]> git.ipfire.org Git - thirdparty/linux.git/blob - drivers/mmc/core/queue.c
io_uring: reset -EBUSY error when io sq thread is waken up
[thirdparty/linux.git] / drivers / mmc / core / queue.c
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 #include <linux/backing-dev.h>
14
15 #include <linux/mmc/card.h>
16 #include <linux/mmc/host.h>
17
18 #include "queue.h"
19 #include "block.h"
20 #include "core.h"
21 #include "card.h"
22 #include "host.h"
23
24 #define MMC_DMA_MAP_MERGE_SEGMENTS 512
25
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
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 mq->cqe_busy &= ~MMC_CQE_QUEUE_FULL;
38 }
39
40 static inline bool mmc_cqe_can_dcmd(struct mmc_host *host)
41 {
42 return host->caps2 & MMC_CAP2_CQE_DCMD;
43 }
44
45 static enum mmc_issue_type mmc_cqe_issue_type(struct mmc_host *host,
46 struct request *req)
47 {
48 switch (req_op(req)) {
49 case REQ_OP_DRV_IN:
50 case REQ_OP_DRV_OUT:
51 case REQ_OP_DISCARD:
52 case REQ_OP_SECURE_ERASE:
53 return MMC_ISSUE_SYNC;
54 case REQ_OP_FLUSH:
55 return mmc_cqe_can_dcmd(host) ? MMC_ISSUE_DCMD : MMC_ISSUE_SYNC;
56 default:
57 return MMC_ISSUE_ASYNC;
58 }
59 }
60
61 enum mmc_issue_type mmc_issue_type(struct mmc_queue *mq, struct request *req)
62 {
63 struct mmc_host *host = mq->card->host;
64
65 if (mq->use_cqe && !host->hsq_enabled)
66 return mmc_cqe_issue_type(host, req);
67
68 if (req_op(req) == REQ_OP_READ || req_op(req) == REQ_OP_WRITE)
69 return MMC_ISSUE_ASYNC;
70
71 return MMC_ISSUE_SYNC;
72 }
73
74 static void __mmc_cqe_recovery_notifier(struct mmc_queue *mq)
75 {
76 if (!mq->recovery_needed) {
77 mq->recovery_needed = true;
78 schedule_work(&mq->recovery_work);
79 }
80 }
81
82 void mmc_cqe_recovery_notifier(struct mmc_request *mrq)
83 {
84 struct mmc_queue_req *mqrq = container_of(mrq, struct mmc_queue_req,
85 brq.mrq);
86 struct request *req = mmc_queue_req_to_req(mqrq);
87 struct request_queue *q = req->q;
88 struct mmc_queue *mq = q->queuedata;
89 unsigned long flags;
90
91 spin_lock_irqsave(&mq->lock, flags);
92 __mmc_cqe_recovery_notifier(mq);
93 spin_unlock_irqrestore(&mq->lock, flags);
94 }
95
96 static enum blk_eh_timer_return mmc_cqe_timed_out(struct request *req)
97 {
98 struct mmc_queue_req *mqrq = req_to_mmc_queue_req(req);
99 struct mmc_request *mrq = &mqrq->brq.mrq;
100 struct mmc_queue *mq = req->q->queuedata;
101 struct mmc_host *host = mq->card->host;
102 enum mmc_issue_type issue_type = mmc_issue_type(mq, req);
103 bool recovery_needed = false;
104
105 switch (issue_type) {
106 case MMC_ISSUE_ASYNC:
107 case MMC_ISSUE_DCMD:
108 if (host->cqe_ops->cqe_timeout(host, mrq, &recovery_needed)) {
109 if (recovery_needed)
110 __mmc_cqe_recovery_notifier(mq);
111 return BLK_EH_RESET_TIMER;
112 }
113 /* No timeout (XXX: huh? comment doesn't make much sense) */
114 blk_mq_complete_request(req);
115 return BLK_EH_DONE;
116 default:
117 /* Timeout is handled by mmc core */
118 return BLK_EH_RESET_TIMER;
119 }
120 }
121
122 static 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 struct mmc_card *card = mq->card;
128 struct mmc_host *host = card->host;
129 unsigned long flags;
130 int ret;
131
132 spin_lock_irqsave(&mq->lock, flags);
133
134 if (mq->recovery_needed || !mq->use_cqe || host->hsq_enabled)
135 ret = BLK_EH_RESET_TIMER;
136 else
137 ret = mmc_cqe_timed_out(req);
138
139 spin_unlock_irqrestore(&mq->lock, flags);
140
141 return ret;
142 }
143
144 static void mmc_mq_recovery_handler(struct work_struct *work)
145 {
146 struct mmc_queue *mq = container_of(work, struct mmc_queue,
147 recovery_work);
148 struct request_queue *q = mq->queue;
149 struct mmc_host *host = mq->card->host;
