Blame view
block/blk-mq-tag.c
13.7 KB
75bb4625b
|
1 2 3 4 5 6 7 8 9 10 11 12 |
/* * Fast and scalable bitmap tagging variant. Uses sparser bitmaps spread * over multiple cachelines to avoid ping-pong between multiple submitters * or submitter and completer. Uses rolling wakeups to avoid falling of * the scaling cliff when we run out of tags and have to start putting * submitters to sleep. * * Uses active queue tracking to support fairer distribution of tags * between multiple submitters when a shared tag map is used. * * Copyright (C) 2013-2014 Jens Axboe */ |
320ae51fe
|
13 14 |
#include <linux/kernel.h> #include <linux/module.h> |
4bb659b15
|
15 |
#include <linux/random.h> |
320ae51fe
|
16 17 18 19 20 |
#include <linux/blk-mq.h> #include "blk.h" #include "blk-mq.h" #include "blk-mq-tag.h" |
4bb659b15
|
21 22 23 24 25 |
static bool bt_has_free_tags(struct blk_mq_bitmap_tags *bt) { int i; for (i = 0; i < bt->map_nr; i++) { |
e93ecf602
|
26 |
struct blk_align_bitmap *bm = &bt->map[i]; |
4bb659b15
|
27 28 29 30 31 32 33 34 |
int ret; ret = find_first_zero_bit(&bm->word, bm->depth); if (ret < bm->depth) return true; } return false; |
320ae51fe
|
35 36 37 38 |
} bool blk_mq_has_free_tags(struct blk_mq_tags *tags) { |
4bb659b15
|
39 40 41 42 43 |
if (!tags) return true; return bt_has_free_tags(&tags->bitmap_tags); } |
8537b1203
|
44 |
static inline int bt_index_inc(int index) |
0d2602ca3
|
45 |
{ |
8537b1203
|
46 47 48 49 50 51 52 53 |
return (index + 1) & (BT_WAIT_QUEUES - 1); } static inline void bt_index_atomic_inc(atomic_t *index) { int old = atomic_read(index); int new = bt_index_inc(old); atomic_cmpxchg(index, old, new); |
0d2602ca3
|
54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 |
} /* * If a previously inactive queue goes active, bump the active user count. */ bool __blk_mq_tag_busy(struct blk_mq_hw_ctx *hctx) { if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state) && !test_and_set_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state)) atomic_inc(&hctx->tags->active_queues); return true; } /* |
e3a2b3f93
|
69 |
* Wakeup all potentially sleeping on normal (non-reserved) tags |
0d2602ca3
|
70 |
*/ |
e3a2b3f93
|
71 |
static void blk_mq_tag_wakeup_all(struct blk_mq_tags *tags) |
0d2602ca3
|
72 |
{ |
0d2602ca3
|
73 74 |
struct blk_mq_bitmap_tags *bt; int i, wake_index; |
0d2602ca3
|
75 |
bt = &tags->bitmap_tags; |
8537b1203
|
76 |
wake_index = atomic_read(&bt->wake_index); |
0d2602ca3
|
77 78 79 80 81 |
for (i = 0; i < BT_WAIT_QUEUES; i++) { struct bt_wait_state *bs = &bt->bs[wake_index]; if (waitqueue_active(&bs->wait)) wake_up(&bs->wait); |
8537b1203
|
82 |
wake_index = bt_index_inc(wake_index); |
0d2602ca3
|
83 84 85 86 |
} } /* |
e3a2b3f93
|
87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 |
* If a previously busy queue goes inactive, potential waiters could now * be allowed to queue. Wake them up and check. */ void __blk_mq_tag_idle(struct blk_mq_hw_ctx *hctx) { struct blk_mq_tags *tags = hctx->tags; if (!test_and_clear_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state)) return; atomic_dec(&tags->active_queues); blk_mq_tag_wakeup_all(tags); } /* |
