Jens Axboe | 771b53d0 | 2019-10-22 10:25:58 -0600 | [diff] [blame] | 1 | // SPDX-License-Identifier: GPL-2.0 |
| 2 | /* |
| 3 | * Basic worker thread pool for io_uring |
| 4 | * |
| 5 | * Copyright (C) 2019 Jens Axboe |
| 6 | * |
| 7 | */ |
| 8 | #include <linux/kernel.h> |
| 9 | #include <linux/init.h> |
| 10 | #include <linux/errno.h> |
| 11 | #include <linux/sched/signal.h> |
| 12 | #include <linux/mm.h> |
| 13 | #include <linux/mmu_context.h> |
| 14 | #include <linux/sched/mm.h> |
| 15 | #include <linux/percpu.h> |
| 16 | #include <linux/slab.h> |
| 17 | #include <linux/kthread.h> |
| 18 | #include <linux/rculist_nulls.h> |
| 19 | |
| 20 | #include "io-wq.h" |
| 21 | |
| 22 | #define WORKER_IDLE_TIMEOUT (5 * HZ) |
| 23 | |
| 24 | enum { |
| 25 | IO_WORKER_F_UP = 1, /* up and active */ |
| 26 | IO_WORKER_F_RUNNING = 2, /* account as running */ |
| 27 | IO_WORKER_F_FREE = 4, /* worker on free list */ |
| 28 | IO_WORKER_F_EXITING = 8, /* worker exiting */ |
| 29 | IO_WORKER_F_FIXED = 16, /* static idle worker */ |
| 30 | }; |
| 31 | |
| 32 | enum { |
| 33 | IO_WQ_BIT_EXIT = 0, /* wq exiting */ |
| 34 | IO_WQ_BIT_CANCEL = 1, /* cancel work on list */ |
| 35 | }; |
| 36 | |
| 37 | enum { |
| 38 | IO_WQE_FLAG_STALLED = 1, /* stalled on hash */ |
| 39 | }; |
| 40 | |
| 41 | /* |
| 42 | * One for each thread in a wqe pool |
| 43 | */ |
| 44 | struct io_worker { |
| 45 | refcount_t ref; |
| 46 | unsigned flags; |
| 47 | struct hlist_nulls_node nulls_node; |
| 48 | struct task_struct *task; |
| 49 | wait_queue_head_t wait; |
| 50 | struct io_wqe *wqe; |
| 51 | struct io_wq_work *cur_work; |
| 52 | |
| 53 | struct rcu_head rcu; |
| 54 | struct mm_struct *mm; |
Jens Axboe | fcb323c | 2019-10-24 12:39:47 -0600 | [diff] [blame^] | 55 | struct files_struct *restore_files; |
Jens Axboe | 771b53d0 | 2019-10-22 10:25:58 -0600 | [diff] [blame] | 56 | }; |
| 57 | |
| 58 | struct io_wq_nulls_list { |
| 59 | struct hlist_nulls_head head; |
| 60 | unsigned long nulls; |
| 61 | }; |
| 62 | |
| 63 | #if BITS_PER_LONG == 64 |
| 64 | #define IO_WQ_HASH_ORDER 6 |
| 65 | #else |
| 66 | #define IO_WQ_HASH_ORDER 5 |
| 67 | #endif |
| 68 | |
| 69 | /* |
| 70 | * Per-node worker thread pool |
| 71 | */ |
| 72 | struct io_wqe { |
| 73 | struct { |
| 74 | spinlock_t lock; |
| 75 | struct list_head work_list; |
| 76 | unsigned long hash_map; |
| 77 | unsigned flags; |
| 78 | } ____cacheline_aligned_in_smp; |
| 79 | |
| 80 | int node; |
| 81 | unsigned nr_workers; |
| 82 | unsigned max_workers; |
| 83 | atomic_t nr_running; |
| 84 | |
| 85 | struct io_wq_nulls_list free_list; |
| 86 | struct io_wq_nulls_list busy_list; |
| 87 | |
| 88 | struct io_wq *wq; |
| 89 | }; |
| 90 | |
| 91 | /* |
| 92 | * Per io_wq state |
| 93 | */ |
| 94 | struct io_wq { |
| 95 | struct io_wqe **wqes; |
| 96 | unsigned long state; |
| 97 | unsigned nr_wqes; |
| 98 | |
| 99 | struct task_struct *manager; |
| 100 | struct mm_struct *mm; |
| 101 | refcount_t refs; |
| 102 | struct completion done; |
| 103 | }; |
| 104 | |
| 105 | static void io_wq_free_worker(struct rcu_head *head) |
| 106 | { |
| 107 | struct io_worker *worker = container_of(head, struct io_worker, rcu); |
| 108 | |
| 109 | kfree(worker); |
| 110 | } |
| 111 | |
| 112 | static bool io_worker_get(struct io_worker *worker) |
| 113 | { |
| 114 | return refcount_inc_not_zero(&worker->ref); |
| 115 | } |
| 116 | |
| 117 | static void io_worker_release(struct io_worker *worker) |
| 118 | { |
| 119 | if (refcount_dec_and_test(&worker->ref)) |
