xref: /openbmc/linux/drivers/mmc/core/queue.c (revision 028822b7)
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 
mmc_cqe_dcmd_busy(struct mmc_queue * mq)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 
mmc_cqe_check_busy(struct mmc_queue * mq)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 
mmc_cqe_can_dcmd(struct mmc_host * host)38 static inline bool mmc_cqe_can_dcmd(struct mmc_host *host)
39 {
40 	return host->caps2 & MMC_CAP2_CQE_DCMD;
41 }
42 
mmc_cqe_issue_type(struct mmc_host * host,struct request * req)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 
mmc_issue_type(struct mmc_queue * mq,struct request * req)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 
__mmc_cqe_recovery_notifier(struct mmc_queue * mq)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 
mmc_cqe_recovery_notifier(struct mmc_request * mrq)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 
mmc_cqe_timed_out(struct request * req)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 
mmc_mq_timed_out(struct request * req)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 
mmc_mq_recovery_handler(struct work_struct * work)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 
mmc_alloc_sg(unsigned short sg_len,gfp_t gfp)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 
mmc_queue_setup_discard(struct request_queue * q,struct mmc_card * card)177 static void mmc_queue_setup_discard(struct request_queue *q,
178 				    struct mmc_card *card)
179 {
180 	unsigned max_discard;
181 
182 	max_discard = mmc_calc_max_discard(card);
183 	if (!max_discard)
184 		return;
185 
186 	blk_queue_max_discard_sectors(q, max_discard);
187 	q->limits.discard_granularity = card->pref_erase << 9;
188 	/* granularity must not be greater than max. discard */
189 	if (card->pref_erase > max_discard)
190 		q->limits.discard_granularity = SECTOR_SIZE;
191 	if (mmc_can_secure_erase_trim(card))
192 		blk_queue_max_secure_erase_sectors(q, max_discard);
193 	if (mmc_can_trim(card) && card->erased_byte == 0)
194 		blk_queue_max_write_zeroes_sectors(q, max_discard);
195 }
196 
mmc_get_max_segments(struct mmc_host * host)197 static unsigned short mmc_get_max_segments(struct mmc_host *host)
198 {
199 	return host->can_dma_map_merge ? MMC_DMA_MAP_MERGE_SEGMENTS :
200 					 host->max_segs;
201 }
202 
mmc_mq_init_request(struct blk_mq_tag_set * set,struct request * req,unsigned int hctx_idx,unsigned int numa_node)203 static int mmc_mq_init_request(struct blk_mq_tag_set *set, struct request *req,
204 			       unsigned int hctx_idx, unsigned int numa_node)
205 {
206 	struct mmc_queue_req *mq_rq = req_to_mmc_queue_req(req);
207 	struct mmc_queue *mq = set->driver_data;
208 	struct mmc_card *card = mq->card;
209 	struct mmc_host *host = card->host;
210 
211 	mq_rq->sg = mmc_alloc_sg(mmc_get_max_segments(host), GFP_KERNEL);
212 	if (!mq_rq->sg)
213 		return -ENOMEM;
214 
215 	return 0;
216 }
217 
mmc_mq_exit_request(struct blk_mq_tag_set * set,struct request * req,unsigned int hctx_idx)218 static void mmc_mq_exit_request(struct blk_mq_tag_set *set, struct request *req,
219 				unsigned int hctx_idx)
220 {
