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 | |
Kosuke Tatsukawa | 8ee1b7b | 2015-10-09 00:35:38 +0000 | [diff] [blame] | 78 | /* |
| 79 | * Make sure all changes prior to this are visible from other CPUs. |
| 80 | */ |
| 81 | smp_mb(); |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 82 | bt = &tags->bitmap_tags; |
Alexander Gordeev | 8537b12 | 2014-06-17 22:12:35 -0700 | [diff] [blame] | 83 | wake_index = atomic_read(&bt->wake_index); |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 84 | for (i = 0; i < BT_WAIT_QUEUES; i++) { |
| 85 | struct bt_wait_state *bs = &bt->bs[wake_index]; |
| 86 | |
| 87 | if (waitqueue_active(&bs->wait)) |
| 88 | wake_up(&bs->wait); |
| 89 | |
Alexander Gordeev | 8537b12 | 2014-06-17 22:12:35 -0700 | [diff] [blame] | 90 | wake_index = bt_index_inc(wake_index); |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 91 | } |
Jens Axboe | aed3ea9 | 2014-12-22 14:04:42 -0700 | [diff] [blame] | 92 | |
| 93 | if (include_reserve) { |
| 94 | bt = &tags->breserved_tags; |
| 95 | if (waitqueue_active(&bt->bs[0].wait)) |
| 96 | wake_up(&bt->bs[0].wait); |
| 97 | } |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 98 | } |
| 99 | |
| 100 | /* |
Jens Axboe | e3a2b3f | 2014-05-20 11:49:02 -0600 | [diff] [blame] | 101 | * If a previously busy queue goes inactive, potential waiters could now |
| 102 | * be allowed to queue. Wake them up and check. |
| 103 | */ |
| 104 | void __blk_mq_tag_idle(struct blk_mq_hw_ctx *hctx) |
| 105 | { |
| 106 | struct blk_mq_tags *tags = hctx->tags; |
| 107 | |
| 108 | if (!test_and_clear_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state)) |
| 109 | return; |
| 110 | |
| 111 | atomic_dec(&tags->active_queues); |
| 112 | |
Jens Axboe | aed3ea9 | 2014-12-22 14:04:42 -0700 | [diff] [blame] | 113 | blk_mq_tag_wakeup_all(tags, false); |
Jens Axboe | e3a2b3f | 2014-05-20 11:49:02 -0600 | [diff] [blame] | 114 | } |
| 115 | |
| 116 | /* |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 117 | * For shared tag users, we track the number of currently active users |
| 118 | * and attempt to provide a fair share of the tag depth for each of them. |
| 119 | */ |
| 120 | static inline bool hctx_may_queue(struct blk_mq_hw_ctx *hctx, |
| 121 | struct blk_mq_bitmap_tags *bt) |
| 122 | { |
| 123 | unsigned int depth, users; |
| 124 | |
| 125 | if (!hctx || !(hctx->flags & BLK_MQ_F_TAG_SHARED)) |
| 126 | return true; |
| 127 | if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state)) |
| 128 | return true; |
| 129 | |
| 130 | /* |
| 131 | * Don't try dividing an ant |
| 132 | */ |
| 133 | if (bt->depth == 1) |
| 134 | return true; |
| 135 | |
| 136 | users = atomic_read(&hctx->tags->active_queues); |
| 137 | if (!users) |
| 138 | return true; |
| 139 | |
| 140 | /* |
| 141 | * Allow at least some tags |
| 142 | */ |
| 143 | depth = max((bt->depth + users - 1) / users, 4U); |
| 144 | return atomic_read(&hctx->nr_active) < depth; |
| 145 | } |
| 146 | |
Shaohua Li | 24391c0 | 2015-01-23 14:18:00 -0700 | [diff] [blame] | 147 | static int __bt_get_word(struct blk_align_bitmap *bm, unsigned int last_tag, |
| 148 | bool nowrap) |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 149 | { |
Jens Axboe | 0bf3649 | 2015-01-14 08:49:55 -0700 | [diff] [blame] | 150 | int tag, org_last_tag = last_tag; |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 151 | |
Jens Axboe | 0bf3649 | 2015-01-14 08:49:55 -0700 | [diff] [blame] | 152 | while (1) { |
| 153 | tag = find_next_zero_bit(&bm->word, bm->depth, last_tag); |
| 154 | if (unlikely(tag >= bm->depth)) { |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 155 | /* |
Jens Axboe | 0bf3649 | 2015-01-14 08:49:55 -0700 | [diff] [blame] | 156 | * We started with an offset, and we didn't reset the |
| 157 | * offset to 0 in a failure case, so start from 0 to |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 158 | * exhaust the map. |
| 159 | */ |
Shaohua Li | 24391c0 | 2015-01-23 14:18:00 -0700 | [diff] [blame] | 160 | if (org_last_tag && last_tag && !nowrap) { |
Jens Axboe | 0bf3649 | 2015-01-14 08:49:55 -0700 | [diff] [blame] | 161 | last_tag = org_last_tag = 0; |
| 162 | continue; |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 163 | } |
| 164 | return -1; |
