1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * Tag allocation using scalable bitmaps. Uses active queue tracking to support 4 * fairer distribution of tags between multiple submitters when a shared tag map 5 * is used. 6 * 7 * Copyright (C) 2013-2014 Jens Axboe 8 */ 9 #include <linux/kernel.h> 10 #include <linux/module.h> 11 12 #include <linux/blk-mq.h> 13 #include <linux/delay.h> 14 #include "blk.h" 15 #include "blk-mq.h" 16 #include "blk-mq-tag.h" 17 18 bool blk_mq_has_free_tags(struct blk_mq_tags *tags) 19 { 20 if (!tags) 21 return true; 22 23 return sbitmap_any_bit_clear(&tags->bitmap_tags.sb); 24 } 25 26 /* 27 * If a previously inactive queue goes active, bump the active user count. 28 * We need to do this before try to allocate driver tag, then even if fail 29 * to get tag when first time, the other shared-tag users could reserve 30 * budget for it. 31 */ 32 bool __blk_mq_tag_busy(struct blk_mq_hw_ctx *hctx) 33 { 34 if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state) && 35 !test_and_set_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state)) 36 atomic_inc(&hctx->tags->active_queues); 37 38 return true; 39 } 40 41 /* 42 * Wakeup all potentially sleeping on tags 43 */ 44 void blk_mq_tag_wakeup_all(struct blk_mq_tags *tags, bool include_reserve) 45 { 46 sbitmap_queue_wake_all(&tags->bitmap_tags); 47 if (include_reserve) 48 sbitmap_queue_wake_all(&tags->breserved_tags); 49 } 50 51 /* 52 * If a previously busy queue goes inactive, potential waiters could now 53 * be allowed to queue. Wake them up and check. 54 */ 55 void __blk_mq_tag_idle(struct blk_mq_hw_ctx *hctx) 56 { 57 struct blk_mq_tags *tags = hctx->tags; 58 59 if (!test_and_clear_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state)) 60 return; 61 62 atomic_dec(&tags->active_queues); 63 64 blk_mq_tag_wakeup_all(tags, false); 65 } 66 67 /* 68 * For shared tag users, we track the number of currently active users 69 * and attempt to provide a fair share of the tag depth for each of them. 70 */ 71 static inline bool hctx_may_queue(struct blk_mq_hw_ctx *hctx, 72 struct sbitmap_queue *bt) 73 { 74 unsigned int depth, users; 75 76 if (!hctx || !(hctx->flags & BLK_MQ_F_TAG_SHARED)) 77 return true; 78 if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state)) 79 return true; 80 81 /* 82 * Don't try dividing an ant 83 */ 84 if (bt->sb.depth == 1) 85 return true; 86 87 users = atomic_read(&hctx->tags->active_queues); 88 if (!users) 89 return true; 90 91 /* 92 * Allow at least some tags 93 */ 94 depth = max((bt->sb.depth + users - 1) / users, 4U); 95 return atomic_read(&hctx->nr_active) < depth; 96 } 97 98 static int __blk_mq_get_tag(struct blk_mq_alloc_data *data, 99 struct sbitmap_queue *bt) 100 { 101 if (!(data->flags & BLK_MQ_REQ_INTERNAL) && 102 !hctx_may_queue(data->hctx, bt)) 103 return -1; 104 if (data->shallow_depth) 105 return __sbitmap_queue_get_shallow(bt, data->shallow_depth); 106 else 107 return __sbitmap_queue_get(bt); 108 } 109 110 unsigned int blk_mq_get_tag(struct blk_mq_alloc_data *data) 111 { 112 struct blk_mq_tags *tags = blk_mq_tags_from_data(data); 113 struct sbitmap_queue *bt; 114 struct sbq_wait_state *ws; 115 DEFINE_SBQ_WAIT(wait); 116 unsigned int tag_offset; 117 int tag; 118 119 if (data->flags & BLK_MQ_REQ_RESERVED) { 120 if (unlikely(!tags->nr_reserved_tags)) { 121 WARN_ON_ONCE(1); 122 return BLK_MQ_TAG_FAIL; 