block/blk-mq.c

   1 /*
   2  * Block multiqueue core code
   3  *
   4  * Copyright (C) 2013-2014 Jens Axboe
   5  * Copyright (C) 2013-2014 Christoph Hellwig
   6  */
   7 #include <linux/kernel.h>
   8 #include <linux/module.h>
   9 #include <linux/backing-dev.h>
  10 #include <linux/bio.h>
  11 #include <linux/blkdev.h>
  12 #include <linux/kmemleak.h>
  13 #include <linux/mm.h>
  14 #include <linux/init.h>
  15 #include <linux/slab.h>
  16 #include <linux/workqueue.h>
  17 #include <linux/smp.h>
  18 #include <linux/llist.h>
  19 #include <linux/list_sort.h>
  20 #include <linux/cpu.h>
  21 #include <linux/cache.h>
  22 #include <linux/sched/sysctl.h>
  23 #include <linux/sched/topology.h>
  24 #include <linux/sched/signal.h>
  25 #include <linux/delay.h>
  26 #include <linux/crash_dump.h>
  27 #include <linux/prefetch.h>
  28
  29 #include <trace/events/block.h>
  30
  31 #include <linux/blk-mq.h>
  32 #include "blk.h"
  33 #include "blk-mq.h"
  34 #include "blk-mq-debugfs.h"
  35 #include "blk-mq-tag.h"
  36 #include "blk-stat.h"
  37 #include "blk-wbt.h"
  38 #include "blk-mq-sched.h"
  39
  40 static void blk_mq_poll_stats_start(struct request_queue *q);
  41 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb);
  42
  43 static int blk_mq_poll_stats_bkt(const struct request *rq)
  44 {
  45         int ddir, bytes, bucket;
  46
  47         ddir = rq_data_dir(rq);
  48         bytes = blk_rq_bytes(rq);
  49
  50         bucket = ddir + 2*(ilog2(bytes) - 9);
  51
  52         if (bucket < 0)
  53                 return -1;
  54         else if (bucket >= BLK_MQ_POLL_STATS_BKTS)
  55                 return ddir + BLK_MQ_POLL_STATS_BKTS - 2;
  56
  57         return bucket;
  58 }
  59
  60 /*
  61  * Check if any of the ctx's have pending work in this hardware queue
  62  */
  63 bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
  64 {
  65         return sbitmap_any_bit_set(&hctx->ctx_map) ||
  66                         !list_empty_careful(&hctx->dispatch) ||
  67                         blk_mq_sched_has_work(hctx);
  68 }
  69
  70 /*
  71  * Mark this ctx as having pending work in this hardware queue
  72  */
  73 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
  74                                      struct blk_mq_ctx *ctx)
  75 {
  76         if (!sbitmap_test_bit(&hctx->ctx_map, ctx->index_hw))
  77                 sbitmap_set_bit(&hctx->ctx_map, ctx->index_hw);
  78 }
  79
  80 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
  81                                       struct blk_mq_ctx *ctx)
  82 {
  83         sbitmap_clear_bit(&hctx->ctx_map, ctx->index_hw);
  84 }
  85
  86 struct mq_inflight {
  87         struct hd_struct *part;
  88         unsigned int *inflight;
  89 };
  90
  91 static void blk_mq_check_inflight(struct blk_mq_hw_ctx *hctx,
  92                                   struct request *rq, void *priv,
  93                                   bool reserved)
  94 {
  95         struct mq_inflight *mi = priv;
  96
  97         if (test_bit(REQ_ATOM_STARTED, &rq->atomic_flags) &&
  98             !test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags)) {
  99                 /*
 100                  * index[0] counts the specific partition that was asked
 101                  * for. index[1] counts the ones that are active on the
 102                  * whole device, so increment that if mi->part is indeed
 103                  * a partition, and not a whole device.
 104                  */
 105                 if (rq->part == mi->part)
 106                         mi->inflight[0]++;
 107                 if (mi->part->partno)
 108                         mi->inflight[1]++;
 109         }
 110 }
 111
 112 void blk_mq_in_flight(struct request_queue *q, struct hd_struct *part,
 113                       unsigned int inflight[2])
 114 {
 115         struct mq_inflight mi = { .part = part, .inflight = inflight, };
 116
 117         inflight[0] = inflight[1] = 0;
 118         blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
 119 }
 120
 121 void blk_freeze_queue_start(struct request_queue *q)
 122 {
 123         int freeze_depth;
 124
 125         freeze_depth = atomic_inc_return(&q->mq_freeze_depth);
 126         if (freeze_depth == 1) {
 127                 percpu_ref_kill(&q->q_usage_counter);
 128                 blk_mq_run_hw_queues(q, false);
 129         }
 130 }
 131 EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
 132
 133 void blk_mq_freeze_queue_wait(struct request_queue *q)
 134 {
 135         wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
 136 }
 137 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
 138
 139 int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
 140                                      unsigned long timeout)
 141 {
 142         return wait_event_timeout(q->mq_freeze_wq,
 143                                         percpu_ref_is_zero(&q->q_usage_counter),
 144                                         timeout);
 145 }
 146 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
 147
 148 /*
 149  * Guarantee no request is in use, so we can change any data structure of
 150  * the queue afterward.
 151  */
 152 void blk_freeze_queue(struct request_queue *q)
 153 {
 154         /*
 155          * In the !blk_mq case we are only calling this to kill the
 156          * q_usage_counter, otherwise this increases the freeze depth
 157          * and waits for it to return to zero.  For this reason there is
 158          * no blk_unfreeze_queue(), and blk_freeze_queue() is not
 159          * exported to drivers as the only user for unfreeze is blk_mq.
 160          */
 161         blk_freeze_queue_start(q);
 162         blk_mq_freeze_queue_wait(q);
 163 }
 164
 165 void blk_mq_freeze_queue(struct request_queue *q)
 166 {
 167         /*
 168          * ...just an alias to keep freeze and unfreeze actions balanced
 169          * in the blk_mq_* namespace
 170          */
 171         blk_freeze_queue(q);
 172 }
 173 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
 174
 175 void blk_mq_unfreeze_queue(struct request_queue *q)
 176 {
 177         int freeze_depth;
 178
 179         freeze_depth = atomic_dec_return(&q->mq_freeze_depth);
 180         WARN_ON_ONCE(freeze_depth < 0);
 181         if (!freeze_depth) {
 182                 percpu_ref_reinit(&q->q_usage_counter);
 183                 wake_up_all(&q->mq_freeze_wq);
 184         }
 185 }
 186 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
 187
 188 /*
 189  * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
 190  * mpt3sas driver such that this function can be removed.
 191  */
 192 void blk_mq_quiesce_queue_nowait(struct request_queue *q)
 193 {
 194         unsigned long flags;
 195
 196         spin_lock_irqsave(q->queue_lock, flags);
 197         queue_flag_set(QUEUE_FLAG_QUIESCED, q);
 198         spin_unlock_irqrestore(q->queue_lock, flags);
 199 }
 200 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
 201
 202 /**
 203  * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
 204  * @q: request queue.
 205  *
 206  * Note: this function does not prevent that the struct request end_io()
 207  * callback function is invoked. Once this function is returned, we make
 208  * sure no dispatch can happen until the queue is unquiesced via
 209  * blk_mq_unquiesce_queue().
 210  */
 211 void blk_mq_quiesce_queue(struct request_queue *q)
 212 {
 213         struct blk_mq_hw_ctx *hctx;
 214         unsigned int i;
 215         bool rcu = false;
 216
 217         blk_mq_quiesce_queue_nowait(q);
 218
 219         queue_for_each_hw_ctx(q, hctx, i) {
 220                 if (hctx->flags & BLK_MQ_F_BLOCKING)
 221                         synchronize_srcu(hctx->queue_rq_srcu);
 222                 else
 223                         rcu = true;
 224         }
 225         if (rcu)
 226                 synchronize_rcu();
 227 }
 228 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
 229
 230 /*
 231  * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
 232  * @q: request queue.
 233  *
 234  * This function recovers queue into the state before quiescing
 235  * which is done by blk_mq_quiesce_queue.
 236  */
 237 void blk_mq_unquiesce_queue(struct request_queue *q)
 238 {
 239         unsigned long flags;
 240
 241         spin_lock_irqsave(q->queue_lock, flags);
 242         queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
 243         spin_unlock_irqrestore(q->queue_lock, flags);
 244
 245         /* dispatch requests which are inserted during quiescing */
 246         blk_mq_run_hw_queues(q, true);
 247 }
 248 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
 249
 250 void blk_mq_wake_waiters(struct request_queue *q)
 251 {
 252         struct blk_mq_hw_ctx *hctx;
 253         unsigned int i;
 254
 255         queue_for_each_hw_ctx(q, hctx, i)
 256                 if (blk_mq_hw_queue_mapped(hctx))
 257                         blk_mq_tag_wakeup_all(hctx->tags, true);
 258
 259         /*
 260          * If we are called because the queue has now been marked as
 261          * dying, we need to ensure that processes currently waiting on
 262          * the queue are notified as well.
 263          */
 264         wake_up_all(&q->mq_freeze_wq);
 265 }
 266
 267 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
 268 {
 269         return blk_mq_has_free_tags(hctx->tags);
 270 }
 271 EXPORT_SYMBOL(blk_mq_can_queue);
 272
 273 static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
 274                 unsigned int tag, unsigned int op)
 275 {
 276         struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
 277         struct request *rq = tags->static_rqs[tag];
 278
 279         rq->rq_flags = 0;
 280
 281         if (data->flags & BLK_MQ_REQ_INTERNAL) {
 282                 rq->tag = -1;
 283                 rq->internal_tag = tag;
 284         } else {
 285                 if (blk_mq_tag_busy(data->hctx)) {
 286                         rq->rq_flags = RQF_MQ_INFLIGHT;
 287                         atomic_inc(&data->hctx->nr_active);
 288                 }
 289                 rq->tag = tag;
 290                 rq->internal_tag = -1;
 291                 data->hctx->tags->rqs[rq->tag] = rq;
 292         }
 293
 294         INIT_LIST_HEAD(&rq->queuelist);
 295         /* csd/requeue_work/fifo_time is initialized before use */
 296         rq->q = data->q;
 297         rq->mq_ctx = data->ctx;
 298         rq->cmd_flags = op;
 299         if (blk_queue_io_stat(data->q))
 300                 rq->rq_flags |= RQF_IO_STAT;
 301         /* do not touch atomic flags, it needs atomic ops against the timer */
 302         rq->cpu = -1;
 303         INIT_HLIST_NODE(&rq->hash);
 304         RB_CLEAR_NODE(&rq->rb_node);
 305         rq->rq_disk = NULL;
 306         rq->part = NULL;
 307         rq->start_time = jiffies;
 308 #ifdef CONFIG_BLK_CGROUP
 309         rq->rl = NULL;
 310         set_start_time_ns(rq);
 311         rq->io_start_time_ns = 0;
 312 #endif
 313         rq->nr_phys_segments = 0;
 314 #if defined(CONFIG_BLK_DEV_INTEGRITY)
 315         rq->nr_integrity_segments = 0;
 316 #endif
 317         rq->special = NULL;
 318         /* tag was already set */
 319         rq->extra_len = 0;
 320
 321         INIT_LIST_HEAD(&rq->timeout_list);
 322         rq->timeout = 0;
 323
 324         rq->end_io = NULL;
 325         rq->end_io_data = NULL;
 326         rq->next_rq = NULL;
 327
 328         data->ctx->rq_dispatched[op_is_sync(op)]++;
 329         return rq;
 330 }
 331
 332 static struct request *blk_mq_get_request(struct request_queue *q,
 333                 struct bio *bio, unsigned int op,
 334                 struct blk_mq_alloc_data *data)
 335 {
 336         struct elevator_queue *e = q->elevator;
 337         struct request *rq;
 338         unsigned int tag;
 339         struct blk_mq_ctx *local_ctx = NULL;
 340
 341         blk_queue_enter_live(q);
 342         data->q = q;
 343         if (likely(!data->ctx))
 344                 data->ctx = local_ctx = blk_mq_get_ctx(q);
 345         if (likely(!data->hctx))
 346                 data->hctx = blk_mq_map_queue(q, data->ctx->cpu);
 347         if (op & REQ_NOWAIT)
 348                 data->flags |= BLK_MQ_REQ_NOWAIT;
 349
 350         if (e) {
 351                 data->flags |= BLK_MQ_REQ_INTERNAL;
 352
 353                 /*
 354                  * Flush requests are special and go directly to the
 355                  * dispatch list.