| 120 | wake_up_process(worker->task); |
| 121 | } |
| 122 | |
| 123 | /* |
| 124 | * Note: drops the wqe->lock if returning true! The caller must re-acquire |
| 125 | * the lock in that case. Some callers need to restart handling if this |
| 126 | * happens, so we can't just re-acquire the lock on behalf of the caller. |
| 127 | */ |
| 128 | static bool __io_worker_unuse(struct io_wqe *wqe, struct io_worker *worker) |
| 129 | { |
Jens Axboe | fcb323c | 2019-10-24 12:39:47 -0600 | [diff] [blame^] | 130 | bool dropped_lock = false; |
| 131 | |
| 132 | if (current->files != worker->restore_files) { |
| 133 | __acquire(&wqe->lock); |
| 134 | spin_unlock_irq(&wqe->lock); |
| 135 | dropped_lock = true; |
| 136 | |
| 137 | task_lock(current); |
| 138 | current->files = worker->restore_files; |
| 139 | task_unlock(current); |
| 140 | } |
| 141 | |
Jens Axboe | 771b53d0 | 2019-10-22 10:25:58 -0600 | [diff] [blame] | 142 | /* |
| 143 | * If we have an active mm, we need to drop the wq lock before unusing |
| 144 | * it. If we do, return true and let the caller retry the idle loop. |
| 145 | */ |
| 146 | if (worker->mm) { |
Jens Axboe | fcb323c | 2019-10-24 12:39:47 -0600 | [diff] [blame^] | 147 | if (!dropped_lock) { |
| 148 | __acquire(&wqe->lock); |
| 149 | spin_unlock_irq(&wqe->lock); |
| 150 | dropped_lock = true; |
| 151 | } |
Jens Axboe | 771b53d0 | 2019-10-22 10:25:58 -0600 | [diff] [blame] | 152 | __set_current_state(TASK_RUNNING); |
| 153 | set_fs(KERNEL_DS); |
| 154 | unuse_mm(worker->mm); |
| 155 | mmput(worker->mm); |
| 156 | worker->mm = NULL; |
Jens Axboe | 771b53d0 | 2019-10-22 10:25:58 -0600 | [diff] [blame] | 157 | } |
| 158 | |
Jens Axboe | fcb323c | 2019-10-24 12:39:47 -0600 | [diff] [blame^] | 159 | return dropped_lock; |
Jens Axboe | 771b53d0 | 2019-10-22 10:25:58 -0600 | [diff] [blame] | 160 | } |
| 161 | |
| 162 | static void io_worker_exit(struct io_worker *worker) |
| 163 | { |
| 164 | struct io_wqe *wqe = worker->wqe; |
| 165 | bool all_done = false; |
| 166 | |
| 167 | /* |
| 168 | * If we're not at zero, someone else is holding a brief reference |
| 169 | * to the worker. Wait for that to go away. |
| 170 | */ |
| 171 | set_current_state(TASK_INTERRUPTIBLE); |
| 172 | if (!refcount_dec_and_test(&worker->ref)) |
| 173 | schedule(); |
| 174 | __set_current_state(TASK_RUNNING); |
| 175 | |
| 176 | preempt_disable(); |
| 177 | current->flags &= ~PF_IO_WORKER; |
| 178 | if (worker->flags & IO_WORKER_F_RUNNING) |
| 179 | atomic_dec(&wqe->nr_running); |
| 180 | worker->flags = 0; |
| 181 | preempt_enable(); |
| 182 | |
| 183 | spin_lock_irq(&wqe->lock); |
| 184 | hlist_nulls_del_rcu(&worker->nulls_node); |
| 185 | if (__io_worker_unuse(wqe, worker)) { |
| 186 | __release(&wqe->lock); |
| 187 | spin_lock_irq(&wqe->lock); |
| 188 | } |
| 189 | wqe->nr_workers--; |
| 190 | all_done = !wqe->nr_workers; |
| 191 | spin_unlock_irq(&wqe->lock); |
| 192 | |
| 193 | /* all workers gone, wq exit can proceed */ |
| 194 | if (all_done && refcount_dec_and_test(&wqe->wq->refs)) |
| 195 | complete(&wqe->wq->done); |
| 196 | |
| 197 | call_rcu(&worker->rcu, io_wq_free_worker); |
| 198 | } |
| 199 | |
| 200 | static void io_worker_start(struct io_wqe *wqe, struct io_worker *worker) |
| 201 | { |
| 202 | allow_kernel_signal(SIGINT); |
| 203 | |
| 204 | current->flags |= PF_IO_WORKER; |
| 205 | |
| 206 | worker->flags |= (IO_WORKER_F_UP | IO_WORKER_F_RUNNING); |
Jens Axboe | fcb323c | 2019-10-24 12:39:47 -0600 | [diff] [blame^] | 207 | worker->restore_files = current->files; |
