Jens Axboe | 75bb462 | 2014-05-28 10:15:41 -0600 | [diff] [blame] | 1 | /* |
| 2 | * Block multiqueue core code |
| 3 | * |
| 4 | * Copyright (C) 2013-2014 Jens Axboe |
| 5 | * Copyright (C) 2013-2014 Christoph Hellwig |
| 6 | */ |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 7 | #include <linux/kernel.h> |
| 8 | #include <linux/module.h> |
| 9 | #include <linux/backing-dev.h> |
| 10 | #include <linux/bio.h> |
| 11 | #include <linux/blkdev.h> |
Catalin Marinas | f75782e | 2015-09-14 18:16:02 +0100 | [diff] [blame] | 12 | #include <linux/kmemleak.h> |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 13 | #include <linux/mm.h> |
| 14 | #include <linux/init.h> |
| 15 | #include <linux/slab.h> |
| 16 | #include <linux/workqueue.h> |
| 17 | #include <linux/smp.h> |
| 18 | #include <linux/llist.h> |
| 19 | #include <linux/list_sort.h> |
| 20 | #include <linux/cpu.h> |
| 21 | #include <linux/cache.h> |
| 22 | #include <linux/sched/sysctl.h> |
| 23 | #include <linux/delay.h> |
Jens Axboe | aedcd72 | 2014-09-17 08:27:03 -0600 | [diff] [blame] | 24 | #include <linux/crash_dump.h> |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 25 | |
| 26 | #include <trace/events/block.h> |
| 27 | |
| 28 | #include <linux/blk-mq.h> |
| 29 | #include "blk.h" |
| 30 | #include "blk-mq.h" |
| 31 | #include "blk-mq-tag.h" |
| 32 | |
| 33 | static DEFINE_MUTEX(all_q_mutex); |
| 34 | static LIST_HEAD(all_q_list); |
| 35 | |
| 36 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx); |
| 37 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 38 | /* |
| 39 | * Check if any of the ctx's have pending work in this hardware queue |
| 40 | */ |
| 41 | static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx) |
| 42 | { |
| 43 | unsigned int i; |
| 44 | |
Jens Axboe | 569fd0c | 2015-04-17 08:28:50 -0600 | [diff] [blame] | 45 | for (i = 0; i < hctx->ctx_map.size; i++) |
Jens Axboe | 1429d7c | 2014-05-19 09:23:55 -0600 | [diff] [blame] | 46 | if (hctx->ctx_map.map[i].word) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 47 | return true; |
| 48 | |
| 49 | return false; |
| 50 | } |
| 51 | |
Jens Axboe | 1429d7c | 2014-05-19 09:23:55 -0600 | [diff] [blame] | 52 | static inline struct blk_align_bitmap *get_bm(struct blk_mq_hw_ctx *hctx, |
| 53 | struct blk_mq_ctx *ctx) |
| 54 | { |
| 55 | return &hctx->ctx_map.map[ctx->index_hw / hctx->ctx_map.bits_per_word]; |
| 56 | } |
| 57 | |
| 58 | #define CTX_TO_BIT(hctx, ctx) \ |
| 59 | ((ctx)->index_hw & ((hctx)->ctx_map.bits_per_word - 1)) |
| 60 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 61 | /* |
| 62 | * Mark this ctx as having pending work in this hardware queue |
| 63 | */ |
| 64 | static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx, |
| 65 | struct blk_mq_ctx *ctx) |
| 66 | { |
Jens Axboe | 1429d7c | 2014-05-19 09:23:55 -0600 | [diff] [blame] | 67 | struct blk_align_bitmap *bm = get_bm(hctx, ctx); |
| 68 | |
| 69 | if (!test_bit(CTX_TO_BIT(hctx, ctx), &bm->word)) |
| 70 | set_bit(CTX_TO_BIT(hctx, ctx), &bm->word); |
| 71 | } |
| 72 | |
| 73 | static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx, |
| 74 | struct blk_mq_ctx *ctx) |
| 75 | { |
| 76 | struct blk_align_bitmap *bm = get_bm(hctx, ctx); |
| 77 | |
| 78 | clear_bit(CTX_TO_BIT(hctx, ctx), &bm->word); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 79 | } |
| 80 | |
Keith Busch | b4c6a02 | 2014-12-19 17:54:14 -0700 | [diff] [blame] | 81 | void blk_mq_freeze_queue_start(struct request_queue *q) |
Ming Lei | 43a5e4e | 2013-12-26 21:31:35 +0800 | [diff] [blame] | 82 | { |
Christoph Hellwig | 4ecd4fe | 2015-05-07 09:38:13 +0200 | [diff] [blame] | 83 | int freeze_depth; |
Tejun Heo | cddd5d1 | 2014-08-16 08:02:24 -0400 | [diff] [blame] | 84 | |
Christoph Hellwig | 4ecd4fe | 2015-05-07 09:38:13 +0200 | [diff] [blame] | 85 | freeze_depth = atomic_inc_return(&q->mq_freeze_depth); |
| 86 | if (freeze_depth == 1) { |
Dan Williams | 3ef28e8 | 2015-10-21 13:20:12 -0400 | [diff] [blame^] | 87 | percpu_ref_kill(&q->q_usage_counter); |
Mike Snitzer | b94ec29 | 2015-03-11 23:56:38 -0400 | [diff] [blame] | 88 | blk_mq_run_hw_queues(q, false); |
Tejun Heo | cddd5d1 | 2014-08-16 08:02:24 -0400 | [diff] [blame] | 89 | } |
Tejun Heo | f3af020 | 2014-11-04 13:52:27 -0500 | [diff] [blame] | 90 | } |
Keith Busch | b4c6a02 | 2014-12-19 17:54:14 -0700 | [diff] [blame] | 91 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_start); |
Tejun Heo | f3af020 | 2014-11-04 13:52:27 -0500 | [diff] [blame] | 92 | |
| 93 | static void blk_mq_freeze_queue_wait(struct request_queue *q) |
| 94 | { |
Dan Williams | 3ef28e8 | 2015-10-21 13:20:12 -0400 | [diff] [blame^] | 95 | wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter)); |
Ming Lei | 43a5e4e | 2013-12-26 21:31:35 +0800 | [diff] [blame] | 96 | } |
| 97 | |
Tejun Heo | f3af020 | 2014-11-04 13:52:27 -0500 | [diff] [blame] | 98 | /* |
| 99 | * Guarantee no request is in use, so we can change any data structure of |
| 100 | * the queue afterward. |
| 101 | */ |
Dan Williams | 3ef28e8 | 2015-10-21 13:20:12 -0400 | [diff] [blame^] | 102 | void blk_freeze_queue(struct request_queue *q) |
Tejun Heo | f3af020 | 2014-11-04 13:52:27 -0500 | [diff] [blame] | 103 | { |
Dan Williams | 3ef28e8 | 2015-10-21 13:20:12 -0400 | [diff] [blame^] | 104 | /* |
| 105 | * In the !blk_mq case we are only calling this to kill the |
| 106 | * q_usage_counter, otherwise this increases the freeze depth |
| 107 | * and waits for it to return to zero. For this reason there is |
| 108 | * no blk_unfreeze_queue(), and blk_freeze_queue() is not |
| 109 | * exported to drivers as the only user for unfreeze is blk_mq. |
| 110 | */ |
Tejun Heo | f3af020 | 2014-11-04 13:52:27 -0500 | [diff] [blame] | 111 | blk_mq_freeze_queue_start(q); |
| 112 | blk_mq_freeze_queue_wait(q); |
| 113 | } |
Dan Williams | 3ef28e8 | 2015-10-21 13:20:12 -0400 | [diff] [blame^] | 114 | |
| 115 | void blk_mq_freeze_queue(struct request_queue *q) |
| 116 | { |
| 117 | /* |
| 118 | * ...just an alias to keep freeze and unfreeze actions balanced |
| 119 | * in the blk_mq_* namespace |
| 120 | */ |
| 121 | blk_freeze_queue(q); |
| 122 | } |
Jens Axboe | c761d96 | 2015-01-02 15:05:12 -0700 | [diff] [blame] | 123 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue); |
Tejun Heo | f3af020 | 2014-11-04 13:52:27 -0500 | [diff] [blame] | 124 | |
Keith Busch | b4c6a02 | 2014-12-19 17:54:14 -0700 | [diff] [blame] | 125 | void blk_mq_unfreeze_queue(struct request_queue *q) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 126 | { |
Christoph Hellwig | 4ecd4fe | 2015-05-07 09:38:13 +0200 | [diff] [blame] | 127 | int freeze_depth; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 128 | |
Christoph Hellwig | 4ecd4fe | 2015-05-07 09:38:13 +0200 | [diff] [blame] | 129 | freeze_depth = atomic_dec_return(&q->mq_freeze_depth); |
| 130 | WARN_ON_ONCE(freeze_depth < 0); |
| 131 | if (!freeze_depth) { |
Dan Williams | 3ef28e8 | 2015-10-21 13:20:12 -0400 | [diff] [blame^] | 132 | percpu_ref_reinit(&q->q_usage_counter); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 133 | wake_up_all(&q->mq_freeze_wq); |
Tejun Heo | add703f | 2014-07-01 10:34:38 -0600 | [diff] [blame] | 134 | } |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 135 | } |
Keith Busch | b4c6a02 | 2014-12-19 17:54:14 -0700 | [diff] [blame] | 136 | EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 137 | |
Jens Axboe | aed3ea9 | 2014-12-22 14:04:42 -0700 | [diff] [blame] | 138 | void blk_mq_wake_waiters(struct request_queue *q) |
| 139 | { |
| 140 | struct blk_mq_hw_ctx *hctx; |
| 141 | unsigned int i; |
| 142 | |
| 143 | queue_for_each_hw_ctx(q, hctx, i) |
| 144 | if (blk_mq_hw_queue_mapped(hctx)) |
| 145 | blk_mq_tag_wakeup_all(hctx->tags, true); |
Keith Busch | 3fd5940 | 2015-01-08 08:53:56 -0700 | [diff] [blame] | 146 | |
| 147 | /* |
| 148 | * If we are called because the queue has now been marked as |
| 149 | * dying, we need to ensure that processes currently waiting on |
| 150 | * the queue are notified as well. |
| 151 | */ |
| 152 | wake_up_all(&q->mq_freeze_wq); |
Jens Axboe | aed3ea9 | 2014-12-22 14:04:42 -0700 | [diff] [blame] | 153 | } |
| 154 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 155 | bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx) |
| 156 | { |
| 157 | return blk_mq_has_free_tags(hctx->tags); |
| 158 | } |
| 159 | EXPORT_SYMBOL(blk_mq_can_queue); |
| 160 | |
Jens Axboe | 94eddfb | 2013-11-19 09:25:07 -0700 | [diff] [blame] | 161 | static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx, |
| 162 | struct request *rq, unsigned int rw_flags) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 163 | { |
Jens Axboe | 94eddfb | 2013-11-19 09:25:07 -0700 | [diff] [blame] | 164 | if (blk_queue_io_stat(q)) |
| 165 | rw_flags |= REQ_IO_STAT; |
| 166 | |
Christoph Hellwig | af76e55 | 2014-05-06 12:12:45 +0200 | [diff] [blame] | 167 | INIT_LIST_HEAD(&rq->queuelist); |
| 168 | /* csd/requeue_work/fifo_time is initialized before use */ |
| 169 | rq->q = q; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 170 | rq->mq_ctx = ctx; |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 171 | rq->cmd_flags |= rw_flags; |
Christoph Hellwig | af76e55 | 2014-05-06 12:12:45 +0200 | [diff] [blame] | 172 | /* do not touch atomic flags, it needs atomic ops against the timer */ |
| 173 | rq->cpu = -1; |
Christoph Hellwig | af76e55 | 2014-05-06 12:12:45 +0200 | [diff] [blame] | 174 | INIT_HLIST_NODE(&rq->hash); |
| 175 | RB_CLEAR_NODE(&rq->rb_node); |
Christoph Hellwig | af76e55 | 2014-05-06 12:12:45 +0200 | [diff] [blame] | 176 | rq->rq_disk = NULL; |
| 177 | rq->part = NULL; |
Jens Axboe | 3ee3237 | 2014-06-09 09:36:53 -0600 | [diff] [blame] | 178 | rq->start_time = jiffies; |
Christoph Hellwig | af76e55 | 2014-05-06 12:12:45 +0200 | [diff] [blame] | 179 | #ifdef CONFIG_BLK_CGROUP |
| 180 | rq->rl = NULL; |
Ming Lei | 0fec08b | 2014-01-03 10:00:08 -0700 | [diff] [blame] | 181 | set_start_time_ns(rq); |
Christoph Hellwig | af76e55 | 2014-05-06 12:12:45 +0200 | [diff] [blame] | 182 | rq->io_start_time_ns = 0; |
| 183 | #endif |
| 184 | rq->nr_phys_segments = 0; |
| 185 | #if defined(CONFIG_BLK_DEV_INTEGRITY) |
| 186 | rq->nr_integrity_segments = 0; |
| 187 | #endif |
Christoph Hellwig | af76e55 | 2014-05-06 12:12:45 +0200 | [diff] [blame] | 188 | rq->special = NULL; |
| 189 | /* tag was already set */ |
| 190 | rq->errors = 0; |
Christoph Hellwig | af76e55 | 2014-05-06 12:12:45 +0200 | [diff] [blame] | 191 | |
Tony Battersby | 6f4a1626 | 2014-08-22 15:53:39 -0400 | [diff] [blame] | 192 | rq->cmd = rq->__cmd; |
| 193 | |
Christoph Hellwig | af76e55 | 2014-05-06 12:12:45 +0200 | [diff] [blame] | 194 | rq->extra_len = 0; |
| 195 | rq->sense_len = 0; |
| 196 | rq->resid_len = 0; |
| 197 | rq->sense = NULL; |
| 198 | |
Christoph Hellwig | af76e55 | 2014-05-06 12:12:45 +0200 | [diff] [blame] | 199 | INIT_LIST_HEAD(&rq->timeout_list); |
Jens Axboe | f6be4fb | 2014-06-06 11:03:48 -0600 | [diff] [blame] | 200 | rq->timeout = 0; |
| 201 | |
Christoph Hellwig | af76e55 | 2014-05-06 12:12:45 +0200 | [diff] [blame] | 202 | rq->end_io = NULL; |
| 203 | rq->end_io_data = NULL; |
| 204 | rq->next_rq = NULL; |
| 205 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 206 | ctx->rq_dispatched[rw_is_sync(rw_flags)]++; |
| 207 | } |
| 208 | |
Christoph Hellwig | 5dee857 | 2014-05-27 20:59:47 +0200 | [diff] [blame] | 209 | static struct request * |
Ming Lei | cb96a42 | 2014-06-01 00:43:37 +0800 | [diff] [blame] | 210 | __blk_mq_alloc_request(struct blk_mq_alloc_data *data, int rw) |
Christoph Hellwig | 5dee857 | 2014-05-27 20:59:47 +0200 | [diff] [blame] | 211 | { |
| 212 | struct request *rq; |
| 213 | unsigned int tag; |
| 214 | |
Ming Lei | cb96a42 | 2014-06-01 00:43:37 +0800 | [diff] [blame] | 215 | tag = blk_mq_get_tag(data); |
Christoph Hellwig | 5dee857 | 2014-05-27 20:59:47 +0200 | [diff] [blame] | 216 | if (tag != BLK_MQ_TAG_FAIL) { |
Ming Lei | cb96a42 | 2014-06-01 00:43:37 +0800 | [diff] [blame] | 217 | rq = data->hctx->tags->rqs[tag]; |
Christoph Hellwig | 5dee857 | 2014-05-27 20:59:47 +0200 | [diff] [blame] | 218 | |
Ming Lei | cb96a42 | 2014-06-01 00:43:37 +0800 | [diff] [blame] | 219 | if (blk_mq_tag_busy(data->hctx)) { |
Christoph Hellwig | 5dee857 | 2014-05-27 20:59:47 +0200 | [diff] [blame] | 220 | rq->cmd_flags = REQ_MQ_INFLIGHT; |
Ming Lei | cb96a42 | 2014-06-01 00:43:37 +0800 | [diff] [blame] | 221 | atomic_inc(&data->hctx->nr_active); |
Christoph Hellwig | 5dee857 | 2014-05-27 20:59:47 +0200 | [diff] [blame] | 222 | } |
| 223 | |
| 224 | rq->tag = tag; |
Ming Lei | cb96a42 | 2014-06-01 00:43:37 +0800 | [diff] [blame] | 225 | blk_mq_rq_ctx_init(data->q, data->ctx, rq, rw); |
Christoph Hellwig | 5dee857 | 2014-05-27 20:59:47 +0200 | [diff] [blame] | 226 | return rq; |
| 227 | } |
| 228 | |
| 229 | return NULL; |
| 230 | } |
| 231 | |
Christoph Hellwig | 4ce01dd | 2014-05-27 20:59:46 +0200 | [diff] [blame] | 232 | struct request *blk_mq_alloc_request(struct request_queue *q, int rw, gfp_t gfp, |
| 233 | bool reserved) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 234 | { |
Christoph Hellwig | d852564 | 2014-05-27 20:59:50 +0200 | [diff] [blame] | 235 | struct blk_mq_ctx *ctx; |
| 236 | struct blk_mq_hw_ctx *hctx; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 237 | struct request *rq; |
Ming Lei | cb96a42 | 2014-06-01 00:43:37 +0800 | [diff] [blame] | 238 | struct blk_mq_alloc_data alloc_data; |
Joe Lawrence | a492f07 | 2014-08-28 08:15:21 -0600 | [diff] [blame] | 239 | int ret; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 240 | |
Dan Williams | 3ef28e8 | 2015-10-21 13:20:12 -0400 | [diff] [blame^] | 241 | ret = blk_queue_enter(q, gfp); |
Joe Lawrence | a492f07 | 2014-08-28 08:15:21 -0600 | [diff] [blame] | 242 | if (ret) |
| 243 | return ERR_PTR(ret); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 244 | |
Christoph Hellwig | d852564 | 2014-05-27 20:59:50 +0200 | [diff] [blame] | 245 | ctx = blk_mq_get_ctx(q); |
| 246 | hctx = q->mq_ops->map_queue(q, ctx->cpu); |
Ming Lei | cb96a42 | 2014-06-01 00:43:37 +0800 | [diff] [blame] | 247 | blk_mq_set_alloc_data(&alloc_data, q, gfp & ~__GFP_WAIT, |
| 248 | reserved, ctx, hctx); |
Christoph Hellwig | d852564 | 2014-05-27 20:59:50 +0200 | [diff] [blame] | 249 | |
Ming Lei | cb96a42 | 2014-06-01 00:43:37 +0800 | [diff] [blame] | 250 | rq = __blk_mq_alloc_request(&alloc_data, rw); |
Christoph Hellwig | d852564 | 2014-05-27 20:59:50 +0200 | [diff] [blame] | 251 | if (!rq && (gfp & __GFP_WAIT)) { |
| 252 | __blk_mq_run_hw_queue(hctx); |
| 253 | blk_mq_put_ctx(ctx); |
| 254 | |
| 255 | ctx = blk_mq_get_ctx(q); |
| 256 | hctx = q->mq_ops->map_queue(q, ctx->cpu); |
Ming Lei | cb96a42 | 2014-06-01 00:43:37 +0800 | [diff] [blame] | 257 | blk_mq_set_alloc_data(&alloc_data, q, gfp, reserved, ctx, |
| 258 | hctx); |
| 259 | rq = __blk_mq_alloc_request(&alloc_data, rw); |
| 260 | ctx = alloc_data.ctx; |
Christoph Hellwig | d852564 | 2014-05-27 20:59:50 +0200 | [diff] [blame] | 261 | } |
| 262 | blk_mq_put_ctx(ctx); |
Keith Busch | c76541a | 2014-12-19 17:54:13 -0700 | [diff] [blame] | 263 | if (!rq) { |
