Jens Axboe | 75bb462 | 2014-05-28 10:15:41 -0600 | [diff] [blame] | 1 | /* |
| 2 | * Fast and scalable bitmap tagging variant. Uses sparser bitmaps spread |
| 3 | * over multiple cachelines to avoid ping-pong between multiple submitters |
| 4 | * or submitter and completer. Uses rolling wakeups to avoid falling of |
| 5 | * the scaling cliff when we run out of tags and have to start putting |
| 6 | * submitters to sleep. |
| 7 | * |
| 8 | * Uses active queue tracking to support fairer distribution of tags |
| 9 | * between multiple submitters when a shared tag map is used. |
| 10 | * |
| 11 | * Copyright (C) 2013-2014 Jens Axboe |
| 12 | */ |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 13 | #include <linux/kernel.h> |
| 14 | #include <linux/module.h> |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 15 | #include <linux/random.h> |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 16 | |
| 17 | #include <linux/blk-mq.h> |
| 18 | #include "blk.h" |
| 19 | #include "blk-mq.h" |
| 20 | #include "blk-mq-tag.h" |
| 21 | |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 22 | static bool bt_has_free_tags(struct blk_mq_bitmap_tags *bt) |
| 23 | { |
| 24 | int i; |
| 25 | |
| 26 | for (i = 0; i < bt->map_nr; i++) { |
Jens Axboe | e93ecf6 | 2014-05-19 09:17:48 -0600 | [diff] [blame] | 27 | struct blk_align_bitmap *bm = &bt->map[i]; |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 28 | int ret; |
| 29 | |
| 30 | ret = find_first_zero_bit(&bm->word, bm->depth); |
| 31 | if (ret < bm->depth) |
| 32 | return true; |
| 33 | } |
| 34 | |
| 35 | return false; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 36 | } |
| 37 | |
| 38 | bool blk_mq_has_free_tags(struct blk_mq_tags *tags) |
| 39 | { |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 40 | if (!tags) |
| 41 | return true; |
| 42 | |
| 43 | return bt_has_free_tags(&tags->bitmap_tags); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 44 | } |
| 45 | |
Alexander Gordeev | 8537b12 | 2014-06-17 22:12:35 -0700 | [diff] [blame] | 46 | static inline int bt_index_inc(int index) |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 47 | { |
Alexander Gordeev | 8537b12 | 2014-06-17 22:12:35 -0700 | [diff] [blame] | 48 | return (index + 1) & (BT_WAIT_QUEUES - 1); |
| 49 | } |
| 50 | |
| 51 | static inline void bt_index_atomic_inc(atomic_t *index) |
| 52 | { |
| 53 | int old = atomic_read(index); |
| 54 | int new = bt_index_inc(old); |
| 55 | atomic_cmpxchg(index, old, new); |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 56 | } |
| 57 | |
| 58 | /* |
| 59 | * If a previously inactive queue goes active, bump the active user count. |
| 60 | */ |
| 61 | bool __blk_mq_tag_busy(struct blk_mq_hw_ctx *hctx) |
| 62 | { |
| 63 | if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state) && |
| 64 | !test_and_set_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state)) |
| 65 | atomic_inc(&hctx->tags->active_queues); |
| 66 | |
| 67 | return true; |
| 68 | } |
| 69 | |
| 70 | /* |
Jens Axboe | aed3ea9 | 2014-12-22 14:04:42 -0700 | [diff] [blame^] | 71 | * Wakeup all potentially sleeping on tags |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 72 | */ |
Jens Axboe | aed3ea9 | 2014-12-22 14:04:42 -0700 | [diff] [blame^] | 73 | void blk_mq_tag_wakeup_all(struct blk_mq_tags *tags, bool include_reserve) |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 74 | { |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 75 | struct blk_mq_bitmap_tags *bt; |
| 76 | int i, wake_index; |
| 77 | |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 78 | bt = &tags->bitmap_tags; |
Alexander Gordeev | 8537b12 | 2014-06-17 22:12:35 -0700 | [diff] [blame] | 79 | wake_index = atomic_read(&bt->wake_index); |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 80 | for (i = 0; i < BT_WAIT_QUEUES; i++) { |
| 81 | struct bt_wait_state *bs = &bt->bs[wake_index]; |
| 82 | |
| 83 | if (waitqueue_active(&bs->wait)) |