Jens Axboe | 771b53d0 | 2019-10-22 10:25:58 -0600 | [diff] [blame] | 208 | atomic_inc(&wqe->nr_running); |
| 209 | } |
| 210 | |
| 211 | /* |
| 212 | * Worker will start processing some work. Move it to the busy list, if |
| 213 | * it's currently on the freelist |
| 214 | */ |
| 215 | static void __io_worker_busy(struct io_wqe *wqe, struct io_worker *worker, |
| 216 | struct io_wq_work *work) |
| 217 | __must_hold(wqe->lock) |
| 218 | { |
| 219 | if (worker->flags & IO_WORKER_F_FREE) { |
| 220 | worker->flags &= ~IO_WORKER_F_FREE; |
| 221 | hlist_nulls_del_init_rcu(&worker->nulls_node); |
| 222 | hlist_nulls_add_head_rcu(&worker->nulls_node, |
| 223 | &wqe->busy_list.head); |
| 224 | } |
| 225 | worker->cur_work = work; |
| 226 | } |
| 227 | |
| 228 | /* |
| 229 | * No work, worker going to sleep. Move to freelist, and unuse mm if we |
| 230 | * have one attached. Dropping the mm may potentially sleep, so we drop |
| 231 | * the lock in that case and return success. Since the caller has to |
| 232 | * retry the loop in that case (we changed task state), we don't regrab |
| 233 | * the lock if we return success. |
| 234 | */ |
| 235 | static bool __io_worker_idle(struct io_wqe *wqe, struct io_worker *worker) |
| 236 | __must_hold(wqe->lock) |
| 237 | { |
| 238 | if (!(worker->flags & IO_WORKER_F_FREE)) { |
| 239 | worker->flags |= IO_WORKER_F_FREE; |
| 240 | hlist_nulls_del_init_rcu(&worker->nulls_node); |
| 241 | hlist_nulls_add_head_rcu(&worker->nulls_node, |
| 242 | &wqe->free_list.head); |
| 243 | } |
| 244 | |
| 245 | return __io_worker_unuse(wqe, worker); |
| 246 | } |
| 247 | |
| 248 | static struct io_wq_work *io_get_next_work(struct io_wqe *wqe, unsigned *hash) |
| 249 | __must_hold(wqe->lock) |
| 250 | { |
| 251 | struct io_wq_work *work; |
| 252 | |
| 253 | list_for_each_entry(work, &wqe->work_list, list) { |
| 254 | /* not hashed, can run anytime */ |
| 255 | if (!(work->flags & IO_WQ_WORK_HASHED)) { |
| 256 | list_del(&work->list); |
| 257 | return work; |
| 258 | } |
| 259 | |
| 260 | /* hashed, can run if not already running */ |
| 261 | *hash = work->flags >> IO_WQ_HASH_SHIFT; |
| 262 | if (!(wqe->hash_map & BIT_ULL(*hash))) { |
| 263 | wqe->hash_map |= BIT_ULL(*hash); |
| 264 | list_del(&work->list); |
| 265 | return work; |
| 266 | } |
| 267 | } |
| 268 | |
| 269 | return NULL; |
| 270 | } |
| 271 | |
| 272 | static void io_worker_handle_work(struct io_worker *worker) |
| 273 | __releases(wqe->lock) |
| 274 | { |
| 275 | struct io_wq_work *work, *old_work; |
| 276 | struct io_wqe *wqe = worker->wqe; |
| 277 | struct io_wq *wq = wqe->wq; |
| 278 | |
| 279 | do { |
| 280 | unsigned hash = -1U; |
| 281 | |
| 282 | /* |
| 283 | * Signals are either sent to cancel specific work, or to just |
| 284 | * cancel all work items. For the former, ->cur_work must |
| 285 | * match. ->cur_work is NULL at this point, since we haven't |
| 286 | * assigned any work, so it's safe to flush signals for that |
| 287 | * case. For the latter case of cancelling all work, the caller |
| 288 | * wil have set IO_WQ_BIT_CANCEL. |
| 289 | */ |
| 290 | if (signal_pending(current)) |
| 291 | flush_signals(current); |
| 292 | |
| 293 | /* |
| 294 | * If we got some work, mark us as busy. If we didn't, but |
| 295 | * the list isn't empty, it means we stalled on hashed work. |
| 296 | * Mark us stalled so we don't keep looking for work when we |
| 297 | * can't make progress, any work completion or insertion will |
