linux/drivers/mmc/core/queue.c
<<
>>
Prefs
   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/kthread.h>
  11#include <linux/scatterlist.h>
  12#include <linux/dma-mapping.h>
  13#include <linux/backing-dev.h>
  14
  15#include <linux/mmc/card.h>
  16#include <linux/mmc/host.h>
  17
  18#include "queue.h"
  19#include "block.h"
  20#include "core.h"
  21#include "card.h"
  22#include "host.h"
  23
  24#define MMC_DMA_MAP_MERGE_SEGMENTS      512
  25
  26static 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
  32void 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        mq->cqe_busy &= ~MMC_CQE_QUEUE_FULL;
  38}
  39
  40static inline bool mmc_cqe_can_dcmd(struct mmc_host *host)
  41{
  42        return host->caps2 & MMC_CAP2_CQE_DCMD;
  43}
  44
  45static enum mmc_issue_type mmc_cqe_issue_type(struct mmc_host *host,
  46                                              struct request *req)
  47{
  48        switch (req_op(req)) {
  49        case REQ_OP_DRV_IN:
  50        case REQ_OP_DRV_OUT:
  51        case REQ_OP_DISCARD:
  52        case REQ_OP_SECURE_ERASE:
  53                return MMC_ISSUE_SYNC;
  54        case REQ_OP_FLUSH:
  55                return mmc_cqe_can_dcmd(host) ? MMC_ISSUE_DCMD : MMC_ISSUE_SYNC;
  56        default:
  57                return MMC_ISSUE_ASYNC;
  58        }
  59}
  60
  61enum mmc_issue_type mmc_issue_type(struct mmc_queue *mq, struct request *req)
  62{
  63        struct mmc_host *host = mq->card->host;
  64
  65        if (mq->use_cqe && !host->hsq_enabled)
  66                return mmc_cqe_issue_type(host, req);
  67
  68        if (req_op(req) == REQ_OP_READ || req_op(req) == REQ_OP_WRITE)
  69                return MMC_ISSUE_ASYNC;
  70
  71        return MMC_ISSUE_SYNC;
  72}
  73
  74static void __mmc_cqe_recovery_notifier(struct mmc_queue *mq)
  75{
  76        if (!mq->recovery_needed) {
  77                mq->recovery_needed = true;
  78                schedule_work(&mq->recovery_work);
  79        }
  80}
  81
  82void mmc_cqe_recovery_notifier(struct mmc_request *mrq)
  83{
  84        struct mmc_queue_req *mqrq = container_of(mrq, struct mmc_queue_req,
  85                                                  brq.mrq);
  86        struct request *req = mmc_queue_req_to_req(mqrq);
  87        struct request_queue *q = req->q;
  88        struct mmc_queue *mq = q->queuedata;
  89        unsigned long flags;
  90
  91        spin_lock_irqsave(&mq->lock, flags);
  92        __mmc_cqe_recovery_notifier(mq);
  93        spin_unlock_irqrestore(&mq->lock, flags);
  94}
  95
  96static enum blk_eh_timer_return mmc_cqe_timed_out(struct request *req)
  97{
  98        struct mmc_queue_req *mqrq = req_to_mmc_queue_req(req);
  99        struct mmc_request *mrq = &mqrq->brq.mrq;
 100        struct mmc_queue *mq = req->q->queuedata;
 101        struct mmc_host *host = mq->card->host;
 102        enum mmc_issue_type issue_type = mmc_issue_type(mq, req);
 103        bool recovery_needed = false;
 104
 105        switch (issue_type) {
 106        case MMC_ISSUE_ASYNC:
 107        case MMC_ISSUE_DCMD:
 108                if (host->cqe_ops->cqe_timeout(host, mrq, &recovery_needed)) {
 109                        if (recovery_needed)
 110                                mmc_cqe_recovery_notifier(mrq);
 111                        return BLK_EH_RESET_TIMER;
 112                }
 113                /* The request has gone already */
 114                return BLK_EH_DONE;
 115        default:
 116                /* Timeout is handled by mmc core */
 117                return BLK_EH_RESET_TIMER;
 118        }
 119}
 120
 121static enum blk_eh_timer_return mmc_mq_timed_out(struct request *req,
 122                                                 bool reserved)
 123{
 124        struct request_queue *q = req->q;
