blk-mq-tag.c 10.2 KB
Newer Older
1 2
#include <linux/kernel.h>
#include <linux/module.h>
3
#include <linux/random.h>
4 5 6 7 8 9

#include <linux/blk-mq.h>
#include "blk.h"
#include "blk-mq.h"
#include "blk-mq-tag.h"

10 11
void blk_mq_wait_for_tags(struct blk_mq_tags *tags, struct blk_mq_hw_ctx *hctx,
			  bool reserved)
12
{
13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32
	int tag, zero = 0;

	tag = blk_mq_get_tag(tags, hctx, &zero, __GFP_WAIT, reserved);
	blk_mq_put_tag(tags, tag, &zero);
}

static bool bt_has_free_tags(struct blk_mq_bitmap_tags *bt)
{
	int i;

	for (i = 0; i < bt->map_nr; i++) {
		struct blk_mq_bitmap *bm = &bt->map[i];
		int ret;

		ret = find_first_zero_bit(&bm->word, bm->depth);
		if (ret < bm->depth)
			return true;
	}

	return false;
33 34 35 36
}

bool blk_mq_has_free_tags(struct blk_mq_tags *tags)
{
37 38 39 40 41 42 43 44 45 46
	if (!tags)
		return true;

	return bt_has_free_tags(&tags->bitmap_tags);
}

static int __bt_get_word(struct blk_mq_bitmap *bm, unsigned int last_tag)
{
	int tag, org_last_tag, end;

47
	org_last_tag = last_tag;
48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86
	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.
 */
static int __bt_get(struct blk_mq_bitmap_tags *bt, unsigned int *tag_cache)
{
	unsigned int last_tag, org_last_tag;
	int index, i, tag;

	last_tag = org_last_tag = *tag_cache;
87
	index = TAG_TO_INDEX(bt, last_tag);
88 89

	for (i = 0; i < bt->map_nr; i++) {
90
		tag = __bt_get_word(&bt->map[index], TAG_TO_BIT(bt, last_tag));
91
		if (tag != -1) {
92
			tag += (index << bt->bits_per_word);
93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135
			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;
}

static inline void bt_index_inc(unsigned int *index)
{
	*index = (*index + 1) & (BT_WAIT_QUEUES - 1);
}

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;

	if (!hctx)
		return &bt->bs[0];

	bs = &bt->bs[hctx->wait_index];
	bt_index_inc(&hctx->wait_index);
	return bs;
136 137
}

138 139
static int bt_get(struct blk_mq_bitmap_tags *bt, struct blk_mq_hw_ctx *hctx,
		  unsigned int *last_tag, gfp_t gfp)
140
{
141 142
	struct bt_wait_state *bs;
	DEFINE_WAIT(wait);
143 144
	int tag;

145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183
	tag = __bt_get(bt, last_tag);
	if (tag != -1)
		return tag;

	if (!(gfp & __GFP_WAIT))
		return -1;

	bs = bt_wait_ptr(bt, hctx);
	do {
		bool was_empty;

		was_empty = list_empty(&wait.task_list);
		prepare_to_wait(&bs->wait, &wait, TASK_UNINTERRUPTIBLE);

		tag = __bt_get(bt, last_tag);
		if (tag != -1)
			break;

		if (was_empty)
			atomic_set(&bs->wait_cnt, bt->wake_cnt);

		io_schedule();
	} while (1);

	finish_wait(&bs->wait, &wait);
	return tag;
}

static unsigned int __blk_mq_get_tag(struct blk_mq_tags *tags,
				     struct blk_mq_hw_ctx *hctx,
				     unsigned int *last_tag, gfp_t gfp)
{
	int tag;

	tag = bt_get(&tags->bitmap_tags, hctx, last_tag, gfp);
	if (tag >= 0)
		return tag + tags->nr_reserved_tags;

	return BLK_MQ_TAG_FAIL;
184 185 186 187 188
}

static unsigned int __blk_mq_get_reserved_tag(struct blk_mq_tags *tags,
					      gfp_t gfp)
{
189
	int tag, zero = 0;
190 191 192 193 194 195

	if (unlikely(!tags->nr_reserved_tags)) {
		WARN_ON_ONCE(1);
		return BLK_MQ_TAG_FAIL;
	}

196
	tag = bt_get(&tags->breserved_tags, NULL, &zero, gfp);
197 198
	if (tag < 0)
		return BLK_MQ_TAG_FAIL;
199

200 201 202
	return tag;
}

203 204 205
unsigned int blk_mq_get_tag(struct blk_mq_tags *tags,
			    struct blk_mq_hw_ctx *hctx, unsigned int *last_tag,
			    gfp_t gfp, bool reserved)
206 207
{
	if (!reserved)
208
		return __blk_mq_get_tag(tags, hctx, last_tag, gfp);
209 210 211 212

	return __blk_mq_get_reserved_tag(tags, gfp);
}

213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235
static struct bt_wait_state *bt_wake_ptr(struct blk_mq_bitmap_tags *bt)
{
	int i, wake_index;

	wake_index = bt->wake_index;
	for (i = 0; i < BT_WAIT_QUEUES; i++) {
		struct bt_wait_state *bs = &bt->bs[wake_index];

		if (waitqueue_active(&bs->wait)) {
			if (wake_index != bt->wake_index)
				bt->wake_index = wake_index;

			return bs;
		}

		bt_index_inc(&wake_index);
	}

	return NULL;
}

static void bt_clear_tag(struct blk_mq_bitmap_tags *bt, unsigned int tag)
{
236
	const int index = TAG_TO_INDEX(bt, tag);
237 238
	struct bt_wait_state *bs;

