dm-rq.c 22 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821822823824825826827828829830831832833834835836837838839840841842843844845846847848849850851852853854855
  1. /*
  2. * Copyright (C) 2016 Red Hat, Inc. All rights reserved.
  3. *
  4. * This file is released under the GPL.
  5. */
  6. #include "dm-core.h"
  7. #include "dm-rq.h"
  8. #include <linux/elevator.h> /* for rq_end_sector() */
  9. #include <linux/blk-mq.h>
  10. #define DM_MSG_PREFIX "core-rq"
  11. #define DM_MQ_NR_HW_QUEUES 1
  12. #define DM_MQ_QUEUE_DEPTH 2048
  13. static unsigned dm_mq_nr_hw_queues = DM_MQ_NR_HW_QUEUES;
  14. static unsigned dm_mq_queue_depth = DM_MQ_QUEUE_DEPTH;
  15. /*
  16. * Request-based DM's mempools' reserved IOs set by the user.
  17. */
  18. #define RESERVED_REQUEST_BASED_IOS 256
  19. static unsigned reserved_rq_based_ios = RESERVED_REQUEST_BASED_IOS;
  20. static bool use_blk_mq = IS_ENABLED(CONFIG_DM_MQ_DEFAULT);
  21. bool dm_use_blk_mq_default(void)
  22. {
  23. return use_blk_mq;
  24. }
  25. bool dm_use_blk_mq(struct mapped_device *md)
  26. {
  27. return md->use_blk_mq;
  28. }
  29. EXPORT_SYMBOL_GPL(dm_use_blk_mq);
  30. unsigned dm_get_reserved_rq_based_ios(void)
  31. {
  32. return __dm_get_module_param(&reserved_rq_based_ios,
  33. RESERVED_REQUEST_BASED_IOS, DM_RESERVED_MAX_IOS);
  34. }
  35. EXPORT_SYMBOL_GPL(dm_get_reserved_rq_based_ios);
  36. static unsigned dm_get_blk_mq_nr_hw_queues(void)
  37. {
  38. return __dm_get_module_param(&dm_mq_nr_hw_queues, 1, 32);
  39. }
  40. static unsigned dm_get_blk_mq_queue_depth(void)
  41. {
  42. return __dm_get_module_param(&dm_mq_queue_depth,
  43. DM_MQ_QUEUE_DEPTH, BLK_MQ_MAX_DEPTH);
  44. }
  45. int dm_request_based(struct mapped_device *md)
  46. {
  47. return queue_is_rq_based(md->queue);
  48. }
  49. static void dm_old_start_queue(struct request_queue *q)
  50. {
  51. unsigned long flags;
  52. spin_lock_irqsave(q->queue_lock, flags);
  53. if (blk_queue_stopped(q))
  54. blk_start_queue(q);
  55. spin_unlock_irqrestore(q->queue_lock, flags);
  56. }
  57. static void dm_mq_start_queue(struct request_queue *q)
  58. {
  59. blk_mq_unquiesce_queue(q);
  60. blk_mq_kick_requeue_list(q);
  61. }
  62. void dm_start_queue(struct request_queue *q)
  63. {
  64. if (!q->mq_ops)
  65. dm_old_start_queue(q);
  66. else
  67. dm_mq_start_queue(q);
  68. }
  69. static void dm_old_stop_queue(struct request_queue *q)
  70. {
  71. unsigned long flags;
  72. spin_lock_irqsave(q->queue_lock, flags);
  73. if (!blk_queue_stopped(q))
  74. blk_stop_queue(q);
  75. spin_unlock_irqrestore(q->queue_lock, flags);
  76. }
  77. static void dm_mq_stop_queue(struct request_queue *q)
  78. {
  79. if (blk_mq_queue_stopped(q))
  80. return;
  81. blk_mq_quiesce_queue(q);
  82. }
  83. void dm_stop_queue(struct request_queue *q)
  84. {
  85. if (!q->mq_ops)
  86. dm_old_stop_queue(q);
  87. else
  88. dm_mq_stop_queue(q);
  89. }
  90. /*
  91. * Partial completion handling for request-based dm
  92. */
  93. static void end_clone_bio(struct bio *clone)
  94. {
  95. struct dm_rq_clone_bio_info *info =
  96. container_of(clone, struct dm_rq_clone_bio_info, clone);
  97. struct dm_rq_target_io *tio = info->tio;
  98. unsigned int nr_bytes = info->orig->bi_iter.bi_size;
  99. blk_status_t error = clone->bi_status;
  100. bool is_last = !clone->bi_next;
  101. bio_put(clone);
  102. if (tio->error)
  103. /*
  104. * An error has already been detected on the request.