| 165 | } |
Jens Axboe | 0bf3649 | 2015-01-14 08:49:55 -0700 | [diff] [blame] | 166 | |
| 167 | if (!test_and_set_bit(tag, &bm->word)) |
| 168 | break; |
| 169 | |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 170 | last_tag = tag + 1; |
Jens Axboe | 0bf3649 | 2015-01-14 08:49:55 -0700 | [diff] [blame] | 171 | if (last_tag >= bm->depth - 1) |
| 172 | last_tag = 0; |
| 173 | } |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 174 | |
| 175 | return tag; |
| 176 | } |
| 177 | |
Shaohua Li | 24391c0 | 2015-01-23 14:18:00 -0700 | [diff] [blame] | 178 | #define BT_ALLOC_RR(tags) (tags->alloc_policy == BLK_TAG_ALLOC_RR) |
| 179 | |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 180 | /* |
| 181 | * Straight forward bitmap tag implementation, where each bit is a tag |
| 182 | * (cleared == free, and set == busy). The small twist is using per-cpu |
| 183 | * last_tag caches, which blk-mq stores in the blk_mq_ctx software queue |
| 184 | * contexts. This enables us to drastically limit the space searched, |
| 185 | * without dirtying an extra shared cacheline like we would if we stored |
| 186 | * the cache value inside the shared blk_mq_bitmap_tags structure. On top |
| 187 | * of that, each word of tags is in a separate cacheline. This means that |
| 188 | * multiple users will tend to stick to different cachelines, at least |
| 189 | * until the map is exhausted. |
| 190 | */ |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 191 | static int __bt_get(struct blk_mq_hw_ctx *hctx, struct blk_mq_bitmap_tags *bt, |
Shaohua Li | 24391c0 | 2015-01-23 14:18:00 -0700 | [diff] [blame] | 192 | unsigned int *tag_cache, struct blk_mq_tags *tags) |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 193 | { |
| 194 | unsigned int last_tag, org_last_tag; |
| 195 | int index, i, tag; |
| 196 | |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 197 | if (!hctx_may_queue(hctx, bt)) |
| 198 | return -1; |
| 199 | |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 200 | last_tag = org_last_tag = *tag_cache; |
Jens Axboe | 59d13bf | 2014-05-09 13:41:15 -0600 | [diff] [blame] | 201 | index = TAG_TO_INDEX(bt, last_tag); |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 202 | |
| 203 | for (i = 0; i < bt->map_nr; i++) { |
Shaohua Li | 24391c0 | 2015-01-23 14:18:00 -0700 | [diff] [blame] | 204 | tag = __bt_get_word(&bt->map[index], TAG_TO_BIT(bt, last_tag), |
| 205 | BT_ALLOC_RR(tags)); |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 206 | if (tag != -1) { |
Jens Axboe | 59d13bf | 2014-05-09 13:41:15 -0600 | [diff] [blame] | 207 | tag += (index << bt->bits_per_word); |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 208 | goto done; |
| 209 | } |
| 210 | |
Jens Axboe | 0bf3649 | 2015-01-14 08:49:55 -0700 | [diff] [blame] | 211 | /* |
| 212 | * Jump to next index, and reset the last tag to be the |
| 213 | * first tag of that index |
| 214 | */ |
| 215 | index++; |
| 216 | last_tag = (index << bt->bits_per_word); |
| 217 | |
| 218 | if (index >= bt->map_nr) { |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 219 | index = 0; |
Jens Axboe | 0bf3649 | 2015-01-14 08:49:55 -0700 | [diff] [blame] | 220 | last_tag = 0; |
| 221 | } |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 222 | } |
| 223 | |
| 224 | *tag_cache = 0; |
| 225 | return -1; |
| 226 | |
| 227 | /* |
| 228 | * Only update the cache from the allocation path, if we ended |
| 229 | * up using the specific cached tag. |
| 230 | */ |
| 231 | done: |
Shaohua Li | 24391c0 | 2015-01-23 14:18:00 -0700 | [diff] [blame] | 232 | if (tag == org_last_tag || unlikely(BT_ALLOC_RR(tags))) { |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 233 | last_tag = tag + 1; |
| 234 | if (last_tag >= bt->depth - 1) |
| 235 | last_tag = 0; |
| 236 | |
| 237 | *tag_cache = last_tag; |
| 238 | } |
| 239 | |
| 240 | return tag; |
| 241 | } |
| 242 | |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 243 | static struct bt_wait_state *bt_wait_ptr(struct blk_mq_bitmap_tags *bt, |
| 244 | struct blk_mq_hw_ctx *hctx) |
| 245 | { |
| 246 | struct bt_wait_state *bs; |
Alexander Gordeev | 8537b12 | 2014-06-17 22:12:35 -0700 | [diff] [blame] | 247 | int wait_index; |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 248 | |
