| 1 | // SPDX-License-Identifier: GPL-2.0 |
| 2 | /* |
| 3 | * Block multiqueue core code |
| 4 | * |
| 5 | * Copyright (C) 2013-2014 Jens Axboe |
| 6 | * Copyright (C) 2013-2014 Christoph Hellwig |
| 7 | */ |
| 8 | #include <linux/kernel.h> |
| 9 | #include <linux/module.h> |
| 10 | #include <linux/backing-dev.h> |
| 11 | #include <linux/bio.h> |
| 12 | #include <linux/blkdev.h> |
| 13 | #include <linux/blk-integrity.h> |
| 14 | #include <linux/kmemleak.h> |
| 15 | #include <linux/mm.h> |
| 16 | #include <linux/init.h> |
| 17 | #include <linux/slab.h> |
| 18 | #include <linux/workqueue.h> |
| 19 | #include <linux/smp.h> |
| 20 | #include <linux/interrupt.h> |
| 21 | #include <linux/llist.h> |
| 22 | #include <linux/cpu.h> |
| 23 | #include <linux/cache.h> |
| 24 | #include <linux/sched/sysctl.h> |
| 25 | #include <linux/sched/topology.h> |
| 26 | #include <linux/sched/signal.h> |
| 27 | #include <linux/delay.h> |
| 28 | #include <linux/crash_dump.h> |
| 29 | #include <linux/prefetch.h> |
| 30 | #include <linux/blk-crypto.h> |
| 31 | #include <linux/part_stat.h> |
| 32 | |
| 33 | #include <trace/events/block.h> |
| 34 | |
| 35 | #include <linux/t10-pi.h> |
| 36 | #include "blk.h" |
| 37 | #include "blk-mq.h" |
| 38 | #include "blk-mq-debugfs.h" |
| 39 | #include "blk-pm.h" |
| 40 | #include "blk-stat.h" |
| 41 | #include "blk-mq-sched.h" |
| 42 | #include "blk-rq-qos.h" |
| 43 | #include "blk-ioprio.h" |
| 44 | |
| 45 | static DEFINE_PER_CPU(struct llist_head, blk_cpu_done); |
| 46 | |
| 47 | static void blk_mq_insert_request(struct request *rq, blk_insert_t flags); |
| 48 | static void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx, |
| 49 | struct list_head *list); |
| 50 | |
| 51 | static inline struct blk_mq_hw_ctx *blk_qc_to_hctx(struct request_queue *q, |
| 52 | blk_qc_t qc) |
| 53 | { |
| 54 | return xa_load(&q->hctx_table, qc); |
| 55 | } |
| 56 | |
| 57 | static inline blk_qc_t blk_rq_to_qc(struct request *rq) |
| 58 | { |
| 59 | return rq->mq_hctx->queue_num; |
| 60 | } |
| 61 | |
| 62 | /* |
| 63 | * Check if any of the ctx, dispatch list or elevator |
| 64 | * have pending work in this hardware queue. |
| 65 | */ |
| 66 | static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx) |
| 67 | { |
| 68 | return !list_empty_careful(&hctx->dispatch) || |
| 69 | sbitmap_any_bit_set(&hctx->ctx_map) || |
| 70 | blk_mq_sched_has_work(hctx); |
| 71 | } |
| 72 | |
| 73 | /* |
| 74 | * Mark this ctx as having pending work in this hardware queue |
| 75 | */ |
| 76 | static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx, |
| 77 | struct blk_mq_ctx *ctx) |
| 78 | { |
| 79 | const int bit = ctx->index_hw[hctx->type]; |
| 80 | |
| 81 | if (!sbitmap_test_bit(&hctx->ctx_map, bit)) |
| 82 | sbitmap_set_bit(&hctx->ctx_map, bit); |
| 83 | } |
| 84 | |
| 85 | static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx, |
| 86 | struct blk_mq_ctx *ctx) |
| 87 | { |
| 88 | const int bit = ctx->index_hw[hctx->type]; |
| 89 | |
| 90 | sbitmap_clear_bit(&hctx->ctx_map, bit); |
| 91 | } |
| 92 | |
| 93 | struct mq_inflight { |
| 94 | struct block_device *part; |
| 95 | unsigned int inflight[2]; |
| 96 | }; |
| 97 | |
| 98 | static bool blk_mq_check_inflight(struct request *rq, void *priv) |
| 99 | { |
| 100 | struct mq_inflight *mi = priv; |
| 101 | |
| 102 | if (rq->part && blk_do_io_stat(rq) && |
| 103 | (!mi->part->bd_partno || rq->part == mi->part) && |
| 104 | blk_mq_rq_state(rq) == MQ_RQ_IN_FLIGHT) |
| 105 | mi->inflight[rq_data_dir(rq)]++; |
| 106 | |
| 107 | return true; |
| 108 | } |
| 109 | |
| 110 | unsigned int blk_mq_in_flight(struct request_queue *q, |
| 111 | struct block_device *part) |
| 112 | { |
| 113 | struct mq_inflight mi = { .part = part }; |
| 114 | |
| 115 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi); |
| 116 | |
| 117 | return mi.inflight[0] + mi.inflight[1]; |
| 118 | } |
| 119 | |
| 120 | void blk_mq_in_flight_rw(struct request_queue *q, struct block_device *part, |
| 121 | unsigned int inflight[2]) |
| 122 | { |
| 123 | struct mq_inflight mi = { .part = part }; |
| 124 | |
| 125 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi); |
| 126 | inflight[0] = mi.inflight[0]; |
| 127 | inflight[1] = mi.inflight[1]; |
| 128 | } |
| 129 | |
| 130 | void blk_freeze_queue_start(struct request_queue *q) |
| 131 | { |
| 132 | mutex_lock(&q->mq_freeze_lock); |
| 133 | if (++q->mq_freeze_depth == 1) { |
| 134 | percpu_ref_kill(&q->q_usage_counter); |
| 135 | mutex_unlock(&q->mq_freeze_lock); |
| 136 | if (queue_is_mq(q)) |
| 137 | blk_mq_run_hw_queues(q, false); |
| 138 | } else { |
| 139 | mutex_unlock(&q->mq_freeze_lock); |
| 140 | } |
| 141 | } |
| 142 | EXPORT_SYMBOL_GPL(blk_freeze_queue_start); |
| 143 | |
| 144 | void blk_mq_freeze_queue_wait(struct request_queue *q) |
| 145 | { |
| 146 | wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter)); |
| 147 | } |
| 148 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait); |
| 149 | |
| 150 | int blk_mq_freeze_queue_wait_timeout(struct request_queue *q, |
| 151 | unsigned long timeout) |
| 152 | { |
| 153 | return wait_event_timeout(q->mq_freeze_wq, |
| 154 | percpu_ref_is_zero(&q->q_usage_counter), |
| 155 | timeout); |
| 156 | } |
| 157 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout); |
| 158 | |
| 159 | /* |
| 160 | * Guarantee no request is in use, so we can change any data structure of |
| 161 | * the queue afterward. |
| 162 | */ |
| 163 | void blk_freeze_queue(struct request_queue *q) |
| 164 | { |
| 165 | /* |
| 166 | * In the !blk_mq case we are only calling this to kill the |
| 167 | * q_usage_counter, otherwise this increases the freeze depth |
| 168 | * and waits for it to return to zero. For this reason there is |
| 169 | * no blk_unfreeze_queue(), and blk_freeze_queue() is not |
| 170 | * exported to drivers as the only user for unfreeze is blk_mq. |
| 171 | */ |
| 172 | blk_freeze_queue_start(q); |
| 173 | blk_mq_freeze_queue_wait(q); |
| 174 | } |
| 175 | |
| 176 | void blk_mq_freeze_queue(struct request_queue *q) |
| 177 | { |
| 178 | /* |
| 179 | * ...just an alias to keep freeze and unfreeze actions balanced |
| 180 | * in the blk_mq_* namespace |
| 181 | */ |
| 182 | blk_freeze_queue(q); |
| 183 | } |
| 184 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue); |
| 185 | |
| 186 | void __blk_mq_unfreeze_queue(struct request_queue *q, bool force_atomic) |
| 187 | { |
| 188 | mutex_lock(&q->mq_freeze_lock); |
| 189 | if (force_atomic) |
| 190 | q->q_usage_counter.data->force_atomic = true; |
| 191 | q->mq_freeze_depth--; |
| 192 | WARN_ON_ONCE(q->mq_freeze_depth < 0); |
| 193 | if (!q->mq_freeze_depth) { |
| 194 | percpu_ref_resurrect(&q->q_usage_counter); |
| 195 | wake_up_all(&q->mq_freeze_wq); |
| 196 | } |
| 197 | mutex_unlock(&q->mq_freeze_lock); |
| 198 | } |
| 199 | |
| 200 | void blk_mq_unfreeze_queue(struct request_queue *q) |
| 201 | { |
| 202 | __blk_mq_unfreeze_queue(q, false); |
| 203 | } |
| 204 | EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue); |
| 205 | |
| 206 | /* |
| 207 | * FIXME: replace the scsi_internal_device_*block_nowait() calls in the |
| 208 | * mpt3sas driver such that this function can be removed. |
| 209 | */ |
| 210 | void blk_mq_quiesce_queue_nowait(struct request_queue *q) |
| 211 | { |
| 212 | unsigned long flags; |
| 213 | |
| 214 | spin_lock_irqsave(&q->queue_lock, flags); |
| 215 | if (!q->quiesce_depth++) |
| 216 | blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q); |
| 217 | spin_unlock_irqrestore(&q->queue_lock, flags); |
| 218 | } |
| 219 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait); |
| 220 | |
| 221 | /** |
| 222 | * blk_mq_wait_quiesce_done() - wait until in-progress quiesce is done |
| 223 | * @set: tag_set to wait on |
| 224 | * |
| 225 | * Note: it is driver's responsibility for making sure that quiesce has |
| 226 | * been started on or more of the request_queues of the tag_set. This |
| 227 | * function only waits for the quiesce on those request_queues that had |
| 228 | * the quiesce flag set using blk_mq_quiesce_queue_nowait. |
| 229 | */ |
| 230 | void blk_mq_wait_quiesce_done(struct blk_mq_tag_set *set) |
| 231 | { |
| 232 | if (set->flags & BLK_MQ_F_BLOCKING) |
| 233 | synchronize_srcu(set->srcu); |
| 234 | else |
| 235 | synchronize_rcu(); |
| 236 | } |
| 237 | EXPORT_SYMBOL_GPL(blk_mq_wait_quiesce_done); |
| 238 | |
| 239 | /** |
| 240 | * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished |
| 241 | * @q: request queue. |
| 242 | * |
| 243 | * Note: this function does not prevent that the struct request end_io() |
| 244 | * callback function is invoked. Once this function is returned, we make |
| 245 | * sure no dispatch can happen until the queue is unquiesced via |
| 246 | * blk_mq_unquiesce_queue(). |
| 247 | */ |
| 248 | void blk_mq_quiesce_queue(struct request_queue *q) |
| 249 | { |
| 250 | blk_mq_quiesce_queue_nowait(q); |
| 251 | /* nothing to wait for non-mq queues */ |
| 252 | if (queue_is_mq(q)) |
| 253 | blk_mq_wait_quiesce_done(q->tag_set); |
| 254 | } |
| 255 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue); |
| 256 | |
| 257 | /* |
| 258 | * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue() |
| 259 | * @q: request queue. |
| 260 | * |
| 261 | * This function recovers queue into the state before quiescing |
| 262 | * which is done by blk_mq_quiesce_queue. |
| 263 | */ |
| 264 | void blk_mq_unquiesce_queue(struct request_queue *q) |
| 265 | { |
| 266 | unsigned long flags; |
| 267 | bool run_queue = false; |
| 268 | |
| 269 | spin_lock_irqsave(&q->queue_lock, flags); |
| 270 | if (WARN_ON_ONCE(q->quiesce_depth <= 0)) { |
| 271 | ; |
| 272 | } else if (!--q->quiesce_depth) { |
| 273 | blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q); |
| 274 | run_queue = true; |
| 275 | } |
| 276 | spin_unlock_irqrestore(&q->queue_lock, flags); |
| 277 | |
| 278 | /* dispatch requests which are inserted during quiescing */ |
| 279 | if (run_queue) |
| 280 | blk_mq_run_hw_queues(q, true); |
| 281 | } |
| 282 | EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue); |
| 283 | |
| 284 | void blk_mq_quiesce_tagset(struct blk_mq_tag_set *set) |
| 285 | { |
| 286 | struct request_queue *q; |
| 287 | |
| 288 | mutex_lock(&set->tag_list_lock); |
| 289 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
| 290 | if (!blk_queue_skip_tagset_quiesce(q)) |
| 291 | blk_mq_quiesce_queue_nowait(q); |
| 292 | } |
| 293 | blk_mq_wait_quiesce_done(set); |
| 294 | mutex_unlock(&set->tag_list_lock); |
| 295 | } |
| 296 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_tagset); |
| 297 | |
| 298 | void blk_mq_unquiesce_tagset(struct blk_mq_tag_set *set) |
| 299 | { |
| 300 | struct request_queue *q; |
| 301 | |
| 302 | mutex_lock(&set->tag_list_lock); |
| 303 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
| 304 | if (!blk_queue_skip_tagset_quiesce(q)) |
| 305 | blk_mq_unquiesce_queue(q); |
| 306 | } |
| 307 | mutex_unlock(&set->tag_list_lock); |
| 308 | } |
| 309 | EXPORT_SYMBOL_GPL(blk_mq_unquiesce_tagset); |
| 310 | |
| 311 | void blk_mq_wake_waiters(struct request_queue *q) |
| 312 | { |
| 313 | struct blk_mq_hw_ctx *hctx; |
| 314 | unsigned long i; |
| 315 | |
| 316 | queue_for_each_hw_ctx(q, hctx, i) |
| 317 | if (blk_mq_hw_queue_mapped(hctx)) |
| 318 | blk_mq_tag_wakeup_all(hctx->tags, true); |
| 319 | } |
| 320 | |
| 321 | void blk_rq_init(struct request_queue *q, struct request *rq) |
| 322 | { |
| 323 | memset(rq, 0, sizeof(*rq)); |
| 324 | |
| 325 | INIT_LIST_HEAD(&rq->queuelist); |
| 326 | rq->q = q; |
| 327 | rq->__sector = (sector_t) -1; |
| 328 | INIT_HLIST_NODE(&rq->hash); |
| 329 | RB_CLEAR_NODE(&rq->rb_node); |
| 330 | rq->tag = BLK_MQ_NO_TAG; |
| 331 | rq->internal_tag = BLK_MQ_NO_TAG; |
| 332 | rq->start_time_ns = ktime_get_ns(); |
| 333 | rq->part = NULL; |
| 334 | blk_crypto_rq_set_defaults(rq); |
| 335 | } |
| 336 | EXPORT_SYMBOL(blk_rq_init); |
| 337 | |
| 338 | static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data, |
| 339 | struct blk_mq_tags *tags, unsigned int tag, u64 alloc_time_ns) |
| 340 | { |
| 341 | struct blk_mq_ctx *ctx = data->ctx; |
| 342 | struct blk_mq_hw_ctx *hctx = data->hctx; |
| 343 | struct request_queue *q = data->q; |
| 344 | struct request *rq = tags->static_rqs[tag]; |
| 345 | |
| 346 | rq->q = q; |
| 347 | rq->mq_ctx = ctx; |
| 348 | rq->mq_hctx = hctx; |
| 349 | rq->cmd_flags = data->cmd_flags; |
| 350 | |
| 351 | if (data->flags & BLK_MQ_REQ_PM) |
| 352 | data->rq_flags |= RQF_PM; |
| 353 | if (blk_queue_io_stat(q)) |
| 354 | data->rq_flags |= RQF_IO_STAT; |
| 355 | rq->rq_flags = data->rq_flags; |
| 356 | |
| 357 | if (!(data->rq_flags & RQF_ELV)) { |
| 358 | rq->tag = tag; |
| 359 | rq->internal_tag = BLK_MQ_NO_TAG; |
| 360 | } else { |
| 361 | rq->tag = BLK_MQ_NO_TAG; |
| 362 | rq->internal_tag = tag; |
| 363 | } |
| 364 | rq->timeout = 0; |
| 365 | |
| 366 | if (blk_mq_need_time_stamp(rq)) |
| 367 | rq->start_time_ns = ktime_get_ns(); |
| 368 | else |
| 369 | rq->start_time_ns = 0; |
| 370 | rq->part = NULL; |
| 371 | #ifdef CONFIG_BLK_RQ_ALLOC_TIME |
| 372 | rq->alloc_time_ns = alloc_time_ns; |
| 373 | #endif |
| 374 | rq->io_start_time_ns = 0; |
| 375 | rq->stats_sectors = 0; |
| 376 | rq->nr_phys_segments = 0; |
| 377 | #if defined(CONFIG_BLK_DEV_INTEGRITY) |
| 378 | rq->nr_integrity_segments = 0; |
| 379 | #endif |
| 380 | rq->end_io = NULL; |
| 381 | rq->end_io_data = NULL; |
| 382 | |
| 383 | blk_crypto_rq_set_defaults(rq); |
| 384 | INIT_LIST_HEAD(&rq->queuelist); |
| 385 | /* tag was already set */ |
| 386 | WRITE_ONCE(rq->deadline, 0); |
| 387 | req_ref_set(rq, 1); |
| 388 | |
| 389 | if (rq->rq_flags & RQF_ELV) { |
| 390 | struct elevator_queue *e = data->q->elevator; |
| 391 | |
| 392 | INIT_HLIST_NODE(&rq->hash); |
| 393 | RB_CLEAR_NODE(&rq->rb_node); |
| 394 | |
| 395 | if (!op_is_flush(data->cmd_flags) && |
| 396 | e->type->ops.prepare_request) { |
| 397 | e->type->ops.prepare_request(rq); |
| 398 | rq->rq_flags |= RQF_ELVPRIV; |
| 399 | } |
| 400 | } |
| 401 | |
| 402 | return rq; |
| 403 | } |
| 404 | |
| 405 | static inline struct request * |
| 406 | __blk_mq_alloc_requests_batch(struct blk_mq_alloc_data *data, |
| 407 | u64 alloc_time_ns) |
| 408 | { |
| 409 | unsigned int tag, tag_offset; |
| 410 | struct blk_mq_tags *tags; |
| 411 | struct request *rq; |
| 412 | unsigned long tag_mask; |
| 413 | int i, nr = 0; |
| 414 | |
| 415 | tag_mask = blk_mq_get_tags(data, data->nr_tags, &tag_offset); |
| 416 | if (unlikely(!tag_mask)) |
| 417 | return NULL; |
| 418 | |
| 419 | tags = blk_mq_tags_from_data(data); |
| 420 | for (i = 0; tag_mask; i++) { |
| 421 | if (!(tag_mask & (1UL << i))) |
| 422 | continue; |
| 423 | tag = tag_offset + i; |
| 424 | prefetch(tags->static_rqs[tag]); |
| 425 | tag_mask &= ~(1UL << i); |
| 426 | rq = blk_mq_rq_ctx_init(data, tags, tag, alloc_time_ns); |
| 427 | rq_list_add(data->cached_rq, rq); |
| 428 | nr++; |
| 429 | } |
| 430 | /* caller already holds a reference, add for remainder */ |
| 431 | percpu_ref_get_many(&data->q->q_usage_counter, nr - 1); |
| 432 | data->nr_tags -= nr; |
| 433 | |
| 434 | return rq_list_pop(data->cached_rq); |
| 435 | } |
| 436 | |
| 437 | static struct request *__blk_mq_alloc_requests(struct blk_mq_alloc_data *data) |
| 438 | { |
| 439 | struct request_queue *q = data->q; |
| 440 | u64 alloc_time_ns = 0; |
| 441 | struct request *rq; |
| 442 | unsigned int tag; |
| 443 | |
| 444 | /* alloc_time includes depth and tag waits */ |
| 445 | if (blk_queue_rq_alloc_time(q)) |
| 446 | alloc_time_ns = ktime_get_ns(); |
| 447 | |
| 448 | if (data->cmd_flags & REQ_NOWAIT) |
| 449 | data->flags |= BLK_MQ_REQ_NOWAIT; |
| 450 | |
| 451 | if (q->elevator) { |
| 452 | struct elevator_queue *e = q->elevator; |
| 453 | |
| 454 | data->rq_flags |= RQF_ELV; |
| 455 | |
| 456 | /* |
| 457 | * Flush/passthrough requests are special and go directly to the |
| 458 | * dispatch list. Don't include reserved tags in the |
| 459 | * limiting, as it isn't useful. |
| 460 | */ |
| 461 | if (!op_is_flush(data->cmd_flags) && |
| 462 | !blk_op_is_passthrough(data->cmd_flags) && |
| 463 | e->type->ops.limit_depth && |
| 464 | !(data->flags & BLK_MQ_REQ_RESERVED)) |
| 465 | e->type->ops.limit_depth(data->cmd_flags, data); |
| 466 | } |
| 467 | |
| 468 | retry: |
| 469 | data->ctx = blk_mq_get_ctx(q); |
| 470 | data->hctx = blk_mq_map_queue(q, data->cmd_flags, data->ctx); |
| 471 | if (!(data->rq_flags & RQF_ELV)) |
| 472 | blk_mq_tag_busy(data->hctx); |
| 473 | |
| 474 | if (data->flags & BLK_MQ_REQ_RESERVED) |
| 475 | data->rq_flags |= RQF_RESV; |
| 476 | |
| 477 | /* |
| 478 | * Try batched alloc if we want more than 1 tag. |
| 479 | */ |
| 480 | if (data->nr_tags > 1) { |
| 481 | rq = __blk_mq_alloc_requests_batch(data, alloc_time_ns); |
| 482 | if (rq) |
| 483 | return rq; |
| 484 | data->nr_tags = 1; |
| 485 | } |
| 486 | |
| 487 | /* |
| 488 | * Waiting allocations only fail because of an inactive hctx. In that |
| 489 | * case just retry the hctx assignment and tag allocation as CPU hotplug |
| 490 | * should have migrated us to an online CPU by now. |
| 491 | */ |
| 492 | tag = blk_mq_get_tag(data); |
| 493 | if (tag == BLK_MQ_NO_TAG) { |
| 494 | if (data->flags & BLK_MQ_REQ_NOWAIT) |
| 495 | return NULL; |
| 496 | /* |
| 497 | * Give up the CPU and sleep for a random short time to |
| 498 | * ensure that thread using a realtime scheduling class |
| 499 | * are migrated off the CPU, and thus off the hctx that |
| 500 | * is going away. |
| 501 | */ |
| 502 | msleep(3); |
| 503 | goto retry; |
| 504 | } |
| 505 | |
| 506 | return blk_mq_rq_ctx_init(data, blk_mq_tags_from_data(data), tag, |
| 507 | alloc_time_ns); |
| 508 | } |
| 509 | |
| 510 | static struct request *blk_mq_rq_cache_fill(struct request_queue *q, |
| 511 | struct blk_plug *plug, |
| 512 | blk_opf_t opf, |
| 513 | blk_mq_req_flags_t flags) |
| 514 | { |
| 515 | struct blk_mq_alloc_data data = { |
| 516 | .q = q, |
| 517 | .flags = flags, |
| 518 | .cmd_flags = opf, |
| 519 | .nr_tags = plug->nr_ios, |
| 520 | .cached_rq = &plug->cached_rq, |
| 521 | }; |
| 522 | struct request *rq; |
| 523 | |
| 524 | if (blk_queue_enter(q, flags)) |
| 525 | return NULL; |
| 526 | |
| 527 | plug->nr_ios = 1; |
| 528 | |
| 529 | rq = __blk_mq_alloc_requests(&data); |
| 530 | if (unlikely(!rq)) |
| 531 | blk_queue_exit(q); |
| 532 | return rq; |
| 533 | } |
| 534 | |
| 535 | static struct request *blk_mq_alloc_cached_request(struct request_queue *q, |
| 536 | blk_opf_t opf, |
| 537 | blk_mq_req_flags_t flags) |
| 538 | { |
| 539 | struct blk_plug *plug = current->plug; |
| 540 | struct request *rq; |
| 541 | |
| 542 | if (!plug) |
| 543 | return NULL; |
| 544 | |
| 545 | if (rq_list_empty(plug->cached_rq)) { |
| 546 | if (plug->nr_ios == 1) |
| 547 | return NULL; |
| 548 | rq = blk_mq_rq_cache_fill(q, plug, opf, flags); |
| 549 | if (!rq) |
| 550 | return NULL; |
| 551 | } else { |
| 552 | rq = rq_list_peek(&plug->cached_rq); |
| 553 | if (!rq || rq->q != q) |
| 554 | return NULL; |
| 555 | |
| 556 | if (blk_mq_get_hctx_type(opf) != rq->mq_hctx->type) |
| 557 | return NULL; |
| 558 | if (op_is_flush(rq->cmd_flags) != op_is_flush(opf)) |
| 559 | return NULL; |
| 560 | |
| 561 | plug->cached_rq = rq_list_next(rq); |
| 562 | } |
| 563 | |
| 564 | rq->cmd_flags = opf; |
| 565 | INIT_LIST_HEAD(&rq->queuelist); |
| 566 | return rq; |
| 567 | } |
| 568 | |
| 569 | struct request *blk_mq_alloc_request(struct request_queue *q, blk_opf_t opf, |
| 570 | blk_mq_req_flags_t flags) |
| 571 | { |
| 572 | struct request *rq; |
| 573 | |
| 574 | rq = blk_mq_alloc_cached_request(q, opf, flags); |
| 575 | if (!rq) { |
| 576 | struct blk_mq_alloc_data data = { |
| 577 | .q = q, |
| 578 | .flags = flags, |
| 579 | .cmd_flags = opf, |
| 580 | .nr_tags = 1, |
| 581 | }; |
| 582 | int ret; |
| 583 | |
| 584 | ret = blk_queue_enter(q, flags); |
| 585 | if (ret) |
| 586 | return ERR_PTR(ret); |
| 587 | |
| 588 | rq = __blk_mq_alloc_requests(&data); |
| 589 | if (!rq) |
| 590 | goto out_queue_exit; |
| 591 | } |
| 592 | rq->__data_len = 0; |
| 593 | rq->__sector = (sector_t) -1; |
| 594 | rq->bio = rq->biotail = NULL; |
| 595 | return rq; |
| 596 | out_queue_exit: |
| 597 | blk_queue_exit(q); |
| 598 | return ERR_PTR(-EWOULDBLOCK); |
| 599 | } |
| 600 | EXPORT_SYMBOL(blk_mq_alloc_request); |
| 601 | |
| 602 | struct request *blk_mq_alloc_request_hctx(struct request_queue *q, |
| 603 | blk_opf_t opf, blk_mq_req_flags_t flags, unsigned int hctx_idx) |
| 604 | { |
| 605 | struct blk_mq_alloc_data data = { |
| 606 | .q = q, |
| 607 | .flags = flags, |
| 608 | .cmd_flags = opf, |
| 609 | .nr_tags = 1, |
| 610 | }; |
| 611 | u64 alloc_time_ns = 0; |
| 612 | struct request *rq; |
| 613 | unsigned int cpu; |
| 614 | unsigned int tag; |
| 615 | int ret; |
| 616 | |
| 617 | /* alloc_time includes depth and tag waits */ |
| 618 | if (blk_queue_rq_alloc_time(q)) |
| 619 | alloc_time_ns = ktime_get_ns(); |
| 620 | |
| 621 | /* |