| 84 | wake_up(&bs->wait); |
| 85 | |
Alexander Gordeev | 8537b12 | 2014-06-17 22:12:35 -0700 | [diff] [blame] | 86 | wake_index = bt_index_inc(wake_index); |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 87 | } |
Jens Axboe | aed3ea9 | 2014-12-22 14:04:42 -0700 | [diff] [blame^] | 88 | |
| 89 | if (include_reserve) { |
| 90 | bt = &tags->breserved_tags; |
| 91 | if (waitqueue_active(&bt->bs[0].wait)) |
| 92 | wake_up(&bt->bs[0].wait); |
| 93 | } |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 94 | } |
| 95 | |
| 96 | /* |
Jens Axboe | e3a2b3f | 2014-05-20 11:49:02 -0600 | [diff] [blame] | 97 | * If a previously busy queue goes inactive, potential waiters could now |
| 98 | * be allowed to queue. Wake them up and check. |
| 99 | */ |
| 100 | void __blk_mq_tag_idle(struct blk_mq_hw_ctx *hctx) |
| 101 | { |
| 102 | struct blk_mq_tags *tags = hctx->tags; |
| 103 | |
| 104 | if (!test_and_clear_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state)) |
| 105 | return; |
| 106 | |
| 107 | atomic_dec(&tags->active_queues); |
| 108 | |
Jens Axboe | aed3ea9 | 2014-12-22 14:04:42 -0700 | [diff] [blame^] | 109 | blk_mq_tag_wakeup_all(tags, false); |
Jens Axboe | e3a2b3f | 2014-05-20 11:49:02 -0600 | [diff] [blame] | 110 | } |
| 111 | |
| 112 | /* |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 113 | * For shared tag users, we track the number of currently active users |
| 114 | * and attempt to provide a fair share of the tag depth for each of them. |
| 115 | */ |
| 116 | static inline bool hctx_may_queue(struct blk_mq_hw_ctx *hctx, |
| 117 | struct blk_mq_bitmap_tags *bt) |
| 118 | { |
| 119 | unsigned int depth, users; |
| 120 | |
| 121 | if (!hctx || !(hctx->flags & BLK_MQ_F_TAG_SHARED)) |
| 122 | return true; |
| 123 | if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state)) |
| 124 | return true; |
| 125 | |
| 126 | /* |
| 127 | * Don't try dividing an ant |
| 128 | */ |
| 129 | if (bt->depth == 1) |
| 130 | return true; |
| 131 | |
| 132 | users = atomic_read(&hctx->tags->active_queues); |
| 133 | if (!users) |
| 134 | return true; |
| 135 | |
| 136 | /* |
| 137 | * Allow at least some tags |
| 138 | */ |
| 139 | depth = max((bt->depth + users - 1) / users, 4U); |
| 140 | return atomic_read(&hctx->nr_active) < depth; |
| 141 | } |
| 142 | |
Jens Axboe | e93ecf6 | 2014-05-19 09:17:48 -0600 | [diff] [blame] | 143 | static int __bt_get_word(struct blk_align_bitmap *bm, unsigned int last_tag) |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 144 | { |
| 145 | int tag, org_last_tag, end; |
Bart Van Assche | 9e98e9d | 2014-12-09 16:58:11 +0100 | [diff] [blame] | 146 | bool wrap = last_tag != 0; |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 147 | |
Jens Axboe | 59d13bf | 2014-05-09 13:41:15 -0600 | [diff] [blame] | 148 | org_last_tag = last_tag; |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 149 | end = bm->depth; |
| 150 | do { |
| 151 | restart: |
| 152 | tag = find_next_zero_bit(&bm->word, end, last_tag); |
| 153 | if (unlikely(tag >= end)) { |
| 154 | /* |
| 155 | * We started with an offset, start from 0 to |
| 156 | * exhaust the map. |
| 157 | */ |
Bart Van Assche | 9e98e9d | 2014-12-09 16:58:11 +0100 | [diff] [blame] | 158 | if (wrap) { |
| 159 | wrap = false; |
| 160 | end = org_last_tag; |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 161 | last_tag = 0; |
| 162 | goto restart; |
| 163 | } |
| 164 | return -1; |
| 165 | } |
| 166 | last_tag = tag + 1; |
Bart Van Assche | c38d185 | 2014-12-09 16:58:35 +0100 | [diff] [blame] | 167 | } while (test_and_set_bit(tag, &bm->word)); |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 168 | |
| 169 | return tag; |
| 170 | } |
| 171 | |
| 172 | /* |
| 173 | * Straight forward bitmap tag implementation, where each bit is a tag |
| 174 | * (cleared == free, and set == busy). The small twist is using per-cpu |