 125        struct mmc_queue *mq = q->queuedata;
 126        struct mmc_card *card = mq->card;
 127        struct mmc_host *host = card->host;
 128        unsigned long flags;
 129        bool ignore_tout;
 130
 131        spin_lock_irqsave(&mq->lock, flags);
 132        ignore_tout = mq->recovery_needed || !mq->use_cqe || host->hsq_enabled;
 133        spin_unlock_irqrestore(&mq->lock, flags);
 134
 135        return ignore_tout ? BLK_EH_RESET_TIMER : mmc_cqe_timed_out(req);
 136}
 137
 138static void mmc_mq_recovery_handler(struct work_struct *work)
 139{
 140        struct mmc_queue *mq = container_of(work, struct mmc_queue,
 141                                            recovery_work);
 142        struct request_queue *q = mq->queue;
 143        struct mmc_host *host = mq->card->host;
 144
 145        mmc_get_card(mq->card, &mq->ctx);
 146
 147        mq->in_recovery = true;
 148
 149        if (mq->use_cqe && !host->hsq_enabled)
 150                mmc_blk_cqe_recovery(mq);
 151        else
 152                mmc_blk_mq_recovery(mq);
 153
 154        mq->in_recovery = false;
 155
 156        spin_lock_irq(&mq->lock);
 157        mq->recovery_needed = false;
 158        spin_unlock_irq(&mq->lock);
 159
 160        if (host->hsq_enabled)
 161                host->cqe_ops->cqe_recovery_finish(host);
 162
 163        mmc_put_card(mq->card, &mq->ctx);
 164
 165        blk_mq_run_hw_queues(q, true);
 166}
 167
 168static struct scatterlist *mmc_alloc_sg(int sg_len, gfp_t gfp)
 169{
 170        struct scatterlist *sg;
 171
 172        sg = kmalloc_array(sg_len, sizeof(*sg), gfp);
 173        if (sg)
 174                sg_init_table(sg, sg_len);
 175
 176        return sg;
 177}
 178
 179static void mmc_queue_setup_discard(struct request_queue *q,
 180                                    struct mmc_card *card)
 181{
 182        unsigned max_discard;
 183
 184        max_discard = mmc_calc_max_discard(card);
 185        if (!max_discard)
 186                return;
 187
 188        blk_queue_flag_set(QUEUE_FLAG_DISCARD, q);
 189        blk_queue_max_discard_sectors(q, max_discard);
 190        q->limits.discard_granularity = card->pref_erase << 9;
 191        /* granularity must not be greater than max. discard */
 192        if (card->pref_erase > max_discard)
 193                q->limits.discard_granularity = 0;
 194        if (mmc_can_secure_erase_trim(card))
 195                blk_queue_flag_set(QUEUE_FLAG_SECERASE, q);
 196}
 197
 198static unsigned int mmc_get_max_segments(struct mmc_host *host)
 199{
 200        return host->can_dma_map_merge ? MMC_DMA_MAP_MERGE_SEGMENTS :
 201                                         host->max_segs;
 202}
 203
 204/**
 205 * mmc_init_request() - initialize the MMC-specific per-request data
 206 * @q: the request queue
 207 * @req: the request
 208 * @gfp: memory allocation policy
 209 */
 210static int __mmc_init_request(struct mmc_queue *mq, struct request *req,
 211                              gfp_t gfp)
 212{
 213        struct mmc_queue_req *mq_rq = req_to_mmc_queue_req(req);
 214        struct mmc_card *card = mq->card;
 215        struct mmc_host *host = card->host;
 216
 217        mq_rq->sg = mmc_alloc_sg(mmc_get_max_segments(host), gfp);
 218        if (!mq_rq->sg)
 219                return -ENOMEM;
 220
 221        return 0;
 222}
 223
 224static void mmc_exit_request(struct request_queue *q, struct request *req)
 225{
 226        struct mmc_queue_req *mq_rq = req_to_mmc_queue_req(req);
 227
 228        kfree(mq_rq->sg);
 229        mq_rq->sg = NULL;
 230}
 231
 232static int mmc_mq_init_request(struct blk_mq_tag_set *set, struct request *req,
 233                               unsigned int hctx_idx, unsigned int numa_node)
 234{
 235        return __mmc_init_request(set->driver_data, req, GFP_KERNEL);