| 622 | * If the tag allocator sleeps we could get an allocation for a |
| 623 | * different hardware context. No need to complicate the low level |
| 624 | * allocator for this for the rare use case of a command tied to |
| 625 | * a specific queue. |
| 626 | */ |
| 627 | if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)) || |
| 628 | WARN_ON_ONCE(!(flags & BLK_MQ_REQ_RESERVED))) |
| 629 | return ERR_PTR(-EINVAL); |
| 630 | |
| 631 | if (hctx_idx >= q->nr_hw_queues) |
| 632 | return ERR_PTR(-EIO); |
| 633 | |
| 634 | ret = blk_queue_enter(q, flags); |
| 635 | if (ret) |
| 636 | return ERR_PTR(ret); |
| 637 | |
| 638 | /* |
| 639 | * Check if the hardware context is actually mapped to anything. |
| 640 | * If not tell the caller that it should skip this queue. |
| 641 | */ |
| 642 | ret = -EXDEV; |
| 643 | data.hctx = xa_load(&q->hctx_table, hctx_idx); |
| 644 | if (!blk_mq_hw_queue_mapped(data.hctx)) |
| 645 | goto out_queue_exit; |
| 646 | cpu = cpumask_first_and(data.hctx->cpumask, cpu_online_mask); |
| 647 | if (cpu >= nr_cpu_ids) |
| 648 | goto out_queue_exit; |
| 649 | data.ctx = __blk_mq_get_ctx(q, cpu); |
| 650 | |
| 651 | if (!q->elevator) |
| 652 | blk_mq_tag_busy(data.hctx); |
| 653 | else |
| 654 | data.rq_flags |= RQF_ELV; |
| 655 | |
| 656 | if (flags & BLK_MQ_REQ_RESERVED) |
| 657 | data.rq_flags |= RQF_RESV; |
| 658 | |
| 659 | ret = -EWOULDBLOCK; |
| 660 | tag = blk_mq_get_tag(&data); |
| 661 | if (tag == BLK_MQ_NO_TAG) |
| 662 | goto out_queue_exit; |
| 663 | rq = blk_mq_rq_ctx_init(&data, blk_mq_tags_from_data(&data), tag, |
| 664 | alloc_time_ns); |
| 665 | rq->__data_len = 0; |
| 666 | rq->__sector = (sector_t) -1; |
| 667 | rq->bio = rq->biotail = NULL; |
| 668 | return rq; |
| 669 | |
| 670 | out_queue_exit: |
| 671 | blk_queue_exit(q); |
| 672 | return ERR_PTR(ret); |
| 673 | } |
| 674 | EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx); |
| 675 | |
| 676 | static void __blk_mq_free_request(struct request *rq) |
| 677 | { |
| 678 | struct request_queue *q = rq->q; |
| 679 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
| 680 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
| 681 | const int sched_tag = rq->internal_tag; |
| 682 | |
| 683 | blk_crypto_free_request(rq); |
| 684 | blk_pm_mark_last_busy(rq); |
| 685 | rq->mq_hctx = NULL; |
| 686 | if (rq->tag != BLK_MQ_NO_TAG) |
| 687 | blk_mq_put_tag(hctx->tags, ctx, rq->tag); |
| 688 | if (sched_tag != BLK_MQ_NO_TAG) |
| 689 | blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag); |
| 690 | blk_mq_sched_restart(hctx); |
| 691 | blk_queue_exit(q); |
| 692 | } |
| 693 | |
| 694 | void blk_mq_free_request(struct request *rq) |
| 695 | { |
| 696 | struct request_queue *q = rq->q; |
| 697 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
| 698 | |
| 699 | if ((rq->rq_flags & RQF_ELVPRIV) && |
| 700 | q->elevator->type->ops.finish_request) |
| 701 | q->elevator->type->ops.finish_request(rq); |
| 702 | |
| 703 | if (rq->rq_flags & RQF_MQ_INFLIGHT) |
| 704 | __blk_mq_dec_active_requests(hctx); |
| 705 | |
| 706 | if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq))) |
| 707 | laptop_io_completion(q->disk->bdi); |
| 708 | |
| 709 | rq_qos_done(q, rq); |
| 710 | |
| 711 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); |
| 712 | if (req_ref_put_and_test(rq)) |
| 713 | __blk_mq_free_request(rq); |
| 714 | } |
| 715 | EXPORT_SYMBOL_GPL(blk_mq_free_request); |
| 716 | |
| 717 | void blk_mq_free_plug_rqs(struct blk_plug *plug) |
| 718 | { |
| 719 | struct request *rq; |
| 720 | |
| 721 | while ((rq = rq_list_pop(&plug->cached_rq)) != NULL) |
| 722 | blk_mq_free_request(rq); |
| 723 | } |
| 724 | |
| 725 | void blk_dump_rq_flags(struct request *rq, char *msg) |
| 726 | { |
| 727 | printk(KERN_INFO "%s: dev %s: flags=%llx\n", msg, |
| 728 | rq->q->disk ? rq->q->disk->disk_name : "?", |
| 729 | (__force unsigned long long) rq->cmd_flags); |
| 730 | |
| 731 | printk(KERN_INFO " sector %llu, nr/cnr %u/%u\n", |
| 732 | (unsigned long long)blk_rq_pos(rq), |
| 733 | blk_rq_sectors(rq), blk_rq_cur_sectors(rq)); |
| 734 | printk(KERN_INFO " bio %p, biotail %p, len %u\n", |
| 735 | rq->bio, rq->biotail, blk_rq_bytes(rq)); |
| 736 | } |
| 737 | EXPORT_SYMBOL(blk_dump_rq_flags); |
| 738 | |
| 739 | static void req_bio_endio(struct request *rq, struct bio *bio, |
| 740 | unsigned int nbytes, blk_status_t error) |
| 741 | { |
| 742 | if (unlikely(error)) { |
| 743 | bio->bi_status = error; |
| 744 | } else if (req_op(rq) == REQ_OP_ZONE_APPEND) { |
| 745 | /* |
| 746 | * Partial zone append completions cannot be supported as the |
| 747 | * BIO fragments may end up not being written sequentially. |
| 748 | */ |
| 749 | if (bio->bi_iter.bi_size != nbytes) |
| 750 | bio->bi_status = BLK_STS_IOERR; |
| 751 | else |
| 752 | bio->bi_iter.bi_sector = rq->__sector; |
| 753 | } |
| 754 | |
| 755 | bio_advance(bio, nbytes); |
| 756 | |
| 757 | if (unlikely(rq->rq_flags & RQF_QUIET)) |
| 758 | bio_set_flag(bio, BIO_QUIET); |
| 759 | /* don't actually finish bio if it's part of flush sequence */ |
| 760 | if (bio->bi_iter.bi_size == 0 && !(rq->rq_flags & RQF_FLUSH_SEQ)) |
| 761 | bio_endio(bio); |
| 762 | } |
| 763 | |
| 764 | static void blk_account_io_completion(struct request *req, unsigned int bytes) |
| 765 | { |
| 766 | if (req->part && blk_do_io_stat(req)) { |
| 767 | const int sgrp = op_stat_group(req_op(req)); |
| 768 | |
| 769 | part_stat_lock(); |
| 770 | part_stat_add(req->part, sectors[sgrp], bytes >> 9); |
| 771 | part_stat_unlock(); |
| 772 | } |
| 773 | } |
| 774 | |
| 775 | static void blk_print_req_error(struct request *req, blk_status_t status) |
| 776 | { |
| 777 | printk_ratelimited(KERN_ERR |
| 778 | "%s error, dev %s, sector %llu op 0x%x:(%s) flags 0x%x " |
| 779 | "phys_seg %u prio class %u\n", |
| 780 | blk_status_to_str(status), |
| 781 | req->q->disk ? req->q->disk->disk_name : "?", |
| 782 | blk_rq_pos(req), (__force u32)req_op(req), |
| 783 | blk_op_str(req_op(req)), |
| 784 | (__force u32)(req->cmd_flags & ~REQ_OP_MASK), |
| 785 | req->nr_phys_segments, |
| 786 | IOPRIO_PRIO_CLASS(req->ioprio)); |
| 787 | } |
| 788 | |
| 789 | /* |
| 790 | * Fully end IO on a request. Does not support partial completions, or |
| 791 | * errors. |
| 792 | */ |
| 793 | static void blk_complete_request(struct request *req) |
| 794 | { |
| 795 | const bool is_flush = (req->rq_flags & RQF_FLUSH_SEQ) != 0; |
| 796 | int total_bytes = blk_rq_bytes(req); |
| 797 | struct bio *bio = req->bio; |
| 798 | |
| 799 | trace_block_rq_complete(req, BLK_STS_OK, total_bytes); |
| 800 | |
| 801 | if (!bio) |
| 802 | return; |
| 803 | |
| 804 | #ifdef CONFIG_BLK_DEV_INTEGRITY |
| 805 | if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ) |
| 806 | req->q->integrity.profile->complete_fn(req, total_bytes); |
| 807 | #endif |
| 808 | |
| 809 | /* |
| 810 | * Upper layers may call blk_crypto_evict_key() anytime after the last |
| 811 | * bio_endio(). Therefore, the keyslot must be released before that. |
| 812 | */ |
| 813 | blk_crypto_rq_put_keyslot(req); |
| 814 | |
| 815 | blk_account_io_completion(req, total_bytes); |
| 816 | |
| 817 | do { |
| 818 | struct bio *next = bio->bi_next; |
| 819 | |
| 820 | /* Completion has already been traced */ |
| 821 | bio_clear_flag(bio, BIO_TRACE_COMPLETION); |
| 822 | |
| 823 | if (req_op(req) == REQ_OP_ZONE_APPEND) |
| 824 | bio->bi_iter.bi_sector = req->__sector; |
| 825 | |
| 826 | if (!is_flush) |
| 827 | bio_endio(bio); |
| 828 | bio = next; |
| 829 | } while (bio); |
| 830 | |
| 831 | /* |
| 832 | * Reset counters so that the request stacking driver |
| 833 | * can find how many bytes remain in the request |
| 834 | * later. |
| 835 | */ |
| 836 | if (!req->end_io) { |
| 837 | req->bio = NULL; |
| 838 | req->__data_len = 0; |
| 839 | } |
| 840 | } |
| 841 | |
| 842 | /** |
| 843 | * blk_update_request - Complete multiple bytes without completing the request |
| 844 | * @req: the request being processed |
| 845 | * @error: block status code |
| 846 | * @nr_bytes: number of bytes to complete for @req |
| 847 | * |
| 848 | * Description: |
| 849 | * Ends I/O on a number of bytes attached to @req, but doesn't complete |
| 850 | * the request structure even if @req doesn't have leftover. |
| 851 | * If @req has leftover, sets it up for the next range of segments. |
| 852 | * |
| 853 | * Passing the result of blk_rq_bytes() as @nr_bytes guarantees |
| 854 | * %false return from this function. |
| 855 | * |
| 856 | * Note: |
| 857 | * The RQF_SPECIAL_PAYLOAD flag is ignored on purpose in this function |
| 858 | * except in the consistency check at the end of this function. |
| 859 | * |
| 860 | * Return: |
| 861 | * %false - this request doesn't have any more data |
| 862 | * %true - this request has more data |
| 863 | **/ |
| 864 | bool blk_update_request(struct request *req, blk_status_t error, |
| 865 | unsigned int nr_bytes) |
| 866 | { |
| 867 | int total_bytes; |
| 868 | |
| 869 | trace_block_rq_complete(req, error, nr_bytes); |
| 870 | |
| 871 | if (!req->bio) |
| 872 | return false; |
| 873 | |
| 874 | #ifdef CONFIG_BLK_DEV_INTEGRITY |
| 875 | if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ && |
| 876 | error == BLK_STS_OK) |
| 877 | req->q->integrity.profile->complete_fn(req, nr_bytes); |
| 878 | #endif |
| 879 | |
| 880 | /* |
| 881 | * Upper layers may call blk_crypto_evict_key() anytime after the last |
| 882 | * bio_endio(). Therefore, the keyslot must be released before that. |
| 883 | */ |
| 884 | if (blk_crypto_rq_has_keyslot(req) && nr_bytes >= blk_rq_bytes(req)) |
| 885 | __blk_crypto_rq_put_keyslot(req); |
| 886 | |
| 887 | if (unlikely(error && !blk_rq_is_passthrough(req) && |
| 888 | !(req->rq_flags & RQF_QUIET)) && |
| 889 | !test_bit(GD_DEAD, &req->q->disk->state)) { |
| 890 | blk_print_req_error(req, error); |
| 891 | trace_block_rq_error(req, error, nr_bytes); |
| 892 | } |
| 893 | |
| 894 | blk_account_io_completion(req, nr_bytes); |
| 895 | |
| 896 | total_bytes = 0; |
| 897 | while (req->bio) { |
| 898 | struct bio *bio = req->bio; |
| 899 | unsigned bio_bytes = min(bio->bi_iter.bi_size, nr_bytes); |
| 900 | |
| 901 | if (bio_bytes == bio->bi_iter.bi_size) |
| 902 | req->bio = bio->bi_next; |
| 903 | |
| 904 | /* Completion has already been traced */ |
| 905 | bio_clear_flag(bio, BIO_TRACE_COMPLETION); |
| 906 | req_bio_endio(req, bio, bio_bytes, error); |
| 907 | |
| 908 | total_bytes += bio_bytes; |
| 909 | nr_bytes -= bio_bytes; |
| 910 | |
| 911 | if (!nr_bytes) |
| 912 | break; |
| 913 | } |
| 914 | |
| 915 | /* |
| 916 | * completely done |
| 917 | */ |
| 918 | if (!req->bio) { |
| 919 | /* |
| 920 | * Reset counters so that the request stacking driver |
| 921 | * can find how many bytes remain in the request |
| 922 | * later. |
| 923 | */ |
| 924 | req->__data_len = 0; |
| 925 | return false; |
| 926 | } |
| 927 | |
| 928 | req->__data_len -= total_bytes; |
| 929 | |
| 930 | /* update sector only for requests with clear definition of sector */ |
| 931 | if (!blk_rq_is_passthrough(req)) |
| 932 | req->__sector += total_bytes >> 9; |
| 933 | |
| 934 | /* mixed attributes always follow the first bio */ |
| 935 | if (req->rq_flags & RQF_MIXED_MERGE) { |
| 936 | req->cmd_flags &= ~REQ_FAILFAST_MASK; |
| 937 | req->cmd_flags |= req->bio->bi_opf & REQ_FAILFAST_MASK; |
| 938 | } |
| 939 | |
| 940 | if (!(req->rq_flags & RQF_SPECIAL_PAYLOAD)) { |
| 941 | /* |
| 942 | * If total number of sectors is less than the first segment |
| 943 | * size, something has gone terribly wrong. |
| 944 | */ |
| 945 | if (blk_rq_bytes(req) < blk_rq_cur_bytes(req)) { |
| 946 | blk_dump_rq_flags(req, "request botched"); |
| 947 | req->__data_len = blk_rq_cur_bytes(req); |
| 948 | } |
| 949 | |
| 950 | /* recalculate the number of segments */ |
| 951 | req->nr_phys_segments = blk_recalc_rq_segments(req); |
| 952 | } |
| 953 | |
| 954 | return true; |
| 955 | } |
| 956 | EXPORT_SYMBOL_GPL(blk_update_request); |
| 957 | |
| 958 | static inline void blk_account_io_done(struct request *req, u64 now) |
| 959 | { |
| 960 | /* |
| 961 | * Account IO completion. flush_rq isn't accounted as a |
| 962 | * normal IO on queueing nor completion. Accounting the |
| 963 | * containing request is enough. |
| 964 | */ |
| 965 | if (blk_do_io_stat(req) && req->part && |
| 966 | !(req->rq_flags & RQF_FLUSH_SEQ)) { |
| 967 | const int sgrp = op_stat_group(req_op(req)); |
| 968 | |
| 969 | part_stat_lock(); |
| 970 | update_io_ticks(req->part, jiffies, true); |
| 971 | part_stat_inc(req->part, ios[sgrp]); |
| 972 | part_stat_add(req->part, nsecs[sgrp], now - req->start_time_ns); |
| 973 | part_stat_unlock(); |
| 974 | } |
| 975 | } |
| 976 | |
| 977 | static inline void blk_account_io_start(struct request *req) |
| 978 | { |
| 979 | if (blk_do_io_stat(req)) { |
| 980 | /* |
| 981 | * All non-passthrough requests are created from a bio with one |
| 982 | * exception: when a flush command that is part of a flush sequence |
| 983 | * generated by the state machine in blk-flush.c is cloned onto the |
| 984 | * lower device by dm-multipath we can get here without a bio. |
| 985 | */ |
| 986 | if (req->bio) |
| 987 | req->part = req->bio->bi_bdev; |
| 988 | else |
| 989 | req->part = req->q->disk->part0; |
| 990 | |
| 991 | part_stat_lock(); |
| 992 | update_io_ticks(req->part, jiffies, false); |
| 993 | part_stat_unlock(); |
| 994 | } |
| 995 | } |
| 996 | |
| 997 | static inline void __blk_mq_end_request_acct(struct request *rq, u64 now) |
| 998 | { |
| 999 | if (rq->rq_flags & RQF_STATS) |
| 1000 | blk_stat_add(rq, now); |
| 1001 | |
| 1002 | blk_mq_sched_completed_request(rq, now); |
| 1003 | blk_account_io_done(rq, now); |
| 1004 | } |
| 1005 | |
| 1006 | inline void __blk_mq_end_request(struct request *rq, blk_status_t error) |
| 1007 | { |
| 1008 | if (blk_mq_need_time_stamp(rq)) |
| 1009 | __blk_mq_end_request_acct(rq, ktime_get_ns()); |
| 1010 | |
| 1011 | if (rq->end_io) { |
| 1012 | rq_qos_done(rq->q, rq); |
| 1013 | if (rq->end_io(rq, error) == RQ_END_IO_FREE) |
| 1014 | blk_mq_free_request(rq); |
| 1015 | } else { |
| 1016 | blk_mq_free_request(rq); |
| 1017 | } |
| 1018 | } |
| 1019 | EXPORT_SYMBOL(__blk_mq_end_request); |
| 1020 | |
| 1021 | void blk_mq_end_request(struct request *rq, blk_status_t error) |
| 1022 | { |
| 1023 | if (blk_update_request(rq, error, blk_rq_bytes(rq))) |
| 1024 | BUG(); |
| 1025 | __blk_mq_end_request(rq, error); |
| 1026 | } |
| 1027 | EXPORT_SYMBOL(blk_mq_end_request); |
| 1028 | |
| 1029 | #define TAG_COMP_BATCH 32 |
| 1030 | |
| 1031 | static inline void blk_mq_flush_tag_batch(struct blk_mq_hw_ctx *hctx, |
| 1032 | int *tag_array, int nr_tags) |
| 1033 | { |
| 1034 | struct request_queue *q = hctx->queue; |
| 1035 | |
| 1036 | /* |
| 1037 | * All requests should have been marked as RQF_MQ_INFLIGHT, so |
| 1038 | * update hctx->nr_active in batch |
| 1039 | */ |
| 1040 | if (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) |
| 1041 | __blk_mq_sub_active_requests(hctx, nr_tags); |
| 1042 | |
| 1043 | blk_mq_put_tags(hctx->tags, tag_array, nr_tags); |
| 1044 | percpu_ref_put_many(&q->q_usage_counter, nr_tags); |
| 1045 | } |
| 1046 | |
| 1047 | void blk_mq_end_request_batch(struct io_comp_batch *iob) |
| 1048 | { |
| 1049 | int tags[TAG_COMP_BATCH], nr_tags = 0; |
| 1050 | struct blk_mq_hw_ctx *cur_hctx = NULL; |
| 1051 | struct request *rq; |
| 1052 | u64 now = 0; |
| 1053 | |
| 1054 | if (iob->need_ts) |
| 1055 | now = ktime_get_ns(); |
| 1056 | |
| 1057 | while ((rq = rq_list_pop(&iob->req_list)) != NULL) { |
| 1058 | prefetch(rq->bio); |
| 1059 | prefetch(rq->rq_next); |
| 1060 | |
| 1061 | blk_complete_request(rq); |
| 1062 | if (iob->need_ts) |
| 1063 | __blk_mq_end_request_acct(rq, now); |
| 1064 | |
| 1065 | rq_qos_done(rq->q, rq); |
| 1066 | |
| 1067 | /* |
| 1068 | * If end_io handler returns NONE, then it still has |
| 1069 | * ownership of the request. |
| 1070 | */ |
| 1071 | if (rq->end_io && rq->end_io(rq, 0) == RQ_END_IO_NONE) |
| 1072 | continue; |
| 1073 | |
| 1074 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); |
| 1075 | if (!req_ref_put_and_test(rq)) |
| 1076 | continue; |
| 1077 | |
| 1078 | blk_crypto_free_request(rq); |
| 1079 | blk_pm_mark_last_busy(rq); |
| 1080 | |
| 1081 | if (nr_tags == TAG_COMP_BATCH || cur_hctx != rq->mq_hctx) { |
| 1082 | if (cur_hctx) |
| 1083 | blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags); |
| 1084 | nr_tags = 0; |
| 1085 | cur_hctx = rq->mq_hctx; |
| 1086 | } |
| 1087 | tags[nr_tags++] = rq->tag; |
| 1088 | } |
| 1089 | |
| 1090 | if (nr_tags) |
| 1091 | blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags); |
| 1092 | } |
| 1093 | EXPORT_SYMBOL_GPL(blk_mq_end_request_batch); |
| 1094 | |
| 1095 | static void blk_complete_reqs(struct llist_head *list) |
| 1096 | { |
| 1097 | struct llist_node *entry = llist_reverse_order(llist_del_all(list)); |
| 1098 | struct request *rq, *next; |
| 1099 | |
| 1100 | llist_for_each_entry_safe(rq, next, entry, ipi_list) |
| 1101 | rq->q->mq_ops->complete(rq); |
| 1102 | } |
| 1103 | |
| 1104 | static __latent_entropy void blk_done_softirq(struct softirq_action *h) |
| 1105 | { |
| 1106 | blk_complete_reqs(this_cpu_ptr(&blk_cpu_done)); |
| 1107 | } |
| 1108 | |
| 1109 | static int blk_softirq_cpu_dead(unsigned int cpu) |
| 1110 | { |
| 1111 | blk_complete_reqs(&per_cpu(blk_cpu_done, cpu)); |
| 1112 | return 0; |
| 1113 | } |
| 1114 | |
| 1115 | static void __blk_mq_complete_request_remote(void *data) |
| 1116 | { |
| 1117 | __raise_softirq_irqoff(BLOCK_SOFTIRQ); |
| 1118 | } |
| 1119 | |
| 1120 | static inline bool blk_mq_complete_need_ipi(struct request *rq) |
| 1121 | { |
| 1122 | int cpu = raw_smp_processor_id(); |
| 1123 | |
| 1124 | if (!IS_ENABLED(CONFIG_SMP) || |
| 1125 | !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) |
| 1126 | return false; |
| 1127 | /* |
| 1128 | * With force threaded interrupts enabled, raising softirq from an SMP |
| 1129 | * function call will always result in waking the ksoftirqd thread. |
| 1130 | * This is probably worse than completing the request on a different |
| 1131 | * cache domain. |
| 1132 | */ |
| 1133 | if (force_irqthreads()) |
| 1134 | return false; |
| 1135 | |
| 1136 | /* same CPU or cache domain? Complete locally */ |
| 1137 | if (cpu == rq->mq_ctx->cpu || |
| 1138 | (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) && |
| 1139 | cpus_share_cache(cpu, rq->mq_ctx->cpu))) |
| 1140 | return false; |
| 1141 | |
| 1142 | /* don't try to IPI to an offline CPU */ |
| 1143 | return cpu_online(rq->mq_ctx->cpu); |
| 1144 | } |
| 1145 | |
| 1146 | static void blk_mq_complete_send_ipi(struct request *rq) |
| 1147 | { |
| 1148 | struct llist_head *list; |
| 1149 | unsigned int cpu; |
| 1150 | |
| 1151 | cpu = rq->mq_ctx->cpu; |
| 1152 | list = &per_cpu(blk_cpu_done, cpu); |
| 1153 | if (llist_add(&rq->ipi_list, list)) { |
| 1154 | INIT_CSD(&rq->csd, __blk_mq_complete_request_remote, rq); |
| 1155 | smp_call_function_single_async(cpu, &rq->csd); |
| 1156 | } |
| 1157 | } |
| 1158 | |
| 1159 | static void blk_mq_raise_softirq(struct request *rq) |
| 1160 | { |
| 1161 | struct llist_head *list; |
| 1162 | |
| 1163 | preempt_disable(); |
| 1164 | list = this_cpu_ptr(&blk_cpu_done); |
| 1165 | if (llist_add(&rq->ipi_list, list)) |
| 1166 | raise_softirq(BLOCK_SOFTIRQ); |
| 1167 | preempt_enable(); |
| 1168 | } |
| 1169 | |
| 1170 | bool blk_mq_complete_request_remote(struct request *rq) |
| 1171 | { |
| 1172 | WRITE_ONCE(rq->state, MQ_RQ_COMPLETE); |
| 1173 | |
| 1174 | /* |
| 1175 | * For request which hctx has only one ctx mapping, |
| 1176 | * or a polled request, always complete locally, |
| 1177 | * it's pointless to redirect the completion. |
| 1178 | */ |
| 1179 | if (rq->mq_hctx->nr_ctx == 1 || |
| 1180 | rq->cmd_flags & REQ_POLLED) |
| 1181 | return false; |
| 1182 | |
| 1183 | if (blk_mq_complete_need_ipi(rq)) { |
| 1184 | blk_mq_complete_send_ipi(rq); |
| 1185 | return true; |
| 1186 | } |
| 1187 | |
| 1188 | if (rq->q->nr_hw_queues == 1) { |
| 1189 | blk_mq_raise_softirq(rq); |
| 1190 | return true; |
| 1191 | } |
| 1192 | return false; |
| 1193 | } |
| 1194 | EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote); |
| 1195 | |
| 1196 | /** |
| 1197 | * blk_mq_complete_request - end I/O on a request |
| 1198 | * @rq: the request being processed |
| 1199 | * |
| 1200 | * Description: |
| 1201 | * Complete a request by scheduling the ->complete_rq operation. |
| 1202 | **/ |
| 1203 | void blk_mq_complete_request(struct request *rq) |
| 1204 | { |
| 1205 | if (!blk_mq_complete_request_remote(rq)) |
| 1206 | rq->q->mq_ops->complete(rq); |
| 1207 | } |
| 1208 | EXPORT_SYMBOL(blk_mq_complete_request); |
| 1209 | |
| 1210 | /** |
| 1211 | * blk_mq_start_request - Start processing a request |
| 1212 | * @rq: Pointer to request to be started |
| 1213 | * |
| 1214 | * Function used by device drivers to notify the block layer that a request |
| 1215 | * is going to be processed now, so blk layer can do proper initializations |
