drivers/mmc/core/queue.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  *  Copyright (C) 2003 Russell King, All Rights Reserved.
   4  *  Copyright 2006-2007 Pierre Ossman
   5  */
   6 #include <linux/slab.h>
   7 #include <linux/module.h>
   8 #include <linux/blkdev.h>
   9 #include <linux/freezer.h>
  10 #include <linux/scatterlist.h>
  11 #include <linux/dma-mapping.h>
  12 #include <linux/backing-dev.h>
  13
  14 #include <linux/mmc/card.h>
  15 #include <linux/mmc/host.h>
  16
  17 #include "queue.h"
  18 #include "block.h"
  19 #include "core.h"
  20 #include "card.h"
  21 #include "crypto.h"
  22 #include "host.h"
  23
  24 #define MMC_DMA_MAP_MERGE_SEGMENTS      512
  25
  26 static inline bool mmc_cqe_dcmd_busy(struct mmc_queue *mq)
  27 {
  28         /* Allow only 1 DCMD at a time */
  29         return mq->in_flight[MMC_ISSUE_DCMD];
  30 }
  31
  32 void mmc_cqe_check_busy(struct mmc_queue *mq)
  33 {
  34         if ((mq->cqe_busy & MMC_CQE_DCMD_BUSY) && !mmc_cqe_dcmd_busy(mq))
  35                 mq->cqe_busy &= ~MMC_CQE_DCMD_BUSY;
  36 }
  37
  38 static inline bool mmc_cqe_can_dcmd(struct mmc_host *host)
  39 {
  40         return host->caps2 & MMC_CAP2_CQE_DCMD;
  41 }
  42
  43 static enum mmc_issue_type mmc_cqe_issue_type(struct mmc_host *host,
  44                                               struct request *req)
  45 {
  46         switch (req_op(req)) {
  47         case REQ_OP_DRV_IN:
  48         case REQ_OP_DRV_OUT:
  49         case REQ_OP_DISCARD:
  50         case REQ_OP_SECURE_ERASE:
  51         case REQ_OP_WRITE_ZEROES:
  52                 return MMC_ISSUE_SYNC;
  53         case REQ_OP_FLUSH:
  54                 return mmc_cqe_can_dcmd(host) ? MMC_ISSUE_DCMD : MMC_ISSUE_SYNC;
  55         default:
  56                 return MMC_ISSUE_ASYNC;
  57         }
  58 }
  59
  60 enum mmc_issue_type mmc_issue_type(struct mmc_queue *mq, struct request *req)
  61 {
  62         struct mmc_host *host = mq->card->host;
  63
  64         if (host->cqe_enabled && !host->hsq_enabled)
  65                 return mmc_cqe_issue_type(host, req);
  66
  67         if (req_op(req) == REQ_OP_READ || req_op(req) == REQ_OP_WRITE)
  68                 return MMC_ISSUE_ASYNC;
  69
  70         return MMC_ISSUE_SYNC;
  71 }
  72
  73 static void __mmc_cqe_recovery_notifier(struct mmc_queue *mq)
  74 {
  75         if (!mq->recovery_needed) {
  76                 mq->recovery_needed = true;
  77                 schedule_work(&mq->recovery_work);
  78         }
  79 }
  80
  81 void mmc_cqe_recovery_notifier(struct mmc_request *mrq)
  82 {
  83         struct mmc_queue_req *mqrq = container_of(mrq, struct mmc_queue_req,
  84                                                   brq.mrq);
  85         struct request *req = mmc_queue_req_to_req(mqrq);
  86         struct request_queue *q = req->q;
  87         struct mmc_queue *mq = q->queuedata;
  88         unsigned long flags;
  89
  90         spin_lock_irqsave(&mq->lock, flags);
  91         __mmc_cqe_recovery_notifier(mq);
  92         spin_unlock_irqrestore(&mq->lock, flags);
  93 }
  94
  95 static enum blk_eh_timer_return mmc_cqe_timed_out(struct request *req)
  96 {
  97         struct mmc_queue_req *mqrq = req_to_mmc_queue_req(req);
  98         struct mmc_request *mrq = &mqrq->brq.mrq;
  99         struct mmc_queue *mq = req->q->queuedata;
 100         struct mmc_host *host = mq->card->host;
 101         enum mmc_issue_type issue_type = mmc_issue_type(mq, req);
 102         bool recovery_needed = false;
 103
 104         switch (issue_type) {
 105         case MMC_ISSUE_ASYNC:
 106         case MMC_ISSUE_DCMD:
 107                 if (host->cqe_ops->cqe_timeout(host, mrq, &recovery_needed)) {
 108                         if (recovery_needed)