0d2602ca3
|
103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 |
* For shared tag users, we track the number of currently active users * and attempt to provide a fair share of the tag depth for each of them. */ static inline bool hctx_may_queue(struct blk_mq_hw_ctx *hctx, struct blk_mq_bitmap_tags *bt) { unsigned int depth, users; if (!hctx || !(hctx->flags & BLK_MQ_F_TAG_SHARED)) return true; if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state)) return true; /* * Don't try dividing an ant */ if (bt->depth == 1) return true; users = atomic_read(&hctx->tags->active_queues); if (!users) return true; /* * Allow at least some tags */ depth = max((bt->depth + users - 1) / users, 4U); return atomic_read(&hctx->nr_active) < depth; } |
e93ecf602
|
132 |
static int __bt_get_word(struct blk_align_bitmap *bm, unsigned int last_tag) |
4bb659b15
|
133 134 |
{ int tag, org_last_tag, end; |
59d13bf5f
|
135 |
org_last_tag = last_tag; |
4bb659b15
|
136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 |
end = bm->depth; do { restart: tag = find_next_zero_bit(&bm->word, end, last_tag); if (unlikely(tag >= end)) { /* * We started with an offset, start from 0 to * exhaust the map. */ if (org_last_tag && last_tag) { end = last_tag; last_tag = 0; goto restart; } return -1; } last_tag = tag + 1; } while (test_and_set_bit_lock(tag, &bm->word)); return tag; } /* * Straight forward bitmap tag implementation, where each bit is a tag * (cleared == free, and set == busy). The small twist is using per-cpu * last_tag caches, which blk-mq stores in the blk_mq_ctx software queue * contexts. This enables us to drastically limit the space searched, * without dirtying an extra shared cacheline like we would if we stored * the cache value inside the shared blk_mq_bitmap_tags structure. On top * of that, each word of tags is in a separate cacheline. This means that * multiple users will tend to stick to different cachelines, at least * until the map is exhausted. */ |
0d2602ca3
|
169 170 |
static int __bt_get(struct blk_mq_hw_ctx *hctx, struct blk_mq_bitmap_tags *bt, unsigned int *tag_cache) |
4bb659b15
|
171 172 173 |
{ unsigned int last_tag, org_last_tag; int index, i, tag; |
0d2602ca3
|
174 175 |
if (!hctx_may_queue(hctx, bt)) return -1; |
4bb659b15
|
176 |
last_tag = org_last_tag = *tag_cache; |
59d13bf5f
|
177 |
index = TAG_TO_INDEX(bt, last_tag); |
4bb659b15
|
178 179 |
for (i = 0; i < bt->map_nr; i++) { |
59d13bf5f
|
180 |
tag = __bt_get_word(&bt->map[index], TAG_TO_BIT(bt, last_tag)); |
4bb659b15
|
181 |
if (tag != -1) { |
59d13bf5f
|
182 |
tag += (index << bt->bits_per_word); |
4bb659b15
|
183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 |
