mmc/core/queue.c

0001 // SPDX-License-Identifier: GPL-2.0-only
0002 /*
0003  *  Copyright (C) 2003 Russell King, All Rights Reserved.
0004  *  Copyright 2006-2007 Pierre Ossman
0005  */
0006 #include <linux/slab.h>
0007 #include <linux/module.h>
0008 #include <linux/blkdev.h>
0009 #include <linux/freezer.h>
0010 #include <linux/scatterlist.h>
0011 #include <linux/dma-mapping.h>
0012 #include <linux/backing-dev.h>
0013
0014 #include <linux/mmc/card.h>
0015 #include <linux/mmc/host.h>
0016
0017 #include "queue.h"
0018 #include "block.h"
0019 #include "core.h"
0020 #include "card.h"
0021 #include "crypto.h"
0022 #include "host.h"
0023
0024 #define MMC_DMA_MAP_MERGE_SEGMENTS  512
0025
0026 static inline bool mmc_cqe_dcmd_busy(struct mmc_queue *mq)
0027 {
0028     /* Allow only 1 DCMD at a time */
0029     return mq->in_flight[MMC_ISSUE_DCMD];
0030 }
0031
0032 void mmc_cqe_check_busy(struct mmc_queue *mq)
0033 {
0034     if ((mq->cqe_busy & MMC_CQE_DCMD_BUSY) && !mmc_cqe_dcmd_busy(mq))
0035         mq->cqe_busy &= ~MMC_CQE_DCMD_BUSY;
0036 }
0037
0038 static inline bool mmc_cqe_can_dcmd(struct mmc_host *host)
0039 {
0040     return host->caps2 & MMC_CAP2_CQE_DCMD;
0041 }
0042
0043 static enum mmc_issue_type mmc_cqe_issue_type(struct mmc_host *host,
0044                           struct request *req)
0045 {
0046     switch (req_op(req)) {
0047     case REQ_OP_DRV_IN:
0048     case REQ_OP_DRV_OUT:
0049     case REQ_OP_DISCARD:
0050     case REQ_OP_SECURE_ERASE:
0051         return MMC_ISSUE_SYNC;
0052     case REQ_OP_FLUSH:
0053         return mmc_cqe_can_dcmd(host) ? MMC_ISSUE_DCMD : MMC_ISSUE_SYNC;
0054     default:
0055         return MMC_ISSUE_ASYNC;
0056     }
0057 }
0058
0059 enum mmc_issue_type mmc_issue_type(struct mmc_queue *mq, struct request *req)
0060 {
0061     struct mmc_host *host = mq->card->host;
0062
0063     if (host->cqe_enabled && !host->hsq_enabled)
0064         return mmc_cqe_issue_type(host, req);
0065
0066     if (req_op(req) == REQ_OP_READ || req_op(req) == REQ_OP_WRITE)
0067         return MMC_ISSUE_ASYNC;
0068
0069     return MMC_ISSUE_SYNC;
0070 }
0071
0072 static void __mmc_cqe_recovery_notifier(struct mmc_queue *mq)
0073 {
0074     if (!mq->recovery_needed) {
0075         mq->recovery_needed = true;
0076         schedule_work(&mq->recovery_work);
0077     }
0078 }
0079
0080 void mmc_cqe_recovery_notifier(struct mmc_request *mrq)
0081 {
0082     struct mmc_queue_req *mqrq = container_of(mrq, struct mmc_queue_req,
0083                           brq.mrq);
0084     struct request *req = mmc_queue_req_to_req(mqrq);
0085     struct request_queue *q = req->q;
0086     struct mmc_queue *mq = q->queuedata;
0087     unsigned long flags;
0088
0089     spin_lock_irqsave(&mq->lock, flags);
0090     __mmc_cqe_recovery_notifier(mq);
0091     spin_unlock_irqrestore(&mq->lock, flags);
0092 }
0093
0094 static enum blk_eh_timer_return mmc_cqe_timed_out(struct request *req)
0095 {
0096     struct mmc_queue_req *mqrq = req_to_mmc_queue_req(req);
0097     struct mmc_request *mrq = &mqrq->brq.mrq;
0098     struct mmc_queue *mq = req->q->queuedata;