| 175 | * last_tag caches, which blk-mq stores in the blk_mq_ctx software queue |
| 176 | * contexts. This enables us to drastically limit the space searched, |
| 177 | * without dirtying an extra shared cacheline like we would if we stored |
| 178 | * the cache value inside the shared blk_mq_bitmap_tags structure. On top |
| 179 | * of that, each word of tags is in a separate cacheline. This means that |
| 180 | * multiple users will tend to stick to different cachelines, at least |
| 181 | * until the map is exhausted. |
| 182 | */ |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 183 | static int __bt_get(struct blk_mq_hw_ctx *hctx, struct blk_mq_bitmap_tags *bt, |
| 184 | unsigned int *tag_cache) |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 185 | { |
| 186 | unsigned int last_tag, org_last_tag; |
| 187 | int index, i, tag; |
| 188 | |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 189 | if (!hctx_may_queue(hctx, bt)) |
| 190 | return -1; |
| 191 | |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 192 | last_tag = org_last_tag = *tag_cache; |
Jens Axboe | 59d13bf | 2014-05-09 13:41:15 -0600 | [diff] [blame] | 193 | index = TAG_TO_INDEX(bt, last_tag); |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 194 | |
| 195 | for (i = 0; i < bt->map_nr; i++) { |
Jens Axboe | 59d13bf | 2014-05-09 13:41:15 -0600 | [diff] [blame] | 196 | tag = __bt_get_word(&bt->map[index], TAG_TO_BIT(bt, last_tag)); |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 197 | if (tag != -1) { |
Jens Axboe | 59d13bf | 2014-05-09 13:41:15 -0600 | [diff] [blame] | 198 | tag += (index << bt->bits_per_word); |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 199 | goto done; |
| 200 | } |
| 201 | |
| 202 | last_tag = 0; |
| 203 | if (++index >= bt->map_nr) |
| 204 | index = 0; |
| 205 | } |
| 206 | |
| 207 | *tag_cache = 0; |
| 208 | return -1; |
| 209 | |
| 210 | /* |
| 211 | * Only update the cache from the allocation path, if we ended |
| 212 | * up using the specific cached tag. |
| 213 | */ |
| 214 | done: |
| 215 | if (tag == org_last_tag) { |
| 216 | last_tag = tag + 1; |
| 217 | if (last_tag >= bt->depth - 1) |
| 218 | last_tag = 0; |
| 219 | |
| 220 | *tag_cache = last_tag; |
| 221 | } |
| 222 | |
| 223 | return tag; |
| 224 | } |
| 225 | |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 226 | static struct bt_wait_state *bt_wait_ptr(struct blk_mq_bitmap_tags *bt, |
| 227 | struct blk_mq_hw_ctx *hctx) |
| 228 | { |
| 229 | struct bt_wait_state *bs; |
Alexander Gordeev | 8537b12 | 2014-06-17 22:12:35 -0700 | [diff] [blame] | 230 | int wait_index; |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 231 | |
| 232 | if (!hctx) |
| 233 | return &bt->bs[0]; |
| 234 | |
Alexander Gordeev | 8537b12 | 2014-06-17 22:12:35 -0700 | [diff] [blame] | 235 | wait_index = atomic_read(&hctx->wait_index); |
| 236 | bs = &bt->bs[wait_index]; |
| 237 | bt_index_atomic_inc(&hctx->wait_index); |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 238 | return bs; |
| 239 | } |
| 240 | |
Ming Lei | cb96a42 | 2014-06-01 00:43:37 +0800 | [diff] [blame] | 241 | static int bt_get(struct blk_mq_alloc_data *data, |
| 242 | struct blk_mq_bitmap_tags *bt, |
| 243 | struct blk_mq_hw_ctx *hctx, |
| 244 | unsigned int *last_tag) |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 245 | { |
| 246 | struct bt_wait_state *bs; |
| 247 | DEFINE_WAIT(wait); |
| 248 | int tag; |
| 249 | |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 250 | tag = __bt_get(hctx, bt, last_tag); |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 251 | if (tag != -1) |
| 252 | return tag; |
| 253 | |
Ming Lei | cb96a42 | 2014-06-01 00:43:37 +0800 | [diff] [blame] | 254 | if (!(data->gfp & __GFP_WAIT)) |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 255 | return -1; |
| 256 | |