 236}
 237
 238static void mmc_mq_exit_request(struct blk_mq_tag_set *set, struct request *req,
 239                                unsigned int hctx_idx)
 240{
 241        struct mmc_queue *mq = set->driver_data;
 242
 243        mmc_exit_request(mq->queue, req);
 244}
 245
 246static blk_status_t mmc_mq_queue_rq(struct blk_mq_hw_ctx *hctx,
 247                                    const struct blk_mq_queue_data *bd)
 248{
 249        struct request *req = bd->rq;
 250        struct request_queue *q = req->q;
 251        struct mmc_queue *mq = q->queuedata;
 252        struct mmc_card *card = mq->card;
 253        struct mmc_host *host = card->host;
 254        enum mmc_issue_type issue_type;
 255        enum mmc_issued issued;
 256        bool get_card, cqe_retune_ok;
 257        int ret;
 258
 259        if (mmc_card_removed(mq->card)) {
 260                req->rq_flags |= RQF_QUIET;
 261                return BLK_STS_IOERR;
 262        }
 263
 264        issue_type = mmc_issue_type(mq, req);
 265
 266        spin_lock_irq(&mq->lock);
 267
 268        if (mq->recovery_needed || mq->busy) {
 269                spin_unlock_irq(&mq->lock);
 270                return BLK_STS_RESOURCE;
 271        }
 272
 273        switch (issue_type) {
 274        case MMC_ISSUE_DCMD:
 275                if (mmc_cqe_dcmd_busy(mq)) {
 276                        mq->cqe_busy |= MMC_CQE_DCMD_BUSY;
 277                        spin_unlock_irq(&mq->lock);
 278                        return BLK_STS_RESOURCE;
 279                }
 280                break;
 281        case MMC_ISSUE_ASYNC:
 282                /*
 283                 * For MMC host software queue, we only allow 2 requests in
 284                 * flight to avoid a long latency.
 285                 */
 286                if (host->hsq_enabled && mq->in_flight[issue_type] > 2) {
 287                        spin_unlock_irq(&mq->lock);
 288                        return BLK_STS_RESOURCE;
 289                }
 290                break;
 291        default:
 292                /*
 293                 * Timeouts are handled by mmc core, and we don't have a host
 294                 * API to abort requests, so we can't handle the timeout anyway.
 295                 * However, when the timeout happens, blk_mq_complete_request()
 296                 * no longer works (to stop the request disappearing under us).
 297                 * To avoid racing with that, set a large timeout.
 298                 */
 299                req->timeout = 600 * HZ;
 300                break;
 301        }
 302
 303        /* Parallel dispatch of requests is not supported at the moment */
 304        mq->busy = true;
 305
 306        mq->in_flight[issue_type] += 1;
 307        get_card = (mmc_tot_in_flight(mq) == 1);
 308        cqe_retune_ok = (mmc_cqe_qcnt(mq) == 1);
 309
 310        spin_unlock_irq(&mq->lock);
 311
 312        if (!(req->rq_flags & RQF_DONTPREP)) {
 313                req_to_mmc_queue_req(req)->retries = 0;
 314                req->rq_flags |= RQF_DONTPREP;
 315        }
 316
 317        if (get_card)
 318                mmc_get_card(card, &mq->ctx);
 319
 320        if (mq->use_cqe) {
 321                host->retune_now = host->need_retune && cqe_retune_ok &&
 322                                   !host->hold_retune;
 323        }
 324
 325        blk_mq_start_request(req);
 326
 327        issued = mmc_blk_mq_issue_rq(mq, req);
 328
 329        switch (issued) {
 330        case MMC_REQ_BUSY:
 331                ret = BLK_STS_RESOURCE;
 332                break;
 333        case MMC_REQ_FAILED_TO_START:
 334                ret = BLK_STS_IOERR;
 335                break;
 336        default:
 337                ret = BLK_STS_OK;
 338                break;
 339        }