| 1216 | * such as starting the timeout timer. |
| 1217 | */ |
| 1218 | void blk_mq_start_request(struct request *rq) |
| 1219 | { |
| 1220 | struct request_queue *q = rq->q; |
| 1221 | |
| 1222 | trace_block_rq_issue(rq); |
| 1223 | |
| 1224 | if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) { |
| 1225 | rq->io_start_time_ns = ktime_get_ns(); |
| 1226 | rq->stats_sectors = blk_rq_sectors(rq); |
| 1227 | rq->rq_flags |= RQF_STATS; |
| 1228 | rq_qos_issue(q, rq); |
| 1229 | } |
| 1230 | |
| 1231 | WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE); |
| 1232 | |
| 1233 | blk_add_timer(rq); |
| 1234 | WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT); |
| 1235 | |
| 1236 | #ifdef CONFIG_BLK_DEV_INTEGRITY |
| 1237 | if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE) |
| 1238 | q->integrity.profile->prepare_fn(rq); |
| 1239 | #endif |
| 1240 | if (rq->bio && rq->bio->bi_opf & REQ_POLLED) |
| 1241 | WRITE_ONCE(rq->bio->bi_cookie, blk_rq_to_qc(rq)); |
| 1242 | } |
| 1243 | EXPORT_SYMBOL(blk_mq_start_request); |
| 1244 | |
| 1245 | /* |
| 1246 | * Allow 2x BLK_MAX_REQUEST_COUNT requests on plug queue for multiple |
| 1247 | * queues. This is important for md arrays to benefit from merging |
| 1248 | * requests. |
| 1249 | */ |
| 1250 | static inline unsigned short blk_plug_max_rq_count(struct blk_plug *plug) |
| 1251 | { |
| 1252 | if (plug->multiple_queues) |
| 1253 | return BLK_MAX_REQUEST_COUNT * 2; |
| 1254 | return BLK_MAX_REQUEST_COUNT; |
| 1255 | } |
| 1256 | |
| 1257 | static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq) |
| 1258 | { |
| 1259 | struct request *last = rq_list_peek(&plug->mq_list); |
| 1260 | |
| 1261 | if (!plug->rq_count) { |
| 1262 | trace_block_plug(rq->q); |
| 1263 | } else if (plug->rq_count >= blk_plug_max_rq_count(plug) || |
| 1264 | (!blk_queue_nomerges(rq->q) && |
| 1265 | blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) { |
| 1266 | blk_mq_flush_plug_list(plug, false); |
| 1267 | last = NULL; |
| 1268 | trace_block_plug(rq->q); |
| 1269 | } |
| 1270 | |
| 1271 | if (!plug->multiple_queues && last && last->q != rq->q) |
| 1272 | plug->multiple_queues = true; |
| 1273 | if (!plug->has_elevator && (rq->rq_flags & RQF_ELV)) |
| 1274 | plug->has_elevator = true; |
| 1275 | rq->rq_next = NULL; |
| 1276 | rq_list_add(&plug->mq_list, rq); |
| 1277 | plug->rq_count++; |
| 1278 | } |
| 1279 | |
| 1280 | /** |
| 1281 | * blk_execute_rq_nowait - insert a request to I/O scheduler for execution |
| 1282 | * @rq: request to insert |
| 1283 | * @at_head: insert request at head or tail of queue |
| 1284 | * |
| 1285 | * Description: |
| 1286 | * Insert a fully prepared request at the back of the I/O scheduler queue |
| 1287 | * for execution. Don't wait for completion. |
| 1288 | * |
| 1289 | * Note: |
| 1290 | * This function will invoke @done directly if the queue is dead. |
| 1291 | */ |
| 1292 | void blk_execute_rq_nowait(struct request *rq, bool at_head) |
| 1293 | { |
| 1294 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
| 1295 | |
| 1296 | WARN_ON(irqs_disabled()); |
| 1297 | WARN_ON(!blk_rq_is_passthrough(rq)); |
| 1298 | |
| 1299 | blk_account_io_start(rq); |
| 1300 | |
| 1301 | /* |
| 1302 | * As plugging can be enabled for passthrough requests on a zoned |
| 1303 | * device, directly accessing the plug instead of using blk_mq_plug() |
| 1304 | * should not have any consequences. |
| 1305 | */ |
| 1306 | if (current->plug && !at_head) { |
| 1307 | blk_add_rq_to_plug(current->plug, rq); |
| 1308 | return; |
| 1309 | } |
| 1310 | |
| 1311 | blk_mq_insert_request(rq, at_head ? BLK_MQ_INSERT_AT_HEAD : 0); |
| 1312 | blk_mq_run_hw_queue(hctx, false); |
| 1313 | } |
| 1314 | EXPORT_SYMBOL_GPL(blk_execute_rq_nowait); |
| 1315 | |
| 1316 | struct blk_rq_wait { |
| 1317 | struct completion done; |
| 1318 | blk_status_t ret; |
| 1319 | }; |
| 1320 | |
| 1321 | static enum rq_end_io_ret blk_end_sync_rq(struct request *rq, blk_status_t ret) |
| 1322 | { |
| 1323 | struct blk_rq_wait *wait = rq->end_io_data; |
| 1324 | |
| 1325 | wait->ret = ret; |
| 1326 | complete(&wait->done); |
| 1327 | return RQ_END_IO_NONE; |
| 1328 | } |
| 1329 | |
| 1330 | bool blk_rq_is_poll(struct request *rq) |
| 1331 | { |
| 1332 | if (!rq->mq_hctx) |
| 1333 | return false; |
| 1334 | if (rq->mq_hctx->type != HCTX_TYPE_POLL) |
| 1335 | return false; |
| 1336 | return true; |
| 1337 | } |
| 1338 | EXPORT_SYMBOL_GPL(blk_rq_is_poll); |
| 1339 | |
| 1340 | static void blk_rq_poll_completion(struct request *rq, struct completion *wait) |
| 1341 | { |
| 1342 | do { |
| 1343 | blk_mq_poll(rq->q, blk_rq_to_qc(rq), NULL, 0); |
| 1344 | cond_resched(); |
| 1345 | } while (!completion_done(wait)); |
| 1346 | } |
| 1347 | |
| 1348 | /** |
| 1349 | * blk_execute_rq - insert a request into queue for execution |
| 1350 | * @rq: request to insert |
| 1351 | * @at_head: insert request at head or tail of queue |
| 1352 | * |
| 1353 | * Description: |
| 1354 | * Insert a fully prepared request at the back of the I/O scheduler queue |
| 1355 | * for execution and wait for completion. |
| 1356 | * Return: The blk_status_t result provided to blk_mq_end_request(). |
| 1357 | */ |
| 1358 | blk_status_t blk_execute_rq(struct request *rq, bool at_head) |
| 1359 | { |
| 1360 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
| 1361 | struct blk_rq_wait wait = { |
| 1362 | .done = COMPLETION_INITIALIZER_ONSTACK(wait.done), |
| 1363 | }; |
| 1364 | |
| 1365 | WARN_ON(irqs_disabled()); |
| 1366 | WARN_ON(!blk_rq_is_passthrough(rq)); |
| 1367 | |
| 1368 | rq->end_io_data = &wait; |
| 1369 | rq->end_io = blk_end_sync_rq; |
| 1370 | |
| 1371 | blk_account_io_start(rq); |
| 1372 | blk_mq_insert_request(rq, at_head ? BLK_MQ_INSERT_AT_HEAD : 0); |
| 1373 | blk_mq_run_hw_queue(hctx, false); |
| 1374 | |
| 1375 | if (blk_rq_is_poll(rq)) { |
| 1376 | blk_rq_poll_completion(rq, &wait.done); |
| 1377 | } else { |
| 1378 | /* |
| 1379 | * Prevent hang_check timer from firing at us during very long |
| 1380 | * I/O |
| 1381 | */ |
| 1382 | unsigned long hang_check = sysctl_hung_task_timeout_secs; |
| 1383 | |
| 1384 | if (hang_check) |
| 1385 | while (!wait_for_completion_io_timeout(&wait.done, |
| 1386 | hang_check * (HZ/2))) |
| 1387 | ; |
| 1388 | else |
| 1389 | wait_for_completion_io(&wait.done); |
| 1390 | } |
| 1391 | |
| 1392 | return wait.ret; |
| 1393 | } |
| 1394 | EXPORT_SYMBOL(blk_execute_rq); |
| 1395 | |
| 1396 | static void __blk_mq_requeue_request(struct request *rq) |
| 1397 | { |
| 1398 | struct request_queue *q = rq->q; |
| 1399 | |
| 1400 | blk_mq_put_driver_tag(rq); |
| 1401 | |
| 1402 | trace_block_rq_requeue(rq); |
| 1403 | rq_qos_requeue(q, rq); |
| 1404 | |
| 1405 | if (blk_mq_request_started(rq)) { |
| 1406 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); |
| 1407 | rq->rq_flags &= ~RQF_TIMED_OUT; |
| 1408 | } |
| 1409 | } |
| 1410 | |
| 1411 | void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list) |
| 1412 | { |
| 1413 | struct request_queue *q = rq->q; |
| 1414 | |
| 1415 | __blk_mq_requeue_request(rq); |
| 1416 | |
| 1417 | /* this request will be re-inserted to io scheduler queue */ |
| 1418 | blk_mq_sched_requeue_request(rq); |
| 1419 | |
| 1420 | blk_mq_add_to_requeue_list(rq, BLK_MQ_INSERT_AT_HEAD); |
| 1421 | |
| 1422 | if (kick_requeue_list) |
| 1423 | blk_mq_kick_requeue_list(q); |
| 1424 | } |
| 1425 | EXPORT_SYMBOL(blk_mq_requeue_request); |
| 1426 | |
| 1427 | static void blk_mq_requeue_work(struct work_struct *work) |
| 1428 | { |
| 1429 | struct request_queue *q = |
| 1430 | container_of(work, struct request_queue, requeue_work.work); |
| 1431 | LIST_HEAD(rq_list); |
| 1432 | struct request *rq, *next; |
| 1433 | |
| 1434 | spin_lock_irq(&q->requeue_lock); |
| 1435 | list_splice_init(&q->requeue_list, &rq_list); |
| 1436 | spin_unlock_irq(&q->requeue_lock); |
| 1437 | |
| 1438 | list_for_each_entry_safe(rq, next, &rq_list, queuelist) { |
| 1439 | /* |
| 1440 | * If RQF_DONTPREP ist set, the request has been started by the |
| 1441 | * driver already and might have driver-specific data allocated |
| 1442 | * already. Insert it into the hctx dispatch list to avoid |
| 1443 | * block layer merges for the request. |
| 1444 | */ |
| 1445 | if (rq->rq_flags & RQF_DONTPREP) { |
| 1446 | rq->rq_flags &= ~RQF_SOFTBARRIER; |
| 1447 | list_del_init(&rq->queuelist); |
| 1448 | blk_mq_request_bypass_insert(rq, 0); |
| 1449 | } else if (rq->rq_flags & RQF_SOFTBARRIER) { |
| 1450 | rq->rq_flags &= ~RQF_SOFTBARRIER; |
| 1451 | list_del_init(&rq->queuelist); |
| 1452 | blk_mq_insert_request(rq, BLK_MQ_INSERT_AT_HEAD); |
| 1453 | } |
| 1454 | } |
| 1455 | |
| 1456 | while (!list_empty(&rq_list)) { |
| 1457 | rq = list_entry(rq_list.next, struct request, queuelist); |
| 1458 | list_del_init(&rq->queuelist); |
| 1459 | blk_mq_insert_request(rq, 0); |
| 1460 | } |
| 1461 | |
| 1462 | blk_mq_run_hw_queues(q, false); |
| 1463 | } |
| 1464 | |
| 1465 | void blk_mq_add_to_requeue_list(struct request *rq, blk_insert_t insert_flags) |
| 1466 | { |
| 1467 | struct request_queue *q = rq->q; |
| 1468 | unsigned long flags; |
| 1469 | |
| 1470 | /* |
| 1471 | * We abuse this flag that is otherwise used by the I/O scheduler to |
| 1472 | * request head insertion from the workqueue. |
| 1473 | */ |
| 1474 | BUG_ON(rq->rq_flags & RQF_SOFTBARRIER); |
| 1475 | |
| 1476 | spin_lock_irqsave(&q->requeue_lock, flags); |
| 1477 | if (insert_flags & BLK_MQ_INSERT_AT_HEAD) { |
| 1478 | rq->rq_flags |= RQF_SOFTBARRIER; |
| 1479 | list_add(&rq->queuelist, &q->requeue_list); |
| 1480 | } else { |
| 1481 | list_add_tail(&rq->queuelist, &q->requeue_list); |
| 1482 | } |
| 1483 | spin_unlock_irqrestore(&q->requeue_lock, flags); |
| 1484 | } |
| 1485 | |
| 1486 | void blk_mq_kick_requeue_list(struct request_queue *q) |
| 1487 | { |
| 1488 | kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0); |
| 1489 | } |
| 1490 | EXPORT_SYMBOL(blk_mq_kick_requeue_list); |
| 1491 | |
| 1492 | void blk_mq_delay_kick_requeue_list(struct request_queue *q, |
| 1493 | unsigned long msecs) |
| 1494 | { |
| 1495 | kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, |
| 1496 | msecs_to_jiffies(msecs)); |
| 1497 | } |
| 1498 | EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list); |
| 1499 | |
| 1500 | static bool blk_mq_rq_inflight(struct request *rq, void *priv) |
| 1501 | { |
| 1502 | /* |
| 1503 | * If we find a request that isn't idle we know the queue is busy |
| 1504 | * as it's checked in the iter. |
| 1505 | * Return false to stop the iteration. |
| 1506 | */ |
| 1507 | if (blk_mq_request_started(rq)) { |
| 1508 | bool *busy = priv; |
| 1509 | |
| 1510 | *busy = true; |
| 1511 | return false; |
| 1512 | } |
| 1513 | |
| 1514 | return true; |
| 1515 | } |
| 1516 | |
| 1517 | bool blk_mq_queue_inflight(struct request_queue *q) |
| 1518 | { |
| 1519 | bool busy = false; |
| 1520 | |
| 1521 | blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy); |
| 1522 | return busy; |
| 1523 | } |
| 1524 | EXPORT_SYMBOL_GPL(blk_mq_queue_inflight); |
| 1525 | |
| 1526 | static void blk_mq_rq_timed_out(struct request *req) |
| 1527 | { |
| 1528 | req->rq_flags |= RQF_TIMED_OUT; |
| 1529 | if (req->q->mq_ops->timeout) { |
| 1530 | enum blk_eh_timer_return ret; |
| 1531 | |
| 1532 | ret = req->q->mq_ops->timeout(req); |
| 1533 | if (ret == BLK_EH_DONE) |
| 1534 | return; |
| 1535 | WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER); |
| 1536 | } |
| 1537 | |
| 1538 | blk_add_timer(req); |
| 1539 | } |
| 1540 | |
| 1541 | struct blk_expired_data { |
| 1542 | bool has_timedout_rq; |
| 1543 | unsigned long next; |
| 1544 | unsigned long timeout_start; |
| 1545 | }; |
| 1546 | |
| 1547 | static bool blk_mq_req_expired(struct request *rq, struct blk_expired_data *expired) |
| 1548 | { |
| 1549 | unsigned long deadline; |
| 1550 | |
| 1551 | if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT) |
| 1552 | return false; |
| 1553 | if (rq->rq_flags & RQF_TIMED_OUT) |
| 1554 | return false; |
| 1555 | |
| 1556 | deadline = READ_ONCE(rq->deadline); |
| 1557 | if (time_after_eq(expired->timeout_start, deadline)) |
| 1558 | return true; |
| 1559 | |
| 1560 | if (expired->next == 0) |
| 1561 | expired->next = deadline; |
| 1562 | else if (time_after(expired->next, deadline)) |
| 1563 | expired->next = deadline; |
| 1564 | return false; |
| 1565 | } |
| 1566 | |
| 1567 | void blk_mq_put_rq_ref(struct request *rq) |
| 1568 | { |
| 1569 | if (is_flush_rq(rq)) { |
| 1570 | if (rq->end_io(rq, 0) == RQ_END_IO_FREE) |
| 1571 | blk_mq_free_request(rq); |
| 1572 | } else if (req_ref_put_and_test(rq)) { |
| 1573 | __blk_mq_free_request(rq); |
| 1574 | } |
| 1575 | } |
| 1576 | |
| 1577 | static bool blk_mq_check_expired(struct request *rq, void *priv) |
| 1578 | { |
| 1579 | struct blk_expired_data *expired = priv; |
| 1580 | |
| 1581 | /* |
| 1582 | * blk_mq_queue_tag_busy_iter() has locked the request, so it cannot |
| 1583 | * be reallocated underneath the timeout handler's processing, then |
| 1584 | * the expire check is reliable. If the request is not expired, then |
| 1585 | * it was completed and reallocated as a new request after returning |
| 1586 | * from blk_mq_check_expired(). |
| 1587 | */ |
| 1588 | if (blk_mq_req_expired(rq, expired)) { |
| 1589 | expired->has_timedout_rq = true; |
| 1590 | return false; |
| 1591 | } |
| 1592 | return true; |
| 1593 | } |
| 1594 | |
| 1595 | static bool blk_mq_handle_expired(struct request *rq, void *priv) |
| 1596 | { |
| 1597 | struct blk_expired_data *expired = priv; |
| 1598 | |
| 1599 | if (blk_mq_req_expired(rq, expired)) |
| 1600 | blk_mq_rq_timed_out(rq); |
| 1601 | return true; |
| 1602 | } |
| 1603 | |
| 1604 | static void blk_mq_timeout_work(struct work_struct *work) |
| 1605 | { |
| 1606 | struct request_queue *q = |
| 1607 | container_of(work, struct request_queue, timeout_work); |
| 1608 | struct blk_expired_data expired = { |
| 1609 | .timeout_start = jiffies, |
| 1610 | }; |
| 1611 | struct blk_mq_hw_ctx *hctx; |
| 1612 | unsigned long i; |
| 1613 | |
| 1614 | /* A deadlock might occur if a request is stuck requiring a |
| 1615 | * timeout at the same time a queue freeze is waiting |
| 1616 | * completion, since the timeout code would not be able to |
| 1617 | * acquire the queue reference here. |
| 1618 | * |
| 1619 | * That's why we don't use blk_queue_enter here; instead, we use |
| 1620 | * percpu_ref_tryget directly, because we need to be able to |
| 1621 | * obtain a reference even in the short window between the queue |
| 1622 | * starting to freeze, by dropping the first reference in |
| 1623 | * blk_freeze_queue_start, and the moment the last request is |
| 1624 | * consumed, marked by the instant q_usage_counter reaches |
| 1625 | * zero. |
| 1626 | */ |
| 1627 | if (!percpu_ref_tryget(&q->q_usage_counter)) |
| 1628 | return; |
| 1629 | |
| 1630 | /* check if there is any timed-out request */ |
| 1631 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &expired); |
| 1632 | if (expired.has_timedout_rq) { |
| 1633 | /* |
| 1634 | * Before walking tags, we must ensure any submit started |
| 1635 | * before the current time has finished. Since the submit |
| 1636 | * uses srcu or rcu, wait for a synchronization point to |
| 1637 | * ensure all running submits have finished |
| 1638 | */ |
| 1639 | blk_mq_wait_quiesce_done(q->tag_set); |
| 1640 | |
| 1641 | expired.next = 0; |
| 1642 | blk_mq_queue_tag_busy_iter(q, blk_mq_handle_expired, &expired); |
| 1643 | } |
| 1644 | |
| 1645 | if (expired.next != 0) { |
| 1646 | mod_timer(&q->timeout, expired.next); |
| 1647 | } else { |
| 1648 | /* |
| 1649 | * Request timeouts are handled as a forward rolling timer. If |
| 1650 | * we end up here it means that no requests are pending and |
| 1651 | * also that no request has been pending for a while. Mark |
| 1652 | * each hctx as idle. |
| 1653 | */ |
| 1654 | queue_for_each_hw_ctx(q, hctx, i) { |
| 1655 | /* the hctx may be unmapped, so check it here */ |
| 1656 | if (blk_mq_hw_queue_mapped(hctx)) |
| 1657 | blk_mq_tag_idle(hctx); |
| 1658 | } |
| 1659 | } |
| 1660 | blk_queue_exit(q); |
| 1661 | } |
| 1662 | |
| 1663 | struct flush_busy_ctx_data { |
| 1664 | struct blk_mq_hw_ctx *hctx; |
| 1665 | struct list_head *list; |
| 1666 | }; |
| 1667 | |
| 1668 | static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data) |
| 1669 | { |
| 1670 | struct flush_busy_ctx_data *flush_data = data; |
| 1671 | struct blk_mq_hw_ctx *hctx = flush_data->hctx; |
| 1672 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; |
| 1673 | enum hctx_type type = hctx->type; |
| 1674 | |
| 1675 | spin_lock(&ctx->lock); |
| 1676 | list_splice_tail_init(&ctx->rq_lists[type], flush_data->list); |
| 1677 | sbitmap_clear_bit(sb, bitnr); |
| 1678 | spin_unlock(&ctx->lock); |
| 1679 | return true; |
| 1680 | } |
| 1681 | |
| 1682 | /* |
| 1683 | * Process software queues that have been marked busy, splicing them |
| 1684 | * to the for-dispatch |
| 1685 | */ |
| 1686 | void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list) |
| 1687 | { |
| 1688 | struct flush_busy_ctx_data data = { |
| 1689 | .hctx = hctx, |
| 1690 | .list = list, |
| 1691 | }; |
| 1692 | |
| 1693 | sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data); |
| 1694 | } |
| 1695 | EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs); |
| 1696 | |
| 1697 | struct dispatch_rq_data { |
| 1698 | struct blk_mq_hw_ctx *hctx; |
| 1699 | struct request *rq; |
| 1700 | }; |
| 1701 | |
| 1702 | static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr, |
| 1703 | void *data) |
| 1704 | { |
| 1705 | struct dispatch_rq_data *dispatch_data = data; |
| 1706 | struct blk_mq_hw_ctx *hctx = dispatch_data->hctx; |
| 1707 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; |
| 1708 | enum hctx_type type = hctx->type; |
| 1709 | |
| 1710 | spin_lock(&ctx->lock); |
| 1711 | if (!list_empty(&ctx->rq_lists[type])) { |
| 1712 | dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next); |
| 1713 | list_del_init(&dispatch_data->rq->queuelist); |
| 1714 | if (list_empty(&ctx->rq_lists[type])) |
| 1715 | sbitmap_clear_bit(sb, bitnr); |
| 1716 | } |
| 1717 | spin_unlock(&ctx->lock); |
| 1718 | |
| 1719 | return !dispatch_data->rq; |
| 1720 | } |
| 1721 | |
| 1722 | struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx, |
| 1723 | struct blk_mq_ctx *start) |
| 1724 | { |
| 1725 | unsigned off = start ? start->index_hw[hctx->type] : 0; |
| 1726 | struct dispatch_rq_data data = { |
| 1727 | .hctx = hctx, |
| 1728 | .rq = NULL, |
| 1729 | }; |
| 1730 | |
| 1731 | __sbitmap_for_each_set(&hctx->ctx_map, off, |
| 1732 | dispatch_rq_from_ctx, &data); |
| 1733 | |
| 1734 | return data.rq; |
| 1735 | } |
| 1736 | |
| 1737 | static bool __blk_mq_alloc_driver_tag(struct request *rq) |
| 1738 | { |
| 1739 | struct sbitmap_queue *bt = &rq->mq_hctx->tags->bitmap_tags; |
| 1740 | unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags; |
| 1741 | int tag; |
| 1742 | |
| 1743 | blk_mq_tag_busy(rq->mq_hctx); |
| 1744 | |
| 1745 | if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) { |
| 1746 | bt = &rq->mq_hctx->tags->breserved_tags; |
| 1747 | tag_offset = 0; |
| 1748 | } else { |
| 1749 | if (!hctx_may_queue(rq->mq_hctx, bt)) |
| 1750 | return false; |
| 1751 | } |
| 1752 | |
| 1753 | tag = __sbitmap_queue_get(bt); |
| 1754 | if (tag == BLK_MQ_NO_TAG) |
| 1755 | return false; |
| 1756 | |
| 1757 | rq->tag = tag + tag_offset; |
| 1758 | return true; |
| 1759 | } |
| 1760 | |
| 1761 | bool __blk_mq_get_driver_tag(struct blk_mq_hw_ctx *hctx, struct request *rq) |
| 1762 | { |
| 1763 | if (rq->tag == BLK_MQ_NO_TAG && !__blk_mq_alloc_driver_tag(rq)) |
| 1764 | return false; |
| 1765 | |
| 1766 | if ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) && |
| 1767 | !(rq->rq_flags & RQF_MQ_INFLIGHT)) { |
| 1768 | rq->rq_flags |= RQF_MQ_INFLIGHT; |
| 1769 | __blk_mq_inc_active_requests(hctx); |
| 1770 | } |
| 1771 | hctx->tags->rqs[rq->tag] = rq; |
| 1772 | return true; |
| 1773 | } |
| 1774 | |
| 1775 | static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode, |
| 1776 | int flags, void *key) |
| 1777 | { |
| 1778 | struct blk_mq_hw_ctx *hctx; |
| 1779 | |
| 1780 | hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait); |
| 1781 | |
| 1782 | spin_lock(&hctx->dispatch_wait_lock); |
| 1783 | if (!list_empty(&wait->entry)) { |
| 1784 | struct sbitmap_queue *sbq; |
| 1785 | |
| 1786 | list_del_init(&wait->entry); |
| 1787 | sbq = &hctx->tags->bitmap_tags; |
| 1788 | atomic_dec(&sbq->ws_active); |
| 1789 | } |
| 1790 | spin_unlock(&hctx->dispatch_wait_lock); |
| 1791 | |
| 1792 | blk_mq_run_hw_queue(hctx, true); |
| 1793 | return 1; |
| 1794 | } |
| 1795 | |
| 1796 | /* |
| 1797 | * Mark us waiting for a tag. For shared tags, this involves hooking us into |
| 1798 | * the tag wakeups. For non-shared tags, we can simply mark us needing a |
| 1799 | * restart. For both cases, take care to check the condition again after |