| 249 | if (!hctx) |
| 250 | return &bt->bs[0]; |
| 251 | |
Alexander Gordeev | 8537b12 | 2014-06-17 22:12:35 -0700 | [diff] [blame] | 252 | wait_index = atomic_read(&hctx->wait_index); |
| 253 | bs = &bt->bs[wait_index]; |
| 254 | bt_index_atomic_inc(&hctx->wait_index); |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 255 | return bs; |
| 256 | } |
| 257 | |
Ming Lei | cb96a42 | 2014-06-01 00:43:37 +0800 | [diff] [blame] | 258 | static int bt_get(struct blk_mq_alloc_data *data, |
| 259 | struct blk_mq_bitmap_tags *bt, |
| 260 | struct blk_mq_hw_ctx *hctx, |
Shaohua Li | 24391c0 | 2015-01-23 14:18:00 -0700 | [diff] [blame] | 261 | unsigned int *last_tag, struct blk_mq_tags *tags) |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 262 | { |
| 263 | struct bt_wait_state *bs; |
| 264 | DEFINE_WAIT(wait); |
| 265 | int tag; |
| 266 | |
Shaohua Li | 24391c0 | 2015-01-23 14:18:00 -0700 | [diff] [blame] | 267 | tag = __bt_get(hctx, bt, last_tag, tags); |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 268 | if (tag != -1) |
| 269 | return tag; |
| 270 | |
Christoph Hellwig | 6f3b0e8 | 2015-11-26 09:13:05 +0100 | [diff] [blame] | 271 | if (data->flags & BLK_MQ_REQ_NOWAIT) |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 272 | return -1; |
| 273 | |
Jens Axboe | 35d37c6 | 2014-12-15 08:30:26 -0700 | [diff] [blame] | 274 | bs = bt_wait_ptr(bt, hctx); |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 275 | do { |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 276 | prepare_to_wait(&bs->wait, &wait, TASK_UNINTERRUPTIBLE); |
| 277 | |
Shaohua Li | 24391c0 | 2015-01-23 14:18:00 -0700 | [diff] [blame] | 278 | tag = __bt_get(hctx, bt, last_tag, tags); |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 279 | if (tag != -1) |
| 280 | break; |
| 281 | |
Bart Van Assche | b322320 | 2014-12-08 08:46:34 -0700 | [diff] [blame] | 282 | /* |
| 283 | * We're out of tags on this hardware queue, kick any |
| 284 | * pending IO submits before going to sleep waiting for |
Sam Bradshaw | bc188d8 | 2015-03-18 17:06:18 -0600 | [diff] [blame] | 285 | * some to complete. Note that hctx can be NULL here for |
| 286 | * reserved tag allocation. |
Bart Van Assche | b322320 | 2014-12-08 08:46:34 -0700 | [diff] [blame] | 287 | */ |
Sam Bradshaw | bc188d8 | 2015-03-18 17:06:18 -0600 | [diff] [blame] | 288 | if (hctx) |
| 289 | blk_mq_run_hw_queue(hctx, false); |
Bart Van Assche | b322320 | 2014-12-08 08:46:34 -0700 | [diff] [blame] | 290 | |
Jens Axboe | 080ff35 | 2014-12-08 08:49:06 -0700 | [diff] [blame] | 291 | /* |
| 292 | * Retry tag allocation after running the hardware queue, |
| 293 | * as running the queue may also have found completions. |
| 294 | */ |
Shaohua Li | 24391c0 | 2015-01-23 14:18:00 -0700 | [diff] [blame] | 295 | tag = __bt_get(hctx, bt, last_tag, tags); |
Jens Axboe | 080ff35 | 2014-12-08 08:49:06 -0700 | [diff] [blame] | 296 | if (tag != -1) |
| 297 | break; |
| 298 | |
Ming Lei | cb96a42 | 2014-06-01 00:43:37 +0800 | [diff] [blame] | 299 | blk_mq_put_ctx(data->ctx); |
| 300 | |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 301 | io_schedule(); |
Ming Lei | cb96a42 | 2014-06-01 00:43:37 +0800 | [diff] [blame] | 302 | |
| 303 | data->ctx = blk_mq_get_ctx(data->q); |
Christoph Hellwig | 7d7e0f9 | 2016-09-14 16:18:54 +0200 | [diff] [blame] | 304 | data->hctx = blk_mq_map_queue(data->q, data->ctx->cpu); |
Christoph Hellwig | 6f3b0e8 | 2015-11-26 09:13:05 +0100 | [diff] [blame] | 305 | if (data->flags & BLK_MQ_REQ_RESERVED) { |
Ming Lei | cb96a42 | 2014-06-01 00:43:37 +0800 | [diff] [blame] | 306 | bt = &data->hctx->tags->breserved_tags; |
| 307 | } else { |
| 308 | last_tag = &data->ctx->last_tag; |
| 309 | hctx = data->hctx; |
| 310 | bt = &hctx->tags->bitmap_tags; |
| 311 | } |
Jens Axboe | 35d37c6 | 2014-12-15 08:30:26 -0700 | [diff] [blame] | 312 | finish_wait(&bs->wait, &wait); |
| 313 | bs = bt_wait_ptr(bt, hctx); |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 314 | } while (1); |
| 315 | |
| 316 | finish_wait(&bs->wait, &wait); |
| 317 | return tag; |
| 318 | } |
| 319 | |
Ming Lei | cb96a42 | 2014-06-01 00:43:37 +0800 | [diff] [blame] | 320 | 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] | 321 | { |
| 322 | int tag; |
| 323 | |