Jens Axboe | 35d37c6 | 2014-12-15 08:30:26 -0700 | [diff] [blame] | 257 | bs = bt_wait_ptr(bt, hctx); |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 258 | do { |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 259 | prepare_to_wait(&bs->wait, &wait, TASK_UNINTERRUPTIBLE); |
| 260 | |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 261 | tag = __bt_get(hctx, bt, last_tag); |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 262 | if (tag != -1) |
| 263 | break; |
| 264 | |
Bart Van Assche | b322320 | 2014-12-08 08:46:34 -0700 | [diff] [blame] | 265 | /* |
| 266 | * We're out of tags on this hardware queue, kick any |
| 267 | * pending IO submits before going to sleep waiting for |
| 268 | * some to complete. |
| 269 | */ |
| 270 | blk_mq_run_hw_queue(hctx, false); |
| 271 | |
Jens Axboe | 080ff35 | 2014-12-08 08:49:06 -0700 | [diff] [blame] | 272 | /* |
| 273 | * Retry tag allocation after running the hardware queue, |
| 274 | * as running the queue may also have found completions. |
| 275 | */ |
| 276 | tag = __bt_get(hctx, bt, last_tag); |
| 277 | if (tag != -1) |
| 278 | break; |
| 279 | |
Ming Lei | cb96a42 | 2014-06-01 00:43:37 +0800 | [diff] [blame] | 280 | blk_mq_put_ctx(data->ctx); |
| 281 | |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 282 | io_schedule(); |
Ming Lei | cb96a42 | 2014-06-01 00:43:37 +0800 | [diff] [blame] | 283 | |
| 284 | data->ctx = blk_mq_get_ctx(data->q); |
| 285 | data->hctx = data->q->mq_ops->map_queue(data->q, |
| 286 | data->ctx->cpu); |
| 287 | if (data->reserved) { |
| 288 | bt = &data->hctx->tags->breserved_tags; |
| 289 | } else { |
| 290 | last_tag = &data->ctx->last_tag; |
| 291 | hctx = data->hctx; |
| 292 | bt = &hctx->tags->bitmap_tags; |
| 293 | } |
Jens Axboe | 35d37c6 | 2014-12-15 08:30:26 -0700 | [diff] [blame] | 294 | finish_wait(&bs->wait, &wait); |
| 295 | bs = bt_wait_ptr(bt, hctx); |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 296 | } while (1); |
| 297 | |
| 298 | finish_wait(&bs->wait, &wait); |
| 299 | return tag; |
| 300 | } |
| 301 | |
Ming Lei | cb96a42 | 2014-06-01 00:43:37 +0800 | [diff] [blame] | 302 | static unsigned int __blk_mq_get_tag(struct blk_mq_alloc_data *data) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 303 | { |
| 304 | int tag; |
| 305 | |
Ming Lei | cb96a42 | 2014-06-01 00:43:37 +0800 | [diff] [blame] | 306 | tag = bt_get(data, &data->hctx->tags->bitmap_tags, data->hctx, |
| 307 | &data->ctx->last_tag); |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 308 | if (tag >= 0) |
Ming Lei | cb96a42 | 2014-06-01 00:43:37 +0800 | [diff] [blame] | 309 | return tag + data->hctx->tags->nr_reserved_tags; |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 310 | |
| 311 | return BLK_MQ_TAG_FAIL; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 312 | } |
| 313 | |
Ming Lei | cb96a42 | 2014-06-01 00:43:37 +0800 | [diff] [blame] | 314 | static unsigned int __blk_mq_get_reserved_tag(struct blk_mq_alloc_data *data) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 315 | { |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 316 | int tag, zero = 0; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 317 | |
Ming Lei | cb96a42 | 2014-06-01 00:43:37 +0800 | [diff] [blame] | 318 | if (unlikely(!data->hctx->tags->nr_reserved_tags)) { |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 319 | WARN_ON_ONCE(1); |
| 320 | return BLK_MQ_TAG_FAIL; |
| 321 | } |
| 322 | |
Ming Lei | cb96a42 | 2014-06-01 00:43:37 +0800 | [diff] [blame] | 323 | tag = bt_get(data, &data->hctx->tags->breserved_tags, NULL, &zero); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 324 | if (tag < 0) |
| 325 | return BLK_MQ_TAG_FAIL; |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 326 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 327 | return tag; |
| 328 | } |
| 329 | |
Ming Lei | cb96a42 | 2014-06-01 00:43:37 +0800 | [diff] [blame] | 330 | unsigned int blk_mq_get_tag(struct blk_mq_alloc_data *data) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 331 | { |
