacks[TASKQUEUE_NUM_CALLBACKS]; 88 void *tq_cb_contexts[TASKQUEUE_NUM_CALLBACKS]; 89 }; 90 91 #define TQ_FLAGS_ACTIVE (1 << 0) 92 #define TQ_FLAGS_BLOCKED (1 << 1) 93 #define TQ_FLAGS_UNLOCKED_ENQUEUE (1 << 2) 94 95 #define DT_CALLOUT_ARMED (1 << 0) 96 97 #define TQ_LOCK(tq) \ 98 do { \ 99 if ((tq)->tq_spin) \ 100 mtx_lock_spin(&(tq)->tq_mutex); \ 101 else \ 102 mtx_lock(&(tq)->tq_mutex); \ 103 } while (0) 104 #define TQ_ASSERT_LOCKED(tq) mtx_assert(&(tq)->tq_mutex, MA_OWNED) 105 106 #define TQ_UNLOCK(tq) \ 107 do { \ 108 if ((tq)->tq_spin) \ 109 mtx_unlock_spin(&(tq)->tq_mutex); \ 110 else \ 111 mtx_unlock(&(tq)->tq_mutex); \ 112 } while (0) 113 #define TQ_ASSERT_UNLOCKED(tq) mtx_assert(&(tq)->tq_mutex, MA_NOTOWNED) 114 115 #ifdef INVARIANTS 116 static void 117 gtask_dump(struct gtask *gtask) 118 { 119 printf("gtask: %p ta_flags=%x ta_priority=%d ta_func=%p ta_context=%p\n", 120 gtask, gtask->ta_flags, gtask->ta_priority, gtask->ta_func, gtask->ta_context); 121 } 122 #endif 123 124 static __inline int 125 TQ_SLEEP(struct gtaskqueue *tq, void *p, struct mtx *m, int pri, const char *wm, 126 int t) 127 { 128 if (tq->tq_spin) 129 return (msleep_spin(p, m, wm, t)); 130 return (msleep(p, m, pri, wm, t)); 131 } 132 133 static struct gtaskqueue * 134 _gtaskqueue_create(const char *name, int mflags, 135 taskqueue_enqueue_fn enqueue, void *context, 136 int mtxflags, const char *mtxname __unused) 137 { 138 struct gtaskqueue *queue; 139 char *tq_name; 140 141 tq_name = malloc(TASKQUEUE_NAMELEN, M_GTASKQUEUE, mflags | M_ZERO); 142 if (!tq_name) 143 return (NULL); 144 145 snprintf(tq_name, TASKQUEUE_NAMELEN, "%s", (name) ? name : "taskqueue"); 146 147 queue = malloc(sizeof(struct gtaskqueue), M_GTASKQUEUE, mflags | M_ZERO); 148 if (!queue) { 149 free(tq_name, M_GTASKQUEUE); 150 return (NULL); 151 } 152 153 STAILQ_INIT(&queue->tq_queue); 154 TAILQ_INIT(&queue->tq_active); 155 queue->tq_enqueue = enqueue; 156 queue->tq_context = context; 157 queue->tq_name = tq_name; 158 queue->tq_spin = (mtxflags & MTX_SPIN) != 0; 159 queue->tq_flags |= TQ_FLAGS_ACTIVE; 160 if (enqueue == gtaskqueue_thread_enqueue) 161 queue->tq_flags |= TQ_FLAGS_UNLOCKED_ENQUEUE; 162 mtx_init(&queue->tq_mutex, tq_name, NULL, mtxflags); 163 164 return (queue); 165 } 166 167 168 /* 169 * Signal a taskqueue thread to terminate. 170 */ 171 static void 172 gtaskqueue_terminate(struct thread **pp, struct gtaskqueue *tq) 173 { 174 175 while (tq->tq_tcount > 0 || tq->tq_callouts > 0) { 176 wakeup(tq); 177 TQ_SLEEP(tq, pp, &tq->tq_mutex, PWAIT, "taskqueue_destroy", 0); 178 } 179 } 180 181 static void 182 gtaskqueue_free(struct gtaskqueue *queue) 183 { 184 185 TQ_LOCK(queue); 186 queue->tq_flags &= ~TQ_FLAGS_ACTIVE; 187 gtaskqueue_terminate(queue->tq_threads, queue); 188 KASSERT(TAILQ_EMPTY(&queue->tq_active), ("Tasks still running?")); 189 KASSERT(queue->tq_callouts == 0, ("Armed timeout tasks")); 190 mtx_destroy(&queue->tq_mutex); 191 free(queue->tq_threads, M_GTASKQUEUE); 192 free(queue->tq_name, M_GTASKQUEUE); 193 free(queue, M_GTASKQUEUE); 194 } 195 196 /* 197 * Wait for all to complete, then prevent it from being enqueued 198 */ 199 void 200 grouptask_block(struct grouptask *grouptask) 201 { 202 struct gtaskqueue *queue = grouptask->gt_taskqueue; 203 struct gtask *gtask = &grouptask->gt_task; 204 205 #ifdef INVARIANTS 206 if (queue == NULL) { 207 gtask_dump(gtask); 208 panic("queue == NULL"); 209 } 210 #endif 211 TQ_LOCK(queue); 212 gtask->ta_flags |= TASK_NOENQUEUE; 213 gtaskqueue_drain_locked(queue, gtask); 214 TQ_UNLOCK(queue); 215 } 216 217 void 218 grouptask_unblock(struct grouptask *grouptask) 219 { 220 struct gtaskqueue *queue = grouptask->gt_taskqueue; 221 struct gtask *gtask = &grouptask->gt_task; 222 223 #ifdef INVARIANTS 224 if (queue == NULL) { 225 gtask_dump(gtask); 226 panic("queue == NULL"); 227 } 228 #endif 229 TQ_LOCK(queue); 230 gtask->ta_flags &= ~TASK_NOENQUEUE; 231 TQ_UNLOCK(queue); 232 } 233 234 int 235 grouptaskqueue_enqueue(struct gtaskqueue *queue, struct gtask *gtask) 236 { 237 #ifdef INVARIANTS 238 if (queue == NULL) { 239 gtask_dump(gtask); 240 panic("queue == NULL"); 241 } 242 #endif 243 TQ_LOCK(queue); 244 if (gtask->ta_flags & TASK_ENQUEUED) { 245 TQ_UNLOCK(queue); 246 return (0); 247 } 248 if (gtask->ta_flags & TASK_NOENQUEUE) { 249 TQ_UNLOCK(queue); 250 return (EAGAIN); 251 } 252 STAILQ_INSERT_TAIL(&queue->tq_queue, gtask, ta_link); 253 gtask->ta_flags |= TASK_ENQUEUED; 254 TQ_UNLOCK(queue); 255 if ((queue->tq_flags & TQ_FLAGS_BLOCKED) == 0) 256 queue->tq_enqueue(queue->tq_context); 257 return (0); 258 } 259 260 static void 261 gtaskqueue_task_nop_fn(void *context) 262 { 263 } 264 265 /* 266 * Block until all currently queued tasks in this taskqueue 267 * have begun execution. Tasks queued during execution of 268 * this function are ignored. 269 */ 270 static void 271 gtaskqueue_drain_tq_queue(struct gtaskqueue *queue) 272 { 273 struct gtask t_barrier; 274 275 if (STAILQ_EMPTY(&queue->tq_queue)) 276 return; 277 278 /* 279 * Enqueue our barrier after all current tasks, but with 280 * the highest priority so that newly queued tasks cannot 281 * pass it. Because of the high priority, we can not use 282 * taskqueue_enqueue_locked directly (which drops the lock 283 * anyway) so just insert it at tail while we have the 284 * queue lock. 285 */ 286 GTASK_INIT(&t_barrier, 0, USHRT_MAX, gtaskqueue_task_nop_fn, &t_barrier); 287 STAILQ_INSERT_TAIL(&queue->tq_queue, &t_barrier, ta_link); 288 t_barrier.ta_flags |= TASK_ENQUEUED; 289 290 /* 291 * Once the barrier has executed, all previously queued tasks 292 * have completed or are currently executing. 293 */ 294 while (t_barrier.ta_flags & TASK_ENQUEUED) 295 TQ_SLEEP(queue, &t_barrier, &queue->tq_mutex, PWAIT, "-", 0); 296 } 297 298 /* 299 * Block until all currently executing tasks for this taskqueue 300 * complete. Tasks that begin execution during the execution 301 * of this function are ignored. 302 */ 303 static void 304 gtaskqueue_drain_tq_active(struct gtaskqueue *queue) 305 { 306 struct gtaskqueue_busy tb_marker, *tb_first; 307 308 if (TAILQ_EMPTY(&queue->tq_active)) 309 return; 310 311 /* Block taskq_terminate().*/ 312 queue->tq_callouts++; 313 314 /* 315 * Wait for all currently executing taskqueue threads 316 * to go idle. 317 */ 318 tb_marker.tb_running = TB_DRAIN_WAITER; 319 TAILQ_INSERT_TAIL(&queue->tq_active, &tb_marker, tb_link); 320 while (TAILQ_FIRST(&queue->tq_active) != &tb_marker) 321 TQ_SLEEP(queue, &tb_marker, &queue->tq_mutex, PWAIT, "-", 0); 322 TAILQ_REMOVE(&queue->tq_active, &tb_marker, tb_link); 323 324 /* 325 * Wakeup any other drain waiter that happened to queue up 326 * without any intervening active thread. 