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> |
| 12 | #include <linux/mm.h> |
| 13 | #include <linux/init.h> |
| 14 | #include <linux/slab.h> |
| 15 | #include <linux/workqueue.h> |
| 16 | #include <linux/smp.h> |
| 17 | #include <linux/llist.h> |
| 18 | #include <linux/list_sort.h> |
| 19 | #include <linux/cpu.h> |
| 20 | #include <linux/cache.h> |
| 21 | #include <linux/sched/sysctl.h> |
| 22 | #include <linux/delay.h> |
Jens Axboe | aedcd72 | 2014-09-17 08:27:03 -0600 | [diff] [blame] | 23 | #include <linux/crash_dump.h> |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 24 | |
| 25 | #include <trace/events/block.h> |
| 26 | |
| 27 | #include <linux/blk-mq.h> |
| 28 | #include "blk.h" |
| 29 | #include "blk-mq.h" |
| 30 | #include "blk-mq-tag.h" |
| 31 | |
| 32 | static DEFINE_MUTEX(all_q_mutex); |
| 33 | static LIST_HEAD(all_q_list); |
| 34 | |
| 35 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx); |
| 36 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 37 | /* |
| 38 | * Check if any of the ctx's have pending work in this hardware queue |
| 39 | */ |
| 40 | static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx) |
| 41 | { |
| 42 | unsigned int i; |
| 43 | |
Jens Axboe | 1429d7c | 2014-05-19 09:23:55 -0600 | [diff] [blame] | 44 | for (i = 0; i < hctx->ctx_map.map_size; i++) |
| 45 | if (hctx->ctx_map.map[i].word) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 46 | return true; |
| 47 | |
| 48 | return false; |
| 49 | } |
| 50 | |
Jens Axboe | 1429d7c | 2014-05-19 09:23:55 -0600 | [diff] [blame] | 51 | static inline struct blk_align_bitmap *get_bm(struct blk_mq_hw_ctx *hctx, |
| 52 | struct blk_mq_ctx *ctx) |
| 53 | { |
| 54 | return &hctx->ctx_map.map[ctx->index_hw / hctx->ctx_map.bits_per_word]; |
| 55 | } |
| 56 | |
| 57 | #define CTX_TO_BIT(hctx, ctx) \ |
| 58 | ((ctx)->index_hw & ((hctx)->ctx_map.bits_per_word - 1)) |
| 59 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 60 | /* |
| 61 | * Mark this ctx as having pending work in this hardware queue |
| 62 | */ |
| 63 | static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx, |
| 64 | struct blk_mq_ctx *ctx) |
| 65 | { |
Jens Axboe | 1429d7c | 2014-05-19 09:23:55 -0600 | [diff] [blame] | 66 | struct blk_align_bitmap *bm = get_bm(hctx, ctx); |
| 67 | |
| 68 | if (!test_bit(CTX_TO_BIT(hctx, ctx), &bm->word)) |
| 69 | set_bit(CTX_TO_BIT(hctx, ctx), &bm->word); |
| 70 | } |
| 71 | |
| 72 | static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx, |
| 73 | struct blk_mq_ctx *ctx) |
| 74 | { |
| 75 | struct blk_align_bitmap *bm = get_bm(hctx, ctx); |
| 76 | |
| 77 | clear_bit(CTX_TO_BIT(hctx, ctx), &bm->word); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 78 | } |
| 79 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 80 | static int blk_mq_queue_enter(struct request_queue *q) |
| 81 | { |
Tejun Heo | add703f | 2014-07-01 10:34:38 -0600 | [diff] [blame] | 82 | while (true) { |
| 83 | int ret; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 84 | |
Tejun Heo | add703f | 2014-07-01 10:34:38 -0600 | [diff] [blame] | 85 | if (percpu_ref_tryget_live(&q->mq_usage_counter)) |
| 86 | return 0; |
Keith Busch | 3b632cf | 2014-06-06 10:22:07 -0600 | [diff] [blame] | 87 | |