goto done; } last_tag = 0; if (++index >= bt->map_nr) index = 0; } *tag_cache = 0; return -1; /* * Only update the cache from the allocation path, if we ended * up using the specific cached tag. */ done: if (tag == org_last_tag) { last_tag = tag + 1; if (last_tag >= bt->depth - 1) last_tag = 0; *tag_cache = last_tag; } return tag; } |
4bb659b15
|
209 210 211 212 |
static struct bt_wait_state *bt_wait_ptr(struct blk_mq_bitmap_tags *bt, struct blk_mq_hw_ctx *hctx) { struct bt_wait_state *bs; |
8537b1203
|
213 |
int wait_index; |
4bb659b15
|
214 215 216 |
if (!hctx) return &bt->bs[0]; |
8537b1203
|
217 218 219 |
wait_index = atomic_read(&hctx->wait_index); bs = &bt->bs[wait_index]; bt_index_atomic_inc(&hctx->wait_index); |
4bb659b15
|
220 |
return bs; |
320ae51fe
|
221 |
} |
cb96a42cc
|
222 223 224 225 |
static int bt_get(struct blk_mq_alloc_data *data, struct blk_mq_bitmap_tags *bt, struct blk_mq_hw_ctx *hctx, unsigned int *last_tag) |
320ae51fe
|
226 |
{ |
4bb659b15
|
227 228 |
struct bt_wait_state *bs; DEFINE_WAIT(wait); |
320ae51fe
|
229 |
int tag; |
0d2602ca3
|
230 |
tag = __bt_get(hctx, bt, last_tag); |
4bb659b15
|
231 232 |
if (tag != -1) return tag; |
cb96a42cc
|
233 |
if (!(data->gfp & __GFP_WAIT)) |
4bb659b15
|
234 235 236 237 |
return -1; bs = bt_wait_ptr(bt, hctx); do { |
4bb659b15
|
238 |
prepare_to_wait(&bs->wait, &wait, TASK_UNINTERRUPTIBLE); |
0d2602ca3
|
239 |
tag = __bt_get(hctx, bt, last_tag); |
4bb659b15
|
240 241 |
if (tag != -1) break; |
cb96a42cc
|
242 |
blk_mq_put_ctx(data->ctx); |
4bb659b15
|
243 |
io_schedule(); |
cb96a42cc
|
244 245 246 247 248 249 250 251 252 253 254 255 256 |
data->ctx = blk_mq_get_ctx(data->q); data->hctx = data->q->mq_ops->map_queue(data->q, data->ctx->cpu); if (data->reserved) { bt = &data->hctx->tags->breserved_tags; } else { last_tag = &data->ctx->last_tag; hctx = data->hctx; bt = &hctx->tags->bitmap_tags; } finish_wait(&bs->wait, &wait); bs = bt_wait_ptr(bt, hctx); |
4bb659b15
|
257 258 259 260 261 |
} while (1); finish_wait(&bs->wait, &wait); return tag; } |
cb96a42cc
|
262 |
static unsigned int __blk_mq_get_tag(struct blk_mq_alloc_data *data) |
4bb659b15
|
263 264 |
{ int tag; |
cb96a42cc
|
265 266 |
tag = bt_get(data, &data->hctx->tags->bitmap_tags, data->hctx, &data->ctx->last_tag); |
4bb659b15
|
267 |
if (tag >= 0) |
cb96a42cc
|
268 |
return tag + data->hctx->tags->nr_reserved_tags; |
4bb659b15
|
269 270 |
return BLK_MQ_TAG_FAIL; |
320ae51fe
|
271 |
} |
cb96a42cc
|
272 |
static unsigned int __blk_mq_get_reserved_tag(struct blk_mq_alloc_data *data) |
320ae51fe
|
273 |
{ |
4bb659b15
|
274 |
int tag, zero = 0; |
320ae51fe
|
275 |
|
cb96a42cc
|
276 |
if (unlikely(!data->hctx->tags->nr_reserved_tags)) { |
320ae51fe
|
277 278 279 |
WARN_ON_ONCE(1); return BLK_MQ_TAG_FAIL; } |
cb96a42cc
|
280 |
tag = bt_get(data, &data->hctx->tags->breserved_tags, NULL, &zero); |
320ae51fe
|
281 282 |
if (tag < 0) return BLK_MQ_TAG_FAIL; |
4bb659b15
|
283 |
|
320ae51fe
|
284 285 |
return tag; } |
cb96a42cc
|
286 |
unsigned int blk_mq_get_tag(struct blk_mq_alloc_data *data) |
320ae51fe
|
287 |
{ |
cb96a42cc
|
288 289 |
if (!data->reserved) return __blk_mq_get_tag(data); |