 356                  */
 357                 if (!op_is_flush(op) && e->type->ops.mq.limit_depth)
 358                         e->type->ops.mq.limit_depth(op, data);
 359         }
 360
 361         tag = blk_mq_get_tag(data);
 362         if (tag == BLK_MQ_TAG_FAIL) {
 363                 if (local_ctx) {
 364                         blk_mq_put_ctx(local_ctx);
 365                         data->ctx = NULL;
 366                 }
 367                 blk_queue_exit(q);
 368                 return NULL;
 369         }
 370
 371         rq = blk_mq_rq_ctx_init(data, tag, op);
 372         if (!op_is_flush(op)) {
 373                 rq->elv.icq = NULL;
 374                 if (e && e->type->ops.mq.prepare_request) {
 375                         if (e->type->icq_cache && rq_ioc(bio))
 376                                 blk_mq_sched_assign_ioc(rq, bio);
 377
 378                         e->type->ops.mq.prepare_request(rq, bio);
 379                         rq->rq_flags |= RQF_ELVPRIV;
 380                 }
 381         }
 382         data->hctx->queued++;
 383         return rq;
 384 }
 385
 386 struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op,
 387                 unsigned int flags)
 388 {
 389         struct blk_mq_alloc_data alloc_data = { .flags = flags };
 390         struct request *rq;
 391         int ret;
 392
 393         ret = blk_queue_enter(q, flags & BLK_MQ_REQ_NOWAIT);
 394         if (ret)
 395                 return ERR_PTR(ret);
 396
 397         rq = blk_mq_get_request(q, NULL, op, &alloc_data);
 398         blk_queue_exit(q);
 399
 400         if (!rq)
 401                 return ERR_PTR(-EWOULDBLOCK);
 402
 403         blk_mq_put_ctx(alloc_data.ctx);
 404
 405         rq->__data_len = 0;
 406         rq->__sector = (sector_t) -1;
 407         rq->bio = rq->biotail = NULL;
 408         return rq;
 409 }
 410 EXPORT_SYMBOL(blk_mq_alloc_request);
 411
 412 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
 413                 unsigned int op, unsigned int flags, unsigned int hctx_idx)
 414 {
 415         struct blk_mq_alloc_data alloc_data = { .flags = flags };
 416         struct request *rq;
 417         unsigned int cpu;
 418         int ret;
 419
 420         /*
 421          * If the tag allocator sleeps we could get an allocation for a
 422          * different hardware context.  No need to complicate the low level
 423          * allocator for this for the rare use case of a command tied to
 424          * a specific queue.
 425          */
 426         if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)))
 427                 return ERR_PTR(-EINVAL);
 428
 429         if (hctx_idx >= q->nr_hw_queues)
 430                 return ERR_PTR(-EIO);
 431
 432         ret = blk_queue_enter(q, true);
 433         if (ret)
 434                 return ERR_PTR(ret);
 435
 436         /*
 437          * Check if the hardware context is actually mapped to anything.
 438          * If not tell the caller that it should skip this queue.
 439          */
 440         alloc_data.hctx = q->queue_hw_ctx[hctx_idx];
 441         if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) {
 442                 blk_queue_exit(q);
 443                 return ERR_PTR(-EXDEV);
 444         }
 445         cpu = cpumask_first(alloc_data.hctx->cpumask);
 446         alloc_data.ctx = __blk_mq_get_ctx(q, cpu);
 447
 448         rq = blk_mq_get_request(q, NULL, op, &alloc_data);
 449         blk_queue_exit(q);
 450
 451         if (!rq)
 452                 return ERR_PTR(-EWOULDBLOCK);
 453
 454         return rq;
 455 }
 456 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
 457
 458 void blk_mq_free_request(struct request *rq)
 459 {
 460         struct request_queue *q = rq->q;
 461         struct elevator_queue *e = q->elevator;
 462         struct blk_mq_ctx *ctx = rq->mq_ctx;
 463         struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
 464         const int sched_tag = rq->internal_tag;
 465
 466         if (rq->rq_flags & RQF_ELVPRIV) {
 467                 if (e && e->type->ops.mq.finish_request)
 468                         e->type->ops.mq.finish_request(rq);
 469                 if (rq->elv.icq) {
 470                         put_io_context(rq->elv.icq->ioc);
 471                         rq->elv.icq = NULL;
 472                 }
 473         }
 474
 475         ctx->rq_completed[rq_is_sync(rq)]++;
 476         if (rq->rq_flags & RQF_MQ_INFLIGHT)
 477                 atomic_dec(&hctx->nr_active);
 478
 479         if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
 480                 laptop_io_completion(q->backing_dev_info);
 481
 482         wbt_done(q->rq_wb, &rq->issue_stat);
 483
 484         if (blk_rq_rl(rq))
 485                 blk_put_rl(blk_rq_rl(rq));
 486
 487         clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
 488         clear_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags);
 489         if (rq->tag != -1)
 490                 blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag);
 491         if (sched_tag != -1)
 492                 blk_mq_put_tag(hctx, hctx->sched_tags, ctx, sched_tag);
 493         blk_mq_sched_restart(hctx);
 494         blk_queue_exit(q);
 495 }
 496 EXPORT_SYMBOL_GPL(blk_mq_free_request);
 497
 498 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
 499 {
 500         blk_account_io_done(rq);
 501
 502         if (rq->end_io) {
 503                 wbt_done(rq->q->rq_wb, &rq->issue_stat);
 504                 rq->end_io(rq, error);
 505         } else {
 506                 if (unlikely(blk_bidi_rq(rq)))
 507                         blk_mq_free_request(rq->next_rq);
 508                 blk_mq_free_request(rq);
 509         }
 510 }
 511 EXPORT_SYMBOL(__blk_mq_end_request);
 512
 513 void blk_mq_end_request(struct request *rq, blk_status_t error)
 514 {
 515         if (blk_update_request(rq, error, blk_rq_bytes(rq)))
 516                 BUG();
 517         __blk_mq_end_request(rq, error);
 518 }
 519 EXPORT_SYMBOL(blk_mq_end_request);
 520
 521 static void __blk_mq_complete_request_remote(void *data)
 522 {
 523         struct request *rq = data;
 524
 525         rq->q->softirq_done_fn(rq);
 526 }
 527
 528 static void __blk_mq_complete_request(struct request *rq)
 529 {
 530         struct blk_mq_ctx *ctx = rq->mq_ctx;
 531         bool shared = false;
 532         int cpu;
 533
 534         if (rq->internal_tag != -1)
 535                 blk_mq_sched_completed_request(rq);
 536         if (rq->rq_flags & RQF_STATS) {
 537                 blk_mq_poll_stats_start(rq->q);
 538                 blk_stat_add(rq);
 539         }
 540
 541         if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
 542                 rq->q->softirq_done_fn(rq);
 543                 return;
 544         }
 545
 546         cpu = get_cpu();
 547         if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
 548                 shared = cpus_share_cache(cpu, ctx->cpu);
 549
 550         if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
 551                 rq->csd.func = __blk_mq_complete_request_remote;
 552                 rq->csd.info = rq;
 553                 rq->csd.flags = 0;
 554                 smp_call_function_single_async(ctx->cpu, &rq->csd);
 555         } else {
 556                 rq->q->softirq_done_fn(rq);
 557         }
 558         put_cpu();
 559 }
 560
 561 /**
 562  * blk_mq_complete_request - end I/O on a request
 563  * @rq:         the request being processed
 564  *
 565  * Description:
 566  *      Ends all I/O on a request. It does not handle partial completions.
 567  *      The actual completion happens out-of-order, through a IPI handler.
 568  **/
 569 void blk_mq_complete_request(struct request *rq)
 570 {
 571         struct request_queue *q = rq->q;
 572
 573         if (unlikely(blk_should_fake_timeout(q)))
 574                 return;
 575         if (!blk_mark_rq_complete(rq))
 576                 __blk_mq_complete_request(rq);
 577 }
 578 EXPORT_SYMBOL(blk_mq_complete_request);
 579
 580 int blk_mq_request_started(struct request *rq)
 581 {
 582         return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
 583 }
 584 EXPORT_SYMBOL_GPL(blk_mq_request_started);
 585
 586 void blk_mq_start_request(struct request *rq)
 587 {
 588         struct request_queue *q = rq->q;
 589
 590         blk_mq_sched_started_request(rq);
 591
 592         trace_block_rq_issue(q, rq);
 593
 594         if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
 595                 blk_stat_set_issue(&rq->issue_stat, blk_rq_sectors(rq));
 596                 rq->rq_flags |= RQF_STATS;
 597                 wbt_issue(q->rq_wb, &rq->issue_stat);
 598         }
 599
 600         blk_add_timer(rq);
 601
 602         WARN_ON_ONCE(test_bit(REQ_ATOM_STARTED, &rq->atomic_flags));
 603
 604         /*
 605          * Mark us as started and clear complete. Complete might have been
 606          * set if requeue raced with timeout, which then marked it as
 607          * complete. So be sure to clear complete again when we start
 608          * the request, otherwise we'll ignore the completion event.
 609          *
 610          * Ensure that ->deadline is visible before we set STARTED, such that
 611          * blk_mq_check_expired() is guaranteed to observe our ->deadline when
 612          * it observes STARTED.
 613          */
 614         smp_wmb();
 615         set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
 616         if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags)) {
 617                 /*
 618                  * Coherence order guarantees these consecutive stores to a
 619                  * single variable propagate in the specified order. Thus the
 620                  * clear_bit() is ordered _after_ the set bit. See
 621                  * blk_mq_check_expired().
 622                  *
 623                  * (the bits must be part of the same byte for this to be
 624                  * true).
 625                  */
 626                 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
 627         }
 628
 629         if (q->dma_drain_size && blk_rq_bytes(rq)) {
 630                 /*
 631                  * Make sure space for the drain appears.  We know we can do
 632                  * this because max_hw_segments has been adjusted to be one
 633                  * fewer than the device can handle.
 634                  */
 635                 rq->nr_phys_segments++;
 636         }
 637 }
 638 EXPORT_SYMBOL(blk_mq_start_request);
 639
 640 /*
 641  * When we reach here because queue is busy, REQ_ATOM_COMPLETE
 642  * flag isn't set yet, so there may be race with timeout handler,
 643  * but given rq->deadline is just set in .queue_rq() under
 644  * this situation, the race won't be possible in reality because
 645  * rq->timeout should be set as big enough to cover the window
 646  * between blk_mq_start_request() called from .queue_rq() and
 647  * clearing REQ_ATOM_STARTED here.
 648  */
 649 static void __blk_mq_requeue_request(struct request *rq)
 650 {
 651         struct request_queue *q = rq->q;
 652
 653         trace_block_rq_requeue(q, rq);
 654         wbt_requeue(q->rq_wb, &rq->issue_stat);
 655         blk_mq_sched_requeue_request(rq);
 656
 657         if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
 658                 if (q->dma_drain_size && blk_rq_bytes(rq))
 659                         rq->nr_phys_segments--;
 660         }
 661 }
 662
 663 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
 664 {
 665         __blk_mq_requeue_request(rq);
 666
 667         BUG_ON(blk_queued_rq(rq));
 668         blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
 669 }
 670 EXPORT_SYMBOL(blk_mq_requeue_request);
 671
 672 static void blk_mq_requeue_work(struct work_struct *work)
 673 {
 674         struct request_queue *q =
 675                 container_of(work, struct request_queue, requeue_work.work);
 676         LIST_HEAD(rq_list);
 677         struct request *rq, *next;
 678
 679         spin_lock_irq(&q->requeue_lock);
 680         list_splice_init(&q->requeue_list, &rq_list);
 681         spin_unlock_irq(&q->requeue_lock);
 682
 683         list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
 684                 if (!(rq->rq_flags & RQF_SOFTBARRIER))
 685                         continue;
 686
 687                 rq->rq_flags &= ~RQF_SOFTBARRIER;
 688                 list_del_init(&rq->queuelist);
 689                 blk_mq_sched_insert_request(rq, true, false, false, true);
 690         }
 691
 692         while (!list_empty(&rq_list)) {
 693                 rq = list_entry(rq_list.next, struct request, queuelist);
 694                 list_del_init(&rq->queuelist);
 695                 blk_mq_sched_insert_request(rq, false, false, false, true);
 696         }
 697
 698         blk_mq_run_hw_queues(q, false);
 699 }
 700
 701 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
 702                                 bool kick_requeue_list)
 703 {
 704         struct request_queue *q = rq->q;
 705         unsigned long flags;
 706
 707         /*
 708          * We abuse this flag that is otherwise used by the I/O scheduler to
 709          * request head insertation from the workqueue.