Dan Williams | 3ef28e8 | 2015-10-21 13:20:12 -0400 | [diff] [blame^] | 264 | blk_queue_exit(q); |
Joe Lawrence | a492f07 | 2014-08-28 08:15:21 -0600 | [diff] [blame] | 265 | return ERR_PTR(-EWOULDBLOCK); |
Keith Busch | c76541a | 2014-12-19 17:54:13 -0700 | [diff] [blame] | 266 | } |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 267 | return rq; |
| 268 | } |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 269 | EXPORT_SYMBOL(blk_mq_alloc_request); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 270 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 271 | static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx, |
| 272 | struct blk_mq_ctx *ctx, struct request *rq) |
| 273 | { |
| 274 | const int tag = rq->tag; |
| 275 | struct request_queue *q = rq->q; |
| 276 | |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 277 | if (rq->cmd_flags & REQ_MQ_INFLIGHT) |
| 278 | atomic_dec(&hctx->nr_active); |
David Hildenbrand | 683d0e1 | 2014-09-18 11:04:31 +0200 | [diff] [blame] | 279 | rq->cmd_flags = 0; |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 280 | |
Christoph Hellwig | af76e55 | 2014-05-06 12:12:45 +0200 | [diff] [blame] | 281 | clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags); |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 282 | blk_mq_put_tag(hctx, tag, &ctx->last_tag); |
Dan Williams | 3ef28e8 | 2015-10-21 13:20:12 -0400 | [diff] [blame^] | 283 | blk_queue_exit(q); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 284 | } |
| 285 | |
Jens Axboe | 7c7f2f2 | 2014-11-17 10:41:57 -0700 | [diff] [blame] | 286 | void blk_mq_free_hctx_request(struct blk_mq_hw_ctx *hctx, struct request *rq) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 287 | { |
| 288 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
Jens Axboe | 7c7f2f2 | 2014-11-17 10:41:57 -0700 | [diff] [blame] | 289 | |
| 290 | ctx->rq_completed[rq_is_sync(rq)]++; |
| 291 | __blk_mq_free_request(hctx, ctx, rq); |
| 292 | |
| 293 | } |
| 294 | EXPORT_SYMBOL_GPL(blk_mq_free_hctx_request); |
| 295 | |
| 296 | void blk_mq_free_request(struct request *rq) |
| 297 | { |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 298 | struct blk_mq_hw_ctx *hctx; |
| 299 | struct request_queue *q = rq->q; |
| 300 | |
Jens Axboe | 7c7f2f2 | 2014-11-17 10:41:57 -0700 | [diff] [blame] | 301 | hctx = q->mq_ops->map_queue(q, rq->mq_ctx->cpu); |
| 302 | blk_mq_free_hctx_request(hctx, rq); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 303 | } |
Jens Axboe | 1a3b595 | 2014-11-17 10:40:48 -0700 | [diff] [blame] | 304 | EXPORT_SYMBOL_GPL(blk_mq_free_request); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 305 | |
Christoph Hellwig | c8a446a | 2014-09-13 16:40:10 -0700 | [diff] [blame] | 306 | inline void __blk_mq_end_request(struct request *rq, int error) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 307 | { |
Ming Lei | 0d11e6a | 2013-12-05 10:50:39 -0700 | [diff] [blame] | 308 | blk_account_io_done(rq); |
| 309 | |
Christoph Hellwig | 91b6363 | 2014-04-16 09:44:53 +0200 | [diff] [blame] | 310 | if (rq->end_io) { |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 311 | rq->end_io(rq, error); |
Christoph Hellwig | 91b6363 | 2014-04-16 09:44:53 +0200 | [diff] [blame] | 312 | } else { |
| 313 | if (unlikely(blk_bidi_rq(rq))) |
| 314 | blk_mq_free_request(rq->next_rq); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 315 | blk_mq_free_request(rq); |
Christoph Hellwig | 91b6363 | 2014-04-16 09:44:53 +0200 | [diff] [blame] | 316 | } |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 317 | } |
Christoph Hellwig | c8a446a | 2014-09-13 16:40:10 -0700 | [diff] [blame] | 318 | EXPORT_SYMBOL(__blk_mq_end_request); |
Christoph Hellwig | 63151a4 | 2014-04-16 09:44:52 +0200 | [diff] [blame] | 319 | |
Christoph Hellwig | c8a446a | 2014-09-13 16:40:10 -0700 | [diff] [blame] | 320 | void blk_mq_end_request(struct request *rq, int error) |
Christoph Hellwig | 63151a4 | 2014-04-16 09:44:52 +0200 | [diff] [blame] | 321 | { |
| 322 | if (blk_update_request(rq, error, blk_rq_bytes(rq))) |
| 323 | BUG(); |
Christoph Hellwig | c8a446a | 2014-09-13 16:40:10 -0700 | [diff] [blame] | 324 | __blk_mq_end_request(rq, error); |
Christoph Hellwig | 63151a4 | 2014-04-16 09:44:52 +0200 | [diff] [blame] | 325 | } |
Christoph Hellwig | c8a446a | 2014-09-13 16:40:10 -0700 | [diff] [blame] | 326 | EXPORT_SYMBOL(blk_mq_end_request); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 327 | |
Christoph Hellwig | 30a91cb | 2014-02-10 03:24:38 -0800 | [diff] [blame] | 328 | static void __blk_mq_complete_request_remote(void *data) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 329 | { |
Christoph Hellwig | 3d6efbf | 2014-01-08 09:33:37 -0800 | [diff] [blame] | 330 | struct request *rq = data; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 331 | |
Christoph Hellwig | 30a91cb | 2014-02-10 03:24:38 -0800 | [diff] [blame] | 332 | rq->q->softirq_done_fn(rq); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 333 | } |
| 334 | |
Jens Axboe | ed85186 | 2014-05-30 21:20:50 -0600 | [diff] [blame] | 335 | static void blk_mq_ipi_complete_request(struct request *rq) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 336 | { |
| 337 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
Christoph Hellwig | 3853520 | 2014-04-25 02:32:53 -0700 | [diff] [blame] | 338 | bool shared = false; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 339 | int cpu; |
| 340 | |
Christoph Hellwig | 3853520 | 2014-04-25 02:32:53 -0700 | [diff] [blame] | 341 | if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) { |
Christoph Hellwig | 30a91cb | 2014-02-10 03:24:38 -0800 | [diff] [blame] | 342 | rq->q->softirq_done_fn(rq); |
| 343 | return; |
| 344 | } |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 345 | |
| 346 | cpu = get_cpu(); |
Christoph Hellwig | 3853520 | 2014-04-25 02:32:53 -0700 | [diff] [blame] | 347 | if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags)) |
| 348 | shared = cpus_share_cache(cpu, ctx->cpu); |
| 349 | |
| 350 | if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) { |
Christoph Hellwig | 30a91cb | 2014-02-10 03:24:38 -0800 | [diff] [blame] | 351 | rq->csd.func = __blk_mq_complete_request_remote; |
Christoph Hellwig | 3d6efbf | 2014-01-08 09:33:37 -0800 | [diff] [blame] | 352 | rq->csd.info = rq; |
| 353 | rq->csd.flags = 0; |
Frederic Weisbecker | c46fff2 | 2014-02-24 16:40:02 +0100 | [diff] [blame] | 354 | smp_call_function_single_async(ctx->cpu, &rq->csd); |
Christoph Hellwig | 3d6efbf | 2014-01-08 09:33:37 -0800 | [diff] [blame] | 355 | } else { |
Christoph Hellwig | 30a91cb | 2014-02-10 03:24:38 -0800 | [diff] [blame] | 356 | rq->q->softirq_done_fn(rq); |
Christoph Hellwig | 3d6efbf | 2014-01-08 09:33:37 -0800 | [diff] [blame] | 357 | } |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 358 | put_cpu(); |
| 359 | } |
Christoph Hellwig | 30a91cb | 2014-02-10 03:24:38 -0800 | [diff] [blame] | 360 | |
Jens Axboe | ed85186 | 2014-05-30 21:20:50 -0600 | [diff] [blame] | 361 | void __blk_mq_complete_request(struct request *rq) |
| 362 | { |
| 363 | struct request_queue *q = rq->q; |
| 364 | |
| 365 | if (!q->softirq_done_fn) |
Christoph Hellwig | c8a446a | 2014-09-13 16:40:10 -0700 | [diff] [blame] | 366 | blk_mq_end_request(rq, rq->errors); |
Jens Axboe | ed85186 | 2014-05-30 21:20:50 -0600 | [diff] [blame] | 367 | else |
| 368 | blk_mq_ipi_complete_request(rq); |
| 369 | } |
| 370 | |
Christoph Hellwig | 30a91cb | 2014-02-10 03:24:38 -0800 | [diff] [blame] | 371 | /** |
| 372 | * blk_mq_complete_request - end I/O on a request |
| 373 | * @rq: the request being processed |
| 374 | * |
| 375 | * Description: |
| 376 | * Ends all I/O on a request. It does not handle partial completions. |
| 377 | * The actual completion happens out-of-order, through a IPI handler. |
| 378 | **/ |
Christoph Hellwig | f4829a9 | 2015-09-27 21:01:50 +0200 | [diff] [blame] | 379 | void blk_mq_complete_request(struct request *rq, int error) |
Christoph Hellwig | 30a91cb | 2014-02-10 03:24:38 -0800 | [diff] [blame] | 380 | { |
Jens Axboe | 95f0968 | 2014-05-27 17:46:48 -0600 | [diff] [blame] | 381 | struct request_queue *q = rq->q; |
| 382 | |
| 383 | if (unlikely(blk_should_fake_timeout(q))) |
Christoph Hellwig | 30a91cb | 2014-02-10 03:24:38 -0800 | [diff] [blame] | 384 | return; |
Christoph Hellwig | f4829a9 | 2015-09-27 21:01:50 +0200 | [diff] [blame] | 385 | if (!blk_mark_rq_complete(rq)) { |
| 386 | rq->errors = error; |
Jens Axboe | ed85186 | 2014-05-30 21:20:50 -0600 | [diff] [blame] | 387 | __blk_mq_complete_request(rq); |
Christoph Hellwig | f4829a9 | 2015-09-27 21:01:50 +0200 | [diff] [blame] | 388 | } |
Christoph Hellwig | 30a91cb | 2014-02-10 03:24:38 -0800 | [diff] [blame] | 389 | } |
| 390 | EXPORT_SYMBOL(blk_mq_complete_request); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 391 | |
Keith Busch | 973c019 | 2015-01-07 18:55:43 -0700 | [diff] [blame] | 392 | int blk_mq_request_started(struct request *rq) |
| 393 | { |
| 394 | return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags); |
| 395 | } |
| 396 | EXPORT_SYMBOL_GPL(blk_mq_request_started); |
| 397 | |
Christoph Hellwig | e249007 | 2014-09-13 16:40:09 -0700 | [diff] [blame] | 398 | void blk_mq_start_request(struct request *rq) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 399 | { |
| 400 | struct request_queue *q = rq->q; |
| 401 | |
| 402 | trace_block_rq_issue(q, rq); |
| 403 | |
Christoph Hellwig | 742ee69 | 2014-04-14 10:30:06 +0200 | [diff] [blame] | 404 | rq->resid_len = blk_rq_bytes(rq); |
Christoph Hellwig | 91b6363 | 2014-04-16 09:44:53 +0200 | [diff] [blame] | 405 | if (unlikely(blk_bidi_rq(rq))) |
| 406 | rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq); |
Christoph Hellwig | 742ee69 | 2014-04-14 10:30:06 +0200 | [diff] [blame] | 407 | |
Ming Lei | 2b8393b | 2014-06-10 00:16:41 +0800 | [diff] [blame] | 408 | blk_add_timer(rq); |
Jens Axboe | 87ee7b1 | 2014-04-24 08:51:47 -0600 | [diff] [blame] | 409 | |
| 410 | /* |
Jens Axboe | 538b753 | 2014-09-16 10:37:37 -0600 | [diff] [blame] | 411 | * Ensure that ->deadline is visible before set the started |
| 412 | * flag and clear the completed flag. |
| 413 | */ |
| 414 | smp_mb__before_atomic(); |
| 415 | |
| 416 | /* |
Jens Axboe | 87ee7b1 | 2014-04-24 08:51:47 -0600 | [diff] [blame] | 417 | * Mark us as started and clear complete. Complete might have been |
| 418 | * set if requeue raced with timeout, which then marked it as |
| 419 | * complete. So be sure to clear complete again when we start |
| 420 | * the request, otherwise we'll ignore the completion event. |
| 421 | */ |
Jens Axboe | 4b57052 | 2014-05-29 11:00:11 -0600 | [diff] [blame] | 422 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) |
| 423 | set_bit(REQ_ATOM_STARTED, &rq->atomic_flags); |
| 424 | if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags)) |
| 425 | clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags); |
Christoph Hellwig | 49f5baa | 2014-02-11 08:27:14 -0800 | [diff] [blame] | 426 | |
| 427 | if (q->dma_drain_size && blk_rq_bytes(rq)) { |
| 428 | /* |
| 429 | * Make sure space for the drain appears. We know we can do |
| 430 | * this because max_hw_segments has been adjusted to be one |
| 431 | * fewer than the device can handle. |
| 432 | */ |
| 433 | rq->nr_phys_segments++; |
| 434 | } |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 435 | } |
Christoph Hellwig | e249007 | 2014-09-13 16:40:09 -0700 | [diff] [blame] | 436 | EXPORT_SYMBOL(blk_mq_start_request); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 437 | |
Christoph Hellwig | ed0791b | 2014-04-16 09:44:57 +0200 | [diff] [blame] | 438 | static void __blk_mq_requeue_request(struct request *rq) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 439 | { |
| 440 | struct request_queue *q = rq->q; |
| 441 | |
| 442 | trace_block_rq_requeue(q, rq); |
Christoph Hellwig | 49f5baa | 2014-02-11 08:27:14 -0800 | [diff] [blame] | 443 | |
Christoph Hellwig | e249007 | 2014-09-13 16:40:09 -0700 | [diff] [blame] | 444 | if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) { |
| 445 | if (q->dma_drain_size && blk_rq_bytes(rq)) |
| 446 | rq->nr_phys_segments--; |
| 447 | } |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 448 | } |
| 449 | |
Christoph Hellwig | ed0791b | 2014-04-16 09:44:57 +0200 | [diff] [blame] | 450 | void blk_mq_requeue_request(struct request *rq) |
| 451 | { |
Christoph Hellwig | ed0791b | 2014-04-16 09:44:57 +0200 | [diff] [blame] | 452 | __blk_mq_requeue_request(rq); |
Christoph Hellwig | ed0791b | 2014-04-16 09:44:57 +0200 | [diff] [blame] | 453 | |
Christoph Hellwig | ed0791b | 2014-04-16 09:44:57 +0200 | [diff] [blame] | 454 | BUG_ON(blk_queued_rq(rq)); |
Christoph Hellwig | 6fca6a6 | 2014-05-28 08:08:02 -0600 | [diff] [blame] | 455 | blk_mq_add_to_requeue_list(rq, true); |
Christoph Hellwig | ed0791b | 2014-04-16 09:44:57 +0200 | [diff] [blame] | 456 | } |
| 457 | EXPORT_SYMBOL(blk_mq_requeue_request); |
| 458 | |
Christoph Hellwig | 6fca6a6 | 2014-05-28 08:08:02 -0600 | [diff] [blame] | 459 | static void blk_mq_requeue_work(struct work_struct *work) |
| 460 | { |
| 461 | struct request_queue *q = |
| 462 | container_of(work, struct request_queue, requeue_work); |
| 463 | LIST_HEAD(rq_list); |
| 464 | struct request *rq, *next; |
| 465 | unsigned long flags; |
| 466 | |
| 467 | spin_lock_irqsave(&q->requeue_lock, flags); |
| 468 | list_splice_init(&q->requeue_list, &rq_list); |
| 469 | spin_unlock_irqrestore(&q->requeue_lock, flags); |
| 470 | |
| 471 | list_for_each_entry_safe(rq, next, &rq_list, queuelist) { |
| 472 | if (!(rq->cmd_flags & REQ_SOFTBARRIER)) |
| 473 | continue; |
| 474 | |
| 475 | rq->cmd_flags &= ~REQ_SOFTBARRIER; |
| 476 | list_del_init(&rq->queuelist); |
| 477 | blk_mq_insert_request(rq, true, false, false); |
| 478 | } |
| 479 | |
| 480 | while (!list_empty(&rq_list)) { |
| 481 | rq = list_entry(rq_list.next, struct request, queuelist); |
| 482 | list_del_init(&rq->queuelist); |
| 483 | blk_mq_insert_request(rq, false, false, false); |
| 484 | } |
| 485 | |
Jens Axboe | 8b95741 | 2014-09-19 13:10:29 -0600 | [diff] [blame] | 486 | /* |
| 487 | * Use the start variant of queue running here, so that running |
| 488 | * the requeue work will kick stopped queues. |
| 489 | */ |
| 490 | blk_mq_start_hw_queues(q); |
Christoph Hellwig | 6fca6a6 | 2014-05-28 08:08:02 -0600 | [diff] [blame] | 491 | } |
| 492 | |
| 493 | void blk_mq_add_to_requeue_list(struct request *rq, bool at_head) |
| 494 | { |
| 495 | struct request_queue *q = rq->q; |
| 496 | unsigned long flags; |
| 497 | |
| 498 | /* |
| 499 | * We abuse this flag that is otherwise used by the I/O scheduler to |
| 500 | * request head insertation from the workqueue. |
| 501 | */ |
| 502 | BUG_ON(rq->cmd_flags & REQ_SOFTBARRIER); |
| 503 | |
| 504 | spin_lock_irqsave(&q->requeue_lock, flags); |
| 505 | if (at_head) { |
| 506 | rq->cmd_flags |= REQ_SOFTBARRIER; |
| 507 | list_add(&rq->queuelist, &q->requeue_list); |
| 508 | } else { |
| 509 | list_add_tail(&rq->queuelist, &q->requeue_list); |
| 510 | } |
| 511 | spin_unlock_irqrestore(&q->requeue_lock, flags); |