320ae51fe
|
290 |
|
cb96a42cc
|
291 |
return __blk_mq_get_reserved_tag(data); |
320ae51fe
|
292 |
} |
4bb659b15
|
293 294 295 |
static struct bt_wait_state *bt_wake_ptr(struct blk_mq_bitmap_tags *bt) { int i, wake_index; |
8537b1203
|
296 |
wake_index = atomic_read(&bt->wake_index); |
4bb659b15
|
297 298 299 300 |
for (i = 0; i < BT_WAIT_QUEUES; i++) { struct bt_wait_state *bs = &bt->bs[wake_index]; if (waitqueue_active(&bs->wait)) { |
8537b1203
|
301 302 303 |
int o = atomic_read(&bt->wake_index); if (wake_index != o) atomic_cmpxchg(&bt->wake_index, o, wake_index); |
4bb659b15
|
304 305 306 |
return bs; } |
8537b1203
|
307 |
wake_index = bt_index_inc(wake_index); |
4bb659b15
|
308 309 310 311 312 313 314 |
} return NULL; } static void bt_clear_tag(struct blk_mq_bitmap_tags *bt, unsigned int tag) { |
59d13bf5f
|
315 |
const int index = TAG_TO_INDEX(bt, tag); |
4bb659b15
|
316 |
struct bt_wait_state *bs; |
2971c35f3
|
317 |
int wait_cnt; |
4bb659b15
|
318 |
|
0289b2e11
|
319 320 321 322 323 |
/* * The unlock memory barrier need to order access to req in free * path and clearing tag bit */ clear_bit_unlock(TAG_TO_BIT(bt, tag), &bt->map[index].word); |
4bb659b15
|
324 325 |
bs = bt_wake_ptr(bt); |
2971c35f3
|
326 327 328 329 330 331 332 |
if (!bs) return; wait_cnt = atomic_dec_return(&bs->wait_cnt); if (wait_cnt == 0) { wake: atomic_add(bt->wake_cnt, &bs->wait_cnt); |
8537b1203
|
333 |
bt_index_atomic_inc(&bt->wake_index); |
4bb659b15
|
334 |
wake_up(&bs->wait); |
2971c35f3
|
335 336 337 338 |
} else if (wait_cnt < 0) { wait_cnt = atomic_inc_return(&bs->wait_cnt); if (!wait_cnt) goto wake; |
4bb659b15
|
339 340 |
} } |
320ae51fe
|
341 342 343 |
static void __blk_mq_put_tag(struct blk_mq_tags *tags, unsigned int tag) { BUG_ON(tag >= tags->nr_tags); |
4bb659b15
|
344 |
bt_clear_tag(&tags->bitmap_tags, tag); |
320ae51fe
|
345 346 347 348 349 350 |
} static void __blk_mq_put_reserved_tag(struct blk_mq_tags *tags, unsigned int tag) { BUG_ON(tag >= tags->nr_reserved_tags); |
4bb659b15
|
351 |
bt_clear_tag(&tags->breserved_tags, tag); |
320ae51fe
|
352 |
} |
0d2602ca3
|
353 |
void blk_mq_put_tag(struct blk_mq_hw_ctx *hctx, unsigned int tag, |
4bb659b15
|
354 |
unsigned int *last_tag) |
320ae51fe
|
355 |
{ |
0d2602ca3
|
356 |
struct blk_mq_tags *tags = hctx->tags; |
4bb659b15
|
357 358 359 360 361 362 |
if (tag >= tags->nr_reserved_tags) { const int real_tag = tag - tags->nr_reserved_tags; __blk_mq_put_tag(tags, real_tag); *last_tag = real_tag; } else |
320ae51fe
|
363 364 |
__blk_mq_put_reserved_tag(tags, tag); } |
4bb659b15
|
365 366 |
static void bt_for_each_free(struct blk_mq_bitmap_tags *bt, unsigned long *free_map, unsigned int off) |
320ae51fe
|
367 |
{ |
4bb659b15
|
368 369 370 |
int i; for (i = 0; i < bt->map_nr; i++) { |
e93ecf602
|
371 |
struct blk_align_bitmap *bm = &bt->map[i]; |
4bb659b15
|
372 373 374 375 376 377 378 379 380 381 |
int bit = 0; do { bit = find_next_zero_bit(&bm->word, bm->depth, bit); if (bit >= bm->depth) break; __set_bit(bit + off, free_map); bit++; } while (1); |