Ming Lei | cb96a42 | 2014-06-01 00:43:37 +0800 | [diff] [blame] | 324 | tag = bt_get(data, &data->hctx->tags->bitmap_tags, data->hctx, |
Shaohua Li | 24391c0 | 2015-01-23 14:18:00 -0700 | [diff] [blame] | 325 | &data->ctx->last_tag, data->hctx->tags); |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 326 | if (tag >= 0) |
Ming Lei | cb96a42 | 2014-06-01 00:43:37 +0800 | [diff] [blame] | 327 | return tag + data->hctx->tags->nr_reserved_tags; |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 328 | |
| 329 | return BLK_MQ_TAG_FAIL; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 330 | } |
| 331 | |
Ming Lei | cb96a42 | 2014-06-01 00:43:37 +0800 | [diff] [blame] | 332 | 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] | 333 | { |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 334 | int tag, zero = 0; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 335 | |
Ming Lei | cb96a42 | 2014-06-01 00:43:37 +0800 | [diff] [blame] | 336 | if (unlikely(!data->hctx->tags->nr_reserved_tags)) { |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 337 | WARN_ON_ONCE(1); |
| 338 | return BLK_MQ_TAG_FAIL; |
| 339 | } |
| 340 | |
Shaohua Li | 24391c0 | 2015-01-23 14:18:00 -0700 | [diff] [blame] | 341 | tag = bt_get(data, &data->hctx->tags->breserved_tags, NULL, &zero, |
| 342 | data->hctx->tags); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 343 | if (tag < 0) |
| 344 | return BLK_MQ_TAG_FAIL; |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 345 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 346 | return tag; |
| 347 | } |
| 348 | |
Ming Lei | cb96a42 | 2014-06-01 00:43:37 +0800 | [diff] [blame] | 349 | unsigned int blk_mq_get_tag(struct blk_mq_alloc_data *data) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 350 | { |
Christoph Hellwig | 6f3b0e8 | 2015-11-26 09:13:05 +0100 | [diff] [blame] | 351 | if (data->flags & BLK_MQ_REQ_RESERVED) |
| 352 | return __blk_mq_get_reserved_tag(data); |
| 353 | return __blk_mq_get_tag(data); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 354 | } |
| 355 | |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 356 | static struct bt_wait_state *bt_wake_ptr(struct blk_mq_bitmap_tags *bt) |
| 357 | { |
| 358 | int i, wake_index; |
| 359 | |
Alexander Gordeev | 8537b12 | 2014-06-17 22:12:35 -0700 | [diff] [blame] | 360 | wake_index = atomic_read(&bt->wake_index); |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 361 | for (i = 0; i < BT_WAIT_QUEUES; i++) { |
| 362 | struct bt_wait_state *bs = &bt->bs[wake_index]; |
| 363 | |
| 364 | if (waitqueue_active(&bs->wait)) { |
Alexander Gordeev | 8537b12 | 2014-06-17 22:12:35 -0700 | [diff] [blame] | 365 | int o = atomic_read(&bt->wake_index); |
| 366 | if (wake_index != o) |
| 367 | atomic_cmpxchg(&bt->wake_index, o, wake_index); |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 368 | |
| 369 | return bs; |
| 370 | } |
| 371 | |
Alexander Gordeev | 8537b12 | 2014-06-17 22:12:35 -0700 | [diff] [blame] | 372 | wake_index = bt_index_inc(wake_index); |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 373 | } |
| 374 | |
| 375 | return NULL; |
| 376 | } |
| 377 | |
| 378 | static void bt_clear_tag(struct blk_mq_bitmap_tags *bt, unsigned int tag) |
| 379 | { |
Jens Axboe | 59d13bf | 2014-05-09 13:41:15 -0600 | [diff] [blame] | 380 | const int index = TAG_TO_INDEX(bt, tag); |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 381 | struct bt_wait_state *bs; |
Alexander Gordeev | 2971c35 | 2014-06-12 17:05:37 +0200 | [diff] [blame] | 382 | int wait_cnt; |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 383 | |
Bart Van Assche | c38d185 | 2014-12-09 16:58:35 +0100 | [diff] [blame] | 384 | clear_bit(TAG_TO_BIT(bt, tag), &bt->map[index].word); |
| 385 | |
| 386 | /* Ensure that the wait list checks occur after clear_bit(). */ |
| 387 | smp_mb(); |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 388 | |
| 389 | bs = bt_wake_ptr(bt); |
Alexander Gordeev | 2971c35 | 2014-06-12 17:05:37 +0200 | [diff] [blame] | 390 | if (!bs) |
| 391 | return; |
| 392 | |
| 393 | wait_cnt = atomic_dec_return(&bs->wait_cnt); |
Bart Van Assche | 9d8f0bc | 2014-10-07 08:45:21 -0600 | [diff] [blame] | 394 | if (unlikely(wait_cnt < 0)) |
| 395 | wait_cnt = atomic_inc_return(&bs->wait_cnt); |
