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 | |
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 | |
38 | static inline bool mmc_cqe_can_dcmd(struct mmc_host *host) |
39 | { |
40 | return host->caps2 & MMC_CAP2_CQE_DCMD; |
41 | } |
42 | |
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 | |
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 | |
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(work: &mq->recovery_work); |
78 | } |
79 | } |
80 | |
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(mqr: 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(lock: &mq->lock, flags); |
93 | } |
94 | |
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(rq: 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 | |
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(lock: &mq->lock, flags); |
132 | |
133 | return ignore_tout ? BLK_EH_RESET_TIMER : mmc_cqe_timed_out(req); |
134 | } |
135 | |
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(card: mq->card, ctx: &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(lock: &mq->lock); |
155 | mq->recovery_needed = false; |
156 | spin_unlock_irq(lock: &mq->lock); |
157 | |
158 | if (host->hsq_enabled) |
159 | host->cqe_ops->cqe_recovery_finish(host); |
160 | |
161 | mmc_put_card(card: mq->card, ctx: &mq->ctx); |
162 | |
163 | blk_mq_run_hw_queues(q, async: true); |
164 | } |
165 | |
166 | static struct scatterlist *mmc_alloc_sg(unsigned short sg_len, gfp_t gfp) |
167 | { |
168 | struct scatterlist *sg; |
169 | |
170 | sg = kmalloc_array(n: sg_len, size: sizeof(*sg), flags: gfp); |
171 | if (sg) |
172 | sg_init_table(sg, sg_len); |
173 | |
174 | return sg; |
175 | } |
176 | |
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_sectors: 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_sectors: max_discard); |
193 | if (mmc_can_trim(card) && card->erased_byte == 0) |
194 | blk_queue_max_write_zeroes_sectors(q, max_write_same_sectors: max_discard); |
195 | } |
196 | |
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 | |
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(rq: 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(sg_len: mmc_get_max_segments(host), GFP_KERNEL); |
212 | if (!mq_rq->sg) |
213 | return -ENOMEM; |
214 | |
215 | return 0; |
216 | } |
217 | |
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(rq: req); |
222 | |
223 | kfree(objp: mq_rq->sg); |
224 | mq_rq->sg = NULL; |
225 | } |
226 | |
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(lock: &mq->lock); |
248 | |
249 | if (mq->recovery_needed || mq->busy) { |
250 | spin_unlock_irq(lock: &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(lock: &mq->lock); |
259 | return BLK_STS_RESOURCE; |
260 | } |
261 | break; |
262 | case MMC_ISSUE_ASYNC: |
263 | if (host->hsq_enabled && mq->in_flight[issue_type] > host->hsq_depth) { |
264 | spin_unlock_irq(lock: &mq->lock); |
265 | return BLK_STS_RESOURCE; |
266 | } |
267 | break; |
268 | default: |
269 | /* |
270 | * Timeouts are handled by mmc core, and we don't have a host |
271 | * API to abort requests, so we can't handle the timeout anyway. |
272 | * However, when the timeout happens, blk_mq_complete_request() |
273 | * no longer works (to stop the request disappearing under us). |
274 | * To avoid racing with that, set a large timeout. |
275 | */ |
276 | req->timeout = 600 * HZ; |
277 | break; |
278 | } |
279 | |
280 | /* Parallel dispatch of requests is not supported at the moment */ |
281 | mq->busy = true; |
282 | |
283 | mq->in_flight[issue_type] += 1; |
284 | get_card = (mmc_tot_in_flight(mq) == 1); |
285 | cqe_retune_ok = (mmc_cqe_qcnt(mq) == 1); |
286 | |
287 | spin_unlock_irq(lock: &mq->lock); |