| 512 | } |
| 513 | EXPORT_SYMBOL(blk_mq_add_to_requeue_list); |
| 514 | |
Keith Busch | c68ed59 | 2015-01-07 18:55:44 -0700 | [diff] [blame] | 515 | void blk_mq_cancel_requeue_work(struct request_queue *q) |
| 516 | { |
| 517 | cancel_work_sync(&q->requeue_work); |
| 518 | } |
| 519 | EXPORT_SYMBOL_GPL(blk_mq_cancel_requeue_work); |
| 520 | |
Christoph Hellwig | 6fca6a6 | 2014-05-28 08:08:02 -0600 | [diff] [blame] | 521 | void blk_mq_kick_requeue_list(struct request_queue *q) |
| 522 | { |
| 523 | kblockd_schedule_work(&q->requeue_work); |
| 524 | } |
| 525 | EXPORT_SYMBOL(blk_mq_kick_requeue_list); |
| 526 | |
Jens Axboe | 1885b24 | 2015-01-07 18:55:45 -0700 | [diff] [blame] | 527 | void blk_mq_abort_requeue_list(struct request_queue *q) |
| 528 | { |
| 529 | unsigned long flags; |
| 530 | LIST_HEAD(rq_list); |
| 531 | |
| 532 | spin_lock_irqsave(&q->requeue_lock, flags); |
| 533 | list_splice_init(&q->requeue_list, &rq_list); |
| 534 | spin_unlock_irqrestore(&q->requeue_lock, flags); |
| 535 | |
| 536 | while (!list_empty(&rq_list)) { |
| 537 | struct request *rq; |
| 538 | |
| 539 | rq = list_first_entry(&rq_list, struct request, queuelist); |
| 540 | list_del_init(&rq->queuelist); |
| 541 | rq->errors = -EIO; |
| 542 | blk_mq_end_request(rq, rq->errors); |
| 543 | } |
| 544 | } |
| 545 | EXPORT_SYMBOL(blk_mq_abort_requeue_list); |
| 546 | |
Jens Axboe | 0e62f51 | 2014-06-04 10:23:49 -0600 | [diff] [blame] | 547 | struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag) |
| 548 | { |
Ming Lei | 0048b48 | 2015-08-09 03:41:51 -0400 | [diff] [blame] | 549 | return tags->rqs[tag]; |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 550 | } |
| 551 | EXPORT_SYMBOL(blk_mq_tag_to_rq); |
| 552 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 553 | struct blk_mq_timeout_data { |
Christoph Hellwig | 46f92d4 | 2014-09-13 16:40:12 -0700 | [diff] [blame] | 554 | unsigned long next; |
| 555 | unsigned int next_set; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 556 | }; |
| 557 | |
Christoph Hellwig | 9041583 | 2014-09-22 10:21:48 -0600 | [diff] [blame] | 558 | void blk_mq_rq_timed_out(struct request *req, bool reserved) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 559 | { |
Christoph Hellwig | 46f92d4 | 2014-09-13 16:40:12 -0700 | [diff] [blame] | 560 | struct blk_mq_ops *ops = req->q->mq_ops; |
| 561 | enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER; |
Jens Axboe | 87ee7b1 | 2014-04-24 08:51:47 -0600 | [diff] [blame] | 562 | |
| 563 | /* |
| 564 | * We know that complete is set at this point. If STARTED isn't set |
| 565 | * anymore, then the request isn't active and the "timeout" should |
| 566 | * just be ignored. This can happen due to the bitflag ordering. |
| 567 | * Timeout first checks if STARTED is set, and if it is, assumes |
| 568 | * the request is active. But if we race with completion, then |
| 569 | * we both flags will get cleared. So check here again, and ignore |
| 570 | * a timeout event with a request that isn't active. |
| 571 | */ |
Christoph Hellwig | 46f92d4 | 2014-09-13 16:40:12 -0700 | [diff] [blame] | 572 | if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags)) |
| 573 | return; |
Jens Axboe | 87ee7b1 | 2014-04-24 08:51:47 -0600 | [diff] [blame] | 574 | |
Christoph Hellwig | 46f92d4 | 2014-09-13 16:40:12 -0700 | [diff] [blame] | 575 | if (ops->timeout) |
Christoph Hellwig | 0152fb6 | 2014-09-13 16:40:13 -0700 | [diff] [blame] | 576 | ret = ops->timeout(req, reserved); |
Jens Axboe | 87ee7b1 | 2014-04-24 08:51:47 -0600 | [diff] [blame] | 577 | |
Christoph Hellwig | 46f92d4 | 2014-09-13 16:40:12 -0700 | [diff] [blame] | 578 | switch (ret) { |
| 579 | case BLK_EH_HANDLED: |
| 580 | __blk_mq_complete_request(req); |
| 581 | break; |
| 582 | case BLK_EH_RESET_TIMER: |
| 583 | blk_add_timer(req); |
| 584 | blk_clear_rq_complete(req); |
| 585 | break; |
| 586 | case BLK_EH_NOT_HANDLED: |
| 587 | break; |
| 588 | default: |
| 589 | printk(KERN_ERR "block: bad eh return: %d\n", ret); |
| 590 | break; |
| 591 | } |
Jens Axboe | 87ee7b1 | 2014-04-24 08:51:47 -0600 | [diff] [blame] | 592 | } |
Keith Busch | 5b3f25f | 2015-01-07 18:55:46 -0700 | [diff] [blame] | 593 | |
Christoph Hellwig | 81481eb | 2014-09-13 16:40:11 -0700 | [diff] [blame] | 594 | static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx, |
| 595 | struct request *rq, void *priv, bool reserved) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 596 | { |
Christoph Hellwig | 81481eb | 2014-09-13 16:40:11 -0700 | [diff] [blame] | 597 | struct blk_mq_timeout_data *data = priv; |
| 598 | |
Keith Busch | eb130db | 2015-01-08 08:59:53 -0700 | [diff] [blame] | 599 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) { |
| 600 | /* |
| 601 | * If a request wasn't started before the queue was |
| 602 | * marked dying, kill it here or it'll go unnoticed. |
| 603 | */ |
Christoph Hellwig | f4829a9 | 2015-09-27 21:01:50 +0200 | [diff] [blame] | 604 | if (unlikely(blk_queue_dying(rq->q))) |
| 605 | blk_mq_complete_request(rq, -EIO); |
Christoph Hellwig | 46f92d4 | 2014-09-13 16:40:12 -0700 | [diff] [blame] | 606 | return; |
Keith Busch | eb130db | 2015-01-08 08:59:53 -0700 | [diff] [blame] | 607 | } |
Keith Busch | 5b3f25f | 2015-01-07 18:55:46 -0700 | [diff] [blame] | 608 | if (rq->cmd_flags & REQ_NO_TIMEOUT) |
| 609 | return; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 610 | |
Christoph Hellwig | 46f92d4 | 2014-09-13 16:40:12 -0700 | [diff] [blame] | 611 | if (time_after_eq(jiffies, rq->deadline)) { |
| 612 | if (!blk_mark_rq_complete(rq)) |
Christoph Hellwig | 0152fb6 | 2014-09-13 16:40:13 -0700 | [diff] [blame] | 613 | blk_mq_rq_timed_out(rq, reserved); |
Christoph Hellwig | 46f92d4 | 2014-09-13 16:40:12 -0700 | [diff] [blame] | 614 | } else if (!data->next_set || time_after(data->next, rq->deadline)) { |
| 615 | data->next = rq->deadline; |
| 616 | data->next_set = 1; |
| 617 | } |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 618 | } |
| 619 | |
Christoph Hellwig | 81481eb | 2014-09-13 16:40:11 -0700 | [diff] [blame] | 620 | static void blk_mq_rq_timer(unsigned long priv) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 621 | { |
Christoph Hellwig | 81481eb | 2014-09-13 16:40:11 -0700 | [diff] [blame] | 622 | struct request_queue *q = (struct request_queue *)priv; |
| 623 | struct blk_mq_timeout_data data = { |
| 624 | .next = 0, |
| 625 | .next_set = 0, |
| 626 | }; |
Christoph Hellwig | 81481eb | 2014-09-13 16:40:11 -0700 | [diff] [blame] | 627 | int i; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 628 | |
Christoph Hellwig | 0bf6cd5 | 2015-09-27 21:01:51 +0200 | [diff] [blame] | 629 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 630 | |
Christoph Hellwig | 81481eb | 2014-09-13 16:40:11 -0700 | [diff] [blame] | 631 | if (data.next_set) { |
| 632 | data.next = blk_rq_timeout(round_jiffies_up(data.next)); |
| 633 | mod_timer(&q->timeout, data.next); |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 634 | } else { |
Christoph Hellwig | 0bf6cd5 | 2015-09-27 21:01:51 +0200 | [diff] [blame] | 635 | struct blk_mq_hw_ctx *hctx; |
| 636 | |
Ming Lei | f054b56 | 2015-04-21 10:00:19 +0800 | [diff] [blame] | 637 | queue_for_each_hw_ctx(q, hctx, i) { |
| 638 | /* the hctx may be unmapped, so check it here */ |
| 639 | if (blk_mq_hw_queue_mapped(hctx)) |
| 640 | blk_mq_tag_idle(hctx); |
| 641 | } |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 642 | } |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 643 | } |
| 644 | |
| 645 | /* |
| 646 | * Reverse check our software queue for entries that we could potentially |
| 647 | * merge with. Currently includes a hand-wavy stop count of 8, to not spend |
| 648 | * too much time checking for merges. |
| 649 | */ |
| 650 | static bool blk_mq_attempt_merge(struct request_queue *q, |
| 651 | struct blk_mq_ctx *ctx, struct bio *bio) |
| 652 | { |
| 653 | struct request *rq; |
| 654 | int checked = 8; |
| 655 | |
| 656 | list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) { |
| 657 | int el_ret; |
| 658 | |
| 659 | if (!checked--) |
| 660 | break; |
| 661 | |
| 662 | if (!blk_rq_merge_ok(rq, bio)) |
| 663 | continue; |
| 664 | |
| 665 | el_ret = blk_try_merge(rq, bio); |
| 666 | if (el_ret == ELEVATOR_BACK_MERGE) { |
| 667 | if (bio_attempt_back_merge(q, rq, bio)) { |
| 668 | ctx->rq_merged++; |
| 669 | return true; |
| 670 | } |
| 671 | break; |
| 672 | } else if (el_ret == ELEVATOR_FRONT_MERGE) { |
| 673 | if (bio_attempt_front_merge(q, rq, bio)) { |
| 674 | ctx->rq_merged++; |
| 675 | return true; |
| 676 | } |
| 677 | break; |
| 678 | } |
| 679 | } |
| 680 | |
| 681 | return false; |
| 682 | } |
| 683 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 684 | /* |
Jens Axboe | 1429d7c | 2014-05-19 09:23:55 -0600 | [diff] [blame] | 685 | * Process software queues that have been marked busy, splicing them |
| 686 | * to the for-dispatch |
| 687 | */ |
| 688 | static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list) |
| 689 | { |
| 690 | struct blk_mq_ctx *ctx; |
| 691 | int i; |
| 692 | |
Jens Axboe | 569fd0c | 2015-04-17 08:28:50 -0600 | [diff] [blame] | 693 | for (i = 0; i < hctx->ctx_map.size; i++) { |
Jens Axboe | 1429d7c | 2014-05-19 09:23:55 -0600 | [diff] [blame] | 694 | struct blk_align_bitmap *bm = &hctx->ctx_map.map[i]; |
| 695 | unsigned int off, bit; |
| 696 | |
| 697 | if (!bm->word) |
| 698 | continue; |
| 699 | |
| 700 | bit = 0; |
| 701 | off = i * hctx->ctx_map.bits_per_word; |
| 702 | do { |
| 703 | bit = find_next_bit(&bm->word, bm->depth, bit); |
| 704 | if (bit >= bm->depth) |
| 705 | break; |
| 706 | |
| 707 | ctx = hctx->ctxs[bit + off]; |
| 708 | clear_bit(bit, &bm->word); |
| 709 | spin_lock(&ctx->lock); |
| 710 | list_splice_tail_init(&ctx->rq_list, list); |
| 711 | spin_unlock(&ctx->lock); |
| 712 | |
| 713 | bit++; |
| 714 | } while (1); |
| 715 | } |
| 716 | } |
| 717 | |
| 718 | /* |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 719 | * Run this hardware queue, pulling any software queues mapped to it in. |
| 720 | * Note that this function currently has various problems around ordering |
| 721 | * of IO. In particular, we'd like FIFO behaviour on handling existing |
| 722 | * items on the hctx->dispatch list. Ignore that for now. |
| 723 | */ |
| 724 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx) |
| 725 | { |
| 726 | struct request_queue *q = hctx->queue; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 727 | struct request *rq; |
| 728 | LIST_HEAD(rq_list); |
Jens Axboe | 74c4505 | 2014-10-29 11:14:52 -0600 | [diff] [blame] | 729 | LIST_HEAD(driver_list); |
| 730 | struct list_head *dptr; |
Jens Axboe | 1429d7c | 2014-05-19 09:23:55 -0600 | [diff] [blame] | 731 | int queued; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 732 | |
Jens Axboe | fd1270d | 2014-04-16 09:23:48 -0600 | [diff] [blame] | 733 | WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask)); |
Jens Axboe | e4043dc | 2014-04-09 10:18:23 -0600 | [diff] [blame] | 734 | |
Jens Axboe | 5d12f90 | 2014-03-19 15:25:02 -0600 | [diff] [blame] | 735 | if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state))) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 736 | return; |
| 737 | |
| 738 | hctx->run++; |
| 739 | |
| 740 | /* |
| 741 | * Touch any software queue that has pending entries. |
| 742 | */ |
Jens Axboe | 1429d7c | 2014-05-19 09:23:55 -0600 | [diff] [blame] | 743 | flush_busy_ctxs(hctx, &rq_list); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 744 | |
| 745 | /* |
| 746 | * If we have previous entries on our dispatch list, grab them |
| 747 | * and stuff them at the front for more fair dispatch. |
| 748 | */ |
| 749 | if (!list_empty_careful(&hctx->dispatch)) { |
| 750 | spin_lock(&hctx->lock); |
| 751 | if (!list_empty(&hctx->dispatch)) |
| 752 | list_splice_init(&hctx->dispatch, &rq_list); |
| 753 | spin_unlock(&hctx->lock); |
| 754 | } |
| 755 | |
| 756 | /* |
Jens Axboe | 74c4505 | 2014-10-29 11:14:52 -0600 | [diff] [blame] | 757 | * Start off with dptr being NULL, so we start the first request |
| 758 | * immediately, even if we have more pending. |
| 759 | */ |
| 760 | dptr = NULL; |
| 761 | |
| 762 | /* |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 763 | * Now process all the entries, sending them to the driver. |
| 764 | */ |
Jens Axboe | 1429d7c | 2014-05-19 09:23:55 -0600 | [diff] [blame] | 765 | queued = 0; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 766 | while (!list_empty(&rq_list)) { |
Jens Axboe | 74c4505 | 2014-10-29 11:14:52 -0600 | [diff] [blame] | 767 | struct blk_mq_queue_data bd; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 768 | int ret; |
| 769 | |
| 770 | rq = list_first_entry(&rq_list, struct request, queuelist); |
| 771 | list_del_init(&rq->queuelist); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 772 | |
Jens Axboe | 74c4505 | 2014-10-29 11:14:52 -0600 | [diff] [blame] | 773 | bd.rq = rq; |
| 774 | bd.list = dptr; |
| 775 | bd.last = list_empty(&rq_list); |
| 776 | |
| 777 | ret = q->mq_ops->queue_rq(hctx, &bd); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 778 | switch (ret) { |
| 779 | case BLK_MQ_RQ_QUEUE_OK: |
| 780 | queued++; |
| 781 | continue; |
| 782 | case BLK_MQ_RQ_QUEUE_BUSY: |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 783 | list_add(&rq->queuelist, &rq_list); |
Christoph Hellwig | ed0791b | 2014-04-16 09:44:57 +0200 | [diff] [blame] | 784 | __blk_mq_requeue_request(rq); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 785 | break; |
| 786 | default: |
| 787 | pr_err("blk-mq: bad return on queue: %d\n", ret); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 788 | case BLK_MQ_RQ_QUEUE_ERROR: |
Christoph Hellwig | 1e93b8c | 2014-02-11 08:27:13 -0800 | [diff] [blame] | 789 | rq->errors = -EIO; |
Christoph Hellwig | c8a446a | 2014-09-13 16:40:10 -0700 | [diff] [blame] | 790 | blk_mq_end_request(rq, rq->errors); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 791 | break; |
| 792 | } |
| 793 | |
| 794 | if (ret == BLK_MQ_RQ_QUEUE_BUSY) |
| 795 | break; |
Jens Axboe | 74c4505 | 2014-10-29 11:14:52 -0600 | [diff] [blame] | 796 | |
| 797 | /* |
| 798 | * We've done the first request. If we have more than 1 |
| 799 | * left in the list, set dptr to defer issue. |
| 800 | */ |
| 801 | if (!dptr && rq_list.next != rq_list.prev) |
| 802 | dptr = &driver_list; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 803 | } |
| 804 | |
| 805 | if (!queued) |
| 806 | hctx->dispatched[0]++; |
| 807 | else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1))) |
| 808 | hctx->dispatched[ilog2(queued) + 1]++; |
| 809 | |
| 810 | /* |
| 811 | * Any items that need requeuing? Stuff them into hctx->dispatch, |
| 812 | * that is where we will continue on next queue run. |
| 813 | */ |
| 814 | if (!list_empty(&rq_list)) { |
| 815 | spin_lock(&hctx->lock); |
| 816 | list_splice(&rq_list, &hctx->dispatch); |
| 817 | spin_unlock(&hctx->lock); |
Shaohua Li | 9ba52e5 | 2015-05-04 14:32:48 -0600 | [diff] [blame] | 818 | /* |
| 819 | * the queue is expected stopped with BLK_MQ_RQ_QUEUE_BUSY, but |
