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