Tejun Heo | add703f | 2014-07-01 10:34:38 -0600 | [diff] [blame] | 88 | ret = wait_event_interruptible(q->mq_freeze_wq, |
| 89 | !q->mq_freeze_depth || blk_queue_dying(q)); |
| 90 | if (blk_queue_dying(q)) |
| 91 | return -ENODEV; |
| 92 | if (ret) |
| 93 | return ret; |
| 94 | } |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 95 | } |
| 96 | |
| 97 | static void blk_mq_queue_exit(struct request_queue *q) |
| 98 | { |
Tejun Heo | add703f | 2014-07-01 10:34:38 -0600 | [diff] [blame] | 99 | percpu_ref_put(&q->mq_usage_counter); |
| 100 | } |
| 101 | |
| 102 | static void blk_mq_usage_counter_release(struct percpu_ref *ref) |
| 103 | { |
| 104 | struct request_queue *q = |
| 105 | container_of(ref, struct request_queue, mq_usage_counter); |
| 106 | |
| 107 | wake_up_all(&q->mq_freeze_wq); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 108 | } |
| 109 | |
Tejun Heo | 72d6f02 | 2014-07-01 10:33:02 -0600 | [diff] [blame] | 110 | /* |
| 111 | * Guarantee no request is in use, so we can change any data structure of |
| 112 | * the queue afterward. |
| 113 | */ |
| 114 | void blk_mq_freeze_queue(struct request_queue *q) |
Ming Lei | 43a5e4e | 2013-12-26 21:31:35 +0800 | [diff] [blame] | 115 | { |
Tejun Heo | cddd5d1 | 2014-08-16 08:02:24 -0400 | [diff] [blame] | 116 | bool freeze; |
| 117 | |
Tejun Heo | 72d6f02 | 2014-07-01 10:33:02 -0600 | [diff] [blame] | 118 | spin_lock_irq(q->queue_lock); |
Tejun Heo | cddd5d1 | 2014-08-16 08:02:24 -0400 | [diff] [blame] | 119 | freeze = !q->mq_freeze_depth++; |
Tejun Heo | 72d6f02 | 2014-07-01 10:33:02 -0600 | [diff] [blame] | 120 | spin_unlock_irq(q->queue_lock); |
| 121 | |
Tejun Heo | cddd5d1 | 2014-08-16 08:02:24 -0400 | [diff] [blame] | 122 | if (freeze) { |
| 123 | percpu_ref_kill(&q->mq_usage_counter); |
| 124 | blk_mq_run_queues(q, false); |
| 125 | } |
Tejun Heo | add703f | 2014-07-01 10:34:38 -0600 | [diff] [blame] | 126 | wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->mq_usage_counter)); |
Ming Lei | 43a5e4e | 2013-12-26 21:31:35 +0800 | [diff] [blame] | 127 | } |
| 128 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 129 | static void blk_mq_unfreeze_queue(struct request_queue *q) |
| 130 | { |
Tejun Heo | cddd5d1 | 2014-08-16 08:02:24 -0400 | [diff] [blame] | 131 | bool wake; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 132 | |
| 133 | spin_lock_irq(q->queue_lock); |
Tejun Heo | 780db20 | 2014-07-01 10:31:13 -0600 | [diff] [blame] | 134 | wake = !--q->mq_freeze_depth; |
| 135 | WARN_ON_ONCE(q->mq_freeze_depth < 0); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 136 | spin_unlock_irq(q->queue_lock); |
Tejun Heo | add703f | 2014-07-01 10:34:38 -0600 | [diff] [blame] | 137 | if (wake) { |
| 138 | percpu_ref_reinit(&q->mq_usage_counter); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 139 | wake_up_all(&q->mq_freeze_wq); |
Tejun Heo | add703f | 2014-07-01 10:34:38 -0600 | [diff] [blame] | 140 | } |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 141 | } |
| 142 | |
| 143 | bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx) |
| 144 | { |
| 145 | return blk_mq_has_free_tags(hctx->tags); |
| 146 | } |
| 147 | EXPORT_SYMBOL(blk_mq_can_queue); |
| 148 | |
