1/*
  2 *  Copyright (C) 2003 Russell King, All Rights Reserved.
  3 *  Copyright 2006-2007 Pierre Ossman
  4 *
  5 * This program is free software; you can redistribute it and/or modify
  6 * it under the terms of the GNU General Public License version 2 as
  7 * published by the Free Software Foundation.
  8 *
  9 */
 10#include <linux/slab.h>
 11#include <linux/module.h>
 12#include <linux/blkdev.h>
 13#include <linux/freezer.h>
 14#include <linux/kthread.h>
 15#include <linux/scatterlist.h>
 16#include <linux/dma-mapping.h>
 17
 18#include <linux/mmc/card.h>
 19#include <linux/mmc/host.h>
 20
 21#include "queue.h"
 22#include "block.h"
 23#include "core.h"
 24#include "card.h"
 25#include "host.h"
 26
 27static inline bool mmc_cqe_dcmd_busy(struct mmc_queue *mq)
 28{
 29	/* Allow only 1 DCMD at a time */
 30	return mq->in_flight[MMC_ISSUE_DCMD];
 31}
 32
 33void mmc_cqe_check_busy(struct mmc_queue *mq)
 34{
 35	if ((mq->cqe_busy & MMC_CQE_DCMD_BUSY) && !mmc_cqe_dcmd_busy(mq))
 36		mq->cqe_busy &= ~MMC_CQE_DCMD_BUSY;
 37
 38	mq->cqe_busy &= ~MMC_CQE_QUEUE_FULL;
 39}
 40
 41static inline bool mmc_cqe_can_dcmd(struct mmc_host *host)
 42{
 43	return host->caps2 & MMC_CAP2_CQE_DCMD;
 44}
 45
 46static enum mmc_issue_type mmc_cqe_issue_type(struct mmc_host *host,
 47					      struct request *req)
 48{
 49	switch (req_op(req)) {
 50	case REQ_OP_DRV_IN:
 51	case REQ_OP_DRV_OUT:
 52	case REQ_OP_DISCARD:
 53	case REQ_OP_SECURE_ERASE:
 54		return MMC_ISSUE_SYNC;
 55	case REQ_OP_FLUSH:
 56		return mmc_cqe_can_dcmd(host) ? MMC_ISSUE_DCMD : MMC_ISSUE_SYNC;
 57	default:
 58		return MMC_ISSUE_ASYNC;
 59	}
 60}
 61
 62enum mmc_issue_type mmc_issue_type(struct mmc_queue *mq, struct request *req)
 63{
 64	struct mmc_host *host = mq->card->host;
 65
 66	if (mq->use_cqe)
 67		return mmc_cqe_issue_type(host, req);
 68
 69	if (req_op(req) == REQ_OP_READ || req_op(req) == REQ_OP_WRITE)
 70		return MMC_ISSUE_ASYNC;
 71
 72	return MMC_ISSUE_SYNC;
 73}
 74
 75static void __mmc_cqe_recovery_notifier(struct mmc_queue *mq)
 76{
 77	if (!mq->recovery_needed) {
 78		mq->recovery_needed = true;
 79		schedule_work(&mq->recovery_work);
 80	}
 81}
 82
 83void mmc_cqe_recovery_notifier(struct mmc_request *mrq)
 84{
 85	struct mmc_queue_req *mqrq = container_of(mrq, struct mmc_queue_req,
 86						  brq.mrq);
 87	struct request *req = mmc_queue_req_to_req(mqrq);
 88	struct request_queue *q = req->q;
 89	struct mmc_queue *mq = q->queuedata;
 90	unsigned long flags;
 91
 92	spin_lock_irqsave(q->queue_lock, flags);
 93	__mmc_cqe_recovery_notifier(mq);
 94	spin_unlock_irqrestore(q->queue_lock, flags);
 95}
 96
 97static enum blk_eh_timer_return mmc_cqe_timed_out(struct request *req)
 98{
 99	struct mmc_queue_req *mqrq = req_to_mmc_queue_req(req);
100	struct mmc_request *mrq = &mqrq->brq.mrq;
101	struct mmc_queue *mq = req->q->queuedata;
102	struct mmc_host *host = mq->card->host;