  105. * Once error occurred, just let clone->end_io() handle
  106. * the remainder.
  107. */
  108. return;
  109. else if (error) {
  110. /*
  111. * Don't notice the error to the upper layer yet.
  112. * The error handling decision is made by the target driver,
  113. * when the request is completed.
  114. */
  115. tio->error = error;
  116. goto exit;
  117. }
  118. /*
  119. * I/O for the bio successfully completed.
  120. * Notice the data completion to the upper layer.
  121. */
  122. tio->completed += nr_bytes;
  123. /*
  124. * Update the original request.
  125. * Do not use blk_end_request() here, because it may complete
  126. * the original request before the clone, and break the ordering.
  127. */
  128. if (is_last)
  129. exit:
  130. blk_update_request(tio->orig, BLK_STS_OK, tio->completed);
  131. }
  132. static struct dm_rq_target_io *tio_from_request(struct request *rq)
  133. {
  134. return blk_mq_rq_to_pdu(rq);
  135. }
  136. static void rq_end_stats(struct mapped_device *md, struct request *orig)
  137. {
  138. if (unlikely(dm_stats_used(&md->stats))) {
  139. struct dm_rq_target_io *tio = tio_from_request(orig);
  140. tio->duration_jiffies = jiffies - tio->duration_jiffies;
  141. dm_stats_account_io(&md->stats, rq_data_dir(orig),
  142. blk_rq_pos(orig), tio->n_sectors, true,
  143. tio->duration_jiffies, &tio->stats_aux);
  144. }
  145. }
  146. /*
  147. * Don't touch any member of the md after calling this function because
  148. * the md may be freed in dm_put() at the end of this function.
  149. * Or do dm_get() before calling this function and dm_put() later.
  150. */
  151. static void rq_completed(struct mapped_device *md, int rw, bool run_queue)
  152. {
  153. struct request_queue *q = md->queue;
  154. unsigned long flags;
  155. atomic_dec(&md->pending[rw]);
  156. /* nudge anyone waiting on suspend queue */
  157. if (!md_in_flight(md))
  158. wake_up(&md->wait);
  159. /*
  160. * Run this off this callpath, as drivers could invoke end_io while
  161. * inside their request_fn (and holding the queue lock). Calling
  162. * back into ->request_fn() could deadlock attempting to grab the
  163. * queue lock again.
  164. */
  165. if (!q->mq_ops && run_queue) {
  166. spin_lock_irqsave(q->queue_lock, flags);
  167. blk_run_queue_async(q);
  168. spin_unlock_irqrestore(q->queue_lock, flags);
  169. }
  170. /*
  171. * dm_put() must be at the end of this function. See the comment above
  172. */
  173. dm_put(md);
  174. }
  175. /*
  176. * Complete the clone and the original request.
  177. * Must be called without clone's queue lock held,
  178. * see end_clone_request() for more details.
  179. */
  180. static void dm_end_request(struct request *clone, blk_status_t error)
  181. {
  182. int rw = rq_data_dir(clone);
  183. struct dm_rq_target_io *tio = clone->end_io_data;
  184. struct mapped_device *md = tio->md;
  185. struct request *rq = tio->orig;
  186. blk_rq_unprep_clone(clone);
  187. tio->ti->type->release_clone_rq(clone);
  188. rq_end_stats(md, rq);
  189. if (!rq->q->mq_ops)
  190. blk_end_request_all(rq, error);
  191. else
  192. blk_mq_end_request(rq, error);
  193. rq_completed(md, rw, true);
  194. }
  195. /*
  196. * Requeue the original request of a clone.
