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