103	enum mmc_issue_type issue_type = mmc_issue_type(mq, req);
104	bool recovery_needed = false;
105
106	switch (issue_type) {
107	case MMC_ISSUE_ASYNC:
108	case MMC_ISSUE_DCMD:
109		if (host->cqe_ops->cqe_timeout(host, mrq, &recovery_needed)) {
110			if (recovery_needed)
111				__mmc_cqe_recovery_notifier(mq);
112			return BLK_EH_RESET_TIMER;
113		}
114		/* No timeout */
115		return BLK_EH_HANDLED;
116	default:
117		/* Timeout is handled by mmc core */
118		return BLK_EH_RESET_TIMER;
119	}
120}
121
122static enum blk_eh_timer_return mmc_mq_timed_out(struct request *req,
123						 bool reserved)
124{
125	struct request_queue *q = req->q;
126	struct mmc_queue *mq = q->queuedata;
127	unsigned long flags;
128	int ret;
129
130	spin_lock_irqsave(q->queue_lock, flags);
131
132	if (mq->recovery_needed || !mq->use_cqe)
133		ret = BLK_EH_RESET_TIMER;
134	else
135		ret = mmc_cqe_timed_out(req);
136
137	spin_unlock_irqrestore(q->queue_lock, flags);
138
139	return ret;
140}
141
142static void mmc_mq_recovery_handler(struct work_struct *work)
143{
144	struct mmc_queue *mq = container_of(work, struct mmc_queue,
145					    recovery_work);
146	struct request_queue *q = mq->queue;
147
148	mmc_get_card(mq->card, &mq->ctx);
149
150	mq->in_recovery = true;
151
152	if (mq->use_cqe)
153		mmc_blk_cqe_recovery(mq);
154	else
155		mmc_blk_mq_recovery(mq);
156
157	mq->in_recovery = false;
158
159	spin_lock_irq(q->queue_lock);
160	mq->recovery_needed = false;
161	spin_unlock_irq(q->queue_lock);
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
198/**
199 * mmc_init_request() - initialize the MMC-specific per-request data
200 * @q: the request queue
201 * @req: the request
202 * @gfp: memory allocation policy
203 */
204static int __mmc_init_request(struct mmc_queue *mq, struct request *req,
205			      gfp_t gfp)
206{
207	struct mmc_queue_req *mq_rq = req_to_mmc_queue_req(req);
208	struct mmc_card *card = mq->card;
209	struct mmc_host *host = card->host;
210
211	mq_rq->sg = mmc_alloc_sg(host->max_segs, gfp);
212	if (!mq_rq->sg)
213		return -ENOMEM;
214
215	return 0;
216}
217
218static void mmc_exit_request(struct request_queue *q, struct request *req)
219{
220	struct mmc_queue_req *mq_rq = req_to_mmc_queue_req(req);
221
222	kfree(mq_rq->sg);
223	mq_rq->sg = NULL;
224}
225
226static int mmc_mq_init_request(struct blk_mq_tag_set *set, struct request *req,
227			       unsigned int hctx_idx, unsigned int numa_node)
228{
229	return __mmc_init_request(set->driver_data, req, GFP_KERNEL);
230}
231
232static void mmc_mq_exit_request(struct blk_mq_tag_set *set, struct request *req,
233				unsigned int hctx_idx)
234{
235	struct mmc_queue *mq = set->driver_data;
236
237	mmc_exit_request(mq->queue, req);
238}
239
240/*
241 * We use BLK_MQ_F_BLOCKING and have only 1 hardware queue, which means requests
242 * will not be dispatched in parallel.