  197. */
  198. static void dm_old_requeue_request(struct request *rq, unsigned long delay_ms)
  199. {
  200. struct request_queue *q = rq->q;
  201. unsigned long flags;
  202. spin_lock_irqsave(q->queue_lock, flags);
  203. blk_requeue_request(q, rq);
  204. blk_delay_queue(q, delay_ms);
  205. spin_unlock_irqrestore(q->queue_lock, flags);
  206. }
  207. static void __dm_mq_kick_requeue_list(struct request_queue *q, unsigned long msecs)
  208. {
  209. blk_mq_delay_kick_requeue_list(q, msecs);
  210. }
  211. void dm_mq_kick_requeue_list(struct mapped_device *md)
  212. {
  213. __dm_mq_kick_requeue_list(dm_get_md_queue(md), 0);
  214. }
  215. EXPORT_SYMBOL(dm_mq_kick_requeue_list);
  216. static void dm_mq_delay_requeue_request(struct request *rq, unsigned long msecs)
  217. {
  218. blk_mq_requeue_request(rq, false);
  219. __dm_mq_kick_requeue_list(rq->q, msecs);
  220. }
  221. static void dm_requeue_original_request(struct dm_rq_target_io *tio, bool delay_requeue)
  222. {
  223. struct mapped_device *md = tio->md;
  224. struct request *rq = tio->orig;
  225. int rw = rq_data_dir(rq);
  226. unsigned long delay_ms = delay_requeue ? 100 : 0;
  227. rq_end_stats(md, rq);
  228. if (tio->clone) {
  229. blk_rq_unprep_clone(tio->clone);
  230. tio->ti->type->release_clone_rq(tio->clone);
  231. }
  232. if (!rq->q->mq_ops)
  233. dm_old_requeue_request(rq, delay_ms);
  234. else
  235. dm_mq_delay_requeue_request(rq, delay_ms);
  236. rq_completed(md, rw, false);
  237. }
  238. static void dm_done(struct request *clone, blk_status_t error, bool mapped)
  239. {
  240. int r = DM_ENDIO_DONE;
  241. struct dm_rq_target_io *tio = clone->end_io_data;
  242. dm_request_endio_fn rq_end_io = NULL;
  243. if (tio->ti) {
  244. rq_end_io = tio->ti->type->rq_end_io;
  245. if (mapped && rq_end_io)
  246. r = rq_end_io(tio->ti, clone, error, &tio->info);
  247. }
  248. if (unlikely(error == BLK_STS_TARGET)) {
  249. if (req_op(clone) == REQ_OP_WRITE_SAME &&
  250. !clone->q->limits.max_write_same_sectors)
  251. disable_write_same(tio->md);
  252. if (req_op(clone) == REQ_OP_WRITE_ZEROES &&
  253. !clone->q->limits.max_write_zeroes_sectors)
  254. disable_write_zeroes(tio->md);
  255. }
  256. switch (r) {
  257. case DM_ENDIO_DONE:
  258. /* The target wants to complete the I/O */
  259. dm_end_request(clone, error);
  260. break;
  261. case DM_ENDIO_INCOMPLETE:
  262. /* The target will handle the I/O */
  263. return;
  264. case DM_ENDIO_REQUEUE:
  265. /* The target wants to requeue the I/O */
  266. dm_requeue_original_request(tio, false);
  267. break;
  268. case DM_ENDIO_DELAY_REQUEUE:
  269. /* The target wants to requeue the I/O after a delay */
  270. dm_requeue_original_request(tio, true);
  271. break;
  272. default:
  273. DMWARN("unimplemented target endio return value: %d", r);
  274. BUG();
  275. }
  276. }
  277. /*
  278. * Request completion handler for request-based dm
  279. */
  280. static void dm_softirq_done(struct request *rq)
  281. {
  282. bool mapped = true;
  283. struct dm_rq_target_io *tio = tio_from_request(rq);
  284. struct request *clone = tio->clone;
  285. int rw;
  286. if (!clone) {
  287. struct mapped_device *md = tio->md;
  288. rq_end_stats(md, rq);
  289. rw = rq_data_dir(rq);
  290. if (!rq->q->mq_ops)
  291. blk_end_request_all(rq, tio->error);
  292. else
  293. blk_mq_end_request(rq, tio->error);
  294. rq_completed(md, rw, false);
  295. return;
  296. }
  297. if (rq->rq_flags & RQF_FAILED)
  298. mapped = false;
  299. dm_done(clone, tio->error, mapped);
  300. }
  301. /*
  302. * Complete the clone and the original request with the error status
  303. * through softirq context.