123 } 124 bt = &tags->breserved_tags; 125 tag_offset = 0; 126 } else { 127 bt = &tags->bitmap_tags; 128 tag_offset = tags->nr_reserved_tags; 129 } 130 131 tag = __blk_mq_get_tag(data, bt); 132 if (tag != -1) 133 goto found_tag; 134 135 if (data->flags & BLK_MQ_REQ_NOWAIT) 136 return BLK_MQ_TAG_FAIL; 137 138 ws = bt_wait_ptr(bt, data->hctx); 139 do { 140 struct sbitmap_queue *bt_prev; 141 142 /* 143 * We're out of tags on this hardware queue, kick any 144 * pending IO submits before going to sleep waiting for 145 * some to complete. 146 */ 147 blk_mq_run_hw_queue(data->hctx, false); 148 149 /* 150 * Retry tag allocation after running the hardware queue, 151 * as running the queue may also have found completions. 152 */ 153 tag = __blk_mq_get_tag(data, bt); 154 if (tag != -1) 155 break; 156 157 sbitmap_prepare_to_wait(bt, ws, &wait, TASK_UNINTERRUPTIBLE); 158 159 tag = __blk_mq_get_tag(data, bt); 160 if (tag != -1) 161 break; 162 163 bt_prev = bt; 164 io_schedule(); 165 166 sbitmap_finish_wait(bt, ws, &wait); 167 168 data->ctx = blk_mq_get_ctx(data->q); 169 data->hctx = blk_mq_map_queue(data->q, data->cmd_flags, 170 data->ctx); 171 tags = blk_mq_tags_from_data(data); 172 if (data->flags & BLK_MQ_REQ_RESERVED) 173 bt = &tags->breserved_tags; 174 else 175 bt = &tags->bitmap_tags; 176 177 /* 178 * If destination hw queue is changed, fake wake up on 179 * previous queue for compensating the wake up miss, so 180 * other allocations on previous queue won't be starved. 181 */ 182 if (bt != bt_prev) 183 sbitmap_queue_wake_up(bt_prev); 184 185 ws = bt_wait_ptr(bt, data->hctx); 186 } while (1); 187 188 sbitmap_finish_wait(bt, ws, &wait); 189 190 found_tag: 191 return tag + tag_offset; 192 } 193 194 void blk_mq_put_tag(struct blk_mq_hw_ctx *hctx, struct blk_mq_tags *tags, 195 struct blk_mq_ctx *ctx, unsigned int tag) 196 { 197 if (!blk_mq_tag_is_reserved(tags, tag)) { 198 const int real_tag = tag - tags->nr_reserved_tags; 199 200 BUG_ON(real_tag >= tags->nr_tags); 201 sbitmap_queue_clear(&tags->bitmap_tags, real_tag, ctx->cpu); 202 } else { 203 BUG_ON(tag >= tags->nr_reserved_tags); 204 sbitmap_queue_clear(&tags->breserved_tags, tag, ctx->cpu); 205 } 206 } 207 208 struct bt_iter_data { 209 struct blk_mq_hw_ctx *hctx; 210 busy_iter_fn *fn; 211 void *data; 212 bool reserved; 213 }; 214 215 static bool bt_iter(struct sbitmap *bitmap, unsigned int bitnr, void *data) 216 { 217 struct bt_iter_data *iter_data = data; 218 struct blk_mq_hw_ctx *hctx = iter_data->hctx; 219 struct blk_mq_tags *tags = hctx->tags; 220 bool reserved = iter_data->reserved; 221 struct request *rq; 222 223 if (!reserved) 224 bitnr += tags->nr_reserved_tags; 225 rq = tags->rqs[bitnr]; 226 227 /* 228 * We can hit rq == NULL here, because the tagging functions 229 * test and set the bit before assigning ->rqs[]. 230 */ 231 if (rq && rq->q == hctx->queue) 232 return iter_data->fn(hctx, rq, iter_data->data, reserved); 233 return true; 234 } 235 236 /** 237 * bt_for_each - iterate over the requests associated with a hardware queue 238 * @hctx: Hardware queue to examine. 