| 298 | * clear the stalled flag. |
| 299 | */ |
| 300 | work = io_get_next_work(wqe, &hash); |
| 301 | if (work) |
| 302 | __io_worker_busy(wqe, worker, work); |
| 303 | else if (!list_empty(&wqe->work_list)) |
| 304 | wqe->flags |= IO_WQE_FLAG_STALLED; |
| 305 | |
| 306 | spin_unlock_irq(&wqe->lock); |
| 307 | if (!work) |
| 308 | break; |
| 309 | next: |
Jens Axboe | fcb323c | 2019-10-24 12:39:47 -0600 | [diff] [blame^] | 310 | if ((work->flags & IO_WQ_WORK_NEEDS_FILES) && |
| 311 | current->files != work->files) { |
| 312 | task_lock(current); |
| 313 | current->files = work->files; |
| 314 | task_unlock(current); |
| 315 | } |
Jens Axboe | 771b53d0 | 2019-10-22 10:25:58 -0600 | [diff] [blame] | 316 | if ((work->flags & IO_WQ_WORK_NEEDS_USER) && !worker->mm && |
| 317 | wq->mm && mmget_not_zero(wq->mm)) { |
| 318 | use_mm(wq->mm); |
| 319 | set_fs(USER_DS); |
| 320 | worker->mm = wq->mm; |
| 321 | } |
| 322 | if (test_bit(IO_WQ_BIT_CANCEL, &wq->state)) |
| 323 | work->flags |= IO_WQ_WORK_CANCEL; |
| 324 | if (worker->mm) |
| 325 | work->flags |= IO_WQ_WORK_HAS_MM; |
| 326 | |
| 327 | old_work = work; |
| 328 | work->func(&work); |
| 329 | |
| 330 | spin_lock_irq(&wqe->lock); |
| 331 | worker->cur_work = NULL; |
| 332 | if (hash != -1U) { |
| 333 | wqe->hash_map &= ~BIT_ULL(hash); |
| 334 | wqe->flags &= ~IO_WQE_FLAG_STALLED; |
| 335 | } |
| 336 | if (work && work != old_work) { |
| 337 | spin_unlock_irq(&wqe->lock); |
| 338 | /* dependent work not hashed */ |
| 339 | hash = -1U; |
| 340 | goto next; |
| 341 | } |
| 342 | } while (1); |
| 343 | } |
| 344 | |
| 345 | static inline bool io_wqe_run_queue(struct io_wqe *wqe) |
| 346 | __must_hold(wqe->lock) |
| 347 | { |
| 348 | if (!list_empty_careful(&wqe->work_list) && |
| 349 | !(wqe->flags & IO_WQE_FLAG_STALLED)) |
| 350 | return true; |
| 351 | return false; |
| 352 | } |
| 353 | |
| 354 | static int io_wqe_worker(void *data) |
| 355 | { |
| 356 | struct io_worker *worker = data; |
| 357 | struct io_wqe *wqe = worker->wqe; |
| 358 | struct io_wq *wq = wqe->wq; |
| 359 | DEFINE_WAIT(wait); |
| 360 | |
| 361 | io_worker_start(wqe, worker); |
| 362 | |
| 363 | while (!test_bit(IO_WQ_BIT_EXIT, &wq->state)) { |
| 364 | prepare_to_wait(&worker->wait, &wait, TASK_INTERRUPTIBLE); |
| 365 | |
| 366 | spin_lock_irq(&wqe->lock); |
| 367 | if (io_wqe_run_queue(wqe)) { |
| 368 | __set_current_state(TASK_RUNNING); |
| 369 | io_worker_handle_work(worker); |
| 370 | continue; |
| 371 | } |
| 372 | /* drops the lock on success, retry */ |
| 373 | if (__io_worker_idle(wqe, worker)) { |
| 374 | __release(&wqe->lock); |
| 375 | continue; |
| 376 | } |
| 377 | spin_unlock_irq(&wqe->lock); |
| 378 | if (signal_pending(current)) |
| 379 | flush_signals(current); |
| 380 | if (schedule_timeout(WORKER_IDLE_TIMEOUT)) |
| 381 | continue; |
| 382 | /* timed out, exit unless we're the fixed worker */ |
| 383 | if (test_bit(IO_WQ_BIT_EXIT, &wq->state) || |
| 384 | !(worker->flags & IO_WORKER_F_FIXED)) |
| 385 | break; |
| 386 | } |
| 387 | |
| 388 | finish_wait(&worker->wait, &wait); |
| 389 | |
| 390 | if (test_bit(IO_WQ_BIT_EXIT, &wq->state)) { |
| 391 | spin_lock_irq(&wqe->lock); |
| 392 | if (!list_empty(&wqe->work_list)) |
| 393 | io_worker_handle_work(worker); |
| 394 | else |
| 395 | spin_unlock_irq(&wqe->lock); |
| 396 | } |
| 397 | |
| 398 | io_worker_exit(worker); |
| 399 | return 0; |
| 400 | } |
| 401 | |
| 402 | /* |
| 403 | * Check head of free list for an available worker. If one isn't available, |