 340
 341        if (issued != MMC_REQ_STARTED) {
 342                bool put_card = false;
 343
 344                spin_lock_irq(&mq->lock);
 345                mq->in_flight[issue_type] -= 1;
 346                if (mmc_tot_in_flight(mq) == 0)
 347                        put_card = true;
 348                mq->busy = false;
 349                spin_unlock_irq(&mq->lock);
 350                if (put_card)
 351                        mmc_put_card(card, &mq->ctx);
 352        } else {
 353                WRITE_ONCE(mq->busy, false);
 354        }
 355
 356        return ret;
 357}
 358
 359static const struct blk_mq_ops mmc_mq_ops = {
 360        .queue_rq       = mmc_mq_queue_rq,
 361        .init_request   = mmc_mq_init_request,
 362        .exit_request   = mmc_mq_exit_request,
 363        .complete       = mmc_blk_mq_complete,
 364        .timeout        = mmc_mq_timed_out,
 365};
 366
 367static void mmc_setup_queue(struct mmc_queue *mq, struct mmc_card *card)
 368{
 369        struct mmc_host *host = card->host;
 370        unsigned block_size = 512;
 371
 372        blk_queue_flag_set(QUEUE_FLAG_NONROT, mq->queue);
 373        blk_queue_flag_clear(QUEUE_FLAG_ADD_RANDOM, mq->queue);
 374        if (mmc_can_erase(card))
 375                mmc_queue_setup_discard(mq->queue, card);
 376
 377        if (!mmc_dev(host)->dma_mask || !*mmc_dev(host)->dma_mask)
 378                blk_queue_bounce_limit(mq->queue, BLK_BOUNCE_HIGH);
 379        blk_queue_max_hw_sectors(mq->queue,
 380                min(host->max_blk_count, host->max_req_size / 512));
 381        if (host->can_dma_map_merge)
 382                WARN(!blk_queue_can_use_dma_map_merging(mq->queue,
 383                                                        mmc_dev(host)),
 384                     "merging was advertised but not possible");
 385        blk_queue_max_segments(mq->queue, mmc_get_max_segments(host));
 386
 387        if (mmc_card_mmc(card))
 388                block_size = card->ext_csd.data_sector_size;
 389
 390        blk_queue_logical_block_size(mq->queue, block_size);
 391        /*
 392         * After blk_queue_can_use_dma_map_merging() was called with succeed,
 393         * since it calls blk_queue_virt_boundary(), the mmc should not call
 394         * both blk_queue_max_segment_size().
 395         */
 396        if (!host->can_dma_map_merge)
 397                blk_queue_max_segment_size(mq->queue,
 398                        round_down(host->max_seg_size, block_size));
 399
 400        dma_set_max_seg_size(mmc_dev(host), queue_max_segment_size(mq->queue));
 401
 402        INIT_WORK(&mq->recovery_work, mmc_mq_recovery_handler);
 403        INIT_WORK(&mq->complete_work, mmc_blk_mq_complete_work);
 404
 405        mutex_init(&mq->complete_lock);
 406
 407        init_waitqueue_head(&mq->wait);
 408}
 409
 410static inline bool mmc_merge_capable(struct mmc_host *host)
 411{
 412        return host->caps2 & MMC_CAP2_MERGE_CAPABLE;
 413}
 414
 415/* Set queue depth to get a reasonable value for q->nr_requests */
 416#define MMC_QUEUE_DEPTH 64
 417
 418/**
 419 * mmc_init_queue - initialise a queue structure.
 420 * @mq: mmc queue
 421 * @card: mmc card to attach this queue
 422 *
 423 * Initialise a MMC card request queue.
 424 */
 425int mmc_init_queue(struct mmc_queue *mq, struct mmc_card *card)
 426{
 427        struct mmc_host *host = card->host;
 428        int ret;
 429
 430        mq->card = card;
 431        mq->use_cqe = host->cqe_enabled;
 432        
 433        spin_lock_init(&mq->lock);
 434
 435        memset(&mq->tag_set, 0, sizeof(mq->tag_set));
 436        mq->tag_set.ops = &mmc_mq_ops;
 437        /*
 438         * The queue depth for CQE must match the hardware because the request
 439         * tag is used to index the hardware queue.