0099     struct mmc_host *host = mq->card->host;
0100     enum mmc_issue_type issue_type = mmc_issue_type(mq, req);
0101     bool recovery_needed = false;
0102
0103     switch (issue_type) {
0104     case MMC_ISSUE_ASYNC:
0105     case MMC_ISSUE_DCMD:
0106         if (host->cqe_ops->cqe_timeout(host, mrq, &recovery_needed)) {
0107             if (recovery_needed)
0108                 mmc_cqe_recovery_notifier(mrq);
0109             return BLK_EH_RESET_TIMER;
0110         }
0111         /* The request has gone already */
0112         return BLK_EH_DONE;
0113     default:
0114         /* Timeout is handled by mmc core */
0115         return BLK_EH_RESET_TIMER;
0116     }
0117 }
0118
0119 static enum blk_eh_timer_return mmc_mq_timed_out(struct request *req)
0120 {
0121     struct request_queue *q = req->q;
0122     struct mmc_queue *mq = q->queuedata;
0123     struct mmc_card *card = mq->card;
0124     struct mmc_host *host = card->host;
0125     unsigned long flags;
0126     bool ignore_tout;
0127
0128     spin_lock_irqsave(&mq->lock, flags);
0129     ignore_tout = mq->recovery_needed || !host->cqe_enabled || host->hsq_enabled;
0130     spin_unlock_irqrestore(&mq->lock, flags);
0131
0132     return ignore_tout ? BLK_EH_RESET_TIMER : mmc_cqe_timed_out(req);
0133 }
0134
0135 static void mmc_mq_recovery_handler(struct work_struct *work)
0136 {
0137     struct mmc_queue *mq = container_of(work, struct mmc_queue,
0138                         recovery_work);
0139     struct request_queue *q = mq->queue;
0140     struct mmc_host *host = mq->card->host;
0141
0142     mmc_get_card(mq->card, &mq->ctx);
0143
0144     mq->in_recovery = true;
0145
0146     if (host->cqe_enabled && !host->hsq_enabled)
0147         mmc_blk_cqe_recovery(mq);
0148     else
0149         mmc_blk_mq_recovery(mq);
0150
0151     mq->in_recovery = false;
0152
0153     spin_lock_irq(&mq->lock);
0154     mq->recovery_needed = false;
0155     spin_unlock_irq(&mq->lock);
0156
0157     if (host->hsq_enabled)
0158         host->cqe_ops->cqe_recovery_finish(host);
0159
0160     mmc_put_card(mq->card, &mq->ctx);
0161
0162     blk_mq_run_hw_queues(q, true);
0163 }
0164
0165 static struct scatterlist *mmc_alloc_sg(unsigned short sg_len, gfp_t gfp)
0166 {
0167     struct scatterlist *sg;
0168
0169     sg = kmalloc_array(sg_len, sizeof(*sg), gfp);
0170     if (sg)
0171         sg_init_table(sg, sg_len);
0172
0173     return sg;
0174 }
0175
0176 static void mmc_queue_setup_discard(struct request_queue *q,
0177                     struct mmc_card *card)
0178 {
0179     unsigned max_discard;
0180
0181     max_discard = mmc_calc_max_discard(card);
0182     if (!max_discard)
0183         return;
0184
0185     blk_queue_max_discard_sectors(q, max_discard);
0186     q->limits.discard_granularity = card->pref_erase << 9;
0187     /* granularity must not be greater than max. discard */
0188     if (card->pref_erase > max_discard)
0189         q->limits.discard_granularity = SECTOR_SIZE;
0190     if (mmc_can_secure_erase_trim(card))
0191         blk_queue_max_secure_erase_sectors(q, max_discard);