  304. */
  305. static void dm_complete_request(struct request *rq, blk_status_t error)
  306. {
  307. struct dm_rq_target_io *tio = tio_from_request(rq);
  308. tio->error = error;
  309. if (!rq->q->mq_ops)
  310. blk_complete_request(rq);
  311. else
  312. blk_mq_complete_request(rq);
  313. }
  314. /*
  315. * Complete the not-mapped clone and the original request with the error status
  316. * through softirq context.
  317. * Target's rq_end_io() function isn't called.
  318. * This may be used when the target's map_rq() or clone_and_map_rq() functions fail.
  319. */
  320. static void dm_kill_unmapped_request(struct request *rq, blk_status_t error)
  321. {
  322. rq->rq_flags |= RQF_FAILED;
  323. dm_complete_request(rq, error);
  324. }
  325. /*
  326. * Called with the clone's queue lock held (in the case of .request_fn)
  327. */
  328. static void end_clone_request(struct request *clone, blk_status_t error)
  329. {
  330. struct dm_rq_target_io *tio = clone->end_io_data;
  331. /*
  332. * Actual request completion is done in a softirq context which doesn't
  333. * hold the clone's queue lock. Otherwise, deadlock could occur because:
  334. * - another request may be submitted by the upper level driver
  335. * of the stacking during the completion
  336. * - the submission which requires queue lock may be done
  337. * against this clone's queue
  338. */
  339. dm_complete_request(tio->orig, error);
  340. }
  341. static blk_status_t dm_dispatch_clone_request(struct request *clone, struct request *rq)
  342. {
  343. blk_status_t r;
  344. if (blk_queue_io_stat(clone->q))
  345. clone->rq_flags |= RQF_IO_STAT;
  346. clone->start_time = jiffies;
  347. r = blk_insert_cloned_request(clone->q, clone);
  348. if (r != BLK_STS_OK && r != BLK_STS_RESOURCE && r != BLK_STS_DEV_RESOURCE)
  349. /* must complete clone in terms of original request */
  350. dm_complete_request(rq, r);
  351. return r;
  352. }
  353. static int dm_rq_bio_constructor(struct bio *bio, struct bio *bio_orig,
  354. void *data)
  355. {
  356. struct dm_rq_target_io *tio = data;
  357. struct dm_rq_clone_bio_info *info =
  358. container_of(bio, struct dm_rq_clone_bio_info, clone);
  359. info->orig = bio_orig;
  360. info->tio = tio;
  361. bio->bi_end_io = end_clone_bio;
  362. return 0;
  363. }
  364. static int setup_clone(struct request *clone, struct request *rq,
  365. struct dm_rq_target_io *tio, gfp_t gfp_mask)
  366. {
  367. int r;
  368. r = blk_rq_prep_clone(clone, rq, tio->md->bs, gfp_mask,
  369. dm_rq_bio_constructor, tio);
  370. if (r)
  371. return r;
  372. clone->end_io = end_clone_request;
  373. clone->end_io_data = tio;
  374. tio->clone = clone;
  375. return 0;
  376. }
  377. static void map_tio_request(struct kthread_work *work);
  378. static void init_tio(struct dm_rq_target_io *tio, struct request *rq,
  379. struct mapped_device *md)
  380. {
  381. tio->md = md;
  382. tio->ti = NULL;
  383. tio->clone = NULL;
  384. tio->orig = rq;
  385. tio->error = 0;
  386. tio->completed = 0;
  387. /*
  388. * Avoid initializing info for blk-mq; it passes