239
	clear_bit(TAG_TO_BIT(bt, tag), &bt->map[index].word);
240 241 242 243 244 245 246 247 248 249

	bs = bt_wake_ptr(bt);
	if (bs && atomic_dec_and_test(&bs->wait_cnt)) {
		smp_mb__after_clear_bit();
		atomic_set(&bs->wait_cnt, bt->wake_cnt);
		bt_index_inc(&bt->wake_index);
		wake_up(&bs->wait);
	}
}

250 251 252 253
static void __blk_mq_put_tag(struct blk_mq_tags *tags, unsigned int tag)
{
	BUG_ON(tag >= tags->nr_tags);

254
	bt_clear_tag(&tags->bitmap_tags, tag);
255 256 257 258 259 260 261
}

static void __blk_mq_put_reserved_tag(struct blk_mq_tags *tags,
				      unsigned int tag)
{
	BUG_ON(tag >= tags->nr_reserved_tags);

262
	bt_clear_tag(&tags->breserved_tags, tag);
263 264
}

265 266
void blk_mq_put_tag(struct blk_mq_tags *tags, unsigned int tag,
		    unsigned int *last_tag)
267
{
268 269 270 271 272 273
	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
274 275 276
		__blk_mq_put_reserved_tag(tags, tag);
}

277 278
static void bt_for_each_free(struct blk_mq_bitmap_tags *bt,
			     unsigned long *free_map, unsigned int off)
279
{
280 281 282 283 284 285 286 287 288 289 290 291 292 293 294
	int i;

	for (i = 0; i < bt->map_nr; i++) {
		struct blk_mq_bitmap *bm = &bt->map[i];
		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);

295
		off += (1 << bt->bits_per_word);
296
	}
297 298 299 300 301 302 303 304 305 306 307 308 309
}

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;

310
	bt_for_each_free(&tags->bitmap_tags, tag_map, tags->nr_reserved_tags);
311
	if (tags->nr_reserved_tags)
312
		bt_for_each_free(&tags->breserved_tags, tag_map, 0);
313 314 315 316 317

	fn(data, tag_map);
	kfree(tag_map);
}

318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335
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++) {
		struct blk_mq_bitmap *bm = &bt->map[i];

		used += bitmap_weight(&bm->word, bm->depth);
	}

	return bt->depth - used;
}

static int bt_alloc(struct blk_mq_bitmap_tags *bt, unsigned int depth,
			int node, bool reserved)
{
	int i;

336 337
	bt->bits_per_word = ilog2(BITS_PER_LONG);

338 339 340 341 342
	/*
	 * Depth can be zero for reserved tags, that's not a failure
	 * condition.
	 */
	if (depth) {
343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358
		unsigned int nr, i, map_depth, tags_per_word;

		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);
			}
		}
359

360
		nr = ALIGN(depth, tags_per_word) / tags_per_word;
361 362 363 364 365 366 367 368
		bt->map = kzalloc_node(nr * sizeof(struct blk_mq_bitmap),
						GFP_KERNEL, node);
		if (!bt->map)
			return -ENOMEM;

		bt->map_nr = nr;
		map_depth = depth;
		for (i = 0; i < nr; i++) {
369 370
			bt->map[i].depth = min(map_depth, tags_per_word);
			map_depth -= tags_per_word;
371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413
		}
	}

	bt->bs = kzalloc(BT_WAIT_QUEUES * sizeof(*bt->bs), GFP_KERNEL);
	if (!bt->bs) {
		kfree(bt->map);
		return -ENOMEM;
	}

	for (i = 0; i < BT_WAIT_QUEUES; i++)
		init_waitqueue_head(&bt->bs[i].wait);

	bt->wake_cnt = BT_WAIT_BATCH;
	if (bt->wake_cnt > depth / 4)
		bt->wake_cnt = max(1U, depth / 4);

	bt->depth = depth;
	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;
}

414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429
struct blk_mq_tags *blk_mq_init_tags(unsigned int total_tags,
				     unsigned int reserved_tags, int node)
{
	unsigned int nr_tags, nr_cache;
	struct blk_mq_tags *tags;

	if (total_tags > BLK_MQ_TAG_MAX) {
		pr_err("blk-mq: tag depth too large\n");
		return NULL;
	}

	tags = kzalloc_node(sizeof(*tags), GFP_KERNEL, node);
	if (!tags)
		return NULL;

	nr_tags = total_tags - reserved_tags;
430
	nr_cache = nr_tags / num_online_cpus();
431 432 433 434

	tags->nr_tags = total_tags;
	tags->nr_reserved_tags = reserved_tags;

435
	return blk_mq_init_bitmap_tags(tags, node);
436 437 438 439
}

void blk_mq_free_tags(struct blk_mq_tags *tags)
{
440 441
	bt_free(&tags->bitmap_tags);
	bt_free(&tags->breserved_tags);
442 443 444
	kfree(tags);
}

445 446 447 448 449 450 451 452 453 454
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;

	if (depth > 1)
		*tag = prandom_u32() % (depth - 1);
	else
		*tag = 0;
}

455 456 457
ssize_t blk_mq_tag_sysfs_show(struct blk_mq_tags *tags, char *page)
{
	char *orig_page = page;
458
	unsigned int free, res;
459 460 461 462

	if (!tags)
		return 0;

463 464 465 466
	page += sprintf(page, "nr_tags=%u, reserved_tags=%u, "
			"bits_per_word=%u\n",
			tags->nr_tags, tags->nr_reserved_tags,
			tags->bitmap_tags.bits_per_word);
467

468 469
	free = bt_unused_tags(&tags->bitmap_tags);
	res = bt_unused_tags(&tags->breserved_tags);
470

471
	page += sprintf(page, "nr_free=%u, nr_reserved=%u\n", free, res);
472 473 474

	return page - orig_page;
}