xref: /openbmc/linux/drivers/mmc/core/queue.c (revision 22d55f02)
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 
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 "host.h"
22 
23 static inline bool mmc_cqe_dcmd_busy(struct mmc_queue *mq)
24 {
25 	/* Allow only 1 DCMD at a time */
26 	return mq->in_flight[MMC_ISSUE_DCMD];
27 }
28 
29 void mmc_cqe_check_busy(struct mmc_queue *mq)
30 {
31 	if ((mq->cqe_busy & MMC_CQE_DCMD_BUSY) && !mmc_cqe_dcmd_busy(mq))
32 		mq->cqe_busy &= ~MMC_CQE_DCMD_BUSY;
33 
34 	mq->cqe_busy &= ~MMC_CQE_QUEUE_FULL;
35 }
36 
37 static inline bool mmc_cqe_can_dcmd(struct mmc_host *host)
38 {
39 	return host->caps2 & MMC_CAP2_CQE_DCMD;
40 }
41 
42 static enum mmc_issue_type mmc_cqe_issue_type(struct mmc_host *host,
43 					      struct request *req)
44 {
45 	switch (req_op(req)) {
46 	case REQ_OP_DRV_IN:
47 	case REQ_OP_DRV_OUT:
48 	case REQ_OP_DISCARD:
49 	case REQ_OP_SECURE_ERASE:
50 		return MMC_ISSUE_SYNC;
51 	case REQ_OP_FLUSH:
52 		return mmc_cqe_can_dcmd(host) ? MMC_ISSUE_DCMD : MMC_ISSUE_SYNC;
53 	default:
54 		return MMC_ISSUE_ASYNC;
55 	}
56 }
57 
58 enum mmc_issue_type mmc_issue_type(struct mmc_queue *mq, struct request *req)
59 {
60 	struct mmc_host *host = mq->card->host;
61 
62 	if (mq->use_cqe)
63 		return mmc_cqe_issue_type(host, req);
64 
65 	if (req_op(req) == REQ_OP_READ || req_op(req) == REQ_OP_WRITE)
66 		return MMC_ISSUE_ASYNC;
67 
68 	return MMC_ISSUE_SYNC;
69 }
70 
71 static void __mmc_cqe_recovery_notifier(struct mmc_queue *mq)
72 {
73 	if (!mq->recovery_needed) {
74 		mq->recovery_needed = true;
75 		schedule_work(&mq->recovery_work);
76 	}
77 }
78 
79 void mmc_cqe_recovery_notifier(struct mmc_request *mrq)
80 {
81 	struct mmc_queue_req *mqrq = container_of(mrq, struct mmc_queue_req,
82 						  brq.mrq);
83 	struct request *req = mmc_queue_req_to_req(mqrq);
84 	struct request_queue *q = req->q;
85 	struct mmc_queue *mq = q->queuedata;
86 	unsigned long flags;
87 
88 	spin_lock_irqsave(&mq->lock, flags);
89 	__mmc_cqe_recovery_notifier(mq);
90 	spin_unlock_irqrestore(&mq->lock, flags);
91 }
92 
93 static enum blk_eh_timer_return mmc_cqe_timed_out(struct request *req)
94 {
95 	struct mmc_queue_req *mqrq = req_to_mmc_queue_req(req);
96 	struct mmc_request *mrq = &mqrq->brq.mrq;
97 	struct mmc_queue *mq = req->q->queuedata;
98 	struct mmc_host *host = mq->card->host;
99 	enum mmc_issue_type issue_type = mmc_issue_type(mq, req);
100 	bool recovery_needed = false;
101 
102 	switch (issue_type) {
103 	case MMC_ISSUE_ASYNC:
104 	case MMC_ISSUE_DCMD:
105 		if (host->cqe_ops->cqe_timeout(host, mrq, &recovery_needed)) {
106 			if (recovery_needed)
107 				__mmc_cqe_recovery_notifier(mq);
108 			return BLK_EH_RESET_TIMER;
109 		}
110 		/* No timeout (XXX: huh? comment doesn't make much sense) */
111 		blk_mq_complete_request(req);
112 		return BLK_EH_DONE;
113 	default:
114 		/* Timeout is handled by mmc core */
115 		return BLK_EH_RESET_TIMER;
116 	}
117 }
118 
119 static enum blk_eh_timer_return mmc_mq_timed_out(struct request *req,
120 						 bool reserved)
121 {
122 	struct request_queue *q = req->q;
123 	struct mmc_queue *mq = q->queuedata;
124 	unsigned long flags;
125 	int ret;
126 
127 	spin_lock_irqsave(&mq->lock, flags);
128 
129 	if (mq->recovery_needed || !mq->use_cqe)
130 		ret = BLK_EH_RESET_TIMER;
131 	else
132 		ret = mmc_cqe_timed_out(req);
133 
134 	spin_unlock_irqrestore(&mq->lock, flags);
135 
136 	return ret;
137 }
138 
139 static void mmc_mq_recovery_handler(struct work_struct *work)