Ming Lei | cb96a42 | 2014-06-01 00:43:37 +0800 | [diff] [blame] | 332 | if (!data->reserved) |
| 333 | return __blk_mq_get_tag(data); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 334 | |
Ming Lei | cb96a42 | 2014-06-01 00:43:37 +0800 | [diff] [blame] | 335 | return __blk_mq_get_reserved_tag(data); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 336 | } |
| 337 | |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 338 | static struct bt_wait_state *bt_wake_ptr(struct blk_mq_bitmap_tags *bt) |
| 339 | { |
| 340 | int i, wake_index; |
| 341 | |
Alexander Gordeev | 8537b12 | 2014-06-17 22:12:35 -0700 | [diff] [blame] | 342 | wake_index = atomic_read(&bt->wake_index); |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 343 | for (i = 0; i < BT_WAIT_QUEUES; i++) { |
| 344 | struct bt_wait_state *bs = &bt->bs[wake_index]; |
| 345 | |
| 346 | if (waitqueue_active(&bs->wait)) { |
Alexander Gordeev | 8537b12 | 2014-06-17 22:12:35 -0700 | [diff] [blame] | 347 | int o = atomic_read(&bt->wake_index); |
| 348 | if (wake_index != o) |
| 349 | atomic_cmpxchg(&bt->wake_index, o, wake_index); |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 350 | |
| 351 | return bs; |
| 352 | } |
| 353 | |
Alexander Gordeev | 8537b12 | 2014-06-17 22:12:35 -0700 | [diff] [blame] | 354 | wake_index = bt_index_inc(wake_index); |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 355 | } |
| 356 | |
| 357 | return NULL; |
| 358 | } |
| 359 | |
| 360 | static void bt_clear_tag(struct blk_mq_bitmap_tags *bt, unsigned int tag) |
| 361 | { |
Jens Axboe | 59d13bf | 2014-05-09 13:41:15 -0600 | [diff] [blame] | 362 | const int index = TAG_TO_INDEX(bt, tag); |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 363 | struct bt_wait_state *bs; |
Alexander Gordeev | 2971c35 | 2014-06-12 17:05:37 +0200 | [diff] [blame] | 364 | int wait_cnt; |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 365 | |
Bart Van Assche | c38d185 | 2014-12-09 16:58:35 +0100 | [diff] [blame] | 366 | clear_bit(TAG_TO_BIT(bt, tag), &bt->map[index].word); |
| 367 | |
| 368 | /* Ensure that the wait list checks occur after clear_bit(). */ |
| 369 | smp_mb(); |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 370 | |
| 371 | bs = bt_wake_ptr(bt); |
Alexander Gordeev | 2971c35 | 2014-06-12 17:05:37 +0200 | [diff] [blame] | 372 | if (!bs) |
| 373 | return; |
| 374 | |
| 375 | wait_cnt = atomic_dec_return(&bs->wait_cnt); |
Bart Van Assche | 9d8f0bc | 2014-10-07 08:45:21 -0600 | [diff] [blame] | 376 | if (unlikely(wait_cnt < 0)) |
| 377 | wait_cnt = atomic_inc_return(&bs->wait_cnt); |
Alexander Gordeev | 2971c35 | 2014-06-12 17:05:37 +0200 | [diff] [blame] | 378 | if (wait_cnt == 0) { |
Alexander Gordeev | 2971c35 | 2014-06-12 17:05:37 +0200 | [diff] [blame] | 379 | atomic_add(bt->wake_cnt, &bs->wait_cnt); |
Alexander Gordeev | 8537b12 | 2014-06-17 22:12:35 -0700 | [diff] [blame] | 380 | bt_index_atomic_inc(&bt->wake_index); |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 381 | wake_up(&bs->wait); |
| 382 | } |
| 383 | } |
| 384 | |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 385 | void blk_mq_put_tag(struct blk_mq_hw_ctx *hctx, unsigned int tag, |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 386 | unsigned int *last_tag) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 387 | { |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 388 | struct blk_mq_tags *tags = hctx->tags; |
| 389 | |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 390 | if (tag >= tags->nr_reserved_tags) { |
| 391 | const int real_tag = tag - tags->nr_reserved_tags; |
| 392 | |
Jens Axboe | 70114c3 | 2014-11-24 15:52:30 -0700 | [diff] [blame] | 393 | BUG_ON(real_tag >= tags->nr_tags); |
| 394 | bt_clear_tag(&tags->bitmap_tags, real_tag); |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 395 | *last_tag = real_tag; |
Jens Axboe | 70114c3 | 2014-11-24 15:52:30 -0700 | [diff] [blame] | 396 | } else { |
| 397 | BUG_ON(tag >= tags->nr_reserved_tags); |