  389. * target-specific data through info.ptr
  390. * (see: dm_mq_init_request)
  391. */
  392. if (!md->init_tio_pdu)
  393. memset(&tio->info, 0, sizeof(tio->info));
  394. if (md->kworker_task)
  395. kthread_init_work(&tio->work, map_tio_request);
  396. }
  397. /*
  398. * Returns:
  399. * DM_MAPIO_* : the request has been processed as indicated
  400. * DM_MAPIO_REQUEUE : the original request needs to be immediately requeued
  401. * < 0 : the request was completed due to failure
  402. */
  403. static int map_request(struct dm_rq_target_io *tio)
  404. {
  405. int r;
  406. struct dm_target *ti = tio->ti;
  407. struct mapped_device *md = tio->md;
  408. struct request *rq = tio->orig;
  409. struct request *clone = NULL;
  410. blk_status_t ret;
  411. r = ti->type->clone_and_map_rq(ti, rq, &tio->info, &clone);
  412. check_again:
  413. switch (r) {
  414. case DM_MAPIO_SUBMITTED:
  415. /* The target has taken the I/O to submit by itself later */
  416. break;
  417. case DM_MAPIO_REMAPPED:
  418. if (setup_clone(clone, rq, tio, GFP_ATOMIC)) {
  419. /* -ENOMEM */
  420. ti->type->release_clone_rq(clone);
  421. return DM_MAPIO_REQUEUE;
  422. }
  423. /* The target has remapped the I/O so dispatch it */
  424. trace_block_rq_remap(clone->q, clone, disk_devt(dm_disk(md)),
  425. blk_rq_pos(rq));
  426. ret = dm_dispatch_clone_request(clone, rq);
  427. if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) {
  428. blk_rq_unprep_clone(clone);
  429. tio->ti->type->release_clone_rq(clone);
  430. tio->clone = NULL;
  431. if (!rq->q->mq_ops)
  432. r = DM_MAPIO_DELAY_REQUEUE;
  433. else
  434. r = DM_MAPIO_REQUEUE;
  435. goto check_again;
  436. }
  437. break;
  438. case DM_MAPIO_REQUEUE:
  439. /* The target wants to requeue the I/O */
  440. break;
  441. case DM_MAPIO_DELAY_REQUEUE:
  442. /* The target wants to requeue the I/O after a delay */
  443. dm_requeue_original_request(tio, true);
  444. break;
  445. case DM_MAPIO_KILL:
  446. /* The target wants to complete the I/O */
  447. dm_kill_unmapped_request(rq, BLK_STS_IOERR);
  448. break;
  449. default:
  450. DMWARN("unimplemented target map return value: %d", r);
  451. BUG();
  452. }
  453. return r;
  454. }
  455. static void dm_start_request(struct mapped_device *md, struct request *orig)
  456. {
  457. if (!orig->q->mq_ops)
  458. blk_start_request(orig);
  459. else
  460. blk_mq_start_request(orig);
  461. atomic_inc(&md->pending[rq_data_dir(orig)]);
  462. if (md->seq_rq_merge_deadline_usecs) {
  463. md->last_rq_pos = rq_end_sector(orig);
  464. md->last_rq_rw = rq_data_dir(orig);
  465. md->last_rq_start_time = ktime_get();
  466. }
  467. if (unlikely(dm_stats_used(&md->stats))) {
  468. struct dm_rq_target_io *tio = tio_from_request(orig);
  469. tio->duration_jiffies = jiffies;
  470. tio->n_sectors = blk_rq_sectors(orig);
  471. dm_stats_account_io(&md->stats, rq_data_dir(orig),
  472. blk_rq_pos(orig), tio->n_sectors, false, 0,
  473. &tio->stats_aux);
  474. }
  475. /*
  476. * Hold the md reference here for the in-flight I/O.