140 {
141 	struct mmc_queue *mq = container_of(work, struct mmc_queue,
142 					    recovery_work);
143 	struct request_queue *q = mq->queue;
144 
145 	mmc_get_card(mq->card, &mq->ctx);
146 
147 	mq->in_recovery = true;
148 
149 	if (mq->use_cqe)
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 	mmc_put_card(mq->card, &mq->ctx);
161 
162 	blk_mq_run_hw_queues(q, true);
163 }
164 
165 static struct scatterlist *mmc_alloc_sg(int sg_len, gfp_t gfp)
166 {
167 	struct scatterlist *sg;
168 
169 	sg = kmalloc_array(sg_len, sizeof(*sg), gfp);
170 	if (sg)
171 		sg_init_table(sg, sg_len);
172 
173 	return sg;
174 }
175 
176 static void mmc_queue_setup_discard(struct request_queue *q,
177 				    struct mmc_card *card)
178 {
179 	unsigned max_discard;
180 
181 	max_discard = mmc_calc_max_discard(card);
182 	if (!max_discard)
183 		return;
184 
185 	blk_queue_flag_set(QUEUE_FLAG_DISCARD, q);
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 = 0;
191 	if (mmc_can_secure_erase_trim(card))
192 		blk_queue_flag_set(QUEUE_FLAG_SECERASE, q);
193 }
194 
195 /**
196  * mmc_init_request() - initialize the MMC-specific per-request data
197  * @q: the request queue
198  * @req: the request
199  * @gfp: memory allocation policy
200  */
201 static int __mmc_init_request(struct mmc_queue *mq, struct request *req,
202 			      gfp_t gfp)
203 {
204 	struct mmc_queue_req *mq_rq = req_to_mmc_queue_req(req);
205 	struct mmc_card *card = mq->card;
206 	struct mmc_host *host = card->host;
207 
208 	mq_rq->sg = mmc_alloc_sg(host->max_segs, gfp);
209 	if (!mq_rq->sg)
210 		return -ENOMEM;
211 
212 	return 0;
213 }
214 
215 static void mmc_exit_request(struct request_queue *q, struct request *req)
216 {
217 	struct mmc_queue_req *mq_rq = req_to_mmc_queue_req(req);
218 
219 	kfree(mq_rq->sg);
220 	mq_rq->sg = NULL;
221 }
222 
223 static int mmc_mq_init_request(struct blk_mq_tag_set *set, struct request *req,
224 			       unsigned int hctx_idx, unsigned int numa_node)
225 {
226 	return __mmc_init_request(set->driver_data, req, GFP_KERNEL);
227 }
228 
229 static void mmc_mq_exit_request(struct blk_mq_tag_set *set, struct request *req,
230 				unsigned int hctx_idx)
231 {
232 	struct mmc_queue *mq = set->driver_data;
233 
234 	mmc_exit_request(mq->queue, req);
235 }
236 
237 static blk_status_t mmc_mq_queue_rq(struct blk_mq_hw_ctx *hctx,
238 				    const struct blk_mq_queue_data *bd)
239 {
240 	struct request *req = bd->rq;
241 	struct request_queue *q = req->q;
242 	struct mmc_queue *mq = q->queuedata;
243 	struct mmc_card *card = mq->card;
244 	struct mmc_host *host = card->host;
245 	enum mmc_issue_type issue_type;
246 	enum mmc_issued issued;
247 	bool get_card, cqe_retune_ok;
248 	int ret;
249 
250 	if (mmc_card_removed(mq->card)) {
251 		req->rq_flags |= RQF_QUIET;
252 		return BLK_STS_IOERR;
253 	}
254 
255 	issue_type = mmc_issue_type(mq, req);
256 
257 	spin_lock_irq(&mq->lock);
258 
259 	if (mq->recovery_needed || mq->busy) {
260 		spin_unlock_irq(&mq->lock);
261 		return BLK_STS_RESOURCE;
262 	}
263 
264 	switch (issue_type) {
265 	case MMC_ISSUE_DCMD:
266 		if (mmc_cqe_dcmd_busy(mq)) {
267 			mq->cqe_busy |= MMC_CQE_DCMD_BUSY;
268 			spin_unlock_irq(&mq->lock);
269 			return BLK_STS_RESOURCE;
270 		}
271 		break;
272 	case MMC_ISSUE_ASYNC:
273 		break;
274 	default:
275 		/*
276 		 * Timeouts are handled by mmc core, and we don't have a host
277 		 * API to abort requests, so we can't handle the timeout anyway.