| 404 | * caller must wake up the wq manager to create one. |
| 405 | */ |
| 406 | static bool io_wqe_activate_free_worker(struct io_wqe *wqe) |
| 407 | __must_hold(RCU) |
| 408 | { |
| 409 | struct hlist_nulls_node *n; |
| 410 | struct io_worker *worker; |
| 411 | |
| 412 | n = rcu_dereference(hlist_nulls_first_rcu(&wqe->free_list.head)); |
| 413 | if (is_a_nulls(n)) |
| 414 | return false; |
| 415 | |
| 416 | worker = hlist_nulls_entry(n, struct io_worker, nulls_node); |
| 417 | if (io_worker_get(worker)) { |
| 418 | wake_up(&worker->wait); |
| 419 | io_worker_release(worker); |
| 420 | return true; |
| 421 | } |
| 422 | |
| 423 | return false; |
| 424 | } |
| 425 | |
| 426 | /* |
| 427 | * We need a worker. If we find a free one, we're good. If not, and we're |
| 428 | * below the max number of workers, wake up the manager to create one. |
| 429 | */ |
| 430 | static void io_wqe_wake_worker(struct io_wqe *wqe) |
| 431 | { |
| 432 | bool ret; |
| 433 | |
| 434 | rcu_read_lock(); |
| 435 | ret = io_wqe_activate_free_worker(wqe); |
| 436 | rcu_read_unlock(); |
| 437 | |
| 438 | if (!ret && wqe->nr_workers < wqe->max_workers) |
| 439 | wake_up_process(wqe->wq->manager); |
| 440 | } |
| 441 | |
| 442 | /* |
| 443 | * Called when a worker is scheduled in. Mark us as currently running. |
| 444 | */ |
| 445 | void io_wq_worker_running(struct task_struct *tsk) |
| 446 | { |
| 447 | struct io_worker *worker = kthread_data(tsk); |
| 448 | struct io_wqe *wqe = worker->wqe; |
| 449 | |
| 450 | if (!(worker->flags & IO_WORKER_F_UP)) |
| 451 | return; |
| 452 | if (worker->flags & IO_WORKER_F_RUNNING) |
| 453 | return; |
| 454 | worker->flags |= IO_WORKER_F_RUNNING; |
| 455 | atomic_inc(&wqe->nr_running); |
| 456 | } |
| 457 | |
| 458 | /* |
| 459 | * Called when worker is going to sleep. If there are no workers currently |
| 460 | * running and we have work pending, wake up a free one or have the manager |
| 461 | * set one up. |
| 462 | */ |
| 463 | void io_wq_worker_sleeping(struct task_struct *tsk) |
| 464 | { |
| 465 | struct io_worker *worker = kthread_data(tsk); |
| 466 | struct io_wqe *wqe = worker->wqe; |
| 467 | |
| 468 | if (!(worker->flags & IO_WORKER_F_UP)) |
| 469 | return; |
| 470 | if (!(worker->flags & IO_WORKER_F_RUNNING)) |
| 471 | return; |
| 472 | |
| 473 | worker->flags &= ~IO_WORKER_F_RUNNING; |
| 474 | |
| 475 | spin_lock_irq(&wqe->lock); |
| 476 | if (atomic_dec_and_test(&wqe->nr_running) && io_wqe_run_queue(wqe)) |
| 477 | io_wqe_wake_worker(wqe); |
| 478 | spin_unlock_irq(&wqe->lock); |
| 479 | } |
| 480 | |
| 481 | static void create_io_worker(struct io_wq *wq, struct io_wqe *wqe) |
| 482 | { |
| 483 | struct io_worker *worker; |
| 484 | |
| 485 | worker = kcalloc_node(1, sizeof(*worker), GFP_KERNEL, wqe->node); |
| 486 | if (!worker) |
| 487 | return; |
| 488 | |
| 489 | refcount_set(&worker->ref, 1); |
| 490 | worker->nulls_node.pprev = NULL; |
| 491 | init_waitqueue_head(&worker->wait); |
| 492 | worker->wqe = wqe; |
| 493 | |
| 494 | worker->task = kthread_create_on_node(io_wqe_worker, worker, wqe->node, |
| 495 | "io_wqe_worker-%d", wqe->node); |
| 496 | if (IS_ERR(worker->task)) { |
| 497 | kfree(worker); |
| 498 | return; |
| 499 | } |
| 500 | |
| 501 | spin_lock_irq(&wqe->lock); |
| 502 | hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list.head); |
| 503 | worker->flags |= IO_WORKER_F_FREE; |
| 504 | if (!wqe->nr_workers) |
| 505 | worker->flags |= IO_WORKER_F_FIXED; |
| 506 | wqe->nr_workers++; |
| 507 | spin_unlock_irq(&wqe->lock); |