221 	struct mmc_queue_req *mq_rq = req_to_mmc_queue_req(req);
222 
223 	kfree(mq_rq->sg);
224 	mq_rq->sg = NULL;
225 }
226 
mmc_mq_queue_rq(struct blk_mq_hw_ctx * hctx,const struct blk_mq_queue_data * bd)227 static blk_status_t mmc_mq_queue_rq(struct blk_mq_hw_ctx *hctx,
228 				    const struct blk_mq_queue_data *bd)
229 {
230 	struct request *req = bd->rq;
231 	struct request_queue *q = req->q;
232 	struct mmc_queue *mq = q->queuedata;
233 	struct mmc_card *card = mq->card;
234 	struct mmc_host *host = card->host;
235 	enum mmc_issue_type issue_type;
236 	enum mmc_issued issued;
237 	bool get_card, cqe_retune_ok;
238 	blk_status_t ret;
239 
240 	if (mmc_card_removed(mq->card)) {
241 		req->rq_flags |= RQF_QUIET;
242 		return BLK_STS_IOERR;
243 	}
244 
245 	issue_type = mmc_issue_type(mq, req);
246 
247 	spin_lock_irq(&mq->lock);
248 
249 	if (mq->recovery_needed || mq->busy) {
250 		spin_unlock_irq(&mq->lock);
251 		return BLK_STS_RESOURCE;
252 	}
253 
254 	switch (issue_type) {
255 	case MMC_ISSUE_DCMD:
256 		if (mmc_cqe_dcmd_busy(mq)) {
257 			mq->cqe_busy |= MMC_CQE_DCMD_BUSY;
258 			spin_unlock_irq(&mq->lock);
259 			return BLK_STS_RESOURCE;
260 		}
261 		break;
262 	case MMC_ISSUE_ASYNC:
263 		/*
264 		 * For MMC host software queue, we only allow 2 requests in
265 		 * flight to avoid a long latency.
266 		 */
267 		if (host->hsq_enabled && mq->in_flight[issue_type] > 2) {
268 			spin_unlock_irq(&mq->lock);
269 			return BLK_STS_RESOURCE;
270 		}
271 		break;
272 	default:
273 		/*
274 		 * Timeouts are handled by mmc core, and we don't have a host
275 		 * API to abort requests, so we can't handle the timeout anyway.
276 		 * However, when the timeout happens, blk_mq_complete_request()
277 		 * no longer works (to stop the request disappearing under us).
278 		 * To avoid racing with that, set a large timeout.
279 		 */
280 		req->timeout = 600 * HZ;
281 		break;
282 	}
283 
284 	/* Parallel dispatch of requests is not supported at the moment */
285 	mq->busy = true;
286 
287 	mq->in_flight[issue_type] += 1;
288 	get_card = (mmc_tot_in_flight(mq) == 1);
289 	cqe_retune_ok = (mmc_cqe_qcnt(mq) == 1);
290 
291 	spin_unlock_irq(&mq->lock);
292 
293 	if (!(req->rq_flags & RQF_DONTPREP)) {
294 		req_to_mmc_queue_req(req)->retries = 0;
295 		req->rq_flags |= RQF_DONTPREP;
296 	}
297 
298 	if (get_card)
299 		mmc_get_card(card, &mq->ctx);
300 
301 	if (host->cqe_enabled) {
302 		host->retune_now = host->need_retune && cqe_retune_ok &&
303 				   !host->hold_retune;
304 	}
305 
306 	blk_mq_start_request(req);
307 
308 	issued = mmc_blk_mq_issue_rq(mq, req);
309 
310 	switch (issued) {
311 	case MMC_REQ_BUSY:
312 		ret = BLK_STS_RESOURCE;
313 		break;
314 	case MMC_REQ_FAILED_TO_START:
315 		ret = BLK_STS_IOERR;
316 		break;
317 	default:
318 		ret = BLK_STS_OK;
319 		break;
320 	}
321 
322 	if (issued != MMC_REQ_STARTED) {
323 		bool put_card = false;
324 
325 		spin_lock_irq(&mq->lock);
326 		mq->in_flight[issue_type] -= 1;
327 		if (mmc_tot_in_flight(mq) == 0)
328 			put_card = true;
329 		mq->busy = false;
330 		spin_unlock_irq(&mq->lock);
331 		if (put_card)
332 			mmc_put_card(card, &mq->ctx);