 710          */
 711         BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
 712
 713         spin_lock_irqsave(&q->requeue_lock, flags);
 714         if (at_head) {
 715                 rq->rq_flags |= RQF_SOFTBARRIER;
 716                 list_add(&rq->queuelist, &q->requeue_list);
 717         } else {
 718                 list_add_tail(&rq->queuelist, &q->requeue_list);
 719         }
 720         spin_unlock_irqrestore(&q->requeue_lock, flags);
 721
 722         if (kick_requeue_list)
 723                 blk_mq_kick_requeue_list(q);
 724 }
 725 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
 726
 727 void blk_mq_kick_requeue_list(struct request_queue *q)
 728 {
 729         kblockd_schedule_delayed_work(&q->requeue_work, 0);
 730 }
 731 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
 732
 733 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
 734                                     unsigned long msecs)
 735 {
 736         kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
 737                                     msecs_to_jiffies(msecs));
 738 }
 739 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
 740
 741 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
 742 {
 743         if (tag < tags->nr_tags) {
 744                 prefetch(tags->rqs[tag]);
 745                 return tags->rqs[tag];
 746         }
 747
 748         return NULL;
 749 }
 750 EXPORT_SYMBOL(blk_mq_tag_to_rq);
 751
 752 struct blk_mq_timeout_data {
 753         unsigned long next;
 754         unsigned int next_set;
 755 };
 756
 757 void blk_mq_rq_timed_out(struct request *req, bool reserved)
 758 {
 759         const struct blk_mq_ops *ops = req->q->mq_ops;
 760         enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
 761
 762         /*
 763          * We know that complete is set at this point. If STARTED isn't set
 764          * anymore, then the request isn't active and the "timeout" should
 765          * just be ignored. This can happen due to the bitflag ordering.
 766          * Timeout first checks if STARTED is set, and if it is, assumes
 767          * the request is active. But if we race with completion, then
 768          * both flags will get cleared. So check here again, and ignore
 769          * a timeout event with a request that isn't active.
 770          */
 771         if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags))
 772                 return;
 773
 774         if (ops->timeout)
 775                 ret = ops->timeout(req, reserved);
 776
 777         switch (ret) {
 778         case BLK_EH_HANDLED:
 779                 __blk_mq_complete_request(req);
 780                 break;
 781         case BLK_EH_RESET_TIMER:
 782                 blk_add_timer(req);
 783                 blk_clear_rq_complete(req);
 784                 break;
 785         case BLK_EH_NOT_HANDLED:
 786                 break;
 787         default:
 788                 printk(KERN_ERR "block: bad eh return: %d\n", ret);
 789                 break;
 790         }
 791 }
 792
 793 static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
 794                 struct request *rq, void *priv, bool reserved)
 795 {
 796         struct blk_mq_timeout_data *data = priv;
 797         unsigned long deadline;
 798
 799         if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
 800                 return;
 801
 802         /*
 803          * Ensures that if we see STARTED we must also see our
 804          * up-to-date deadline, see blk_mq_start_request().
 805          */
 806         smp_rmb();
 807
 808         deadline = READ_ONCE(rq->deadline);
 809
 810         /*
 811          * The rq being checked may have been freed and reallocated
 812          * out already here, we avoid this race by checking rq->deadline
 813          * and REQ_ATOM_COMPLETE flag together:
 814          *
 815          * - if rq->deadline is observed as new value because of
 816          *   reusing, the rq won't be timed out because of timing.
 817          * - if rq->deadline is observed as previous value,
 818          *   REQ_ATOM_COMPLETE flag won't be cleared in reuse path
 819          *   because we put a barrier between setting rq->deadline
 820          *   and clearing the flag in blk_mq_start_request(), so
 821          *   this rq won't be timed out too.
 822          */
 823         if (time_after_eq(jiffies, deadline)) {
 824                 if (!blk_mark_rq_complete(rq)) {
 825                         /*
 826                          * Again coherence order ensures that consecutive reads
 827                          * from the same variable must be in that order. This
 828                          * ensures that if we see COMPLETE clear, we must then
 829                          * see STARTED set and we'll ignore this timeout.
 830                          *
 831                          * (There's also the MB implied by the test_and_clear())
 832                          */
 833                         blk_mq_rq_timed_out(rq, reserved);
 834                 }
 835         } else if (!data->next_set || time_after(data->next, deadline)) {
 836                 data->next = deadline;
 837                 data->next_set = 1;
 838         }
 839 }
 840
 841 static void blk_mq_timeout_work(struct work_struct *work)
 842 {
 843         struct request_queue *q =
 844                 container_of(work, struct request_queue, timeout_work);
 845         struct blk_mq_timeout_data data = {
 846                 .next           = 0,
 847                 .next_set       = 0,
 848         };
 849         int i;
 850
 851         /* A deadlock might occur if a request is stuck requiring a
 852          * timeout at the same time a queue freeze is waiting
 853          * completion, since the timeout code would not be able to
 854          * acquire the queue reference here.
 855          *
 856          * That's why we don't use blk_queue_enter here; instead, we use
 857          * percpu_ref_tryget directly, because we need to be able to
 858          * obtain a reference even in the short window between the queue
 859          * starting to freeze, by dropping the first reference in
 860          * blk_freeze_queue_start, and the moment the last request is
 861          * consumed, marked by the instant q_usage_counter reaches
 862          * zero.
 863          */
 864         if (!percpu_ref_tryget(&q->q_usage_counter))
 865                 return;
 866
 867         blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data);
 868
 869         if (data.next_set) {
 870                 data.next = blk_rq_timeout(round_jiffies_up(data.next));
 871                 mod_timer(&q->timeout, data.next);
 872         } else {
 873                 struct blk_mq_hw_ctx *hctx;
 874
 875                 queue_for_each_hw_ctx(q, hctx, i) {
 876                         /* the hctx may be unmapped, so check it here */
 877                         if (blk_mq_hw_queue_mapped(hctx))
 878                                 blk_mq_tag_idle(hctx);
 879                 }
 880         }
 881         blk_queue_exit(q);
 882 }
 883
 884 struct flush_busy_ctx_data {
 885         struct blk_mq_hw_ctx *hctx;
 886         struct list_head *list;
 887 };
 888
 889 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
 890 {
 891         struct flush_busy_ctx_data *flush_data = data;
 892         struct blk_mq_hw_ctx *hctx = flush_data->hctx;
 893         struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
 894
 895         sbitmap_clear_bit(sb, bitnr);
 896         spin_lock(&ctx->lock);
 897         list_splice_tail_init(&ctx->rq_list, flush_data->list);
 898         spin_unlock(&ctx->lock);
 899         return true;
 900 }
 901
 902 /*
 903  * Process software queues that have been marked busy, splicing them
 904  * to the for-dispatch
 905  */
 906 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
 907 {
 908         struct flush_busy_ctx_data data = {
 909                 .hctx = hctx,
 910                 .list = list,
 911         };
 912
 913         sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
 914 }
 915 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
 916
 917 static inline unsigned int queued_to_index(unsigned int queued)
 918 {
 919         if (!queued)
 920                 return 0;
 921
 922         return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
 923 }
 924
 925 bool blk_mq_get_driver_tag(struct request *rq, struct blk_mq_hw_ctx **hctx,
 926                            bool wait)
 927 {
 928         struct blk_mq_alloc_data data = {
 929                 .q = rq->q,
 930                 .hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu),
 931                 .flags = wait ? 0 : BLK_MQ_REQ_NOWAIT,
 932         };
 933
 934         might_sleep_if(wait);
 935
 936         if (rq->tag != -1)
 937                 goto done;
 938
 939         if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag))
 940                 data.flags |= BLK_MQ_REQ_RESERVED;
 941
 942         rq->tag = blk_mq_get_tag(&data);
 943         if (rq->tag >= 0) {
 944                 if (blk_mq_tag_busy(data.hctx)) {
 945                         rq->rq_flags |= RQF_MQ_INFLIGHT;
 946                         atomic_inc(&data.hctx->nr_active);
 947                 }
 948                 data.hctx->tags->rqs[rq->tag] = rq;
 949         }
 950
 951 done:
 952         if (hctx)
 953                 *hctx = data.hctx;
 954         return rq->tag != -1;
 955 }
 956
 957 static void __blk_mq_put_driver_tag(struct blk_mq_hw_ctx *hctx,
 958                                     struct request *rq)
 959 {
 960         blk_mq_put_tag(hctx, hctx->tags, rq->mq_ctx, rq->tag);
 961         rq->tag = -1;
 962
 963         if (rq->rq_flags & RQF_MQ_INFLIGHT) {
 964                 rq->rq_flags &= ~RQF_MQ_INFLIGHT;
 965                 atomic_dec(&hctx->nr_active);
 966         }
 967 }
 968
 969 static void blk_mq_put_driver_tag_hctx(struct blk_mq_hw_ctx *hctx,
 970                                        struct request *rq)
 971 {
 972         if (rq->tag == -1 || rq->internal_tag == -1)
 973                 return;
 974
 975         __blk_mq_put_driver_tag(hctx, rq);
 976 }
 977
 978 static void blk_mq_put_driver_tag(struct request *rq)
 979 {
 980         struct blk_mq_hw_ctx *hctx;
 981
 982         if (rq->tag == -1 || rq->internal_tag == -1)
 983                 return;
 984
 985         hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu);
 986         __blk_mq_put_driver_tag(hctx, rq);
 987 }
 988
 989 /*
 990  * If we fail getting a driver tag because all the driver tags are already
 991  * assigned and on the dispatch list, BUT the first entry does not have a
 992  * tag, then we could deadlock. For that case, move entries with assigned
 993  * driver tags to the front, leaving the set of tagged requests in the
 994  * same order, and the untagged set in the same order.
 995  */
 996 static bool reorder_tags_to_front(struct list_head *list)
 997 {
 998         struct request *rq, *tmp, *first = NULL;
 999
1000         list_for_each_entry_safe_reverse(rq, tmp, list, queuelist) {
1001                 if (rq == first)
1002                         break;
1003                 if (rq->tag != -1) {
1004                         list_move(&rq->queuelist, list);
1005                         if (!first)
1006                                 first = rq;
1007                 }
1008         }
1009
1010         return first != NULL;
1011 }
1012
1013 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode, int flags,
1014                                 void *key)
1015 {
1016         struct blk_mq_hw_ctx *hctx;
1017
1018         hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1019
1020         list_del(&wait->entry);
1021         clear_bit_unlock(BLK_MQ_S_TAG_WAITING, &hctx->state);
1022         blk_mq_run_hw_queue(hctx, true);
1023         return 1;
1024 }
1025
1026 static bool blk_mq_dispatch_wait_add(struct blk_mq_hw_ctx *hctx)
1027 {
1028         struct sbq_wait_state *ws;
1029
1030         /*
1031          * The TAG_WAITING bit serves as a lock protecting hctx->dispatch_wait.
1032          * The thread which wins the race to grab this bit adds the hardware
1033          * queue to the wait queue.
1034          */
1035         if (test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state) ||
1036             test_and_set_bit_lock(BLK_MQ_S_TAG_WAITING, &hctx->state))
1037                 return false;
1038
1039         init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
1040         ws = bt_wait_ptr(&hctx->tags->bitmap_tags, hctx);
1041
1042         /*
1043          * As soon as this returns, it's no longer safe to fiddle with
1044          * hctx->dispatch_wait, since a completion can wake up the wait queue
1045          * and unlock the bit.