| 508 | |
| 509 | wake_up_process(worker->task); |
| 510 | } |
| 511 | |
| 512 | static inline bool io_wqe_need_new_worker(struct io_wqe *wqe) |
| 513 | __must_hold(wqe->lock) |
| 514 | { |
| 515 | if (!wqe->nr_workers) |
| 516 | return true; |
| 517 | if (hlist_nulls_empty(&wqe->free_list.head) && |
| 518 | wqe->nr_workers < wqe->max_workers && io_wqe_run_queue(wqe)) |
| 519 | return true; |
| 520 | |
| 521 | return false; |
| 522 | } |
| 523 | |
| 524 | /* |
| 525 | * Manager thread. Tasked with creating new workers, if we need them. |
| 526 | */ |
| 527 | static int io_wq_manager(void *data) |
| 528 | { |
| 529 | struct io_wq *wq = data; |
| 530 | |
| 531 | while (!kthread_should_stop()) { |
| 532 | int i; |
| 533 | |
| 534 | for (i = 0; i < wq->nr_wqes; i++) { |
| 535 | struct io_wqe *wqe = wq->wqes[i]; |
| 536 | bool fork_worker = false; |
| 537 | |
| 538 | spin_lock_irq(&wqe->lock); |
| 539 | fork_worker = io_wqe_need_new_worker(wqe); |
| 540 | spin_unlock_irq(&wqe->lock); |
| 541 | if (fork_worker) |
| 542 | create_io_worker(wq, wqe); |
| 543 | } |
| 544 | set_current_state(TASK_INTERRUPTIBLE); |
| 545 | schedule_timeout(HZ); |
| 546 | } |
| 547 | |
| 548 | return 0; |
| 549 | } |
| 550 | |
| 551 | static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work) |
| 552 | { |
| 553 | unsigned long flags; |
| 554 | |
| 555 | spin_lock_irqsave(&wqe->lock, flags); |
| 556 | list_add_tail(&work->list, &wqe->work_list); |
| 557 | wqe->flags &= ~IO_WQE_FLAG_STALLED; |
| 558 | spin_unlock_irqrestore(&wqe->lock, flags); |
| 559 | |
| 560 | if (!atomic_read(&wqe->nr_running)) |
| 561 | io_wqe_wake_worker(wqe); |
| 562 | } |
| 563 | |
| 564 | void io_wq_enqueue(struct io_wq *wq, struct io_wq_work *work) |
| 565 | { |
| 566 | struct io_wqe *wqe = wq->wqes[numa_node_id()]; |
| 567 | |
| 568 | io_wqe_enqueue(wqe, work); |
| 569 | } |
| 570 | |
| 571 | /* |
| 572 | * Enqueue work, hashed by some key. Work items that hash to the same value |
| 573 | * will not be done in parallel. Used to limit concurrent writes, generally |
| 574 | * hashed by inode. |
| 575 | */ |
| 576 | void io_wq_enqueue_hashed(struct io_wq *wq, struct io_wq_work *work, void *val) |
| 577 | { |
| 578 | struct io_wqe *wqe = wq->wqes[numa_node_id()]; |
| 579 | unsigned bit; |
| 580 | |
| 581 | |
| 582 | bit = hash_ptr(val, IO_WQ_HASH_ORDER); |
| 583 | work->flags |= (IO_WQ_WORK_HASHED | (bit << IO_WQ_HASH_SHIFT)); |
| 584 | io_wqe_enqueue(wqe, work); |
| 585 | } |
| 586 | |
| 587 | static bool io_wqe_worker_send_sig(struct io_worker *worker, void *data) |
| 588 | { |
| 589 | send_sig(SIGINT, worker->task, 1); |
| 590 | return false; |
| 591 | } |
| 592 | |
| 593 | /* |
| 594 | * Iterate the passed in list and call the specific function for each |
| 595 | * worker that isn't exiting |
| 596 | */ |
| 597 | static bool io_wq_for_each_worker(struct io_wqe *wqe, |
| 598 | struct io_wq_nulls_list *list, |
| 599 | bool (*func)(struct io_worker *, void *), |
| 600 | void *data) |
| 601 | { |
| 602 | struct hlist_nulls_node *n; |
| 603 | struct io_worker *worker; |
| 604 | bool ret = false; |
| 605 | |
| 606 | restart: |
| 607 | hlist_nulls_for_each_entry_rcu(worker, n, &list->head, nulls_node) { |
| 608 | if (io_worker_get(worker)) { |
| 609 | ret = func(worker, data); |
| 610 | io_worker_release(worker); |
| 611 | if (ret) |
| 612 | break; |
| 613 | } |
| 614 | } |
| 615 | if (!ret && get_nulls_value(n) != list->nulls) |
| 616 | goto restart; |
| 617 | return ret; |
| 618 | } |
| 619 | |