327 */ 328 tb_first = TAILQ_FIRST(&queue->tq_active); 329 if (tb_first != NULL && tb_first->tb_running == TB_DRAIN_WAITER) 330 wakeup(tb_first); 331 332 /* Release taskqueue_terminate(). */ 333 queue->tq_callouts--; 334 if ((queue->tq_flags & TQ_FLAGS_ACTIVE) == 0) 335 wakeup_one(queue->tq_threads); 336 } 337 338 void 339 gtaskqueue_block(struct gtaskqueue *queue) 340 { 341 342 TQ_LOCK(queue); 343 queue->tq_flags |= TQ_FLAGS_BLOCKED; 344 TQ_UNLOCK(queue); 345 } 346 347 void 348 gtaskqueue_unblock(struct gtaskqueue *queue) 349 { 350 351 TQ_LOCK(queue); 352 queue->tq_flags &= ~TQ_FLAGS_BLOCKED; 353 if (!STAILQ_EMPTY(&queue->tq_queue)) 354 queue->tq_enqueue(queue->tq_context); 355 TQ_UNLOCK(queue); 356 } 357 358 static void 359 gtaskqueue_run_locked(struct gtaskqueue *queue) 360 { 361 struct gtaskqueue_busy tb; 362 struct gtaskqueue_busy *tb_first; 363 struct gtask *gtask; 364 365 KASSERT(queue != NULL, ("tq is NULL")); 366 TQ_ASSERT_LOCKED(queue); 367 tb.tb_running = NULL; 368 369 while (STAILQ_FIRST(&queue->tq_queue)) { 370 TAILQ_INSERT_TAIL(&queue->tq_active, &tb, tb_link); 371 372 /* 373 * Carefully remove the first task from the queue and 374 * clear its TASK_ENQUEUED flag 375 */ 376 gtask = STAILQ_FIRST(&queue->tq_queue); 377 KASSERT(gtask != NULL, ("task is NULL")); 378 STAILQ_REMOVE_HEAD(&queue->tq_queue, ta_link); 379 gtask->ta_flags &= ~TASK_ENQUEUED; 380 tb.tb_running = gtask; 381 TQ_UNLOCK(queue); 382 383 KASSERT(gtask->ta_func != NULL, ("task->ta_func is NULL")); 384 gtask->ta_func(gtask->ta_context); 385 386 TQ_LOCK(queue); 387 tb.tb_running = NULL; 388 wakeup(gtask); 389 390 TAILQ_REMOVE(&queue->tq_active, &tb, tb_link); 391 tb_first = TAILQ_FIRST(&queue->tq_active); 392 if (tb_first != NULL && 393 tb_first->tb_running == TB_DRAIN_WAITER) 394 wakeup(tb_first); 395 } 396 } 397 398 static int 399 task_is_running(struct gtaskqueue *queue, struct gtask *gtask) 400 { 401 struct gtaskqueue_busy *tb; 402 403 TQ_ASSERT_LOCKED(queue); 404 TAILQ_FOREACH(tb, &queue->tq_active, tb_link) { 405 if (tb->tb_running == gtask) 406 return (1); 407 } 408 return (0); 409 } 410 411 static int 412 gtaskqueue_cancel_locked(struct gtaskqueue *queue, struct gtask *gtask) 413 { 414 415 if (gtask->ta_flags & TASK_ENQUEUED) 416 STAILQ_REMOVE(&queue->tq_queue, gtask, gtask, ta_link); 417 gtask->ta_flags &= ~TASK_ENQUEUED; 418 return (task_is_running(queue, gtask) ? EBUSY : 0); 419 } 420 421 int 422 gtaskqueue_cancel(struct gtaskqueue *queue, struct gtask *gtask) 423 { 424 int error; 425 426 TQ_LOCK(queue); 427 error = gtaskqueue_cancel_locked(queue, gtask); 428 TQ_UNLOCK(queue); 429 430 return (error); 431 } 432 433 static void 434 gtaskqueue_drain_locked(struct gtaskqueue *queue, struct gtask *gtask) 435 { 436 while ((gtask->ta_flags & TASK_ENQUEUED) || task_is_running(queue, gtask)) 437 TQ_SLEEP(queue, gtask, &queue->tq_mutex, PWAIT, "-", 0); 438 } 439 440 void 441 gtaskqueue_drain(struct gtaskqueue *queue, struct gtask *gtask) 442 { 443 #ifndef __HAIKU__ 444 if (!queue->tq_spin) 445 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, __func__); 446 #endif 447 448 TQ_LOCK(queue); 449 gtaskqueue_drain_locked(queue, gtask); 450 TQ_UNLOCK(queue); 451 } 452 453 void 454 gtaskqueue_drain_all(struct gtaskqueue *queue) 455 { 456 #ifndef __HAIKU__ 457 if (!queue->tq_spin) 458 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, __func__); 459 #endif 460 461 TQ_LOCK(queue); 462 gtaskqueue_drain_tq_queue(queue); 463 gtaskqueue_drain_tq_active(queue); 464 TQ_UNLOCK(queue); 465 } 466 467 static int 