150
151 mmc_get_card(mq->card, &mq->ctx);
152
153 mq->in_recovery = true;
154
155 if (mq->use_cqe && !host->hsq_enabled)
156 mmc_blk_cqe_recovery(mq);
157 else
158 mmc_blk_mq_recovery(mq);
159
160 mq->in_recovery = false;
161
162 spin_lock_irq(&mq->lock);
163 mq->recovery_needed = false;
164 spin_unlock_irq(&mq->lock);
165
166 if (host->hsq_enabled)
167 host->cqe_ops->cqe_recovery_finish(host);
168
169 mmc_put_card(mq->card, &mq->ctx);
170
171 blk_mq_run_hw_queues(q, true);
172 }
173
174 static struct scatterlist *mmc_alloc_sg(int sg_len, gfp_t gfp)
175 {
176 struct scatterlist *sg;
177
178 sg = kmalloc_array(sg_len, sizeof(*sg), gfp);
179 if (sg)
180 sg_init_table(sg, sg_len);
181
182 return sg;
183 }
184
185 static void mmc_queue_setup_discard(struct request_queue *q,
186 struct mmc_card *card)
187 {
188 unsigned max_discard;
189
190 max_discard = mmc_calc_max_discard(card);
191 if (!max_discard)
192 return;
193
194 blk_queue_flag_set(QUEUE_FLAG_DISCARD, q);
195 blk_queue_max_discard_sectors(q, max_discard);
196 q->limits.discard_granularity = card->pref_erase << 9;
197 /* granularity must not be greater than max. discard */
198 if (card->pref_erase > max_discard)
199 q->limits.discard_granularity = 0;
200 if (mmc_can_secure_erase_trim(card))
201 blk_queue_flag_set(QUEUE_FLAG_SECERASE, q);
202 }
203
204 static unsigned int 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
210 /**
211 * mmc_init_request() - initialize the MMC-specific per-request data
212 * @q: the request queue
213 * @req: the request
214 * @gfp: memory allocation policy
215 */
216 static int __mmc_init_request(struct mmc_queue *mq, struct request *req,
217 gfp_t gfp)
218 {
219 struct mmc_queue_req *mq_rq = req_to_mmc_queue_req(req);
220 struct mmc_card *card = mq->card;
221 struct mmc_host *host = card->host;
222
223 mq_rq->sg = mmc_alloc_sg(mmc_get_max_segments(host), gfp);
224 if (!mq_rq->sg)
225 return -ENOMEM;
226
227 return 0;
228 }
229
230 static void mmc_exit_request(struct request_queue *q, struct request *req)
231 {
232 struct mmc_queue_req *mq_rq = req_to_mmc_queue_req(req);
233
234 kfree(mq_rq->sg);
235 mq_rq->sg = NULL;
236 }
237
238 static int mmc_mq_init_request(struct blk_mq_tag_set *set, struct request *req,
239 unsigned int hctx_idx, unsigned int numa_node)
240 {
241 return __mmc_init_request(set->driver_data, req, GFP_KERNEL);
242 }
243
244 static void mmc_mq_exit_request(struct blk_mq_tag_set *set, struct request *req,
245 unsigned int hctx_idx)
246 {
247 struct mmc_queue *mq = set->driver_data;
248
249 mmc_exit_request(mq->queue, req);
250 }
251
252 static blk_status_t mmc_mq_queue_rq(struct blk_mq_hw_ctx *hctx,
253 const struct blk_mq_queue_data *bd)
254 {
255 struct request *req = bd->rq;
256 struct request_queue *q = req->q;
257 struct mmc_queue *mq = q->queuedata;
258 struct mmc_card *card = mq->card;
259 struct mmc_host *host = card->host;
260 enum mmc_issue_type issue_type;
261 enum mmc_issued issued;
262 bool get_card, cqe_retune_ok;
263 int ret;
264
265 if (mmc_card_removed(mq->card)) {
266 req->rq_flags |= RQF_QUIET;
267 return BLK_STS_IOERR;
268 }
269
270 issue_type = mmc_issue_type(mq, req);
271
272 spin_lock_irq(&mq->lock);
273
274 if (mq->recovery_needed || mq->busy) {
275 spin_unlock_irq(&mq->lock);
276 return BLK_STS_RESOURCE;
277 }
278
279 switch (issue_type) {
280 case MMC_ISSUE_DCMD:
281 if (mmc_cqe_dcmd_busy(mq)) {
282 mq->cqe_busy |= MMC_CQE_DCMD_BUSY;
283 spin_unlock_irq(&mq->lock);
284 return BLK_STS_RESOURCE;
285 }
286 break;
287 case MMC_ISSUE_ASYNC:
288 /*
289 * For MMC host software queue, we only allow 2 requests in
290 * flight to avoid a long latency.