243 */
244static blk_status_t mmc_mq_queue_rq(struct blk_mq_hw_ctx *hctx,
245				    const struct blk_mq_queue_data *bd)
246{
247	struct request *req = bd->rq;
248	struct request_queue *q = req->q;
249	struct mmc_queue *mq = q->queuedata;
250	struct mmc_card *card = mq->card;
251	struct mmc_host *host = card->host;
252	enum mmc_issue_type issue_type;
253	enum mmc_issued issued;
254	bool get_card, cqe_retune_ok;
255	int ret;
256
257	if (mmc_card_removed(mq->card)) {
258		req->rq_flags |= RQF_QUIET;
259		return BLK_STS_IOERR;
260	}
261
262	issue_type = mmc_issue_type(mq, req);
263
264	spin_lock_irq(q->queue_lock);
265
266	if (mq->recovery_needed) {
267		spin_unlock_irq(q->queue_lock);
268		return BLK_STS_RESOURCE;
269	}
270
271	switch (issue_type) {
272	case MMC_ISSUE_DCMD:
273		if (mmc_cqe_dcmd_busy(mq)) {
274			mq->cqe_busy |= MMC_CQE_DCMD_BUSY;
275			spin_unlock_irq(q->queue_lock);
276			return BLK_STS_RESOURCE;
277		}
278		break;
279	case MMC_ISSUE_ASYNC:
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	mq->in_flight[issue_type] += 1;
294	get_card = (mmc_tot_in_flight(mq) == 1);
295	cqe_retune_ok = (mmc_cqe_qcnt(mq) == 1);
296
297	spin_unlock_irq(q->queue_lock);
298
299	if (!(req->rq_flags & RQF_DONTPREP)) {
300		req_to_mmc_queue_req(req)->retries = 0;
301		req->rq_flags |= RQF_DONTPREP;
302	}
303
304	if (get_card)
305		mmc_get_card(card, &mq->ctx);
306
307	if (mq->use_cqe) {
308		host->retune_now = host->need_retune && cqe_retune_ok &&
309				   !host->hold_retune;
310	}
311
312	blk_mq_start_request(req);
313
314	issued = mmc_blk_mq_issue_rq(mq, req);
315
316	switch (issued) {
317	case MMC_REQ_BUSY:
318		ret = BLK_STS_RESOURCE;
319		break;
320	case MMC_REQ_FAILED_TO_START:
321		ret = BLK_STS_IOERR;
322		break;
323	default:
324		ret = BLK_STS_OK;
325		break;
326	}
327
328	if (issued != MMC_REQ_STARTED) {
329		bool put_card = false;
330
331		spin_lock_irq(q->queue_lock);
332		mq->in_flight[issue_type] -= 1;
333		if (mmc_tot_in_flight(mq) == 0)
334			put_card = true;
335		spin_unlock_irq(q->queue_lock);
336		if (put_card)
337			mmc_put_card(card, &mq->ctx);
338	}
339
340	return ret;
341}
342
343static const struct blk_mq_ops mmc_mq_ops = {
344	.queue_rq	= mmc_mq_queue_rq,
345	.init_request	= mmc_mq_init_request,
346	.exit_request	= mmc_mq_exit_request,
347	.complete	= mmc_blk_mq_complete,
348	.timeout	= mmc_mq_timed_out,
349};
350
351static void mmc_setup_queue(struct mmc_queue *mq, struct mmc_card *card)
352{
353	struct mmc_host *host = card->host;
354	u64 limit = BLK_BOUNCE_HIGH;
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	blk_queue_max_segment_size(mq->queue, host->max_seg_size);
369
370	INIT_WORK(&mq->recovery_work, mmc_mq_recovery_handler);
371	INIT_WORK(&mq->complete_work, mmc_blk_mq_complete_work);
372
373	mutex_init(&mq->complete_lock);
374
375	init_waitqueue_head(&mq->wait);
376}
377
378static int mmc_mq_init_queue(struct mmc_queue *mq, int q_depth,
379			     const struct blk_mq_ops *mq_ops, spinlock_t *lock)
380{
381	int ret;
382
383	memset(&mq->tag_set, 0, sizeof(mq->tag_set));
384	mq->tag_set.ops = mq_ops;
385	mq->tag_set.queue_depth = q_depth;
386	mq->tag_set.numa_node = NUMA_NO_NODE;