59d13bf5f
|
382 |
off += (1 << bt->bits_per_word); |
4bb659b15
|
383 |
} |
320ae51fe
|
384 385 386 387 388 389 390 391 392 393 394 395 |
} void blk_mq_tag_busy_iter(struct blk_mq_tags *tags, void (*fn)(void *, unsigned long *), void *data) { unsigned long *tag_map; size_t map_size; map_size = ALIGN(tags->nr_tags, BITS_PER_LONG) / BITS_PER_LONG; tag_map = kzalloc(map_size * sizeof(unsigned long), GFP_ATOMIC); if (!tag_map) return; |
4bb659b15
|
396 |
bt_for_each_free(&tags->bitmap_tags, tag_map, tags->nr_reserved_tags); |
320ae51fe
|
397 |
if (tags->nr_reserved_tags) |
4bb659b15
|
398 |
bt_for_each_free(&tags->breserved_tags, tag_map, 0); |
320ae51fe
|
399 400 401 402 |
fn(data, tag_map); kfree(tag_map); } |
edf866b38
|
403 |
EXPORT_SYMBOL(blk_mq_tag_busy_iter); |
320ae51fe
|
404 |
|
4bb659b15
|
405 406 407 408 409 |
static unsigned int bt_unused_tags(struct blk_mq_bitmap_tags *bt) { unsigned int i, used; for (i = 0, used = 0; i < bt->map_nr; i++) { |
e93ecf602
|
410 |
struct blk_align_bitmap *bm = &bt->map[i]; |
4bb659b15
|
411 412 413 414 415 416 |
used += bitmap_weight(&bm->word, bm->depth); } return bt->depth - used; } |
e3a2b3f93
|
417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 |
static void bt_update_count(struct blk_mq_bitmap_tags *bt, unsigned int depth) { unsigned int tags_per_word = 1U << bt->bits_per_word; unsigned int map_depth = depth; if (depth) { int i; for (i = 0; i < bt->map_nr; i++) { bt->map[i].depth = min(map_depth, tags_per_word); map_depth -= bt->map[i].depth; } } bt->wake_cnt = BT_WAIT_BATCH; if (bt->wake_cnt > depth / 4) bt->wake_cnt = max(1U, depth / 4); bt->depth = depth; } |
4bb659b15
|
438 439 440 441 |
static int bt_alloc(struct blk_mq_bitmap_tags *bt, unsigned int depth, int node, bool reserved) { int i; |
59d13bf5f
|
442 |
bt->bits_per_word = ilog2(BITS_PER_LONG); |
4bb659b15
|
443 444 445 446 447 |
/* * Depth can be zero for reserved tags, that's not a failure * condition. */ if (depth) { |
e3a2b3f93
|
448 |
unsigned int nr, tags_per_word; |
59d13bf5f
|
449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 |
tags_per_word = (1 << bt->bits_per_word); /* * If the tag space is small, shrink the number of tags * per word so we spread over a few cachelines, at least. * If less than 4 tags, just forget about it, it's not * going to work optimally anyway. */ if (depth >= 4) { while (tags_per_word * 4 > depth) { bt->bits_per_word--; tags_per_word = (1 << bt->bits_per_word); } } |
4bb659b15
|
464 |
|
59d13bf5f
|
465 |
nr = ALIGN(depth, tags_per_word) / tags_per_word; |
e93ecf602
|
466 |
bt->map = kzalloc_node(nr * sizeof(struct blk_align_bitmap), |
4bb659b15
|
467 468 469 470 471 |
GFP_KERNEL, node); if (!bt->map) return -ENOMEM; bt->map_nr = nr; |
4bb659b15
|
472 473 474 475 476 477 478 |
} bt->bs = kzalloc(BT_WAIT_QUEUES * sizeof(*bt->bs), GFP_KERNEL); if (!bt->bs) { kfree(bt->map); return -ENOMEM; } |
86fb5c56c
|
479 480 481 |