288 | |
289 | if (!(req->rq_flags & RQF_DONTPREP)) { |
290 | req_to_mmc_queue_req(rq: req)->retries = 0; |
291 | req->rq_flags |= RQF_DONTPREP; |
292 | } |
293 | |
294 | if (get_card) |
295 | mmc_get_card(card, ctx: &mq->ctx); |
296 | |
297 | if (host->cqe_enabled) { |
298 | host->retune_now = host->need_retune && cqe_retune_ok && |
299 | !host->hold_retune; |
300 | } |
301 | |
302 | blk_mq_start_request(rq: req); |
303 | |
304 | issued = mmc_blk_mq_issue_rq(mq, req); |
305 | |
306 | switch (issued) { |
307 | case MMC_REQ_BUSY: |
308 | ret = BLK_STS_RESOURCE; |
309 | break; |
310 | case MMC_REQ_FAILED_TO_START: |
311 | ret = BLK_STS_IOERR; |
312 | break; |
313 | default: |
314 | ret = BLK_STS_OK; |
315 | break; |
316 | } |
317 | |
318 | if (issued != MMC_REQ_STARTED) { |
319 | bool put_card = false; |
320 | |
321 | spin_lock_irq(lock: &mq->lock); |
322 | mq->in_flight[issue_type] -= 1; |
323 | if (mmc_tot_in_flight(mq) == 0) |
324 | put_card = true; |
325 | mq->busy = false; |
326 | spin_unlock_irq(lock: &mq->lock); |
327 | if (put_card) |
328 | mmc_put_card(card, ctx: &mq->ctx); |
329 | } else { |
330 | WRITE_ONCE(mq->busy, false); |
331 | } |
332 | |
333 | return ret; |
334 | } |
335 | |
336 | static const struct blk_mq_ops mmc_mq_ops = { |
337 | .queue_rq = mmc_mq_queue_rq, |
338 | .init_request = mmc_mq_init_request, |
339 | .exit_request = mmc_mq_exit_request, |
340 | .complete = mmc_blk_mq_complete, |
341 | .timeout = mmc_mq_timed_out, |
342 | }; |
343 | |
344 | static void mmc_setup_queue(struct mmc_queue *mq, struct mmc_card *card) |
345 | { |
346 | struct mmc_host *host = card->host; |
347 | unsigned block_size = 512; |
348 | |
349 | blk_queue_flag_set(QUEUE_FLAG_NONROT, q: mq->queue); |
350 | blk_queue_flag_clear(QUEUE_FLAG_ADD_RANDOM, q: mq->queue); |
351 | if (mmc_can_erase(card)) |
352 | mmc_queue_setup_discard(q: mq->queue, card); |
353 | |
354 | if (!mmc_dev(host)->dma_mask || !*mmc_dev(host)->dma_mask) |
355 | blk_queue_bounce_limit(q: mq->queue, limit: BLK_BOUNCE_HIGH); |
356 | blk_queue_max_hw_sectors(mq->queue, |
357 | min(host->max_blk_count, host->max_req_size / 512)); |
358 | if (host->can_dma_map_merge) |
359 | WARN(!blk_queue_can_use_dma_map_merging(mq->queue, |
360 | mmc_dev(host)), |
361 | "merging was advertised but not possible" ); |
362 | blk_queue_max_segments(mq->queue, mmc_get_max_segments(host)); |
363 | |
364 | if (mmc_card_mmc(card) && card->ext_csd.data_sector_size) { |
365 | block_size = card->ext_csd.data_sector_size; |
366 | WARN_ON(block_size != 512 && block_size != 4096); |
367 | } |
368 | |
369 | blk_queue_logical_block_size(mq->queue, block_size); |
370 | /* |
371 | * After blk_queue_can_use_dma_map_merging() was called with succeed, |
372 | * since it calls blk_queue_virt_boundary(), the mmc should not call |
373 | * both blk_queue_max_segment_size(). |
374 | */ |
375 | if (!host->can_dma_map_merge) |
376 | blk_queue_max_segment_size(mq->queue, |
377 | round_down(host->max_seg_size, block_size)); |
378 | |
379 | dma_set_max_seg_size(mmc_dev(host), size: queue_max_segment_size(q: mq->queue)); |
380 | |
381 | INIT_WORK(&mq->recovery_work, mmc_mq_recovery_handler); |
382 | INIT_WORK(&mq->complete_work, mmc_blk_mq_complete_work); |
383 | |
384 | mutex_init(&mq->complete_lock); |
385 | |
386 | init_waitqueue_head(&mq->wait); |
387 | |
388 | mmc_crypto_setup_queue(q: mq->queue, host); |
389 | } |
390 | |
391 | static inline bool mmc_merge_capable(struct mmc_host *host) |
392 | { |
393 | return host->caps2 & MMC_CAP2_MERGE_CAPABLE; |
394 | } |
395 | |
396 | /* Set queue depth to get a reasonable value for q->nr_requests */ |
397 | #define MMC_QUEUE_DEPTH 64 |
398 | |