333 	} else {
334 		WRITE_ONCE(mq->busy, false);
335 	}
336 
337 	return ret;
338 }
339 
340 static const struct blk_mq_ops mmc_mq_ops = {
341 	.queue_rq	= mmc_mq_queue_rq,
342 	.init_request	= mmc_mq_init_request,
343 	.exit_request	= mmc_mq_exit_request,
344 	.complete	= mmc_blk_mq_complete,
345 	.timeout	= mmc_mq_timed_out,
346 };
347 
mmc_setup_queue(struct mmc_queue * mq,struct mmc_card * card)348 static void mmc_setup_queue(struct mmc_queue *mq, struct mmc_card *card)
349 {
350 	struct mmc_host *host = card->host;
351 	unsigned block_size = 512;
352 
353 	blk_queue_flag_set(QUEUE_FLAG_NONROT, mq->queue);
354 	blk_queue_flag_clear(QUEUE_FLAG_ADD_RANDOM, mq->queue);
355 	if (mmc_can_erase(card))
356 		mmc_queue_setup_discard(mq->queue, card);
357 
358 	if (!mmc_dev(host)->dma_mask || !*mmc_dev(host)->dma_mask)
359 		blk_queue_bounce_limit(mq->queue, BLK_BOUNCE_HIGH);
360 	blk_queue_max_hw_sectors(mq->queue,
361 		min(host->max_blk_count, host->max_req_size / 512));
362 	if (host->can_dma_map_merge)
363 		WARN(!blk_queue_can_use_dma_map_merging(mq->queue,
364 							mmc_dev(host)),
365 		     "merging was advertised but not possible");
366 	blk_queue_max_segments(mq->queue, mmc_get_max_segments(host));
367 
368 	if (mmc_card_mmc(card) && card->ext_csd.data_sector_size) {
369 		block_size = card->ext_csd.data_sector_size;
370 		WARN_ON(block_size != 512 && block_size != 4096);
371 	}
372 
373 	blk_queue_logical_block_size(mq->queue, block_size);
374 	/*
375 	 * After blk_queue_can_use_dma_map_merging() was called with succeed,
376 	 * since it calls blk_queue_virt_boundary(), the mmc should not call
377 	 * both blk_queue_max_segment_size().
378 	 */
379 	if (!host->can_dma_map_merge)
380 		blk_queue_max_segment_size(mq->queue,
381 			round_down(host->max_seg_size, block_size));
382 
383 	dma_set_max_seg_size(mmc_dev(host), queue_max_segment_size(mq->queue));
384 
385 	INIT_WORK(&mq->recovery_work, mmc_mq_recovery_handler);
386 	INIT_WORK(&mq->complete_work, mmc_blk_mq_complete_work);
387 
388 	mutex_init(&mq->complete_lock);
389 
390 	init_waitqueue_head(&mq->wait);
391 
392 	mmc_crypto_setup_queue(mq->queue, host);
393 }
394 
mmc_merge_capable(struct mmc_host * host)395 static inline bool mmc_merge_capable(struct mmc_host *host)
396 {
397 	return host->caps2 & MMC_CAP2_MERGE_CAPABLE;
398 }
399 
400 /* Set queue depth to get a reasonable value for q->nr_requests */
401 #define MMC_QUEUE_DEPTH 64
402 
403 /**
404  * mmc_init_queue - initialise a queue structure.
405  * @mq: mmc queue
406  * @card: mmc card to attach this queue
407  *
408  * Initialise a MMC card request queue.
409  */
mmc_init_queue(struct mmc_queue * mq,struct mmc_card * card)410 struct gendisk *mmc_init_queue(struct mmc_queue *mq, struct mmc_card *card)
411 {
412 	struct mmc_host *host = card->host;
413 	struct gendisk *disk;
414 	int ret;
415 
416 	mq->card = card;
417 
418 	spin_lock_init(&mq->lock);
419 
420 	memset(&mq->tag_set, 0, sizeof(mq->tag_set));
421 	mq->tag_set.ops = &mmc_mq_ops;
422 	/*
423 	 * The queue depth for CQE must match the hardware because the request
424 	 * tag is used to index the hardware queue.