Jens Axboe | 94eddfb | 2013-11-19 09:25:07 -0700 | [diff] [blame] | 149 | static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx, |
| 150 | struct request *rq, unsigned int rw_flags) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 151 | { |
Jens Axboe | 94eddfb | 2013-11-19 09:25:07 -0700 | [diff] [blame] | 152 | if (blk_queue_io_stat(q)) |
| 153 | rw_flags |= REQ_IO_STAT; |
| 154 | |
Christoph Hellwig | af76e55 | 2014-05-06 12:12:45 +0200 | [diff] [blame] | 155 | INIT_LIST_HEAD(&rq->queuelist); |
| 156 | /* csd/requeue_work/fifo_time is initialized before use */ |
| 157 | rq->q = q; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 158 | rq->mq_ctx = ctx; |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 159 | rq->cmd_flags |= rw_flags; |
Christoph Hellwig | af76e55 | 2014-05-06 12:12:45 +0200 | [diff] [blame] | 160 | /* do not touch atomic flags, it needs atomic ops against the timer */ |
| 161 | rq->cpu = -1; |
Christoph Hellwig | af76e55 | 2014-05-06 12:12:45 +0200 | [diff] [blame] | 162 | INIT_HLIST_NODE(&rq->hash); |
| 163 | RB_CLEAR_NODE(&rq->rb_node); |
Christoph Hellwig | af76e55 | 2014-05-06 12:12:45 +0200 | [diff] [blame] | 164 | rq->rq_disk = NULL; |
| 165 | rq->part = NULL; |
Jens Axboe | 3ee3237 | 2014-06-09 09:36:53 -0600 | [diff] [blame] | 166 | rq->start_time = jiffies; |
Christoph Hellwig | af76e55 | 2014-05-06 12:12:45 +0200 | [diff] [blame] | 167 | #ifdef CONFIG_BLK_CGROUP |
| 168 | rq->rl = NULL; |
Ming Lei | 0fec08b | 2014-01-03 10:00:08 -0700 | [diff] [blame] | 169 | set_start_time_ns(rq); |
Christoph Hellwig | af76e55 | 2014-05-06 12:12:45 +0200 | [diff] [blame] | 170 | rq->io_start_time_ns = 0; |
| 171 | #endif |
| 172 | rq->nr_phys_segments = 0; |
| 173 | #if defined(CONFIG_BLK_DEV_INTEGRITY) |
| 174 | rq->nr_integrity_segments = 0; |
| 175 | #endif |
Christoph Hellwig | af76e55 | 2014-05-06 12:12:45 +0200 | [diff] [blame] | 176 | rq->special = NULL; |
| 177 | /* tag was already set */ |
| 178 | rq->errors = 0; |
Christoph Hellwig | af76e55 | 2014-05-06 12:12:45 +0200 | [diff] [blame] | 179 | |
Tony Battersby | 6f4a1626 | 2014-08-22 15:53:39 -0400 | [diff] [blame] | 180 | rq->cmd = rq->__cmd; |
| 181 | |
Christoph Hellwig | af76e55 | 2014-05-06 12:12:45 +0200 | [diff] [blame] | 182 | rq->extra_len = 0; |
| 183 | rq->sense_len = 0; |
| 184 | rq->resid_len = 0; |
| 185 | rq->sense = NULL; |
| 186 | |
Christoph Hellwig | af76e55 | 2014-05-06 12:12:45 +0200 | [diff] [blame] | 187 | INIT_LIST_HEAD(&rq->timeout_list); |
Jens Axboe | f6be4fb | 2014-06-06 11:03:48 -0600 | [diff] [blame] | 188 | rq->timeout = 0; |
| 189 | |
Christoph Hellwig | af76e55 | 2014-05-06 12:12:45 +0200 | [diff] [blame] | 190 | rq->end_io = NULL; |
| 191 | rq->end_io_data = NULL; |
| 192 | rq->next_rq = NULL; |
| 193 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 194 | ctx->rq_dispatched[rw_is_sync(rw_flags)]++; |
| 195 | } |
| 196 | |
Christoph Hellwig | 5dee857 | 2014-05-27 20:59:47 +0200 | [diff] [blame] | 197 | static struct request * |
Ming Lei | cb96a42 | 2014-06-01 00:43:37 +0800 | [diff] [blame] | 198 | __blk_mq_alloc_request(struct blk_mq_alloc_data *data, int rw) |
Christoph Hellwig | 5dee857 | 2014-05-27 20:59:47 +0200 | [diff] [blame] | 199 | { |
