projects/linux·blob master · refs

538 lines · 13.2 KiB

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
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
38static inline bool mmc_cqe_can_dcmd(struct mmc_host *host)
39{
40 return host->caps2 & MMC_CAP2_CQE_DCMD;
41}
42
43static 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
60enum 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
73static 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
81void 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
95static 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
120static 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
136static 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
166static struct scatterlist *mmc_alloc_sg(unsigned short sg_len, gfp_t gfp)
167{
168 struct scatterlist *sg;
169
170 sg = kmalloc_objs(*sg, sg_len, gfp);
171 if (sg)
172 sg_init_table(sg, sg_len);
173
174 return sg;
175}
176
177static 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_card_can_secure_erase_trim(card)) {
188 if (mmc_card_fixed_secure_erase_trim_time(card))
189 lim->max_secure_erase_sectors = UINT_MAX >> card->erase_shift;
190 else
191 lim->max_secure_erase_sectors = max_discard;
192 }
193
194 if (mmc_card_can_trim(card) && card->erased_byte == 0)
195 lim->max_write_zeroes_sectors = max_discard;
196
197 /* granularity must not be greater than max. discard */
198 if (card->pref_erase > max_discard)
199 lim->discard_granularity = SECTOR_SIZE;
200 else
201 lim->discard_granularity = card->pref_erase << 9;
202}
203
204static unsigned short mmc_get_max_segments(struct mmc_host *host)
205{
206 return host->can_dma_map_merge ? MMC_DMA_MAP_MERGE_SEGMENTS :
207 host->max_segs;
208}
209
210static int mmc_mq_init_request(struct blk_mq_tag_set *set, struct request *req,
211 unsigned int hctx_idx, int numa_node)
212{
213 struct mmc_queue_req *mq_rq = req_to_mmc_queue_req(req);
214 struct mmc_queue *mq = set->driver_data;
215 struct mmc_card *card = mq->card;
216 struct mmc_host *host = card->host;
217 u16 sg_len = mmc_get_max_segments(host);
218
219 if (!sg_len) {
220 dev_err(mmc_dev(host), "Wrong max_segs assigned\n");
221 return -EINVAL;
222 }
223
224 mq_rq->sg = mmc_alloc_sg(sg_len, GFP_KERNEL);
225 if (!mq_rq->sg)
226 return -ENOMEM;
227
228 return 0;
229}
230
231static void mmc_mq_exit_request(struct blk_mq_tag_set *set, struct request *req,
232 unsigned int hctx_idx)
233{
234 struct mmc_queue_req *mq_rq = req_to_mmc_queue_req(req);
235
236 kfree(mq_rq->sg);
237 mq_rq->sg = NULL;
238}
239
240static blk_status_t mmc_mq_queue_rq(struct blk_mq_hw_ctx *hctx,
241 const struct blk_mq_queue_data *bd)
242{
243 struct request *req = bd->rq;
244 struct request_queue *q = req->q;
245 struct mmc_queue *mq = q->queuedata;
246 struct mmc_card *card = mq->card;
247 struct mmc_host *host = card->host;
248 enum mmc_issue_type issue_type;
249 enum mmc_issued issued;
250 bool get_card, cqe_retune_ok;
251 blk_status_t ret;
252
253 if (mmc_card_removed(mq->card)) {
254 req->rq_flags |= RQF_QUIET;
255 return BLK_STS_IOERR;
256 }
257
258 issue_type = mmc_issue_type(mq, req);
259
260 spin_lock_irq(&mq->lock);
261
262 if (mq->recovery_needed || mq->busy) {
263 spin_unlock_irq(&mq->lock);
264 return BLK_STS_RESOURCE;
265 }
266
267 switch (issue_type) {
268 case MMC_ISSUE_DCMD:
269 if (mmc_cqe_dcmd_busy(mq)) {
270 mq->cqe_busy |= MMC_CQE_DCMD_BUSY;
271 spin_unlock_irq(&mq->lock);
272 return BLK_STS_RESOURCE;
273 }
274 break;
275 case MMC_ISSUE_ASYNC:
276 if (host->hsq_enabled && mq->in_flight[issue_type] > host->hsq_depth) {
277 spin_unlock_irq(&mq->lock);
278 return BLK_STS_RESOURCE;
279 }
280 break;
281 default:
282 /*
283 * Timeouts are handled by mmc core, and we don't have a host
284 * API to abort requests, so we can't handle the timeout anyway.