 109                                 mmc_cqe_recovery_notifier(mrq);
 110                         return BLK_EH_RESET_TIMER;
 111                 }
 112                 /* The request has gone already */
 113                 return BLK_EH_DONE;
 114         default:
 115                 /* Timeout is handled by mmc core */
 116                 return BLK_EH_RESET_TIMER;
 117         }
 118 }
 119
 120 static enum blk_eh_timer_return mmc_mq_timed_out(struct request *req)
 121 {
 122         struct request_queue *q = req->q;
 123         struct mmc_queue *mq = q->queuedata;
 124         struct mmc_card *card = mq->card;
 125         struct mmc_host *host = card->host;
 126         unsigned long flags;
 127         bool ignore_tout;
 128
 129         spin_lock_irqsave(&mq->lock, flags);
 130         ignore_tout = mq->recovery_needed || !host->cqe_enabled || host->hsq_enabled;
 131         spin_unlock_irqrestore(&mq->lock, flags);
 132
 133         return ignore_tout ? BLK_EH_RESET_TIMER : mmc_cqe_timed_out(req);
 134 }
 135
 136 static void mmc_mq_recovery_handler(struct work_struct *work)
 137 {
 138         struct mmc_queue *mq = container_of(work, struct mmc_queue,
 139                                             recovery_work);
 140         struct request_queue *q = mq->queue;
 141         struct mmc_host *host = mq->card->host;
 142
 143         mmc_get_card(mq->card, &mq->ctx);
 144
 145         mq->in_recovery = true;
 146
 147         if (host->cqe_enabled && !host->hsq_enabled)
 148                 mmc_blk_cqe_recovery(mq);
 149         else
 150                 mmc_blk_mq_recovery(mq);
 151
 152         mq->in_recovery = false;
 153
 154         spin_lock_irq(&mq->lock);
 155         mq->recovery_needed = false;
 156         spin_unlock_irq(&mq->lock);
 157
 158         if (host->hsq_enabled)
 159                 host->cqe_ops->cqe_recovery_finish(host);
 160
 161         mmc_put_card(mq->card, &mq->ctx);
 162
 163         blk_mq_run_hw_queues(q, true);
 164 }
 165
 166 static struct scatterlist *mmc_alloc_sg(unsigned short sg_len, gfp_t gfp)
 167 {
 168         struct scatterlist *sg;
 169
 170         sg = kmalloc_array(sg_len, sizeof(*sg), gfp);
 171         if (sg)
 172                 sg_init_table(sg, sg_len);
 173
 174         return sg;
 175 }
 176
 177 static void mmc_queue_setup_discard(struct mmc_card *card,
 178                 struct queue_limits *lim)
 179 {
 180         unsigned max_discard;
 181
 182         max_discard = mmc_calc_max_discard(card);
 183         if (!max_discard)
 184                 return;
 185
 186         lim->max_hw_discard_sectors = max_discard;
 187         if (mmc_can_secure_erase_trim(card))
 188                 lim->max_secure_erase_sectors = max_discard;
 189         if (mmc_can_trim(card) && card->erased_byte == 0)
 190                 lim->max_write_zeroes_sectors = max_discard;
 191
 192         /* granularity must not be greater than max. discard */
 193         if (card->pref_erase > max_discard)
 194                 lim->discard_granularity = SECTOR_SIZE;
 195         else
 196                 lim->discard_granularity = card->pref_erase << 9;
 197 }
 198
 199 static unsigned short mmc_get_max_segments(struct mmc_host *host)
 200 {
 201         return host->can_dma_map_merge ? MMC_DMA_MAP_MERGE_SEGMENTS :
 202                                          host->max_segs;
 203 }
 204
 205 static int mmc_mq_init_request(struct blk_mq_tag_set *set, struct request *req,
 206                                unsigned int hctx_idx, unsigned int numa_node)
 207 {
 208         struct mmc_queue_req *mq_rq = req_to_mmc_queue_req(req);
 209         struct mmc_queue *mq = set->driver_data;
 210         struct mmc_card *card = mq->card;
 211         struct mmc_host *host = card->host;
 212