1046          */
1047         add_wait_queue(&ws->wait, &hctx->dispatch_wait);
1048         return true;
1049 }
1050
1051 bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list)
1052 {
1053         struct blk_mq_hw_ctx *hctx;
1054         struct request *rq;
1055         int errors, queued;
1056
1057         if (list_empty(list))
1058                 return false;
1059
1060         /*
1061          * Now process all the entries, sending them to the driver.
1062          */
1063         errors = queued = 0;
1064         do {
1065                 struct blk_mq_queue_data bd;
1066                 blk_status_t ret;
1067
1068                 rq = list_first_entry(list, struct request, queuelist);
1069                 if (!blk_mq_get_driver_tag(rq, &hctx, false)) {
1070                         if (!queued && reorder_tags_to_front(list))
1071                                 continue;
1072
1073                         /*
1074                          * The initial allocation attempt failed, so we need to
1075                          * rerun the hardware queue when a tag is freed.
1076                          */
1077                         if (!blk_mq_dispatch_wait_add(hctx))
1078                                 break;
1079
1080                         /*
1081                          * It's possible that a tag was freed in the window
1082                          * between the allocation failure and adding the
1083                          * hardware queue to the wait queue.
1084                          */
1085                         if (!blk_mq_get_driver_tag(rq, &hctx, false))
1086                                 break;
1087                 }
1088
1089                 list_del_init(&rq->queuelist);
1090
1091                 bd.rq = rq;
1092
1093                 /*
1094                  * Flag last if we have no more requests, or if we have more
1095                  * but can't assign a driver tag to it.
1096                  */
1097                 if (list_empty(list))
1098                         bd.last = true;
1099                 else {
1100                         struct request *nxt;
1101
1102                         nxt = list_first_entry(list, struct request, queuelist);
1103                         bd.last = !blk_mq_get_driver_tag(nxt, NULL, false);
1104                 }
1105
1106                 ret = q->mq_ops->queue_rq(hctx, &bd);
1107                 if (ret == BLK_STS_RESOURCE) {
1108                         blk_mq_put_driver_tag_hctx(hctx, rq);
1109                         list_add(&rq->queuelist, list);
1110                         __blk_mq_requeue_request(rq);
1111                         break;
1112                 }
1113
1114                 if (unlikely(ret != BLK_STS_OK)) {
1115                         errors++;
1116                         blk_mq_end_request(rq, BLK_STS_IOERR);
1117                         continue;
1118                 }
1119
1120                 queued++;
1121         } while (!list_empty(list));
1122
1123         hctx->dispatched[queued_to_index(queued)]++;
1124
1125         /*
1126          * Any items that need requeuing? Stuff them into hctx->dispatch,
1127          * that is where we will continue on next queue run.
1128          */
1129         if (!list_empty(list)) {
1130                 /*
1131                  * If an I/O scheduler has been configured and we got a driver
1132                  * tag for the next request already, free it again.
1133                  */
1134                 rq = list_first_entry(list, struct request, queuelist);
1135                 blk_mq_put_driver_tag(rq);
1136
1137                 spin_lock(&hctx->lock);
1138                 list_splice_init(list, &hctx->dispatch);
1139                 spin_unlock(&hctx->lock);
1140
1141                 /*
1142                  * If SCHED_RESTART was set by the caller of this function and
1143                  * it is no longer set that means that it was cleared by another
1144                  * thread and hence that a queue rerun is needed.
1145                  *
1146                  * If TAG_WAITING is set that means that an I/O scheduler has
1147                  * been configured and another thread is waiting for a driver
1148                  * tag. To guarantee fairness, do not rerun this hardware queue
1149                  * but let the other thread grab the driver tag.
1150                  *
1151                  * If no I/O scheduler has been configured it is possible that
1152                  * the hardware queue got stopped and restarted before requests
1153                  * were pushed back onto the dispatch list. Rerun the queue to
1154                  * avoid starvation. Notes:
1155                  * - blk_mq_run_hw_queue() checks whether or not a queue has
1156                  *   been stopped before rerunning a queue.
1157                  * - Some but not all block drivers stop a queue before
1158                  *   returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1159                  *   and dm-rq.
1160                  */
1161                 if (!blk_mq_sched_needs_restart(hctx) &&
1162                     !test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state))
1163                         blk_mq_run_hw_queue(hctx, true);
1164         }
1165
1166         return (queued + errors) != 0;
1167 }
1168
1169 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1170 {
1171         int srcu_idx;
1172
1173         /*
1174          * We should be running this queue from one of the CPUs that
1175          * are mapped to it.
1176          */
1177         WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1178                 cpu_online(hctx->next_cpu));
1179
1180         /*
1181          * We can't run the queue inline with ints disabled. Ensure that
1182          * we catch bad users of this early.
1183          */
1184         WARN_ON_ONCE(in_interrupt());
1185
1186         if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
1187                 rcu_read_lock();
1188                 blk_mq_sched_dispatch_requests(hctx);
1189                 rcu_read_unlock();
1190         } else {
1191                 might_sleep();
1192
1193                 srcu_idx = srcu_read_lock(hctx->queue_rq_srcu);
1194                 blk_mq_sched_dispatch_requests(hctx);
1195                 srcu_read_unlock(hctx->queue_rq_srcu, srcu_idx);
1196         }
1197 }
1198
1199 /*
1200  * It'd be great if the workqueue API had a way to pass
1201  * in a mask and had some smarts for more clever placement.
1202  * For now we just round-robin here, switching for every
1203  * BLK_MQ_CPU_WORK_BATCH queued items.
1204  */
1205 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1206 {
1207         if (hctx->queue->nr_hw_queues == 1)
1208                 return WORK_CPU_UNBOUND;
1209
1210         if (--hctx->next_cpu_batch <= 0) {
1211                 int next_cpu;
1212
1213                 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
1214                 if (next_cpu >= nr_cpu_ids)
1215                         next_cpu = cpumask_first(hctx->cpumask);
1216
1217                 hctx->next_cpu = next_cpu;
1218                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1219         }
1220
1221         return hctx->next_cpu;
1222 }
1223
1224 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1225                                         unsigned long msecs)
1226 {
1227         if (WARN_ON_ONCE(!blk_mq_hw_queue_mapped(hctx)))
1228                 return;
1229
1230         if (unlikely(blk_mq_hctx_stopped(hctx)))
1231                 return;
1232
1233         if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1234                 int cpu = get_cpu();
1235                 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1236                         __blk_mq_run_hw_queue(hctx);
1237                         put_cpu();
1238                         return;
1239                 }
1240
1241                 put_cpu();
1242         }
1243
1244         kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
1245                                          &hctx->run_work,
1246                                          msecs_to_jiffies(msecs));
1247 }
1248
1249 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1250 {
1251         __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1252 }
1253 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1254
1255 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1256 {
1257         __blk_mq_delay_run_hw_queue(hctx, async, 0);
1258 }
1259 EXPORT_SYMBOL(blk_mq_run_hw_queue);
1260
1261 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1262 {
1263         struct blk_mq_hw_ctx *hctx;
1264         int i;
1265
1266         queue_for_each_hw_ctx(q, hctx, i) {
1267                 if (!blk_mq_hctx_has_pending(hctx) ||
1268                     blk_mq_hctx_stopped(hctx))
1269                         continue;
1270
1271                 blk_mq_run_hw_queue(hctx, async);
1272         }
1273 }
1274 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1275
1276 /**
1277  * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1278  * @q: request queue.
1279  *
1280  * The caller is responsible for serializing this function against
1281  * blk_mq_{start,stop}_hw_queue().
1282  */
1283 bool blk_mq_queue_stopped(struct request_queue *q)
1284 {
1285         struct blk_mq_hw_ctx *hctx;
1286         int i;
1287
1288         queue_for_each_hw_ctx(q, hctx, i)
1289                 if (blk_mq_hctx_stopped(hctx))
1290                         return true;
1291
1292         return false;
1293 }
1294 EXPORT_SYMBOL(blk_mq_queue_stopped);
1295
1296 /*
1297  * This function is often used for pausing .queue_rq() by driver when
1298  * there isn't enough resource or some conditions aren't satisfied, and
1299  * BLK_STS_RESOURCE is usually returned.
1300  *
1301  * We do not guarantee that dispatch can be drained or blocked
1302  * after blk_mq_stop_hw_queue() returns. Please use
1303  * blk_mq_quiesce_queue() for that requirement.
1304  */
1305 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1306 {
1307         cancel_delayed_work(&hctx->run_work);
1308
1309         set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1310 }
1311 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1312
1313 /*
1314  * This function is often used for pausing .queue_rq() by driver when
1315  * there isn't enough resource or some conditions aren't satisfied, and
1316  * BLK_STS_RESOURCE is usually returned.
1317  *
1318  * We do not guarantee that dispatch can be drained or blocked
1319  * after blk_mq_stop_hw_queues() returns. Please use
1320  * blk_mq_quiesce_queue() for that requirement.
1321  */
1322 void blk_mq_stop_hw_queues(struct request_queue *q)
1323 {
1324         struct blk_mq_hw_ctx *hctx;
1325         int i;
1326
1327         queue_for_each_hw_ctx(q, hctx, i)
1328                 blk_mq_stop_hw_queue(hctx);
1329 }
1330 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1331
1332 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1333 {
1334         clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1335
1336         blk_mq_run_hw_queue(hctx, false);
1337 }
1338 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1339
1340 void blk_mq_start_hw_queues(struct request_queue *q)
1341 {
1342         struct blk_mq_hw_ctx *hctx;
1343         int i;
1344
1345         queue_for_each_hw_ctx(q, hctx, i)
1346                 blk_mq_start_hw_queue(hctx);
1347 }
1348 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1349
1350 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1351 {
1352         if (!blk_mq_hctx_stopped(hctx))
1353                 return;
1354
1355         clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1356         blk_mq_run_hw_queue(hctx, async);
1357 }
1358 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1359
1360 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1361 {
1362         struct blk_mq_hw_ctx *hctx;
1363         int i;
1364
1365         queue_for_each_hw_ctx(q, hctx, i)
1366                 blk_mq_start_stopped_hw_queue(hctx, async);
1367 }
1368 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1369
1370 static void blk_mq_run_work_fn(struct work_struct *work)
1371 {
1372         struct blk_mq_hw_ctx *hctx;
1373
1374         hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1375
1376         /*
1377          * If we are stopped, don't run the queue. The exception is if
1378          * BLK_MQ_S_START_ON_RUN is set. For that case, we auto-clear
1379          * the STOPPED bit and run it.
1380          */
1381         if (test_bit(BLK_MQ_S_STOPPED, &hctx->state)) {
1382                 if (!test_bit(BLK_MQ_S_START_ON_RUN, &hctx->state))
1383                         return;
1384
1385                 clear_bit(BLK_MQ_S_START_ON_RUN, &hctx->state);
1386                 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1387         }
1388
1389         __blk_mq_run_hw_queue(hctx);
1390 }
1391
1392
1393 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1394 {
1395         if (WARN_ON_ONCE(!blk_mq_hw_queue_mapped(hctx)))
1396                 return;
1397
1398         /*
1399          * Stop the hw queue, then modify currently delayed work.
1400          * This should prevent us from running the queue prematurely.
1401          * Mark the queue as auto-clearing STOPPED when it runs.
1402          */
1403         blk_mq_stop_hw_queue(hctx);
1404         set_bit(BLK_MQ_S_START_ON_RUN, &hctx->state);
1405         kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
1406                                         &hctx->run_work,
1407                                         msecs_to_jiffies(msecs));
1408 }
1409 EXPORT_SYMBOL(blk_mq_delay_queue);
1410
1411 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1412                                             struct request *rq,
1413                                             bool at_head)
1414 {
1415         struct blk_mq_ctx *ctx = rq->mq_ctx;
1416
1417         lockdep_assert_held(&ctx->lock);
1418
1419         trace_block_rq_insert(hctx->queue, rq);
1420
1421         if (at_head)
1422                 list_add(&rq->queuelist, &ctx->rq_list);
1423         else
1424                 list_add_tail(&rq->queuelist, &ctx->rq_list);
1425 }
1426
1427 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1428                              bool at_head)
1429 {
1430         struct blk_mq_ctx *ctx = rq->mq_ctx;
1431
1432         lockdep_assert_held(&ctx->lock);
1433
1434         __blk_mq_insert_req_list(hctx, rq, at_head);
1435         blk_mq_hctx_mark_pending(hctx, ctx);
1436 }
1437
1438 /*
1439  * Should only be used carefully, when the caller knows we want to
1440  * bypass a potential IO scheduler on the target device.