387	mq->tag_set.flags = BLK_MQ_F_SHOULD_MERGE | BLK_MQ_F_SG_MERGE |
388			    BLK_MQ_F_BLOCKING;
389	mq->tag_set.nr_hw_queues = 1;
390	mq->tag_set.cmd_size = sizeof(struct mmc_queue_req);
391	mq->tag_set.driver_data = mq;
392
393	ret = blk_mq_alloc_tag_set(&mq->tag_set);
394	if (ret)
395		return ret;
396
397	mq->queue = blk_mq_init_queue(&mq->tag_set);
398	if (IS_ERR(mq->queue)) {
399		ret = PTR_ERR(mq->queue);
400		goto free_tag_set;
401	}
402
403	mq->queue->queue_lock = lock;
404	mq->queue->queuedata = mq;
405
406	return 0;
407
408free_tag_set:
409	blk_mq_free_tag_set(&mq->tag_set);
410
411	return ret;
412}
413
414/* Set queue depth to get a reasonable value for q->nr_requests */
415#define MMC_QUEUE_DEPTH 64
416
417static int mmc_mq_init(struct mmc_queue *mq, struct mmc_card *card,
418			 spinlock_t *lock)
419{
420	struct mmc_host *host = card->host;
421	int q_depth;
422	int ret;
423
424	/*
425	 * The queue depth for CQE must match the hardware because the request
426	 * tag is used to index the hardware queue.
427	 */
428	if (mq->use_cqe)
429		q_depth = min_t(int, card->ext_csd.cmdq_depth, host->cqe_qdepth);
430	else
431		q_depth = MMC_QUEUE_DEPTH;
432
433	ret = mmc_mq_init_queue(mq, q_depth, &mmc_mq_ops, lock);
434	if (ret)
435		return ret;
436
437	blk_queue_rq_timeout(mq->queue, 60 * HZ);
438
439	mmc_setup_queue(mq, card);
440
441	return 0;
442}
443
444/**
445 * mmc_init_queue - initialise a queue structure.
446 * @mq: mmc queue
447 * @card: mmc card to attach this queue
448 * @lock: queue lock
449 * @subname: partition subname
450 *
451 * Initialise a MMC card request queue.
452 */
453int mmc_init_queue(struct mmc_queue *mq, struct mmc_card *card,
454		   spinlock_t *lock, const char *subname)
455{
456	struct mmc_host *host = card->host;
457
458	mq->card = card;
459
460	mq->use_cqe = host->cqe_enabled;
461
462	return mmc_mq_init(mq, card, lock);
463}
464
465void mmc_queue_suspend(struct mmc_queue *mq)
466{
467	blk_mq_quiesce_queue(mq->queue);
468
469	/*
470	 * The host remains claimed while there are outstanding requests, so
471	 * simply claiming and releasing here ensures there are none.
472	 */
473	mmc_claim_host(mq->card->host);
474	mmc_release_host(mq->card->host);
475}
476
477void mmc_queue_resume(struct mmc_queue *mq)
478{
479	blk_mq_unquiesce_queue(mq->queue);
480}
481
482void mmc_cleanup_queue(struct mmc_queue *mq)
483{
484	struct request_queue *q = mq->queue;
485
486	/*
487	 * The legacy code handled the possibility of being suspended,
488	 * so do that here too.
489	 */
490	if (blk_queue_quiesced(q))
491		blk_mq_unquiesce_queue(q);
492
493	blk_cleanup_queue(q);
494
495	/*
496	 * A request can be completed before the next request, potentially
497	 * leaving a complete_work with nothing to do. Such a work item might
498	 * still be queued at this point. Flush it.
499	 */
500	flush_work(&mq->complete_work);
501
502	mq->card = NULL;
503}
504
505/*
506 * Prepare the sg list(s) to be handed of to the host driver
507 */
508unsigned int mmc_queue_map_sg(struct mmc_queue *mq, struct mmc_queue_req *mqrq)
509{
510	struct request *req = mmc_queue_req_to_req(mqrq);
511
512	return blk_rq_map_sg(mq->queue, req, mqrq->sg);
513}