| 398 | bt_clear_tag(&tags->breserved_tags, tag); |
| 399 | } |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 400 | } |
| 401 | |
Christoph Hellwig | 81481eb | 2014-09-13 16:40:11 -0700 | [diff] [blame] | 402 | static void bt_for_each(struct blk_mq_hw_ctx *hctx, |
| 403 | struct blk_mq_bitmap_tags *bt, unsigned int off, |
| 404 | busy_iter_fn *fn, void *data, bool reserved) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 405 | { |
Christoph Hellwig | 81481eb | 2014-09-13 16:40:11 -0700 | [diff] [blame] | 406 | struct request *rq; |
| 407 | int bit, i; |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 408 | |
| 409 | for (i = 0; i < bt->map_nr; i++) { |
Jens Axboe | e93ecf6 | 2014-05-19 09:17:48 -0600 | [diff] [blame] | 410 | struct blk_align_bitmap *bm = &bt->map[i]; |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 411 | |
Christoph Hellwig | 81481eb | 2014-09-13 16:40:11 -0700 | [diff] [blame] | 412 | for (bit = find_first_bit(&bm->word, bm->depth); |
| 413 | bit < bm->depth; |
| 414 | bit = find_next_bit(&bm->word, bm->depth, bit + 1)) { |
| 415 | rq = blk_mq_tag_to_rq(hctx->tags, off + bit); |
| 416 | if (rq->q == hctx->queue) |
| 417 | fn(hctx, rq, data, reserved); |
| 418 | } |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 419 | |
Jens Axboe | 59d13bf | 2014-05-09 13:41:15 -0600 | [diff] [blame] | 420 | off += (1 << bt->bits_per_word); |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 421 | } |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 422 | } |
| 423 | |
Christoph Hellwig | 81481eb | 2014-09-13 16:40:11 -0700 | [diff] [blame] | 424 | void blk_mq_tag_busy_iter(struct blk_mq_hw_ctx *hctx, busy_iter_fn *fn, |
| 425 | void *priv) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 426 | { |
Christoph Hellwig | 81481eb | 2014-09-13 16:40:11 -0700 | [diff] [blame] | 427 | struct blk_mq_tags *tags = hctx->tags; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 428 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 429 | if (tags->nr_reserved_tags) |
Christoph Hellwig | 81481eb | 2014-09-13 16:40:11 -0700 | [diff] [blame] | 430 | bt_for_each(hctx, &tags->breserved_tags, 0, fn, priv, true); |
| 431 | bt_for_each(hctx, &tags->bitmap_tags, tags->nr_reserved_tags, fn, priv, |
| 432 | false); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 433 | } |
Sam Bradshaw | edf866b | 2014-05-23 13:30:16 -0600 | [diff] [blame] | 434 | EXPORT_SYMBOL(blk_mq_tag_busy_iter); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 435 | |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 436 | static unsigned int bt_unused_tags(struct blk_mq_bitmap_tags *bt) |
| 437 | { |
| 438 | unsigned int i, used; |
| 439 | |
| 440 | for (i = 0, used = 0; i < bt->map_nr; i++) { |
Jens Axboe | e93ecf6 | 2014-05-19 09:17:48 -0600 | [diff] [blame] | 441 | struct blk_align_bitmap *bm = &bt->map[i]; |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 442 | |
| 443 | used += bitmap_weight(&bm->word, bm->depth); |
| 444 | } |
| 445 | |
| 446 | return bt->depth - used; |
| 447 | } |
| 448 | |
Jens Axboe | e3a2b3f | 2014-05-20 11:49:02 -0600 | [diff] [blame] | 449 | static void bt_update_count(struct blk_mq_bitmap_tags *bt, |
| 450 | unsigned int depth) |
| 451 | { |
| 452 | unsigned int tags_per_word = 1U << bt->bits_per_word; |
| 453 | unsigned int map_depth = depth; |
| 454 | |
| 455 | if (depth) { |
| 456 | int i; |
| 457 | |
| 458 | for (i = 0; i < bt->map_nr; i++) { |
| 459 | bt->map[i].depth = min(map_depth, tags_per_word); |
| 460 | map_depth -= bt->map[i].depth; |
| 461 | } |
| 462 | } |
| 463 | |
| 464 | bt->wake_cnt = BT_WAIT_BATCH; |
Jens Axboe | abab13b | 2014-10-07 08:39:20 -0600 | [diff] [blame] | 465 | if (bt->wake_cnt > depth / BT_WAIT_QUEUES) |
| 466 | bt->wake_cnt = max(1U, depth / BT_WAIT_QUEUES); |
Jens Axboe | e3a2b3f | 2014-05-20 11:49:02 -0600 | [diff] [blame] | 467 | |
| 468 | bt->depth = depth; |
| 469 | } |