425 	 */
426 	if (host->cqe_enabled && !host->hsq_enabled)
427 		mq->tag_set.queue_depth =
428 			min_t(int, card->ext_csd.cmdq_depth, host->cqe_qdepth);
429 	else
430 		mq->tag_set.queue_depth = MMC_QUEUE_DEPTH;
431 	mq->tag_set.numa_node = NUMA_NO_NODE;
432 	mq->tag_set.flags = BLK_MQ_F_SHOULD_MERGE | BLK_MQ_F_BLOCKING;
433 	mq->tag_set.nr_hw_queues = 1;
434 	mq->tag_set.cmd_size = sizeof(struct mmc_queue_req);
435 	mq->tag_set.driver_data = mq;
436 
437 	/*
438 	 * Since blk_mq_alloc_tag_set() calls .init_request() of mmc_mq_ops,
439 	 * the host->can_dma_map_merge should be set before to get max_segs
440 	 * from mmc_get_max_segments().
441 	 */
442 	if (mmc_merge_capable(host) &&
443 	    host->max_segs < MMC_DMA_MAP_MERGE_SEGMENTS &&
444 	    dma_get_merge_boundary(mmc_dev(host)))
445 		host->can_dma_map_merge = 1;
446 	else
447 		host->can_dma_map_merge = 0;
448 
449 	ret = blk_mq_alloc_tag_set(&mq->tag_set);
450 	if (ret)
451 		return ERR_PTR(ret);
452 
453 
454 	disk = blk_mq_alloc_disk(&mq->tag_set, mq);
455 	if (IS_ERR(disk)) {
456 		blk_mq_free_tag_set(&mq->tag_set);
457 		return disk;
458 	}
459 	mq->queue = disk->queue;
460 
461 	if (mmc_host_is_spi(host) && host->use_spi_crc)
462 		blk_queue_flag_set(QUEUE_FLAG_STABLE_WRITES, mq->queue);
463 	blk_queue_rq_timeout(mq->queue, 60 * HZ);
464 
465 	mmc_setup_queue(mq, card);
466 	return disk;
467 }
468 
mmc_queue_suspend(struct mmc_queue * mq)469 void mmc_queue_suspend(struct mmc_queue *mq)
470 {
471 	blk_mq_quiesce_queue(mq->queue);
472 
473 	/*
474 	 * The host remains claimed while there are outstanding requests, so
475 	 * simply claiming and releasing here ensures there are none.
476 	 */
477 	mmc_claim_host(mq->card->host);
478 	mmc_release_host(mq->card->host);
479 }
480 
mmc_queue_resume(struct mmc_queue * mq)481 void mmc_queue_resume(struct mmc_queue *mq)
482 {
483 	blk_mq_unquiesce_queue(mq->queue);
484 }
485 
mmc_cleanup_queue(struct mmc_queue * mq)486 void mmc_cleanup_queue(struct mmc_queue *mq)
487 {
488 	struct request_queue *q = mq->queue;
489 
490 	/*
491 	 * The legacy code handled the possibility of being suspended,
492 	 * so do that here too.
493 	 */
494 	if (blk_queue_quiesced(q))
495 		blk_mq_unquiesce_queue(q);
496 
497 	/*
498 	 * If the recovery completes the last (and only remaining) request in
499 	 * the queue, and the card has been removed, we could end up here with
500 	 * the recovery not quite finished yet, so cancel it.
501 	 */
502 	cancel_work_sync(&mq->recovery_work);
503 
504 	blk_mq_free_tag_set(&mq->tag_set);
505 
506 	/*
507 	 * A request can be completed before the next request, potentially
508 	 * leaving a complete_work with nothing to do. Such a work item might
509 	 * still be queued at this point. Flush it.
510 	 */
511 	flush_work(&mq->complete_work);
512 
513 	mq->card = NULL;
514 }
515 
516 /*
517  * Prepare the sg list(s) to be handed of to the host driver
518  */
mmc_queue_map_sg(struct mmc_queue * mq,struct mmc_queue_req * mqrq)519 unsigned int mmc_queue_map_sg(struct mmc_queue *mq, struct mmc_queue_req *mqrq)
520 {
521 	struct request *req = mmc_queue_req_to_req(mqrq);
522 
523 	return blk_rq_map_sg(mq->queue, req, mqrq->sg);
524 }
525