Alexander Gordeev | 2971c35 | 2014-06-12 17:05:37 +0200 | [diff] [blame] | 396 | if (wait_cnt == 0) { |
Alexander Gordeev | 2971c35 | 2014-06-12 17:05:37 +0200 | [diff] [blame] | 397 | atomic_add(bt->wake_cnt, &bs->wait_cnt); |
Alexander Gordeev | 8537b12 | 2014-06-17 22:12:35 -0700 | [diff] [blame] | 398 | bt_index_atomic_inc(&bt->wake_index); |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 399 | wake_up(&bs->wait); |
| 400 | } |
| 401 | } |
| 402 | |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 403 | 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] | 404 | unsigned int *last_tag) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 405 | { |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 406 | struct blk_mq_tags *tags = hctx->tags; |
| 407 | |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 408 | if (tag >= tags->nr_reserved_tags) { |
| 409 | const int real_tag = tag - tags->nr_reserved_tags; |
| 410 | |
Jens Axboe | 70114c3 | 2014-11-24 15:52:30 -0700 | [diff] [blame] | 411 | BUG_ON(real_tag >= tags->nr_tags); |
| 412 | bt_clear_tag(&tags->bitmap_tags, real_tag); |
Shaohua Li | 24391c0 | 2015-01-23 14:18:00 -0700 | [diff] [blame] | 413 | if (likely(tags->alloc_policy == BLK_TAG_ALLOC_FIFO)) |
| 414 | *last_tag = real_tag; |
Jens Axboe | 70114c3 | 2014-11-24 15:52:30 -0700 | [diff] [blame] | 415 | } else { |
| 416 | BUG_ON(tag >= tags->nr_reserved_tags); |
| 417 | bt_clear_tag(&tags->breserved_tags, tag); |
| 418 | } |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 419 | } |
| 420 | |
Christoph Hellwig | 81481eb | 2014-09-13 16:40:11 -0700 | [diff] [blame] | 421 | static void bt_for_each(struct blk_mq_hw_ctx *hctx, |
| 422 | struct blk_mq_bitmap_tags *bt, unsigned int off, |
| 423 | busy_iter_fn *fn, void *data, bool reserved) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 424 | { |
Christoph Hellwig | 81481eb | 2014-09-13 16:40:11 -0700 | [diff] [blame] | 425 | struct request *rq; |
| 426 | int bit, i; |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 427 | |
| 428 | for (i = 0; i < bt->map_nr; i++) { |
Jens Axboe | e93ecf6 | 2014-05-19 09:17:48 -0600 | [diff] [blame] | 429 | struct blk_align_bitmap *bm = &bt->map[i]; |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 430 | |
Christoph Hellwig | 81481eb | 2014-09-13 16:40:11 -0700 | [diff] [blame] | 431 | for (bit = find_first_bit(&bm->word, bm->depth); |
| 432 | bit < bm->depth; |
| 433 | bit = find_next_bit(&bm->word, bm->depth, bit + 1)) { |
Ming Lei | 0048b48 | 2015-08-09 03:41:51 -0400 | [diff] [blame] | 434 | rq = hctx->tags->rqs[off + bit]; |
Christoph Hellwig | 81481eb | 2014-09-13 16:40:11 -0700 | [diff] [blame] | 435 | if (rq->q == hctx->queue) |
| 436 | fn(hctx, rq, data, reserved); |
| 437 | } |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 438 | |
Jens Axboe | 59d13bf | 2014-05-09 13:41:15 -0600 | [diff] [blame] | 439 | off += (1 << bt->bits_per_word); |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 440 | } |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 441 | } |
| 442 | |
Keith Busch | f26cdc8 | 2015-06-01 09:29:53 -0600 | [diff] [blame] | 443 | static void bt_tags_for_each(struct blk_mq_tags *tags, |
| 444 | struct blk_mq_bitmap_tags *bt, unsigned int off, |
| 445 | busy_tag_iter_fn *fn, void *data, bool reserved) |
| 446 | { |
| 447 | struct request *rq; |
| 448 | int bit, i; |
| 449 | |
| 450 | if (!tags->rqs) |
| 451 | return; |
| 452 | for (i = 0; i < bt->map_nr; i++) { |
| 453 | struct blk_align_bitmap *bm = &bt->map[i]; |
| 454 | |
| 455 | for (bit = find_first_bit(&bm->word, bm->depth); |
| 456 | bit < bm->depth; |
| 457 | bit = find_next_bit(&bm->word, bm->depth, bit + 1)) { |
Ming Lei | 0048b48 | 2015-08-09 03:41:51 -0400 | [diff] [blame] | 458 | rq = tags->rqs[off + bit]; |
Keith Busch | f26cdc8 | 2015-06-01 09:29:53 -0600 | [diff] [blame] | 459 | fn(rq, data, reserved); |
| 460 | } |
| 461 | |
| 462 | off += (1 << bt->bits_per_word); |
| 463 | } |
| 464 | } |
| 465 | |
Sagi Grimberg | e8f1e16 | 2016-03-10 13:58:49 +0200 | [diff] [blame] | 466 | static void blk_mq_all_tag_busy_iter(struct blk_mq_tags *tags, |
| 467 | busy_tag_iter_fn *fn, void *priv) |
Keith Busch | f26cdc8 | 2015-06-01 09:29:53 -0600 | [diff] [blame] | 468 | { |