291 */
292 if (host->hsq_enabled && mq->in_flight[issue_type] > 2) {
293 spin_unlock_irq(&mq->lock);
294 return BLK_STS_RESOURCE;
295 }
296 break;
297 default:
298 /*
299 * Timeouts are handled by mmc core, and we don't have a host
300 * API to abort requests, so we can't handle the timeout anyway.
301 * However, when the timeout happens, blk_mq_complete_request()
302 * no longer works (to stop the request disappearing under us).
303 * To avoid racing with that, set a large timeout.
304 */
305 req->timeout = 600 * HZ;
306 break;
307 }
308
309 /* Parallel dispatch of requests is not supported at the moment */
310 mq->busy = true;
311
312 mq->in_flight[issue_type] += 1;
313 get_card = (mmc_tot_in_flight(mq) == 1);
314 cqe_retune_ok = (mmc_cqe_qcnt(mq) == 1);
315
316 spin_unlock_irq(&mq->lock);
317
318 if (!(req->rq_flags & RQF_DONTPREP)) {
319 req_to_mmc_queue_req(req)->retries = 0;
320 req->rq_flags |= RQF_DONTPREP;
321 }
322
323 if (get_card)
324 mmc_get_card(card, &mq->ctx);
325
326 if (mq->use_cqe) {
327 host->retune_now = host->need_retune && cqe_retune_ok &&
328 !host->hold_retune;
329 }
330
331 blk_mq_start_request(req);
332
333 issued = mmc_blk_mq_issue_rq(mq, req);
334
335 switch (issued) {
336 case MMC_REQ_BUSY:
337 ret = BLK_STS_RESOURCE;
338 break;
339 case MMC_REQ_FAILED_TO_START:
340 ret = BLK_STS_IOERR;
341 break;
342 default:
343 ret = BLK_STS_OK;
344 break;
345 }
346
347 if (issued != MMC_REQ_STARTED) {
348 bool put_card = false;
349
350 spin_lock_irq(&mq->lock);
351 mq->in_flight[issue_type] -= 1;
352 if (mmc_tot_in_flight(mq) == 0)
353 put_card = true;
354 mq->busy = false;
355 spin_unlock_irq(&mq->lock);
356 if (put_card)
357 mmc_put_card(card, &mq->ctx);
358 } else {
359 WRITE_ONCE(mq->busy, false);
360 }
361
362 return ret;
363 }
364
365 static const struct blk_mq_ops mmc_mq_ops = {
366 .queue_rq = mmc_mq_queue_rq,
367 .init_request = mmc_mq_init_request,
368 .exit_request = mmc_mq_exit_request,
369 .complete = mmc_blk_mq_complete,
370 .timeout = mmc_mq_timed_out,
371 };
372
373 static void mmc_setup_queue(struct mmc_queue *mq, struct mmc_card *card)
374 {
375 struct mmc_host *host = card->host;
376 unsigned block_size = 512;
377
378 blk_queue_flag_set(QUEUE_FLAG_NONROT, mq->queue);
379 blk_queue_flag_clear(QUEUE_FLAG_ADD_RANDOM, mq->queue);
380 if (mmc_can_erase(card))
381 mmc_queue_setup_discard(mq->queue, card);
382
383 if (!mmc_dev(host)->dma_mask || !*mmc_dev(host)->dma_mask)
384 blk_queue_bounce_limit(mq->queue, BLK_BOUNCE_HIGH);
385 blk_queue_max_hw_sectors(mq->queue,
386 min(host->max_blk_count, host->max_req_size / 512));
387 if (host->can_dma_map_merge)
388 WARN(!blk_queue_can_use_dma_map_merging(mq->queue,
389 mmc_dev(host)),
390 "merging was advertised but not possible");
391 blk_queue_max_segments(mq->queue, mmc_get_max_segments(host));
392
393 if (mmc_card_mmc(card))
394 block_size = card->ext_csd.data_sector_size;
395
396 blk_queue_logical_block_size(mq->queue, block_size);
397 /*
398 * After blk_queue_can_use_dma_map_merging() was called with succeed,
399 * since it calls blk_queue_virt_boundary(), the mmc should not call
400 * both blk_queue_max_segment_size().