239 * @bt: sbitmap to examine. This is either the breserved_tags member 240 * or the bitmap_tags member of struct blk_mq_tags. 241 * @fn: Pointer to the function that will be called for each request 242 * associated with @hctx that has been assigned a driver tag. 243 * @fn will be called as follows: @fn(@hctx, rq, @data, @reserved) 244 * where rq is a pointer to a request. Return true to continue 245 * iterating tags, false to stop. 246 * @data: Will be passed as third argument to @fn. 247 * @reserved: Indicates whether @bt is the breserved_tags member or the 248 * bitmap_tags member of struct blk_mq_tags. 249 */ 250 static void bt_for_each(struct blk_mq_hw_ctx *hctx, struct sbitmap_queue *bt, 251 busy_iter_fn *fn, void *data, bool reserved) 252 { 253 struct bt_iter_data iter_data = { 254 .hctx = hctx, 255 .fn = fn, 256 .data = data, 257 .reserved = reserved, 258 }; 259 260 sbitmap_for_each_set(&bt->sb, bt_iter, &iter_data); 261 } 262 263 struct bt_tags_iter_data { 264 struct blk_mq_tags *tags; 265 busy_tag_iter_fn *fn; 266 void *data; 267 bool reserved; 268 }; 269 270 static bool bt_tags_iter(struct sbitmap *bitmap, unsigned int bitnr, void *data) 271 { 272 struct bt_tags_iter_data *iter_data = data; 273 struct blk_mq_tags *tags = iter_data->tags; 274 bool reserved = iter_data->reserved; 275 struct request *rq; 276 277 if (!reserved) 278 bitnr += tags->nr_reserved_tags; 279 280 /* 281 * We can hit rq == NULL here, because the tagging functions 282 * test and set the bit before assining ->rqs[]. 283 */ 284 rq = tags->rqs[bitnr]; 285 if (rq && blk_mq_request_started(rq)) 286 return iter_data->fn(rq, iter_data->data, reserved); 287 288 return true; 289 } 290 291 /** 292 * bt_tags_for_each - iterate over the requests in a tag map 293 * @tags: Tag map to iterate over. 294 * @bt: sbitmap to examine. This is either the breserved_tags member 295 * or the bitmap_tags member of struct blk_mq_tags. 296 * @fn: Pointer to the function that will be called for each started 297 * request. @fn will be called as follows: @fn(rq, @data, 298 * @reserved) where rq is a pointer to a request. Return true 299 * to continue iterating tags, false to stop. 300 * @data: Will be passed as second argument to @fn. 301 * @reserved: Indicates whether @bt is the breserved_tags member or the 302 * bitmap_tags member of struct blk_mq_tags. 303 */ 304 static void bt_tags_for_each(struct blk_mq_tags *tags, struct sbitmap_queue *bt, 305 busy_tag_iter_fn *fn, void *data, bool reserved) 306 { 307 struct bt_tags_iter_data iter_data = { 308 .tags = tags, 309 .fn = fn, 310 .data = data, 311 .reserved = reserved, 312 }; 313 314 if (tags->rqs) 315 sbitmap_for_each_set(&bt->sb, bt_tags_iter, &iter_data); 316 } 317 318 /** 319 * blk_mq_all_tag_busy_iter - iterate over all started requests in a tag map 320 * @tags: Tag map to iterate over. 321 * @fn: Pointer to the function that will be called for each started 322 * request. @fn will be called as follows: @fn(rq, @priv, 323 * reserved) where rq is a pointer to a request. 'reserved' 324 * indicates whether or not @rq is a reserved request. Return 325 * true to continue iterating tags, false to stop. 326 * @priv: Will be passed as second argument to @fn. 327 */ 328 static void blk_mq_all_tag_busy_iter(struct blk_mq_tags *tags, 329 busy_tag_iter_fn *fn, void *priv) 330 { 331 if (tags->nr_reserved_tags) 332 bt_tags_for_each(tags, &tags->breserved_tags, fn, priv, true); 333 bt_tags_for_each(tags, &tags->bitmap_tags, fn, priv, false); 334 } 335 336 /** 337 * blk_mq_tagset_busy_iter - iterate over all started requests in a tag set 338 * @tagset: Tag set to iterate over. 339 * @fn: Pointer to the function that will be called for each started 340 * request. @fn will be called as follows: @fn(rq, @priv, 341 * reserved) where rq is a pointer to a request. 