| 469 | if (tags->nr_reserved_tags) |
| 470 | bt_tags_for_each(tags, &tags->breserved_tags, 0, fn, priv, true); |
| 471 | bt_tags_for_each(tags, &tags->bitmap_tags, tags->nr_reserved_tags, fn, priv, |
| 472 | false); |
| 473 | } |
Keith Busch | f26cdc8 | 2015-06-01 09:29:53 -0600 | [diff] [blame] | 474 | |
Sagi Grimberg | e048948 | 2016-03-10 13:58:46 +0200 | [diff] [blame] | 475 | void blk_mq_tagset_busy_iter(struct blk_mq_tag_set *tagset, |
| 476 | busy_tag_iter_fn *fn, void *priv) |
| 477 | { |
| 478 | int i; |
| 479 | |
| 480 | for (i = 0; i < tagset->nr_hw_queues; i++) { |
| 481 | if (tagset->tags && tagset->tags[i]) |
| 482 | blk_mq_all_tag_busy_iter(tagset->tags[i], fn, priv); |
| 483 | } |
| 484 | } |
| 485 | EXPORT_SYMBOL(blk_mq_tagset_busy_iter); |
| 486 | |
Sagi Grimberg | 486cf98 | 2016-07-06 21:55:48 +0900 | [diff] [blame] | 487 | int blk_mq_reinit_tagset(struct blk_mq_tag_set *set) |
| 488 | { |
| 489 | int i, j, ret = 0; |
| 490 | |
| 491 | if (!set->ops->reinit_request) |
| 492 | goto out; |
| 493 | |
| 494 | for (i = 0; i < set->nr_hw_queues; i++) { |
| 495 | struct blk_mq_tags *tags = set->tags[i]; |
| 496 | |
| 497 | for (j = 0; j < tags->nr_tags; j++) { |
| 498 | if (!tags->rqs[j]) |
| 499 | continue; |
| 500 | |
| 501 | ret = set->ops->reinit_request(set->driver_data, |
| 502 | tags->rqs[j]); |
| 503 | if (ret) |
| 504 | goto out; |
| 505 | } |
| 506 | } |
| 507 | |
| 508 | out: |
| 509 | return ret; |
| 510 | } |
| 511 | EXPORT_SYMBOL_GPL(blk_mq_reinit_tagset); |
| 512 | |
Christoph Hellwig | 0bf6cd5 | 2015-09-27 21:01:51 +0200 | [diff] [blame] | 513 | void blk_mq_queue_tag_busy_iter(struct request_queue *q, busy_iter_fn *fn, |
Christoph Hellwig | 81481eb | 2014-09-13 16:40:11 -0700 | [diff] [blame] | 514 | void *priv) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 515 | { |
Christoph Hellwig | 0bf6cd5 | 2015-09-27 21:01:51 +0200 | [diff] [blame] | 516 | struct blk_mq_hw_ctx *hctx; |
| 517 | int i; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 518 | |
Christoph Hellwig | 0bf6cd5 | 2015-09-27 21:01:51 +0200 | [diff] [blame] | 519 | |
| 520 | queue_for_each_hw_ctx(q, hctx, i) { |
| 521 | struct blk_mq_tags *tags = hctx->tags; |
| 522 | |
| 523 | /* |
| 524 | * If not software queues are currently mapped to this |
| 525 | * hardware queue, there's nothing to check |
| 526 | */ |
| 527 | if (!blk_mq_hw_queue_mapped(hctx)) |
| 528 | continue; |
| 529 | |
| 530 | if (tags->nr_reserved_tags) |
| 531 | bt_for_each(hctx, &tags->breserved_tags, 0, fn, priv, true); |
| 532 | bt_for_each(hctx, &tags->bitmap_tags, tags->nr_reserved_tags, fn, priv, |
| 533 | false); |
| 534 | } |
| 535 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 536 | } |
| 537 | |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 538 | static unsigned int bt_unused_tags(struct blk_mq_bitmap_tags *bt) |
| 539 | { |
| 540 | unsigned int i, used; |
| 541 | |
| 542 | for (i = 0, used = 0; i < bt->map_nr; i++) { |
Jens Axboe | e93ecf6 | 2014-05-19 09:17:48 -0600 | [diff] [blame] | 543 | struct blk_align_bitmap *bm = &bt->map[i]; |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 544 | |
| 545 | used += bitmap_weight(&bm->word, bm->depth); |
| 546 | } |
| 547 | |
| 548 | return bt->depth - used; |
| 549 | } |
| 550 | |
Jens Axboe | e3a2b3f | 2014-05-20 11:49:02 -0600 | [diff] [blame] | 551 | static void bt_update_count(struct blk_mq_bitmap_tags *bt, |
| 552 | unsigned int depth) |
| 553 | { |
| 554 | unsigned int tags_per_word = 1U << bt->bits_per_word; |
| 555 | unsigned int map_depth = depth; |
| 556 | |
| 557 | if (depth) { |
| 558 | int i; |
| 559 | |
| 560 | for (i = 0; i < bt->map_nr; i++) { |
| 561 | bt->map[i].depth = min(map_depth, tags_per_word); |
| 562 | map_depth -= bt->map[i].depth; |
| 563 | } |
| 564 | } |
| 565 | |
| 566 | bt->wake_cnt = BT_WAIT_BATCH; |
Jens Axboe | abab13b | 2014-10-07 08:39:20 -0600 | [diff] [blame] | 567 | if (bt->wake_cnt > depth / BT_WAIT_QUEUES) |
| 568 | bt->wake_cnt = max(1U, depth / BT_WAIT_QUEUES); |
Jens Axboe | e3a2b3f | 2014-05-20 11:49:02 -0600 | [diff] [blame] | 569 | |
| 570 | bt->depth = depth; |
| 571 | } |
| 572 | |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 573 | static int bt_alloc(struct blk_mq_bitmap_tags *bt, unsigned int depth, |