468 _gtaskqueue_start_threads(struct gtaskqueue **tqp, int count, int pri, 469 void* mask, const char *name, va_list ap) 470 { 471 char ktname[19 + 1]; 472 struct thread *td; 473 struct gtaskqueue *tq; 474 int i, error; 475 476 if (count <= 0) 477 return (EINVAL); 478 479 vsnprintf(ktname, sizeof(ktname), name, ap); 480 tq = *tqp; 481 482 tq->tq_threads = malloc(sizeof(struct thread *) * count, M_GTASKQUEUE, 483 M_NOWAIT | M_ZERO); 484 if (tq->tq_threads == NULL) { 485 printf("%s: no memory for %s threads\n", __func__, ktname); 486 return (ENOMEM); 487 } 488 489 for (i = 0; i < count; i++) { 490 if (count == 1) 491 error = kthread_add(gtaskqueue_thread_loop, tqp, NULL, 492 &tq->tq_threads[i], 0, 0, "%s", ktname); 493 else 494 error = kthread_add(gtaskqueue_thread_loop, tqp, NULL, 495 &tq->tq_threads[i], 0, 0, 496 "%s_%d", ktname, i); 497 if (error) { 498 /* should be ok to continue, taskqueue_free will dtrt */ 499 printf("%s: kthread_add(%s): error %d", __func__, 500 ktname, error); 501 tq->tq_threads[i] = NULL; /* paranoid */ 502 } else 503 tq->tq_tcount++; 504 } 505 for (i = 0; i < count; i++) { 506 if (tq->tq_threads[i] == NULL) 507 continue; 508 td = tq->tq_threads[i]; 509 #ifndef __HAIKU__ 510 if (mask) { 511 error = cpuset_setthread(td->td_tid, mask); 512 /* 513 * Failing to pin is rarely an actual fatal error; 514 * it'll just affect performance. 515 */ 516 if (error) 517 printf("%s: curthread=%llu: can't pin; " 518 "error=%d\n", 519 __func__, 520 (unsigned long long) td->td_tid, 521 error); 522 } 523 #endif 524 thread_lock(td); 525 sched_prio(td, pri); 526 sched_add(td, SRQ_BORING); 527 thread_unlock(td); 528 } 529 530 return (0); 531 } 532 533 static int 534 gtaskqueue_start_threads(struct gtaskqueue **tqp, int count, int pri, 535 const char *name, ...) 536 { 537 va_list ap; 538 int error; 539 540 va_start(ap, name); 541 error = _gtaskqueue_start_threads(tqp, count, pri, NULL, name, ap); 542 va_end(ap); 543 return (error); 544 } 545 546 static inline void 547 gtaskqueue_run_callback(struct gtaskqueue *tq, 548 enum taskqueue_callback_type cb_type) 549 { 550 taskqueue_callback_fn tq_callback; 551 552 TQ_ASSERT_UNLOCKED(tq); 553 tq_callback = tq->tq_callbacks[cb_type]; 554 if (tq_callback != NULL) 555 tq_callback(tq->tq_cb_contexts[cb_type]); 556 } 557 558 static void 559 gtaskqueue_thread_loop(void *arg) 560 { 561 struct gtaskqueue **tqp, *tq; 562 563 tqp = arg; 564 tq = *tqp; 565 gtaskqueue_run_callback(tq, TASKQUEUE_CALLBACK_TYPE_INIT); 566 TQ_LOCK(tq); 567 while ((tq->tq_flags & TQ_FLAGS_ACTIVE) != 0) { 568 /* XXX ? */ 569 gtaskqueue_run_locked(tq); 570 /* 571 * Because taskqueue_run() can drop tq_mutex, we need to 572 * check if the TQ_FLAGS_ACTIVE flag wasn't removed in the 573 * meantime, which means we missed a wakeup. 574 */ 575 if ((tq->tq_flags & TQ_FLAGS_ACTIVE) == 0) 576 break; 577 TQ_SLEEP(tq, tq, &tq->tq_mutex, 0, "-", 0); 578 } 579 gtaskqueue_run_locked(tq); 580 /* 581 * This thread is on its way out, so just drop the lock temporarily 582 * in order to call the shutdown callback. This allows the callback 583 * to look at the taskqueue, even just before it dies. 584 */ 585 TQ_UNLOCK(tq); 586 gtaskqueue_run_callback(tq, TASKQUEUE_CALLBACK_TYPE_SHUTDOWN); 587 TQ_LOCK(tq); 588 589 /* rendezvous with thread that asked us to terminate */ 590 tq->tq_tcount--; 591 wakeup_one(tq->tq_threads); 592 TQ_UNLOCK(tq); 593 kthread_exit(); 