1441  */
1442 void blk_mq_request_bypass_insert(struct request *rq)
1443 {
1444         struct blk_mq_ctx *ctx = rq->mq_ctx;
1445         struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(rq->q, ctx->cpu);
1446
1447         spin_lock(&hctx->lock);
1448         list_add_tail(&rq->queuelist, &hctx->dispatch);
1449         spin_unlock(&hctx->lock);
1450
1451         blk_mq_run_hw_queue(hctx, false);
1452 }
1453
1454 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1455                             struct list_head *list)
1456
1457 {
1458         /*
1459          * preemption doesn't flush plug list, so it's possible ctx->cpu is
1460          * offline now
1461          */
1462         spin_lock(&ctx->lock);
1463         while (!list_empty(list)) {
1464                 struct request *rq;
1465
1466                 rq = list_first_entry(list, struct request, queuelist);
1467                 BUG_ON(rq->mq_ctx != ctx);
1468                 list_del_init(&rq->queuelist);
1469                 __blk_mq_insert_req_list(hctx, rq, false);
1470         }
1471         blk_mq_hctx_mark_pending(hctx, ctx);
1472         spin_unlock(&ctx->lock);
1473 }
1474
1475 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1476 {
1477         struct request *rqa = container_of(a, struct request, queuelist);
1478         struct request *rqb = container_of(b, struct request, queuelist);
1479
1480         return !(rqa->mq_ctx < rqb->mq_ctx ||
1481                  (rqa->mq_ctx == rqb->mq_ctx &&
1482                   blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1483 }
1484
1485 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1486 {
1487         struct blk_mq_ctx *this_ctx;
1488         struct request_queue *this_q;
1489         struct request *rq;
1490         LIST_HEAD(list);
1491         LIST_HEAD(ctx_list);
1492         unsigned int depth;
1493
1494         list_splice_init(&plug->mq_list, &list);
1495
1496         list_sort(NULL, &list, plug_ctx_cmp);
1497
1498         this_q = NULL;
1499         this_ctx = NULL;
1500         depth = 0;
1501
1502         while (!list_empty(&list)) {
1503                 rq = list_entry_rq(list.next);
1504                 list_del_init(&rq->queuelist);
1505                 BUG_ON(!rq->q);
1506                 if (rq->mq_ctx != this_ctx) {
1507                         if (this_ctx) {
1508                                 trace_block_unplug(this_q, depth, from_schedule);
1509                                 blk_mq_sched_insert_requests(this_q, this_ctx,
1510                                                                 &ctx_list,
1511                                                                 from_schedule);
1512                         }
1513
1514                         this_ctx = rq->mq_ctx;
1515                         this_q = rq->q;
1516                         depth = 0;
1517                 }
1518
1519                 depth++;
1520                 list_add_tail(&rq->queuelist, &ctx_list);
1521         }
1522
1523         /*
1524          * If 'this_ctx' is set, we know we have entries to complete
1525          * on 'ctx_list'. Do those.
1526          */
1527         if (this_ctx) {
1528                 trace_block_unplug(this_q, depth, from_schedule);
1529                 blk_mq_sched_insert_requests(this_q, this_ctx, &ctx_list,
1530                                                 from_schedule);
1531         }
1532 }
1533
1534 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1535 {
1536         blk_init_request_from_bio(rq, bio);
1537
1538         blk_rq_set_rl(rq, blk_get_rl(rq->q, bio));
1539
1540         blk_account_io_start(rq, true);
1541 }
1542
1543 static inline void blk_mq_queue_io(struct blk_mq_hw_ctx *hctx,
1544                                    struct blk_mq_ctx *ctx,
1545                                    struct request *rq)
1546 {
1547         spin_lock(&ctx->lock);
1548         __blk_mq_insert_request(hctx, rq, false);
1549         spin_unlock(&ctx->lock);
1550 }
1551
1552 static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq)
1553 {
1554         if (rq->tag != -1)
1555                 return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false);
1556
1557         return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true);
1558 }
1559
1560 static void __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1561                                         struct request *rq,
1562                                         blk_qc_t *cookie, bool may_sleep)
1563 {
1564         struct request_queue *q = rq->q;
1565         struct blk_mq_queue_data bd = {
1566                 .rq = rq,
1567                 .last = true,
1568         };
1569         blk_qc_t new_cookie;
1570         blk_status_t ret;
1571         bool run_queue = true;
1572
1573         /* RCU or SRCU read lock is needed before checking quiesced flag */
1574         if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
1575                 run_queue = false;
1576                 goto insert;
1577         }
1578
1579         if (q->elevator)
1580                 goto insert;
1581
1582         if (!blk_mq_get_driver_tag(rq, NULL, false))
1583                 goto insert;
1584
1585         new_cookie = request_to_qc_t(hctx, rq);
1586
1587         /*
1588          * For OK queue, we are done. For error, kill it. Any other
1589          * error (busy), just add it to our list as we previously
1590          * would have done
1591          */
1592         ret = q->mq_ops->queue_rq(hctx, &bd);
1593         switch (ret) {
1594         case BLK_STS_OK:
1595                 *cookie = new_cookie;
1596                 return;
1597         case BLK_STS_RESOURCE:
1598                 __blk_mq_requeue_request(rq);
1599                 goto insert;
1600         default:
1601                 *cookie = BLK_QC_T_NONE;
1602                 blk_mq_end_request(rq, ret);
1603                 return;
1604         }
1605
1606 insert:
1607         blk_mq_sched_insert_request(rq, false, run_queue, false, may_sleep);
1608 }
1609
1610 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1611                 struct request *rq, blk_qc_t *cookie)
1612 {
1613         if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
1614                 rcu_read_lock();
1615                 __blk_mq_try_issue_directly(hctx, rq, cookie, false);
1616                 rcu_read_unlock();
1617         } else {
1618                 unsigned int srcu_idx;
1619
1620                 might_sleep();
1621
1622                 srcu_idx = srcu_read_lock(hctx->queue_rq_srcu);
1623                 __blk_mq_try_issue_directly(hctx, rq, cookie, true);
1624                 srcu_read_unlock(hctx->queue_rq_srcu, srcu_idx);
1625         }
1626 }
1627
1628 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1629 {
1630         const int is_sync = op_is_sync(bio->bi_opf);
1631         const int is_flush_fua = op_is_flush(bio->bi_opf);
1632         struct blk_mq_alloc_data data = { .flags = 0 };
1633         struct request *rq;
1634         unsigned int request_count = 0;
1635         struct blk_plug *plug;
1636         struct request *same_queue_rq = NULL;
1637         blk_qc_t cookie;
1638         unsigned int wb_acct;
1639
1640         blk_queue_bounce(q, &bio);
1641
1642         blk_queue_split(q, &bio);
1643
1644         if (!bio_integrity_prep(bio))
1645                 return BLK_QC_T_NONE;
1646
1647         if (!is_flush_fua && !blk_queue_nomerges(q) &&
1648             blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq))
1649                 return BLK_QC_T_NONE;
1650
1651         if (blk_mq_sched_bio_merge(q, bio))
1652                 return BLK_QC_T_NONE;
1653
1654         wb_acct = wbt_wait(q->rq_wb, bio, NULL);
1655
1656         trace_block_getrq(q, bio, bio->bi_opf);
1657
1658         rq = blk_mq_get_request(q, bio, bio->bi_opf, &data);
1659         if (unlikely(!rq)) {
1660                 __wbt_done(q->rq_wb, wb_acct);
1661                 if (bio->bi_opf & REQ_NOWAIT)
1662                         bio_wouldblock_error(bio);
1663                 return BLK_QC_T_NONE;
1664         }
1665
1666         wbt_track(&rq->issue_stat, wb_acct);
1667
1668         cookie = request_to_qc_t(data.hctx, rq);
1669
1670         plug = current->plug;
1671         if (unlikely(is_flush_fua)) {
1672                 blk_mq_put_ctx(data.ctx);
1673                 blk_mq_bio_to_request(rq, bio);
1674                 if (q->elevator) {
1675                         blk_mq_sched_insert_request(rq, false, true, true,
1676                                         true);
1677                 } else {
1678                         blk_insert_flush(rq);
1679                         blk_mq_run_hw_queue(data.hctx, true);
1680                 }
1681         } else if (plug && q->nr_hw_queues == 1) {
1682                 struct request *last = NULL;
1683
1684                 blk_mq_put_ctx(data.ctx);
1685                 blk_mq_bio_to_request(rq, bio);
1686
1687                 /*
1688                  * @request_count may become stale because of schedule
1689                  * out, so check the list again.
1690                  */
1691                 if (list_empty(&plug->mq_list))
1692                         request_count = 0;
1693                 else if (blk_queue_nomerges(q))
1694                         request_count = blk_plug_queued_count(q);
1695
1696                 if (!request_count)
1697                         trace_block_plug(q);
1698                 else
1699                         last = list_entry_rq(plug->mq_list.prev);
1700
1701                 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
1702                     blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1703                         blk_flush_plug_list(plug, false);
1704                         trace_block_plug(q);
1705                 }
1706
1707                 list_add_tail(&rq->queuelist, &plug->mq_list);
1708         } else if (plug && !blk_queue_nomerges(q)) {
1709                 blk_mq_bio_to_request(rq, bio);
1710
1711                 /*
1712                  * We do limited plugging. If the bio can be merged, do that.
1713                  * Otherwise the existing request in the plug list will be
1714                  * issued. So the plug list will have one request at most
1715                  * The plug list might get flushed before this. If that happens,
1716                  * the plug list is empty, and same_queue_rq is invalid.
1717                  */
1718                 if (list_empty(&plug->mq_list))
1719                         same_queue_rq = NULL;
1720                 if (same_queue_rq)
1721                         list_del_init(&same_queue_rq->queuelist);
1722                 list_add_tail(&rq->queuelist, &plug->mq_list);
1723
1724                 blk_mq_put_ctx(data.ctx);
1725
1726                 if (same_queue_rq) {
1727                         data.hctx = blk_mq_map_queue(q,
1728                                         same_queue_rq->mq_ctx->cpu);
1729                         blk_mq_try_issue_directly(data.hctx, same_queue_rq,
1730                                         &cookie);
1731                 }
1732         } else if (q->nr_hw_queues > 1 && is_sync) {
1733                 blk_mq_put_ctx(data.ctx);
1734                 blk_mq_bio_to_request(rq, bio);
1735                 blk_mq_try_issue_directly(data.hctx, rq, &cookie);
1736         } else if (q->elevator) {
1737                 blk_mq_put_ctx(data.ctx);
1738                 blk_mq_bio_to_request(rq, bio);
1739                 blk_mq_sched_insert_request(rq, false, true, true, true);
1740         } else {
1741                 blk_mq_put_ctx(data.ctx);
1742                 blk_mq_bio_to_request(rq, bio);
1743                 blk_mq_queue_io(data.hctx, data.ctx, rq);
1744                 blk_mq_run_hw_queue(data.hctx, true);
1745         }
1746
1747         return cookie;
1748 }
1749
1750 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1751                      unsigned int hctx_idx)
1752 {
1753         struct page *page;
1754
1755         if (tags->rqs && set->ops->exit_request) {
1756                 int i;
1757
1758                 for (i = 0; i < tags->nr_tags; i++) {
1759                         struct request *rq = tags->static_rqs[i];
1760
1761                         if (!rq)
1762                                 continue;
1763                         set->ops->exit_request(set, rq, hctx_idx);
1764                         tags->static_rqs[i] = NULL;
1765                 }
1766         }
1767
1768         while (!list_empty(&tags->page_list)) {
1769                 page = list_first_entry(&tags->page_list, struct page, lru);
1770                 list_del_init(&page->lru);
1771                 /*
1772                  * Remove kmemleak object previously allocated in
1773                  * blk_mq_init_rq_map().