285 * However, when the timeout happens, blk_mq_complete_request()
286 * no longer works (to stop the request disappearing under us).
287 * To avoid racing with that, set a large timeout.
288 */
289 req->timeout = 600 * HZ;
290 break;
291 }
292
293 /* Parallel dispatch of requests is not supported at the moment */
294 mq->busy = true;
295
296 mq->in_flight[issue_type] += 1;
297 get_card = (mmc_tot_in_flight(mq) == 1);
298 cqe_retune_ok = (mmc_cqe_qcnt(mq) == 1);
299
300 spin_unlock_irq(&mq->lock);
301
302 if (!(req->rq_flags & RQF_DONTPREP)) {
303 req_to_mmc_queue_req(req)->retries = 0;
304 req->rq_flags |= RQF_DONTPREP;
305 }
306
307 if (get_card)
308 mmc_get_card(card, &mq->ctx);
309
310 if (host->cqe_enabled) {
311 host->retune_now = host->need_retune && cqe_retune_ok &&
312 !host->hold_retune;
313 }
314
315 blk_mq_start_request(req);
316
317 issued = mmc_blk_mq_issue_rq(mq, req);
318
319 switch (issued) {
320 case MMC_REQ_BUSY:
321 ret = BLK_STS_RESOURCE;
322 break;
323 case MMC_REQ_FAILED_TO_START:
324 ret = BLK_STS_IOERR;
325 break;
326 default:
327 ret = BLK_STS_OK;
328 break;
329 }
330
331 if (issued != MMC_REQ_STARTED) {
332 bool put_card = false;
333
334 spin_lock_irq(&mq->lock);
335 mq->in_flight[issue_type] -= 1;
336 if (mmc_tot_in_flight(mq) == 0)
337 put_card = true;
338 mq->busy = false;
339 spin_unlock_irq(&mq->lock);
340 if (put_card)
341 mmc_put_card(card, &mq->ctx);
342 } else {
343 WRITE_ONCE(mq->busy, false);
344 }
345
346 return ret;
347}
348
349static const struct blk_mq_ops mmc_mq_ops = {
350 .queue_rq = mmc_mq_queue_rq,
351 .init_request = mmc_mq_init_request,
352 .exit_request = mmc_mq_exit_request,
353 .complete = mmc_blk_mq_complete,
354 .timeout = mmc_mq_timed_out,
355};
356
357static struct gendisk *mmc_alloc_disk(struct mmc_queue *mq,
358 struct mmc_card *card, unsigned int features)
359{
360 struct mmc_host *host = card->host;
361 struct queue_limits lim = {
362 .features = features,
363 };
364 struct gendisk *disk;
365
366 if (mmc_card_can_erase(card))
367 mmc_queue_setup_discard(card, &lim);
368
369 lim.max_hw_sectors = min(host->max_blk_count, host->max_req_size / 512);
370
371 if (mmc_card_mmc(card) && card->ext_csd.data_sector_size)
372 lim.logical_block_size = card->ext_csd.data_sector_size;
373 else
374 lim.logical_block_size = 512;
375
376 WARN_ON_ONCE(lim.logical_block_size != 512 &&
377 lim.logical_block_size != 4096);
378
379 /*
380 * Setting a virt_boundary implicity sets a max_segment_size, so try
381 * to set the hardware one here.