| 820 | * it's possible the queue is stopped and restarted again |
| 821 | * before this. Queue restart will dispatch requests. And since |
| 822 | * requests in rq_list aren't added into hctx->dispatch yet, |
| 823 | * the requests in rq_list might get lost. |
| 824 | * |
| 825 | * blk_mq_run_hw_queue() already checks the STOPPED bit |
| 826 | **/ |
| 827 | blk_mq_run_hw_queue(hctx, true); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 828 | } |
| 829 | } |
| 830 | |
Jens Axboe | 506e931 | 2014-05-07 10:26:44 -0600 | [diff] [blame] | 831 | /* |
| 832 | * It'd be great if the workqueue API had a way to pass |
| 833 | * in a mask and had some smarts for more clever placement. |
| 834 | * For now we just round-robin here, switching for every |
| 835 | * BLK_MQ_CPU_WORK_BATCH queued items. |
| 836 | */ |
| 837 | static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx) |
| 838 | { |
Christoph Hellwig | b657d7e | 2014-11-24 09:27:23 +0100 | [diff] [blame] | 839 | if (hctx->queue->nr_hw_queues == 1) |
| 840 | return WORK_CPU_UNBOUND; |
Jens Axboe | 506e931 | 2014-05-07 10:26:44 -0600 | [diff] [blame] | 841 | |
| 842 | if (--hctx->next_cpu_batch <= 0) { |
Christoph Hellwig | b657d7e | 2014-11-24 09:27:23 +0100 | [diff] [blame] | 843 | int cpu = hctx->next_cpu, next_cpu; |
Jens Axboe | 506e931 | 2014-05-07 10:26:44 -0600 | [diff] [blame] | 844 | |
| 845 | next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask); |
| 846 | if (next_cpu >= nr_cpu_ids) |
| 847 | next_cpu = cpumask_first(hctx->cpumask); |
| 848 | |
| 849 | hctx->next_cpu = next_cpu; |
| 850 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; |
Christoph Hellwig | b657d7e | 2014-11-24 09:27:23 +0100 | [diff] [blame] | 851 | |
| 852 | return cpu; |
Jens Axboe | 506e931 | 2014-05-07 10:26:44 -0600 | [diff] [blame] | 853 | } |
| 854 | |
Christoph Hellwig | b657d7e | 2014-11-24 09:27:23 +0100 | [diff] [blame] | 855 | return hctx->next_cpu; |
Jens Axboe | 506e931 | 2014-05-07 10:26:44 -0600 | [diff] [blame] | 856 | } |
| 857 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 858 | void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
| 859 | { |
Ming Lei | 19c66e5 | 2014-12-03 19:38:04 +0800 | [diff] [blame] | 860 | if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state) || |
| 861 | !blk_mq_hw_queue_mapped(hctx))) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 862 | return; |
| 863 | |
Paolo Bonzini | 398205b | 2014-11-07 23:03:59 +0100 | [diff] [blame] | 864 | if (!async) { |
Paolo Bonzini | 2a90d4a | 2014-11-07 23:04:00 +0100 | [diff] [blame] | 865 | int cpu = get_cpu(); |
| 866 | if (cpumask_test_cpu(cpu, hctx->cpumask)) { |
Paolo Bonzini | 398205b | 2014-11-07 23:03:59 +0100 | [diff] [blame] | 867 | __blk_mq_run_hw_queue(hctx); |
Paolo Bonzini | 2a90d4a | 2014-11-07 23:04:00 +0100 | [diff] [blame] | 868 | put_cpu(); |
Paolo Bonzini | 398205b | 2014-11-07 23:03:59 +0100 | [diff] [blame] | 869 | return; |
| 870 | } |
Jens Axboe | e4043dc | 2014-04-09 10:18:23 -0600 | [diff] [blame] | 871 | |
Paolo Bonzini | 2a90d4a | 2014-11-07 23:04:00 +0100 | [diff] [blame] | 872 | put_cpu(); |
Jens Axboe | e4043dc | 2014-04-09 10:18:23 -0600 | [diff] [blame] | 873 | } |
Paolo Bonzini | 398205b | 2014-11-07 23:03:59 +0100 | [diff] [blame] | 874 | |
Christoph Hellwig | b657d7e | 2014-11-24 09:27:23 +0100 | [diff] [blame] | 875 | kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx), |
| 876 | &hctx->run_work, 0); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 877 | } |
| 878 | |
Mike Snitzer | b94ec29 | 2015-03-11 23:56:38 -0400 | [diff] [blame] | 879 | void blk_mq_run_hw_queues(struct request_queue *q, bool async) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 880 | { |
| 881 | struct blk_mq_hw_ctx *hctx; |
| 882 | int i; |
| 883 | |
| 884 | queue_for_each_hw_ctx(q, hctx, i) { |
| 885 | if ((!blk_mq_hctx_has_pending(hctx) && |
| 886 | list_empty_careful(&hctx->dispatch)) || |
Jens Axboe | 5d12f90 | 2014-03-19 15:25:02 -0600 | [diff] [blame] | 887 | test_bit(BLK_MQ_S_STOPPED, &hctx->state)) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 888 | continue; |
| 889 | |
Mike Snitzer | b94ec29 | 2015-03-11 23:56:38 -0400 | [diff] [blame] | 890 | blk_mq_run_hw_queue(hctx, async); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 891 | } |
| 892 | } |
Mike Snitzer | b94ec29 | 2015-03-11 23:56:38 -0400 | [diff] [blame] | 893 | EXPORT_SYMBOL(blk_mq_run_hw_queues); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 894 | |
| 895 | void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) |
| 896 | { |
Christoph Hellwig | 70f4db6 | 2014-04-16 10:48:08 -0600 | [diff] [blame] | 897 | cancel_delayed_work(&hctx->run_work); |
| 898 | cancel_delayed_work(&hctx->delay_work); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 899 | set_bit(BLK_MQ_S_STOPPED, &hctx->state); |
| 900 | } |
| 901 | EXPORT_SYMBOL(blk_mq_stop_hw_queue); |
| 902 | |
Christoph Hellwig | 280d45f | 2013-10-25 14:45:58 +0100 | [diff] [blame] | 903 | void blk_mq_stop_hw_queues(struct request_queue *q) |
| 904 | { |
| 905 | struct blk_mq_hw_ctx *hctx; |
| 906 | int i; |
| 907 | |
| 908 | queue_for_each_hw_ctx(q, hctx, i) |
| 909 | blk_mq_stop_hw_queue(hctx); |
| 910 | } |
| 911 | EXPORT_SYMBOL(blk_mq_stop_hw_queues); |
| 912 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 913 | void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) |
| 914 | { |
| 915 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); |
Jens Axboe | e4043dc | 2014-04-09 10:18:23 -0600 | [diff] [blame] | 916 | |
Jens Axboe | 0ffbce8 | 2014-06-25 08:22:34 -0600 | [diff] [blame] | 917 | blk_mq_run_hw_queue(hctx, false); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 918 | } |
| 919 | EXPORT_SYMBOL(blk_mq_start_hw_queue); |
| 920 | |
Christoph Hellwig | 2f26855 | 2014-04-16 09:44:56 +0200 | [diff] [blame] | 921 | void blk_mq_start_hw_queues(struct request_queue *q) |
| 922 | { |
| 923 | struct blk_mq_hw_ctx *hctx; |
| 924 | int i; |
| 925 | |
| 926 | queue_for_each_hw_ctx(q, hctx, i) |
| 927 | blk_mq_start_hw_queue(hctx); |
| 928 | } |
| 929 | EXPORT_SYMBOL(blk_mq_start_hw_queues); |
| 930 | |
Christoph Hellwig | 1b4a325 | 2014-04-16 09:44:54 +0200 | [diff] [blame] | 931 | void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 932 | { |
| 933 | struct blk_mq_hw_ctx *hctx; |
| 934 | int i; |
| 935 | |
| 936 | queue_for_each_hw_ctx(q, hctx, i) { |
| 937 | if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state)) |
| 938 | continue; |
| 939 | |
| 940 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); |
Christoph Hellwig | 1b4a325 | 2014-04-16 09:44:54 +0200 | [diff] [blame] | 941 | blk_mq_run_hw_queue(hctx, async); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 942 | } |
| 943 | } |
| 944 | EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); |
| 945 | |
Christoph Hellwig | 70f4db6 | 2014-04-16 10:48:08 -0600 | [diff] [blame] | 946 | static void blk_mq_run_work_fn(struct work_struct *work) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 947 | { |
| 948 | struct blk_mq_hw_ctx *hctx; |
| 949 | |
Christoph Hellwig | 70f4db6 | 2014-04-16 10:48:08 -0600 | [diff] [blame] | 950 | hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work); |
Jens Axboe | e4043dc | 2014-04-09 10:18:23 -0600 | [diff] [blame] | 951 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 952 | __blk_mq_run_hw_queue(hctx); |
| 953 | } |
| 954 | |
Christoph Hellwig | 70f4db6 | 2014-04-16 10:48:08 -0600 | [diff] [blame] | 955 | static void blk_mq_delay_work_fn(struct work_struct *work) |
| 956 | { |
| 957 | struct blk_mq_hw_ctx *hctx; |
| 958 | |
| 959 | hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work); |
| 960 | |
| 961 | if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state)) |
| 962 | __blk_mq_run_hw_queue(hctx); |
| 963 | } |
| 964 | |
| 965 | void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) |
| 966 | { |
Ming Lei | 19c66e5 | 2014-12-03 19:38:04 +0800 | [diff] [blame] | 967 | if (unlikely(!blk_mq_hw_queue_mapped(hctx))) |
| 968 | return; |
Christoph Hellwig | 70f4db6 | 2014-04-16 10:48:08 -0600 | [diff] [blame] | 969 | |
Christoph Hellwig | b657d7e | 2014-11-24 09:27:23 +0100 | [diff] [blame] | 970 | kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx), |
| 971 | &hctx->delay_work, msecs_to_jiffies(msecs)); |
Christoph Hellwig | 70f4db6 | 2014-04-16 10:48:08 -0600 | [diff] [blame] | 972 | } |
| 973 | EXPORT_SYMBOL(blk_mq_delay_queue); |
| 974 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 975 | static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, |
Christoph Hellwig | 72a0a36 | 2014-02-07 10:22:36 -0800 | [diff] [blame] | 976 | struct request *rq, bool at_head) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 977 | { |
| 978 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
| 979 | |
Jens Axboe | 01b983c | 2013-11-19 18:59:10 -0700 | [diff] [blame] | 980 | trace_block_rq_insert(hctx->queue, rq); |
| 981 | |
Christoph Hellwig | 72a0a36 | 2014-02-07 10:22:36 -0800 | [diff] [blame] | 982 | if (at_head) |
| 983 | list_add(&rq->queuelist, &ctx->rq_list); |
| 984 | else |
| 985 | list_add_tail(&rq->queuelist, &ctx->rq_list); |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 986 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 987 | blk_mq_hctx_mark_pending(hctx, ctx); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 988 | } |
| 989 | |
Christoph Hellwig | eeabc85 | 2014-03-21 08:57:37 -0600 | [diff] [blame] | 990 | void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue, |
| 991 | bool async) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 992 | { |
| 993 | struct request_queue *q = rq->q; |
| 994 | struct blk_mq_hw_ctx *hctx; |
Christoph Hellwig | eeabc85 | 2014-03-21 08:57:37 -0600 | [diff] [blame] | 995 | struct blk_mq_ctx *ctx = rq->mq_ctx, *current_ctx; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 996 | |
| 997 | current_ctx = blk_mq_get_ctx(q); |
Christoph Hellwig | eeabc85 | 2014-03-21 08:57:37 -0600 | [diff] [blame] | 998 | if (!cpu_online(ctx->cpu)) |
| 999 | rq->mq_ctx = ctx = current_ctx; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1000 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1001 | hctx = q->mq_ops->map_queue(q, ctx->cpu); |
| 1002 | |
Christoph Hellwig | a57a178 | 2014-09-16 14:44:07 -0700 | [diff] [blame] | 1003 | spin_lock(&ctx->lock); |
| 1004 | __blk_mq_insert_request(hctx, rq, at_head); |
| 1005 | spin_unlock(&ctx->lock); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1006 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1007 | if (run_queue) |
| 1008 | blk_mq_run_hw_queue(hctx, async); |
Jens Axboe | e4043dc | 2014-04-09 10:18:23 -0600 | [diff] [blame] | 1009 | |
| 1010 | blk_mq_put_ctx(current_ctx); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1011 | } |
| 1012 | |
| 1013 | static void blk_mq_insert_requests(struct request_queue *q, |
| 1014 | struct blk_mq_ctx *ctx, |
| 1015 | struct list_head *list, |
| 1016 | int depth, |
| 1017 | bool from_schedule) |
| 1018 | |
| 1019 | { |
| 1020 | struct blk_mq_hw_ctx *hctx; |
| 1021 | struct blk_mq_ctx *current_ctx; |
| 1022 | |
| 1023 | trace_block_unplug(q, depth, !from_schedule); |
| 1024 | |
| 1025 | current_ctx = blk_mq_get_ctx(q); |
| 1026 | |
| 1027 | if (!cpu_online(ctx->cpu)) |
| 1028 | ctx = current_ctx; |
| 1029 | hctx = q->mq_ops->map_queue(q, ctx->cpu); |
| 1030 | |
| 1031 | /* |
| 1032 | * preemption doesn't flush plug list, so it's possible ctx->cpu is |
| 1033 | * offline now |
| 1034 | */ |
| 1035 | spin_lock(&ctx->lock); |
| 1036 | while (!list_empty(list)) { |
| 1037 | struct request *rq; |
| 1038 | |
| 1039 | rq = list_first_entry(list, struct request, queuelist); |
| 1040 | list_del_init(&rq->queuelist); |
| 1041 | rq->mq_ctx = ctx; |
Christoph Hellwig | 72a0a36 | 2014-02-07 10:22:36 -0800 | [diff] [blame] | 1042 | __blk_mq_insert_request(hctx, rq, false); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1043 | } |
| 1044 | spin_unlock(&ctx->lock); |
| 1045 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1046 | blk_mq_run_hw_queue(hctx, from_schedule); |
Jens Axboe | e4043dc | 2014-04-09 10:18:23 -0600 | [diff] [blame] | 1047 | blk_mq_put_ctx(current_ctx); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1048 | } |
| 1049 | |
| 1050 | static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b) |
| 1051 | { |
| 1052 | struct request *rqa = container_of(a, struct request, queuelist); |
| 1053 | struct request *rqb = container_of(b, struct request, queuelist); |
| 1054 | |
| 1055 | return !(rqa->mq_ctx < rqb->mq_ctx || |
| 1056 | (rqa->mq_ctx == rqb->mq_ctx && |
| 1057 | blk_rq_pos(rqa) < blk_rq_pos(rqb))); |
| 1058 | } |
| 1059 | |
| 1060 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) |
| 1061 | { |
| 1062 | struct blk_mq_ctx *this_ctx; |
| 1063 | struct request_queue *this_q; |
| 1064 | struct request *rq; |
| 1065 | LIST_HEAD(list); |
| 1066 | LIST_HEAD(ctx_list); |
| 1067 | unsigned int depth; |
| 1068 | |
| 1069 | list_splice_init(&plug->mq_list, &list); |
| 1070 | |
| 1071 | list_sort(NULL, &list, plug_ctx_cmp); |
| 1072 | |
| 1073 | this_q = NULL; |
| 1074 | this_ctx = NULL; |
| 1075 | depth = 0; |
| 1076 | |
| 1077 | while (!list_empty(&list)) { |
| 1078 | rq = list_entry_rq(list.next); |
| 1079 | list_del_init(&rq->queuelist); |
| 1080 | BUG_ON(!rq->q); |
| 1081 | if (rq->mq_ctx != this_ctx) { |
| 1082 | if (this_ctx) { |
| 1083 | blk_mq_insert_requests(this_q, this_ctx, |
| 1084 | &ctx_list, depth, |
| 1085 | from_schedule); |
| 1086 | } |
| 1087 | |
| 1088 | this_ctx = rq->mq_ctx; |
| 1089 | this_q = rq->q; |
| 1090 | depth = 0; |
| 1091 | } |
| 1092 | |
| 1093 | depth++; |
| 1094 | list_add_tail(&rq->queuelist, &ctx_list); |
| 1095 | } |
| 1096 | |
| 1097 | /* |
| 1098 | * If 'this_ctx' is set, we know we have entries to complete |
| 1099 | * on 'ctx_list'. Do those. |
| 1100 | */ |
| 1101 | if (this_ctx) { |
| 1102 | blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth, |
| 1103 | from_schedule); |
| 1104 | } |
| 1105 | } |
| 1106 | |
| 1107 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio) |
| 1108 | { |
| 1109 | init_request_from_bio(rq, bio); |
Jens Axboe | 4b57052 | 2014-05-29 11:00:11 -0600 | [diff] [blame] | 1110 | |
Jens Axboe | 3ee3237 | 2014-06-09 09:36:53 -0600 | [diff] [blame] | 1111 | if (blk_do_io_stat(rq)) |
Jens Axboe | 4b57052 | 2014-05-29 11:00:11 -0600 | [diff] [blame] | 1112 | blk_account_io_start(rq, 1); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1113 | } |
| 1114 | |
Jens Axboe | 274a584 | 2014-08-15 12:44:08 -0600 | [diff] [blame] | 1115 | static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx) |
| 1116 | { |
| 1117 | return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) && |
| 1118 | !blk_queue_nomerges(hctx->queue); |
| 1119 | } |
| 1120 | |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1121 | static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx, |
| 1122 | struct blk_mq_ctx *ctx, |
| 1123 | struct request *rq, struct bio *bio) |
| 1124 | { |