| 470 | |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 471 | static int bt_alloc(struct blk_mq_bitmap_tags *bt, unsigned int depth, |
| 472 | int node, bool reserved) |
| 473 | { |
| 474 | int i; |
| 475 | |
Jens Axboe | 59d13bf | 2014-05-09 13:41:15 -0600 | [diff] [blame] | 476 | bt->bits_per_word = ilog2(BITS_PER_LONG); |
| 477 | |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 478 | /* |
| 479 | * Depth can be zero for reserved tags, that's not a failure |
| 480 | * condition. |
| 481 | */ |
| 482 | if (depth) { |
Jens Axboe | e3a2b3f | 2014-05-20 11:49:02 -0600 | [diff] [blame] | 483 | unsigned int nr, tags_per_word; |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 484 | |
Jens Axboe | 59d13bf | 2014-05-09 13:41:15 -0600 | [diff] [blame] | 485 | tags_per_word = (1 << bt->bits_per_word); |
| 486 | |
| 487 | /* |
| 488 | * If the tag space is small, shrink the number of tags |
| 489 | * per word so we spread over a few cachelines, at least. |
| 490 | * If less than 4 tags, just forget about it, it's not |
| 491 | * going to work optimally anyway. |
| 492 | */ |
| 493 | if (depth >= 4) { |
| 494 | while (tags_per_word * 4 > depth) { |
| 495 | bt->bits_per_word--; |
| 496 | tags_per_word = (1 << bt->bits_per_word); |
| 497 | } |
| 498 | } |
| 499 | |
| 500 | nr = ALIGN(depth, tags_per_word) / tags_per_word; |
Jens Axboe | e93ecf6 | 2014-05-19 09:17:48 -0600 | [diff] [blame] | 501 | bt->map = kzalloc_node(nr * sizeof(struct blk_align_bitmap), |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 502 | GFP_KERNEL, node); |
| 503 | if (!bt->map) |
| 504 | return -ENOMEM; |
| 505 | |
| 506 | bt->map_nr = nr; |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 507 | } |
| 508 | |
| 509 | bt->bs = kzalloc(BT_WAIT_QUEUES * sizeof(*bt->bs), GFP_KERNEL); |
| 510 | if (!bt->bs) { |
| 511 | kfree(bt->map); |
| 512 | return -ENOMEM; |
| 513 | } |
| 514 | |
Jens Axboe | e3a2b3f | 2014-05-20 11:49:02 -0600 | [diff] [blame] | 515 | bt_update_count(bt, depth); |
Alexander Gordeev | 86fb5c5 | 2014-06-17 22:37:23 +0200 | [diff] [blame] | 516 | |
| 517 | for (i = 0; i < BT_WAIT_QUEUES; i++) { |
| 518 | init_waitqueue_head(&bt->bs[i].wait); |
| 519 | atomic_set(&bt->bs[i].wait_cnt, bt->wake_cnt); |
| 520 | } |
| 521 | |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 522 | return 0; |
| 523 | } |
| 524 | |
| 525 | static void bt_free(struct blk_mq_bitmap_tags *bt) |
| 526 | { |
| 527 | kfree(bt->map); |
| 528 | kfree(bt->bs); |
| 529 | } |
| 530 | |
| 531 | static struct blk_mq_tags *blk_mq_init_bitmap_tags(struct blk_mq_tags *tags, |
| 532 | int node) |
| 533 | { |
| 534 | unsigned int depth = tags->nr_tags - tags->nr_reserved_tags; |
| 535 | |
| 536 | if (bt_alloc(&tags->bitmap_tags, depth, node, false)) |
| 537 | goto enomem; |
| 538 | if (bt_alloc(&tags->breserved_tags, tags->nr_reserved_tags, node, true)) |
| 539 | goto enomem; |
| 540 | |
| 541 | return tags; |
| 542 | enomem: |
| 543 | bt_free(&tags->bitmap_tags); |
| 544 | kfree(tags); |
| 545 | return NULL; |
| 546 | } |
| 547 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 548 | struct blk_mq_tags *blk_mq_init_tags(unsigned int total_tags, |
| 549 | unsigned int reserved_tags, int node) |
| 550 | { |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 551 | struct blk_mq_tags *tags; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 552 | |
| 553 | if (total_tags > BLK_MQ_TAG_MAX) { |
| 554 | pr_err("blk-mq: tag depth too large\n"); |
| 555 | return NULL; |
| 556 | } |
| 557 | |
| 558 | tags = kzalloc_node(sizeof(*tags), GFP_KERNEL, node); |
| 559 | if (!tags) |
| 560 | return NULL; |
| 561 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 562 | tags->nr_tags = total_tags; |
| 563 | tags->nr_reserved_tags = reserved_tags; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 564 | |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 565 | return blk_mq_init_bitmap_tags(tags, node); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 566 | } |