| 620 | void io_wq_cancel_all(struct io_wq *wq) |
| 621 | { |
| 622 | int i; |
| 623 | |
| 624 | set_bit(IO_WQ_BIT_CANCEL, &wq->state); |
| 625 | |
| 626 | /* |
| 627 | * Browse both lists, as there's a gap between handing work off |
| 628 | * to a worker and the worker putting itself on the busy_list |
| 629 | */ |
| 630 | rcu_read_lock(); |
| 631 | for (i = 0; i < wq->nr_wqes; i++) { |
| 632 | struct io_wqe *wqe = wq->wqes[i]; |
| 633 | |
| 634 | io_wq_for_each_worker(wqe, &wqe->busy_list, |
| 635 | io_wqe_worker_send_sig, NULL); |
| 636 | io_wq_for_each_worker(wqe, &wqe->free_list, |
| 637 | io_wqe_worker_send_sig, NULL); |
| 638 | } |
| 639 | rcu_read_unlock(); |
| 640 | } |
| 641 | |
| 642 | static bool io_wq_worker_cancel(struct io_worker *worker, void *data) |
| 643 | { |
| 644 | struct io_wq_work *work = data; |
| 645 | |
| 646 | if (worker->cur_work == work) { |
| 647 | send_sig(SIGINT, worker->task, 1); |
| 648 | return true; |
| 649 | } |
| 650 | |
| 651 | return false; |
| 652 | } |
| 653 | |
| 654 | static enum io_wq_cancel io_wqe_cancel_work(struct io_wqe *wqe, |
| 655 | struct io_wq_work *cwork) |
| 656 | { |
| 657 | struct io_wq_work *work; |
| 658 | bool found = false; |
| 659 | |
| 660 | cwork->flags |= IO_WQ_WORK_CANCEL; |
| 661 | |
| 662 | /* |
| 663 | * First check pending list, if we're lucky we can just remove it |
| 664 | * from there. CANCEL_OK means that the work is returned as-new, |
| 665 | * no completion will be posted for it. |
| 666 | */ |
| 667 | spin_lock_irq(&wqe->lock); |
| 668 | list_for_each_entry(work, &wqe->work_list, list) { |
| 669 | if (work == cwork) { |
| 670 | list_del(&work->list); |
| 671 | found = true; |
| 672 | break; |
| 673 | } |
| 674 | } |
| 675 | spin_unlock_irq(&wqe->lock); |
| 676 | |
| 677 | if (found) { |
| 678 | work->flags |= IO_WQ_WORK_CANCEL; |
| 679 | work->func(&work); |
| 680 | return IO_WQ_CANCEL_OK; |
| 681 | } |
| 682 | |
| 683 | /* |
| 684 | * Now check if a free (going busy) or busy worker has the work |
| 685 | * currently running. If we find it there, we'll return CANCEL_RUNNING |
| 686 | * as an indication that we attempte to signal cancellation. The |
| 687 | * completion will run normally in this case. |
| 688 | */ |
| 689 | rcu_read_lock(); |
| 690 | found = io_wq_for_each_worker(wqe, &wqe->free_list, io_wq_worker_cancel, |
| 691 | cwork); |
| 692 | if (found) |
| 693 | goto done; |
| 694 | |
| 695 | found = io_wq_for_each_worker(wqe, &wqe->busy_list, io_wq_worker_cancel, |
| 696 | cwork); |
| 697 | done: |
| 698 | rcu_read_unlock(); |
| 699 | return found ? IO_WQ_CANCEL_RUNNING : IO_WQ_CANCEL_NOTFOUND; |
| 700 | } |
| 701 | |
| 702 | enum io_wq_cancel io_wq_cancel_work(struct io_wq *wq, struct io_wq_work *cwork) |
| 703 | { |
| 704 | enum io_wq_cancel ret = IO_WQ_CANCEL_NOTFOUND; |
| 705 | int i; |
| 706 | |
| 707 | for (i = 0; i < wq->nr_wqes; i++) { |
| 708 | struct io_wqe *wqe = wq->wqes[i]; |
| 709 | |
| 710 | ret = io_wqe_cancel_work(wqe, cwork); |
| 711 | if (ret != IO_WQ_CANCEL_NOTFOUND) |
| 712 | break; |
| 713 | } |
| 714 | |
| 715 | return ret; |
| 716 | } |
| 717 | |
| 718 | struct io_wq_flush_data { |
| 719 | struct io_wq_work work; |
| 720 | struct completion done; |
| 721 | }; |
| 722 | |
| 723 | static void io_wq_flush_func(struct io_wq_work **workptr) |
| 724 | { |
| 725 | struct io_wq_work *work = *workptr; |
| 726 | struct io_wq_flush_data *data; |
| 727 | |
| 728 | data = container_of(work, struct io_wq_flush_data, work); |
| 729 | complete(&data->done); |