| 1800 | * marking us as waiting. |
| 1801 | */ |
| 1802 | static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx, |
| 1803 | struct request *rq) |
| 1804 | { |
| 1805 | struct sbitmap_queue *sbq; |
| 1806 | struct wait_queue_head *wq; |
| 1807 | wait_queue_entry_t *wait; |
| 1808 | bool ret; |
| 1809 | |
| 1810 | if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) && |
| 1811 | !(blk_mq_is_shared_tags(hctx->flags))) { |
| 1812 | blk_mq_sched_mark_restart_hctx(hctx); |
| 1813 | |
| 1814 | /* |
| 1815 | * It's possible that a tag was freed in the window between the |
| 1816 | * allocation failure and adding the hardware queue to the wait |
| 1817 | * queue. |
| 1818 | * |
| 1819 | * Don't clear RESTART here, someone else could have set it. |
| 1820 | * At most this will cost an extra queue run. |
| 1821 | */ |
| 1822 | return blk_mq_get_driver_tag(rq); |
| 1823 | } |
| 1824 | |
| 1825 | wait = &hctx->dispatch_wait; |
| 1826 | if (!list_empty_careful(&wait->entry)) |
| 1827 | return false; |
| 1828 | |
| 1829 | if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) |
| 1830 | sbq = &hctx->tags->breserved_tags; |
| 1831 | else |
| 1832 | sbq = &hctx->tags->bitmap_tags; |
| 1833 | wq = &bt_wait_ptr(sbq, hctx)->wait; |
| 1834 | |
| 1835 | spin_lock_irq(&wq->lock); |
| 1836 | spin_lock(&hctx->dispatch_wait_lock); |
| 1837 | if (!list_empty(&wait->entry)) { |
| 1838 | spin_unlock(&hctx->dispatch_wait_lock); |
| 1839 | spin_unlock_irq(&wq->lock); |
| 1840 | return false; |
| 1841 | } |
| 1842 | |
| 1843 | atomic_inc(&sbq->ws_active); |
| 1844 | wait->flags &= ~WQ_FLAG_EXCLUSIVE; |
| 1845 | __add_wait_queue(wq, wait); |
| 1846 | |
| 1847 | /* |
| 1848 | * It's possible that a tag was freed in the window between the |
| 1849 | * allocation failure and adding the hardware queue to the wait |
| 1850 | * queue. |
| 1851 | */ |
| 1852 | ret = blk_mq_get_driver_tag(rq); |
| 1853 | if (!ret) { |
| 1854 | spin_unlock(&hctx->dispatch_wait_lock); |
| 1855 | spin_unlock_irq(&wq->lock); |
| 1856 | return false; |
| 1857 | } |
| 1858 | |
| 1859 | /* |
| 1860 | * We got a tag, remove ourselves from the wait queue to ensure |
| 1861 | * someone else gets the wakeup. |
| 1862 | */ |
| 1863 | list_del_init(&wait->entry); |
| 1864 | atomic_dec(&sbq->ws_active); |
| 1865 | spin_unlock(&hctx->dispatch_wait_lock); |
| 1866 | spin_unlock_irq(&wq->lock); |
| 1867 | |
| 1868 | return true; |
| 1869 | } |
| 1870 | |
| 1871 | #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8 |
| 1872 | #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4 |
| 1873 | /* |
| 1874 | * Update dispatch busy with the Exponential Weighted Moving Average(EWMA): |
| 1875 | * - EWMA is one simple way to compute running average value |
| 1876 | * - weight(7/8 and 1/8) is applied so that it can decrease exponentially |
| 1877 | * - take 4 as factor for avoiding to get too small(0) result, and this |
| 1878 | * factor doesn't matter because EWMA decreases exponentially |
| 1879 | */ |
| 1880 | static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy) |
| 1881 | { |
| 1882 | unsigned int ewma; |
| 1883 | |
| 1884 | ewma = hctx->dispatch_busy; |
| 1885 | |
| 1886 | if (!ewma && !busy) |
| 1887 | return; |
| 1888 | |
| 1889 | ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1; |
| 1890 | if (busy) |
| 1891 | ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR; |
| 1892 | ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT; |
| 1893 | |
| 1894 | hctx->dispatch_busy = ewma; |
| 1895 | } |
| 1896 | |
| 1897 | #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */ |
| 1898 | |
| 1899 | static void blk_mq_handle_dev_resource(struct request *rq, |
| 1900 | struct list_head *list) |
| 1901 | { |
| 1902 | list_add(&rq->queuelist, list); |
| 1903 | __blk_mq_requeue_request(rq); |
| 1904 | } |
| 1905 | |
| 1906 | static void blk_mq_handle_zone_resource(struct request *rq, |
| 1907 | struct list_head *zone_list) |
| 1908 | { |
| 1909 | /* |
| 1910 | * If we end up here it is because we cannot dispatch a request to a |
| 1911 | * specific zone due to LLD level zone-write locking or other zone |
| 1912 | * related resource not being available. In this case, set the request |
| 1913 | * aside in zone_list for retrying it later. |
| 1914 | */ |
| 1915 | list_add(&rq->queuelist, zone_list); |
| 1916 | __blk_mq_requeue_request(rq); |
| 1917 | } |
| 1918 | |
| 1919 | enum prep_dispatch { |
| 1920 | PREP_DISPATCH_OK, |
| 1921 | PREP_DISPATCH_NO_TAG, |
| 1922 | PREP_DISPATCH_NO_BUDGET, |
| 1923 | }; |
| 1924 | |
| 1925 | static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq, |
| 1926 | bool need_budget) |
| 1927 | { |
| 1928 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
| 1929 | int budget_token = -1; |
| 1930 | |
| 1931 | if (need_budget) { |
| 1932 | budget_token = blk_mq_get_dispatch_budget(rq->q); |
| 1933 | if (budget_token < 0) { |
| 1934 | blk_mq_put_driver_tag(rq); |
| 1935 | return PREP_DISPATCH_NO_BUDGET; |
| 1936 | } |
| 1937 | blk_mq_set_rq_budget_token(rq, budget_token); |
| 1938 | } |
| 1939 | |
| 1940 | if (!blk_mq_get_driver_tag(rq)) { |
| 1941 | /* |
| 1942 | * The initial allocation attempt failed, so we need to |
| 1943 | * rerun the hardware queue when a tag is freed. The |
| 1944 | * waitqueue takes care of that. If the queue is run |
| 1945 | * before we add this entry back on the dispatch list, |
| 1946 | * we'll re-run it below. |
| 1947 | */ |
| 1948 | if (!blk_mq_mark_tag_wait(hctx, rq)) { |
| 1949 | /* |
| 1950 | * All budgets not got from this function will be put |
| 1951 | * together during handling partial dispatch |
| 1952 | */ |
| 1953 | if (need_budget) |
| 1954 | blk_mq_put_dispatch_budget(rq->q, budget_token); |
| 1955 | return PREP_DISPATCH_NO_TAG; |
| 1956 | } |
| 1957 | } |
| 1958 | |
| 1959 | return PREP_DISPATCH_OK; |
| 1960 | } |
| 1961 | |
| 1962 | /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */ |
| 1963 | static void blk_mq_release_budgets(struct request_queue *q, |
| 1964 | struct list_head *list) |
| 1965 | { |
| 1966 | struct request *rq; |
| 1967 | |
| 1968 | list_for_each_entry(rq, list, queuelist) { |
| 1969 | int budget_token = blk_mq_get_rq_budget_token(rq); |
| 1970 | |
| 1971 | if (budget_token >= 0) |
| 1972 | blk_mq_put_dispatch_budget(q, budget_token); |
| 1973 | } |
| 1974 | } |
| 1975 | |
| 1976 | /* |
| 1977 | * blk_mq_commit_rqs will notify driver using bd->last that there is no |
| 1978 | * more requests. (See comment in struct blk_mq_ops for commit_rqs for |
| 1979 | * details) |
| 1980 | * Attention, we should explicitly call this in unusual cases: |
| 1981 | * 1) did not queue everything initially scheduled to queue |
| 1982 | * 2) the last attempt to queue a request failed |
| 1983 | */ |
| 1984 | static void blk_mq_commit_rqs(struct blk_mq_hw_ctx *hctx, int queued, |
| 1985 | bool from_schedule) |
| 1986 | { |
| 1987 | if (hctx->queue->mq_ops->commit_rqs && queued) { |
| 1988 | trace_block_unplug(hctx->queue, queued, !from_schedule); |
| 1989 | hctx->queue->mq_ops->commit_rqs(hctx); |
| 1990 | } |
| 1991 | } |
| 1992 | |
| 1993 | /* |
| 1994 | * Returns true if we did some work AND can potentially do more. |
| 1995 | */ |
| 1996 | bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list, |
| 1997 | unsigned int nr_budgets) |
| 1998 | { |
| 1999 | enum prep_dispatch prep; |
| 2000 | struct request_queue *q = hctx->queue; |
| 2001 | struct request *rq; |
| 2002 | int queued; |
| 2003 | blk_status_t ret = BLK_STS_OK; |
| 2004 | LIST_HEAD(zone_list); |
| 2005 | bool needs_resource = false; |
| 2006 | |
| 2007 | if (list_empty(list)) |
| 2008 | return false; |
| 2009 | |
| 2010 | /* |
| 2011 | * Now process all the entries, sending them to the driver. |
| 2012 | */ |
| 2013 | queued = 0; |
| 2014 | do { |
| 2015 | struct blk_mq_queue_data bd; |
| 2016 | |
| 2017 | rq = list_first_entry(list, struct request, queuelist); |
| 2018 | |
| 2019 | WARN_ON_ONCE(hctx != rq->mq_hctx); |
| 2020 | prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets); |
| 2021 | if (prep != PREP_DISPATCH_OK) |
| 2022 | break; |
| 2023 | |
| 2024 | list_del_init(&rq->queuelist); |
| 2025 | |
| 2026 | bd.rq = rq; |
| 2027 | bd.last = list_empty(list); |
| 2028 | |
| 2029 | /* |
| 2030 | * once the request is queued to lld, no need to cover the |
| 2031 | * budget any more |
| 2032 | */ |
| 2033 | if (nr_budgets) |
| 2034 | nr_budgets--; |
| 2035 | ret = q->mq_ops->queue_rq(hctx, &bd); |
| 2036 | switch (ret) { |
| 2037 | case BLK_STS_OK: |
| 2038 | queued++; |
| 2039 | break; |
| 2040 | case BLK_STS_RESOURCE: |
| 2041 | needs_resource = true; |
| 2042 | fallthrough; |
| 2043 | case BLK_STS_DEV_RESOURCE: |
| 2044 | blk_mq_handle_dev_resource(rq, list); |
| 2045 | goto out; |
| 2046 | case BLK_STS_ZONE_RESOURCE: |
| 2047 | /* |
| 2048 | * Move the request to zone_list and keep going through |
| 2049 | * the dispatch list to find more requests the drive can |
| 2050 | * accept. |
| 2051 | */ |
| 2052 | blk_mq_handle_zone_resource(rq, &zone_list); |
| 2053 | needs_resource = true; |
| 2054 | break; |
| 2055 | default: |
| 2056 | blk_mq_end_request(rq, ret); |
| 2057 | } |
| 2058 | } while (!list_empty(list)); |
| 2059 | out: |
| 2060 | if (!list_empty(&zone_list)) |
| 2061 | list_splice_tail_init(&zone_list, list); |
| 2062 | |
| 2063 | /* If we didn't flush the entire list, we could have told the driver |
| 2064 | * there was more coming, but that turned out to be a lie. |
| 2065 | */ |
| 2066 | if (!list_empty(list) || ret != BLK_STS_OK) |
| 2067 | blk_mq_commit_rqs(hctx, queued, false); |
| 2068 | |
| 2069 | /* |
| 2070 | * Any items that need requeuing? Stuff them into hctx->dispatch, |
| 2071 | * that is where we will continue on next queue run. |
| 2072 | */ |
| 2073 | if (!list_empty(list)) { |
| 2074 | bool needs_restart; |
| 2075 | /* For non-shared tags, the RESTART check will suffice */ |
| 2076 | bool no_tag = prep == PREP_DISPATCH_NO_TAG && |
| 2077 | ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) || |
| 2078 | blk_mq_is_shared_tags(hctx->flags)); |
| 2079 | |
| 2080 | if (nr_budgets) |
| 2081 | blk_mq_release_budgets(q, list); |
| 2082 | |
| 2083 | spin_lock(&hctx->lock); |
| 2084 | list_splice_tail_init(list, &hctx->dispatch); |
| 2085 | spin_unlock(&hctx->lock); |
| 2086 | |
| 2087 | /* |
| 2088 | * Order adding requests to hctx->dispatch and checking |
| 2089 | * SCHED_RESTART flag. The pair of this smp_mb() is the one |
| 2090 | * in blk_mq_sched_restart(). Avoid restart code path to |
| 2091 | * miss the new added requests to hctx->dispatch, meantime |
| 2092 | * SCHED_RESTART is observed here. |
| 2093 | */ |
| 2094 | smp_mb(); |
| 2095 | |
| 2096 | /* |
| 2097 | * If SCHED_RESTART was set by the caller of this function and |
| 2098 | * it is no longer set that means that it was cleared by another |
| 2099 | * thread and hence that a queue rerun is needed. |
| 2100 | * |
| 2101 | * If 'no_tag' is set, that means that we failed getting |
| 2102 | * a driver tag with an I/O scheduler attached. If our dispatch |
| 2103 | * waitqueue is no longer active, ensure that we run the queue |
| 2104 | * AFTER adding our entries back to the list. |
| 2105 | * |
| 2106 | * If no I/O scheduler has been configured it is possible that |
| 2107 | * the hardware queue got stopped and restarted before requests |
| 2108 | * were pushed back onto the dispatch list. Rerun the queue to |
| 2109 | * avoid starvation. Notes: |
| 2110 | * - blk_mq_run_hw_queue() checks whether or not a queue has |
| 2111 | * been stopped before rerunning a queue. |
| 2112 | * - Some but not all block drivers stop a queue before |
| 2113 | * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq |
| 2114 | * and dm-rq. |
| 2115 | * |
| 2116 | * If driver returns BLK_STS_RESOURCE and SCHED_RESTART |
| 2117 | * bit is set, run queue after a delay to avoid IO stalls |
| 2118 | * that could otherwise occur if the queue is idle. We'll do |
| 2119 | * similar if we couldn't get budget or couldn't lock a zone |
| 2120 | * and SCHED_RESTART is set. |
| 2121 | */ |
| 2122 | needs_restart = blk_mq_sched_needs_restart(hctx); |
| 2123 | if (prep == PREP_DISPATCH_NO_BUDGET) |
| 2124 | needs_resource = true; |
| 2125 | if (!needs_restart || |
| 2126 | (no_tag && list_empty_careful(&hctx->dispatch_wait.entry))) |
| 2127 | blk_mq_run_hw_queue(hctx, true); |
| 2128 | else if (needs_resource) |
| 2129 | blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY); |
| 2130 | |
| 2131 | blk_mq_update_dispatch_busy(hctx, true); |
| 2132 | return false; |
| 2133 | } |
| 2134 | |
| 2135 | blk_mq_update_dispatch_busy(hctx, false); |
| 2136 | return true; |
| 2137 | } |
| 2138 | |
| 2139 | static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx) |
| 2140 | { |
| 2141 | int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask); |
| 2142 | |
| 2143 | if (cpu >= nr_cpu_ids) |
| 2144 | cpu = cpumask_first(hctx->cpumask); |
| 2145 | return cpu; |
| 2146 | } |
| 2147 | |
| 2148 | /* |
| 2149 | * It'd be great if the workqueue API had a way to pass |
| 2150 | * in a mask and had some smarts for more clever placement. |
| 2151 | * For now we just round-robin here, switching for every |
| 2152 | * BLK_MQ_CPU_WORK_BATCH queued items. |
| 2153 | */ |
| 2154 | static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx) |
| 2155 | { |
| 2156 | bool tried = false; |
| 2157 | int next_cpu = hctx->next_cpu; |
| 2158 | |
| 2159 | if (hctx->queue->nr_hw_queues == 1) |
| 2160 | return WORK_CPU_UNBOUND; |
| 2161 | |
| 2162 | if (--hctx->next_cpu_batch <= 0) { |
| 2163 | select_cpu: |
| 2164 | next_cpu = cpumask_next_and(next_cpu, hctx->cpumask, |
| 2165 | cpu_online_mask); |
| 2166 | if (next_cpu >= nr_cpu_ids) |
| 2167 | next_cpu = blk_mq_first_mapped_cpu(hctx); |
| 2168 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; |
| 2169 | } |
| 2170 | |
| 2171 | /* |
| 2172 | * Do unbound schedule if we can't find a online CPU for this hctx, |
| 2173 | * and it should only happen in the path of handling CPU DEAD. |
| 2174 | */ |
| 2175 | if (!cpu_online(next_cpu)) { |
| 2176 | if (!tried) { |
| 2177 | tried = true; |
| 2178 | goto select_cpu; |
| 2179 | } |
| 2180 | |
| 2181 | /* |
| 2182 | * Make sure to re-select CPU next time once after CPUs |
| 2183 | * in hctx->cpumask become online again. |
| 2184 | */ |
| 2185 | hctx->next_cpu = next_cpu; |
| 2186 | hctx->next_cpu_batch = 1; |
| 2187 | return WORK_CPU_UNBOUND; |
| 2188 | } |
| 2189 | |
| 2190 | hctx->next_cpu = next_cpu; |
| 2191 | return next_cpu; |
| 2192 | } |
| 2193 | |
| 2194 | /** |
| 2195 | * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously. |
| 2196 | * @hctx: Pointer to the hardware queue to run. |
| 2197 | * @msecs: Milliseconds of delay to wait before running the queue. |
| 2198 | * |
| 2199 | * Run a hardware queue asynchronously with a delay of @msecs. |
| 2200 | */ |
| 2201 | void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) |
| 2202 | { |
| 2203 | if (unlikely(blk_mq_hctx_stopped(hctx))) |
| 2204 | return; |
| 2205 | kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work, |
| 2206 | msecs_to_jiffies(msecs)); |
| 2207 | } |
| 2208 | EXPORT_SYMBOL(blk_mq_delay_run_hw_queue); |
| 2209 | |
| 2210 | /** |
| 2211 | * blk_mq_run_hw_queue - Start to run a hardware queue. |
| 2212 | * @hctx: Pointer to the hardware queue to run. |
| 2213 | * @async: If we want to run the queue asynchronously. |
| 2214 | * |
| 2215 | * Check if the request queue is not in a quiesced state and if there are |
| 2216 | * pending requests to be sent. If this is true, run the queue to send requests |
| 2217 | * to hardware. |
| 2218 | */ |
| 2219 | void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
| 2220 | { |
| 2221 | bool need_run; |
| 2222 | |
| 2223 | /* |
| 2224 | * We can't run the queue inline with interrupts disabled. |
| 2225 | */ |
| 2226 | WARN_ON_ONCE(!async && in_interrupt()); |
| 2227 | |
| 2228 | /* |
| 2229 | * When queue is quiesced, we may be switching io scheduler, or |
| 2230 | * updating nr_hw_queues, or other things, and we can't run queue |
| 2231 | * any more, even __blk_mq_hctx_has_pending() can't be called safely. |
| 2232 | * |
| 2233 | * And queue will be rerun in blk_mq_unquiesce_queue() if it is |
| 2234 | * quiesced. |
| 2235 | */ |
| 2236 | __blk_mq_run_dispatch_ops(hctx->queue, false, |
| 2237 | need_run = !blk_queue_quiesced(hctx->queue) && |
| 2238 | blk_mq_hctx_has_pending(hctx)); |
| 2239 | |
| 2240 | if (!need_run) |
| 2241 | return; |
| 2242 | |
| 2243 | if (async || (hctx->flags & BLK_MQ_F_BLOCKING) || |
| 2244 | !cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask)) { |
| 2245 | blk_mq_delay_run_hw_queue(hctx, 0); |
| 2246 | return; |
| 2247 | } |
| 2248 | |
| 2249 | blk_mq_run_dispatch_ops(hctx->queue, |
| 2250 | blk_mq_sched_dispatch_requests(hctx)); |
| 2251 | } |
| 2252 | EXPORT_SYMBOL(blk_mq_run_hw_queue); |
| 2253 | |
| 2254 | /* |
| 2255 | * Return prefered queue to dispatch from (if any) for non-mq aware IO |
| 2256 | * scheduler. |
| 2257 | */ |
| 2258 | static struct blk_mq_hw_ctx *blk_mq_get_sq_hctx(struct request_queue *q) |
| 2259 | { |
| 2260 | struct blk_mq_ctx *ctx = blk_mq_get_ctx(q); |
| 2261 | /* |
| 2262 | * If the IO scheduler does not respect hardware queues when |
| 2263 | * dispatching, we just don't bother with multiple HW queues and |
| 2264 | * dispatch from hctx for the current CPU since running multiple queues |
| 2265 | * just causes lock contention inside the scheduler and pointless cache |
| 2266 | * bouncing. |
| 2267 | */ |
| 2268 | struct blk_mq_hw_ctx *hctx = ctx->hctxs[HCTX_TYPE_DEFAULT]; |
| 2269 | |
| 2270 | if (!blk_mq_hctx_stopped(hctx)) |
| 2271 | return hctx; |
| 2272 | return NULL; |
| 2273 | } |
| 2274 | |
| 2275 | /** |
| 2276 | * blk_mq_run_hw_queues - Run all hardware queues in a request queue. |
| 2277 | * @q: Pointer to the request queue to run. |
| 2278 | * @async: If we want to run the queue asynchronously. |
| 2279 | */ |
| 2280 | void blk_mq_run_hw_queues(struct request_queue *q, bool async) |
| 2281 | { |
| 2282 | struct blk_mq_hw_ctx *hctx, *sq_hctx; |
| 2283 | unsigned long i; |
| 2284 | |
| 2285 | sq_hctx = NULL; |
| 2286 | if (blk_queue_sq_sched(q)) |
| 2287 | sq_hctx = blk_mq_get_sq_hctx(q); |
| 2288 | queue_for_each_hw_ctx(q, hctx, i) { |
| 2289 | if (blk_mq_hctx_stopped(hctx)) |
| 2290 | continue; |
| 2291 | /* |
| 2292 | * Dispatch from this hctx either if there's no hctx preferred |
| 2293 | * by IO scheduler or if it has requests that bypass the |
| 2294 | * scheduler. |
| 2295 | */ |
| 2296 | if (!sq_hctx || sq_hctx == hctx || |
| 2297 | !list_empty_careful(&hctx->dispatch)) |
| 2298 | blk_mq_run_hw_queue(hctx, async); |
| 2299 | } |
| 2300 | } |
| 2301 | EXPORT_SYMBOL(blk_mq_run_hw_queues); |
| 2302 | |
| 2303 | /** |
| 2304 | * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously. |
| 2305 | * @q: Pointer to the request queue to run. |
| 2306 | * @msecs: Milliseconds of delay to wait before running the queues. |
| 2307 | */ |
| 2308 | void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs) |
| 2309 | { |
| 2310 | struct blk_mq_hw_ctx *hctx, *sq_hctx; |
| 2311 | unsigned long i; |
| 2312 | |
| 2313 | sq_hctx = NULL; |
| 2314 | if (blk_queue_sq_sched(q)) |
| 2315 | sq_hctx = blk_mq_get_sq_hctx(q); |
| 2316 | queue_for_each_hw_ctx(q, hctx, i) { |
| 2317 | if (blk_mq_hctx_stopped(hctx)) |
| 2318 | continue; |
| 2319 | /* |
| 2320 | * If there is already a run_work pending, leave the |
| 2321 | * pending delay untouched. Otherwise, a hctx can stall |
| 2322 | * if another hctx is re-delaying the other's work |
| 2323 | * before the work executes. |
| 2324 | */ |
| 2325 | if (delayed_work_pending(&hctx->run_work)) |
| 2326 | continue; |
| 2327 | /* |
| 2328 | * Dispatch from this hctx either if there's no hctx preferred |
| 2329 | * by IO scheduler or if it has requests that bypass the |
| 2330 | * scheduler. |
| 2331 | */ |
| 2332 | if (!sq_hctx || sq_hctx == hctx || |
| 2333 | !list_empty_careful(&hctx->dispatch)) |
| 2334 | blk_mq_delay_run_hw_queue(hctx, msecs); |
| 2335 | } |
| 2336 | } |
| 2337 | EXPORT_SYMBOL(blk_mq_delay_run_hw_queues); |
| 2338 | |
| 2339 | /* |
| 2340 | * This function is often used for pausing .queue_rq() by driver when |
| 2341 | * there isn't enough resource or some conditions aren't satisfied, and |
| 2342 | * BLK_STS_RESOURCE is usually returned. |
| 2343 | * |
| 2344 | * We do not guarantee that dispatch can be drained or blocked |
| 2345 | * after blk_mq_stop_hw_queue() returns. Please use |
| 2346 | * blk_mq_quiesce_queue() for that requirement. |
| 2347 | */ |
| 2348 | void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) |
| 2349 | { |
| 2350 | cancel_delayed_work(&hctx->run_work); |
| 2351 | |
| 2352 | set_bit(BLK_MQ_S_STOPPED, &hctx->state); |
| 2353 | } |
| 2354 | EXPORT_SYMBOL(blk_mq_stop_hw_queue); |
| 2355 | |
| 2356 | /* |
| 2357 | * This function is often used for pausing .queue_rq() by driver when |