  477. * We can't rely on the reference count by device opener,
  478. * because the device may be closed during the request completion
  479. * when all bios are completed.
  480. * See the comment in rq_completed() too.
  481. */
  482. dm_get(md);
  483. }
  484. static int __dm_rq_init_rq(struct mapped_device *md, struct request *rq)
  485. {
  486. struct dm_rq_target_io *tio = blk_mq_rq_to_pdu(rq);
  487. /*
  488. * Must initialize md member of tio, otherwise it won't
  489. * be available in dm_mq_queue_rq.
  490. */
  491. tio->md = md;
  492. if (md->init_tio_pdu) {
  493. /* target-specific per-io data is immediately after the tio */
  494. tio->info.ptr = tio + 1;
  495. }
  496. return 0;
  497. }
  498. static int dm_rq_init_rq(struct request_queue *q, struct request *rq, gfp_t gfp)
  499. {
  500. return __dm_rq_init_rq(q->rq_alloc_data, rq);
  501. }
  502. static void map_tio_request(struct kthread_work *work)
  503. {
  504. struct dm_rq_target_io *tio = container_of(work, struct dm_rq_target_io, work);
  505. if (map_request(tio) == DM_MAPIO_REQUEUE)
  506. dm_requeue_original_request(tio, false);
  507. }
  508. ssize_t dm_attr_rq_based_seq_io_merge_deadline_show(struct mapped_device *md, char *buf)
  509. {
  510. return sprintf(buf, "%u\n", md->seq_rq_merge_deadline_usecs);
  511. }
  512. #define MAX_SEQ_RQ_MERGE_DEADLINE_USECS 100000
  513. ssize_t dm_attr_rq_based_seq_io_merge_deadline_store(struct mapped_device *md,
  514. const char *buf, size_t count)
  515. {
  516. unsigned deadline;
  517. if (dm_get_md_type(md) != DM_TYPE_REQUEST_BASED)
  518. return count;
  519. if (kstrtouint(buf, 10, &deadline))
  520. return -EINVAL;
  521. if (deadline > MAX_SEQ_RQ_MERGE_DEADLINE_USECS)
  522. deadline = MAX_SEQ_RQ_MERGE_DEADLINE_USECS;
  523. md->seq_rq_merge_deadline_usecs = deadline;
  524. return count;
  525. }
  526. static bool dm_old_request_peeked_before_merge_deadline(struct mapped_device *md)
  527. {
  528. ktime_t kt_deadline;
  529. if (!md->seq_rq_merge_deadline_usecs)
  530. return false;
  531. kt_deadline = ns_to_ktime((u64)md->seq_rq_merge_deadline_usecs * NSEC_PER_USEC);
  532. kt_deadline = ktime_add_safe(md->last_rq_start_time, kt_deadline);
  533. return !ktime_after(ktime_get(), kt_deadline);
  534. }
  535. /*
  536. * q->request_fn for old request-based dm.
  537. * Called with the queue lock held.
  538. */
  539. static void dm_old_request_fn(struct request_queue *q)
  540. {
  541. struct mapped_device *md = q->queuedata;
  542. struct dm_target *ti = md->immutable_target;
  543. struct request *rq;
  544. struct dm_rq_target_io *tio;
  545. sector_t pos = 0;
  546. if (unlikely(!ti)) {
  547. int srcu_idx;
  548. struct dm_table *map = dm_get_live_table(md, &srcu_idx);
  549. if (unlikely(!map)) {
  550. dm_put_live_table(md, srcu_idx);
  551. return;
  552. }
  553. ti = dm_table_find_target(map, pos);
  554. dm_put_live_table(md, srcu_idx);
  555. }
  556. /*
  557. * For suspend, check blk_queue_stopped() and increment
  558. * ->pending within a single queue_lock not to increment the
  559. * number of in-flight I/Os after the queue is stopped in
  560. * dm_suspend().