| 574 | int node, bool reserved) |
| 575 | { |
| 576 | int i; |
| 577 | |
Jens Axboe | 59d13bf | 2014-05-09 13:41:15 -0600 | [diff] [blame] | 578 | bt->bits_per_word = ilog2(BITS_PER_LONG); |
| 579 | |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 580 | /* |
| 581 | * Depth can be zero for reserved tags, that's not a failure |
| 582 | * condition. |
| 583 | */ |
| 584 | if (depth) { |
Jens Axboe | e3a2b3f | 2014-05-20 11:49:02 -0600 | [diff] [blame] | 585 | unsigned int nr, tags_per_word; |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 586 | |
Jens Axboe | 59d13bf | 2014-05-09 13:41:15 -0600 | [diff] [blame] | 587 | tags_per_word = (1 << bt->bits_per_word); |
| 588 | |
| 589 | /* |
| 590 | * If the tag space is small, shrink the number of tags |
| 591 | * per word so we spread over a few cachelines, at least. |
| 592 | * If less than 4 tags, just forget about it, it's not |
| 593 | * going to work optimally anyway. |
| 594 | */ |
| 595 | if (depth >= 4) { |
| 596 | while (tags_per_word * 4 > depth) { |
| 597 | bt->bits_per_word--; |
| 598 | tags_per_word = (1 << bt->bits_per_word); |
| 599 | } |
| 600 | } |
| 601 | |
| 602 | nr = ALIGN(depth, tags_per_word) / tags_per_word; |
Jens Axboe | e93ecf6 | 2014-05-19 09:17:48 -0600 | [diff] [blame] | 603 | bt->map = kzalloc_node(nr * sizeof(struct blk_align_bitmap), |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 604 | GFP_KERNEL, node); |
| 605 | if (!bt->map) |
| 606 | return -ENOMEM; |
| 607 | |
| 608 | bt->map_nr = nr; |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 609 | } |
| 610 | |
| 611 | bt->bs = kzalloc(BT_WAIT_QUEUES * sizeof(*bt->bs), GFP_KERNEL); |
| 612 | if (!bt->bs) { |
| 613 | kfree(bt->map); |
Tony Battersby | 564e559 | 2015-02-11 11:32:30 -0500 | [diff] [blame] | 614 | bt->map = NULL; |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 615 | return -ENOMEM; |
| 616 | } |
| 617 | |
Jens Axboe | e3a2b3f | 2014-05-20 11:49:02 -0600 | [diff] [blame] | 618 | bt_update_count(bt, depth); |
Alexander Gordeev | 86fb5c5 | 2014-06-17 22:37:23 +0200 | [diff] [blame] | 619 | |
| 620 | for (i = 0; i < BT_WAIT_QUEUES; i++) { |
| 621 | init_waitqueue_head(&bt->bs[i].wait); |
| 622 | atomic_set(&bt->bs[i].wait_cnt, bt->wake_cnt); |
| 623 | } |
| 624 | |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 625 | return 0; |
| 626 | } |
| 627 | |
| 628 | static void bt_free(struct blk_mq_bitmap_tags *bt) |
| 629 | { |
| 630 | kfree(bt->map); |
| 631 | kfree(bt->bs); |
| 632 | } |
| 633 | |
| 634 | static struct blk_mq_tags *blk_mq_init_bitmap_tags(struct blk_mq_tags *tags, |
Shaohua Li | 24391c0 | 2015-01-23 14:18:00 -0700 | [diff] [blame] | 635 | int node, int alloc_policy) |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 636 | { |
| 637 | unsigned int depth = tags->nr_tags - tags->nr_reserved_tags; |
| 638 | |
Shaohua Li | 24391c0 | 2015-01-23 14:18:00 -0700 | [diff] [blame] | 639 | tags->alloc_policy = alloc_policy; |
| 640 | |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 641 | if (bt_alloc(&tags->bitmap_tags, depth, node, false)) |
| 642 | goto enomem; |
| 643 | if (bt_alloc(&tags->breserved_tags, tags->nr_reserved_tags, node, true)) |
| 644 | goto enomem; |
| 645 | |
| 646 | return tags; |
| 647 | enomem: |
| 648 | bt_free(&tags->bitmap_tags); |
| 649 | kfree(tags); |
| 650 | return NULL; |
| 651 | } |
| 652 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 653 | struct blk_mq_tags *blk_mq_init_tags(unsigned int total_tags, |
Shaohua Li | 24391c0 | 2015-01-23 14:18:00 -0700 | [diff] [blame] | 654 | unsigned int reserved_tags, |
| 655 | int node, int alloc_policy) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 656 | { |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 657 | struct blk_mq_tags *tags; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 658 | |
| 659 | if (total_tags > BLK_MQ_TAG_MAX) { |
| 660 | pr_err("blk-mq: tag depth too large\n"); |
| 661 | return NULL; |
| 662 | } |
| 663 | |
| 664 | tags = kzalloc_node(sizeof(*tags), GFP_KERNEL, node); |
| 665 | if (!tags) |
| 666 | return NULL; |
| 667 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 668 | tags->nr_tags = total_tags; |