0192     if (mmc_can_trim(card) && card->erased_byte == 0)
0193         blk_queue_max_write_zeroes_sectors(q, max_discard);
0194 }
0195
0196 static unsigned short mmc_get_max_segments(struct mmc_host *host)
0197 {
0198     return host->can_dma_map_merge ? MMC_DMA_MAP_MERGE_SEGMENTS :
0199                      host->max_segs;
0200 }
0201
0202 static int mmc_mq_init_request(struct blk_mq_tag_set *set, struct request *req,
0203                    unsigned int hctx_idx, unsigned int numa_node)
0204 {
0205     struct mmc_queue_req *mq_rq = req_to_mmc_queue_req(req);
0206     struct mmc_queue *mq = set->driver_data;
0207     struct mmc_card *card = mq->card;
0208     struct mmc_host *host = card->host;
0209
0210     mq_rq->sg = mmc_alloc_sg(mmc_get_max_segments(host), GFP_KERNEL);
0211     if (!mq_rq->sg)
0212         return -ENOMEM;
0213
0214     return 0;
0215 }
0216
0217 static void mmc_mq_exit_request(struct blk_mq_tag_set *set, struct request *req,
0218                 unsigned int hctx_idx)
0219 {
0220     struct mmc_queue_req *mq_rq = req_to_mmc_queue_req(req);
0221
0222     kfree(mq_rq->sg);
0223     mq_rq->sg = NULL;
0224 }
0225
0226 static blk_status_t mmc_mq_queue_rq(struct blk_mq_hw_ctx *hctx,
0227                     const struct blk_mq_queue_data *bd)
0228 {
0229     struct request *req = bd->rq;
0230     struct request_queue *q = req->q;
0231     struct mmc_queue *mq = q->queuedata;
0232     struct mmc_card *card = mq->card;
0233     struct mmc_host *host = card->host;
0234     enum mmc_issue_type issue_type;
0235     enum mmc_issued issued;
0236     bool get_card, cqe_retune_ok;
0237     blk_status_t ret;
0238
0239     if (mmc_card_removed(mq->card)) {
0240         req->rq_flags |= RQF_QUIET;
0241         return BLK_STS_IOERR;
0242     }
0243
0244     issue_type = mmc_issue_type(mq, req);
0245
0246     spin_lock_irq(&mq->lock);
0247
0248     if (mq->recovery_needed || mq->busy) {
0249         spin_unlock_irq(&mq->lock);
0250         return BLK_STS_RESOURCE;
0251     }
0252
0253     switch (issue_type) {
0254     case MMC_ISSUE_DCMD:
0255         if (mmc_cqe_dcmd_busy(mq)) {
0256             mq->cqe_busy |= MMC_CQE_DCMD_BUSY;
0257             spin_unlock_irq(&mq->lock);
0258             return BLK_STS_RESOURCE;
0259         }
0260         break;
0261     case MMC_ISSUE_ASYNC:
0262         /*
0263          * For MMC host software queue, we only allow 2 requests in
0264          * flight to avoid a long latency.
0265          */
0266         if (host->hsq_enabled && mq->in_flight[issue_type] > 2) {
0267             spin_unlock_irq(&mq->lock);
0268             return BLK_STS_RESOURCE;
0269         }
0270         break;
0271     default:
0272         /*
0273          * Timeouts are handled by mmc core, and we don't have a host
0274          * API to abort requests, so we can't handle the timeout anyway.
0275          * However, when the timeout happens, blk_mq_complete_request()
0276          * no longer works (to stop the request disappearing under us).
0277          * To avoid racing with that, set a large timeout.