401 */
402 if (!host->can_dma_map_merge)
403 blk_queue_max_segment_size(mq->queue,
404 round_down(host->max_seg_size, block_size));
405
406 dma_set_max_seg_size(mmc_dev(host), queue_max_segment_size(mq->queue));
407
408 INIT_WORK(&mq->recovery_work, mmc_mq_recovery_handler);
409 INIT_WORK(&mq->complete_work, mmc_blk_mq_complete_work);
410
411 mutex_init(&mq->complete_lock);
412
413 init_waitqueue_head(&mq->wait);
414 }
415
416 static 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 *
429 * Initialise a MMC card request queue.
430 */
431 int mmc_init_queue(struct mmc_queue *mq, struct mmc_card *card)
432 {
433 struct mmc_host *host = card->host;
434 int ret;
435
436 mq->card = card;
437 mq->use_cqe = host->cqe_enabled;
438
439 spin_lock_init(&mq->lock);
440
441 memset(&mq->tag_set, 0, sizeof(mq->tag_set));
442 mq->tag_set.ops = &mmc_mq_ops;
443 /*
444 * The queue depth for CQE must match the hardware because the request
445 * tag is used to index the hardware queue.
446 */
447 if (mq->use_cqe && !host->hsq_enabled)
448 mq->tag_set.queue_depth =
449 min_t(int, card->ext_csd.cmdq_depth, host->cqe_qdepth);
450 else
451 mq->tag_set.queue_depth = MMC_QUEUE_DEPTH;
452 mq->tag_set.numa_node = NUMA_NO_NODE;
453 mq->tag_set.flags = BLK_MQ_F_SHOULD_MERGE | BLK_MQ_F_BLOCKING;
454 mq->tag_set.nr_hw_queues = 1;
455 mq->tag_set.cmd_size = sizeof(struct mmc_queue_req);
456 mq->tag_set.driver_data = mq;
457
458 /*
459 * Since blk_mq_alloc_tag_set() calls .init_request() of mmc_mq_ops,
460 * the host->can_dma_map_merge should be set before to get max_segs
461 * from mmc_get_max_segments().
462 */
463 if (mmc_merge_capable(host) &&
464 host->max_segs < MMC_DMA_MAP_MERGE_SEGMENTS &&
465 dma_get_merge_boundary(mmc_dev(host)))
466 host->can_dma_map_merge = 1;
467 else
468 host->can_dma_map_merge = 0;
469
470 ret = blk_mq_alloc_tag_set(&mq->tag_set);
471 if (ret)
472 return ret;
473
474 mq->queue = blk_mq_init_queue(&mq->tag_set);
475 if (IS_ERR(mq->queue)) {
476 ret = PTR_ERR(mq->queue);
477 goto free_tag_set;
478 }
479
480 if (mmc_host_is_spi(host) && host->use_spi_crc)
481 mq->queue->backing_dev_info->capabilities |=
482 BDI_CAP_STABLE_WRITES;
483
484 mq->queue->queuedata = mq;
485 blk_queue_rq_timeout(mq->queue, 60 * HZ);
486
487 mmc_setup_queue(mq, card);
488 return 0;
489
490 free_tag_set:
491 blk_mq_free_tag_set(&mq->tag_set);
492 return ret;
493 }
494
495 void mmc_queue_suspend(struct mmc_queue *mq)
496 {
497 blk_mq_quiesce_queue(mq->queue);
498
499 /*
500 * The host remains claimed while there are outstanding requests, so
501 * simply claiming and releasing here ensures there are none.
502 */
503 mmc_claim_host(mq->card->host);
504 mmc_release_host(mq->card->host);
505 }
506
507 void mmc_queue_resume(struct mmc_queue *mq)
508 {
509 blk_mq_unquiesce_queue(mq->queue);
510 }
511
512 void mmc_cleanup_queue(struct mmc_queue *mq)
513 {
514 struct request_queue *q = mq->queue;
515
516 /*
517 * The legacy code handled the possibility of being suspended,
518 * so do that here too.
519 */
520 if (blk_queue_quiesced(q))
521 blk_mq_unquiesce_queue(q);
522
523 blk_cleanup_queue(q);
524 blk_mq_free_tag_set(&mq->tag_set);
525
526 /*
527 * A request can be completed before the next request, potentially
528 * leaving a complete_work with nothing to do. Such a work item might
529 * still be queued at this point. Flush it.
530 */
531 flush_work(&mq->complete_work);
532
533 mq->card = NULL;
534 }
535
536 /*
537 * Prepare the sg list(s) to be handed of to the host driver
538 */
539 unsigned int mmc_queue_map_sg(struct mmc_queue *mq, struct mmc_queue_req *mqrq)
540 {
541 struct request *req = mmc_queue_req_to_req(mqrq);
542
543 return blk_rq_map_sg(mq->queue, req, mqrq->sg);
544 }