| 669 | tags->nr_reserved_tags = reserved_tags; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 670 | |
Shaohua Li | 24391c0 | 2015-01-23 14:18:00 -0700 | [diff] [blame] | 671 | return blk_mq_init_bitmap_tags(tags, node, alloc_policy); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 672 | } |
| 673 | |
| 674 | void blk_mq_free_tags(struct blk_mq_tags *tags) |
| 675 | { |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 676 | bt_free(&tags->bitmap_tags); |
| 677 | bt_free(&tags->breserved_tags); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 678 | kfree(tags); |
| 679 | } |
| 680 | |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 681 | void blk_mq_tag_init_last_tag(struct blk_mq_tags *tags, unsigned int *tag) |
| 682 | { |
| 683 | unsigned int depth = tags->nr_tags - tags->nr_reserved_tags; |
| 684 | |
Ming Lei | 9d3d21a | 2014-05-10 15:43:14 -0600 | [diff] [blame] | 685 | *tag = prandom_u32() % depth; |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 686 | } |
| 687 | |
Jens Axboe | e3a2b3f | 2014-05-20 11:49:02 -0600 | [diff] [blame] | 688 | int blk_mq_tag_update_depth(struct blk_mq_tags *tags, unsigned int tdepth) |
| 689 | { |
| 690 | tdepth -= tags->nr_reserved_tags; |
| 691 | if (tdepth > tags->nr_tags) |
| 692 | return -EINVAL; |
| 693 | |
| 694 | /* |
| 695 | * Don't need (or can't) update reserved tags here, they remain |
| 696 | * static and should never need resizing. |
| 697 | */ |
| 698 | bt_update_count(&tags->bitmap_tags, tdepth); |
Jens Axboe | aed3ea9 | 2014-12-22 14:04:42 -0700 | [diff] [blame] | 699 | blk_mq_tag_wakeup_all(tags, false); |
Jens Axboe | e3a2b3f | 2014-05-20 11:49:02 -0600 | [diff] [blame] | 700 | return 0; |
| 701 | } |
| 702 | |
Bart Van Assche | 205fb5f | 2014-10-30 14:45:11 +0100 | [diff] [blame] | 703 | /** |
| 704 | * blk_mq_unique_tag() - return a tag that is unique queue-wide |
| 705 | * @rq: request for which to compute a unique tag |
| 706 | * |
| 707 | * The tag field in struct request is unique per hardware queue but not over |
| 708 | * all hardware queues. Hence this function that returns a tag with the |
| 709 | * hardware context index in the upper bits and the per hardware queue tag in |
| 710 | * the lower bits. |
| 711 | * |
| 712 | * Note: When called for a request that is queued on a non-multiqueue request |
| 713 | * queue, the hardware context index is set to zero. |
| 714 | */ |
| 715 | u32 blk_mq_unique_tag(struct request *rq) |
| 716 | { |
| 717 | struct request_queue *q = rq->q; |
| 718 | struct blk_mq_hw_ctx *hctx; |
| 719 | int hwq = 0; |
| 720 | |
| 721 | if (q->mq_ops) { |
Christoph Hellwig | 7d7e0f9 | 2016-09-14 16:18:54 +0200 | [diff] [blame] | 722 | hctx = blk_mq_map_queue(q, rq->mq_ctx->cpu); |
Bart Van Assche | 205fb5f | 2014-10-30 14:45:11 +0100 | [diff] [blame] | 723 | hwq = hctx->queue_num; |
| 724 | } |
| 725 | |
| 726 | return (hwq << BLK_MQ_UNIQUE_TAG_BITS) | |
| 727 | (rq->tag & BLK_MQ_UNIQUE_TAG_MASK); |
| 728 | } |
| 729 | EXPORT_SYMBOL(blk_mq_unique_tag); |
| 730 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 731 | ssize_t blk_mq_tag_sysfs_show(struct blk_mq_tags *tags, char *page) |
| 732 | { |
| 733 | char *orig_page = page; |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 734 | unsigned int free, res; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 735 | |
| 736 | if (!tags) |
| 737 | return 0; |
| 738 | |
Jens Axboe | 59d13bf | 2014-05-09 13:41:15 -0600 | [diff] [blame] | 739 | page += sprintf(page, "nr_tags=%u, reserved_tags=%u, " |
| 740 | "bits_per_word=%u\n", |
| 741 | tags->nr_tags, tags->nr_reserved_tags, |
| 742 | tags->bitmap_tags.bits_per_word); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 743 | |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 744 | free = bt_unused_tags(&tags->bitmap_tags); |
| 745 | res = bt_unused_tags(&tags->breserved_tags); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 746 | |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 747 | page += sprintf(page, "nr_free=%u, nr_reserved=%u\n", free, res); |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 748 | page += sprintf(page, "active_queues=%u\n", atomic_read(&tags->active_queues)); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 749 | |
| 750 | return page - orig_page; |
| 751 | } |