594 } 595 596 static void 597 gtaskqueue_thread_enqueue(void *context) 598 { 599 struct gtaskqueue **tqp, *tq; 600 601 tqp = context; 602 tq = *tqp; 603 wakeup_one(tq); 604 } 605 606 607 static struct gtaskqueue * 608 gtaskqueue_create_fast(const char *name, int mflags, 609 taskqueue_enqueue_fn enqueue, void *context) 610 { 611 return _gtaskqueue_create(name, mflags, enqueue, context, 612 MTX_SPIN, "fast_taskqueue"); 613 } 614 615 616 struct taskqgroup_cpu { 617 LIST_HEAD(, grouptask) tgc_tasks; 618 struct gtaskqueue *tgc_taskq; 619 int tgc_cnt; 620 int tgc_cpu; 621 }; 622 623 struct taskqgroup { 624 struct taskqgroup_cpu tqg_queue[MAXCPU]; 625 struct mtx tqg_lock; 626 const char * tqg_name; 627 int tqg_adjusting; 628 int tqg_stride; 629 int tqg_cnt; 630 }; 631 632 struct taskq_bind_task { 633 struct gtask bt_task; 634 int bt_cpuid; 635 }; 636 637 static void 638 taskqgroup_cpu_create(struct taskqgroup *qgroup, int idx, int cpu) 639 { 640 struct taskqgroup_cpu *qcpu; 641 642 qcpu = &qgroup->tqg_queue[idx]; 643 LIST_INIT(&qcpu->tgc_tasks); 644 qcpu->tgc_taskq = gtaskqueue_create_fast(NULL, M_WAITOK, 645 gtaskqueue_thread_enqueue, &qcpu->tgc_taskq); 646 MPASS(qcpu->tgc_taskq); 647 gtaskqueue_start_threads(&qcpu->tgc_taskq, 1, PI_SOFT, 648 "%s_%d", qgroup->tqg_name, idx); 649 qcpu->tgc_cpu = cpu; 650 } 651 652 static void 653 taskqgroup_cpu_remove(struct taskqgroup *qgroup, int idx) 654 { 655 656 gtaskqueue_free(qgroup->tqg_queue[idx].tgc_taskq); 657 } 658 659 /* 660 * Find the taskq with least # of tasks that doesn't currently have any 661 * other queues from the uniq identifier. 662 */ 663 static int 664 taskqgroup_find(struct taskqgroup *qgroup, void *uniq) 665 { 666 struct grouptask *n; 667 int i, idx, mincnt; 668 int strict; 669 670 mtx_assert(&qgroup->tqg_lock, MA_OWNED); 671 #ifndef __HAIKU__ 672 if (qgroup->tqg_cnt == 0) 673 #else 674 KASSERT(qgroup->tqg_cnt > 0, ("qgroup(%p)->tqg_cnt is %d!", qgroup, qgroup->tqg_cnt)); 675 if (qgroup->tqg_cnt == 1) 676 #endif 677 return (0); 678 idx = -1; 679 mincnt = INT_MAX; 680 /* 681 * Two passes; First scan for a queue with the least tasks that 682 * does not already service this uniq id. If that fails simply find 683 * the queue with the least total tasks; 684 */ 685 for (strict = 1; mincnt == INT_MAX; strict = 0) { 686 for (i = 0; i < qgroup->tqg_cnt; i++) { 687 if (qgroup->tqg_queue[i].tgc_cnt > mincnt) 688 continue; 689 if (strict) { 690 LIST_FOREACH(n, 691 &qgroup->tqg_queue[i].tgc_tasks, gt_list) 692 if (n->gt_uniq == uniq) 693 break; 694 if (n != NULL) 695 continue; 696 } 697 mincnt = qgroup->tqg_queue[i].tgc_cnt; 698 idx = i; 699 } 700 } 701 if (idx == -1) 702 panic("%s: failed to pick a qid.", __func__); 703 704 return (idx); 705 } 706 707 /* 708 * smp_started is unusable since it is not set for UP kernels or even for 709 * SMP kernels when there is 1 CPU. This is usually handled by adding a 710 * (mp_ncpus == 1) test, but that would be broken here since we need to 711 * to synchronize with the SI_SUB_SMP ordering. Even in the pure SMP case 712 * smp_started only gives a fuzzy ordering relative to SI_SUB_SMP. 713 * 714 * So maintain our own flag. It must be set after all CPUs are started 715 * and before SI_SUB_SMP:SI_ORDER_ANY so that the SYSINIT for delayed 716 * adjustment is properly delayed. SI_ORDER_FOURTH is clearly before 717 * SI_ORDER_ANY and unclearly after the CPUs are started. It would be 718 * simpler for adjustment to pass a flag indicating if it is delayed. 