'reserved' 342 * indicates whether or not @rq is a reserved request. Return 343 * true to continue iterating tags, false to stop. 344 * @priv: Will be passed as second argument to @fn. 345 */ 346 void blk_mq_tagset_busy_iter(struct blk_mq_tag_set *tagset, 347 busy_tag_iter_fn *fn, void *priv) 348 { 349 int i; 350 351 for (i = 0; i < tagset->nr_hw_queues; i++) { 352 if (tagset->tags && tagset->tags[i]) 353 blk_mq_all_tag_busy_iter(tagset->tags[i], fn, priv); 354 } 355 } 356 EXPORT_SYMBOL(blk_mq_tagset_busy_iter); 357 358 static bool blk_mq_tagset_count_completed_rqs(struct request *rq, 359 void *data, bool reserved) 360 { 361 unsigned *count = data; 362 363 if (blk_mq_request_completed(rq)) 364 (*count)++; 365 return true; 366 } 367 368 /** 369 * blk_mq_tagset_wait_completed_request - wait until all completed req's 370 * complete funtion is run 371 * @tagset: Tag set to drain completed request 372 * 373 * Note: This function has to be run after all IO queues are shutdown 374 */ 375 void blk_mq_tagset_wait_completed_request(struct blk_mq_tag_set *tagset) 376 { 377 while (true) { 378 unsigned count = 0; 379 380 blk_mq_tagset_busy_iter(tagset, 381 blk_mq_tagset_count_completed_rqs, &count); 382 if (!count) 383 break; 384 msleep(5); 385 } 386 } 387 EXPORT_SYMBOL(blk_mq_tagset_wait_completed_request); 388 389 /** 390 * blk_mq_queue_tag_busy_iter - iterate over all requests with a driver tag 391 * @q: Request queue to examine. 392 * @fn: Pointer to the function that will be called for each request 393 * on @q. @fn will be called as follows: @fn(hctx, rq, @priv, 394 * reserved) where rq is a pointer to a request and hctx points 395 * to the hardware queue associated with the request. 'reserved' 396 * indicates whether or not @rq is a reserved request. 397 * @priv: Will be passed as third argument to @fn. 398 * 399 * Note: if @q->tag_set is shared with other request queues then @fn will be 400 * called for all requests on all queues that share that tag set and not only 401 * for requests associated with @q. 402 */ 403 void blk_mq_queue_tag_busy_iter(struct request_queue *q, busy_iter_fn *fn, 404 void *priv) 405 { 406 struct blk_mq_hw_ctx *hctx; 407 int i; 408 409 /* 410 * __blk_mq_update_nr_hw_queues() updates nr_hw_queues and queue_hw_ctx 411 * while the queue is frozen. So we can use q_usage_counter to avoid 412 * racing with it. __blk_mq_update_nr_hw_queues() uses 413 * synchronize_rcu() to ensure this function left the critical section 414 * below. 415 */ 416 if (!percpu_ref_tryget(&q->q_usage_counter)) 417 return; 418 419 queue_for_each_hw_ctx(q, hctx, i) { 420 struct blk_mq_tags *tags = hctx->tags; 421 422 /* 423 * If no software queues are currently mapped to this 424 * hardware queue, there's nothing to check 425 */ 426 if (!blk_mq_hw_queue_mapped(hctx)) 427 continue; 428 429 if (tags->nr_reserved_tags) 430 bt_for_each(hctx, &tags->breserved_tags, fn, priv, true); 431 bt_for_each(hctx, &tags->bitmap_tags, fn, priv, false); 432 } 433 blk_queue_exit(q); 434 } 435 436 static int bt_alloc(struct sbitmap_queue *bt, unsigned int depth, 437 bool round_robin, int node) 438 { 439 return sbitmap_queue_init_node(bt, depth, -1, round_robin, GFP_KERNEL, 440 node); 441 } 442 443 static struct blk_mq_tags *blk_mq_init_bitmap_tags(struct blk_mq_tags *tags, 444 int node, int alloc_policy) 445 { 446 unsigned int depth = tags->nr_tags - tags->nr_reserved_tags; 447 bool round_robin = alloc_policy == BLK_TAG_ALLOC_RR; 448 449 if (bt_alloc(&tags->bitmap_tags, depth, round_robin, node)) 450 goto free_tags; 451 if (bt_alloc(&tags->breserved_tags, tags->nr_reserved_tags, round_robin, 452 node)) 453 goto free_bitmap_tags; 454 455 return tags; 456 free_bitmap_tags: 457 sbitmap_queue_free(&tags->bitmap_tags); 458 free_tags: 459 kfree(tags); 460 return NULL; 461 } 462 463 struct blk_mq_tags *blk_mq_init_tags(unsigned int total_tags, 464 unsigned int reserved_tags, 465 int node, int alloc_policy) 466 { 467 struct blk_mq_tags *tags; 468 469 if (total_tags > BLK_MQ_TAG_MAX) { 470 pr_err("blk-mq: tag depth too large\n"); 471 return NULL; 472 } 473 474 tags = kzalloc_node(sizeof(*tags), GFP_KERNEL, node); 475 if (!tags) 476 return NULL; 477 478 tags->nr_tags = total_tags; 479 tags->nr_reserved_tags = reserved_tags; 480 481 return blk_mq_init_bitmap_tags(tags, node, alloc_policy); 482 } 483 484 void blk_mq_free_tags(struct blk_mq_tags *tags) 485 { 486 sbitmap_queue_free(&tags->bitmap_tags); 487 sbitmap_queue_free(&tags->breserved_tags); 488 kfree(tags); 489 } 490 491 int blk_mq_tag_update_depth(struct blk_mq_hw_ctx *hctx, 492 struct blk_mq_tags **tagsptr, unsigned int tdepth, 493 bool can_grow) 494 { 495 struct blk_mq_tags *tags = *tagsptr; 496 497 if (tdepth <= tags->nr_reserved_tags) 498 return -EINVAL; 499 500 /* 501 * If we are allowed to grow beyond the original size, allocate 502 * a new set of tags before freeing the old one. 503 */ 504 if (tdepth > tags->nr_tags) { 505 struct blk_mq_tag_set *set = hctx->queue->tag_set; 506 struct blk_mq_tags *new; 507 bool ret; 508 509 if (!can_grow) 510 return -EINVAL; 511 512 /* 513 * We need some sort of upper limit, set it high enough that 514 * no valid use cases should require more. 515 */ 516 if (tdepth > 16 * BLKDEV_MAX_RQ) 517 return -EINVAL; 518 519 new = blk_mq_alloc_rq_map(set, hctx->queue_num, tdepth, 520 tags->nr_reserved_tags); 521 if (!new) 522 return -ENOMEM; 523 ret = blk_mq_alloc_rqs(set, new, hctx->queue_num, tdepth); 524 if (ret) { 525 blk_mq_free_rq_map(new); 526 return -ENOMEM; 527 } 528 529 blk_mq_free_rqs(set, *tagsptr, hctx->queue_num); 530 blk_mq_free_rq_map(*tagsptr); 531 *tagsptr = new; 532 } else { 533 /* 534 * Don't need (or can't) update reserved tags here, they 535 * remain static and should never need resizing. 536 */ 537 sbitmap_queue_resize(&tags->bitmap_tags, 538 tdepth - tags->nr_reserved_tags); 539 } 540 541 return 0; 542 } 543 544 /** 545 * blk_mq_unique_tag() - return a tag that is unique queue-wide 546 * @rq: request for which to compute a unique tag 547 * 548 * The tag field in struct request is unique per hardware queue but not over 549 * all hardware queues. Hence this function that returns a tag with the 550 * hardware context index in the upper bits and the per hardware queue tag in 551 * the lower bits. 552 * 553 * Note: When called for a request that is queued on a non-multiqueue request 554 * queue, the hardware context index is set to zero. 555 */ 556 u32 blk_mq_unique_tag(struct request *rq) 557 { 558 return (rq->mq_hctx->queue_num << BLK_MQ_UNIQUE_TAG_BITS) | 559 (rq->tag & BLK_MQ_UNIQUE_TAG_MASK); 560 } 561 EXPORT_SYMBOL(blk_mq_unique_tag); 562