Jens Axboe | 274a584 | 2014-08-15 12:44:08 -0600 | [diff] [blame] | 1125 | if (!hctx_allow_merges(hctx)) { |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1126 | blk_mq_bio_to_request(rq, bio); |
| 1127 | spin_lock(&ctx->lock); |
| 1128 | insert_rq: |
| 1129 | __blk_mq_insert_request(hctx, rq, false); |
| 1130 | spin_unlock(&ctx->lock); |
| 1131 | return false; |
| 1132 | } else { |
Jens Axboe | 274a584 | 2014-08-15 12:44:08 -0600 | [diff] [blame] | 1133 | struct request_queue *q = hctx->queue; |
| 1134 | |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1135 | spin_lock(&ctx->lock); |
| 1136 | if (!blk_mq_attempt_merge(q, ctx, bio)) { |
| 1137 | blk_mq_bio_to_request(rq, bio); |
| 1138 | goto insert_rq; |
| 1139 | } |
| 1140 | |
| 1141 | spin_unlock(&ctx->lock); |
| 1142 | __blk_mq_free_request(hctx, ctx, rq); |
| 1143 | return true; |
| 1144 | } |
| 1145 | } |
| 1146 | |
| 1147 | struct blk_map_ctx { |
| 1148 | struct blk_mq_hw_ctx *hctx; |
| 1149 | struct blk_mq_ctx *ctx; |
| 1150 | }; |
| 1151 | |
| 1152 | static struct request *blk_mq_map_request(struct request_queue *q, |
| 1153 | struct bio *bio, |
| 1154 | struct blk_map_ctx *data) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1155 | { |
| 1156 | struct blk_mq_hw_ctx *hctx; |
| 1157 | struct blk_mq_ctx *ctx; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1158 | struct request *rq; |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1159 | int rw = bio_data_dir(bio); |
Ming Lei | cb96a42 | 2014-06-01 00:43:37 +0800 | [diff] [blame] | 1160 | struct blk_mq_alloc_data alloc_data; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1161 | |
Dan Williams | 3ef28e8 | 2015-10-21 13:20:12 -0400 | [diff] [blame^] | 1162 | blk_queue_enter_live(q); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1163 | ctx = blk_mq_get_ctx(q); |
| 1164 | hctx = q->mq_ops->map_queue(q, ctx->cpu); |
| 1165 | |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1166 | if (rw_is_sync(bio->bi_rw)) |
Shaohua Li | 27fbf4e | 2014-02-19 20:20:21 +0800 | [diff] [blame] | 1167 | rw |= REQ_SYNC; |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1168 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1169 | trace_block_getrq(q, bio, rw); |
Ming Lei | cb96a42 | 2014-06-01 00:43:37 +0800 | [diff] [blame] | 1170 | blk_mq_set_alloc_data(&alloc_data, q, GFP_ATOMIC, false, ctx, |
| 1171 | hctx); |
| 1172 | rq = __blk_mq_alloc_request(&alloc_data, rw); |
Christoph Hellwig | 5dee857 | 2014-05-27 20:59:47 +0200 | [diff] [blame] | 1173 | if (unlikely(!rq)) { |
Christoph Hellwig | 793597a | 2014-05-27 20:59:49 +0200 | [diff] [blame] | 1174 | __blk_mq_run_hw_queue(hctx); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1175 | blk_mq_put_ctx(ctx); |
| 1176 | trace_block_sleeprq(q, bio, rw); |
Christoph Hellwig | 793597a | 2014-05-27 20:59:49 +0200 | [diff] [blame] | 1177 | |
| 1178 | ctx = blk_mq_get_ctx(q); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1179 | hctx = q->mq_ops->map_queue(q, ctx->cpu); |
Ming Lei | cb96a42 | 2014-06-01 00:43:37 +0800 | [diff] [blame] | 1180 | blk_mq_set_alloc_data(&alloc_data, q, |
| 1181 | __GFP_WAIT|GFP_ATOMIC, false, ctx, hctx); |
| 1182 | rq = __blk_mq_alloc_request(&alloc_data, rw); |
| 1183 | ctx = alloc_data.ctx; |
| 1184 | hctx = alloc_data.hctx; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1185 | } |
| 1186 | |
| 1187 | hctx->queued++; |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1188 | data->hctx = hctx; |
| 1189 | data->ctx = ctx; |
| 1190 | return rq; |
| 1191 | } |
| 1192 | |
Shaohua Li | f984df1 | 2015-05-08 10:51:32 -0700 | [diff] [blame] | 1193 | static int blk_mq_direct_issue_request(struct request *rq) |
| 1194 | { |
| 1195 | int ret; |
| 1196 | struct request_queue *q = rq->q; |
| 1197 | struct blk_mq_hw_ctx *hctx = q->mq_ops->map_queue(q, |
| 1198 | rq->mq_ctx->cpu); |
| 1199 | struct blk_mq_queue_data bd = { |
| 1200 | .rq = rq, |
| 1201 | .list = NULL, |
| 1202 | .last = 1 |
| 1203 | }; |
| 1204 | |
| 1205 | /* |
| 1206 | * For OK queue, we are done. For error, kill it. Any other |
| 1207 | * error (busy), just add it to our list as we previously |
| 1208 | * would have done |
| 1209 | */ |
| 1210 | ret = q->mq_ops->queue_rq(hctx, &bd); |
| 1211 | if (ret == BLK_MQ_RQ_QUEUE_OK) |
| 1212 | return 0; |
| 1213 | else { |
| 1214 | __blk_mq_requeue_request(rq); |
| 1215 | |
| 1216 | if (ret == BLK_MQ_RQ_QUEUE_ERROR) { |
| 1217 | rq->errors = -EIO; |
| 1218 | blk_mq_end_request(rq, rq->errors); |
| 1219 | return 0; |
| 1220 | } |
| 1221 | return -1; |
| 1222 | } |
| 1223 | } |
| 1224 | |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1225 | /* |
| 1226 | * Multiple hardware queue variant. This will not use per-process plugs, |
| 1227 | * but will attempt to bypass the hctx queueing if we can go straight to |
| 1228 | * hardware for SYNC IO. |
| 1229 | */ |
| 1230 | static void blk_mq_make_request(struct request_queue *q, struct bio *bio) |
| 1231 | { |
| 1232 | const int is_sync = rw_is_sync(bio->bi_rw); |
| 1233 | const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA); |
| 1234 | struct blk_map_ctx data; |
| 1235 | struct request *rq; |
Shaohua Li | f984df1 | 2015-05-08 10:51:32 -0700 | [diff] [blame] | 1236 | unsigned int request_count = 0; |
| 1237 | struct blk_plug *plug; |
Shaohua Li | 5b3f341 | 2015-05-08 10:51:33 -0700 | [diff] [blame] | 1238 | struct request *same_queue_rq = NULL; |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1239 | |
| 1240 | blk_queue_bounce(q, &bio); |
| 1241 | |
| 1242 | if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) { |
Christoph Hellwig | 4246a0b | 2015-07-20 15:29:37 +0200 | [diff] [blame] | 1243 | bio_io_error(bio); |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1244 | return; |
| 1245 | } |
| 1246 | |
Kent Overstreet | 54efd50 | 2015-04-23 22:37:18 -0700 | [diff] [blame] | 1247 | blk_queue_split(q, &bio, q->bio_split); |
| 1248 | |
Shaohua Li | f984df1 | 2015-05-08 10:51:32 -0700 | [diff] [blame] | 1249 | if (!is_flush_fua && !blk_queue_nomerges(q) && |
Shaohua Li | 5b3f341 | 2015-05-08 10:51:33 -0700 | [diff] [blame] | 1250 | blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq)) |
Shaohua Li | f984df1 | 2015-05-08 10:51:32 -0700 | [diff] [blame] | 1251 | return; |
| 1252 | |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1253 | rq = blk_mq_map_request(q, bio, &data); |
| 1254 | if (unlikely(!rq)) |
| 1255 | return; |
| 1256 | |
| 1257 | if (unlikely(is_flush_fua)) { |
| 1258 | blk_mq_bio_to_request(rq, bio); |
| 1259 | blk_insert_flush(rq); |
| 1260 | goto run_queue; |
| 1261 | } |
| 1262 | |
Shaohua Li | f984df1 | 2015-05-08 10:51:32 -0700 | [diff] [blame] | 1263 | plug = current->plug; |
Jens Axboe | e167dfb | 2014-10-29 11:18:26 -0600 | [diff] [blame] | 1264 | /* |
| 1265 | * If the driver supports defer issued based on 'last', then |
| 1266 | * queue it up like normal since we can potentially save some |
| 1267 | * CPU this way. |
| 1268 | */ |
Shaohua Li | f984df1 | 2015-05-08 10:51:32 -0700 | [diff] [blame] | 1269 | if (((plug && !blk_queue_nomerges(q)) || is_sync) && |
| 1270 | !(data.hctx->flags & BLK_MQ_F_DEFER_ISSUE)) { |
| 1271 | struct request *old_rq = NULL; |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1272 | |
| 1273 | blk_mq_bio_to_request(rq, bio); |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1274 | |
| 1275 | /* |
Shaohua Li | f984df1 | 2015-05-08 10:51:32 -0700 | [diff] [blame] | 1276 | * we do limited pluging. If bio can be merged, do merge. |
| 1277 | * Otherwise the existing request in the plug list will be |
| 1278 | * issued. So the plug list will have one request at most |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1279 | */ |
Shaohua Li | f984df1 | 2015-05-08 10:51:32 -0700 | [diff] [blame] | 1280 | if (plug) { |
Shaohua Li | 5b3f341 | 2015-05-08 10:51:33 -0700 | [diff] [blame] | 1281 | /* |
| 1282 | * The plug list might get flushed before this. If that |
| 1283 | * happens, same_queue_rq is invalid and plug list is empty |
| 1284 | **/ |
| 1285 | if (same_queue_rq && !list_empty(&plug->mq_list)) { |
| 1286 | old_rq = same_queue_rq; |
Shaohua Li | f984df1 | 2015-05-08 10:51:32 -0700 | [diff] [blame] | 1287 | list_del_init(&old_rq->queuelist); |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1288 | } |
Shaohua Li | f984df1 | 2015-05-08 10:51:32 -0700 | [diff] [blame] | 1289 | list_add_tail(&rq->queuelist, &plug->mq_list); |
| 1290 | } else /* is_sync */ |
| 1291 | old_rq = rq; |
| 1292 | blk_mq_put_ctx(data.ctx); |
| 1293 | if (!old_rq) |
Shaohua Li | 239ad21 | 2015-05-08 10:51:31 -0700 | [diff] [blame] | 1294 | return; |
Shaohua Li | f984df1 | 2015-05-08 10:51:32 -0700 | [diff] [blame] | 1295 | if (!blk_mq_direct_issue_request(old_rq)) |
| 1296 | return; |
| 1297 | blk_mq_insert_request(old_rq, false, true, true); |
| 1298 | return; |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1299 | } |
| 1300 | |
| 1301 | if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) { |
| 1302 | /* |
| 1303 | * For a SYNC request, send it to the hardware immediately. For |
| 1304 | * an ASYNC request, just ensure that we run it later on. The |
| 1305 | * latter allows for merging opportunities and more efficient |
| 1306 | * dispatching. |
| 1307 | */ |
| 1308 | run_queue: |
| 1309 | blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua); |
| 1310 | } |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1311 | blk_mq_put_ctx(data.ctx); |
| 1312 | } |
| 1313 | |
| 1314 | /* |
| 1315 | * Single hardware queue variant. This will attempt to use any per-process |
| 1316 | * plug for merging and IO deferral. |
| 1317 | */ |
| 1318 | static void blk_sq_make_request(struct request_queue *q, struct bio *bio) |
| 1319 | { |
| 1320 | const int is_sync = rw_is_sync(bio->bi_rw); |
| 1321 | const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA); |
Jeff Moyer | e6c4438 | 2015-05-08 10:51:30 -0700 | [diff] [blame] | 1322 | struct blk_plug *plug; |
| 1323 | unsigned int request_count = 0; |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1324 | struct blk_map_ctx data; |
| 1325 | struct request *rq; |
| 1326 | |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1327 | blk_queue_bounce(q, &bio); |
| 1328 | |
| 1329 | if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) { |
Christoph Hellwig | 4246a0b | 2015-07-20 15:29:37 +0200 | [diff] [blame] | 1330 | bio_io_error(bio); |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1331 | return; |
| 1332 | } |
| 1333 | |
Kent Overstreet | 54efd50 | 2015-04-23 22:37:18 -0700 | [diff] [blame] | 1334 | blk_queue_split(q, &bio, q->bio_split); |
| 1335 | |
Jeff Moyer | e6c4438 | 2015-05-08 10:51:30 -0700 | [diff] [blame] | 1336 | if (!is_flush_fua && !blk_queue_nomerges(q) && |
Shaohua Li | 5b3f341 | 2015-05-08 10:51:33 -0700 | [diff] [blame] | 1337 | blk_attempt_plug_merge(q, bio, &request_count, NULL)) |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1338 | return; |
| 1339 | |
| 1340 | rq = blk_mq_map_request(q, bio, &data); |
Jens Axboe | ff87bce | 2014-06-03 11:59:49 -0600 | [diff] [blame] | 1341 | if (unlikely(!rq)) |
| 1342 | return; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1343 | |
| 1344 | if (unlikely(is_flush_fua)) { |
| 1345 | blk_mq_bio_to_request(rq, bio); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1346 | blk_insert_flush(rq); |
| 1347 | goto run_queue; |
| 1348 | } |
| 1349 | |
| 1350 | /* |
| 1351 | * A task plug currently exists. Since this is completely lockless, |
| 1352 | * utilize that to temporarily store requests until the task is |
| 1353 | * either done or scheduled away. |
| 1354 | */ |
Jeff Moyer | e6c4438 | 2015-05-08 10:51:30 -0700 | [diff] [blame] | 1355 | plug = current->plug; |
| 1356 | if (plug) { |
| 1357 | blk_mq_bio_to_request(rq, bio); |
| 1358 | if (list_empty(&plug->mq_list)) |
| 1359 | trace_block_plug(q); |
| 1360 | else if (request_count >= BLK_MAX_REQUEST_COUNT) { |
| 1361 | blk_flush_plug_list(plug, false); |
| 1362 | trace_block_plug(q); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1363 | } |
Jeff Moyer | e6c4438 | 2015-05-08 10:51:30 -0700 | [diff] [blame] | 1364 | list_add_tail(&rq->queuelist, &plug->mq_list); |
| 1365 | blk_mq_put_ctx(data.ctx); |
| 1366 | return; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1367 | } |
| 1368 | |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1369 | if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) { |
| 1370 | /* |
| 1371 | * For a SYNC request, send it to the hardware immediately. For |
| 1372 | * an ASYNC request, just ensure that we run it later on. The |
| 1373 | * latter allows for merging opportunities and more efficient |
| 1374 | * dispatching. |
| 1375 | */ |
| 1376 | run_queue: |
| 1377 | blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1378 | } |
| 1379 | |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1380 | blk_mq_put_ctx(data.ctx); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1381 | } |
| 1382 | |
| 1383 | /* |
| 1384 | * Default mapping to a software queue, since we use one per CPU. |
| 1385 | */ |
| 1386 | struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu) |
| 1387 | { |
| 1388 | return q->queue_hw_ctx[q->mq_map[cpu]]; |
| 1389 | } |
| 1390 | EXPORT_SYMBOL(blk_mq_map_queue); |
| 1391 | |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1392 | static void blk_mq_free_rq_map(struct blk_mq_tag_set *set, |
| 1393 | struct blk_mq_tags *tags, unsigned int hctx_idx) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1394 | { |
| 1395 | struct page *page; |
| 1396 | |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1397 | if (tags->rqs && set->ops->exit_request) { |
Christoph Hellwig | e9b267d | 2014-04-15 13:59:10 -0600 | [diff] [blame] | 1398 | int i; |
| 1399 | |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1400 | for (i = 0; i < tags->nr_tags; i++) { |
| 1401 | if (!tags->rqs[i]) |
Christoph Hellwig | e9b267d | 2014-04-15 13:59:10 -0600 | [diff] [blame] | 1402 | continue; |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1403 | set->ops->exit_request(set->driver_data, tags->rqs[i], |
| 1404 | hctx_idx, i); |
Jens Axboe | a516440 | 2014-09-10 09:02:03 -0600 | [diff] [blame] | 1405 | tags->rqs[i] = NULL; |
Christoph Hellwig | e9b267d | 2014-04-15 13:59:10 -0600 | [diff] [blame] | 1406 | } |
| 1407 | } |
| 1408 | |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1409 | while (!list_empty(&tags->page_list)) { |
| 1410 | page = list_first_entry(&tags->page_list, struct page, lru); |
Dave Hansen | 6753471 | 2014-01-08 20:17:46 -0700 | [diff] [blame] | 1411 | list_del_init(&page->lru); |
Catalin Marinas | f75782e | 2015-09-14 18:16:02 +0100 | [diff] [blame] | 1412 | /* |
| 1413 | * Remove kmemleak object previously allocated in |
| 1414 | * blk_mq_init_rq_map(). |
| 1415 | */ |
| 1416 | kmemleak_free(page_address(page)); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1417 | __free_pages(page, page->private); |
| 1418 | } |
| 1419 | |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1420 | kfree(tags->rqs); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1421 | |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1422 | blk_mq_free_tags(tags); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1423 | } |