| 200 | struct request *rq; |
| 201 | unsigned int tag; |
| 202 | |
Ming Lei | cb96a42 | 2014-06-01 00:43:37 +0800 | [diff] [blame] | 203 | tag = blk_mq_get_tag(data); |
Christoph Hellwig | 5dee857 | 2014-05-27 20:59:47 +0200 | [diff] [blame] | 204 | if (tag != BLK_MQ_TAG_FAIL) { |
Ming Lei | cb96a42 | 2014-06-01 00:43:37 +0800 | [diff] [blame] | 205 | rq = data->hctx->tags->rqs[tag]; |
Christoph Hellwig | 5dee857 | 2014-05-27 20:59:47 +0200 | [diff] [blame] | 206 | |
Ming Lei | cb96a42 | 2014-06-01 00:43:37 +0800 | [diff] [blame] | 207 | if (blk_mq_tag_busy(data->hctx)) { |
Christoph Hellwig | 5dee857 | 2014-05-27 20:59:47 +0200 | [diff] [blame] | 208 | rq->cmd_flags = REQ_MQ_INFLIGHT; |
Ming Lei | cb96a42 | 2014-06-01 00:43:37 +0800 | [diff] [blame] | 209 | atomic_inc(&data->hctx->nr_active); |
Christoph Hellwig | 5dee857 | 2014-05-27 20:59:47 +0200 | [diff] [blame] | 210 | } |
| 211 | |
| 212 | rq->tag = tag; |
Ming Lei | cb96a42 | 2014-06-01 00:43:37 +0800 | [diff] [blame] | 213 | blk_mq_rq_ctx_init(data->q, data->ctx, rq, rw); |
Christoph Hellwig | 5dee857 | 2014-05-27 20:59:47 +0200 | [diff] [blame] | 214 | return rq; |
| 215 | } |
| 216 | |
| 217 | return NULL; |
| 218 | } |
| 219 | |
Christoph Hellwig | 4ce01dd | 2014-05-27 20:59:46 +0200 | [diff] [blame] | 220 | struct request *blk_mq_alloc_request(struct request_queue *q, int rw, gfp_t gfp, |
| 221 | bool reserved) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 222 | { |
Christoph Hellwig | d852564 | 2014-05-27 20:59:50 +0200 | [diff] [blame] | 223 | struct blk_mq_ctx *ctx; |
| 224 | struct blk_mq_hw_ctx *hctx; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 225 | struct request *rq; |
Ming Lei | cb96a42 | 2014-06-01 00:43:37 +0800 | [diff] [blame] | 226 | struct blk_mq_alloc_data alloc_data; |
Joe Lawrence | a492f07 | 2014-08-28 08:15:21 -0600 | [diff] [blame] | 227 | int ret; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 228 | |
Joe Lawrence | a492f07 | 2014-08-28 08:15:21 -0600 | [diff] [blame] | 229 | ret = blk_mq_queue_enter(q); |
| 230 | if (ret) |
| 231 | return ERR_PTR(ret); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 232 | |
Christoph Hellwig | d852564 | 2014-05-27 20:59:50 +0200 | [diff] [blame] | 233 | ctx = blk_mq_get_ctx(q); |
| 234 | hctx = q->mq_ops->map_queue(q, ctx->cpu); |
Ming Lei | cb96a42 | 2014-06-01 00:43:37 +0800 | [diff] [blame] | 235 | blk_mq_set_alloc_data(&alloc_data, q, gfp & ~__GFP_WAIT, |
| 236 | reserved, ctx, hctx); |
Christoph Hellwig | d852564 | 2014-05-27 20:59:50 +0200 | [diff] [blame] | 237 | |
Ming Lei | cb96a42 | 2014-06-01 00:43:37 +0800 | [diff] [blame] | 238 | rq = __blk_mq_alloc_request(&alloc_data, rw); |
Christoph Hellwig | d852564 | 2014-05-27 20:59:50 +0200 | [diff] [blame] | 239 | if (!rq && (gfp & __GFP_WAIT)) { |
| 240 | __blk_mq_run_hw_queue(hctx); |
| 241 | blk_mq_put_ctx(ctx); |
| 242 | |
| 243 | ctx = blk_mq_get_ctx(q); |
| 244 | hctx = q->mq_ops->map_queue(q, ctx->cpu); |
Ming Lei | cb96a42 | 2014-06-01 00:43:37 +0800 | [diff] [blame] | 245 | blk_mq_set_alloc_data(&alloc_data, q, gfp, reserved, ctx, |
| 246 | hctx); |
| 247 | rq = __blk_mq_alloc_request(&alloc_data, rw); |
| 248 | ctx = alloc_data.ctx; |