  561. */
  562. while (!blk_queue_stopped(q)) {
  563. rq = blk_peek_request(q);
  564. if (!rq)
  565. return;
  566. /* always use block 0 to find the target for flushes for now */
  567. pos = 0;
  568. if (req_op(rq) != REQ_OP_FLUSH)
  569. pos = blk_rq_pos(rq);
  570. if ((dm_old_request_peeked_before_merge_deadline(md) &&
  571. md_in_flight(md) && rq->bio && !bio_multiple_segments(rq->bio) &&
  572. md->last_rq_pos == pos && md->last_rq_rw == rq_data_dir(rq)) ||
  573. (ti->type->busy && ti->type->busy(ti))) {
  574. blk_delay_queue(q, 10);
  575. return;
  576. }
  577. dm_start_request(md, rq);
  578. tio = tio_from_request(rq);
  579. init_tio(tio, rq, md);
  580. /* Establish tio->ti before queuing work (map_tio_request) */
  581. tio->ti = ti;
  582. kthread_queue_work(&md->kworker, &tio->work);
  583. BUG_ON(!irqs_disabled());
  584. }
  585. }
  586. /*
  587. * Fully initialize a .request_fn request-based queue.
  588. */
  589. int dm_old_init_request_queue(struct mapped_device *md, struct dm_table *t)
  590. {
  591. struct dm_target *immutable_tgt;
  592. /* Fully initialize the queue */
  593. md->queue->cmd_size = sizeof(struct dm_rq_target_io);
  594. md->queue->rq_alloc_data = md;
  595. md->queue->request_fn = dm_old_request_fn;
  596. md->queue->init_rq_fn = dm_rq_init_rq;
  597. immutable_tgt = dm_table_get_immutable_target(t);
  598. if (immutable_tgt && immutable_tgt->per_io_data_size) {
  599. /* any target-specific per-io data is immediately after the tio */
  600. md->queue->cmd_size += immutable_tgt->per_io_data_size;
  601. md->init_tio_pdu = true;
  602. }
  603. if (blk_init_allocated_queue(md->queue) < 0)
  604. return -EINVAL;
  605. /* disable dm_old_request_fn's merge heuristic by default */
  606. md->seq_rq_merge_deadline_usecs = 0;
  607. blk_queue_softirq_done(md->queue, dm_softirq_done);
  608. /* Initialize the request-based DM worker thread */
  609. kthread_init_worker(&md->kworker);
  610. md->kworker_task = kthread_run(kthread_worker_fn, &md->kworker,
  611. "kdmwork-%s", dm_device_name(md));
  612. if (IS_ERR(md->kworker_task)) {
  613. int error = PTR_ERR(md->kworker_task);
  614. md->kworker_task = NULL;
  615. return error;
  616. }
  617. return 0;
  618. }
  619. static int dm_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
  620. unsigned int hctx_idx, unsigned int numa_node)
  621. {
  622. return __dm_rq_init_rq(set->driver_data, rq);
  623. }
  624. static blk_status_t dm_mq_queue_rq(struct blk_mq_hw_ctx *hctx,
  625. const struct blk_mq_queue_data *bd)
  626. {
  627. struct request *rq = bd->rq;
  628. struct dm_rq_target_io *tio = blk_mq_rq_to_pdu(rq);
  629. struct mapped_device *md = tio->md;
  630. struct dm_target *ti = md->immutable_target;
  631. if (unlikely(!ti)) {
  632. int srcu_idx;
  633. struct dm_table *map = dm_get_live_table(md, &srcu_idx);
  634. ti = dm_table_find_target(map, 0);
  635. dm_put_live_table(md, srcu_idx);
  636. }
  637. if (ti->type->busy && ti->type->busy(ti))
  638. return BLK_STS_RESOURCE;
  639. dm_start_request(md, rq);
  640. /* Init tio using md established in .init_request */
  641. init_tio(tio, rq, md);
  642. /*
  643. * Establish tio->ti before calling map_request().