 440         */
 441        if (mq->use_cqe && !host->hsq_enabled)
 442                mq->tag_set.queue_depth =
 443                        min_t(int, card->ext_csd.cmdq_depth, host->cqe_qdepth);
 444        else
 445                mq->tag_set.queue_depth = MMC_QUEUE_DEPTH;
 446        mq->tag_set.numa_node = NUMA_NO_NODE;
 447        mq->tag_set.flags = BLK_MQ_F_SHOULD_MERGE | BLK_MQ_F_BLOCKING;
 448        mq->tag_set.nr_hw_queues = 1;
 449        mq->tag_set.cmd_size = sizeof(struct mmc_queue_req);
 450        mq->tag_set.driver_data = mq;
 451
 452        /*
 453         * Since blk_mq_alloc_tag_set() calls .init_request() of mmc_mq_ops,
 454         * the host->can_dma_map_merge should be set before to get max_segs
 455         * from mmc_get_max_segments().
 456         */
 457        if (mmc_merge_capable(host) &&
 458            host->max_segs < MMC_DMA_MAP_MERGE_SEGMENTS &&
 459            dma_get_merge_boundary(mmc_dev(host)))
 460                host->can_dma_map_merge = 1;
 461        else
 462                host->can_dma_map_merge = 0;
 463
 464        ret = blk_mq_alloc_tag_set(&mq->tag_set);
 465        if (ret)
 466                return ret;
 467
 468        mq->queue = blk_mq_init_queue(&mq->tag_set);
 469        if (IS_ERR(mq->queue)) {
 470                ret = PTR_ERR(mq->queue);
 471                goto free_tag_set;
 472        }
 473
 474        if (mmc_host_is_spi(host) && host->use_spi_crc)
 475                mq->queue->backing_dev_info->capabilities |=
 476                        BDI_CAP_STABLE_WRITES;
 477
 478        mq->queue->queuedata = mq;
 479        blk_queue_rq_timeout(mq->queue, 60 * HZ);
 480
 481        mmc_setup_queue(mq, card);
 482        return 0;
 483
 484free_tag_set:
 485        blk_mq_free_tag_set(&mq->tag_set);
 486        return ret;
 487}
 488
 489void mmc_queue_suspend(struct mmc_queue *mq)
 490{
 491        blk_mq_quiesce_queue(mq->queue);
 492
 493        /*
 494         * The host remains claimed while there are outstanding requests, so
 495         * simply claiming and releasing here ensures there are none.
 496         */
 497        mmc_claim_host(mq->card->host);
 498        mmc_release_host(mq->card->host);
 499}
 500
 501void mmc_queue_resume(struct mmc_queue *mq)
 502{
 503        blk_mq_unquiesce_queue(mq->queue);
 504}
 505
 506void mmc_cleanup_queue(struct mmc_queue *mq)
 507{
 508        struct request_queue *q = mq->queue;
 509
 510        /*
 511         * The legacy code handled the possibility of being suspended,
 512         * so do that here too.
 513         */
 514        if (blk_queue_quiesced(q))
 515                blk_mq_unquiesce_queue(q);
 516
 517        blk_cleanup_queue(q);
 518        blk_mq_free_tag_set(&mq->tag_set);
 519
 520        /*
 521         * A request can be completed before the next request, potentially
 522         * leaving a complete_work with nothing to do. Such a work item might
 523         * still be queued at this point. Flush it.
 524         */
 525        flush_work(&mq->complete_work);
 526
 527        mq->card = NULL;
 528}
 529
 530/*
 531 * Prepare the sg list(s) to be handed of to the host driver
 532 */
 533unsigned int mmc_queue_map_sg(struct mmc_queue *mq, struct mmc_queue_req *mqrq)
 534{
 535        struct request *req = mmc_queue_req_to_req(mqrq);
 536
 537        return blk_rq_map_sg(mq->queue, req, mqrq->sg);
 538}
 539