Christoph Hellwig | d852564 | 2014-05-27 20:59:50 +0200 | [diff] [blame] | 249 | } |
| 250 | blk_mq_put_ctx(ctx); |
Joe Lawrence | a492f07 | 2014-08-28 08:15:21 -0600 | [diff] [blame] | 251 | if (!rq) |
| 252 | return ERR_PTR(-EWOULDBLOCK); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 253 | return rq; |
| 254 | } |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 255 | EXPORT_SYMBOL(blk_mq_alloc_request); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 256 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 257 | static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx, |
| 258 | struct blk_mq_ctx *ctx, struct request *rq) |
| 259 | { |
| 260 | const int tag = rq->tag; |
| 261 | struct request_queue *q = rq->q; |
| 262 | |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 263 | if (rq->cmd_flags & REQ_MQ_INFLIGHT) |
| 264 | atomic_dec(&hctx->nr_active); |
David Hildenbrand | 683d0e1 | 2014-09-18 11:04:31 +0200 | [diff] [blame] | 265 | rq->cmd_flags = 0; |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 266 | |
Christoph Hellwig | af76e55 | 2014-05-06 12:12:45 +0200 | [diff] [blame] | 267 | clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags); |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 268 | blk_mq_put_tag(hctx, tag, &ctx->last_tag); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 269 | blk_mq_queue_exit(q); |
| 270 | } |
| 271 | |
| 272 | void blk_mq_free_request(struct request *rq) |
| 273 | { |
| 274 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
| 275 | struct blk_mq_hw_ctx *hctx; |
| 276 | struct request_queue *q = rq->q; |
| 277 | |
| 278 | ctx->rq_completed[rq_is_sync(rq)]++; |
| 279 | |
| 280 | hctx = q->mq_ops->map_queue(q, ctx->cpu); |
| 281 | __blk_mq_free_request(hctx, ctx, rq); |
| 282 | } |
| 283 | |
Christoph Hellwig | 8727af4 | 2014-04-14 10:30:08 +0200 | [diff] [blame] | 284 | /* |
| 285 | * Clone all relevant state from a request that has been put on hold in |
| 286 | * the flush state machine into the preallocated flush request that hangs |
| 287 | * off the request queue. |
| 288 | * |
| 289 | * For a driver the flush request should be invisible, that's why we are |
| 290 | * impersonating the original request here. |
| 291 | */ |
| 292 | void blk_mq_clone_flush_request(struct request *flush_rq, |
| 293 | struct request *orig_rq) |
| 294 | { |
| 295 | struct blk_mq_hw_ctx *hctx = |
| 296 | orig_rq->q->mq_ops->map_queue(orig_rq->q, orig_rq->mq_ctx->cpu); |
| 297 | |
| 298 | flush_rq->mq_ctx = orig_rq->mq_ctx; |
| 299 | flush_rq->tag = orig_rq->tag; |
| 300 | memcpy(blk_mq_rq_to_pdu(flush_rq), blk_mq_rq_to_pdu(orig_rq), |
| 301 | hctx->cmd_size); |
| 302 | } |
| 303 | |
Christoph Hellwig | c8a446a | 2014-09-13 16:40:10 -0700 | [diff] [blame] | 304 | inline void __blk_mq_end_request(struct request *rq, int error) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 305 | { |
Ming Lei | 0d11e6a | 2013-12-05 10:50:39 -0700 | [diff] [blame] | 306 | blk_account_io_done(rq); |
| 307 | |
Christoph Hellwig | 91b6363 | 2014-04-16 09:44:53 +0200 | [diff] [blame] | 308 | if (rq->end_io) { |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 309 | rq->end_io(rq, error); |
Christoph Hellwig | 91b6363 | 2014-04-16 09:44:53 +0200 | [diff] [blame] | 310 | } else { |