| 2358 | * there isn't enough resource or some conditions aren't satisfied, and |
| 2359 | * BLK_STS_RESOURCE is usually returned. |
| 2360 | * |
| 2361 | * We do not guarantee that dispatch can be drained or blocked |
| 2362 | * after blk_mq_stop_hw_queues() returns. Please use |
| 2363 | * blk_mq_quiesce_queue() for that requirement. |
| 2364 | */ |
| 2365 | void blk_mq_stop_hw_queues(struct request_queue *q) |
| 2366 | { |
| 2367 | struct blk_mq_hw_ctx *hctx; |
| 2368 | unsigned long i; |
| 2369 | |
| 2370 | queue_for_each_hw_ctx(q, hctx, i) |
| 2371 | blk_mq_stop_hw_queue(hctx); |
| 2372 | } |
| 2373 | EXPORT_SYMBOL(blk_mq_stop_hw_queues); |
| 2374 | |
| 2375 | void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) |
| 2376 | { |
| 2377 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); |
| 2378 | |
| 2379 | blk_mq_run_hw_queue(hctx, false); |
| 2380 | } |
| 2381 | EXPORT_SYMBOL(blk_mq_start_hw_queue); |
| 2382 | |
| 2383 | void blk_mq_start_hw_queues(struct request_queue *q) |
| 2384 | { |
| 2385 | struct blk_mq_hw_ctx *hctx; |
| 2386 | unsigned long i; |
| 2387 | |
| 2388 | queue_for_each_hw_ctx(q, hctx, i) |
| 2389 | blk_mq_start_hw_queue(hctx); |
| 2390 | } |
| 2391 | EXPORT_SYMBOL(blk_mq_start_hw_queues); |
| 2392 | |
| 2393 | void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
| 2394 | { |
| 2395 | if (!blk_mq_hctx_stopped(hctx)) |
| 2396 | return; |
| 2397 | |
| 2398 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); |
| 2399 | blk_mq_run_hw_queue(hctx, async); |
| 2400 | } |
| 2401 | EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue); |
| 2402 | |
| 2403 | void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async) |
| 2404 | { |
| 2405 | struct blk_mq_hw_ctx *hctx; |
| 2406 | unsigned long i; |
| 2407 | |
| 2408 | queue_for_each_hw_ctx(q, hctx, i) |
| 2409 | blk_mq_start_stopped_hw_queue(hctx, async); |
| 2410 | } |
| 2411 | EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); |
| 2412 | |
| 2413 | static void blk_mq_run_work_fn(struct work_struct *work) |
| 2414 | { |
| 2415 | struct blk_mq_hw_ctx *hctx = |
| 2416 | container_of(work, struct blk_mq_hw_ctx, run_work.work); |
| 2417 | |
| 2418 | blk_mq_run_dispatch_ops(hctx->queue, |
| 2419 | blk_mq_sched_dispatch_requests(hctx)); |
| 2420 | } |
| 2421 | |
| 2422 | /** |
| 2423 | * blk_mq_request_bypass_insert - Insert a request at dispatch list. |
| 2424 | * @rq: Pointer to request to be inserted. |
| 2425 | * @flags: BLK_MQ_INSERT_* |
| 2426 | * |
| 2427 | * Should only be used carefully, when the caller knows we want to |
| 2428 | * bypass a potential IO scheduler on the target device. |
| 2429 | */ |
| 2430 | void blk_mq_request_bypass_insert(struct request *rq, blk_insert_t flags) |
| 2431 | { |
| 2432 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
| 2433 | |
| 2434 | spin_lock(&hctx->lock); |
| 2435 | if (flags & BLK_MQ_INSERT_AT_HEAD) |
| 2436 | list_add(&rq->queuelist, &hctx->dispatch); |
| 2437 | else |
| 2438 | list_add_tail(&rq->queuelist, &hctx->dispatch); |
| 2439 | spin_unlock(&hctx->lock); |
| 2440 | } |
| 2441 | |
| 2442 | static void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, |
| 2443 | struct blk_mq_ctx *ctx, struct list_head *list, |
| 2444 | bool run_queue_async) |
| 2445 | { |
| 2446 | struct request *rq; |
| 2447 | enum hctx_type type = hctx->type; |
| 2448 | |
| 2449 | /* |
| 2450 | * Try to issue requests directly if the hw queue isn't busy to save an |
| 2451 | * extra enqueue & dequeue to the sw queue. |
| 2452 | */ |
| 2453 | if (!hctx->dispatch_busy && !run_queue_async) { |
| 2454 | blk_mq_run_dispatch_ops(hctx->queue, |
| 2455 | blk_mq_try_issue_list_directly(hctx, list)); |
| 2456 | if (list_empty(list)) |
| 2457 | goto out; |
| 2458 | } |
| 2459 | |
| 2460 | /* |
| 2461 | * preemption doesn't flush plug list, so it's possible ctx->cpu is |
| 2462 | * offline now |
| 2463 | */ |
| 2464 | list_for_each_entry(rq, list, queuelist) { |
| 2465 | BUG_ON(rq->mq_ctx != ctx); |
| 2466 | trace_block_rq_insert(rq); |
| 2467 | } |
| 2468 | |
| 2469 | spin_lock(&ctx->lock); |
| 2470 | list_splice_tail_init(list, &ctx->rq_lists[type]); |
| 2471 | blk_mq_hctx_mark_pending(hctx, ctx); |
| 2472 | spin_unlock(&ctx->lock); |
| 2473 | out: |
| 2474 | blk_mq_run_hw_queue(hctx, run_queue_async); |
| 2475 | } |
| 2476 | |
| 2477 | static void blk_mq_insert_request(struct request *rq, blk_insert_t flags) |
| 2478 | { |
| 2479 | struct request_queue *q = rq->q; |
| 2480 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
| 2481 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
| 2482 | |
| 2483 | if (blk_rq_is_passthrough(rq)) { |
| 2484 | /* |
| 2485 | * Passthrough request have to be added to hctx->dispatch |
| 2486 | * directly. The device may be in a situation where it can't |
| 2487 | * handle FS request, and always returns BLK_STS_RESOURCE for |
| 2488 | * them, which gets them added to hctx->dispatch. |
| 2489 | * |
| 2490 | * If a passthrough request is required to unblock the queues, |
| 2491 | * and it is added to the scheduler queue, there is no chance to |
| 2492 | * dispatch it given we prioritize requests in hctx->dispatch. |
| 2493 | */ |
| 2494 | blk_mq_request_bypass_insert(rq, flags); |
| 2495 | } else if (rq->rq_flags & RQF_FLUSH_SEQ) { |
| 2496 | /* |
| 2497 | * Firstly normal IO request is inserted to scheduler queue or |
| 2498 | * sw queue, meantime we add flush request to dispatch queue( |
| 2499 | * hctx->dispatch) directly and there is at most one in-flight |
| 2500 | * flush request for each hw queue, so it doesn't matter to add |
| 2501 | * flush request to tail or front of the dispatch queue. |
| 2502 | * |
| 2503 | * Secondly in case of NCQ, flush request belongs to non-NCQ |
| 2504 | * command, and queueing it will fail when there is any |
| 2505 | * in-flight normal IO request(NCQ command). When adding flush |
| 2506 | * rq to the front of hctx->dispatch, it is easier to introduce |
| 2507 | * extra time to flush rq's latency because of S_SCHED_RESTART |
| 2508 | * compared with adding to the tail of dispatch queue, then |
| 2509 | * chance of flush merge is increased, and less flush requests |
| 2510 | * will be issued to controller. It is observed that ~10% time |
| 2511 | * is saved in blktests block/004 on disk attached to AHCI/NCQ |
| 2512 | * drive when adding flush rq to the front of hctx->dispatch. |
| 2513 | * |
| 2514 | * Simply queue flush rq to the front of hctx->dispatch so that |
| 2515 | * intensive flush workloads can benefit in case of NCQ HW. |
| 2516 | */ |
| 2517 | blk_mq_request_bypass_insert(rq, BLK_MQ_INSERT_AT_HEAD); |
| 2518 | } else if (q->elevator) { |
| 2519 | LIST_HEAD(list); |
| 2520 | |
| 2521 | WARN_ON_ONCE(rq->tag != BLK_MQ_NO_TAG); |
| 2522 | |
| 2523 | list_add(&rq->queuelist, &list); |
| 2524 | q->elevator->type->ops.insert_requests(hctx, &list, flags); |
| 2525 | } else { |
| 2526 | trace_block_rq_insert(rq); |
| 2527 | |
| 2528 | spin_lock(&ctx->lock); |
| 2529 | if (flags & BLK_MQ_INSERT_AT_HEAD) |
| 2530 | list_add(&rq->queuelist, &ctx->rq_lists[hctx->type]); |
| 2531 | else |
| 2532 | list_add_tail(&rq->queuelist, |
| 2533 | &ctx->rq_lists[hctx->type]); |
| 2534 | blk_mq_hctx_mark_pending(hctx, ctx); |
| 2535 | spin_unlock(&ctx->lock); |
| 2536 | } |
| 2537 | } |
| 2538 | |
| 2539 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio, |
| 2540 | unsigned int nr_segs) |
| 2541 | { |
| 2542 | int err; |
| 2543 | |
| 2544 | if (bio->bi_opf & REQ_RAHEAD) |
| 2545 | rq->cmd_flags |= REQ_FAILFAST_MASK; |
| 2546 | |
| 2547 | rq->__sector = bio->bi_iter.bi_sector; |
| 2548 | blk_rq_bio_prep(rq, bio, nr_segs); |
| 2549 | |
| 2550 | /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */ |
| 2551 | err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO); |
| 2552 | WARN_ON_ONCE(err); |
| 2553 | |
| 2554 | blk_account_io_start(rq); |
| 2555 | } |
| 2556 | |
| 2557 | static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx, |
| 2558 | struct request *rq, bool last) |
| 2559 | { |
| 2560 | struct request_queue *q = rq->q; |
| 2561 | struct blk_mq_queue_data bd = { |
| 2562 | .rq = rq, |
| 2563 | .last = last, |
| 2564 | }; |
| 2565 | blk_status_t ret; |
| 2566 | |
| 2567 | /* |
| 2568 | * For OK queue, we are done. For error, caller may kill it. |
| 2569 | * Any other error (busy), just add it to our list as we |
| 2570 | * previously would have done. |
| 2571 | */ |
| 2572 | ret = q->mq_ops->queue_rq(hctx, &bd); |
| 2573 | switch (ret) { |
| 2574 | case BLK_STS_OK: |
| 2575 | blk_mq_update_dispatch_busy(hctx, false); |
| 2576 | break; |
| 2577 | case BLK_STS_RESOURCE: |
| 2578 | case BLK_STS_DEV_RESOURCE: |
| 2579 | blk_mq_update_dispatch_busy(hctx, true); |
| 2580 | __blk_mq_requeue_request(rq); |
| 2581 | break; |
| 2582 | default: |
| 2583 | blk_mq_update_dispatch_busy(hctx, false); |
| 2584 | break; |
| 2585 | } |
| 2586 | |
| 2587 | return ret; |
| 2588 | } |
| 2589 | |
| 2590 | static bool blk_mq_get_budget_and_tag(struct request *rq) |
| 2591 | { |
| 2592 | int budget_token; |
| 2593 | |
| 2594 | budget_token = blk_mq_get_dispatch_budget(rq->q); |
| 2595 | if (budget_token < 0) |
| 2596 | return false; |
| 2597 | blk_mq_set_rq_budget_token(rq, budget_token); |
| 2598 | if (!blk_mq_get_driver_tag(rq)) { |
| 2599 | blk_mq_put_dispatch_budget(rq->q, budget_token); |
| 2600 | return false; |
| 2601 | } |
| 2602 | return true; |
| 2603 | } |
| 2604 | |
| 2605 | /** |
| 2606 | * blk_mq_try_issue_directly - Try to send a request directly to device driver. |
| 2607 | * @hctx: Pointer of the associated hardware queue. |
| 2608 | * @rq: Pointer to request to be sent. |
| 2609 | * |
| 2610 | * If the device has enough resources to accept a new request now, send the |
| 2611 | * request directly to device driver. Else, insert at hctx->dispatch queue, so |
| 2612 | * we can try send it another time in the future. Requests inserted at this |
| 2613 | * queue have higher priority. |
| 2614 | */ |
| 2615 | static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, |
| 2616 | struct request *rq) |
| 2617 | { |
| 2618 | blk_status_t ret; |
| 2619 | |
| 2620 | if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(rq->q)) { |
| 2621 | blk_mq_insert_request(rq, 0); |
| 2622 | return; |
| 2623 | } |
| 2624 | |
| 2625 | if ((rq->rq_flags & RQF_ELV) || !blk_mq_get_budget_and_tag(rq)) { |
| 2626 | blk_mq_insert_request(rq, 0); |
| 2627 | blk_mq_run_hw_queue(hctx, false); |
| 2628 | return; |
| 2629 | } |
| 2630 | |
| 2631 | ret = __blk_mq_issue_directly(hctx, rq, true); |
| 2632 | switch (ret) { |
| 2633 | case BLK_STS_OK: |
| 2634 | break; |
| 2635 | case BLK_STS_RESOURCE: |
| 2636 | case BLK_STS_DEV_RESOURCE: |
| 2637 | blk_mq_request_bypass_insert(rq, 0); |
| 2638 | blk_mq_run_hw_queue(hctx, false); |
| 2639 | break; |
| 2640 | default: |
| 2641 | blk_mq_end_request(rq, ret); |
| 2642 | break; |
| 2643 | } |
| 2644 | } |
| 2645 | |
| 2646 | static blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last) |
| 2647 | { |
| 2648 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
| 2649 | |
| 2650 | if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(rq->q)) { |
| 2651 | blk_mq_insert_request(rq, 0); |
| 2652 | return BLK_STS_OK; |
| 2653 | } |
| 2654 | |
| 2655 | if (!blk_mq_get_budget_and_tag(rq)) |
| 2656 | return BLK_STS_RESOURCE; |
| 2657 | return __blk_mq_issue_directly(hctx, rq, last); |
| 2658 | } |
| 2659 | |
| 2660 | static void blk_mq_plug_issue_direct(struct blk_plug *plug) |
| 2661 | { |
| 2662 | struct blk_mq_hw_ctx *hctx = NULL; |
| 2663 | struct request *rq; |
| 2664 | int queued = 0; |
| 2665 | blk_status_t ret = BLK_STS_OK; |
| 2666 | |
| 2667 | while ((rq = rq_list_pop(&plug->mq_list))) { |
| 2668 | bool last = rq_list_empty(plug->mq_list); |
| 2669 | |
| 2670 | if (hctx != rq->mq_hctx) { |
| 2671 | if (hctx) { |
| 2672 | blk_mq_commit_rqs(hctx, queued, false); |
| 2673 | queued = 0; |
| 2674 | } |
| 2675 | hctx = rq->mq_hctx; |
| 2676 | } |
| 2677 | |
| 2678 | ret = blk_mq_request_issue_directly(rq, last); |
| 2679 | switch (ret) { |
| 2680 | case BLK_STS_OK: |
| 2681 | queued++; |
| 2682 | break; |
| 2683 | case BLK_STS_RESOURCE: |
| 2684 | case BLK_STS_DEV_RESOURCE: |
| 2685 | blk_mq_request_bypass_insert(rq, 0); |
| 2686 | blk_mq_run_hw_queue(hctx, false); |
| 2687 | goto out; |
| 2688 | default: |
| 2689 | blk_mq_end_request(rq, ret); |
| 2690 | break; |
| 2691 | } |
| 2692 | } |
| 2693 | |
| 2694 | out: |
| 2695 | if (ret != BLK_STS_OK) |
| 2696 | blk_mq_commit_rqs(hctx, queued, false); |
| 2697 | } |
| 2698 | |
| 2699 | static void __blk_mq_flush_plug_list(struct request_queue *q, |
| 2700 | struct blk_plug *plug) |
| 2701 | { |
| 2702 | if (blk_queue_quiesced(q)) |
| 2703 | return; |
| 2704 | q->mq_ops->queue_rqs(&plug->mq_list); |
| 2705 | } |
| 2706 | |
| 2707 | static void blk_mq_dispatch_plug_list(struct blk_plug *plug, bool from_sched) |
| 2708 | { |
| 2709 | struct blk_mq_hw_ctx *this_hctx = NULL; |
| 2710 | struct blk_mq_ctx *this_ctx = NULL; |
| 2711 | struct request *requeue_list = NULL; |
| 2712 | struct request **requeue_lastp = &requeue_list; |
| 2713 | unsigned int depth = 0; |
| 2714 | LIST_HEAD(list); |
| 2715 | |
| 2716 | do { |
| 2717 | struct request *rq = rq_list_pop(&plug->mq_list); |
| 2718 | |
| 2719 | if (!this_hctx) { |
| 2720 | this_hctx = rq->mq_hctx; |
| 2721 | this_ctx = rq->mq_ctx; |
| 2722 | } else if (this_hctx != rq->mq_hctx || this_ctx != rq->mq_ctx) { |
| 2723 | rq_list_add_tail(&requeue_lastp, rq); |
| 2724 | continue; |
| 2725 | } |
| 2726 | list_add(&rq->queuelist, &list); |
| 2727 | depth++; |
| 2728 | } while (!rq_list_empty(plug->mq_list)); |
| 2729 | |
| 2730 | plug->mq_list = requeue_list; |
| 2731 | trace_block_unplug(this_hctx->queue, depth, !from_sched); |
| 2732 | |
| 2733 | percpu_ref_get(&this_hctx->queue->q_usage_counter); |
| 2734 | if (this_hctx->queue->elevator) { |
| 2735 | this_hctx->queue->elevator->type->ops.insert_requests(this_hctx, |
| 2736 | &list, 0); |
| 2737 | blk_mq_run_hw_queue(this_hctx, from_sched); |
| 2738 | } else { |
| 2739 | blk_mq_insert_requests(this_hctx, this_ctx, &list, from_sched); |
| 2740 | } |
| 2741 | percpu_ref_put(&this_hctx->queue->q_usage_counter); |
| 2742 | } |
| 2743 | |
| 2744 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) |
| 2745 | { |
| 2746 | struct request *rq; |
| 2747 | |
| 2748 | if (rq_list_empty(plug->mq_list)) |
| 2749 | return; |
| 2750 | plug->rq_count = 0; |
| 2751 | |
| 2752 | if (!plug->multiple_queues && !plug->has_elevator && !from_schedule) { |
| 2753 | struct request_queue *q; |
| 2754 | |
| 2755 | rq = rq_list_peek(&plug->mq_list); |
| 2756 | q = rq->q; |
| 2757 | |
| 2758 | /* |
| 2759 | * Peek first request and see if we have a ->queue_rqs() hook. |
| 2760 | * If we do, we can dispatch the whole plug list in one go. We |
| 2761 | * already know at this point that all requests belong to the |
| 2762 | * same queue, caller must ensure that's the case. |
| 2763 | * |
| 2764 | * Since we pass off the full list to the driver at this point, |
| 2765 | * we do not increment the active request count for the queue. |
| 2766 | * Bypass shared tags for now because of that. |
| 2767 | */ |
| 2768 | if (q->mq_ops->queue_rqs && |
| 2769 | !(rq->mq_hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) { |
| 2770 | blk_mq_run_dispatch_ops(q, |
| 2771 | __blk_mq_flush_plug_list(q, plug)); |
| 2772 | if (rq_list_empty(plug->mq_list)) |
| 2773 | return; |
| 2774 | } |
| 2775 | |
| 2776 | blk_mq_run_dispatch_ops(q, |
| 2777 | blk_mq_plug_issue_direct(plug)); |
| 2778 | if (rq_list_empty(plug->mq_list)) |
| 2779 | return; |
| 2780 | } |
| 2781 | |
| 2782 | do { |
| 2783 | blk_mq_dispatch_plug_list(plug, from_schedule); |
| 2784 | } while (!rq_list_empty(plug->mq_list)); |
| 2785 | } |
| 2786 | |
| 2787 | static void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx, |
| 2788 | struct list_head *list) |
| 2789 | { |
| 2790 | int queued = 0; |
| 2791 | blk_status_t ret = BLK_STS_OK; |
| 2792 | |
| 2793 | while (!list_empty(list)) { |
| 2794 | struct request *rq = list_first_entry(list, struct request, |
| 2795 | queuelist); |
| 2796 | |
| 2797 | list_del_init(&rq->queuelist); |
| 2798 | ret = blk_mq_request_issue_directly(rq, list_empty(list)); |
| 2799 | switch (ret) { |
| 2800 | case BLK_STS_OK: |
| 2801 | queued++; |
| 2802 | break; |
| 2803 | case BLK_STS_RESOURCE: |
| 2804 | case BLK_STS_DEV_RESOURCE: |
| 2805 | blk_mq_request_bypass_insert(rq, 0); |
| 2806 | if (list_empty(list)) |
| 2807 | blk_mq_run_hw_queue(hctx, false); |
| 2808 | goto out; |
| 2809 | default: |
| 2810 | blk_mq_end_request(rq, ret); |
| 2811 | break; |
| 2812 | } |
| 2813 | } |
| 2814 | |
| 2815 | out: |
| 2816 | if (ret != BLK_STS_OK) |
| 2817 | blk_mq_commit_rqs(hctx, queued, false); |
| 2818 | } |
| 2819 | |
| 2820 | static bool blk_mq_attempt_bio_merge(struct request_queue *q, |
| 2821 | struct bio *bio, unsigned int nr_segs) |
| 2822 | { |
| 2823 | if (!blk_queue_nomerges(q) && bio_mergeable(bio)) { |
| 2824 | if (blk_attempt_plug_merge(q, bio, nr_segs)) |
| 2825 | return true; |
| 2826 | if (blk_mq_sched_bio_merge(q, bio, nr_segs)) |
| 2827 | return true; |
| 2828 | } |
| 2829 | return false; |
| 2830 | } |
| 2831 | |
| 2832 | static struct request *blk_mq_get_new_requests(struct request_queue *q, |
| 2833 | struct blk_plug *plug, |
| 2834 | struct bio *bio, |
| 2835 | unsigned int nsegs) |
| 2836 | { |
| 2837 | struct blk_mq_alloc_data data = { |
| 2838 | .q = q, |
| 2839 | .nr_tags = 1, |
| 2840 | .cmd_flags = bio->bi_opf, |
| 2841 | }; |
| 2842 | struct request *rq; |
| 2843 | |
| 2844 | if (unlikely(bio_queue_enter(bio))) |
| 2845 | return NULL; |
| 2846 | |
| 2847 | if (blk_mq_attempt_bio_merge(q, bio, nsegs)) |
| 2848 | goto queue_exit; |
| 2849 | |
| 2850 | rq_qos_throttle(q, bio); |
| 2851 | |
| 2852 | if (plug) { |
| 2853 | data.nr_tags = plug->nr_ios; |
| 2854 | plug->nr_ios = 1; |
| 2855 | data.cached_rq = &plug->cached_rq; |
| 2856 | } |
| 2857 | |
| 2858 | rq = __blk_mq_alloc_requests(&data); |
| 2859 | if (rq) |
| 2860 | return rq; |
| 2861 | rq_qos_cleanup(q, bio); |
| 2862 | if (bio->bi_opf & REQ_NOWAIT) |
| 2863 | bio_wouldblock_error(bio); |
| 2864 | queue_exit: |
| 2865 | blk_queue_exit(q); |
| 2866 | return NULL; |
| 2867 | } |
| 2868 | |
| 2869 | static inline struct request *blk_mq_get_cached_request(struct request_queue *q, |
| 2870 | struct blk_plug *plug, struct bio **bio, unsigned int nsegs) |
| 2871 | { |
| 2872 | struct request *rq; |
| 2873 | enum hctx_type type, hctx_type; |
| 2874 | |
| 2875 | if (!plug) |
| 2876 | return NULL; |
| 2877 | rq = rq_list_peek(&plug->cached_rq); |
| 2878 | if (!rq || rq->q != q) |
| 2879 | return NULL; |
| 2880 | |
| 2881 | if (blk_mq_attempt_bio_merge(q, *bio, nsegs)) { |
| 2882 | *bio = NULL; |
| 2883 | return NULL; |
| 2884 | } |
| 2885 | |
| 2886 | type = blk_mq_get_hctx_type((*bio)->bi_opf); |
| 2887 | hctx_type = rq->mq_hctx->type; |
| 2888 | if (type != hctx_type && |
| 2889 | !(type == HCTX_TYPE_READ && hctx_type == HCTX_TYPE_DEFAULT)) |
| 2890 | return NULL; |
| 2891 | if (op_is_flush(rq->cmd_flags) != op_is_flush((*bio)->bi_opf)) |
| 2892 | return NULL; |
| 2893 | |
| 2894 | /* |
| 2895 | * If any qos ->throttle() end up blocking, we will have flushed the |
| 2896 | * plug and hence killed the cached_rq list as well. Pop this entry |
| 2897 | * before we throttle. |
| 2898 | */ |
| 2899 | plug->cached_rq = rq_list_next(rq); |
| 2900 | rq_qos_throttle(q, *bio); |
| 2901 | |
| 2902 | rq->cmd_flags = (*bio)->bi_opf; |
| 2903 | INIT_LIST_HEAD(&rq->queuelist); |
| 2904 | return rq; |
| 2905 | } |
| 2906 | |
| 2907 | static void bio_set_ioprio(struct bio *bio) |
| 2908 | { |
| 2909 | /* Nobody set ioprio so far? Initialize it based on task's nice value */ |
| 2910 | if (IOPRIO_PRIO_CLASS(bio->bi_ioprio) == IOPRIO_CLASS_NONE) |
| 2911 | bio->bi_ioprio = get_current_ioprio(); |
| 2912 | blkcg_set_ioprio(bio); |
| 2913 | } |
| 2914 | |
| 2915 | /** |
| 2916 | * blk_mq_submit_bio - Create and send a request to block device. |
| 2917 | * @bio: Bio pointer. |
| 2918 | * |
| 2919 | * Builds up a request structure from @q and @bio and send to the device. The |
| 2920 | * request may not be queued directly to hardware if: |
| 2921 | * * This request can be merged with another one |
| 2922 | * * We want to place request at plug queue for possible future merging |
| 2923 | * * There is an IO scheduler active at this queue |
| 2924 | * |
| 2925 | * It will not queue the request if there is an error with the bio, or at the |
| 2926 | * request creation. |
| 2927 | */ |
| 2928 | void blk_mq_submit_bio(struct bio *bio) |
| 2929 | { |
| 2930 | struct request_queue *q = bdev_get_queue(bio->bi_bdev); |
| 2931 | struct blk_plug *plug = blk_mq_plug(bio); |
| 2932 | const int is_sync = op_is_sync(bio->bi_opf); |
| 2933 | struct blk_mq_hw_ctx *hctx; |
| 2934 | struct request *rq; |
| 2935 | unsigned int nr_segs = 1; |
| 2936 | blk_status_t ret; |
| 2937 | |