  644. */
  645. tio->ti = ti;
  646. /* Direct call is fine since .queue_rq allows allocations */
  647. if (map_request(tio) == DM_MAPIO_REQUEUE) {
  648. /* Undo dm_start_request() before requeuing */
  649. rq_end_stats(md, rq);
  650. rq_completed(md, rq_data_dir(rq), false);
  651. return BLK_STS_RESOURCE;
  652. }
  653. return BLK_STS_OK;
  654. }
  655. static const struct blk_mq_ops dm_mq_ops = {
  656. .queue_rq = dm_mq_queue_rq,
  657. .complete = dm_softirq_done,
  658. .init_request = dm_mq_init_request,
  659. };
  660. int dm_mq_init_request_queue(struct mapped_device *md, struct dm_table *t)
  661. {
  662. struct request_queue *q;
  663. struct dm_target *immutable_tgt;
  664. int err;
  665. if (!dm_table_all_blk_mq_devices(t)) {
  666. DMERR("request-based dm-mq may only be stacked on blk-mq device(s)");
  667. return -EINVAL;
  668. }
  669. md->tag_set = kzalloc_node(sizeof(struct blk_mq_tag_set), GFP_KERNEL, md->numa_node_id);
  670. if (!md->tag_set)
  671. return -ENOMEM;
  672. md->tag_set->ops = &dm_mq_ops;
  673. md->tag_set->queue_depth = dm_get_blk_mq_queue_depth();
  674. md->tag_set->numa_node = md->numa_node_id;
  675. md->tag_set->flags = BLK_MQ_F_SHOULD_MERGE | BLK_MQ_F_SG_MERGE;
  676. md->tag_set->nr_hw_queues = dm_get_blk_mq_nr_hw_queues();
  677. md->tag_set->driver_data = md;
  678. md->tag_set->cmd_size = sizeof(struct dm_rq_target_io);
  679. immutable_tgt = dm_table_get_immutable_target(t);
  680. if (immutable_tgt && immutable_tgt->per_io_data_size) {
  681. /* any target-specific per-io data is immediately after the tio */
  682. md->tag_set->cmd_size += immutable_tgt->per_io_data_size;
  683. md->init_tio_pdu = true;
  684. }
  685. err = blk_mq_alloc_tag_set(md->tag_set);
  686. if (err)
  687. goto out_kfree_tag_set;
  688. q = blk_mq_init_allocated_queue(md->tag_set, md->queue);
  689. if (IS_ERR(q)) {
  690. err = PTR_ERR(q);
  691. goto out_tag_set;
  692. }
  693. return 0;
  694. out_tag_set:
  695. blk_mq_free_tag_set(md->tag_set);
  696. out_kfree_tag_set:
  697. kfree(md->tag_set);
  698. return err;
  699. }
  700. void dm_mq_cleanup_mapped_device(struct mapped_device *md)
  701. {
  702. if (md->tag_set) {
  703. blk_mq_free_tag_set(md->tag_set);
  704. kfree(md->tag_set);
  705. }
  706. }
  707. module_param(reserved_rq_based_ios, uint, S_IRUGO | S_IWUSR);
  708. MODULE_PARM_DESC(reserved_rq_based_ios, "Reserved IOs in request-based mempools");
  709. module_param(use_blk_mq, bool, S_IRUGO | S_IWUSR);
  710. MODULE_PARM_DESC(use_blk_mq, "Use block multiqueue for request-based DM devices");
  711. module_param(dm_mq_nr_hw_queues, uint, S_IRUGO | S_IWUSR);
  712. MODULE_PARM_DESC(dm_mq_nr_hw_queues, "Number of hardware queues for request-based dm-mq devices");
  713. module_param(dm_mq_queue_depth, uint, S_IRUGO | S_IWUSR);
  714. MODULE_PARM_DESC(dm_mq_queue_depth, "Queue depth for request-based dm-mq devices");