| 1424 | |
| 1425 | static size_t order_to_size(unsigned int order) |
| 1426 | { |
Ming Lei | 4ca0850 | 2014-04-19 18:00:18 +0800 | [diff] [blame] | 1427 | return (size_t)PAGE_SIZE << order; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1428 | } |
| 1429 | |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1430 | static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set, |
| 1431 | unsigned int hctx_idx) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1432 | { |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1433 | struct blk_mq_tags *tags; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1434 | unsigned int i, j, entries_per_page, max_order = 4; |
| 1435 | size_t rq_size, left; |
| 1436 | |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1437 | tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags, |
Shaohua Li | 24391c0 | 2015-01-23 14:18:00 -0700 | [diff] [blame] | 1438 | set->numa_node, |
| 1439 | BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags)); |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1440 | if (!tags) |
| 1441 | return NULL; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1442 | |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1443 | INIT_LIST_HEAD(&tags->page_list); |
| 1444 | |
Jens Axboe | a516440 | 2014-09-10 09:02:03 -0600 | [diff] [blame] | 1445 | tags->rqs = kzalloc_node(set->queue_depth * sizeof(struct request *), |
| 1446 | GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY, |
| 1447 | set->numa_node); |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1448 | if (!tags->rqs) { |
| 1449 | blk_mq_free_tags(tags); |
| 1450 | return NULL; |
| 1451 | } |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1452 | |
| 1453 | /* |
| 1454 | * rq_size is the size of the request plus driver payload, rounded |
| 1455 | * to the cacheline size |
| 1456 | */ |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1457 | rq_size = round_up(sizeof(struct request) + set->cmd_size, |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1458 | cache_line_size()); |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1459 | left = rq_size * set->queue_depth; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1460 | |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1461 | for (i = 0; i < set->queue_depth; ) { |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1462 | int this_order = max_order; |
| 1463 | struct page *page; |
| 1464 | int to_do; |
| 1465 | void *p; |
| 1466 | |
| 1467 | while (left < order_to_size(this_order - 1) && this_order) |
| 1468 | this_order--; |
| 1469 | |
| 1470 | do { |
Jens Axboe | a516440 | 2014-09-10 09:02:03 -0600 | [diff] [blame] | 1471 | page = alloc_pages_node(set->numa_node, |
Linus Torvalds | ac21117 | 2015-04-09 14:12:22 -0700 | [diff] [blame] | 1472 | GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO, |
Jens Axboe | a516440 | 2014-09-10 09:02:03 -0600 | [diff] [blame] | 1473 | this_order); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1474 | if (page) |
| 1475 | break; |
| 1476 | if (!this_order--) |
| 1477 | break; |
| 1478 | if (order_to_size(this_order) < rq_size) |
| 1479 | break; |
| 1480 | } while (1); |
| 1481 | |
| 1482 | if (!page) |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1483 | goto fail; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1484 | |
| 1485 | page->private = this_order; |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1486 | list_add_tail(&page->lru, &tags->page_list); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1487 | |
| 1488 | p = page_address(page); |
Catalin Marinas | f75782e | 2015-09-14 18:16:02 +0100 | [diff] [blame] | 1489 | /* |
| 1490 | * Allow kmemleak to scan these pages as they contain pointers |
| 1491 | * to additional allocations like via ops->init_request(). |
| 1492 | */ |
| 1493 | kmemleak_alloc(p, order_to_size(this_order), 1, GFP_KERNEL); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1494 | entries_per_page = order_to_size(this_order) / rq_size; |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1495 | to_do = min(entries_per_page, set->queue_depth - i); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1496 | left -= to_do * rq_size; |
| 1497 | for (j = 0; j < to_do; j++) { |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1498 | tags->rqs[i] = p; |
| 1499 | if (set->ops->init_request) { |
| 1500 | if (set->ops->init_request(set->driver_data, |
| 1501 | tags->rqs[i], hctx_idx, i, |
Jens Axboe | a516440 | 2014-09-10 09:02:03 -0600 | [diff] [blame] | 1502 | set->numa_node)) { |
| 1503 | tags->rqs[i] = NULL; |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1504 | goto fail; |
Jens Axboe | a516440 | 2014-09-10 09:02:03 -0600 | [diff] [blame] | 1505 | } |
Christoph Hellwig | e9b267d | 2014-04-15 13:59:10 -0600 | [diff] [blame] | 1506 | } |
| 1507 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1508 | p += rq_size; |
| 1509 | i++; |
| 1510 | } |
| 1511 | } |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1512 | return tags; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1513 | |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1514 | fail: |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1515 | blk_mq_free_rq_map(set, tags, hctx_idx); |
| 1516 | return NULL; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1517 | } |
| 1518 | |
Jens Axboe | 1429d7c | 2014-05-19 09:23:55 -0600 | [diff] [blame] | 1519 | static void blk_mq_free_bitmap(struct blk_mq_ctxmap *bitmap) |
| 1520 | { |
| 1521 | kfree(bitmap->map); |
| 1522 | } |
| 1523 | |
| 1524 | static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap *bitmap, int node) |
| 1525 | { |
| 1526 | unsigned int bpw = 8, total, num_maps, i; |
| 1527 | |
| 1528 | bitmap->bits_per_word = bpw; |
| 1529 | |
| 1530 | num_maps = ALIGN(nr_cpu_ids, bpw) / bpw; |
| 1531 | bitmap->map = kzalloc_node(num_maps * sizeof(struct blk_align_bitmap), |
| 1532 | GFP_KERNEL, node); |
| 1533 | if (!bitmap->map) |
| 1534 | return -ENOMEM; |
| 1535 | |
Jens Axboe | 1429d7c | 2014-05-19 09:23:55 -0600 | [diff] [blame] | 1536 | total = nr_cpu_ids; |
| 1537 | for (i = 0; i < num_maps; i++) { |
| 1538 | bitmap->map[i].depth = min(total, bitmap->bits_per_word); |
| 1539 | total -= bitmap->map[i].depth; |
| 1540 | } |
| 1541 | |
| 1542 | return 0; |
| 1543 | } |
| 1544 | |
Jens Axboe | 484b406 | 2014-05-21 14:01:15 -0600 | [diff] [blame] | 1545 | static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx *hctx, int cpu) |
| 1546 | { |
| 1547 | struct request_queue *q = hctx->queue; |
| 1548 | struct blk_mq_ctx *ctx; |
| 1549 | LIST_HEAD(tmp); |
| 1550 | |
| 1551 | /* |
| 1552 | * Move ctx entries to new CPU, if this one is going away. |
| 1553 | */ |
| 1554 | ctx = __blk_mq_get_ctx(q, cpu); |
| 1555 | |
| 1556 | spin_lock(&ctx->lock); |
| 1557 | if (!list_empty(&ctx->rq_list)) { |
| 1558 | list_splice_init(&ctx->rq_list, &tmp); |
| 1559 | blk_mq_hctx_clear_pending(hctx, ctx); |
| 1560 | } |
| 1561 | spin_unlock(&ctx->lock); |
| 1562 | |
| 1563 | if (list_empty(&tmp)) |
| 1564 | return NOTIFY_OK; |
| 1565 | |
| 1566 | ctx = blk_mq_get_ctx(q); |
| 1567 | spin_lock(&ctx->lock); |
| 1568 | |
| 1569 | while (!list_empty(&tmp)) { |
| 1570 | struct request *rq; |
| 1571 | |
| 1572 | rq = list_first_entry(&tmp, struct request, queuelist); |
| 1573 | rq->mq_ctx = ctx; |
| 1574 | list_move_tail(&rq->queuelist, &ctx->rq_list); |
| 1575 | } |
| 1576 | |
| 1577 | hctx = q->mq_ops->map_queue(q, ctx->cpu); |
| 1578 | blk_mq_hctx_mark_pending(hctx, ctx); |
| 1579 | |
| 1580 | spin_unlock(&ctx->lock); |
| 1581 | |
| 1582 | blk_mq_run_hw_queue(hctx, true); |
| 1583 | blk_mq_put_ctx(ctx); |
| 1584 | return NOTIFY_OK; |
| 1585 | } |
| 1586 | |
Jens Axboe | 484b406 | 2014-05-21 14:01:15 -0600 | [diff] [blame] | 1587 | static int blk_mq_hctx_notify(void *data, unsigned long action, |
| 1588 | unsigned int cpu) |
| 1589 | { |
| 1590 | struct blk_mq_hw_ctx *hctx = data; |
| 1591 | |
| 1592 | if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) |
| 1593 | return blk_mq_hctx_cpu_offline(hctx, cpu); |
Ming Lei | 2a34c08 | 2015-04-21 10:00:20 +0800 | [diff] [blame] | 1594 | |
| 1595 | /* |
| 1596 | * In case of CPU online, tags may be reallocated |
| 1597 | * in blk_mq_map_swqueue() after mapping is updated. |
| 1598 | */ |
Jens Axboe | 484b406 | 2014-05-21 14:01:15 -0600 | [diff] [blame] | 1599 | |
| 1600 | return NOTIFY_OK; |
| 1601 | } |
| 1602 | |
Ming Lei | c3b4afc | 2015-06-04 22:25:04 +0800 | [diff] [blame] | 1603 | /* hctx->ctxs will be freed in queue's release handler */ |
Ming Lei | 08e98fc | 2014-09-25 23:23:38 +0800 | [diff] [blame] | 1604 | static void blk_mq_exit_hctx(struct request_queue *q, |
| 1605 | struct blk_mq_tag_set *set, |
| 1606 | struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) |
| 1607 | { |
Ming Lei | f70ced0 | 2014-09-25 23:23:47 +0800 | [diff] [blame] | 1608 | unsigned flush_start_tag = set->queue_depth; |
| 1609 | |
Ming Lei | 08e98fc | 2014-09-25 23:23:38 +0800 | [diff] [blame] | 1610 | blk_mq_tag_idle(hctx); |
| 1611 | |
Ming Lei | f70ced0 | 2014-09-25 23:23:47 +0800 | [diff] [blame] | 1612 | if (set->ops->exit_request) |
| 1613 | set->ops->exit_request(set->driver_data, |
| 1614 | hctx->fq->flush_rq, hctx_idx, |
| 1615 | flush_start_tag + hctx_idx); |
| 1616 | |
Ming Lei | 08e98fc | 2014-09-25 23:23:38 +0800 | [diff] [blame] | 1617 | if (set->ops->exit_hctx) |
| 1618 | set->ops->exit_hctx(hctx, hctx_idx); |
| 1619 | |
| 1620 | blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier); |
Ming Lei | f70ced0 | 2014-09-25 23:23:47 +0800 | [diff] [blame] | 1621 | blk_free_flush_queue(hctx->fq); |
Ming Lei | 08e98fc | 2014-09-25 23:23:38 +0800 | [diff] [blame] | 1622 | blk_mq_free_bitmap(&hctx->ctx_map); |
| 1623 | } |
| 1624 | |
Ming Lei | 624dbe4 | 2014-05-27 23:35:13 +0800 | [diff] [blame] | 1625 | static void blk_mq_exit_hw_queues(struct request_queue *q, |
| 1626 | struct blk_mq_tag_set *set, int nr_queue) |
| 1627 | { |
| 1628 | struct blk_mq_hw_ctx *hctx; |
| 1629 | unsigned int i; |
| 1630 | |
| 1631 | queue_for_each_hw_ctx(q, hctx, i) { |
| 1632 | if (i == nr_queue) |
| 1633 | break; |
Ming Lei | 08e98fc | 2014-09-25 23:23:38 +0800 | [diff] [blame] | 1634 | blk_mq_exit_hctx(q, set, hctx, i); |
Ming Lei | 624dbe4 | 2014-05-27 23:35:13 +0800 | [diff] [blame] | 1635 | } |
Ming Lei | 624dbe4 | 2014-05-27 23:35:13 +0800 | [diff] [blame] | 1636 | } |
| 1637 | |
| 1638 | static void blk_mq_free_hw_queues(struct request_queue *q, |
| 1639 | struct blk_mq_tag_set *set) |
| 1640 | { |
| 1641 | struct blk_mq_hw_ctx *hctx; |
| 1642 | unsigned int i; |
| 1643 | |
Ming Lei | e09aae7 | 2015-01-29 20:17:27 +0800 | [diff] [blame] | 1644 | queue_for_each_hw_ctx(q, hctx, i) |
Ming Lei | 624dbe4 | 2014-05-27 23:35:13 +0800 | [diff] [blame] | 1645 | free_cpumask_var(hctx->cpumask); |
Ming Lei | 624dbe4 | 2014-05-27 23:35:13 +0800 | [diff] [blame] | 1646 | } |
| 1647 | |
Ming Lei | 08e98fc | 2014-09-25 23:23:38 +0800 | [diff] [blame] | 1648 | static int blk_mq_init_hctx(struct request_queue *q, |
| 1649 | struct blk_mq_tag_set *set, |
| 1650 | struct blk_mq_hw_ctx *hctx, unsigned hctx_idx) |
| 1651 | { |
| 1652 | int node; |
Ming Lei | f70ced0 | 2014-09-25 23:23:47 +0800 | [diff] [blame] | 1653 | unsigned flush_start_tag = set->queue_depth; |
Ming Lei | 08e98fc | 2014-09-25 23:23:38 +0800 | [diff] [blame] | 1654 | |
| 1655 | node = hctx->numa_node; |
| 1656 | if (node == NUMA_NO_NODE) |
| 1657 | node = hctx->numa_node = set->numa_node; |
| 1658 | |
| 1659 | INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn); |
| 1660 | INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn); |
| 1661 | spin_lock_init(&hctx->lock); |
| 1662 | INIT_LIST_HEAD(&hctx->dispatch); |
| 1663 | hctx->queue = q; |
| 1664 | hctx->queue_num = hctx_idx; |
| 1665 | hctx->flags = set->flags; |
Ming Lei | 08e98fc | 2014-09-25 23:23:38 +0800 | [diff] [blame] | 1666 | |
| 1667 | blk_mq_init_cpu_notifier(&hctx->cpu_notifier, |
| 1668 | blk_mq_hctx_notify, hctx); |
| 1669 | blk_mq_register_cpu_notifier(&hctx->cpu_notifier); |
| 1670 | |
| 1671 | hctx->tags = set->tags[hctx_idx]; |
| 1672 | |
| 1673 | /* |
| 1674 | * Allocate space for all possible cpus to avoid allocation at |
| 1675 | * runtime |
| 1676 | */ |
| 1677 | hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *), |
| 1678 | GFP_KERNEL, node); |
| 1679 | if (!hctx->ctxs) |
| 1680 | goto unregister_cpu_notifier; |
| 1681 | |
| 1682 | if (blk_mq_alloc_bitmap(&hctx->ctx_map, node)) |
| 1683 | goto free_ctxs; |
| 1684 | |
| 1685 | hctx->nr_ctx = 0; |
| 1686 | |
| 1687 | if (set->ops->init_hctx && |
| 1688 | set->ops->init_hctx(hctx, set->driver_data, hctx_idx)) |
| 1689 | goto free_bitmap; |
| 1690 | |
Ming Lei | f70ced0 | 2014-09-25 23:23:47 +0800 | [diff] [blame] | 1691 | hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size); |
| 1692 | if (!hctx->fq) |
| 1693 | goto exit_hctx; |
| 1694 | |
| 1695 | if (set->ops->init_request && |
| 1696 | set->ops->init_request(set->driver_data, |
| 1697 | hctx->fq->flush_rq, hctx_idx, |
| 1698 | flush_start_tag + hctx_idx, node)) |
| 1699 | goto free_fq; |
| 1700 | |
Ming Lei | 08e98fc | 2014-09-25 23:23:38 +0800 | [diff] [blame] | 1701 | return 0; |
| 1702 | |
Ming Lei | f70ced0 | 2014-09-25 23:23:47 +0800 | [diff] [blame] | 1703 | free_fq: |
| 1704 | kfree(hctx->fq); |
| 1705 | exit_hctx: |
| 1706 | if (set->ops->exit_hctx) |
| 1707 | set->ops->exit_hctx(hctx, hctx_idx); |
Ming Lei | 08e98fc | 2014-09-25 23:23:38 +0800 | [diff] [blame] | 1708 | free_bitmap: |
| 1709 | blk_mq_free_bitmap(&hctx->ctx_map); |
| 1710 | free_ctxs: |
| 1711 | kfree(hctx->ctxs); |
| 1712 | unregister_cpu_notifier: |
| 1713 | blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier); |
| 1714 | |
| 1715 | return -1; |
| 1716 | } |
| 1717 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1718 | static int blk_mq_init_hw_queues(struct request_queue *q, |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1719 | struct blk_mq_tag_set *set) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1720 | { |
| 1721 | struct blk_mq_hw_ctx *hctx; |
Ming Lei | 624dbe4 | 2014-05-27 23:35:13 +0800 | [diff] [blame] | 1722 | unsigned int i; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1723 | |
| 1724 | /* |
| 1725 | * Initialize hardware queues |
| 1726 | */ |
| 1727 | queue_for_each_hw_ctx(q, hctx, i) { |
Ming Lei | 08e98fc | 2014-09-25 23:23:38 +0800 | [diff] [blame] | 1728 | if (blk_mq_init_hctx(q, set, hctx, i)) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1729 | break; |
| 1730 | } |
| 1731 | |
| 1732 | if (i == q->nr_hw_queues) |
| 1733 | return 0; |
| 1734 | |
| 1735 | /* |
| 1736 | * Init failed |
| 1737 | */ |
Ming Lei | 624dbe4 | 2014-05-27 23:35:13 +0800 | [diff] [blame] | 1738 | blk_mq_exit_hw_queues(q, set, i); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1739 | |
| 1740 | return 1; |
| 1741 | } |
| 1742 | |
| 1743 | static void blk_mq_init_cpu_queues(struct request_queue *q, |
| 1744 | unsigned int nr_hw_queues) |
| 1745 | { |
| 1746 | unsigned int i; |
| 1747 | |