| 2938 | bio = blk_queue_bounce(bio, q); |
| 2939 | if (bio_may_exceed_limits(bio, &q->limits)) { |
| 2940 | bio = __bio_split_to_limits(bio, &q->limits, &nr_segs); |
| 2941 | if (!bio) |
| 2942 | return; |
| 2943 | } |
| 2944 | |
| 2945 | if (!bio_integrity_prep(bio)) |
| 2946 | return; |
| 2947 | |
| 2948 | bio_set_ioprio(bio); |
| 2949 | |
| 2950 | rq = blk_mq_get_cached_request(q, plug, &bio, nr_segs); |
| 2951 | if (!rq) { |
| 2952 | if (!bio) |
| 2953 | return; |
| 2954 | rq = blk_mq_get_new_requests(q, plug, bio, nr_segs); |
| 2955 | if (unlikely(!rq)) |
| 2956 | return; |
| 2957 | } |
| 2958 | |
| 2959 | trace_block_getrq(bio); |
| 2960 | |
| 2961 | rq_qos_track(q, rq, bio); |
| 2962 | |
| 2963 | blk_mq_bio_to_request(rq, bio, nr_segs); |
| 2964 | |
| 2965 | ret = blk_crypto_rq_get_keyslot(rq); |
| 2966 | if (ret != BLK_STS_OK) { |
| 2967 | bio->bi_status = ret; |
| 2968 | bio_endio(bio); |
| 2969 | blk_mq_free_request(rq); |
| 2970 | return; |
| 2971 | } |
| 2972 | |
| 2973 | if (op_is_flush(bio->bi_opf)) { |
| 2974 | blk_insert_flush(rq); |
| 2975 | return; |
| 2976 | } |
| 2977 | |
| 2978 | if (plug) { |
| 2979 | blk_add_rq_to_plug(plug, rq); |
| 2980 | return; |
| 2981 | } |
| 2982 | |
| 2983 | hctx = rq->mq_hctx; |
| 2984 | if ((rq->rq_flags & RQF_ELV) || |
| 2985 | (hctx->dispatch_busy && (q->nr_hw_queues == 1 || !is_sync))) { |
| 2986 | blk_mq_insert_request(rq, 0); |
| 2987 | blk_mq_run_hw_queue(hctx, true); |
| 2988 | } else { |
| 2989 | blk_mq_run_dispatch_ops(q, blk_mq_try_issue_directly(hctx, rq)); |
| 2990 | } |
| 2991 | } |
| 2992 | |
| 2993 | #ifdef CONFIG_BLK_MQ_STACKING |
| 2994 | /** |
| 2995 | * blk_insert_cloned_request - Helper for stacking drivers to submit a request |
| 2996 | * @rq: the request being queued |
| 2997 | */ |
| 2998 | blk_status_t blk_insert_cloned_request(struct request *rq) |
| 2999 | { |
| 3000 | struct request_queue *q = rq->q; |
| 3001 | unsigned int max_sectors = blk_queue_get_max_sectors(q, req_op(rq)); |
| 3002 | unsigned int max_segments = blk_rq_get_max_segments(rq); |
| 3003 | blk_status_t ret; |
| 3004 | |
| 3005 | if (blk_rq_sectors(rq) > max_sectors) { |
| 3006 | /* |
| 3007 | * SCSI device does not have a good way to return if |
| 3008 | * Write Same/Zero is actually supported. If a device rejects |
| 3009 | * a non-read/write command (discard, write same,etc.) the |
| 3010 | * low-level device driver will set the relevant queue limit to |
| 3011 | * 0 to prevent blk-lib from issuing more of the offending |
| 3012 | * operations. Commands queued prior to the queue limit being |
| 3013 | * reset need to be completed with BLK_STS_NOTSUPP to avoid I/O |
| 3014 | * errors being propagated to upper layers. |
| 3015 | */ |
| 3016 | if (max_sectors == 0) |
| 3017 | return BLK_STS_NOTSUPP; |
| 3018 | |
| 3019 | printk(KERN_ERR "%s: over max size limit. (%u > %u)\n", |
| 3020 | __func__, blk_rq_sectors(rq), max_sectors); |
| 3021 | return BLK_STS_IOERR; |
| 3022 | } |
| 3023 | |
| 3024 | /* |
| 3025 | * The queue settings related to segment counting may differ from the |
| 3026 | * original queue. |
| 3027 | */ |
| 3028 | rq->nr_phys_segments = blk_recalc_rq_segments(rq); |
| 3029 | if (rq->nr_phys_segments > max_segments) { |
| 3030 | printk(KERN_ERR "%s: over max segments limit. (%u > %u)\n", |
| 3031 | __func__, rq->nr_phys_segments, max_segments); |
| 3032 | return BLK_STS_IOERR; |
| 3033 | } |
| 3034 | |
| 3035 | if (q->disk && should_fail_request(q->disk->part0, blk_rq_bytes(rq))) |
| 3036 | return BLK_STS_IOERR; |
| 3037 | |
| 3038 | ret = blk_crypto_rq_get_keyslot(rq); |
| 3039 | if (ret != BLK_STS_OK) |
| 3040 | return ret; |
| 3041 | |
| 3042 | blk_account_io_start(rq); |
| 3043 | |
| 3044 | /* |
| 3045 | * Since we have a scheduler attached on the top device, |
| 3046 | * bypass a potential scheduler on the bottom device for |
| 3047 | * insert. |
| 3048 | */ |
| 3049 | blk_mq_run_dispatch_ops(q, |
| 3050 | ret = blk_mq_request_issue_directly(rq, true)); |
| 3051 | if (ret) |
| 3052 | blk_account_io_done(rq, ktime_get_ns()); |
| 3053 | return ret; |
| 3054 | } |
| 3055 | EXPORT_SYMBOL_GPL(blk_insert_cloned_request); |
| 3056 | |
| 3057 | /** |
| 3058 | * blk_rq_unprep_clone - Helper function to free all bios in a cloned request |
| 3059 | * @rq: the clone request to be cleaned up |
| 3060 | * |
| 3061 | * Description: |
| 3062 | * Free all bios in @rq for a cloned request. |
| 3063 | */ |
| 3064 | void blk_rq_unprep_clone(struct request *rq) |
| 3065 | { |
| 3066 | struct bio *bio; |
| 3067 | |
| 3068 | while ((bio = rq->bio) != NULL) { |
| 3069 | rq->bio = bio->bi_next; |
| 3070 | |
| 3071 | bio_put(bio); |
| 3072 | } |
| 3073 | } |
| 3074 | EXPORT_SYMBOL_GPL(blk_rq_unprep_clone); |
| 3075 | |
| 3076 | /** |
| 3077 | * blk_rq_prep_clone - Helper function to setup clone request |
| 3078 | * @rq: the request to be setup |
| 3079 | * @rq_src: original request to be cloned |
| 3080 | * @bs: bio_set that bios for clone are allocated from |
| 3081 | * @gfp_mask: memory allocation mask for bio |
| 3082 | * @bio_ctr: setup function to be called for each clone bio. |
| 3083 | * Returns %0 for success, non %0 for failure. |
| 3084 | * @data: private data to be passed to @bio_ctr |
| 3085 | * |
| 3086 | * Description: |
| 3087 | * Clones bios in @rq_src to @rq, and copies attributes of @rq_src to @rq. |
| 3088 | * Also, pages which the original bios are pointing to are not copied |
| 3089 | * and the cloned bios just point same pages. |
| 3090 | * So cloned bios must be completed before original bios, which means |
| 3091 | * the caller must complete @rq before @rq_src. |
| 3092 | */ |
| 3093 | int blk_rq_prep_clone(struct request *rq, struct request *rq_src, |
| 3094 | struct bio_set *bs, gfp_t gfp_mask, |
| 3095 | int (*bio_ctr)(struct bio *, struct bio *, void *), |
| 3096 | void *data) |
| 3097 | { |
| 3098 | struct bio *bio, *bio_src; |
| 3099 | |
| 3100 | if (!bs) |
| 3101 | bs = &fs_bio_set; |
| 3102 | |
| 3103 | __rq_for_each_bio(bio_src, rq_src) { |
| 3104 | bio = bio_alloc_clone(rq->q->disk->part0, bio_src, gfp_mask, |
| 3105 | bs); |
| 3106 | if (!bio) |
| 3107 | goto free_and_out; |
| 3108 | |
| 3109 | if (bio_ctr && bio_ctr(bio, bio_src, data)) |
| 3110 | goto free_and_out; |
| 3111 | |
| 3112 | if (rq->bio) { |
| 3113 | rq->biotail->bi_next = bio; |
| 3114 | rq->biotail = bio; |
| 3115 | } else { |
| 3116 | rq->bio = rq->biotail = bio; |
| 3117 | } |
| 3118 | bio = NULL; |
| 3119 | } |
| 3120 | |
| 3121 | /* Copy attributes of the original request to the clone request. */ |
| 3122 | rq->__sector = blk_rq_pos(rq_src); |
| 3123 | rq->__data_len = blk_rq_bytes(rq_src); |
| 3124 | if (rq_src->rq_flags & RQF_SPECIAL_PAYLOAD) { |
| 3125 | rq->rq_flags |= RQF_SPECIAL_PAYLOAD; |
| 3126 | rq->special_vec = rq_src->special_vec; |
| 3127 | } |
| 3128 | rq->nr_phys_segments = rq_src->nr_phys_segments; |
| 3129 | rq->ioprio = rq_src->ioprio; |
| 3130 | |
| 3131 | if (rq->bio && blk_crypto_rq_bio_prep(rq, rq->bio, gfp_mask) < 0) |
| 3132 | goto free_and_out; |
| 3133 | |
| 3134 | return 0; |
| 3135 | |
| 3136 | free_and_out: |
| 3137 | if (bio) |
| 3138 | bio_put(bio); |
| 3139 | blk_rq_unprep_clone(rq); |
| 3140 | |
| 3141 | return -ENOMEM; |
| 3142 | } |
| 3143 | EXPORT_SYMBOL_GPL(blk_rq_prep_clone); |
| 3144 | #endif /* CONFIG_BLK_MQ_STACKING */ |
| 3145 | |
| 3146 | /* |
| 3147 | * Steal bios from a request and add them to a bio list. |
| 3148 | * The request must not have been partially completed before. |
| 3149 | */ |
| 3150 | void blk_steal_bios(struct bio_list *list, struct request *rq) |
| 3151 | { |
| 3152 | if (rq->bio) { |
| 3153 | if (list->tail) |
| 3154 | list->tail->bi_next = rq->bio; |
| 3155 | else |
| 3156 | list->head = rq->bio; |
| 3157 | list->tail = rq->biotail; |
| 3158 | |
| 3159 | rq->bio = NULL; |
| 3160 | rq->biotail = NULL; |
| 3161 | } |
| 3162 | |
| 3163 | rq->__data_len = 0; |
| 3164 | } |
| 3165 | EXPORT_SYMBOL_GPL(blk_steal_bios); |
| 3166 | |
| 3167 | static size_t order_to_size(unsigned int order) |
| 3168 | { |
| 3169 | return (size_t)PAGE_SIZE << order; |
| 3170 | } |
| 3171 | |
| 3172 | /* called before freeing request pool in @tags */ |
| 3173 | static void blk_mq_clear_rq_mapping(struct blk_mq_tags *drv_tags, |
| 3174 | struct blk_mq_tags *tags) |
| 3175 | { |
| 3176 | struct page *page; |
| 3177 | unsigned long flags; |
| 3178 | |
| 3179 | /* |
| 3180 | * There is no need to clear mapping if driver tags is not initialized |
| 3181 | * or the mapping belongs to the driver tags. |
| 3182 | */ |
| 3183 | if (!drv_tags || drv_tags == tags) |
| 3184 | return; |
| 3185 | |
| 3186 | list_for_each_entry(page, &tags->page_list, lru) { |
| 3187 | unsigned long start = (unsigned long)page_address(page); |
| 3188 | unsigned long end = start + order_to_size(page->private); |
| 3189 | int i; |
| 3190 | |
| 3191 | for (i = 0; i < drv_tags->nr_tags; i++) { |
| 3192 | struct request *rq = drv_tags->rqs[i]; |
| 3193 | unsigned long rq_addr = (unsigned long)rq; |
| 3194 | |
| 3195 | if (rq_addr >= start && rq_addr < end) { |
| 3196 | WARN_ON_ONCE(req_ref_read(rq) != 0); |
| 3197 | cmpxchg(&drv_tags->rqs[i], rq, NULL); |
| 3198 | } |
| 3199 | } |
| 3200 | } |
| 3201 | |
| 3202 | /* |
| 3203 | * Wait until all pending iteration is done. |
| 3204 | * |
| 3205 | * Request reference is cleared and it is guaranteed to be observed |
| 3206 | * after the ->lock is released. |
| 3207 | */ |
| 3208 | spin_lock_irqsave(&drv_tags->lock, flags); |
| 3209 | spin_unlock_irqrestore(&drv_tags->lock, flags); |
| 3210 | } |
| 3211 | |
| 3212 | void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, |
| 3213 | unsigned int hctx_idx) |
| 3214 | { |
| 3215 | struct blk_mq_tags *drv_tags; |
| 3216 | struct page *page; |
| 3217 | |
| 3218 | if (list_empty(&tags->page_list)) |
| 3219 | return; |
| 3220 | |
| 3221 | if (blk_mq_is_shared_tags(set->flags)) |
| 3222 | drv_tags = set->shared_tags; |
| 3223 | else |
| 3224 | drv_tags = set->tags[hctx_idx]; |
| 3225 | |
| 3226 | if (tags->static_rqs && set->ops->exit_request) { |
| 3227 | int i; |
| 3228 | |
| 3229 | for (i = 0; i < tags->nr_tags; i++) { |
| 3230 | struct request *rq = tags->static_rqs[i]; |
| 3231 | |
| 3232 | if (!rq) |
| 3233 | continue; |
| 3234 | set->ops->exit_request(set, rq, hctx_idx); |
| 3235 | tags->static_rqs[i] = NULL; |
| 3236 | } |
| 3237 | } |
| 3238 | |
| 3239 | blk_mq_clear_rq_mapping(drv_tags, tags); |
| 3240 | |
| 3241 | while (!list_empty(&tags->page_list)) { |
| 3242 | page = list_first_entry(&tags->page_list, struct page, lru); |
| 3243 | list_del_init(&page->lru); |
| 3244 | /* |
| 3245 | * Remove kmemleak object previously allocated in |
| 3246 | * blk_mq_alloc_rqs(). |
| 3247 | */ |
| 3248 | kmemleak_free(page_address(page)); |
| 3249 | __free_pages(page, page->private); |
| 3250 | } |
| 3251 | } |
| 3252 | |
| 3253 | void blk_mq_free_rq_map(struct blk_mq_tags *tags) |
| 3254 | { |
| 3255 | kfree(tags->rqs); |
| 3256 | tags->rqs = NULL; |
| 3257 | kfree(tags->static_rqs); |
| 3258 | tags->static_rqs = NULL; |
| 3259 | |
| 3260 | blk_mq_free_tags(tags); |
| 3261 | } |
| 3262 | |
| 3263 | static enum hctx_type hctx_idx_to_type(struct blk_mq_tag_set *set, |
| 3264 | unsigned int hctx_idx) |
| 3265 | { |
| 3266 | int i; |
| 3267 | |
| 3268 | for (i = 0; i < set->nr_maps; i++) { |
| 3269 | unsigned int start = set->map[i].queue_offset; |
| 3270 | unsigned int end = start + set->map[i].nr_queues; |
| 3271 | |
| 3272 | if (hctx_idx >= start && hctx_idx < end) |
| 3273 | break; |
| 3274 | } |
| 3275 | |
| 3276 | if (i >= set->nr_maps) |
| 3277 | i = HCTX_TYPE_DEFAULT; |
| 3278 | |
| 3279 | return i; |
| 3280 | } |
| 3281 | |
| 3282 | static int blk_mq_get_hctx_node(struct blk_mq_tag_set *set, |
| 3283 | unsigned int hctx_idx) |
| 3284 | { |
| 3285 | enum hctx_type type = hctx_idx_to_type(set, hctx_idx); |
| 3286 | |
| 3287 | return blk_mq_hw_queue_to_node(&set->map[type], hctx_idx); |
| 3288 | } |
| 3289 | |
| 3290 | static struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, |
| 3291 | unsigned int hctx_idx, |
| 3292 | unsigned int nr_tags, |
| 3293 | unsigned int reserved_tags) |
| 3294 | { |
| 3295 | int node = blk_mq_get_hctx_node(set, hctx_idx); |
| 3296 | struct blk_mq_tags *tags; |
| 3297 | |
| 3298 | if (node == NUMA_NO_NODE) |
| 3299 | node = set->numa_node; |
| 3300 | |
| 3301 | tags = blk_mq_init_tags(nr_tags, reserved_tags, node, |
| 3302 | BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags)); |
| 3303 | if (!tags) |
| 3304 | return NULL; |
| 3305 | |
| 3306 | tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *), |
| 3307 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, |
| 3308 | node); |
| 3309 | if (!tags->rqs) |
| 3310 | goto err_free_tags; |
| 3311 | |
| 3312 | tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *), |
| 3313 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, |
| 3314 | node); |
| 3315 | if (!tags->static_rqs) |
| 3316 | goto err_free_rqs; |
| 3317 | |
| 3318 | return tags; |
| 3319 | |
| 3320 | err_free_rqs: |
| 3321 | kfree(tags->rqs); |
| 3322 | err_free_tags: |
| 3323 | blk_mq_free_tags(tags); |
| 3324 | return NULL; |
| 3325 | } |
| 3326 | |
| 3327 | static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq, |
| 3328 | unsigned int hctx_idx, int node) |
| 3329 | { |
| 3330 | int ret; |
| 3331 | |
| 3332 | if (set->ops->init_request) { |
| 3333 | ret = set->ops->init_request(set, rq, hctx_idx, node); |
| 3334 | if (ret) |
| 3335 | return ret; |
| 3336 | } |
| 3337 | |
| 3338 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); |
| 3339 | return 0; |
| 3340 | } |
| 3341 | |
| 3342 | static int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, |
| 3343 | struct blk_mq_tags *tags, |
| 3344 | unsigned int hctx_idx, unsigned int depth) |
| 3345 | { |
| 3346 | unsigned int i, j, entries_per_page, max_order = 4; |
| 3347 | int node = blk_mq_get_hctx_node(set, hctx_idx); |
| 3348 | size_t rq_size, left; |
| 3349 | |
| 3350 | if (node == NUMA_NO_NODE) |
| 3351 | node = set->numa_node; |
| 3352 | |
| 3353 | INIT_LIST_HEAD(&tags->page_list); |
| 3354 | |
| 3355 | /* |
| 3356 | * rq_size is the size of the request plus driver payload, rounded |
| 3357 | * to the cacheline size |
| 3358 | */ |
| 3359 | rq_size = round_up(sizeof(struct request) + set->cmd_size, |
| 3360 | cache_line_size()); |
| 3361 | left = rq_size * depth; |
| 3362 | |
| 3363 | for (i = 0; i < depth; ) { |
| 3364 | int this_order = max_order; |
| 3365 | struct page *page; |
| 3366 | int to_do; |
| 3367 | void *p; |
| 3368 | |
| 3369 | while (this_order && left < order_to_size(this_order - 1)) |
| 3370 | this_order--; |
| 3371 | |
| 3372 | do { |
| 3373 | page = alloc_pages_node(node, |
| 3374 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO, |
| 3375 | this_order); |
| 3376 | if (page) |
| 3377 | break; |
| 3378 | if (!this_order--) |
| 3379 | break; |
| 3380 | if (order_to_size(this_order) < rq_size) |
| 3381 | break; |
| 3382 | } while (1); |
| 3383 | |
| 3384 | if (!page) |
| 3385 | goto fail; |
| 3386 | |
| 3387 | page->private = this_order; |
| 3388 | list_add_tail(&page->lru, &tags->page_list); |
| 3389 | |
| 3390 | p = page_address(page); |
| 3391 | /* |
| 3392 | * Allow kmemleak to scan these pages as they contain pointers |
| 3393 | * to additional allocations like via ops->init_request(). |
| 3394 | */ |
| 3395 | kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO); |
| 3396 | entries_per_page = order_to_size(this_order) / rq_size; |
| 3397 | to_do = min(entries_per_page, depth - i); |
| 3398 | left -= to_do * rq_size; |
| 3399 | for (j = 0; j < to_do; j++) { |
| 3400 | struct request *rq = p; |
| 3401 | |
| 3402 | tags->static_rqs[i] = rq; |
| 3403 | if (blk_mq_init_request(set, rq, hctx_idx, node)) { |
| 3404 | tags->static_rqs[i] = NULL; |
| 3405 | goto fail; |
| 3406 | } |
| 3407 | |
| 3408 | p += rq_size; |
| 3409 | i++; |
| 3410 | } |
| 3411 | } |
| 3412 | return 0; |
| 3413 | |
| 3414 | fail: |
| 3415 | blk_mq_free_rqs(set, tags, hctx_idx); |
| 3416 | return -ENOMEM; |
| 3417 | } |
| 3418 | |
| 3419 | struct rq_iter_data { |
| 3420 | struct blk_mq_hw_ctx *hctx; |
| 3421 | bool has_rq; |
| 3422 | }; |
| 3423 | |
| 3424 | static bool blk_mq_has_request(struct request *rq, void *data) |
| 3425 | { |
| 3426 | struct rq_iter_data *iter_data = data; |
| 3427 | |
| 3428 | if (rq->mq_hctx != iter_data->hctx) |
| 3429 | return true; |
| 3430 | iter_data->has_rq = true; |
| 3431 | return false; |
| 3432 | } |
| 3433 | |
| 3434 | static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx) |
| 3435 | { |
| 3436 | struct blk_mq_tags *tags = hctx->sched_tags ? |
| 3437 | hctx->sched_tags : hctx->tags; |
| 3438 | struct rq_iter_data data = { |
| 3439 | .hctx = hctx, |
| 3440 | }; |
| 3441 | |
| 3442 | blk_mq_all_tag_iter(tags, blk_mq_has_request, &data); |
| 3443 | return data.has_rq; |
| 3444 | } |
| 3445 | |
| 3446 | static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu, |
| 3447 | struct blk_mq_hw_ctx *hctx) |
| 3448 | { |
| 3449 | if (cpumask_first_and(hctx->cpumask, cpu_online_mask) != cpu) |
| 3450 | return false; |
| 3451 | if (cpumask_next_and(cpu, hctx->cpumask, cpu_online_mask) < nr_cpu_ids) |
| 3452 | return false; |
| 3453 | return true; |
| 3454 | } |
| 3455 | |
| 3456 | static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node) |
| 3457 | { |
| 3458 | struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node, |
| 3459 | struct blk_mq_hw_ctx, cpuhp_online); |
| 3460 | |
| 3461 | if (!cpumask_test_cpu(cpu, hctx->cpumask) || |
| 3462 | !blk_mq_last_cpu_in_hctx(cpu, hctx)) |
| 3463 | return 0; |
| 3464 | |
| 3465 | /* |
| 3466 | * Prevent new request from being allocated on the current hctx. |
| 3467 | * |
| 3468 | * The smp_mb__after_atomic() Pairs with the implied barrier in |
| 3469 | * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is |
| 3470 | * seen once we return from the tag allocator. |
| 3471 | */ |
| 3472 | set_bit(BLK_MQ_S_INACTIVE, &hctx->state); |
| 3473 | smp_mb__after_atomic(); |
| 3474 | |
| 3475 | /* |
| 3476 | * Try to grab a reference to the queue and wait for any outstanding |
| 3477 | * requests. If we could not grab a reference the queue has been |
| 3478 | * frozen and there are no requests. |
| 3479 | */ |
| 3480 | if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) { |
| 3481 | while (blk_mq_hctx_has_requests(hctx)) |
| 3482 | msleep(5); |
| 3483 | percpu_ref_put(&hctx->queue->q_usage_counter); |
| 3484 | } |
| 3485 | |
| 3486 | return 0; |
| 3487 | } |
| 3488 | |
| 3489 | static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node) |
| 3490 | { |
| 3491 | struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node, |
| 3492 | struct blk_mq_hw_ctx, cpuhp_online); |
| 3493 | |
| 3494 | if (cpumask_test_cpu(cpu, hctx->cpumask)) |
| 3495 | clear_bit(BLK_MQ_S_INACTIVE, &hctx->state); |
| 3496 | return 0; |
| 3497 | } |
| 3498 | |
| 3499 | /* |
| 3500 | * 'cpu' is going away. splice any existing rq_list entries from this |
| 3501 | * software queue to the hw queue dispatch list, and ensure that it |
| 3502 | * gets run. |
| 3503 | */ |
| 3504 | static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node) |
| 3505 | { |
| 3506 | struct blk_mq_hw_ctx *hctx; |
| 3507 | struct blk_mq_ctx *ctx; |
| 3508 | LIST_HEAD(tmp); |
| 3509 | enum hctx_type type; |
| 3510 | |
| 3511 | hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead); |
| 3512 | if (!cpumask_test_cpu(cpu, hctx->cpumask)) |
| 3513 | return 0; |
| 3514 | |
| 3515 | ctx = __blk_mq_get_ctx(hctx->queue, cpu); |
| 3516 | type = hctx->type; |
| 3517 | |
| 3518 | spin_lock(&ctx->lock); |
| 3519 | if (!list_empty(&ctx->rq_lists[type])) { |
| 3520 | list_splice_init(&ctx->rq_lists[type], &tmp); |
| 3521 | blk_mq_hctx_clear_pending(hctx, ctx); |
| 3522 | } |
| 3523 | spin_unlock(&ctx->lock); |
| 3524 | |
| 3525 | if (list_empty(&tmp)) |
| 3526 | return 0; |
| 3527 | |
| 3528 | spin_lock(&hctx->lock); |
| 3529 | list_splice_tail_init(&tmp, &hctx->dispatch); |
| 3530 | spin_unlock(&hctx->lock); |
| 3531 | |
| 3532 | blk_mq_run_hw_queue(hctx, true); |
| 3533 | return 0; |
| 3534 | } |
| 3535 | |
| 3536 | static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx) |
| 3537 | { |
| 3538 | if (!(hctx->flags & BLK_MQ_F_STACKING)) |
| 3539 | cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE, |
| 3540 | &hctx->cpuhp_online); |
| 3541 | cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD, |
| 3542 | &hctx->cpuhp_dead); |
| 3543 | } |
| 3544 | |
| 3545 | /* |
| 3546 | * Before freeing hw queue, clearing the flush request reference in |
| 3547 | * tags->rqs[] for avoiding potential UAF. |
| 3548 | */ |
| 3549 | static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags *tags, |
| 3550 | unsigned int queue_depth, struct request *flush_rq) |
| 3551 | { |
| 3552 | int i; |
| 3553 | unsigned long flags; |
| 3554 | |
| 3555 | /* The hw queue may not be mapped yet */ |
| 3556 | if (!tags) |
| 3557 | return; |
| 3558 | |