399 | /** |
400 | * mmc_init_queue - initialise a queue structure. |
401 | * @mq: mmc queue |
402 | * @card: mmc card to attach this queue |
403 | * |
404 | * Initialise a MMC card request queue. |
405 | */ |
406 | struct gendisk *mmc_init_queue(struct mmc_queue *mq, struct mmc_card *card) |
407 | { |
408 | struct mmc_host *host = card->host; |
409 | struct gendisk *disk; |
410 | int ret; |
411 | |
412 | mq->card = card; |
413 | |
414 | spin_lock_init(&mq->lock); |
415 | |
416 | memset(&mq->tag_set, 0, sizeof(mq->tag_set)); |
417 | mq->tag_set.ops = &mmc_mq_ops; |
418 | /* |
419 | * The queue depth for CQE must match the hardware because the request |
420 | * tag is used to index the hardware queue. |
421 | */ |
422 | if (host->cqe_enabled && !host->hsq_enabled) |
423 | mq->tag_set.queue_depth = |
424 | min_t(int, card->ext_csd.cmdq_depth, host->cqe_qdepth); |
425 | else |
426 | mq->tag_set.queue_depth = MMC_QUEUE_DEPTH; |
427 | mq->tag_set.numa_node = NUMA_NO_NODE; |
428 | mq->tag_set.flags = BLK_MQ_F_SHOULD_MERGE | BLK_MQ_F_BLOCKING; |
429 | mq->tag_set.nr_hw_queues = 1; |
430 | mq->tag_set.cmd_size = sizeof(struct mmc_queue_req); |
431 | mq->tag_set.driver_data = mq; |
432 | |
433 | /* |
434 | * Since blk_mq_alloc_tag_set() calls .init_request() of mmc_mq_ops, |
435 | * the host->can_dma_map_merge should be set before to get max_segs |
436 | * from mmc_get_max_segments(). |
437 | */ |
438 | if (mmc_merge_capable(host) && |
439 | host->max_segs < MMC_DMA_MAP_MERGE_SEGMENTS && |
440 | dma_get_merge_boundary(mmc_dev(host))) |
441 | host->can_dma_map_merge = 1; |
442 | else |
443 | host->can_dma_map_merge = 0; |
444 | |
445 | ret = blk_mq_alloc_tag_set(set: &mq->tag_set); |
446 | if (ret) |
447 | return ERR_PTR(error: ret); |
448 | |
449 | |
450 | disk = blk_mq_alloc_disk(&mq->tag_set, mq); |
451 | if (IS_ERR(ptr: disk)) { |
452 | blk_mq_free_tag_set(set: &mq->tag_set); |
453 | return disk; |
454 | } |
455 | mq->queue = disk->queue; |
456 | |
457 | if (mmc_host_is_spi(host) && host->use_spi_crc) |
458 | blk_queue_flag_set(QUEUE_FLAG_STABLE_WRITES, q: mq->queue); |
459 | blk_queue_rq_timeout(mq->queue, 60 * HZ); |
460 | |
461 | mmc_setup_queue(mq, card); |
462 | return disk; |
463 | } |
464 | |
465 | void mmc_queue_suspend(struct mmc_queue *mq) |
466 | { |
467 | blk_mq_quiesce_queue(q: 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(host: mq->card->host); |
474 | mmc_release_host(host: mq->card->host); |
475 | } |
476 | |
477 | void mmc_queue_resume(struct mmc_queue *mq) |
478 | { |
479 | blk_mq_unquiesce_queue(q: mq->queue); |
480 | } |
481 | |
482 | void 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 | /* |
494 | * If the recovery completes the last (and only remaining) request in |
495 | * the queue, and the card has been removed, we could end up here with |
496 | * the recovery not quite finished yet, so cancel it. |
497 | */ |
498 | cancel_work_sync(work: &mq->recovery_work); |
499 | |
500 | blk_mq_free_tag_set(set: &mq->tag_set); |
501 | |
502 | /* |
503 | * A request can be completed before the next request, potentially |
504 | * leaving a complete_work with nothing to do. Such a work item might |
505 | * still be queued at this point. Flush it. |
506 | */ |
507 | flush_work(work: &mq->complete_work); |
508 | |
509 | mq->card = NULL; |
510 | } |
511 | |
512 | /* |
513 | * Prepare the sg list(s) to be handed of to the host driver |
514 | */ |
515 | unsigned int mmc_queue_map_sg(struct mmc_queue *mq, struct mmc_queue_req *mqrq) |
516 | { |
517 | struct request *req = mmc_queue_req_to_req(mqr: mqrq); |
518 | |
519 | return blk_rq_map_sg(q: mq->queue, rq: req, sglist: mqrq->sg); |
520 | } |
521 | |