| 567 | |
| 568 | void blk_mq_free_tags(struct blk_mq_tags *tags) |
| 569 | { |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 570 | bt_free(&tags->bitmap_tags); |
| 571 | bt_free(&tags->breserved_tags); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 572 | kfree(tags); |
| 573 | } |
| 574 | |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 575 | void blk_mq_tag_init_last_tag(struct blk_mq_tags *tags, unsigned int *tag) |
| 576 | { |
| 577 | unsigned int depth = tags->nr_tags - tags->nr_reserved_tags; |
| 578 | |
Ming Lei | 9d3d21a | 2014-05-10 15:43:14 -0600 | [diff] [blame] | 579 | *tag = prandom_u32() % depth; |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 580 | } |
| 581 | |
Jens Axboe | e3a2b3f | 2014-05-20 11:49:02 -0600 | [diff] [blame] | 582 | int blk_mq_tag_update_depth(struct blk_mq_tags *tags, unsigned int tdepth) |
| 583 | { |
| 584 | tdepth -= tags->nr_reserved_tags; |
| 585 | if (tdepth > tags->nr_tags) |
| 586 | return -EINVAL; |
| 587 | |
| 588 | /* |
| 589 | * Don't need (or can't) update reserved tags here, they remain |
| 590 | * static and should never need resizing. |
| 591 | */ |
| 592 | bt_update_count(&tags->bitmap_tags, tdepth); |
Jens Axboe | aed3ea9 | 2014-12-22 14:04:42 -0700 | [diff] [blame^] | 593 | blk_mq_tag_wakeup_all(tags, false); |
Jens Axboe | e3a2b3f | 2014-05-20 11:49:02 -0600 | [diff] [blame] | 594 | return 0; |
| 595 | } |
| 596 | |
Bart Van Assche | 205fb5f | 2014-10-30 14:45:11 +0100 | [diff] [blame] | 597 | /** |
| 598 | * blk_mq_unique_tag() - return a tag that is unique queue-wide |
| 599 | * @rq: request for which to compute a unique tag |
| 600 | * |
| 601 | * The tag field in struct request is unique per hardware queue but not over |
| 602 | * all hardware queues. Hence this function that returns a tag with the |
| 603 | * hardware context index in the upper bits and the per hardware queue tag in |
| 604 | * the lower bits. |
| 605 | * |
| 606 | * Note: When called for a request that is queued on a non-multiqueue request |
| 607 | * queue, the hardware context index is set to zero. |
| 608 | */ |
| 609 | u32 blk_mq_unique_tag(struct request *rq) |
| 610 | { |
| 611 | struct request_queue *q = rq->q; |
| 612 | struct blk_mq_hw_ctx *hctx; |
| 613 | int hwq = 0; |
| 614 | |
| 615 | if (q->mq_ops) { |
| 616 | hctx = q->mq_ops->map_queue(q, rq->mq_ctx->cpu); |
| 617 | hwq = hctx->queue_num; |
| 618 | } |
| 619 | |
| 620 | return (hwq << BLK_MQ_UNIQUE_TAG_BITS) | |
| 621 | (rq->tag & BLK_MQ_UNIQUE_TAG_MASK); |
| 622 | } |
| 623 | EXPORT_SYMBOL(blk_mq_unique_tag); |
| 624 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 625 | ssize_t blk_mq_tag_sysfs_show(struct blk_mq_tags *tags, char *page) |
| 626 | { |
| 627 | char *orig_page = page; |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 628 | unsigned int free, res; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 629 | |
| 630 | if (!tags) |
| 631 | return 0; |
| 632 | |
Jens Axboe | 59d13bf | 2014-05-09 13:41:15 -0600 | [diff] [blame] | 633 | page += sprintf(page, "nr_tags=%u, reserved_tags=%u, " |
| 634 | "bits_per_word=%u\n", |
| 635 | tags->nr_tags, tags->nr_reserved_tags, |
| 636 | tags->bitmap_tags.bits_per_word); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 637 | |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 638 | free = bt_unused_tags(&tags->bitmap_tags); |
| 639 | res = bt_unused_tags(&tags->breserved_tags); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 640 | |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 641 | page += sprintf(page, "nr_free=%u, nr_reserved=%u\n", free, res); |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 642 | page += sprintf(page, "active_queues=%u\n", atomic_read(&tags->active_queues)); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 643 | |
| 644 | return page - orig_page; |
| 645 | } |