278 		 * However, when the timeout happens, blk_mq_complete_request()
279 		 * no longer works (to stop the request disappearing under us).
280 		 * To avoid racing with that, set a large timeout.
281 		 */
282 		req->timeout = 600 * HZ;
283 		break;
284 	}
285 
286 	/* Parallel dispatch of requests is not supported at the moment */
287 	mq->busy = true;
288 
289 	mq->in_flight[issue_type] += 1;
290 	get_card = (mmc_tot_in_flight(mq) == 1);
291 	cqe_retune_ok = (mmc_cqe_qcnt(mq) == 1);
292 
293 	spin_unlock_irq(&mq->lock);
294 
295 	if (!(req->rq_flags & RQF_DONTPREP)) {
296 		req_to_mmc_queue_req(req)->retries = 0;
297 		req->rq_flags |= RQF_DONTPREP;
298 	}
299 
300 	if (get_card)
301 		mmc_get_card(card, &mq->ctx);
302 
303 	if (mq->use_cqe) {
304 		host->retune_now = host->need_retune && cqe_retune_ok &&
305 				   !host->hold_retune;
306 	}
307 
308 	blk_mq_start_request(req);
309 
310 	issued = mmc_blk_mq_issue_rq(mq, req);
311 
312 	switch (issued) {
313 	case MMC_REQ_BUSY:
314 		ret = BLK_STS_RESOURCE;
315 		break;
316 	case MMC_REQ_FAILED_TO_START:
317 		ret = BLK_STS_IOERR;
318 		break;
319 	default:
320 		ret = BLK_STS_OK;
321 		break;
322 	}
323 
324 	if (issued != MMC_REQ_STARTED) {
325 		bool put_card = false;
326 
327 		spin_lock_irq(&mq->lock);
328 		mq->in_flight[issue_type] -= 1;
329 		if (mmc_tot_in_flight(mq) == 0)
330 			put_card = true;
331 		mq->busy = false;
332 		spin_unlock_irq(&mq->lock);
333 		if (put_card)
334 			mmc_put_card(card, &mq->ctx);
335 	} else {
336 		WRITE_ONCE(mq->busy, false);
337 	}
338 
339 	return ret;
340 }
341 
342 static const struct blk_mq_ops mmc_mq_ops = {
343 	.queue_rq	= mmc_mq_queue_rq,
344 	.init_request	= mmc_mq_init_request,
345 	.exit_request	= mmc_mq_exit_request,
346 	.complete	= mmc_blk_mq_complete,
347 	.timeout	= mmc_mq_timed_out,
348 };
349 
350 static void mmc_setup_queue(struct mmc_queue *mq, struct mmc_card *card)
351 {
352 	struct mmc_host *host = card->host;
353 	u64 limit = BLK_BOUNCE_HIGH;
354 	unsigned block_size = 512;
355 
356 	if (mmc_dev(host)->dma_mask && *mmc_dev(host)->dma_mask)
357 		limit = (u64)dma_max_pfn(mmc_dev(host)) << PAGE_SHIFT;
358 
359 	blk_queue_flag_set(QUEUE_FLAG_NONROT, mq->queue);
360 	blk_queue_flag_clear(QUEUE_FLAG_ADD_RANDOM, mq->queue);
361 	if (mmc_can_erase(card))
362 		mmc_queue_setup_discard(mq->queue, card);
363 
364 	blk_queue_bounce_limit(mq->queue, limit);
365 	blk_queue_max_hw_sectors(mq->queue,
366 		min(host->max_blk_count, host->max_req_size / 512));
367 	blk_queue_max_segments(mq->queue, host->max_segs);
368 
369 	if (mmc_card_mmc(card))
370 		block_size = card->ext_csd.data_sector_size;
371 
372 	blk_queue_logical_block_size(mq->queue, block_size);
373 	blk_queue_max_segment_size(mq->queue,
374 			round_down(host->max_seg_size, block_size));
375 
376 	dma_set_max_seg_size(mmc_dev(host), queue_max_segment_size(mq->queue));
377 
378 	INIT_WORK(&mq->recovery_work, mmc_mq_recovery_handler);
379 	INIT_WORK(&mq->complete_work, mmc_blk_mq_complete_work);
380 
381 	mutex_init(&mq->complete_lock);
382 
383 	init_waitqueue_head(&mq->wait);
384 }
385 
386 /* Set queue depth to get a reasonable value for q->nr_requests */
387 #define MMC_QUEUE_DEPTH 64
388 
389 /**
390  * mmc_init_queue - initialise a queue structure.