1774                  */
1775                 kmemleak_free(page_address(page));
1776                 __free_pages(page, page->private);
1777         }
1778 }
1779
1780 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
1781 {
1782         kfree(tags->rqs);
1783         tags->rqs = NULL;
1784         kfree(tags->static_rqs);
1785         tags->static_rqs = NULL;
1786
1787         blk_mq_free_tags(tags);
1788 }
1789
1790 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
1791                                         unsigned int hctx_idx,
1792                                         unsigned int nr_tags,
1793                                         unsigned int reserved_tags)
1794 {
1795         struct blk_mq_tags *tags;
1796         int node;
1797
1798         node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
1799         if (node == NUMA_NO_NODE)
1800                 node = set->numa_node;
1801
1802         tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
1803                                 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
1804         if (!tags)
1805                 return NULL;
1806
1807         tags->rqs = kzalloc_node(nr_tags * sizeof(struct request *),
1808                                  GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1809                                  node);
1810         if (!tags->rqs) {
1811                 blk_mq_free_tags(tags);
1812                 return NULL;
1813         }
1814
1815         tags->static_rqs = kzalloc_node(nr_tags * sizeof(struct request *),
1816                                  GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1817                                  node);
1818         if (!tags->static_rqs) {
1819                 kfree(tags->rqs);
1820                 blk_mq_free_tags(tags);
1821                 return NULL;
1822         }
1823
1824         return tags;
1825 }
1826
1827 static size_t order_to_size(unsigned int order)
1828 {
1829         return (size_t)PAGE_SIZE << order;
1830 }
1831
1832 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1833                      unsigned int hctx_idx, unsigned int depth)
1834 {
1835         unsigned int i, j, entries_per_page, max_order = 4;
1836         size_t rq_size, left;
1837         int node;
1838
1839         node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
1840         if (node == NUMA_NO_NODE)
1841                 node = set->numa_node;
1842
1843         INIT_LIST_HEAD(&tags->page_list);
1844
1845         /*
1846          * rq_size is the size of the request plus driver payload, rounded
1847          * to the cacheline size
1848          */
1849         rq_size = round_up(sizeof(struct request) + set->cmd_size,
1850                                 cache_line_size());
1851         left = rq_size * depth;
1852
1853         for (i = 0; i < depth; ) {
1854                 int this_order = max_order;
1855                 struct page *page;
1856                 int to_do;
1857                 void *p;
1858
1859                 while (this_order && left < order_to_size(this_order - 1))
1860                         this_order--;
1861
1862                 do {
1863                         page = alloc_pages_node(node,
1864                                 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
1865                                 this_order);
1866                         if (page)
1867                                 break;
1868                         if (!this_order--)
1869                                 break;
1870                         if (order_to_size(this_order) < rq_size)
1871                                 break;
1872                 } while (1);
1873
1874                 if (!page)
1875                         goto fail;
1876
1877                 page->private = this_order;
1878                 list_add_tail(&page->lru, &tags->page_list);
1879
1880                 p = page_address(page);
1881                 /*
1882                  * Allow kmemleak to scan these pages as they contain pointers
1883                  * to additional allocations like via ops->init_request().
1884                  */
1885                 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
1886                 entries_per_page = order_to_size(this_order) / rq_size;
1887                 to_do = min(entries_per_page, depth - i);
1888                 left -= to_do * rq_size;
1889                 for (j = 0; j < to_do; j++) {
1890                         struct request *rq = p;
1891
1892                         tags->static_rqs[i] = rq;
1893                         if (set->ops->init_request) {
1894                                 if (set->ops->init_request(set, rq, hctx_idx,
1895                                                 node)) {
1896                                         tags->static_rqs[i] = NULL;
1897                                         goto fail;
1898                                 }
1899                         }
1900
1901                         p += rq_size;
1902                         i++;
1903                 }
1904         }
1905         return 0;
1906
1907 fail:
1908         blk_mq_free_rqs(set, tags, hctx_idx);
1909         return -ENOMEM;
1910 }
1911
1912 /*
1913  * 'cpu' is going away. splice any existing rq_list entries from this
1914  * software queue to the hw queue dispatch list, and ensure that it
1915  * gets run.
1916  */
1917 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
1918 {
1919         struct blk_mq_hw_ctx *hctx;
1920         struct blk_mq_ctx *ctx;
1921         LIST_HEAD(tmp);
1922
1923         hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
1924         ctx = __blk_mq_get_ctx(hctx->queue, cpu);
1925
1926         spin_lock(&ctx->lock);
1927         if (!list_empty(&ctx->rq_list)) {
1928                 list_splice_init(&ctx->rq_list, &tmp);
1929                 blk_mq_hctx_clear_pending(hctx, ctx);
1930         }
1931         spin_unlock(&ctx->lock);
1932
1933         if (list_empty(&tmp))
1934                 return 0;
1935
1936         spin_lock(&hctx->lock);
1937         list_splice_tail_init(&tmp, &hctx->dispatch);
1938         spin_unlock(&hctx->lock);
1939
1940         blk_mq_run_hw_queue(hctx, true);
1941         return 0;
1942 }
1943
1944 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
1945 {
1946         cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
1947                                             &hctx->cpuhp_dead);
1948 }
1949
1950 /* hctx->ctxs will be freed in queue's release handler */
1951 static void blk_mq_exit_hctx(struct request_queue *q,
1952                 struct blk_mq_tag_set *set,
1953                 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
1954 {
1955         blk_mq_debugfs_unregister_hctx(hctx);
1956
1957         blk_mq_tag_idle(hctx);
1958
1959         if (set->ops->exit_request)
1960                 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
1961
1962         blk_mq_sched_exit_hctx(q, hctx, hctx_idx);
1963
1964         if (set->ops->exit_hctx)
1965                 set->ops->exit_hctx(hctx, hctx_idx);
1966
1967         if (hctx->flags & BLK_MQ_F_BLOCKING)
1968                 cleanup_srcu_struct(hctx->queue_rq_srcu);
1969
1970         blk_mq_remove_cpuhp(hctx);
1971         blk_free_flush_queue(hctx->fq);
1972         sbitmap_free(&hctx->ctx_map);
1973 }
1974
1975 static void blk_mq_exit_hw_queues(struct request_queue *q,
1976                 struct blk_mq_tag_set *set, int nr_queue)
1977 {
1978         struct blk_mq_hw_ctx *hctx;
1979         unsigned int i;
1980
1981         queue_for_each_hw_ctx(q, hctx, i) {
1982                 if (i == nr_queue)
1983                         break;
1984                 blk_mq_exit_hctx(q, set, hctx, i);
1985         }
1986 }
1987
1988 static int blk_mq_init_hctx(struct request_queue *q,
1989                 struct blk_mq_tag_set *set,
1990                 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
1991 {
1992         int node;
1993
1994         node = hctx->numa_node;
1995         if (node == NUMA_NO_NODE)
1996                 node = hctx->numa_node = set->numa_node;
1997
1998         INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
1999         spin_lock_init(&hctx->lock);
2000         INIT_LIST_HEAD(&hctx->dispatch);
2001         hctx->queue = q;
2002         hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
2003
2004         cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
2005
2006         hctx->tags = set->tags[hctx_idx];
2007
2008         /*
2009          * Allocate space for all possible cpus to avoid allocation at
2010          * runtime
2011          */
2012         hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
2013                                         GFP_KERNEL, node);
2014         if (!hctx->ctxs)
2015                 goto unregister_cpu_notifier;
2016
2017         if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL,
2018                               node))
2019                 goto free_ctxs;
2020
2021         hctx->nr_ctx = 0;
2022
2023         if (set->ops->init_hctx &&
2024             set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
2025                 goto free_bitmap;
2026
2027         if (blk_mq_sched_init_hctx(q, hctx, hctx_idx))
2028                 goto exit_hctx;
2029
2030         hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
2031         if (!hctx->fq)
2032                 goto sched_exit_hctx;
2033
2034         if (set->ops->init_request &&
2035             set->ops->init_request(set, hctx->fq->flush_rq, hctx_idx,
2036                                    node))
2037                 goto free_fq;
2038
2039         if (hctx->flags & BLK_MQ_F_BLOCKING)
2040                 init_srcu_struct(hctx->queue_rq_srcu);
2041
2042         blk_mq_debugfs_register_hctx(q, hctx);
2043
2044         return 0;
2045
2046  free_fq:
2047         kfree(hctx->fq);
2048  sched_exit_hctx:
2049         blk_mq_sched_exit_hctx(q, hctx, hctx_idx);
2050  exit_hctx:
2051         if (set->ops->exit_hctx)
2052                 set->ops->exit_hctx(hctx, hctx_idx);
2053  free_bitmap:
2054         sbitmap_free(&hctx->ctx_map);
2055  free_ctxs:
2056         kfree(hctx->ctxs);
2057  unregister_cpu_notifier:
2058         blk_mq_remove_cpuhp(hctx);
2059         return -1;
2060 }
2061
2062 static void blk_mq_init_cpu_queues(struct request_queue *q,
2063                                    unsigned int nr_hw_queues)
2064 {
2065         unsigned int i;
2066
2067         for_each_possible_cpu(i) {
2068                 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
2069                 struct blk_mq_hw_ctx *hctx;
2070
2071                 __ctx->cpu = i;
2072                 spin_lock_init(&__ctx->lock);
2073                 INIT_LIST_HEAD(&__ctx->rq_list);
2074                 __ctx->queue = q;
2075
2076                 /* If the cpu isn't present, the cpu is mapped to first hctx */
2077                 if (!cpu_present(i))
2078                         continue;
2079
2080                 hctx = blk_mq_map_queue(q, i);
2081
2082                 /*
2083                  * Set local node, IFF we have more than one hw queue. If
2084                  * not, we remain on the home node of the device
2085                  */
2086                 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
2087                         hctx->numa_node = local_memory_node(cpu_to_node(i));
2088         }
2089 }
2090
2091 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx)
2092 {
2093         int ret = 0;
2094
2095         set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
2096                                         set->queue_depth, set->reserved_tags);
2097         if (!set->tags[hctx_idx])
2098                 return false;
2099
2100         ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2101                                 set->queue_depth);
2102         if (!ret)
2103                 return true;
2104
2105         blk_mq_free_rq_map(set->tags[hctx_idx]);
2106         set->tags[hctx_idx] = NULL;
2107         return false;
2108 }
2109
2110 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2111                                          unsigned int hctx_idx)
2112 {
2113         if (set->tags[hctx_idx]) {
2114                 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2115                 blk_mq_free_rq_map(set->tags[hctx_idx]);
2116                 set->tags[hctx_idx] = NULL;
2117         }
2118 }
2119
2120 static void blk_mq_map_swqueue(struct request_queue *q)
2121 {
2122         unsigned int i, hctx_idx;
2123         struct blk_mq_hw_ctx *hctx;
2124         struct blk_mq_ctx *ctx;
2125         struct blk_mq_tag_set *set = q->tag_set;
2126
2127         /*
2128          * Avoid others reading imcomplete hctx->cpumask through sysfs
2129          */
2130         mutex_lock(&q->sysfs_lock);
2131
2132         queue_for_each_hw_ctx(q, hctx, i) {
2133                 cpumask_clear(hctx->cpumask);
2134                 hctx->nr_ctx = 0;
2135         }
2136
2137         /*
2138          * Map software to hardware queues.
2139          *
2140          * If the cpu isn't present, the cpu is mapped to first hctx.
2141          */
2142         for_each_present_cpu(i) {
2143                 hctx_idx = q->mq_map[i];
2144                 /* unmapped hw queue can be remapped after CPU topo changed */
2145                 if (!set->tags[hctx_idx] &&
2146                     !__blk_mq_alloc_rq_map(set, hctx_idx)) {
2147                         /*
2148                          * If tags initialization fail for some hctx,
2149                          * that hctx won't be brought online.  In this
2150                          * case, remap the current ctx to hctx[0] which
2151                          * is guaranteed to always have tags allocated
2152                          */
2153                         q->mq_map[i] = 0;
2154                 }
2155
2156                 ctx = per_cpu_ptr(q->queue_ctx, i);
2157                 hctx = blk_mq_map_queue(q, i);
2158
2159                 cpumask_set_cpu(i, hctx->cpumask);
2160                 ctx->index_hw = hctx->nr_ctx;
2161                 hctx->ctxs[hctx->nr_ctx++] = ctx;
2162         }
2163
2164         mutex_unlock(&q->sysfs_lock);
2165
2166         queue_for_each_hw_ctx(q, hctx, i) {
2167                 /*
2168                  * If no software queues are mapped to this hardware queue,
2169                  * disable it and free the request entries.