bt_update_count(bt, depth); for (i = 0; i < BT_WAIT_QUEUES; i++) { |
4bb659b15
|
482 |
init_waitqueue_head(&bt->bs[i].wait); |
86fb5c56c
|
483 484 |
atomic_set(&bt->bs[i].wait_cnt, bt->wake_cnt); } |
4bb659b15
|
485 |
|
4bb659b15
|
486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 |
return 0; } static void bt_free(struct blk_mq_bitmap_tags *bt) { kfree(bt->map); kfree(bt->bs); } static struct blk_mq_tags *blk_mq_init_bitmap_tags(struct blk_mq_tags *tags, int node) { unsigned int depth = tags->nr_tags - tags->nr_reserved_tags; if (bt_alloc(&tags->bitmap_tags, depth, node, false)) goto enomem; if (bt_alloc(&tags->breserved_tags, tags->nr_reserved_tags, node, true)) goto enomem; return tags; enomem: bt_free(&tags->bitmap_tags); kfree(tags); return NULL; } |
320ae51fe
|
511 512 513 |
struct blk_mq_tags *blk_mq_init_tags(unsigned int total_tags, unsigned int reserved_tags, int node) { |
320ae51fe
|
514 |
struct blk_mq_tags *tags; |
320ae51fe
|
515 516 517 518 519 520 521 522 523 524 |
if (total_tags > BLK_MQ_TAG_MAX) { pr_err("blk-mq: tag depth too large "); return NULL; } tags = kzalloc_node(sizeof(*tags), GFP_KERNEL, node); if (!tags) return NULL; |
320ae51fe
|
525 526 |
tags->nr_tags = total_tags; tags->nr_reserved_tags = reserved_tags; |
320ae51fe
|
527 |
|
4bb659b15
|
528 |
return blk_mq_init_bitmap_tags(tags, node); |
320ae51fe
|
529 530 531 532 |
} void blk_mq_free_tags(struct blk_mq_tags *tags) { |
4bb659b15
|
533 534 |
bt_free(&tags->bitmap_tags); bt_free(&tags->breserved_tags); |
320ae51fe
|
535 536 |
kfree(tags); } |
4bb659b15
|
537 538 539 |
void blk_mq_tag_init_last_tag(struct blk_mq_tags *tags, unsigned int *tag) { unsigned int depth = tags->nr_tags - tags->nr_reserved_tags; |
9d3d21aeb
|
540 |
*tag = prandom_u32() % depth; |
4bb659b15
|
541 |
} |
e3a2b3f93
|
542 543 544 545 546 547 548 549 550 551 552 553 554 555 |
int blk_mq_tag_update_depth(struct blk_mq_tags *tags, unsigned int tdepth) { tdepth -= tags->nr_reserved_tags; if (tdepth > tags->nr_tags) return -EINVAL; /* * Don't need (or can't) update reserved tags here, they remain * static and should never need resizing. */ bt_update_count(&tags->bitmap_tags, tdepth); blk_mq_tag_wakeup_all(tags); return 0; } |
320ae51fe
|
556 557 558 |
ssize_t blk_mq_tag_sysfs_show(struct blk_mq_tags *tags, char *page) { char *orig_page = page; |
4bb659b15
|
559 |
unsigned int free, res; |
320ae51fe
|
560 561 562 |
if (!tags) return 0; |
59d13bf5f
|
563 564 565 566 567 |
page += sprintf(page, "nr_tags=%u, reserved_tags=%u, " "bits_per_word=%u ", tags->nr_tags, tags->nr_reserved_tags, tags->bitmap_tags.bits_per_word); |
320ae51fe
|
568 |
|
4bb659b15
|
569 570 |
free = bt_unused_tags(&tags->bitmap_tags); res = bt_unused_tags(&tags->breserved_tags); |
320ae51fe
|
571 |
|
4bb659b15
|
572 573 |
page += sprintf(page, "nr_free=%u, nr_reserved=%u ", free, res); |
0d2602ca3
|
574 575 |
page += sprintf(page, "active_queues=%u ", atomic_read(&tags->active_queues)); |
320ae51fe
|
576 577 578 |
return page - orig_page; } |