| 1748 | for_each_possible_cpu(i) { |
| 1749 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); |
| 1750 | struct blk_mq_hw_ctx *hctx; |
| 1751 | |
| 1752 | memset(__ctx, 0, sizeof(*__ctx)); |
| 1753 | __ctx->cpu = i; |
| 1754 | spin_lock_init(&__ctx->lock); |
| 1755 | INIT_LIST_HEAD(&__ctx->rq_list); |
| 1756 | __ctx->queue = q; |
| 1757 | |
| 1758 | /* If the cpu isn't online, the cpu is mapped to first hctx */ |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1759 | if (!cpu_online(i)) |
| 1760 | continue; |
| 1761 | |
Jens Axboe | e4043dc | 2014-04-09 10:18:23 -0600 | [diff] [blame] | 1762 | hctx = q->mq_ops->map_queue(q, i); |
Jens Axboe | e4043dc | 2014-04-09 10:18:23 -0600 | [diff] [blame] | 1763 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1764 | /* |
| 1765 | * Set local node, IFF we have more than one hw queue. If |
| 1766 | * not, we remain on the home node of the device |
| 1767 | */ |
| 1768 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) |
| 1769 | hctx->numa_node = cpu_to_node(i); |
| 1770 | } |
| 1771 | } |
| 1772 | |
Akinobu Mita | 5778322 | 2015-09-27 02:09:23 +0900 | [diff] [blame] | 1773 | static void blk_mq_map_swqueue(struct request_queue *q, |
| 1774 | const struct cpumask *online_mask) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1775 | { |
| 1776 | unsigned int i; |
| 1777 | struct blk_mq_hw_ctx *hctx; |
| 1778 | struct blk_mq_ctx *ctx; |
Ming Lei | 2a34c08 | 2015-04-21 10:00:20 +0800 | [diff] [blame] | 1779 | struct blk_mq_tag_set *set = q->tag_set; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1780 | |
Akinobu Mita | 60de074 | 2015-09-27 02:09:25 +0900 | [diff] [blame] | 1781 | /* |
| 1782 | * Avoid others reading imcomplete hctx->cpumask through sysfs |
| 1783 | */ |
| 1784 | mutex_lock(&q->sysfs_lock); |
| 1785 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1786 | queue_for_each_hw_ctx(q, hctx, i) { |
Jens Axboe | e4043dc | 2014-04-09 10:18:23 -0600 | [diff] [blame] | 1787 | cpumask_clear(hctx->cpumask); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1788 | hctx->nr_ctx = 0; |
| 1789 | } |
| 1790 | |
| 1791 | /* |
| 1792 | * Map software to hardware queues |
| 1793 | */ |
| 1794 | queue_for_each_ctx(q, ctx, i) { |
| 1795 | /* If the cpu isn't online, the cpu is mapped to first hctx */ |
Akinobu Mita | 5778322 | 2015-09-27 02:09:23 +0900 | [diff] [blame] | 1796 | if (!cpumask_test_cpu(i, online_mask)) |
Jens Axboe | e4043dc | 2014-04-09 10:18:23 -0600 | [diff] [blame] | 1797 | continue; |
| 1798 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1799 | hctx = q->mq_ops->map_queue(q, i); |
Jens Axboe | e4043dc | 2014-04-09 10:18:23 -0600 | [diff] [blame] | 1800 | cpumask_set_cpu(i, hctx->cpumask); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1801 | ctx->index_hw = hctx->nr_ctx; |
| 1802 | hctx->ctxs[hctx->nr_ctx++] = ctx; |
| 1803 | } |
Jens Axboe | 506e931 | 2014-05-07 10:26:44 -0600 | [diff] [blame] | 1804 | |
Akinobu Mita | 60de074 | 2015-09-27 02:09:25 +0900 | [diff] [blame] | 1805 | mutex_unlock(&q->sysfs_lock); |
| 1806 | |
Jens Axboe | 506e931 | 2014-05-07 10:26:44 -0600 | [diff] [blame] | 1807 | queue_for_each_hw_ctx(q, hctx, i) { |
Chong Yuan | 889fa31 | 2015-04-15 11:39:29 -0600 | [diff] [blame] | 1808 | struct blk_mq_ctxmap *map = &hctx->ctx_map; |
| 1809 | |
Jens Axboe | 484b406 | 2014-05-21 14:01:15 -0600 | [diff] [blame] | 1810 | /* |
Jens Axboe | a68aafa | 2014-08-15 13:19:15 -0600 | [diff] [blame] | 1811 | * If no software queues are mapped to this hardware queue, |
| 1812 | * disable it and free the request entries. |
Jens Axboe | 484b406 | 2014-05-21 14:01:15 -0600 | [diff] [blame] | 1813 | */ |
| 1814 | if (!hctx->nr_ctx) { |
Jens Axboe | 484b406 | 2014-05-21 14:01:15 -0600 | [diff] [blame] | 1815 | if (set->tags[i]) { |
| 1816 | blk_mq_free_rq_map(set, set->tags[i], i); |
| 1817 | set->tags[i] = NULL; |
Jens Axboe | 484b406 | 2014-05-21 14:01:15 -0600 | [diff] [blame] | 1818 | } |
Ming Lei | 2a34c08 | 2015-04-21 10:00:20 +0800 | [diff] [blame] | 1819 | hctx->tags = NULL; |
Jens Axboe | 484b406 | 2014-05-21 14:01:15 -0600 | [diff] [blame] | 1820 | continue; |
| 1821 | } |
| 1822 | |
Ming Lei | 2a34c08 | 2015-04-21 10:00:20 +0800 | [diff] [blame] | 1823 | /* unmapped hw queue can be remapped after CPU topo changed */ |
| 1824 | if (!set->tags[i]) |
| 1825 | set->tags[i] = blk_mq_init_rq_map(set, i); |
| 1826 | hctx->tags = set->tags[i]; |
| 1827 | WARN_ON(!hctx->tags); |
| 1828 | |
Jens Axboe | 484b406 | 2014-05-21 14:01:15 -0600 | [diff] [blame] | 1829 | /* |
Chong Yuan | 889fa31 | 2015-04-15 11:39:29 -0600 | [diff] [blame] | 1830 | * Set the map size to the number of mapped software queues. |
| 1831 | * This is more accurate and more efficient than looping |
| 1832 | * over all possibly mapped software queues. |
| 1833 | */ |
Jens Axboe | 569fd0c | 2015-04-17 08:28:50 -0600 | [diff] [blame] | 1834 | map->size = DIV_ROUND_UP(hctx->nr_ctx, map->bits_per_word); |
Chong Yuan | 889fa31 | 2015-04-15 11:39:29 -0600 | [diff] [blame] | 1835 | |
| 1836 | /* |
Jens Axboe | 484b406 | 2014-05-21 14:01:15 -0600 | [diff] [blame] | 1837 | * Initialize batch roundrobin counts |
| 1838 | */ |
Jens Axboe | 506e931 | 2014-05-07 10:26:44 -0600 | [diff] [blame] | 1839 | hctx->next_cpu = cpumask_first(hctx->cpumask); |
| 1840 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; |
| 1841 | } |
Akinobu Mita | 1356aae | 2015-09-27 02:09:19 +0900 | [diff] [blame] | 1842 | |
| 1843 | queue_for_each_ctx(q, ctx, i) { |
Akinobu Mita | 5778322 | 2015-09-27 02:09:23 +0900 | [diff] [blame] | 1844 | if (!cpumask_test_cpu(i, online_mask)) |
Akinobu Mita | 1356aae | 2015-09-27 02:09:19 +0900 | [diff] [blame] | 1845 | continue; |
| 1846 | |
| 1847 | hctx = q->mq_ops->map_queue(q, i); |
| 1848 | cpumask_set_cpu(i, hctx->tags->cpumask); |
| 1849 | } |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1850 | } |
| 1851 | |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 1852 | static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set) |
| 1853 | { |
| 1854 | struct blk_mq_hw_ctx *hctx; |
| 1855 | struct request_queue *q; |
| 1856 | bool shared; |
| 1857 | int i; |
| 1858 | |
| 1859 | if (set->tag_list.next == set->tag_list.prev) |
| 1860 | shared = false; |
| 1861 | else |
| 1862 | shared = true; |
| 1863 | |
| 1864 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
| 1865 | blk_mq_freeze_queue(q); |
| 1866 | |
| 1867 | queue_for_each_hw_ctx(q, hctx, i) { |
| 1868 | if (shared) |
| 1869 | hctx->flags |= BLK_MQ_F_TAG_SHARED; |
| 1870 | else |
| 1871 | hctx->flags &= ~BLK_MQ_F_TAG_SHARED; |
| 1872 | } |
| 1873 | blk_mq_unfreeze_queue(q); |
| 1874 | } |
| 1875 | } |
| 1876 | |
| 1877 | static void blk_mq_del_queue_tag_set(struct request_queue *q) |
| 1878 | { |
| 1879 | struct blk_mq_tag_set *set = q->tag_set; |
| 1880 | |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 1881 | mutex_lock(&set->tag_list_lock); |
| 1882 | list_del_init(&q->tag_set_list); |
| 1883 | blk_mq_update_tag_set_depth(set); |
| 1884 | mutex_unlock(&set->tag_list_lock); |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 1885 | } |
| 1886 | |
| 1887 | static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, |
| 1888 | struct request_queue *q) |
| 1889 | { |
| 1890 | q->tag_set = set; |
| 1891 | |
| 1892 | mutex_lock(&set->tag_list_lock); |
| 1893 | list_add_tail(&q->tag_set_list, &set->tag_list); |
| 1894 | blk_mq_update_tag_set_depth(set); |
| 1895 | mutex_unlock(&set->tag_list_lock); |
| 1896 | } |
| 1897 | |
Ming Lei | e09aae7 | 2015-01-29 20:17:27 +0800 | [diff] [blame] | 1898 | /* |
| 1899 | * It is the actual release handler for mq, but we do it from |
| 1900 | * request queue's release handler for avoiding use-after-free |
| 1901 | * and headache because q->mq_kobj shouldn't have been introduced, |
| 1902 | * but we can't group ctx/kctx kobj without it. |
| 1903 | */ |
| 1904 | void blk_mq_release(struct request_queue *q) |
| 1905 | { |
| 1906 | struct blk_mq_hw_ctx *hctx; |
| 1907 | unsigned int i; |
| 1908 | |
| 1909 | /* hctx kobj stays in hctx */ |
Ming Lei | c3b4afc | 2015-06-04 22:25:04 +0800 | [diff] [blame] | 1910 | queue_for_each_hw_ctx(q, hctx, i) { |
| 1911 | if (!hctx) |
| 1912 | continue; |
| 1913 | kfree(hctx->ctxs); |
Ming Lei | e09aae7 | 2015-01-29 20:17:27 +0800 | [diff] [blame] | 1914 | kfree(hctx); |
Ming Lei | c3b4afc | 2015-06-04 22:25:04 +0800 | [diff] [blame] | 1915 | } |
Ming Lei | e09aae7 | 2015-01-29 20:17:27 +0800 | [diff] [blame] | 1916 | |
Akinobu Mita | a723bab | 2015-09-27 02:09:21 +0900 | [diff] [blame] | 1917 | kfree(q->mq_map); |
| 1918 | q->mq_map = NULL; |
| 1919 | |
Ming Lei | e09aae7 | 2015-01-29 20:17:27 +0800 | [diff] [blame] | 1920 | kfree(q->queue_hw_ctx); |
| 1921 | |
| 1922 | /* ctx kobj stays in queue_ctx */ |
| 1923 | free_percpu(q->queue_ctx); |
| 1924 | } |
| 1925 | |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1926 | struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1927 | { |
Mike Snitzer | b62c21b | 2015-03-12 23:56:02 -0400 | [diff] [blame] | 1928 | struct request_queue *uninit_q, *q; |
| 1929 | |
| 1930 | uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node); |
| 1931 | if (!uninit_q) |
| 1932 | return ERR_PTR(-ENOMEM); |
| 1933 | |
| 1934 | q = blk_mq_init_allocated_queue(set, uninit_q); |
| 1935 | if (IS_ERR(q)) |
| 1936 | blk_cleanup_queue(uninit_q); |
| 1937 | |
| 1938 | return q; |
| 1939 | } |
| 1940 | EXPORT_SYMBOL(blk_mq_init_queue); |
| 1941 | |
| 1942 | struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set, |
| 1943 | struct request_queue *q) |
| 1944 | { |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1945 | struct blk_mq_hw_ctx **hctxs; |
Ming Lei | e6cdb09 | 2014-06-03 11:24:06 +0800 | [diff] [blame] | 1946 | struct blk_mq_ctx __percpu *ctx; |
Jens Axboe | f14bbe7 | 2014-05-27 12:06:53 -0600 | [diff] [blame] | 1947 | unsigned int *map; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1948 | int i; |
| 1949 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1950 | ctx = alloc_percpu(struct blk_mq_ctx); |
| 1951 | if (!ctx) |
| 1952 | return ERR_PTR(-ENOMEM); |
| 1953 | |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1954 | hctxs = kmalloc_node(set->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL, |
| 1955 | set->numa_node); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1956 | |
| 1957 | if (!hctxs) |
| 1958 | goto err_percpu; |
| 1959 | |
Jens Axboe | f14bbe7 | 2014-05-27 12:06:53 -0600 | [diff] [blame] | 1960 | map = blk_mq_make_queue_map(set); |
| 1961 | if (!map) |
| 1962 | goto err_map; |
| 1963 | |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1964 | for (i = 0; i < set->nr_hw_queues; i++) { |
Jens Axboe | f14bbe7 | 2014-05-27 12:06:53 -0600 | [diff] [blame] | 1965 | int node = blk_mq_hw_queue_to_node(map, i); |
| 1966 | |
Christoph Hellwig | cdef54d | 2014-05-28 18:11:06 +0200 | [diff] [blame] | 1967 | hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx), |
| 1968 | GFP_KERNEL, node); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1969 | if (!hctxs[i]) |
| 1970 | goto err_hctxs; |
| 1971 | |
Jens Axboe | a86073e | 2014-10-13 15:41:54 -0600 | [diff] [blame] | 1972 | if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL, |
| 1973 | node)) |
Jens Axboe | e4043dc | 2014-04-09 10:18:23 -0600 | [diff] [blame] | 1974 | goto err_hctxs; |
| 1975 | |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 1976 | atomic_set(&hctxs[i]->nr_active, 0); |
Jens Axboe | f14bbe7 | 2014-05-27 12:06:53 -0600 | [diff] [blame] | 1977 | hctxs[i]->numa_node = node; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1978 | hctxs[i]->queue_num = i; |
| 1979 | } |
| 1980 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1981 | setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q); |
Ming Lei | e56f698 | 2015-07-16 19:53:22 +0800 | [diff] [blame] | 1982 | blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1983 | |
| 1984 | q->nr_queues = nr_cpu_ids; |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1985 | q->nr_hw_queues = set->nr_hw_queues; |
Jens Axboe | f14bbe7 | 2014-05-27 12:06:53 -0600 | [diff] [blame] | 1986 | q->mq_map = map; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1987 | |
| 1988 | q->queue_ctx = ctx; |
| 1989 | q->queue_hw_ctx = hctxs; |
| 1990 | |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1991 | q->mq_ops = set->ops; |
Jens Axboe | 94eddfb | 2013-11-19 09:25:07 -0700 | [diff] [blame] | 1992 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1993 | |
Jens Axboe | 05f1dd5 | 2014-05-29 09:53:32 -0600 | [diff] [blame] | 1994 | if (!(set->flags & BLK_MQ_F_SG_MERGE)) |
| 1995 | q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE; |
| 1996 | |
Christoph Hellwig | 1be036e | 2014-02-07 10:22:39 -0800 | [diff] [blame] | 1997 | q->sg_reserved_size = INT_MAX; |
| 1998 | |
Christoph Hellwig | 6fca6a6 | 2014-05-28 08:08:02 -0600 | [diff] [blame] | 1999 | INIT_WORK(&q->requeue_work, blk_mq_requeue_work); |
| 2000 | INIT_LIST_HEAD(&q->requeue_list); |
| 2001 | spin_lock_init(&q->requeue_lock); |
| 2002 | |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 2003 | if (q->nr_hw_queues > 1) |
| 2004 | blk_queue_make_request(q, blk_mq_make_request); |
| 2005 | else |
| 2006 | blk_queue_make_request(q, blk_sq_make_request); |
| 2007 | |
Jens Axboe | eba7176 | 2014-05-20 15:17:27 -0600 | [diff] [blame] | 2008 | /* |
| 2009 | * Do this after blk_queue_make_request() overrides it... |
| 2010 | */ |
| 2011 | q->nr_requests = set->queue_depth; |
| 2012 | |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 2013 | if (set->ops->complete) |
| 2014 | blk_queue_softirq_done(q, set->ops->complete); |
Christoph Hellwig | 30a91cb | 2014-02-10 03:24:38 -0800 | [diff] [blame] | 2015 | |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 2016 | blk_mq_init_cpu_queues(q, set->nr_hw_queues); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2017 | |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 2018 | if (blk_mq_init_hw_queues(q, set)) |
Mike Snitzer | b62c21b | 2015-03-12 23:56:02 -0400 | [diff] [blame] | 2019 | goto err_hctxs; |
Christoph Hellwig | 1874198 | 2014-02-10 09:29:00 -0700 | [diff] [blame] | 2020 | |
Akinobu Mita | 5778322 | 2015-09-27 02:09:23 +0900 | [diff] [blame] | 2021 | get_online_cpus(); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2022 | mutex_lock(&all_q_mutex); |
Akinobu Mita | 4593fdb | 2015-09-27 02:09:20 +0900 | [diff] [blame] | 2023 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2024 | list_add_tail(&q->all_q_node, &all_q_list); |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 2025 | blk_mq_add_queue_tag_set(set, q); |
Akinobu Mita | 5778322 | 2015-09-27 02:09:23 +0900 | [diff] [blame] | 2026 | blk_mq_map_swqueue(q, cpu_online_mask); |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 2027 | |
Akinobu Mita | 4593fdb | 2015-09-27 02:09:20 +0900 | [diff] [blame] | 2028 | mutex_unlock(&all_q_mutex); |
Akinobu Mita | 5778322 | 2015-09-27 02:09:23 +0900 | [diff] [blame] | 2029 | put_online_cpus(); |
Jens Axboe | 484b406 | 2014-05-21 14:01:15 -0600 | [diff] [blame] | 2030 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2031 | return q; |