2170                  */
2171                 if (!hctx->nr_ctx) {
2172                         /* Never unmap queue 0.  We need it as a
2173                          * fallback in case of a new remap fails
2174                          * allocation
2175                          */
2176                         if (i && set->tags[i])
2177                                 blk_mq_free_map_and_requests(set, i);
2178
2179                         hctx->tags = NULL;
2180                         continue;
2181                 }
2182
2183                 hctx->tags = set->tags[i];
2184                 WARN_ON(!hctx->tags);
2185
2186                 /*
2187                  * Set the map size to the number of mapped software queues.
2188                  * This is more accurate and more efficient than looping
2189                  * over all possibly mapped software queues.
2190                  */
2191                 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2192
2193                 /*
2194                  * Initialize batch roundrobin counts
2195                  */
2196                 hctx->next_cpu = cpumask_first(hctx->cpumask);
2197                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2198         }
2199 }
2200
2201 /*
2202  * Caller needs to ensure that we're either frozen/quiesced, or that
2203  * the queue isn't live yet.
2204  */
2205 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2206 {
2207         struct blk_mq_hw_ctx *hctx;
2208         int i;
2209
2210         queue_for_each_hw_ctx(q, hctx, i) {
2211                 if (shared) {
2212                         if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
2213                                 atomic_inc(&q->shared_hctx_restart);
2214                         hctx->flags |= BLK_MQ_F_TAG_SHARED;
2215                 } else {
2216                         if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
2217                                 atomic_dec(&q->shared_hctx_restart);
2218                         hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2219                 }
2220         }
2221 }
2222
2223 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set,
2224                                         bool shared)
2225 {
2226         struct request_queue *q;
2227
2228         lockdep_assert_held(&set->tag_list_lock);
2229
2230         list_for_each_entry(q, &set->tag_list, tag_set_list) {
2231                 blk_mq_freeze_queue(q);
2232                 queue_set_hctx_shared(q, shared);
2233                 blk_mq_unfreeze_queue(q);
2234         }
2235 }
2236
2237 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2238 {
2239         struct blk_mq_tag_set *set = q->tag_set;
2240
2241         mutex_lock(&set->tag_list_lock);
2242         list_del_rcu(&q->tag_set_list);
2243         INIT_LIST_HEAD(&q->tag_set_list);
2244         if (list_is_singular(&set->tag_list)) {
2245                 /* just transitioned to unshared */
2246                 set->flags &= ~BLK_MQ_F_TAG_SHARED;
2247                 /* update existing queue */
2248                 blk_mq_update_tag_set_depth(set, false);
2249         }
2250         mutex_unlock(&set->tag_list_lock);
2251
2252         synchronize_rcu();
2253 }
2254
2255 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2256                                      struct request_queue *q)
2257 {
2258         q->tag_set = set;
2259
2260         mutex_lock(&set->tag_list_lock);
2261
2262         /* Check to see if we're transitioning to shared (from 1 to 2 queues). */
2263         if (!list_empty(&set->tag_list) && !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2264                 set->flags |= BLK_MQ_F_TAG_SHARED;
2265                 /* update existing queue */
2266                 blk_mq_update_tag_set_depth(set, true);
2267         }
2268         if (set->flags & BLK_MQ_F_TAG_SHARED)
2269                 queue_set_hctx_shared(q, true);
2270         list_add_tail_rcu(&q->tag_set_list, &set->tag_list);
2271
2272         mutex_unlock(&set->tag_list_lock);
2273 }
2274
2275 /*
2276  * It is the actual release handler for mq, but we do it from
2277  * request queue's release handler for avoiding use-after-free
2278  * and headache because q->mq_kobj shouldn't have been introduced,
2279  * but we can't group ctx/kctx kobj without it.
2280  */
2281 void blk_mq_release(struct request_queue *q)
2282 {
2283         struct blk_mq_hw_ctx *hctx;
2284         unsigned int i;
2285
2286         /* hctx kobj stays in hctx */
2287         queue_for_each_hw_ctx(q, hctx, i) {
2288                 if (!hctx)
2289                         continue;
2290                 kobject_put(&hctx->kobj);
2291         }
2292
2293         q->mq_map = NULL;
2294
2295         kfree(q->queue_hw_ctx);
2296
2297         /*
2298          * release .mq_kobj and sw queue's kobject now because
2299          * both share lifetime with request queue.
2300          */
2301         blk_mq_sysfs_deinit(q);
2302
2303         free_percpu(q->queue_ctx);
2304 }
2305
2306 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
2307 {
2308         struct request_queue *uninit_q, *q;
2309
2310         uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
2311         if (!uninit_q)
2312                 return ERR_PTR(-ENOMEM);
2313
2314         q = blk_mq_init_allocated_queue(set, uninit_q);
2315         if (IS_ERR(q))
2316                 blk_cleanup_queue(uninit_q);
2317
2318         return q;
2319 }
2320 EXPORT_SYMBOL(blk_mq_init_queue);
2321
2322 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set)
2323 {
2324         int hw_ctx_size = sizeof(struct blk_mq_hw_ctx);
2325
2326         BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, queue_rq_srcu),
2327                            __alignof__(struct blk_mq_hw_ctx)) !=
2328                      sizeof(struct blk_mq_hw_ctx));
2329
2330         if (tag_set->flags & BLK_MQ_F_BLOCKING)
2331                 hw_ctx_size += sizeof(struct srcu_struct);
2332
2333         return hw_ctx_size;
2334 }
2335
2336 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
2337                                                 struct request_queue *q)
2338 {
2339         int i, j;
2340         struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
2341
2342         blk_mq_sysfs_unregister(q);
2343         for (i = 0; i < set->nr_hw_queues; i++) {
2344                 int node;
2345
2346                 if (hctxs[i])
2347                         continue;
2348
2349                 node = blk_mq_hw_queue_to_node(q->mq_map, i);
2350                 hctxs[i] = kzalloc_node(blk_mq_hw_ctx_size(set),
2351                                         GFP_KERNEL, node);
2352                 if (!hctxs[i])
2353                         break;
2354
2355                 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
2356                                                 node)) {
2357                         kfree(hctxs[i]);
2358                         hctxs[i] = NULL;
2359                         break;
2360                 }
2361
2362                 atomic_set(&hctxs[i]->nr_active, 0);
2363                 hctxs[i]->numa_node = node;
2364                 hctxs[i]->queue_num = i;
2365
2366                 if (blk_mq_init_hctx(q, set, hctxs[i], i)) {
2367                         free_cpumask_var(hctxs[i]->cpumask);
2368                         kfree(hctxs[i]);
2369                         hctxs[i] = NULL;
2370                         break;
2371                 }
2372                 blk_mq_hctx_kobj_init(hctxs[i]);
2373         }
2374         for (j = i; j < q->nr_hw_queues; j++) {
2375                 struct blk_mq_hw_ctx *hctx = hctxs[j];
2376
2377                 if (hctx) {
2378                         if (hctx->tags)
2379                                 blk_mq_free_map_and_requests(set, j);
2380                         blk_mq_exit_hctx(q, set, hctx, j);
2381                         kobject_put(&hctx->kobj);
2382                         hctxs[j] = NULL;
2383
2384                 }
2385         }
2386         q->nr_hw_queues = i;
2387         blk_mq_sysfs_register(q);
2388 }
2389
2390 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2391                                                   struct request_queue *q)
2392 {
2393         /* mark the queue as mq asap */
2394         q->mq_ops = set->ops;
2395
2396         q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
2397                                              blk_mq_poll_stats_bkt,
2398                                              BLK_MQ_POLL_STATS_BKTS, q);
2399         if (!q->poll_cb)
2400                 goto err_exit;
2401
2402         q->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2403         if (!q->queue_ctx)
2404                 goto err_exit;
2405
2406         /* init q->mq_kobj and sw queues' kobjects */
2407         blk_mq_sysfs_init(q);
2408
2409         q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)),
2410                                                 GFP_KERNEL, set->numa_node);
2411         if (!q->queue_hw_ctx)
2412                 goto err_percpu;
2413
2414         q->mq_map = set->mq_map;
2415
2416         blk_mq_realloc_hw_ctxs(set, q);
2417         if (!q->nr_hw_queues)
2418                 goto err_hctxs;
2419
2420         INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2421         blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2422
2423         q->nr_queues = nr_cpu_ids;
2424
2425         q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2426
2427         if (!(set->flags & BLK_MQ_F_SG_MERGE))
2428                 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
2429
2430         q->sg_reserved_size = INT_MAX;
2431
2432         INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
2433         INIT_LIST_HEAD(&q->requeue_list);
2434         spin_lock_init(&q->requeue_lock);
2435
2436         blk_queue_make_request(q, blk_mq_make_request);
2437
2438         /*
2439          * Do this after blk_queue_make_request() overrides it...
2440          */
2441         q->nr_requests = set->queue_depth;
2442
2443         /*
2444          * Default to classic polling
2445          */
2446         q->poll_nsec = -1;
2447
2448         if (set->ops->complete)
2449                 blk_queue_softirq_done(q, set->ops->complete);
2450
2451         blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2452         blk_mq_add_queue_tag_set(set, q);
2453         blk_mq_map_swqueue(q);
2454
2455         if (!(set->flags & BLK_MQ_F_NO_SCHED)) {
2456                 int ret;
2457
2458                 ret = blk_mq_sched_init(q);
2459                 if (ret)
2460                         return ERR_PTR(ret);
2461         }
2462
2463         return q;
2464
2465 err_hctxs:
2466         kfree(q->queue_hw_ctx);
2467 err_percpu:
2468         free_percpu(q->queue_ctx);
2469 err_exit:
2470         q->mq_ops = NULL;
2471         return ERR_PTR(-ENOMEM);
2472 }
2473 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2474
2475 void blk_mq_free_queue(struct request_queue *q)
2476 {
2477         struct blk_mq_tag_set   *set = q->tag_set;
2478
2479         blk_mq_del_queue_tag_set(q);
2480         blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2481 }
2482
2483 /* Basically redo blk_mq_init_queue with queue frozen */
2484 static void blk_mq_queue_reinit(struct request_queue *q)
2485 {
2486         WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth));
2487
2488         blk_mq_debugfs_unregister_hctxs(q);
2489         blk_mq_sysfs_unregister(q);
2490
2491         /*
2492          * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2493          * we should change hctx numa_node according to new topology (this
2494          * involves free and re-allocate memory, worthy doing?)
2495          */
2496
2497         blk_mq_map_swqueue(q);
2498
2499         blk_mq_sysfs_register(q);
2500         blk_mq_debugfs_register_hctxs(q);
2501 }
2502
2503 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2504 {
2505         int i;
2506
2507         for (i = 0; i < set->nr_hw_queues; i++)
2508                 if (!__blk_mq_alloc_rq_map(set, i))
2509                         goto out_unwind;
2510
2511         return 0;
2512
2513 out_unwind:
2514         while (--i >= 0)
2515                 blk_mq_free_rq_map(set->tags[i]);
2516
2517         return -ENOMEM;
2518 }
2519
2520 /*
2521  * Allocate the request maps associated with this tag_set. Note that this
2522  * may reduce the depth asked for, if memory is tight. set->queue_depth
2523  * will be updated to reflect the allocated depth.