| 311 | if (unlikely(blk_bidi_rq(rq))) |
| 312 | blk_mq_free_request(rq->next_rq); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 313 | blk_mq_free_request(rq); |
Christoph Hellwig | 91b6363 | 2014-04-16 09:44:53 +0200 | [diff] [blame] | 314 | } |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 315 | } |
Christoph Hellwig | c8a446a | 2014-09-13 16:40:10 -0700 | [diff] [blame] | 316 | EXPORT_SYMBOL(__blk_mq_end_request); |
Christoph Hellwig | 63151a4 | 2014-04-16 09:44:52 +0200 | [diff] [blame] | 317 | |
Christoph Hellwig | c8a446a | 2014-09-13 16:40:10 -0700 | [diff] [blame] | 318 | void blk_mq_end_request(struct request *rq, int error) |
Christoph Hellwig | 63151a4 | 2014-04-16 09:44:52 +0200 | [diff] [blame] | 319 | { |
| 320 | if (blk_update_request(rq, error, blk_rq_bytes(rq))) |
| 321 | BUG(); |
Christoph Hellwig | c8a446a | 2014-09-13 16:40:10 -0700 | [diff] [blame] | 322 | __blk_mq_end_request(rq, error); |
Christoph Hellwig | 63151a4 | 2014-04-16 09:44:52 +0200 | [diff] [blame] | 323 | } |
Christoph Hellwig | c8a446a | 2014-09-13 16:40:10 -0700 | [diff] [blame] | 324 | EXPORT_SYMBOL(blk_mq_end_request); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 325 | |
Christoph Hellwig | 30a91cb | 2014-02-10 03:24:38 -0800 | [diff] [blame] | 326 | static void __blk_mq_complete_request_remote(void *data) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 327 | { |
Christoph Hellwig | 3d6efbf | 2014-01-08 09:33:37 -0800 | [diff] [blame] | 328 | struct request *rq = data; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 329 | |
Christoph Hellwig | 30a91cb | 2014-02-10 03:24:38 -0800 | [diff] [blame] | 330 | rq->q->softirq_done_fn(rq); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 331 | } |
| 332 | |
Jens Axboe | ed85186 | 2014-05-30 21:20:50 -0600 | [diff] [blame] | 333 | static void blk_mq_ipi_complete_request(struct request *rq) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 334 | { |
| 335 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
Christoph Hellwig | 3853520 | 2014-04-25 02:32:53 -0700 | [diff] [blame] | 336 | bool shared = false; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 337 | int cpu; |
| 338 | |
Christoph Hellwig | 3853520 | 2014-04-25 02:32:53 -0700 | [diff] [blame] | 339 | if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) { |
Christoph Hellwig | 30a91cb | 2014-02-10 03:24:38 -0800 | [diff] [blame] | 340 | rq->q->softirq_done_fn(rq); |
| 341 | return; |
| 342 | } |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 343 | |
| 344 | cpu = get_cpu(); |
Christoph Hellwig | 3853520 | 2014-04-25 02:32:53 -0700 | [diff] [blame] | 345 | if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags)) |
| 346 | shared = cpus_share_cache(cpu, ctx->cpu); |
| 347 | |
| 348 | if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) { |
Christoph Hellwig | 30a91cb | 2014-02-10 03:24:38 -0800 | [diff] [blame] | 349 | rq->csd.func = __blk_mq_complete_request_remote; |
Christoph Hellwig | 3d6efbf | 2014-01-08 09:33:37 -0800 | [diff] [blame] | 350 | rq->csd.info = rq; |
| 351 | rq->csd.flags = 0; |
Frederic Weisbecker | c46fff2 | 2014-02-24 16:40:02 +0100 | [diff] [blame] | 352 | smp_call_function_single_async(ctx->cpu, &rq->csd); |