| 3559 | WARN_ON_ONCE(req_ref_read(flush_rq) != 0); |
| 3560 | |
| 3561 | for (i = 0; i < queue_depth; i++) |
| 3562 | cmpxchg(&tags->rqs[i], flush_rq, NULL); |
| 3563 | |
| 3564 | /* |
| 3565 | * Wait until all pending iteration is done. |
| 3566 | * |
| 3567 | * Request reference is cleared and it is guaranteed to be observed |
| 3568 | * after the ->lock is released. |
| 3569 | */ |
| 3570 | spin_lock_irqsave(&tags->lock, flags); |
| 3571 | spin_unlock_irqrestore(&tags->lock, flags); |
| 3572 | } |
| 3573 | |
| 3574 | /* hctx->ctxs will be freed in queue's release handler */ |
| 3575 | static void blk_mq_exit_hctx(struct request_queue *q, |
| 3576 | struct blk_mq_tag_set *set, |
| 3577 | struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) |
| 3578 | { |
| 3579 | struct request *flush_rq = hctx->fq->flush_rq; |
| 3580 | |
| 3581 | if (blk_mq_hw_queue_mapped(hctx)) |
| 3582 | blk_mq_tag_idle(hctx); |
| 3583 | |
| 3584 | if (blk_queue_init_done(q)) |
| 3585 | blk_mq_clear_flush_rq_mapping(set->tags[hctx_idx], |
| 3586 | set->queue_depth, flush_rq); |
| 3587 | if (set->ops->exit_request) |
| 3588 | set->ops->exit_request(set, flush_rq, hctx_idx); |
| 3589 | |
| 3590 | if (set->ops->exit_hctx) |
| 3591 | set->ops->exit_hctx(hctx, hctx_idx); |
| 3592 | |
| 3593 | blk_mq_remove_cpuhp(hctx); |
| 3594 | |
| 3595 | xa_erase(&q->hctx_table, hctx_idx); |
| 3596 | |
| 3597 | spin_lock(&q->unused_hctx_lock); |
| 3598 | list_add(&hctx->hctx_list, &q->unused_hctx_list); |
| 3599 | spin_unlock(&q->unused_hctx_lock); |
| 3600 | } |
| 3601 | |
| 3602 | static void blk_mq_exit_hw_queues(struct request_queue *q, |
| 3603 | struct blk_mq_tag_set *set, int nr_queue) |
| 3604 | { |
| 3605 | struct blk_mq_hw_ctx *hctx; |
| 3606 | unsigned long i; |
| 3607 | |
| 3608 | queue_for_each_hw_ctx(q, hctx, i) { |
| 3609 | if (i == nr_queue) |
| 3610 | break; |
| 3611 | blk_mq_exit_hctx(q, set, hctx, i); |
| 3612 | } |
| 3613 | } |
| 3614 | |
| 3615 | static int blk_mq_init_hctx(struct request_queue *q, |
| 3616 | struct blk_mq_tag_set *set, |
| 3617 | struct blk_mq_hw_ctx *hctx, unsigned hctx_idx) |
| 3618 | { |
| 3619 | hctx->queue_num = hctx_idx; |
| 3620 | |
| 3621 | if (!(hctx->flags & BLK_MQ_F_STACKING)) |
| 3622 | cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE, |
| 3623 | &hctx->cpuhp_online); |
| 3624 | cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead); |
| 3625 | |
| 3626 | hctx->tags = set->tags[hctx_idx]; |
| 3627 | |
| 3628 | if (set->ops->init_hctx && |
| 3629 | set->ops->init_hctx(hctx, set->driver_data, hctx_idx)) |
| 3630 | goto unregister_cpu_notifier; |
| 3631 | |
| 3632 | if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx, |
| 3633 | hctx->numa_node)) |
| 3634 | goto exit_hctx; |
| 3635 | |
| 3636 | if (xa_insert(&q->hctx_table, hctx_idx, hctx, GFP_KERNEL)) |
| 3637 | goto exit_flush_rq; |
| 3638 | |
| 3639 | return 0; |
| 3640 | |
| 3641 | exit_flush_rq: |
| 3642 | if (set->ops->exit_request) |
| 3643 | set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx); |
| 3644 | exit_hctx: |
| 3645 | if (set->ops->exit_hctx) |
| 3646 | set->ops->exit_hctx(hctx, hctx_idx); |
| 3647 | unregister_cpu_notifier: |
| 3648 | blk_mq_remove_cpuhp(hctx); |
| 3649 | return -1; |
| 3650 | } |
| 3651 | |
| 3652 | static struct blk_mq_hw_ctx * |
| 3653 | blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set, |
| 3654 | int node) |
| 3655 | { |
| 3656 | struct blk_mq_hw_ctx *hctx; |
| 3657 | gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY; |
| 3658 | |
| 3659 | hctx = kzalloc_node(sizeof(struct blk_mq_hw_ctx), gfp, node); |
| 3660 | if (!hctx) |
| 3661 | goto fail_alloc_hctx; |
| 3662 | |
| 3663 | if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node)) |
| 3664 | goto free_hctx; |
| 3665 | |
| 3666 | atomic_set(&hctx->nr_active, 0); |
| 3667 | if (node == NUMA_NO_NODE) |
| 3668 | node = set->numa_node; |
| 3669 | hctx->numa_node = node; |
| 3670 | |
| 3671 | INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn); |
| 3672 | spin_lock_init(&hctx->lock); |
| 3673 | INIT_LIST_HEAD(&hctx->dispatch); |
| 3674 | hctx->queue = q; |
| 3675 | hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED; |
| 3676 | |
| 3677 | INIT_LIST_HEAD(&hctx->hctx_list); |
| 3678 | |
| 3679 | /* |
| 3680 | * Allocate space for all possible cpus to avoid allocation at |
| 3681 | * runtime |
| 3682 | */ |
| 3683 | hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *), |
| 3684 | gfp, node); |
| 3685 | if (!hctx->ctxs) |
| 3686 | goto free_cpumask; |
| 3687 | |
| 3688 | if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), |
| 3689 | gfp, node, false, false)) |
| 3690 | goto free_ctxs; |
| 3691 | hctx->nr_ctx = 0; |
| 3692 | |
| 3693 | spin_lock_init(&hctx->dispatch_wait_lock); |
| 3694 | init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake); |
| 3695 | INIT_LIST_HEAD(&hctx->dispatch_wait.entry); |
| 3696 | |
| 3697 | hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp); |
| 3698 | if (!hctx->fq) |
| 3699 | goto free_bitmap; |
| 3700 | |
| 3701 | blk_mq_hctx_kobj_init(hctx); |
| 3702 | |
| 3703 | return hctx; |
| 3704 | |
| 3705 | free_bitmap: |
| 3706 | sbitmap_free(&hctx->ctx_map); |
| 3707 | free_ctxs: |
| 3708 | kfree(hctx->ctxs); |
| 3709 | free_cpumask: |
| 3710 | free_cpumask_var(hctx->cpumask); |
| 3711 | free_hctx: |
| 3712 | kfree(hctx); |
| 3713 | fail_alloc_hctx: |
| 3714 | return NULL; |
| 3715 | } |
| 3716 | |
| 3717 | static void blk_mq_init_cpu_queues(struct request_queue *q, |
| 3718 | unsigned int nr_hw_queues) |
| 3719 | { |
| 3720 | struct blk_mq_tag_set *set = q->tag_set; |
| 3721 | unsigned int i, j; |
| 3722 | |
| 3723 | for_each_possible_cpu(i) { |
| 3724 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); |
| 3725 | struct blk_mq_hw_ctx *hctx; |
| 3726 | int k; |
| 3727 | |
| 3728 | __ctx->cpu = i; |
| 3729 | spin_lock_init(&__ctx->lock); |
| 3730 | for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++) |
| 3731 | INIT_LIST_HEAD(&__ctx->rq_lists[k]); |
| 3732 | |
| 3733 | __ctx->queue = q; |
| 3734 | |
| 3735 | /* |
| 3736 | * Set local node, IFF we have more than one hw queue. If |
| 3737 | * not, we remain on the home node of the device |
| 3738 | */ |
| 3739 | for (j = 0; j < set->nr_maps; j++) { |
| 3740 | hctx = blk_mq_map_queue_type(q, j, i); |
| 3741 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) |
| 3742 | hctx->numa_node = cpu_to_node(i); |
| 3743 | } |
| 3744 | } |
| 3745 | } |
| 3746 | |
| 3747 | struct blk_mq_tags *blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set, |
| 3748 | unsigned int hctx_idx, |
| 3749 | unsigned int depth) |
| 3750 | { |
| 3751 | struct blk_mq_tags *tags; |
| 3752 | int ret; |
| 3753 | |
| 3754 | tags = blk_mq_alloc_rq_map(set, hctx_idx, depth, set->reserved_tags); |
| 3755 | if (!tags) |
| 3756 | return NULL; |
| 3757 | |
| 3758 | ret = blk_mq_alloc_rqs(set, tags, hctx_idx, depth); |
| 3759 | if (ret) { |
| 3760 | blk_mq_free_rq_map(tags); |
| 3761 | return NULL; |
| 3762 | } |
| 3763 | |
| 3764 | return tags; |
| 3765 | } |
| 3766 | |
| 3767 | static bool __blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set, |
| 3768 | int hctx_idx) |
| 3769 | { |
| 3770 | if (blk_mq_is_shared_tags(set->flags)) { |
| 3771 | set->tags[hctx_idx] = set->shared_tags; |
| 3772 | |
| 3773 | return true; |
| 3774 | } |
| 3775 | |
| 3776 | set->tags[hctx_idx] = blk_mq_alloc_map_and_rqs(set, hctx_idx, |
| 3777 | set->queue_depth); |
| 3778 | |
| 3779 | return set->tags[hctx_idx]; |
| 3780 | } |
| 3781 | |
| 3782 | void blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set, |
| 3783 | struct blk_mq_tags *tags, |
| 3784 | unsigned int hctx_idx) |
| 3785 | { |
| 3786 | if (tags) { |
| 3787 | blk_mq_free_rqs(set, tags, hctx_idx); |
| 3788 | blk_mq_free_rq_map(tags); |
| 3789 | } |
| 3790 | } |
| 3791 | |
| 3792 | static void __blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set, |
| 3793 | unsigned int hctx_idx) |
| 3794 | { |
| 3795 | if (!blk_mq_is_shared_tags(set->flags)) |
| 3796 | blk_mq_free_map_and_rqs(set, set->tags[hctx_idx], hctx_idx); |
| 3797 | |
| 3798 | set->tags[hctx_idx] = NULL; |
| 3799 | } |
| 3800 | |
| 3801 | static void blk_mq_map_swqueue(struct request_queue *q) |
| 3802 | { |
| 3803 | unsigned int j, hctx_idx; |
| 3804 | unsigned long i; |
| 3805 | struct blk_mq_hw_ctx *hctx; |
| 3806 | struct blk_mq_ctx *ctx; |
| 3807 | struct blk_mq_tag_set *set = q->tag_set; |
| 3808 | |
| 3809 | queue_for_each_hw_ctx(q, hctx, i) { |
| 3810 | cpumask_clear(hctx->cpumask); |
| 3811 | hctx->nr_ctx = 0; |
| 3812 | hctx->dispatch_from = NULL; |
| 3813 | } |
| 3814 | |
| 3815 | /* |
| 3816 | * Map software to hardware queues. |
| 3817 | * |
| 3818 | * If the cpu isn't present, the cpu is mapped to first hctx. |
| 3819 | */ |
| 3820 | for_each_possible_cpu(i) { |
| 3821 | |
| 3822 | ctx = per_cpu_ptr(q->queue_ctx, i); |
| 3823 | for (j = 0; j < set->nr_maps; j++) { |
| 3824 | if (!set->map[j].nr_queues) { |
| 3825 | ctx->hctxs[j] = blk_mq_map_queue_type(q, |
| 3826 | HCTX_TYPE_DEFAULT, i); |
| 3827 | continue; |
| 3828 | } |
| 3829 | hctx_idx = set->map[j].mq_map[i]; |
| 3830 | /* unmapped hw queue can be remapped after CPU topo changed */ |
| 3831 | if (!set->tags[hctx_idx] && |
| 3832 | !__blk_mq_alloc_map_and_rqs(set, hctx_idx)) { |
| 3833 | /* |
| 3834 | * If tags initialization fail for some hctx, |
| 3835 | * that hctx won't be brought online. In this |
| 3836 | * case, remap the current ctx to hctx[0] which |
| 3837 | * is guaranteed to always have tags allocated |
| 3838 | */ |
| 3839 | set->map[j].mq_map[i] = 0; |
| 3840 | } |
| 3841 | |
| 3842 | hctx = blk_mq_map_queue_type(q, j, i); |
| 3843 | ctx->hctxs[j] = hctx; |
| 3844 | /* |
| 3845 | * If the CPU is already set in the mask, then we've |
| 3846 | * mapped this one already. This can happen if |
| 3847 | * devices share queues across queue maps. |
| 3848 | */ |
| 3849 | if (cpumask_test_cpu(i, hctx->cpumask)) |
| 3850 | continue; |
| 3851 | |
| 3852 | cpumask_set_cpu(i, hctx->cpumask); |
| 3853 | hctx->type = j; |
| 3854 | ctx->index_hw[hctx->type] = hctx->nr_ctx; |
| 3855 | hctx->ctxs[hctx->nr_ctx++] = ctx; |
| 3856 | |
| 3857 | /* |
| 3858 | * If the nr_ctx type overflows, we have exceeded the |
| 3859 | * amount of sw queues we can support. |
| 3860 | */ |
| 3861 | BUG_ON(!hctx->nr_ctx); |
| 3862 | } |
| 3863 | |
| 3864 | for (; j < HCTX_MAX_TYPES; j++) |
| 3865 | ctx->hctxs[j] = blk_mq_map_queue_type(q, |
| 3866 | HCTX_TYPE_DEFAULT, i); |
| 3867 | } |
| 3868 | |
| 3869 | queue_for_each_hw_ctx(q, hctx, i) { |
| 3870 | /* |
| 3871 | * If no software queues are mapped to this hardware queue, |
| 3872 | * disable it and free the request entries. |
| 3873 | */ |
| 3874 | if (!hctx->nr_ctx) { |
| 3875 | /* Never unmap queue 0. We need it as a |
| 3876 | * fallback in case of a new remap fails |
| 3877 | * allocation |
| 3878 | */ |
| 3879 | if (i) |
| 3880 | __blk_mq_free_map_and_rqs(set, i); |
| 3881 | |
| 3882 | hctx->tags = NULL; |
| 3883 | continue; |
| 3884 | } |
| 3885 | |
| 3886 | hctx->tags = set->tags[i]; |
| 3887 | WARN_ON(!hctx->tags); |
| 3888 | |
| 3889 | /* |
| 3890 | * Set the map size to the number of mapped software queues. |
| 3891 | * This is more accurate and more efficient than looping |
| 3892 | * over all possibly mapped software queues. |
| 3893 | */ |
| 3894 | sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx); |
| 3895 | |
| 3896 | /* |
| 3897 | * Initialize batch roundrobin counts |
| 3898 | */ |
| 3899 | hctx->next_cpu = blk_mq_first_mapped_cpu(hctx); |
| 3900 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; |
| 3901 | } |
| 3902 | } |
| 3903 | |
| 3904 | /* |
| 3905 | * Caller needs to ensure that we're either frozen/quiesced, or that |
| 3906 | * the queue isn't live yet. |
| 3907 | */ |
| 3908 | static void queue_set_hctx_shared(struct request_queue *q, bool shared) |
| 3909 | { |
| 3910 | struct blk_mq_hw_ctx *hctx; |
| 3911 | unsigned long i; |
| 3912 | |
| 3913 | queue_for_each_hw_ctx(q, hctx, i) { |
| 3914 | if (shared) { |
| 3915 | hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED; |
| 3916 | } else { |
| 3917 | blk_mq_tag_idle(hctx); |
| 3918 | hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED; |
| 3919 | } |
| 3920 | } |
| 3921 | } |
| 3922 | |
| 3923 | static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set, |
| 3924 | bool shared) |
| 3925 | { |
| 3926 | struct request_queue *q; |
| 3927 | |
| 3928 | lockdep_assert_held(&set->tag_list_lock); |
| 3929 | |
| 3930 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
| 3931 | blk_mq_freeze_queue(q); |
| 3932 | queue_set_hctx_shared(q, shared); |
| 3933 | blk_mq_unfreeze_queue(q); |
| 3934 | } |
| 3935 | } |
| 3936 | |
| 3937 | static void blk_mq_del_queue_tag_set(struct request_queue *q) |
| 3938 | { |
| 3939 | struct blk_mq_tag_set *set = q->tag_set; |
| 3940 | |
| 3941 | mutex_lock(&set->tag_list_lock); |
| 3942 | list_del(&q->tag_set_list); |
| 3943 | if (list_is_singular(&set->tag_list)) { |
| 3944 | /* just transitioned to unshared */ |
| 3945 | set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED; |
| 3946 | /* update existing queue */ |
| 3947 | blk_mq_update_tag_set_shared(set, false); |
| 3948 | } |
| 3949 | mutex_unlock(&set->tag_list_lock); |
| 3950 | INIT_LIST_HEAD(&q->tag_set_list); |
| 3951 | } |
| 3952 | |
| 3953 | static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, |
| 3954 | struct request_queue *q) |
| 3955 | { |
| 3956 | mutex_lock(&set->tag_list_lock); |
| 3957 | |
| 3958 | /* |
| 3959 | * Check to see if we're transitioning to shared (from 1 to 2 queues). |
| 3960 | */ |
| 3961 | if (!list_empty(&set->tag_list) && |
| 3962 | !(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) { |
| 3963 | set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED; |
| 3964 | /* update existing queue */ |
| 3965 | blk_mq_update_tag_set_shared(set, true); |
| 3966 | } |
| 3967 | if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED) |
| 3968 | queue_set_hctx_shared(q, true); |
| 3969 | list_add_tail(&q->tag_set_list, &set->tag_list); |
| 3970 | |
| 3971 | mutex_unlock(&set->tag_list_lock); |
| 3972 | } |
| 3973 | |
| 3974 | /* All allocations will be freed in release handler of q->mq_kobj */ |
| 3975 | static int blk_mq_alloc_ctxs(struct request_queue *q) |
| 3976 | { |
| 3977 | struct blk_mq_ctxs *ctxs; |
| 3978 | int cpu; |
| 3979 | |
| 3980 | ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL); |
| 3981 | if (!ctxs) |
| 3982 | return -ENOMEM; |
| 3983 | |
| 3984 | ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx); |
| 3985 | if (!ctxs->queue_ctx) |
| 3986 | goto fail; |
| 3987 | |
| 3988 | for_each_possible_cpu(cpu) { |
| 3989 | struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu); |
| 3990 | ctx->ctxs = ctxs; |
| 3991 | } |
| 3992 | |
| 3993 | q->mq_kobj = &ctxs->kobj; |
| 3994 | q->queue_ctx = ctxs->queue_ctx; |
| 3995 | |
| 3996 | return 0; |
| 3997 | fail: |
| 3998 | kfree(ctxs); |
| 3999 | return -ENOMEM; |
| 4000 | } |
| 4001 | |
| 4002 | /* |
| 4003 | * It is the actual release handler for mq, but we do it from |
| 4004 | * request queue's release handler for avoiding use-after-free |
| 4005 | * and headache because q->mq_kobj shouldn't have been introduced, |
| 4006 | * but we can't group ctx/kctx kobj without it. |
| 4007 | */ |
| 4008 | void blk_mq_release(struct request_queue *q) |
| 4009 | { |
| 4010 | struct blk_mq_hw_ctx *hctx, *next; |
| 4011 | unsigned long i; |
| 4012 | |
| 4013 | queue_for_each_hw_ctx(q, hctx, i) |
| 4014 | WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list)); |
| 4015 | |
| 4016 | /* all hctx are in .unused_hctx_list now */ |
| 4017 | list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) { |
| 4018 | list_del_init(&hctx->hctx_list); |
| 4019 | kobject_put(&hctx->kobj); |
| 4020 | } |
| 4021 | |
| 4022 | xa_destroy(&q->hctx_table); |
| 4023 | |
| 4024 | /* |
| 4025 | * release .mq_kobj and sw queue's kobject now because |
| 4026 | * both share lifetime with request queue. |
| 4027 | */ |
| 4028 | blk_mq_sysfs_deinit(q); |
| 4029 | } |
| 4030 | |
| 4031 | static struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set, |
| 4032 | void *queuedata) |
| 4033 | { |
| 4034 | struct request_queue *q; |
| 4035 | int ret; |
| 4036 | |
| 4037 | q = blk_alloc_queue(set->numa_node); |
| 4038 | if (!q) |
| 4039 | return ERR_PTR(-ENOMEM); |
| 4040 | q->queuedata = queuedata; |
| 4041 | ret = blk_mq_init_allocated_queue(set, q); |
| 4042 | if (ret) { |
| 4043 | blk_put_queue(q); |
| 4044 | return ERR_PTR(ret); |
| 4045 | } |
| 4046 | return q; |
| 4047 | } |
| 4048 | |
| 4049 | struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set) |
| 4050 | { |
| 4051 | return blk_mq_init_queue_data(set, NULL); |
| 4052 | } |
| 4053 | EXPORT_SYMBOL(blk_mq_init_queue); |
| 4054 | |
| 4055 | /** |
| 4056 | * blk_mq_destroy_queue - shutdown a request queue |
| 4057 | * @q: request queue to shutdown |
| 4058 | * |
| 4059 | * This shuts down a request queue allocated by blk_mq_init_queue(). All future |
| 4060 | * requests will be failed with -ENODEV. The caller is responsible for dropping |
| 4061 | * the reference from blk_mq_init_queue() by calling blk_put_queue(). |
| 4062 | * |
| 4063 | * Context: can sleep |
| 4064 | */ |
| 4065 | void blk_mq_destroy_queue(struct request_queue *q) |
| 4066 | { |
| 4067 | WARN_ON_ONCE(!queue_is_mq(q)); |
| 4068 | WARN_ON_ONCE(blk_queue_registered(q)); |
| 4069 | |
| 4070 | might_sleep(); |
| 4071 | |
| 4072 | blk_queue_flag_set(QUEUE_FLAG_DYING, q); |
| 4073 | blk_queue_start_drain(q); |
| 4074 | blk_mq_freeze_queue_wait(q); |
| 4075 | |
| 4076 | blk_sync_queue(q); |
| 4077 | blk_mq_cancel_work_sync(q); |
| 4078 | blk_mq_exit_queue(q); |
| 4079 | } |
| 4080 | EXPORT_SYMBOL(blk_mq_destroy_queue); |
| 4081 | |
| 4082 | struct gendisk *__blk_mq_alloc_disk(struct blk_mq_tag_set *set, void *queuedata, |
| 4083 | struct lock_class_key *lkclass) |
| 4084 | { |
| 4085 | struct request_queue *q; |
| 4086 | struct gendisk *disk; |
| 4087 | |
| 4088 | q = blk_mq_init_queue_data(set, queuedata); |
| 4089 | if (IS_ERR(q)) |
| 4090 | return ERR_CAST(q); |
| 4091 | |
| 4092 | disk = __alloc_disk_node(q, set->numa_node, lkclass); |
| 4093 | if (!disk) { |
| 4094 | blk_mq_destroy_queue(q); |
| 4095 | blk_put_queue(q); |
| 4096 | return ERR_PTR(-ENOMEM); |
| 4097 | } |
| 4098 | set_bit(GD_OWNS_QUEUE, &disk->state); |
| 4099 | return disk; |
| 4100 | } |
| 4101 | EXPORT_SYMBOL(__blk_mq_alloc_disk); |
| 4102 | |
| 4103 | struct gendisk *blk_mq_alloc_disk_for_queue(struct request_queue *q, |
| 4104 | struct lock_class_key *lkclass) |
| 4105 | { |
| 4106 | struct gendisk *disk; |
| 4107 | |
| 4108 | if (!blk_get_queue(q)) |
| 4109 | return NULL; |
| 4110 | disk = __alloc_disk_node(q, NUMA_NO_NODE, lkclass); |
| 4111 | if (!disk) |
| 4112 | blk_put_queue(q); |
| 4113 | return disk; |
| 4114 | } |
| 4115 | EXPORT_SYMBOL(blk_mq_alloc_disk_for_queue); |
| 4116 | |
| 4117 | static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx( |
| 4118 | struct blk_mq_tag_set *set, struct request_queue *q, |
| 4119 | int hctx_idx, int node) |
| 4120 | { |
| 4121 | struct blk_mq_hw_ctx *hctx = NULL, *tmp; |
| 4122 | |
| 4123 | /* reuse dead hctx first */ |
| 4124 | spin_lock(&q->unused_hctx_lock); |
| 4125 | list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) { |
| 4126 | if (tmp->numa_node == node) { |
| 4127 | hctx = tmp; |
| 4128 | break; |
| 4129 | } |
| 4130 | } |
| 4131 | if (hctx) |
| 4132 | list_del_init(&hctx->hctx_list); |
| 4133 | spin_unlock(&q->unused_hctx_lock); |
| 4134 | |
| 4135 | if (!hctx) |
| 4136 | hctx = blk_mq_alloc_hctx(q, set, node); |
| 4137 | if (!hctx) |
| 4138 | goto fail; |
| 4139 | |
| 4140 | if (blk_mq_init_hctx(q, set, hctx, hctx_idx)) |
| 4141 | goto free_hctx; |
| 4142 | |
| 4143 | return hctx; |
| 4144 | |
| 4145 | free_hctx: |
| 4146 | kobject_put(&hctx->kobj); |
| 4147 | fail: |
| 4148 | return NULL; |
| 4149 | } |
| 4150 | |
| 4151 | static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set, |
| 4152 | struct request_queue *q) |
| 4153 | { |
| 4154 | struct blk_mq_hw_ctx *hctx; |
| 4155 | unsigned long i, j; |
| 4156 | |
| 4157 | /* protect against switching io scheduler */ |
| 4158 | mutex_lock(&q->sysfs_lock); |
| 4159 | for (i = 0; i < set->nr_hw_queues; i++) { |
| 4160 | int old_node; |
| 4161 | int node = blk_mq_get_hctx_node(set, i); |
| 4162 | struct blk_mq_hw_ctx *old_hctx = xa_load(&q->hctx_table, i); |
| 4163 | |
| 4164 | if (old_hctx) { |
| 4165 | old_node = old_hctx->numa_node; |
| 4166 | blk_mq_exit_hctx(q, set, old_hctx, i); |
| 4167 | } |
| 4168 | |
| 4169 | if (!blk_mq_alloc_and_init_hctx(set, q, i, node)) { |
| 4170 | if (!old_hctx) |
| 4171 | break; |
| 4172 | pr_warn("Allocate new hctx on node %d fails, fallback to previous one on node %d\n", |
| 4173 | node, old_node); |
| 4174 | hctx = blk_mq_alloc_and_init_hctx(set, q, i, old_node); |
| 4175 | WARN_ON_ONCE(!hctx); |
| 4176 | } |
| 4177 | } |