2524  */
2525 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2526 {
2527         unsigned int depth;
2528         int err;
2529
2530         depth = set->queue_depth;
2531         do {
2532                 err = __blk_mq_alloc_rq_maps(set);
2533                 if (!err)
2534                         break;
2535
2536                 set->queue_depth >>= 1;
2537                 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2538                         err = -ENOMEM;
2539                         break;
2540                 }
2541         } while (set->queue_depth);
2542
2543         if (!set->queue_depth || err) {
2544                 pr_err("blk-mq: failed to allocate request map\n");
2545                 return -ENOMEM;
2546         }
2547
2548         if (depth != set->queue_depth)
2549                 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2550                                                 depth, set->queue_depth);
2551
2552         return 0;
2553 }
2554
2555 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
2556 {
2557         if (set->ops->map_queues)
2558                 return set->ops->map_queues(set);
2559         else
2560                 return blk_mq_map_queues(set);
2561 }
2562
2563 /*
2564  * Alloc a tag set to be associated with one or more request queues.
2565  * May fail with EINVAL for various error conditions. May adjust the
2566  * requested depth down, if if it too large. In that case, the set
2567  * value will be stored in set->queue_depth.
2568  */
2569 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2570 {
2571         int ret;
2572
2573         BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2574
2575         if (!set->nr_hw_queues)
2576                 return -EINVAL;
2577         if (!set->queue_depth)
2578                 return -EINVAL;
2579         if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2580                 return -EINVAL;
2581
2582         if (!set->ops->queue_rq)
2583                 return -EINVAL;
2584
2585         if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2586                 pr_info("blk-mq: reduced tag depth to %u\n",
2587                         BLK_MQ_MAX_DEPTH);
2588                 set->queue_depth = BLK_MQ_MAX_DEPTH;
2589         }
2590
2591         /*
2592          * If a crashdump is active, then we are potentially in a very
2593          * memory constrained environment. Limit us to 1 queue and
2594          * 64 tags to prevent using too much memory.
2595          */
2596         if (is_kdump_kernel()) {
2597                 set->nr_hw_queues = 1;
2598                 set->queue_depth = min(64U, set->queue_depth);
2599         }
2600         /*
2601          * There is no use for more h/w queues than cpus.
2602          */
2603         if (set->nr_hw_queues > nr_cpu_ids)
2604                 set->nr_hw_queues = nr_cpu_ids;
2605
2606         set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *),
2607                                  GFP_KERNEL, set->numa_node);
2608         if (!set->tags)
2609                 return -ENOMEM;
2610
2611         ret = -ENOMEM;
2612         set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids,
2613                         GFP_KERNEL, set->numa_node);
2614         if (!set->mq_map)
2615                 goto out_free_tags;
2616
2617         ret = blk_mq_update_queue_map(set);
2618         if (ret)
2619                 goto out_free_mq_map;
2620
2621         ret = blk_mq_alloc_rq_maps(set);
2622         if (ret)
2623                 goto out_free_mq_map;
2624
2625         mutex_init(&set->tag_list_lock);
2626         INIT_LIST_HEAD(&set->tag_list);
2627
2628         return 0;
2629
2630 out_free_mq_map:
2631         kfree(set->mq_map);
2632         set->mq_map = NULL;
2633 out_free_tags:
2634         kfree(set->tags);
2635         set->tags = NULL;
2636         return ret;
2637 }
2638 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2639
2640 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2641 {
2642         int i;
2643
2644         for (i = 0; i < nr_cpu_ids; i++)
2645                 blk_mq_free_map_and_requests(set, i);
2646
2647         kfree(set->mq_map);
2648         set->mq_map = NULL;
2649
2650         kfree(set->tags);
2651         set->tags = NULL;
2652 }
2653 EXPORT_SYMBOL(blk_mq_free_tag_set);
2654
2655 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2656 {
2657         struct blk_mq_tag_set *set = q->tag_set;
2658         struct blk_mq_hw_ctx *hctx;
2659         int i, ret;
2660
2661         if (!set)
2662                 return -EINVAL;
2663
2664         blk_mq_freeze_queue(q);
2665
2666         ret = 0;
2667         queue_for_each_hw_ctx(q, hctx, i) {
2668                 if (!hctx->tags)
2669                         continue;
2670                 /*
2671                  * If we're using an MQ scheduler, just update the scheduler
2672                  * queue depth. This is similar to what the old code would do.
2673                  */
2674                 if (!hctx->sched_tags) {
2675                         ret = blk_mq_tag_update_depth(hctx, &hctx->tags,
2676                                                         min(nr, set->queue_depth),
2677                                                         false);
2678                 } else {
2679                         ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
2680                                                         nr, true);
2681                 }
2682                 if (ret)
2683                         break;
2684         }
2685
2686         if (!ret)
2687                 q->nr_requests = nr;
2688
2689         blk_mq_unfreeze_queue(q);
2690
2691         return ret;
2692 }
2693
2694 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
2695                                                         int nr_hw_queues)
2696 {
2697         struct request_queue *q;
2698
2699         lockdep_assert_held(&set->tag_list_lock);
2700
2701         if (nr_hw_queues > nr_cpu_ids)
2702                 nr_hw_queues = nr_cpu_ids;
2703         if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
2704                 return;
2705
2706         list_for_each_entry(q, &set->tag_list, tag_set_list)
2707                 blk_mq_freeze_queue(q);
2708
2709         set->nr_hw_queues = nr_hw_queues;
2710         blk_mq_update_queue_map(set);
2711         list_for_each_entry(q, &set->tag_list, tag_set_list) {
2712                 blk_mq_realloc_hw_ctxs(set, q);
2713                 blk_mq_queue_reinit(q);
2714         }
2715
2716         list_for_each_entry(q, &set->tag_list, tag_set_list)
2717                 blk_mq_unfreeze_queue(q);
2718 }
2719
2720 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
2721 {
2722         mutex_lock(&set->tag_list_lock);
2723         __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
2724         mutex_unlock(&set->tag_list_lock);
2725 }
2726 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
2727
2728 /* Enable polling stats and return whether they were already enabled. */
2729 static bool blk_poll_stats_enable(struct request_queue *q)
2730 {
2731         if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
2732             test_and_set_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags))
2733                 return true;
2734         blk_stat_add_callback(q, q->poll_cb);
2735         return false;
2736 }
2737
2738 static void blk_mq_poll_stats_start(struct request_queue *q)
2739 {
2740         /*
2741          * We don't arm the callback if polling stats are not enabled or the
2742          * callback is already active.
2743          */
2744         if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
2745             blk_stat_is_active(q->poll_cb))
2746                 return;
2747
2748         blk_stat_activate_msecs(q->poll_cb, 100);
2749 }
2750
2751 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
2752 {
2753         struct request_queue *q = cb->data;
2754         int bucket;
2755
2756         for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
2757                 if (cb->stat[bucket].nr_samples)
2758                         q->poll_stat[bucket] = cb->stat[bucket];
2759         }
2760 }
2761
2762 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
2763                                        struct blk_mq_hw_ctx *hctx,
2764                                        struct request *rq)
2765 {
2766         unsigned long ret = 0;
2767         int bucket;
2768
2769         /*
2770          * If stats collection isn't on, don't sleep but turn it on for
2771          * future users
2772          */
2773         if (!blk_poll_stats_enable(q))
2774                 return 0;
2775
2776         /*
2777          * As an optimistic guess, use half of the mean service time
2778          * for this type of request. We can (and should) make this smarter.
2779          * For instance, if the completion latencies are tight, we can
2780          * get closer than just half the mean. This is especially
2781          * important on devices where the completion latencies are longer
2782          * than ~10 usec. We do use the stats for the relevant IO size
2783          * if available which does lead to better estimates.
2784          */
2785         bucket = blk_mq_poll_stats_bkt(rq);
2786         if (bucket < 0)
2787                 return ret;
2788
2789         if (q->poll_stat[bucket].nr_samples)
2790                 ret = (q->poll_stat[bucket].mean + 1) / 2;
2791
2792         return ret;
2793 }
2794
2795 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
2796                                      struct blk_mq_hw_ctx *hctx,
2797                                      struct request *rq)
2798 {
2799         struct hrtimer_sleeper hs;
2800         enum hrtimer_mode mode;
2801         unsigned int nsecs;
2802         ktime_t kt;
2803
2804         if (test_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags))
2805                 return false;
2806
2807         /*
2808          * poll_nsec can be:
2809          *
2810          * -1:  don't ever hybrid sleep
2811          *  0:  use half of prev avg
2812          * >0:  use this specific value
2813          */
2814         if (q->poll_nsec == -1)
2815                 return false;
2816         else if (q->poll_nsec > 0)
2817                 nsecs = q->poll_nsec;
2818         else
2819                 nsecs = blk_mq_poll_nsecs(q, hctx, rq);
2820
2821         if (!nsecs)
2822                 return false;
2823
2824         set_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags);
2825
2826         /*
2827          * This will be replaced with the stats tracking code, using
2828          * 'avg_completion_time / 2' as the pre-sleep target.
2829          */
2830         kt = nsecs;
2831
2832         mode = HRTIMER_MODE_REL;
2833         hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode);
2834         hrtimer_set_expires(&hs.timer, kt);
2835
2836         hrtimer_init_sleeper(&hs, current);
2837         do {
2838                 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
2839                         break;
2840                 set_current_state(TASK_UNINTERRUPTIBLE);
2841                 hrtimer_start_expires(&hs.timer, mode);
2842                 if (hs.task)
2843                         io_schedule();
2844                 hrtimer_cancel(&hs.timer);
2845                 mode = HRTIMER_MODE_ABS;
2846         } while (hs.task && !signal_pending(current));
2847
2848         __set_current_state(TASK_RUNNING);
2849         destroy_hrtimer_on_stack(&hs.timer);
2850         return true;
2851 }
2852
2853 static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq)
2854 {
2855         struct request_queue *q = hctx->queue;
2856         long state;
2857
2858         /*
2859          * If we sleep, have the caller restart the poll loop to reset
2860          * the state. Like for the other success return cases, the
2861          * caller is responsible for checking if the IO completed. If
2862          * the IO isn't complete, we'll get called again and will go
2863          * straight to the busy poll loop.
2864          */
2865         if (blk_mq_poll_hybrid_sleep(q, hctx, rq))
2866                 return true;
2867
2868         hctx->poll_considered++;
2869
2870         state = current->state;
2871         while (!need_resched()) {
2872                 int ret;
2873
2874                 hctx->poll_invoked++;
2875
2876                 ret = q->mq_ops->poll(hctx, rq->tag);
2877                 if (ret > 0) {
2878                         hctx->poll_success++;
2879                         set_current_state(TASK_RUNNING);
2880                         return true;
2881                 }
2882
2883                 if (signal_pending_state(state, current))
2884                         set_current_state(TASK_RUNNING);
2885
2886                 if (current->state == TASK_RUNNING)
2887                         return true;
2888                 if (ret < 0)
2889                         break;
2890                 cpu_relax();
2891         }
2892
2893         return false;
2894 }
2895
2896 bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie)
2897 {
2898         struct blk_mq_hw_ctx *hctx;
2899         struct blk_plug *plug;
2900         struct request *rq;
2901
2902         if (!q->mq_ops || !q->mq_ops->poll || !blk_qc_t_valid(cookie) ||
2903             !test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
2904                 return false;
2905
2906         plug = current->plug;
2907         if (plug)
2908                 blk_flush_plug_list(plug, false);
2909
2910         hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
2911         if (!blk_qc_t_is_internal(cookie))
2912                 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
2913         else {
2914                 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
2915                 /*
2916                  * With scheduling, if the request has completed, we'll
2917                  * get a NULL return here, as we clear the sched tag when
2918                  * that happens. The request still remains valid, like always,
2919                  * so we should be safe with just the NULL check.
2920                  */
2921                 if (!rq)
2922                         return false;
2923         }
2924
2925         return __blk_mq_poll(hctx, rq);
2926 }
2927 EXPORT_SYMBOL_GPL(blk_mq_poll);
2928
2929 static int __init blk_mq_init(void)
2930 {
2931         /*
2932          * See comment in block/blk.h rq_atomic_flags enum
2933          */
2934         BUILD_BUG_ON((REQ_ATOM_STARTED / BITS_PER_BYTE) !=
2935                         (REQ_ATOM_COMPLETE / BITS_PER_BYTE));
2936
2937         cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
2938                                 blk_mq_hctx_notify_dead);
2939         return 0;
2940 }
2941 subsys_initcall(blk_mq_init);