719 */ 720 721 static int tqg_smp_started; 722 723 static void 724 tqg_record_smp_started(void *arg) 725 { 726 tqg_smp_started = 1; 727 } 728 729 SYSINIT(tqg_record_smp_started, SI_SUB_SMP, SI_ORDER_FOURTH, 730 tqg_record_smp_started, NULL); 731 732 void 733 taskqgroup_attach(struct taskqgroup *qgroup, struct grouptask *gtask, 734 void *uniq, device_t dev, struct resource *irq, const char *name) 735 { 736 int cpu, qid, error; 737 738 gtask->gt_uniq = uniq; 739 snprintf(gtask->gt_name, GROUPTASK_NAMELEN, "%s", name ? name : "grouptask"); 740 gtask->gt_dev = dev; 741 gtask->gt_irq = irq; 742 gtask->gt_cpu = -1; 743 mtx_lock(&qgroup->tqg_lock); 744 qid = taskqgroup_find(qgroup, uniq); 745 qgroup->tqg_queue[qid].tgc_cnt++; 746 LIST_INSERT_HEAD(&qgroup->tqg_queue[qid].tgc_tasks, gtask, gt_list); 747 gtask->gt_taskqueue = qgroup->tqg_queue[qid].tgc_taskq; 748 if (dev != NULL && irq != NULL && tqg_smp_started) { 749 cpu = qgroup->tqg_queue[qid].tgc_cpu; 750 gtask->gt_cpu = cpu; 751 mtx_unlock(&qgroup->tqg_lock); 752 error = bus_bind_intr(dev, irq, cpu); 753 if (error) 754 printf("%s: binding interrupt failed for %s: %d\n", 755 __func__, gtask->gt_name, error); 756 } else 757 mtx_unlock(&qgroup->tqg_lock); 758 } 759 760 static void 761 taskqgroup_attach_deferred(struct taskqgroup *qgroup, struct grouptask *gtask) 762 { 763 int qid, cpu, error; 764 765 mtx_lock(&qgroup->tqg_lock); 766 qid = taskqgroup_find(qgroup, gtask->gt_uniq); 767 cpu = qgroup->tqg_queue[qid].tgc_cpu; 768 if (gtask->gt_dev != NULL && gtask->gt_irq != NULL) { 769 mtx_unlock(&qgroup->tqg_lock); 770 error = bus_bind_intr(gtask->gt_dev, gtask->gt_irq, cpu); 771 mtx_lock(&qgroup->tqg_lock); 772 if (error) 773 printf("%s: binding interrupt failed for %s: %d\n", 774 __func__, gtask->gt_name, error); 775 776 } 777 qgroup->tqg_queue[qid].tgc_cnt++; 778 LIST_INSERT_HEAD(&qgroup->tqg_queue[qid].tgc_tasks, gtask, gt_list); 779 MPASS(qgroup->tqg_queue[qid].tgc_taskq != NULL); 780 gtask->gt_taskqueue = qgroup->tqg_queue[qid].tgc_taskq; 781 mtx_unlock(&qgroup->tqg_lock); 782 } 783 784 int 785 taskqgroup_attach_cpu(struct taskqgroup *qgroup, struct grouptask *gtask, 786 void *uniq, int cpu, device_t dev, struct resource *irq, const char *name) 787 { 788 int i, qid, error; 789 790 qid = -1; 791 gtask->gt_uniq = uniq; 792 snprintf(gtask->gt_name, GROUPTASK_NAMELEN, "%s", name ? name : "grouptask"); 793 gtask->gt_dev = dev; 794 gtask->gt_irq = irq; 795 gtask->gt_cpu = cpu; 796 mtx_lock(&qgroup->tqg_lock); 797 if (tqg_smp_started) { 798 for (i = 0; i < qgroup->tqg_cnt; i++) 799 if (qgroup->tqg_queue[i].tgc_cpu == cpu) { 800 qid = i; 801 break; 802 } 803 if (qid == -1) { 804 mtx_unlock(&qgroup->tqg_lock); 805 printf("%s: qid not found for %s cpu=%d\n", __func__, gtask->gt_name, cpu); 806 return (EINVAL); 807 } 808 } else 809 qid = 0; 810 qgroup->tqg_queue[qid].tgc_cnt++; 811 LIST_INSERT_HEAD(&qgroup->tqg_queue[qid].tgc_tasks, gtask, gt_list); 812 gtask->gt_taskqueue = qgroup->tqg_queue[qid].tgc_taskq; 813 cpu = qgroup->tqg_queue[qid].tgc_cpu; 814 mtx_unlock(&qgroup->tqg_lock); 815 816 if (dev != NULL && irq != NULL && tqg_smp_started) { 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 static int 826 taskqgroup_attach_cpu_deferred(struct taskqgroup *qgroup, struct grouptask *gtask) 827 { 828 device_t dev; 829 struct resource *irq; 830 int cpu, error, i, qid; 831 832 qid = -1; 833 dev = gtask->gt_dev; 834 irq = gtask->gt_irq; 835 cpu = gtask->gt_cpu; 836 MPASS(tqg_smp_started); 837 mtx_lock(&qgroup->tqg_lock); 838 for (i = 0; i < qgroup->tqg_cnt; i++) 839 if (qgroup->tqg_queue[i].tgc_cpu == cpu) { 840 qid = i; 841 break; 842 } 843 if (qid == -1) { 844 mtx_unlock(&qgroup->tqg_lock); 845 printf("%s: qid not found for %s cpu=%d\n", __func__, gtask->gt_name, cpu); 846 return (EINVAL); 847 } 848 qgroup->tqg_queue[qid].tgc_cnt++; 849 LIST_INSERT_HEAD(&qgroup->tqg_queue[qid].tgc_tasks, gtask, gt_list); 850 MPASS(qgroup->tqg_queue[qid].tgc_taskq != NULL); 851 gtask->gt_taskqueue = qgroup->tqg_queue[qid].tgc_taskq; 852 mtx_unlock(&qgroup->tqg_lock); 853 854 if (dev != NULL && irq != NULL) { 855 error = bus_bind_intr(dev, irq, cpu); 856 if (error) 857 printf("%s: binding interrupt failed for %s: %d\n", 858 __func__, gtask->gt_name, error); 859 } 860 return (0); 861 } 862 863 void 864 taskqgroup_detach(struct taskqgroup *qgroup, struct grouptask *gtask) 865 { 866 int i; 867 868 grouptask_block(gtask); 869 mtx_lock(&qgroup->tqg_lock); 870 for (i = 0; i < qgroup->tqg_cnt; i++) 871 if (qgroup->tqg_queue[i].tgc_taskq == gtask->gt_taskqueue) 872 break; 873 if (i == qgroup->tqg_cnt) 874 panic("%s: task %s not in group", __func__, gtask->gt_name); 875 qgroup->tqg_queue[i].tgc_cnt--; 876 LIST_REMOVE(gtask, gt_list); 877 mtx_unlock(&qgroup->tqg_lock); 878 gtask->gt_taskqueue = NULL; 879 gtask->gt_task.ta_flags &= ~TASK_NOENQUEUE; 880 } 881 882 static void 883 taskqgroup_binder(void *ctx) 884 { 885 struct taskq_bind_task *gtask = (struct taskq_bind_task *)ctx; 886 #ifndef __HAIKU__ 887 cpuset_t mask; 888 int error; 889 890 CPU_ZERO(&mask); 891 CPU_SET(gtask->bt_cpuid, &mask); 892 error = cpuset_setthread(curthread->td_tid, &mask); 893 thread_lock(curthread); 894 sched_bind(curthread, gtask->bt_cpuid); 895 thread_unlock(curthread); 896 897 if (error) 898 printf("%s: binding curthread failed: %d\n", __func__, error); 899 #endif 900 free(gtask, M_DEVBUF); 901 } 902 903 static void 904 taskqgroup_bind(struct taskqgroup *qgroup) 905 { 906 struct taskq_bind_task *gtask; 907 int i; 908 909 /* 910 * Bind taskqueue threads to specific CPUs, if they have been assigned 911 * one. 912 */ 913 if (qgroup->tqg_cnt == 1) 914 return; 915 916 for (i = 0; i < qgroup->tqg_cnt; i++) { 917 gtask = malloc(sizeof (*gtask), M_DEVBUF, M_WAITOK); 918 GTASK_INIT(>ask->bt_task, 0, 0, taskqgroup_binder, gtask); 919 gtask->bt_cpuid = qgroup->tqg_queue[i].tgc_cpu; 920 grouptaskqueue_enqueue(qgroup->tqg_queue[i].tgc_taskq, 921 >ask->bt_task); 922 } 923 } 924 925 static void 926 taskqgroup_config_init(void *arg) 927 { 928 struct taskqgroup *qgroup = qgroup_config; 929 LIST_HEAD(, grouptask) gtask_head = LIST_HEAD_INITIALIZER(NULL); 930 931 LIST_SWAP(>ask_head, &qgroup->tqg_queue[0].tgc_tasks, 932 grouptask, gt_list); 933 qgroup->tqg_queue[0].tgc_cnt = 0; 934 taskqgroup_cpu_create(qgroup, 0, 0); 935 936 qgroup->tqg_cnt = 1; 937 qgroup->tqg_stride = 1; 938 } 939 940 SYSINIT(taskqgroup_config_init, SI_SUB_TASKQ, SI_ORDER_SECOND, 941 taskqgroup_config_init, NULL); 942 943 static int 944 _taskqgroup_adjust(struct taskqgroup *qgroup, int cnt, int stride) 945 { 946 LIST_HEAD(, grouptask) gtask_head = LIST_HEAD_INITIALIZER(NULL); 947 struct grouptask *gtask; 948 int i, k, old_cnt, old_cpu, cpu; 949 950 mtx_assert(&qgroup->tqg_lock, MA_OWNED); 