391  * @mq: mmc queue
392  * @card: mmc card to attach this queue
393  *
394  * Initialise a MMC card request queue.
395  */
396 int mmc_init_queue(struct mmc_queue *mq, struct mmc_card *card)
397 {
398 	struct mmc_host *host = card->host;
399 	int ret;
400 
401 	mq->card = card;
402 	mq->use_cqe = host->cqe_enabled;
403 
404 	spin_lock_init(&mq->lock);
405 
406 	memset(&mq->tag_set, 0, sizeof(mq->tag_set));
407 	mq->tag_set.ops = &mmc_mq_ops;
408 	/*
409 	 * The queue depth for CQE must match the hardware because the request
410 	 * tag is used to index the hardware queue.
411 	 */
412 	if (mq->use_cqe)
413 		mq->tag_set.queue_depth =
414 			min_t(int, card->ext_csd.cmdq_depth, host->cqe_qdepth);
415 	else
416 		mq->tag_set.queue_depth = MMC_QUEUE_DEPTH;
417 	mq->tag_set.numa_node = NUMA_NO_NODE;
418 	mq->tag_set.flags = BLK_MQ_F_SHOULD_MERGE | BLK_MQ_F_BLOCKING;
419 	mq->tag_set.nr_hw_queues = 1;
420 	mq->tag_set.cmd_size = sizeof(struct mmc_queue_req);
421 	mq->tag_set.driver_data = mq;
422 
423 	ret = blk_mq_alloc_tag_set(&mq->tag_set);
424 	if (ret)
425 		return ret;
426 
427 	mq->queue = blk_mq_init_queue(&mq->tag_set);
428 	if (IS_ERR(mq->queue)) {
429 		ret = PTR_ERR(mq->queue);
430 		goto free_tag_set;
431 	}
432 
433 	mq->queue->queuedata = mq;
434 	blk_queue_rq_timeout(mq->queue, 60 * HZ);
435 
436 	mmc_setup_queue(mq, card);
437 	return 0;
438 
439 free_tag_set:
440 	blk_mq_free_tag_set(&mq->tag_set);
441 	return ret;
442 }
443 
444 void mmc_queue_suspend(struct mmc_queue *mq)
445 {
446 	blk_mq_quiesce_queue(mq->queue);
447 
448 	/*
449 	 * The host remains claimed while there are outstanding requests, so
450 	 * simply claiming and releasing here ensures there are none.
451 	 */
452 	mmc_claim_host(mq->card->host);
453 	mmc_release_host(mq->card->host);
454 }
455 
456 void mmc_queue_resume(struct mmc_queue *mq)
457 {
458 	blk_mq_unquiesce_queue(mq->queue);
459 }
460 
461 void mmc_cleanup_queue(struct mmc_queue *mq)
462 {
463 	struct request_queue *q = mq->queue;
464 
465 	/*
466 	 * The legacy code handled the possibility of being suspended,
467 	 * so do that here too.
468 	 */
469 	if (blk_queue_quiesced(q))
470 		blk_mq_unquiesce_queue(q);
471 
472 	blk_cleanup_queue(q);
473 	blk_mq_free_tag_set(&mq->tag_set);
474 
475 	/*
476 	 * A request can be completed before the next request, potentially
477 	 * leaving a complete_work with nothing to do. Such a work item might
478 	 * still be queued at this point. Flush it.
479 	 */
480 	flush_work(&mq->complete_work);
481 
482 	mq->card = NULL;
483 }
484 
485 /*
486  * Prepare the sg list(s) to be handed of to the host driver
487  */
488 unsigned int mmc_queue_map_sg(struct mmc_queue *mq, struct mmc_queue_req *mqrq)
489 {
490 	struct request *req = mmc_queue_req_to_req(mqrq);
491 
492 	return blk_rq_map_sg(mq->queue, req, mqrq->sg);
493 }
494