0278          */
0279         req->timeout = 600 * HZ;
0280         break;
0281     }
0282
0283     /* Parallel dispatch of requests is not supported at the moment */
0284     mq->busy = true;
0285
0286     mq->in_flight[issue_type] += 1;
0287     get_card = (mmc_tot_in_flight(mq) == 1);
0288     cqe_retune_ok = (mmc_cqe_qcnt(mq) == 1);
0289
0290     spin_unlock_irq(&mq->lock);
0291
0292     if (!(req->rq_flags & RQF_DONTPREP)) {
0293         req_to_mmc_queue_req(req)->retries = 0;
0294         req->rq_flags |= RQF_DONTPREP;
0295     }
0296
0297     if (get_card)
0298         mmc_get_card(card, &mq->ctx);
0299
0300     if (host->cqe_enabled) {
0301         host->retune_now = host->need_retune && cqe_retune_ok &&
0302                    !host->hold_retune;
0303     }
0304
0305     blk_mq_start_request(req);
0306
0307     issued = mmc_blk_mq_issue_rq(mq, req);
0308
0309     switch (issued) {
0310     case MMC_REQ_BUSY:
0311         ret = BLK_STS_RESOURCE;
0312         break;
0313     case MMC_REQ_FAILED_TO_START:
0314         ret = BLK_STS_IOERR;
0315         break;
0316     default:
0317         ret = BLK_STS_OK;
0318         break;
0319     }
0320
0321     if (issued != MMC_REQ_STARTED) {
0322         bool put_card = false;
0323
0324         spin_lock_irq(&mq->lock);
0325         mq->in_flight[issue_type] -= 1;
0326         if (mmc_tot_in_flight(mq) == 0)
0327             put_card = true;
0328         mq->busy = false;
0329         spin_unlock_irq(&mq->lock);
0330         if (put_card)
0331             mmc_put_card(card, &mq->ctx);
0332     } else {
0333         WRITE_ONCE(mq->busy, false);
0334     }
0335
0336     return ret;
0337 }
0338
0339 static const struct blk_mq_ops mmc_mq_ops = {
0340     .queue_rq   = mmc_mq_queue_rq,
0341     .init_request   = mmc_mq_init_request,
0342     .exit_request   = mmc_mq_exit_request,
0343     .complete   = mmc_blk_mq_complete,
0344     .timeout    = mmc_mq_timed_out,
0345 };
0346
0347 static void mmc_setup_queue(struct mmc_queue *mq, struct mmc_card *card)
0348 {
0349     struct mmc_host *host = card->host;
0350     unsigned block_size = 512;
0351
0352     blk_queue_flag_set(QUEUE_FLAG_NONROT, mq->queue);
0353     blk_queue_flag_clear(QUEUE_FLAG_ADD_RANDOM, mq->queue);
0354     if (mmc_can_erase(card))
0355         mmc_queue_setup_discard(mq->queue, card);
0356
0357     if (!mmc_dev(host)->dma_mask || !*mmc_dev(host)->dma_mask)
0358         blk_queue_bounce_limit(mq->queue, BLK_BOUNCE_HIGH);
0359     blk_queue_max_hw_sectors(mq->queue,
0360         min(host->max_blk_count, host->max_req_size / 512));
0361     if (host->can_dma_map_merge)
0362         WARN(!blk_queue_can_use_dma_map_merging(mq->queue,
0363                             mmc_dev(host)),
0364              "merging was advertised but not possible");
0365     blk_queue_max_segments(mq->queue, mmc_get_max_segments(host));
0366
0367     if (mmc_card_mmc(card) && card->ext_csd.data_sector_size) {
0368         block_size = card->ext_csd.data_sector_size;
0369         WARN_ON(block_size != 512 && block_size != 4096);
0370     }
0371
0372     blk_queue_logical_block_size(mq->queue, block_size);
0373     /*
0374      * After blk_queue_can_use_dma_map_merging() was called with succeed,
0375      * since it calls blk_queue_virt_boundary(), the mmc should not call
0376      * both blk_queue_max_segment_size().
0377      */
0378     if (!host->can_dma_map_merge)
0379         blk_queue_max_segment_size(mq->queue,
0380             round_down(host->max_seg_size, block_size));
0381
0382     dma_set_max_seg_size(mmc_dev(host), queue_max_segment_size(mq->queue));
0383
0384     INIT_WORK(&mq->recovery_work, mmc_mq_recovery_handler);
0385     INIT_WORK(&mq->complete_work, mmc_blk_mq_complete_work);
0386
0387     mutex_init(&mq->complete_lock);
0388
0389     init_waitqueue_head(&mq->wait);
0390
0391     mmc_crypto_setup_queue(mq->queue, host);
0392 }
0393
0394 static inline bool mmc_merge_capable(struct mmc_host *host)
0395 {
0396     return host->caps2 & MMC_CAP2_MERGE_CAPABLE;
0397 }
0398
0399 /* Set queue depth to get a reasonable value for q->nr_requests */
0400 #define MMC_QUEUE_DEPTH 64
0401
0402 /**
0403  * mmc_init_queue - initialise a queue structure.
0404  * @mq: mmc queue
0405  * @card: mmc card to attach this queue
0406  *
0407  * Initialise a MMC card request queue.