Christoph Hellwig | 3d6efbf | 2014-01-08 09:33:37 -0800 | [diff] [blame] | 353 | } else { |
Christoph Hellwig | 30a91cb | 2014-02-10 03:24:38 -0800 | [diff] [blame] | 354 | rq->q->softirq_done_fn(rq); |
Christoph Hellwig | 3d6efbf | 2014-01-08 09:33:37 -0800 | [diff] [blame] | 355 | } |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 356 | put_cpu(); |
| 357 | } |
Christoph Hellwig | 30a91cb | 2014-02-10 03:24:38 -0800 | [diff] [blame] | 358 | |
Jens Axboe | ed85186 | 2014-05-30 21:20:50 -0600 | [diff] [blame] | 359 | void __blk_mq_complete_request(struct request *rq) |
| 360 | { |
| 361 | struct request_queue *q = rq->q; |
| 362 | |
| 363 | if (!q->softirq_done_fn) |
Christoph Hellwig | c8a446a | 2014-09-13 16:40:10 -0700 | [diff] [blame] | 364 | blk_mq_end_request(rq, rq->errors); |
Jens Axboe | ed85186 | 2014-05-30 21:20:50 -0600 | [diff] [blame] | 365 | else |
| 366 | blk_mq_ipi_complete_request(rq); |
| 367 | } |
| 368 | |
Christoph Hellwig | 30a91cb | 2014-02-10 03:24:38 -0800 | [diff] [blame] | 369 | /** |
| 370 | * blk_mq_complete_request - end I/O on a request |
| 371 | * @rq: the request being processed |
| 372 | * |
| 373 | * Description: |
| 374 | * Ends all I/O on a request. It does not handle partial completions. |
| 375 | * The actual completion happens out-of-order, through a IPI handler. |
| 376 | **/ |
| 377 | void blk_mq_complete_request(struct request *rq) |
| 378 | { |
Jens Axboe | 95f0968 | 2014-05-27 17:46:48 -0600 | [diff] [blame] | 379 | struct request_queue *q = rq->q; |
| 380 | |
| 381 | if (unlikely(blk_should_fake_timeout(q))) |
Christoph Hellwig | 30a91cb | 2014-02-10 03:24:38 -0800 | [diff] [blame] | 382 | return; |
Jens Axboe | ed85186 | 2014-05-30 21:20:50 -0600 | [diff] [blame] | 383 | if (!blk_mark_rq_complete(rq)) |
| 384 | __blk_mq_complete_request(rq); |
Christoph Hellwig | 30a91cb | 2014-02-10 03:24:38 -0800 | [diff] [blame] | 385 | } |
| 386 | EXPORT_SYMBOL(blk_mq_complete_request); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 387 | |
Christoph Hellwig | e249007 | 2014-09-13 16:40:09 -0700 | [diff] [blame] | 388 | void blk_mq_start_request(struct request *rq) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 389 | { |
| 390 | struct request_queue *q = rq->q; |
| 391 | |
| 392 | trace_block_rq_issue(q, rq); |
| 393 | |
Christoph Hellwig | 742ee69 | 2014-04-14 10:30:06 +0200 | [diff] [blame] | 394 | rq->resid_len = blk_rq_bytes(rq); |
Christoph Hellwig | 91b6363 | 2014-04-16 09:44:53 +0200 | [diff] [blame] | 395 | if (unlikely(blk_bidi_rq(rq))) |
| 396 | rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq); |
Christoph Hellwig | 742ee69 | 2014-04-14 10:30:06 +0200 | [diff] [blame] | 397 | |
Ming Lei | 2b8393b | 2014-06-10 00:16:41 +0800 | [diff] [blame] | 398 | blk_add_timer(rq); |
Jens Axboe | 87ee7b1 | 2014-04-24 08:51:47 -0600 | [diff] [blame] | 399 | |
| 400 | /* |
Jens Axboe | 538b753 | 2014-09-16 10:37:37 -0600 | [diff] [blame] | 401 | * Ensure that ->deadline is visible before set the started |
| 402 | * flag and clear the completed flag. |
| 403 | */ |
| 404 | smp_mb__before_atomic(); |
| 405 | |
| 406 | /* |