951 952 if (cnt < 1 || cnt * stride > mp_ncpus || !tqg_smp_started) { 953 printf("%s: failed cnt: %d stride: %d " 954 "mp_ncpus: %d tqg_smp_started: %d\n", 955 __func__, cnt, stride, mp_ncpus, tqg_smp_started); 956 return (EINVAL); 957 } 958 if (qgroup->tqg_adjusting) { 959 printf("%s failed: adjusting\n", __func__); 960 return (EBUSY); 961 } 962 qgroup->tqg_adjusting = 1; 963 old_cnt = qgroup->tqg_cnt; 964 old_cpu = 0; 965 if (old_cnt < cnt) 966 old_cpu = qgroup->tqg_queue[old_cnt].tgc_cpu; 967 mtx_unlock(&qgroup->tqg_lock); 968 /* 969 * Set up queue for tasks added before boot. 970 */ 971 if (old_cnt == 0) { 972 LIST_SWAP(>ask_head, &qgroup->tqg_queue[0].tgc_tasks, 973 grouptask, gt_list); 974 qgroup->tqg_queue[0].tgc_cnt = 0; 975 } 976 977 /* 978 * If new taskq threads have been added. 979 */ 980 cpu = old_cpu; 981 for (i = old_cnt; i < cnt; i++) { 982 taskqgroup_cpu_create(qgroup, i, cpu); 983 984 for (k = 0; k < stride; k++) 985 cpu = CPU_NEXT(cpu); 986 } 987 mtx_lock(&qgroup->tqg_lock); 988 qgroup->tqg_cnt = cnt; 989 qgroup->tqg_stride = stride; 990 991 /* 992 * Adjust drivers to use new taskqs. 993 */ 994 for (i = 0; i < old_cnt; i++) { 995 while ((gtask = LIST_FIRST(&qgroup->tqg_queue[i].tgc_tasks))) { 996 LIST_REMOVE(gtask, gt_list); 997 qgroup->tqg_queue[i].tgc_cnt--; 998 LIST_INSERT_HEAD(>ask_head, gtask, gt_list); 999 } 1000 } 1001 mtx_unlock(&qgroup->tqg_lock); 1002 1003 while ((gtask = LIST_FIRST(>ask_head))) { 1004 LIST_REMOVE(gtask, gt_list); 1005 if (gtask->gt_cpu == -1) 1006 taskqgroup_attach_deferred(qgroup, gtask); 1007 else if (taskqgroup_attach_cpu_deferred(qgroup, gtask)) 1008 taskqgroup_attach_deferred(qgroup, gtask); 1009 } 1010 1011 #ifdef INVARIANTS 1012 mtx_lock(&qgroup->tqg_lock); 1013 for (i = 0; i < qgroup->tqg_cnt; i++) { 1014 MPASS(qgroup->tqg_queue[i].tgc_taskq != NULL); 1015 LIST_FOREACH(gtask, &qgroup->tqg_queue[i].tgc_tasks, gt_list) 1016 MPASS(gtask->gt_taskqueue != NULL); 1017 } 1018 mtx_unlock(&qgroup->tqg_lock); 1019 #endif 1020 /* 1021 * If taskq thread count has been reduced. 1022 */ 1023 for (i = cnt; i < old_cnt; i++) 1024 taskqgroup_cpu_remove(qgroup, i); 1025 1026 taskqgroup_bind(qgroup); 1027 1028 mtx_lock(&qgroup->tqg_lock); 1029 qgroup->tqg_adjusting = 0; 1030 1031 return (0); 1032 } 1033 1034 int 1035 taskqgroup_adjust(struct taskqgroup *qgroup, int cnt, int stride) 1036 { 1037 int error; 1038 1039 mtx_lock(&qgroup->tqg_lock); 1040 error = _taskqgroup_adjust(qgroup, cnt, stride); 1041 mtx_unlock(&qgroup->tqg_lock); 1042 1043 return (error); 1044 } 1045 1046 struct taskqgroup * 1047 taskqgroup_create(const char *name) 1048 { 1049 struct taskqgroup *qgroup; 1050 1051 qgroup = malloc(sizeof(*qgroup), M_GTASKQUEUE, M_WAITOK | M_ZERO); 1052 mtx_init(&qgroup->tqg_lock, "taskqgroup", NULL, MTX_DEF); 1053 qgroup->tqg_name = name; 1054 LIST_INIT(&qgroup->tqg_queue[0].tgc_tasks); 1055 1056 return (qgroup); 1057 } 1058 1059 void 1060 taskqgroup_destroy(struct taskqgroup *qgroup) 1061 { 1062 1063 } 1064 1065 void 1066 taskqgroup_config_gtask_init(void *ctx, struct grouptask *gtask, gtask_fn_t *fn, 1067 const char *name) 1068 { 1069 1070 GROUPTASK_INIT(gtask, 0, fn, ctx); 1071 taskqgroup_attach(qgroup_config, gtask, gtask, NULL, NULL, name); 1072 } 1073 1074 void 1075 taskqgroup_config_gtask_deinit(struct grouptask *gtask) 1076 { 1077 1078 taskqgroup_detach(qgroup_config, gtask); 1079 } 1080