382 */
383 if (host->can_dma_map_merge) {
384 lim.virt_boundary_mask = dma_get_merge_boundary(mmc_dev(host));
385 lim.max_segments = MMC_DMA_MAP_MERGE_SEGMENTS;
386 } else {
387 lim.max_segment_size =
388 round_down(host->max_seg_size, lim.logical_block_size);
389 lim.max_segments = host->max_segs;
390 }
391
392 if (mmc_host_is_spi(host) && host->use_spi_crc)
393 lim.features |= BLK_FEAT_STABLE_WRITES;
394
395 disk = blk_mq_alloc_disk(&mq->tag_set, &lim, mq);
396 if (IS_ERR(disk))
397 return disk;
398 mq->queue = disk->queue;
399
400 blk_queue_rq_timeout(mq->queue, 60 * HZ);
401
402 if (mmc_dev(host)->dma_parms)
403 dma_set_max_seg_size(mmc_dev(host), queue_max_segment_size(mq->queue));
404
405 INIT_WORK(&mq->recovery_work, mmc_mq_recovery_handler);
406 INIT_WORK(&mq->complete_work, mmc_blk_mq_complete_work);
407
408 mutex_init(&mq->complete_lock);
409
410 init_waitqueue_head(&mq->wait);
411
412 mmc_crypto_setup_queue(mq->queue, host);
413 return disk;
414}
415
416static inline bool mmc_merge_capable(struct mmc_host *host)
417{
418 return host->caps2 & MMC_CAP2_MERGE_CAPABLE;
419}
420
421/* Set queue depth to get a reasonable value for q->nr_requests */
422#define MMC_QUEUE_DEPTH 64
423
424/**
425 * mmc_init_queue - initialise a queue structure.
426 * @mq: mmc queue
427 * @card: mmc card to attach this queue
428 * @features: block layer features (BLK_FEAT_*)
429 *
430 * Initialise a MMC card request queue.
431 */
432struct gendisk *mmc_init_queue(struct mmc_queue *mq, struct mmc_card *card,
433 unsigned int features)
434{
435 struct mmc_host *host = card->host;
436 struct gendisk *disk;
437 int ret;
438
439 mq->card = card;
440
441 spin_lock_init(&mq->lock);
442
443 memset(&mq->tag_set, 0, sizeof(mq->tag_set));
444 mq->tag_set.ops = &mmc_mq_ops;
445 /*
446 * The queue depth for CQE must match the hardware because the request
447 * tag is used to index the hardware queue.
448 */
449 if (host->cqe_enabled && !host->hsq_enabled)
450 mq->tag_set.queue_depth =
451 min_t(int, card->ext_csd.cmdq_depth, host->cqe_qdepth);
452 else
453 mq->tag_set.queue_depth = MMC_QUEUE_DEPTH;
454 mq->tag_set.numa_node = NUMA_NO_NODE;
455 mq->tag_set.flags = BLK_MQ_F_BLOCKING;
456 mq->tag_set.nr_hw_queues = 1;
457 mq->tag_set.cmd_size = sizeof(struct mmc_queue_req);
458 mq->tag_set.driver_data = mq;
459
460 /*
461 * Since blk_mq_alloc_tag_set() calls .init_request() of mmc_mq_ops,
462 * the host->can_dma_map_merge should be set before to get max_segs
463 * from mmc_get_max_segments().
464 */
465 if (mmc_merge_capable(host) &&
466 host->max_segs < MMC_DMA_MAP_MERGE_SEGMENTS &&
467 dma_get_merge_boundary(mmc_dev(host)))
468 host->can_dma_map_merge = 1;
469 else
470 host->can_dma_map_merge = 0;
471
472 ret = blk_mq_alloc_tag_set(&mq->tag_set);
473 if (ret)
474 return ERR_PTR(ret);
475
476
477 disk = mmc_alloc_disk(mq, card, features);
478 if (IS_ERR(disk))
479 blk_mq_free_tag_set(&mq->tag_set);
480 return disk;
481}
482
483void mmc_queue_suspend(struct mmc_queue *mq)
484{
485 blk_mq_quiesce_queue(mq->queue);
486
487 /*
488 * The host remains claimed while there are outstanding requests, so
489 * simply claiming and releasing here ensures there are none.
490 */
491 mmc_claim_host(mq->card->host);
492 mmc_release_host(mq->card->host);
493}
494
495void mmc_queue_resume(struct mmc_queue *mq)
496{
497 blk_mq_unquiesce_queue(mq->queue);
498}
499
500void mmc_cleanup_queue(struct mmc_queue *mq)
501{
502 struct request_queue *q = mq->queue;
503
504 /*
505 * The legacy code handled the possibility of being suspended,
506 * so do that here too.
507 */
508 if (blk_queue_quiesced(q))
509 blk_mq_unquiesce_queue(q);
510
511 /*
512 * If the recovery completes the last (and only remaining) request in
513 * the queue, and the card has been removed, we could end up here with
514 * the recovery not quite finished yet, so cancel it.
515 */
516 cancel_work_sync(&mq->recovery_work);
517
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(req, mqrq->sg);
538}