| 730 | } |
| 731 | |
| 732 | /* |
| 733 | * Doesn't wait for previously queued work to finish. When this completes, |
| 734 | * it just means that previously queued work was started. |
| 735 | */ |
| 736 | void io_wq_flush(struct io_wq *wq) |
| 737 | { |
| 738 | struct io_wq_flush_data data; |
| 739 | int i; |
| 740 | |
| 741 | for (i = 0; i < wq->nr_wqes; i++) { |
| 742 | struct io_wqe *wqe = wq->wqes[i]; |
| 743 | |
| 744 | init_completion(&data.done); |
| 745 | INIT_IO_WORK(&data.work, io_wq_flush_func); |
| 746 | io_wqe_enqueue(wqe, &data.work); |
| 747 | wait_for_completion(&data.done); |
| 748 | } |
| 749 | } |
| 750 | |
| 751 | struct io_wq *io_wq_create(unsigned concurrency, struct mm_struct *mm) |
| 752 | { |
| 753 | int ret = -ENOMEM, i, node; |
| 754 | struct io_wq *wq; |
| 755 | |
| 756 | wq = kcalloc(1, sizeof(*wq), GFP_KERNEL); |
| 757 | if (!wq) |
| 758 | return ERR_PTR(-ENOMEM); |
| 759 | |
| 760 | wq->nr_wqes = num_online_nodes(); |
| 761 | wq->wqes = kcalloc(wq->nr_wqes, sizeof(struct io_wqe *), GFP_KERNEL); |
| 762 | if (!wq->wqes) { |
| 763 | kfree(wq); |
| 764 | return ERR_PTR(-ENOMEM); |
| 765 | } |
| 766 | |
| 767 | i = 0; |
| 768 | refcount_set(&wq->refs, wq->nr_wqes); |
| 769 | for_each_online_node(node) { |
| 770 | struct io_wqe *wqe; |
| 771 | |
| 772 | wqe = kcalloc_node(1, sizeof(struct io_wqe), GFP_KERNEL, node); |
| 773 | if (!wqe) |
| 774 | break; |
| 775 | wq->wqes[i] = wqe; |
| 776 | wqe->node = node; |
| 777 | wqe->max_workers = concurrency; |
| 778 | wqe->node = node; |
| 779 | wqe->wq = wq; |
| 780 | spin_lock_init(&wqe->lock); |
| 781 | INIT_LIST_HEAD(&wqe->work_list); |
| 782 | INIT_HLIST_NULLS_HEAD(&wqe->free_list.head, 0); |
| 783 | wqe->free_list.nulls = 0; |
| 784 | INIT_HLIST_NULLS_HEAD(&wqe->busy_list.head, 1); |
| 785 | wqe->busy_list.nulls = 1; |
| 786 | atomic_set(&wqe->nr_running, 0); |
| 787 | |
| 788 | i++; |
| 789 | } |
| 790 | |
| 791 | init_completion(&wq->done); |
| 792 | |
| 793 | if (i != wq->nr_wqes) |
| 794 | goto err; |
| 795 | |
| 796 | /* caller must have already done mmgrab() on this mm */ |
| 797 | wq->mm = mm; |
| 798 | |
| 799 | wq->manager = kthread_create(io_wq_manager, wq, "io_wq_manager"); |
| 800 | if (!IS_ERR(wq->manager)) { |
| 801 | wake_up_process(wq->manager); |
| 802 | return wq; |
| 803 | } |
| 804 | |
| 805 | ret = PTR_ERR(wq->manager); |
| 806 | wq->manager = NULL; |
| 807 | err: |
| 808 | complete(&wq->done); |
| 809 | io_wq_destroy(wq); |
| 810 | return ERR_PTR(ret); |
| 811 | } |
| 812 | |
| 813 | static bool io_wq_worker_wake(struct io_worker *worker, void *data) |
| 814 | { |
| 815 | wake_up_process(worker->task); |
| 816 | return false; |
| 817 | } |
| 818 | |
| 819 | void io_wq_destroy(struct io_wq *wq) |
| 820 | { |
| 821 | int i; |
| 822 | |
| 823 | if (wq->manager) { |
| 824 | set_bit(IO_WQ_BIT_EXIT, &wq->state); |
| 825 | kthread_stop(wq->manager); |
| 826 | } |
| 827 | |
| 828 | rcu_read_lock(); |
| 829 | for (i = 0; i < wq->nr_wqes; i++) { |
| 830 | struct io_wqe *wqe = wq->wqes[i]; |
| 831 | |
| 832 | if (!wqe) |
| 833 | continue; |
| 834 | io_wq_for_each_worker(wqe, &wqe->free_list, io_wq_worker_wake, |
| 835 | NULL); |
| 836 | io_wq_for_each_worker(wqe, &wqe->busy_list, io_wq_worker_wake, |
| 837 | NULL); |
| 838 | } |
| 839 | rcu_read_unlock(); |
| 840 | |
| 841 | wait_for_completion(&wq->done); |
| 842 | |
| 843 | for (i = 0; i < wq->nr_wqes; i++) |
| 844 | kfree(wq->wqes[i]); |
| 845 | kfree(wq->wqes); |
| 846 | kfree(wq); |
| 847 | } |