 213         mq_rq->sg = mmc_alloc_sg(mmc_get_max_segments(host), GFP_KERNEL);
 214         if (!mq_rq->sg)
 215                 return -ENOMEM;
 216
 217         return 0;
 218 }
 219
 220 static void mmc_mq_exit_request(struct blk_mq_tag_set *set, struct request *req,
 221                                 unsigned int hctx_idx)
 222 {
 223         struct mmc_queue_req *mq_rq = req_to_mmc_queue_req(req);
 224
 225         kfree(mq_rq->sg);
 226         mq_rq->sg = NULL;
 227 }
 228
 229 static blk_status_t mmc_mq_queue_rq(struct blk_mq_hw_ctx *hctx,
 230                                     const struct blk_mq_queue_data *bd)
 231 {
 232         struct request *req = bd->rq;
 233         struct request_queue *q = req->q;
 234         struct mmc_queue *mq = q->queuedata;
 235         struct mmc_card *card = mq->card;
 236         struct mmc_host *host = card->host;
 237         enum mmc_issue_type issue_type;
 238         enum mmc_issued issued;
 239         bool get_card, cqe_retune_ok;
 240         blk_status_t ret;
 241
 242         if (mmc_card_removed(mq->card)) {
 243                 req->rq_flags |= RQF_QUIET;
 244                 return BLK_STS_IOERR;
 245         }
 246
 247         issue_type = mmc_issue_type(mq, req);
 248
 249         spin_lock_irq(&mq->lock);
 250
 251         if (mq->recovery_needed || mq->busy) {
 252                 spin_unlock_irq(&mq->lock);
 253                 return BLK_STS_RESOURCE;
 254         }
 255
 256         switch (issue_type) {
 257         case MMC_ISSUE_DCMD:
 258                 if (mmc_cqe_dcmd_busy(mq)) {
 259                         mq->cqe_busy |= MMC_CQE_DCMD_BUSY;
 260                         spin_unlock_irq(&mq->lock);
 261                         return BLK_STS_RESOURCE;
 262                 }
 263                 break;
 264         case MMC_ISSUE_ASYNC:
 265                 if (host->hsq_enabled && mq->in_flight[issue_type] > host->hsq_depth) {
 266                         spin_unlock_irq(&mq->lock);
 267                         return BLK_STS_RESOURCE;
 268                 }
 269                 break;
 270         default:
 271                 /*
 272                  * Timeouts are handled by mmc core, and we don't have a host
 273                  * API to abort requests, so we can't handle the timeout anyway.
 274                  * However, when the timeout happens, blk_mq_complete_request()
 275                  * no longer works (to stop the request disappearing under us).
 276                  * To avoid racing with that, set a large timeout.
 277                  */
 278                 req->timeout = 600 * HZ;
 279                 break;
 280         }
 281
 282         /* Parallel dispatch of requests is not supported at the moment */
 283         mq->busy = true;
 284
 285         mq->in_flight[issue_type] += 1;
 286         get_card = (mmc_tot_in_flight(mq) == 1);
 287         cqe_retune_ok = (mmc_cqe_qcnt(mq) == 1);
 288
 289         spin_unlock_irq(&mq->lock);
 290
 291         if (!(req->rq_flags & RQF_DONTPREP)) {
 292                 req_to_mmc_queue_req(req)->retries = 0;
 293                 req->rq_flags |= RQF_DONTPREP;
 294         }
 295
 296         if (get_card)
 297                 mmc_get_card(card, &mq->ctx);
 298
 299         if (host->cqe_enabled) {
 300                 host->retune_now = host->need_retune && cqe_retune_ok &&
 301                                    !host->hold_retune;
 302         }
 303
 304         blk_mq_start_request(req);
 305
 306         issued = mmc_blk_mq_issue_rq(mq, req);
 307
 308         switch (issued) {
 309         case MMC_REQ_BUSY:
 310                 ret = BLK_STS_RESOURCE;
 311                 break;
 312         case MMC_REQ_FAILED_TO_START:
 313                 ret = BLK_STS_IOERR;
 314                 break;
 315         default:
 316                 ret = BLK_STS_OK;
 317                 break;
 318         }
 319
 320         if (issued != MMC_REQ_STARTED) {
 321                 bool put_card = false;
 322
 323                 spin_lock_irq(&mq->lock);
 324                 mq->in_flight[issue_type] -= 1;
 325                 if (mmc_tot_in_flight(mq) == 0)
 326                         put_card = true;
 327                 mq->busy = false;
 328                 spin_unlock_irq(&mq->lock);
 329                 if (put_card)
 330                         mmc_put_card(card, &mq->ctx);
 331         } else {
 332                 WRITE_ONCE(mq->busy, false);
 333         }
 334
 335         return ret;
 336 }
 337
 338 static const struct blk_mq_ops mmc_mq_ops = {
 339         .queue_rq       = mmc_mq_queue_rq,
 340         .init_request   = mmc_mq_init_request,
 341         .exit_request   = mmc_mq_exit_request,
 342         .complete       = mmc_blk_mq_complete,
 343         .timeout        = mmc_mq_timed_out,
 344 };
 345
 346 static struct gendisk *mmc_alloc_disk(struct mmc_queue *mq,
 347                 struct mmc_card *card, unsigned int features)
 348 {
 349         struct mmc_host *host = card->host;
 350         struct queue_limits lim = {
 351                 .features               = features,
 352         };
 353         struct gendisk *disk;
 354
 355         if (mmc_can_erase(card))
 356                 mmc_queue_setup_discard(card, &lim);
 357
 358         lim.max_hw_sectors = min(host->max_blk_count, host->max_req_size / 512);
 359
 360         if (mmc_card_mmc(card) && card->ext_csd.data_sector_size)
 361                 lim.logical_block_size = card->ext_csd.data_sector_size;
 362         else
 363                 lim.logical_block_size = 512;
 364
 365         WARN_ON_ONCE(lim.logical_block_size != 512 &&
 366                      lim.logical_block_size != 4096);
 367
 368         /*
 369          * Setting a virt_boundary implicity sets a max_segment_size, so try
 370          * to set the hardware one here.
 371          */
 372         if (host->can_dma_map_merge) {
 373                 lim.virt_boundary_mask = dma_get_merge_boundary(mmc_dev(host));
 374                 lim.max_segments = MMC_DMA_MAP_MERGE_SEGMENTS;
 375         } else {
 376                 lim.max_segment_size =
 377                         round_down(host->max_seg_size, lim.logical_block_size);
 378                 lim.max_segments = host->max_segs;
 379         }
 380
 381         if (mmc_host_is_spi(host) && host->use_spi_crc)
 382                 lim.features |= BLK_FEAT_STABLE_WRITES;
 383
 384         disk = blk_mq_alloc_disk(&mq->tag_set, &lim, mq);
 385         if (IS_ERR(disk))
 386                 return disk;
 387         mq->queue = disk->queue;
 388
 389         blk_queue_rq_timeout(mq->queue, 60 * HZ);
 390
 391         if (mmc_dev(host)->dma_parms)
 392                 dma_set_max_seg_size(mmc_dev(host), queue_max_segment_size(mq->queue));
 393
 394         INIT_WORK(&mq->recovery_work, mmc_mq_recovery_handler);
 395         INIT_WORK(&mq->complete_work, mmc_blk_mq_complete_work);
 396
 397         mutex_init(&mq->complete_lock);
 398
 399         init_waitqueue_head(&mq->wait);
 400
 401         mmc_crypto_setup_queue(mq->queue, host);
 402         return disk;
 403 }
 404
 405 static inline bool mmc_merge_capable(struct mmc_host *host)
 406 {
 407         return host->caps2 & MMC_CAP2_MERGE_CAPABLE;
 408 }
 409
 410 /* Set queue depth to get a reasonable value for q->nr_requests */
 411 #define MMC_QUEUE_DEPTH 64
 412
 413 /**
 414  * mmc_init_queue - initialise a queue structure.
 415  * @mq: mmc queue
 416  * @card: mmc card to attach this queue
 417  * @features: block layer features (BLK_FEAT_*)
 418  *
 419  * Initialise a MMC card request queue.