| 4178 | /* |
| 4179 | * Increasing nr_hw_queues fails. Free the newly allocated |
| 4180 | * hctxs and keep the previous q->nr_hw_queues. |
| 4181 | */ |
| 4182 | if (i != set->nr_hw_queues) { |
| 4183 | j = q->nr_hw_queues; |
| 4184 | } else { |
| 4185 | j = i; |
| 4186 | q->nr_hw_queues = set->nr_hw_queues; |
| 4187 | } |
| 4188 | |
| 4189 | xa_for_each_start(&q->hctx_table, j, hctx, j) |
| 4190 | blk_mq_exit_hctx(q, set, hctx, j); |
| 4191 | mutex_unlock(&q->sysfs_lock); |
| 4192 | } |
| 4193 | |
| 4194 | static void blk_mq_update_poll_flag(struct request_queue *q) |
| 4195 | { |
| 4196 | struct blk_mq_tag_set *set = q->tag_set; |
| 4197 | |
| 4198 | if (set->nr_maps > HCTX_TYPE_POLL && |
| 4199 | set->map[HCTX_TYPE_POLL].nr_queues) |
| 4200 | blk_queue_flag_set(QUEUE_FLAG_POLL, q); |
| 4201 | else |
| 4202 | blk_queue_flag_clear(QUEUE_FLAG_POLL, q); |
| 4203 | } |
| 4204 | |
| 4205 | int blk_mq_init_allocated_queue(struct blk_mq_tag_set *set, |
| 4206 | struct request_queue *q) |
| 4207 | { |
| 4208 | /* mark the queue as mq asap */ |
| 4209 | q->mq_ops = set->ops; |
| 4210 | |
| 4211 | if (blk_mq_alloc_ctxs(q)) |
| 4212 | goto err_exit; |
| 4213 | |
| 4214 | /* init q->mq_kobj and sw queues' kobjects */ |
| 4215 | blk_mq_sysfs_init(q); |
| 4216 | |
| 4217 | INIT_LIST_HEAD(&q->unused_hctx_list); |
| 4218 | spin_lock_init(&q->unused_hctx_lock); |
| 4219 | |
| 4220 | xa_init(&q->hctx_table); |
| 4221 | |
| 4222 | blk_mq_realloc_hw_ctxs(set, q); |
| 4223 | if (!q->nr_hw_queues) |
| 4224 | goto err_hctxs; |
| 4225 | |
| 4226 | INIT_WORK(&q->timeout_work, blk_mq_timeout_work); |
| 4227 | blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ); |
| 4228 | |
| 4229 | q->tag_set = set; |
| 4230 | |
| 4231 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; |
| 4232 | blk_mq_update_poll_flag(q); |
| 4233 | |
| 4234 | INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work); |
| 4235 | INIT_LIST_HEAD(&q->requeue_list); |
| 4236 | spin_lock_init(&q->requeue_lock); |
| 4237 | |
| 4238 | q->nr_requests = set->queue_depth; |
| 4239 | |
| 4240 | blk_mq_init_cpu_queues(q, set->nr_hw_queues); |
| 4241 | blk_mq_add_queue_tag_set(set, q); |
| 4242 | blk_mq_map_swqueue(q); |
| 4243 | return 0; |
| 4244 | |
| 4245 | err_hctxs: |
| 4246 | blk_mq_release(q); |
| 4247 | err_exit: |
| 4248 | q->mq_ops = NULL; |
| 4249 | return -ENOMEM; |
| 4250 | } |
| 4251 | EXPORT_SYMBOL(blk_mq_init_allocated_queue); |
| 4252 | |
| 4253 | /* tags can _not_ be used after returning from blk_mq_exit_queue */ |
| 4254 | void blk_mq_exit_queue(struct request_queue *q) |
| 4255 | { |
| 4256 | struct blk_mq_tag_set *set = q->tag_set; |
| 4257 | |
| 4258 | /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */ |
| 4259 | blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); |
| 4260 | /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */ |
| 4261 | blk_mq_del_queue_tag_set(q); |
| 4262 | } |
| 4263 | |
| 4264 | static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) |
| 4265 | { |
| 4266 | int i; |
| 4267 | |
| 4268 | if (blk_mq_is_shared_tags(set->flags)) { |
| 4269 | set->shared_tags = blk_mq_alloc_map_and_rqs(set, |
| 4270 | BLK_MQ_NO_HCTX_IDX, |
| 4271 | set->queue_depth); |
| 4272 | if (!set->shared_tags) |
| 4273 | return -ENOMEM; |
| 4274 | } |
| 4275 | |
| 4276 | for (i = 0; i < set->nr_hw_queues; i++) { |
| 4277 | if (!__blk_mq_alloc_map_and_rqs(set, i)) |
| 4278 | goto out_unwind; |
| 4279 | cond_resched(); |
| 4280 | } |
| 4281 | |
| 4282 | return 0; |
| 4283 | |
| 4284 | out_unwind: |
| 4285 | while (--i >= 0) |
| 4286 | __blk_mq_free_map_and_rqs(set, i); |
| 4287 | |
| 4288 | if (blk_mq_is_shared_tags(set->flags)) { |
| 4289 | blk_mq_free_map_and_rqs(set, set->shared_tags, |
| 4290 | BLK_MQ_NO_HCTX_IDX); |
| 4291 | } |
| 4292 | |
| 4293 | return -ENOMEM; |
| 4294 | } |
| 4295 | |
| 4296 | /* |
| 4297 | * Allocate the request maps associated with this tag_set. Note that this |
| 4298 | * may reduce the depth asked for, if memory is tight. set->queue_depth |
| 4299 | * will be updated to reflect the allocated depth. |
| 4300 | */ |
| 4301 | static int blk_mq_alloc_set_map_and_rqs(struct blk_mq_tag_set *set) |
| 4302 | { |
| 4303 | unsigned int depth; |
| 4304 | int err; |
| 4305 | |
| 4306 | depth = set->queue_depth; |
| 4307 | do { |
| 4308 | err = __blk_mq_alloc_rq_maps(set); |
| 4309 | if (!err) |
| 4310 | break; |
| 4311 | |
| 4312 | set->queue_depth >>= 1; |
| 4313 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) { |
| 4314 | err = -ENOMEM; |
| 4315 | break; |
| 4316 | } |
| 4317 | } while (set->queue_depth); |
| 4318 | |
| 4319 | if (!set->queue_depth || err) { |
| 4320 | pr_err("blk-mq: failed to allocate request map\n"); |
| 4321 | return -ENOMEM; |
| 4322 | } |
| 4323 | |
| 4324 | if (depth != set->queue_depth) |
| 4325 | pr_info("blk-mq: reduced tag depth (%u -> %u)\n", |
| 4326 | depth, set->queue_depth); |
| 4327 | |
| 4328 | return 0; |
| 4329 | } |
| 4330 | |
| 4331 | static void blk_mq_update_queue_map(struct blk_mq_tag_set *set) |
| 4332 | { |
| 4333 | /* |
| 4334 | * blk_mq_map_queues() and multiple .map_queues() implementations |
| 4335 | * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the |
| 4336 | * number of hardware queues. |
| 4337 | */ |
| 4338 | if (set->nr_maps == 1) |
| 4339 | set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues; |
| 4340 | |
| 4341 | if (set->ops->map_queues && !is_kdump_kernel()) { |
| 4342 | int i; |
| 4343 | |
| 4344 | /* |
| 4345 | * transport .map_queues is usually done in the following |
| 4346 | * way: |
| 4347 | * |
| 4348 | * for (queue = 0; queue < set->nr_hw_queues; queue++) { |
| 4349 | * mask = get_cpu_mask(queue) |
| 4350 | * for_each_cpu(cpu, mask) |
| 4351 | * set->map[x].mq_map[cpu] = queue; |
| 4352 | * } |
| 4353 | * |
| 4354 | * When we need to remap, the table has to be cleared for |
| 4355 | * killing stale mapping since one CPU may not be mapped |
| 4356 | * to any hw queue. |
| 4357 | */ |
| 4358 | for (i = 0; i < set->nr_maps; i++) |
| 4359 | blk_mq_clear_mq_map(&set->map[i]); |
| 4360 | |
| 4361 | set->ops->map_queues(set); |
| 4362 | } else { |
| 4363 | BUG_ON(set->nr_maps > 1); |
| 4364 | blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]); |
| 4365 | } |
| 4366 | } |
| 4367 | |
| 4368 | static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set, |
| 4369 | int new_nr_hw_queues) |
| 4370 | { |
| 4371 | struct blk_mq_tags **new_tags; |
| 4372 | |
| 4373 | if (set->nr_hw_queues >= new_nr_hw_queues) |
| 4374 | goto done; |
| 4375 | |
| 4376 | new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *), |
| 4377 | GFP_KERNEL, set->numa_node); |
| 4378 | if (!new_tags) |
| 4379 | return -ENOMEM; |
| 4380 | |
| 4381 | if (set->tags) |
| 4382 | memcpy(new_tags, set->tags, set->nr_hw_queues * |
| 4383 | sizeof(*set->tags)); |
| 4384 | kfree(set->tags); |
| 4385 | set->tags = new_tags; |
| 4386 | done: |
| 4387 | set->nr_hw_queues = new_nr_hw_queues; |
| 4388 | return 0; |
| 4389 | } |
| 4390 | |
| 4391 | /* |
| 4392 | * Alloc a tag set to be associated with one or more request queues. |
| 4393 | * May fail with EINVAL for various error conditions. May adjust the |
| 4394 | * requested depth down, if it's too large. In that case, the set |
| 4395 | * value will be stored in set->queue_depth. |
| 4396 | */ |
| 4397 | int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) |
| 4398 | { |
| 4399 | int i, ret; |
| 4400 | |
| 4401 | BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS); |
| 4402 | |
| 4403 | if (!set->nr_hw_queues) |
| 4404 | return -EINVAL; |
| 4405 | if (!set->queue_depth) |
| 4406 | return -EINVAL; |
| 4407 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) |
| 4408 | return -EINVAL; |
| 4409 | |
| 4410 | if (!set->ops->queue_rq) |
| 4411 | return -EINVAL; |
| 4412 | |
| 4413 | if (!set->ops->get_budget ^ !set->ops->put_budget) |
| 4414 | return -EINVAL; |
| 4415 | |
| 4416 | if (set->queue_depth > BLK_MQ_MAX_DEPTH) { |
| 4417 | pr_info("blk-mq: reduced tag depth to %u\n", |
| 4418 | BLK_MQ_MAX_DEPTH); |
| 4419 | set->queue_depth = BLK_MQ_MAX_DEPTH; |
| 4420 | } |
| 4421 | |
| 4422 | if (!set->nr_maps) |
| 4423 | set->nr_maps = 1; |
| 4424 | else if (set->nr_maps > HCTX_MAX_TYPES) |
| 4425 | return -EINVAL; |
| 4426 | |
| 4427 | /* |
| 4428 | * If a crashdump is active, then we are potentially in a very |
| 4429 | * memory constrained environment. Limit us to 1 queue and |
| 4430 | * 64 tags to prevent using too much memory. |
| 4431 | */ |
| 4432 | if (is_kdump_kernel()) { |
| 4433 | set->nr_hw_queues = 1; |
| 4434 | set->nr_maps = 1; |
| 4435 | set->queue_depth = min(64U, set->queue_depth); |
| 4436 | } |
| 4437 | /* |
| 4438 | * There is no use for more h/w queues than cpus if we just have |
| 4439 | * a single map |
| 4440 | */ |
| 4441 | if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids) |
| 4442 | set->nr_hw_queues = nr_cpu_ids; |
| 4443 | |
| 4444 | if (set->flags & BLK_MQ_F_BLOCKING) { |
| 4445 | set->srcu = kmalloc(sizeof(*set->srcu), GFP_KERNEL); |
| 4446 | if (!set->srcu) |
| 4447 | return -ENOMEM; |
| 4448 | ret = init_srcu_struct(set->srcu); |
| 4449 | if (ret) |
| 4450 | goto out_free_srcu; |
| 4451 | } |
| 4452 | |
| 4453 | ret = -ENOMEM; |
| 4454 | set->tags = kcalloc_node(set->nr_hw_queues, |
| 4455 | sizeof(struct blk_mq_tags *), GFP_KERNEL, |
| 4456 | set->numa_node); |
| 4457 | if (!set->tags) |
| 4458 | goto out_cleanup_srcu; |
| 4459 | |
| 4460 | for (i = 0; i < set->nr_maps; i++) { |
| 4461 | set->map[i].mq_map = kcalloc_node(nr_cpu_ids, |
| 4462 | sizeof(set->map[i].mq_map[0]), |
| 4463 | GFP_KERNEL, set->numa_node); |
| 4464 | if (!set->map[i].mq_map) |
| 4465 | goto out_free_mq_map; |
| 4466 | set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues; |
| 4467 | } |
| 4468 | |
| 4469 | blk_mq_update_queue_map(set); |
| 4470 | |
| 4471 | ret = blk_mq_alloc_set_map_and_rqs(set); |
| 4472 | if (ret) |
| 4473 | goto out_free_mq_map; |
| 4474 | |
| 4475 | mutex_init(&set->tag_list_lock); |
| 4476 | INIT_LIST_HEAD(&set->tag_list); |
| 4477 | |
| 4478 | return 0; |
| 4479 | |
| 4480 | out_free_mq_map: |
| 4481 | for (i = 0; i < set->nr_maps; i++) { |
| 4482 | kfree(set->map[i].mq_map); |
| 4483 | set->map[i].mq_map = NULL; |
| 4484 | } |
| 4485 | kfree(set->tags); |
| 4486 | set->tags = NULL; |
| 4487 | out_cleanup_srcu: |
| 4488 | if (set->flags & BLK_MQ_F_BLOCKING) |
| 4489 | cleanup_srcu_struct(set->srcu); |
| 4490 | out_free_srcu: |
| 4491 | if (set->flags & BLK_MQ_F_BLOCKING) |
| 4492 | kfree(set->srcu); |
| 4493 | return ret; |
| 4494 | } |
| 4495 | EXPORT_SYMBOL(blk_mq_alloc_tag_set); |
| 4496 | |
| 4497 | /* allocate and initialize a tagset for a simple single-queue device */ |
| 4498 | int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set *set, |
| 4499 | const struct blk_mq_ops *ops, unsigned int queue_depth, |
| 4500 | unsigned int set_flags) |
| 4501 | { |
| 4502 | memset(set, 0, sizeof(*set)); |
| 4503 | set->ops = ops; |
| 4504 | set->nr_hw_queues = 1; |
| 4505 | set->nr_maps = 1; |
| 4506 | set->queue_depth = queue_depth; |
| 4507 | set->numa_node = NUMA_NO_NODE; |
| 4508 | set->flags = set_flags; |
| 4509 | return blk_mq_alloc_tag_set(set); |
| 4510 | } |
| 4511 | EXPORT_SYMBOL_GPL(blk_mq_alloc_sq_tag_set); |
| 4512 | |
| 4513 | void blk_mq_free_tag_set(struct blk_mq_tag_set *set) |
| 4514 | { |
| 4515 | int i, j; |
| 4516 | |
| 4517 | for (i = 0; i < set->nr_hw_queues; i++) |
| 4518 | __blk_mq_free_map_and_rqs(set, i); |
| 4519 | |
| 4520 | if (blk_mq_is_shared_tags(set->flags)) { |
| 4521 | blk_mq_free_map_and_rqs(set, set->shared_tags, |
| 4522 | BLK_MQ_NO_HCTX_IDX); |
| 4523 | } |
| 4524 | |
| 4525 | for (j = 0; j < set->nr_maps; j++) { |
| 4526 | kfree(set->map[j].mq_map); |
| 4527 | set->map[j].mq_map = NULL; |
| 4528 | } |
| 4529 | |
| 4530 | kfree(set->tags); |
| 4531 | set->tags = NULL; |
| 4532 | if (set->flags & BLK_MQ_F_BLOCKING) { |
| 4533 | cleanup_srcu_struct(set->srcu); |
| 4534 | kfree(set->srcu); |
| 4535 | } |
| 4536 | } |
| 4537 | EXPORT_SYMBOL(blk_mq_free_tag_set); |
| 4538 | |
| 4539 | int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr) |
| 4540 | { |
| 4541 | struct blk_mq_tag_set *set = q->tag_set; |
| 4542 | struct blk_mq_hw_ctx *hctx; |
| 4543 | int ret; |
| 4544 | unsigned long i; |
| 4545 | |
| 4546 | if (!set) |
| 4547 | return -EINVAL; |
| 4548 | |
| 4549 | if (q->nr_requests == nr) |
| 4550 | return 0; |
| 4551 | |
| 4552 | blk_mq_freeze_queue(q); |
| 4553 | blk_mq_quiesce_queue(q); |
| 4554 | |
| 4555 | ret = 0; |
| 4556 | queue_for_each_hw_ctx(q, hctx, i) { |
| 4557 | if (!hctx->tags) |
| 4558 | continue; |
| 4559 | /* |
| 4560 | * If we're using an MQ scheduler, just update the scheduler |
| 4561 | * queue depth. This is similar to what the old code would do. |
| 4562 | */ |
| 4563 | if (hctx->sched_tags) { |
| 4564 | ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags, |
| 4565 | nr, true); |
| 4566 | } else { |
| 4567 | ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr, |
| 4568 | false); |
| 4569 | } |
| 4570 | if (ret) |
| 4571 | break; |
| 4572 | if (q->elevator && q->elevator->type->ops.depth_updated) |
| 4573 | q->elevator->type->ops.depth_updated(hctx); |
| 4574 | } |
| 4575 | if (!ret) { |
| 4576 | q->nr_requests = nr; |
| 4577 | if (blk_mq_is_shared_tags(set->flags)) { |
| 4578 | if (q->elevator) |
| 4579 | blk_mq_tag_update_sched_shared_tags(q); |
| 4580 | else |
| 4581 | blk_mq_tag_resize_shared_tags(set, nr); |
| 4582 | } |
| 4583 | } |
| 4584 | |
| 4585 | blk_mq_unquiesce_queue(q); |
| 4586 | blk_mq_unfreeze_queue(q); |
| 4587 | |
| 4588 | return ret; |
| 4589 | } |
| 4590 | |
| 4591 | /* |
| 4592 | * request_queue and elevator_type pair. |
| 4593 | * It is just used by __blk_mq_update_nr_hw_queues to cache |
| 4594 | * the elevator_type associated with a request_queue. |
| 4595 | */ |
| 4596 | struct blk_mq_qe_pair { |
| 4597 | struct list_head node; |
| 4598 | struct request_queue *q; |
| 4599 | struct elevator_type *type; |
| 4600 | }; |
| 4601 | |
| 4602 | /* |
| 4603 | * Cache the elevator_type in qe pair list and switch the |
| 4604 | * io scheduler to 'none' |
| 4605 | */ |
| 4606 | static bool blk_mq_elv_switch_none(struct list_head *head, |
| 4607 | struct request_queue *q) |
| 4608 | { |
| 4609 | struct blk_mq_qe_pair *qe; |
| 4610 | |
| 4611 | if (!q->elevator) |
| 4612 | return true; |
| 4613 | |
| 4614 | qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY); |
| 4615 | if (!qe) |
| 4616 | return false; |
| 4617 | |
| 4618 | /* q->elevator needs protection from ->sysfs_lock */ |
| 4619 | mutex_lock(&q->sysfs_lock); |
| 4620 | |
| 4621 | INIT_LIST_HEAD(&qe->node); |
| 4622 | qe->q = q; |
| 4623 | qe->type = q->elevator->type; |
| 4624 | /* keep a reference to the elevator module as we'll switch back */ |
| 4625 | __elevator_get(qe->type); |
| 4626 | list_add(&qe->node, head); |
| 4627 | elevator_disable(q); |
| 4628 | mutex_unlock(&q->sysfs_lock); |
| 4629 | |
| 4630 | return true; |
| 4631 | } |
| 4632 | |
| 4633 | static struct blk_mq_qe_pair *blk_lookup_qe_pair(struct list_head *head, |
| 4634 | struct request_queue *q) |
| 4635 | { |
| 4636 | struct blk_mq_qe_pair *qe; |
| 4637 | |
| 4638 | list_for_each_entry(qe, head, node) |
| 4639 | if (qe->q == q) |
| 4640 | return qe; |
| 4641 | |
| 4642 | return NULL; |
| 4643 | } |
| 4644 | |
| 4645 | static void blk_mq_elv_switch_back(struct list_head *head, |
| 4646 | struct request_queue *q) |
| 4647 | { |
| 4648 | struct blk_mq_qe_pair *qe; |
| 4649 | struct elevator_type *t; |
| 4650 | |
| 4651 | qe = blk_lookup_qe_pair(head, q); |
| 4652 | if (!qe) |
| 4653 | return; |
| 4654 | t = qe->type; |
| 4655 | list_del(&qe->node); |
| 4656 | kfree(qe); |
| 4657 | |
| 4658 | mutex_lock(&q->sysfs_lock); |
| 4659 | elevator_switch(q, t); |
| 4660 | /* drop the reference acquired in blk_mq_elv_switch_none */ |
| 4661 | elevator_put(t); |
| 4662 | mutex_unlock(&q->sysfs_lock); |
| 4663 | } |
| 4664 | |
| 4665 | static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, |
| 4666 | int nr_hw_queues) |
| 4667 | { |
| 4668 | struct request_queue *q; |
| 4669 | LIST_HEAD(head); |
| 4670 | int prev_nr_hw_queues; |
| 4671 | |
| 4672 | lockdep_assert_held(&set->tag_list_lock); |
| 4673 | |
| 4674 | if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids) |
| 4675 | nr_hw_queues = nr_cpu_ids; |
| 4676 | if (nr_hw_queues < 1) |
| 4677 | return; |
| 4678 | if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues) |
| 4679 | return; |
| 4680 | |
| 4681 | list_for_each_entry(q, &set->tag_list, tag_set_list) |
| 4682 | blk_mq_freeze_queue(q); |
| 4683 | /* |
| 4684 | * Switch IO scheduler to 'none', cleaning up the data associated |
| 4685 | * with the previous scheduler. We will switch back once we are done |
| 4686 | * updating the new sw to hw queue mappings. |
| 4687 | */ |
| 4688 | list_for_each_entry(q, &set->tag_list, tag_set_list) |
| 4689 | if (!blk_mq_elv_switch_none(&head, q)) |
| 4690 | goto switch_back; |
| 4691 | |
| 4692 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
| 4693 | blk_mq_debugfs_unregister_hctxs(q); |
| 4694 | blk_mq_sysfs_unregister_hctxs(q); |
| 4695 | } |
| 4696 | |
| 4697 | prev_nr_hw_queues = set->nr_hw_queues; |
| 4698 | if (blk_mq_realloc_tag_set_tags(set, nr_hw_queues) < 0) |
| 4699 | goto reregister; |
| 4700 | |
| 4701 | fallback: |
| 4702 | blk_mq_update_queue_map(set); |
| 4703 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
| 4704 | blk_mq_realloc_hw_ctxs(set, q); |
| 4705 | blk_mq_update_poll_flag(q); |
| 4706 | if (q->nr_hw_queues != set->nr_hw_queues) { |
| 4707 | int i = prev_nr_hw_queues; |
| 4708 | |
| 4709 | pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n", |
| 4710 | nr_hw_queues, prev_nr_hw_queues); |
| 4711 | for (; i < set->nr_hw_queues; i++) |
| 4712 | __blk_mq_free_map_and_rqs(set, i); |
| 4713 | |
| 4714 | set->nr_hw_queues = prev_nr_hw_queues; |
| 4715 | blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]); |
| 4716 | goto fallback; |
| 4717 | } |
| 4718 | blk_mq_map_swqueue(q); |
| 4719 | } |
| 4720 | |
| 4721 | reregister: |
| 4722 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
| 4723 | blk_mq_sysfs_register_hctxs(q); |
| 4724 | blk_mq_debugfs_register_hctxs(q); |
| 4725 | } |
| 4726 | |
| 4727 | switch_back: |
| 4728 | list_for_each_entry(q, &set->tag_list, tag_set_list) |
| 4729 | blk_mq_elv_switch_back(&head, q); |
| 4730 | |
| 4731 | list_for_each_entry(q, &set->tag_list, tag_set_list) |
| 4732 | blk_mq_unfreeze_queue(q); |
| 4733 | } |
| 4734 | |
| 4735 | void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues) |
| 4736 | { |
| 4737 | mutex_lock(&set->tag_list_lock); |
| 4738 | __blk_mq_update_nr_hw_queues(set, nr_hw_queues); |
| 4739 | mutex_unlock(&set->tag_list_lock); |
| 4740 | } |
| 4741 | EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues); |
| 4742 | |
| 4743 | int blk_mq_poll(struct request_queue *q, blk_qc_t cookie, struct io_comp_batch *iob, |
| 4744 | unsigned int flags) |
| 4745 | { |
| 4746 | struct blk_mq_hw_ctx *hctx = blk_qc_to_hctx(q, cookie); |
| 4747 | long state = get_current_state(); |
| 4748 | int ret; |
| 4749 | |
| 4750 | do { |
| 4751 | ret = q->mq_ops->poll(hctx, iob); |
| 4752 | if (ret > 0) { |
| 4753 | __set_current_state(TASK_RUNNING); |
| 4754 | return ret; |
| 4755 | } |
| 4756 | |
| 4757 | if (signal_pending_state(state, current)) |
| 4758 | __set_current_state(TASK_RUNNING); |
| 4759 | if (task_is_running(current)) |
| 4760 | return 1; |
| 4761 | |
| 4762 | if (ret < 0 || (flags & BLK_POLL_ONESHOT)) |
| 4763 | break; |
| 4764 | cpu_relax(); |
| 4765 | } while (!need_resched()); |
| 4766 | |
| 4767 | __set_current_state(TASK_RUNNING); |
| 4768 | return 0; |
| 4769 | } |
| 4770 | |
| 4771 | unsigned int blk_mq_rq_cpu(struct request *rq) |
| 4772 | { |
| 4773 | return rq->mq_ctx->cpu; |
| 4774 | } |
| 4775 | EXPORT_SYMBOL(blk_mq_rq_cpu); |
| 4776 | |
| 4777 | void blk_mq_cancel_work_sync(struct request_queue *q) |
| 4778 | { |
| 4779 | struct blk_mq_hw_ctx *hctx; |
| 4780 | unsigned long i; |
| 4781 | |
| 4782 | cancel_delayed_work_sync(&q->requeue_work); |
| 4783 | |
| 4784 | queue_for_each_hw_ctx(q, hctx, i) |
| 4785 | cancel_delayed_work_sync(&hctx->run_work); |
| 4786 | } |
| 4787 | |
| 4788 | static int __init blk_mq_init(void) |
| 4789 | { |
| 4790 | int i; |
| 4791 | |
| 4792 | for_each_possible_cpu(i) |
| 4793 | init_llist_head(&per_cpu(blk_cpu_done, i)); |
| 4794 | open_softirq(BLOCK_SOFTIRQ, blk_done_softirq); |
| 4795 | |
| 4796 | cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD, |
| 4797 | "block/softirq:dead", NULL, |
| 4798 | blk_softirq_cpu_dead); |
| 4799 | cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL, |
| 4800 | blk_mq_hctx_notify_dead); |
| 4801 | cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online", |
| 4802 | blk_mq_hctx_notify_online, |
| 4803 | blk_mq_hctx_notify_offline); |
| 4804 | return 0; |
| 4805 | } |
| 4806 | subsys_initcall(blk_mq_init); |