0408  */
0409 struct gendisk *mmc_init_queue(struct mmc_queue *mq, struct mmc_card *card)
0410 {
0411     struct mmc_host *host = card->host;
0412     struct gendisk *disk;
0413     int ret;
0414
0415     mq->card = card;
0416
0417     spin_lock_init(&mq->lock);
0418
0419     memset(&mq->tag_set, 0, sizeof(mq->tag_set));
0420     mq->tag_set.ops = &mmc_mq_ops;
0421     /*
0422      * The queue depth for CQE must match the hardware because the request
0423      * tag is used to index the hardware queue.
0424      */
0425     if (host->cqe_enabled && !host->hsq_enabled)
0426         mq->tag_set.queue_depth =
0427             min_t(int, card->ext_csd.cmdq_depth, host->cqe_qdepth);
0428     else
0429         mq->tag_set.queue_depth = MMC_QUEUE_DEPTH;
0430     mq->tag_set.numa_node = NUMA_NO_NODE;
0431     mq->tag_set.flags = BLK_MQ_F_SHOULD_MERGE | BLK_MQ_F_BLOCKING;
0432     mq->tag_set.nr_hw_queues = 1;
0433     mq->tag_set.cmd_size = sizeof(struct mmc_queue_req);
0434     mq->tag_set.driver_data = mq;
0435
0436     /*
0437      * Since blk_mq_alloc_tag_set() calls .init_request() of mmc_mq_ops,
0438      * the host->can_dma_map_merge should be set before to get max_segs
0439      * from mmc_get_max_segments().
0440      */
0441     if (mmc_merge_capable(host) &&
0442         host->max_segs < MMC_DMA_MAP_MERGE_SEGMENTS &&
0443         dma_get_merge_boundary(mmc_dev(host)))
0444         host->can_dma_map_merge = 1;
0445     else
0446         host->can_dma_map_merge = 0;
0447
0448     ret = blk_mq_alloc_tag_set(&mq->tag_set);
0449     if (ret)
0450         return ERR_PTR(ret);
0451
0452
0453     disk = blk_mq_alloc_disk(&mq->tag_set, mq);
0454     if (IS_ERR(disk)) {
0455         blk_mq_free_tag_set(&mq->tag_set);
0456         return disk;
0457     }
0458     mq->queue = disk->queue;
0459
0460     if (mmc_host_is_spi(host) && host->use_spi_crc)
0461         blk_queue_flag_set(QUEUE_FLAG_STABLE_WRITES, mq->queue);
0462     blk_queue_rq_timeout(mq->queue, 60 * HZ);
0463
0464     mmc_setup_queue(mq, card);
0465     return disk;
0466 }
0467
0468 void mmc_queue_suspend(struct mmc_queue *mq)
0469 {
0470     blk_mq_quiesce_queue(mq->queue);
0471
0472     /*
0473      * The host remains claimed while there are outstanding requests, so
0474      * simply claiming and releasing here ensures there are none.
0475      */
0476     mmc_claim_host(mq->card->host);
0477     mmc_release_host(mq->card->host);
0478 }
0479
0480 void mmc_queue_resume(struct mmc_queue *mq)
0481 {
0482     blk_mq_unquiesce_queue(mq->queue);
0483 }
0484
0485 void mmc_cleanup_queue(struct mmc_queue *mq)
0486 {
0487     struct request_queue *q = mq->queue;
0488
0489     /*
0490      * The legacy code handled the possibility of being suspended,
0491      * so do that here too.
0492      */
0493     if (blk_queue_quiesced(q))
0494         blk_mq_unquiesce_queue(q);
0495
0496     blk_mq_free_tag_set(&mq->tag_set);
0497
0498     /*
0499      * A request can be completed before the next request, potentially
0500      * leaving a complete_work with nothing to do. Such a work item might
0501      * still be queued at this point. Flush it.
0502      */
0503     flush_work(&mq->complete_work);
0504
0505     mq->card = NULL;
0506 }
0507
0508 /*
0509  * Prepare the sg list(s) to be handed of to the host driver
0510  */
0511 unsigned int mmc_queue_map_sg(struct mmc_queue *mq, struct mmc_queue_req *mqrq)
0512 {
0513     struct request *req = mmc_queue_req_to_req(mqrq);
0514
0515     return blk_rq_map_sg(mq->queue, req, mqrq->sg);
0516 }