Christoph Hellwig | 1874198 | 2014-02-10 09:29:00 -0700 | [diff] [blame] | 2032 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2033 | err_hctxs: |
Jens Axboe | f14bbe7 | 2014-05-27 12:06:53 -0600 | [diff] [blame] | 2034 | kfree(map); |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 2035 | for (i = 0; i < set->nr_hw_queues; i++) { |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2036 | if (!hctxs[i]) |
| 2037 | break; |
Jens Axboe | e4043dc | 2014-04-09 10:18:23 -0600 | [diff] [blame] | 2038 | free_cpumask_var(hctxs[i]->cpumask); |
Christoph Hellwig | cdef54d | 2014-05-28 18:11:06 +0200 | [diff] [blame] | 2039 | kfree(hctxs[i]); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2040 | } |
Jens Axboe | f14bbe7 | 2014-05-27 12:06:53 -0600 | [diff] [blame] | 2041 | err_map: |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2042 | kfree(hctxs); |
| 2043 | err_percpu: |
| 2044 | free_percpu(ctx); |
| 2045 | return ERR_PTR(-ENOMEM); |
| 2046 | } |
Mike Snitzer | b62c21b | 2015-03-12 23:56:02 -0400 | [diff] [blame] | 2047 | EXPORT_SYMBOL(blk_mq_init_allocated_queue); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2048 | |
| 2049 | void blk_mq_free_queue(struct request_queue *q) |
| 2050 | { |
Ming Lei | 624dbe4 | 2014-05-27 23:35:13 +0800 | [diff] [blame] | 2051 | struct blk_mq_tag_set *set = q->tag_set; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2052 | |
Akinobu Mita | 0e62636 | 2015-09-27 02:09:22 +0900 | [diff] [blame] | 2053 | mutex_lock(&all_q_mutex); |
| 2054 | list_del_init(&q->all_q_node); |
| 2055 | mutex_unlock(&all_q_mutex); |
| 2056 | |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 2057 | blk_mq_del_queue_tag_set(q); |
| 2058 | |
Ming Lei | 624dbe4 | 2014-05-27 23:35:13 +0800 | [diff] [blame] | 2059 | blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); |
| 2060 | blk_mq_free_hw_queues(q, set); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2061 | } |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2062 | |
| 2063 | /* Basically redo blk_mq_init_queue with queue frozen */ |
Akinobu Mita | 5778322 | 2015-09-27 02:09:23 +0900 | [diff] [blame] | 2064 | static void blk_mq_queue_reinit(struct request_queue *q, |
| 2065 | const struct cpumask *online_mask) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2066 | { |
Christoph Hellwig | 4ecd4fe | 2015-05-07 09:38:13 +0200 | [diff] [blame] | 2067 | WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth)); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2068 | |
Jens Axboe | 67aec14 | 2014-05-30 08:25:36 -0600 | [diff] [blame] | 2069 | blk_mq_sysfs_unregister(q); |
| 2070 | |
Akinobu Mita | 5778322 | 2015-09-27 02:09:23 +0900 | [diff] [blame] | 2071 | blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues, online_mask); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2072 | |
| 2073 | /* |
| 2074 | * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe |
| 2075 | * we should change hctx numa_node according to new topology (this |
| 2076 | * involves free and re-allocate memory, worthy doing?) |
| 2077 | */ |
| 2078 | |
Akinobu Mita | 5778322 | 2015-09-27 02:09:23 +0900 | [diff] [blame] | 2079 | blk_mq_map_swqueue(q, online_mask); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2080 | |
Jens Axboe | 67aec14 | 2014-05-30 08:25:36 -0600 | [diff] [blame] | 2081 | blk_mq_sysfs_register(q); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2082 | } |
| 2083 | |
Paul Gortmaker | f618ef7 | 2013-11-14 08:26:02 -0700 | [diff] [blame] | 2084 | static int blk_mq_queue_reinit_notify(struct notifier_block *nb, |
| 2085 | unsigned long action, void *hcpu) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2086 | { |
| 2087 | struct request_queue *q; |
Akinobu Mita | 5778322 | 2015-09-27 02:09:23 +0900 | [diff] [blame] | 2088 | int cpu = (unsigned long)hcpu; |
| 2089 | /* |
| 2090 | * New online cpumask which is going to be set in this hotplug event. |
| 2091 | * Declare this cpumasks as global as cpu-hotplug operation is invoked |
| 2092 | * one-by-one and dynamically allocating this could result in a failure. |
| 2093 | */ |
| 2094 | static struct cpumask online_new; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2095 | |
| 2096 | /* |
Akinobu Mita | 5778322 | 2015-09-27 02:09:23 +0900 | [diff] [blame] | 2097 | * Before hotadded cpu starts handling requests, new mappings must |
| 2098 | * be established. Otherwise, these requests in hw queue might |
| 2099 | * never be dispatched. |
| 2100 | * |
| 2101 | * For example, there is a single hw queue (hctx) and two CPU queues |
| 2102 | * (ctx0 for CPU0, and ctx1 for CPU1). |
| 2103 | * |
| 2104 | * Now CPU1 is just onlined and a request is inserted into |
| 2105 | * ctx1->rq_list and set bit0 in pending bitmap as ctx1->index_hw is |
| 2106 | * still zero. |
| 2107 | * |
| 2108 | * And then while running hw queue, flush_busy_ctxs() finds bit0 is |
| 2109 | * set in pending bitmap and tries to retrieve requests in |
| 2110 | * hctx->ctxs[0]->rq_list. But htx->ctxs[0] is a pointer to ctx0, |
| 2111 | * so the request in ctx1->rq_list is ignored. |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2112 | */ |
Akinobu Mita | 5778322 | 2015-09-27 02:09:23 +0900 | [diff] [blame] | 2113 | switch (action & ~CPU_TASKS_FROZEN) { |
| 2114 | case CPU_DEAD: |
| 2115 | case CPU_UP_CANCELED: |
| 2116 | cpumask_copy(&online_new, cpu_online_mask); |
| 2117 | break; |
| 2118 | case CPU_UP_PREPARE: |
| 2119 | cpumask_copy(&online_new, cpu_online_mask); |
| 2120 | cpumask_set_cpu(cpu, &online_new); |
| 2121 | break; |
| 2122 | default: |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2123 | return NOTIFY_OK; |
Akinobu Mita | 5778322 | 2015-09-27 02:09:23 +0900 | [diff] [blame] | 2124 | } |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2125 | |
| 2126 | mutex_lock(&all_q_mutex); |
Tejun Heo | f3af020 | 2014-11-04 13:52:27 -0500 | [diff] [blame] | 2127 | |
| 2128 | /* |
| 2129 | * We need to freeze and reinit all existing queues. Freezing |
| 2130 | * involves synchronous wait for an RCU grace period and doing it |
| 2131 | * one by one may take a long time. Start freezing all queues in |
| 2132 | * one swoop and then wait for the completions so that freezing can |
| 2133 | * take place in parallel. |
| 2134 | */ |
| 2135 | list_for_each_entry(q, &all_q_list, all_q_node) |
| 2136 | blk_mq_freeze_queue_start(q); |
Ming Lei | f054b56 | 2015-04-21 10:00:19 +0800 | [diff] [blame] | 2137 | list_for_each_entry(q, &all_q_list, all_q_node) { |
Tejun Heo | f3af020 | 2014-11-04 13:52:27 -0500 | [diff] [blame] | 2138 | blk_mq_freeze_queue_wait(q); |
| 2139 | |
Ming Lei | f054b56 | 2015-04-21 10:00:19 +0800 | [diff] [blame] | 2140 | /* |
| 2141 | * timeout handler can't touch hw queue during the |
| 2142 | * reinitialization |
| 2143 | */ |
| 2144 | del_timer_sync(&q->timeout); |
| 2145 | } |
| 2146 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2147 | list_for_each_entry(q, &all_q_list, all_q_node) |
Akinobu Mita | 5778322 | 2015-09-27 02:09:23 +0900 | [diff] [blame] | 2148 | blk_mq_queue_reinit(q, &online_new); |
Tejun Heo | f3af020 | 2014-11-04 13:52:27 -0500 | [diff] [blame] | 2149 | |
| 2150 | list_for_each_entry(q, &all_q_list, all_q_node) |
| 2151 | blk_mq_unfreeze_queue(q); |
| 2152 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2153 | mutex_unlock(&all_q_mutex); |
| 2154 | return NOTIFY_OK; |
| 2155 | } |
| 2156 | |
Jens Axboe | a516440 | 2014-09-10 09:02:03 -0600 | [diff] [blame] | 2157 | static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) |
| 2158 | { |
| 2159 | int i; |
| 2160 | |
| 2161 | for (i = 0; i < set->nr_hw_queues; i++) { |
| 2162 | set->tags[i] = blk_mq_init_rq_map(set, i); |
| 2163 | if (!set->tags[i]) |
| 2164 | goto out_unwind; |
| 2165 | } |
| 2166 | |
| 2167 | return 0; |
| 2168 | |
| 2169 | out_unwind: |
| 2170 | while (--i >= 0) |
| 2171 | blk_mq_free_rq_map(set, set->tags[i], i); |
| 2172 | |
Jens Axboe | a516440 | 2014-09-10 09:02:03 -0600 | [diff] [blame] | 2173 | return -ENOMEM; |
| 2174 | } |
| 2175 | |
| 2176 | /* |
| 2177 | * Allocate the request maps associated with this tag_set. Note that this |
| 2178 | * may reduce the depth asked for, if memory is tight. set->queue_depth |
| 2179 | * will be updated to reflect the allocated depth. |
| 2180 | */ |
| 2181 | static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) |
| 2182 | { |
| 2183 | unsigned int depth; |
| 2184 | int err; |
| 2185 | |
| 2186 | depth = set->queue_depth; |
| 2187 | do { |
| 2188 | err = __blk_mq_alloc_rq_maps(set); |
| 2189 | if (!err) |
| 2190 | break; |
| 2191 | |
| 2192 | set->queue_depth >>= 1; |
| 2193 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) { |
| 2194 | err = -ENOMEM; |
| 2195 | break; |
| 2196 | } |
| 2197 | } while (set->queue_depth); |
| 2198 | |
| 2199 | if (!set->queue_depth || err) { |
| 2200 | pr_err("blk-mq: failed to allocate request map\n"); |
| 2201 | return -ENOMEM; |
| 2202 | } |
| 2203 | |
| 2204 | if (depth != set->queue_depth) |
| 2205 | pr_info("blk-mq: reduced tag depth (%u -> %u)\n", |
| 2206 | depth, set->queue_depth); |
| 2207 | |
| 2208 | return 0; |
| 2209 | } |
| 2210 | |
Keith Busch | f26cdc8 | 2015-06-01 09:29:53 -0600 | [diff] [blame] | 2211 | struct cpumask *blk_mq_tags_cpumask(struct blk_mq_tags *tags) |
| 2212 | { |
| 2213 | return tags->cpumask; |
| 2214 | } |
| 2215 | EXPORT_SYMBOL_GPL(blk_mq_tags_cpumask); |
| 2216 | |
Jens Axboe | a4391c6 | 2014-06-05 15:21:56 -0600 | [diff] [blame] | 2217 | /* |
| 2218 | * Alloc a tag set to be associated with one or more request queues. |
| 2219 | * May fail with EINVAL for various error conditions. May adjust the |
| 2220 | * requested depth down, if if it too large. In that case, the set |
| 2221 | * value will be stored in set->queue_depth. |
| 2222 | */ |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 2223 | int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) |
| 2224 | { |
Bart Van Assche | 205fb5f | 2014-10-30 14:45:11 +0100 | [diff] [blame] | 2225 | BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS); |
| 2226 | |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 2227 | if (!set->nr_hw_queues) |
| 2228 | return -EINVAL; |
Jens Axboe | a4391c6 | 2014-06-05 15:21:56 -0600 | [diff] [blame] | 2229 | if (!set->queue_depth) |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 2230 | return -EINVAL; |
| 2231 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) |
| 2232 | return -EINVAL; |
| 2233 | |
Xiaoguang Wang | f9018ac | 2015-03-30 13:19:14 +0800 | [diff] [blame] | 2234 | if (!set->ops->queue_rq || !set->ops->map_queue) |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 2235 | return -EINVAL; |
| 2236 | |
Jens Axboe | a4391c6 | 2014-06-05 15:21:56 -0600 | [diff] [blame] | 2237 | if (set->queue_depth > BLK_MQ_MAX_DEPTH) { |
| 2238 | pr_info("blk-mq: reduced tag depth to %u\n", |
| 2239 | BLK_MQ_MAX_DEPTH); |
| 2240 | set->queue_depth = BLK_MQ_MAX_DEPTH; |
| 2241 | } |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 2242 | |
Shaohua Li | 6637fad | 2014-11-30 16:00:58 -0800 | [diff] [blame] | 2243 | /* |
| 2244 | * If a crashdump is active, then we are potentially in a very |
| 2245 | * memory constrained environment. Limit us to 1 queue and |
| 2246 | * 64 tags to prevent using too much memory. |
| 2247 | */ |
| 2248 | if (is_kdump_kernel()) { |
| 2249 | set->nr_hw_queues = 1; |
| 2250 | set->queue_depth = min(64U, set->queue_depth); |
| 2251 | } |
| 2252 | |
Ming Lei | 4847900 | 2014-04-19 18:00:17 +0800 | [diff] [blame] | 2253 | set->tags = kmalloc_node(set->nr_hw_queues * |
| 2254 | sizeof(struct blk_mq_tags *), |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 2255 | GFP_KERNEL, set->numa_node); |
| 2256 | if (!set->tags) |
Jens Axboe | a516440 | 2014-09-10 09:02:03 -0600 | [diff] [blame] | 2257 | return -ENOMEM; |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 2258 | |
Jens Axboe | a516440 | 2014-09-10 09:02:03 -0600 | [diff] [blame] | 2259 | if (blk_mq_alloc_rq_maps(set)) |
| 2260 | goto enomem; |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 2261 | |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 2262 | mutex_init(&set->tag_list_lock); |
| 2263 | INIT_LIST_HEAD(&set->tag_list); |
| 2264 | |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 2265 | return 0; |
Jens Axboe | a516440 | 2014-09-10 09:02:03 -0600 | [diff] [blame] | 2266 | enomem: |
Robert Elliott | 5676e7b | 2014-09-02 11:38:44 -0500 | [diff] [blame] | 2267 | kfree(set->tags); |
| 2268 | set->tags = NULL; |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 2269 | return -ENOMEM; |
| 2270 | } |
| 2271 | EXPORT_SYMBOL(blk_mq_alloc_tag_set); |
| 2272 | |
| 2273 | void blk_mq_free_tag_set(struct blk_mq_tag_set *set) |
| 2274 | { |
| 2275 | int i; |
| 2276 | |
Jens Axboe | 484b406 | 2014-05-21 14:01:15 -0600 | [diff] [blame] | 2277 | for (i = 0; i < set->nr_hw_queues; i++) { |
Keith Busch | f26cdc8 | 2015-06-01 09:29:53 -0600 | [diff] [blame] | 2278 | if (set->tags[i]) { |
Jens Axboe | 484b406 | 2014-05-21 14:01:15 -0600 | [diff] [blame] | 2279 | blk_mq_free_rq_map(set, set->tags[i], i); |
Keith Busch | f26cdc8 | 2015-06-01 09:29:53 -0600 | [diff] [blame] | 2280 | free_cpumask_var(set->tags[i]->cpumask); |
| 2281 | } |
Jens Axboe | 484b406 | 2014-05-21 14:01:15 -0600 | [diff] [blame] | 2282 | } |
| 2283 | |
Ming Lei | 981bd18 | 2014-04-24 00:07:34 +0800 | [diff] [blame] | 2284 | kfree(set->tags); |
Robert Elliott | 5676e7b | 2014-09-02 11:38:44 -0500 | [diff] [blame] | 2285 | set->tags = NULL; |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 2286 | } |
| 2287 | EXPORT_SYMBOL(blk_mq_free_tag_set); |
| 2288 | |
Jens Axboe | e3a2b3f | 2014-05-20 11:49:02 -0600 | [diff] [blame] | 2289 | int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr) |
| 2290 | { |
| 2291 | struct blk_mq_tag_set *set = q->tag_set; |
| 2292 | struct blk_mq_hw_ctx *hctx; |
| 2293 | int i, ret; |
| 2294 | |
| 2295 | if (!set || nr > set->queue_depth) |
| 2296 | return -EINVAL; |
| 2297 | |
| 2298 | ret = 0; |
| 2299 | queue_for_each_hw_ctx(q, hctx, i) { |
| 2300 | ret = blk_mq_tag_update_depth(hctx->tags, nr); |
| 2301 | if (ret) |
| 2302 | break; |
| 2303 | } |
| 2304 | |
| 2305 | if (!ret) |
| 2306 | q->nr_requests = nr; |
| 2307 | |
| 2308 | return ret; |
| 2309 | } |
| 2310 | |
Jens Axboe | 676141e | 2014-03-20 13:29:18 -0600 | [diff] [blame] | 2311 | void blk_mq_disable_hotplug(void) |
| 2312 | { |
| 2313 | mutex_lock(&all_q_mutex); |
| 2314 | } |
| 2315 | |
| 2316 | void blk_mq_enable_hotplug(void) |
| 2317 | { |
| 2318 | mutex_unlock(&all_q_mutex); |
| 2319 | } |
| 2320 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2321 | static int __init blk_mq_init(void) |
| 2322 | { |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2323 | blk_mq_cpu_init(); |
| 2324 | |
Tejun Heo | add703f | 2014-07-01 10:34:38 -0600 | [diff] [blame] | 2325 | hotcpu_notifier(blk_mq_queue_reinit_notify, 0); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2326 | |
| 2327 | return 0; |
| 2328 | } |
| 2329 | subsys_initcall(blk_mq_init); |