 420  */
 421 struct gendisk *mmc_init_queue(struct mmc_queue *mq, struct mmc_card *card,
 422                 unsigned int features)
 423 {
 424         struct mmc_host *host = card->host;
 425         struct gendisk *disk;
 426         int ret;
 427
 428         mq->card = card;
 429
 430         spin_lock_init(&mq->lock);
 431
 432         memset(&mq->tag_set, 0, sizeof(mq->tag_set));
 433         mq->tag_set.ops = &mmc_mq_ops;
 434         /*
 435          * The queue depth for CQE must match the hardware because the request
 436          * tag is used to index the hardware queue.
 437          */
 438         if (host->cqe_enabled && !host->hsq_enabled)
 439                 mq->tag_set.queue_depth =
 440                         min_t(int, card->ext_csd.cmdq_depth, host->cqe_qdepth);
 441         else
 442                 mq->tag_set.queue_depth = MMC_QUEUE_DEPTH;
 443         mq->tag_set.numa_node = NUMA_NO_NODE;
 444         mq->tag_set.flags = BLK_MQ_F_SHOULD_MERGE | BLK_MQ_F_BLOCKING;
 445         mq->tag_set.nr_hw_queues = 1;
 446         mq->tag_set.cmd_size = sizeof(struct mmc_queue_req);
 447         mq->tag_set.driver_data = mq;
 448
 449         /*
 450          * Since blk_mq_alloc_tag_set() calls .init_request() of mmc_mq_ops,
 451          * the host->can_dma_map_merge should be set before to get max_segs
 452          * from mmc_get_max_segments().
 453          */
 454         if (mmc_merge_capable(host) &&
 455             host->max_segs < MMC_DMA_MAP_MERGE_SEGMENTS &&
 456             dma_get_merge_boundary(mmc_dev(host)))
 457                 host->can_dma_map_merge = 1;
 458         else
 459                 host->can_dma_map_merge = 0;
 460
 461         ret = blk_mq_alloc_tag_set(&mq->tag_set);
 462         if (ret)
 463                 return ERR_PTR(ret);
 464
 465
 466         disk = mmc_alloc_disk(mq, card, features);
 467         if (IS_ERR(disk))
 468                 blk_mq_free_tag_set(&mq->tag_set);
 469         return disk;
 470 }
 471
 472 void mmc_queue_suspend(struct mmc_queue *mq)
 473 {
 474         blk_mq_quiesce_queue(mq->queue);
 475
 476         /*
 477          * The host remains claimed while there are outstanding requests, so
 478          * simply claiming and releasing here ensures there are none.
 479          */
 480         mmc_claim_host(mq->card->host);
 481         mmc_release_host(mq->card->host);
 482 }
 483
 484 void mmc_queue_resume(struct mmc_queue *mq)
 485 {
 486         blk_mq_unquiesce_queue(mq->queue);
 487 }
 488
 489 void mmc_cleanup_queue(struct mmc_queue *mq)
 490 {
 491         struct request_queue *q = mq->queue;
 492
 493         /*
 494          * The legacy code handled the possibility of being suspended,
 495          * so do that here too.
 496          */
 497         if (blk_queue_quiesced(q))
 498                 blk_mq_unquiesce_queue(q);
 499
 500         /*
 501          * If the recovery completes the last (and only remaining) request in
 502          * the queue, and the card has been removed, we could end up here with
 503          * the recovery not quite finished yet, so cancel it.
 504          */
 505         cancel_work_sync(&mq->recovery_work);
 506
 507         blk_mq_free_tag_set(&mq->tag_set);
 508
 509         /*
 510          * A request can be completed before the next request, potentially
 511          * leaving a complete_work with nothing to do. Such a work item might
 512          * still be queued at this point. Flush it.
 513          */
 514         flush_work(&mq->complete_work);
 515
 516         mq->card = NULL;
 517 }
 518
 519 /*
 520  * Prepare the sg list(s) to be handed of to the host driver
 521  */
 522 unsigned int mmc_queue_map_sg(struct mmc_queue *mq, struct mmc_queue_req *mqrq)
 523 {
 524         struct request *req = mmc_queue_req_to_req(mqrq);
 525
 526         return blk_rq_map_sg(mq->queue, req, mqrq->sg);
 527 }