Jens Axboe | 87ee7b1 | 2014-04-24 08:51:47 -0600 | [diff] [blame] | 407 | * Mark us as started and clear complete. Complete might have been |
| 408 | * set if requeue raced with timeout, which then marked it as |
| 409 | * complete. So be sure to clear complete again when we start |
| 410 | * the request, otherwise we'll ignore the completion event. |
| 411 | */ |
Jens Axboe | 4b57052 | 2014-05-29 11:00:11 -0600 | [diff] [blame] | 412 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) |
| 413 | set_bit(REQ_ATOM_STARTED, &rq->atomic_flags); |
| 414 | if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags)) |
| 415 | clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags); |
Christoph Hellwig | 49f5baa | 2014-02-11 08:27:14 -0800 | [diff] [blame] | 416 | |
| 417 | if (q->dma_drain_size && blk_rq_bytes(rq)) { |
| 418 | /* |
| 419 | * Make sure space for the drain appears. We know we can do |
| 420 | * this because max_hw_segments has been adjusted to be one |
| 421 | * fewer than the device can handle. |
| 422 | */ |
| 423 | rq->nr_phys_segments++; |
| 424 | } |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 425 | } |
Christoph Hellwig | e249007 | 2014-09-13 16:40:09 -0700 | [diff] [blame] | 426 | EXPORT_SYMBOL(blk_mq_start_request); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 427 | |
Christoph Hellwig | ed0791b | 2014-04-16 09:44:57 +0200 | [diff] [blame] | 428 | static void __blk_mq_requeue_request(struct request *rq) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 429 | { |
| 430 | struct request_queue *q = rq->q; |
| 431 | |
| 432 | trace_block_rq_requeue(q, rq); |
Christoph Hellwig | 49f5baa | 2014-02-11 08:27:14 -0800 | [diff] [blame] | 433 | |
Christoph Hellwig | e249007 | 2014-09-13 16:40:09 -0700 | [diff] [blame] | 434 | if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) { |
| 435 | if (q->dma_drain_size && blk_rq_bytes(rq)) |
| 436 | rq->nr_phys_segments--; |
| 437 | } |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 438 | } |
| 439 | |
Christoph Hellwig | ed0791b | 2014-04-16 09:44:57 +0200 | [diff] [blame] | 440 | void blk_mq_requeue_request(struct request *rq) |
| 441 | { |
Christoph Hellwig | ed0791b | 2014-04-16 09:44:57 +0200 | [diff] [blame] | 442 | __blk_mq_requeue_request(rq); |
Christoph Hellwig | ed0791b | 2014-04-16 09:44:57 +0200 | [diff] [blame] | 443 | |
Christoph Hellwig | ed0791b | 2014-04-16 09:44:57 +0200 | [diff] [blame] | 444 | BUG_ON(blk_queued_rq(rq)); |
Christoph Hellwig | 6fca6a6 | 2014-05-28 08:08:02 -0600 | [diff] [blame] | 445 | blk_mq_add_to_requeue_list(rq, true); |
Christoph Hellwig | ed0791b | 2014-04-16 09:44:57 +0200 | [diff] [blame] | 446 | } |
| 447 | EXPORT_SYMBOL(blk_mq_requeue_request); |
| 448 | |
Christoph Hellwig | 6fca6a6 | 2014-05-28 08:08:02 -0600 | [diff] [blame] | 449 | static void blk_mq_requeue_work(struct work_struct *work) |
| 450 | { |
| 451 | struct request_queue *q = |
| 452 | container_of(work, struct request_queue, requeue_work); |
| 453 | LIST_HEAD(rq_list); |
| 454 | struct request *rq, *next; |
| 455 | unsigned long flags; |
| 456 | |
| 457 | spin_lock_irqsave(&q->requeue_lock, flags); |
| 458 | list_splice_init(&q->requeue_list, &rq_list); |
| 459 | spin_unlock_irqrestore(&q->requeue_lock, flags); |
| 460 | |
| 461 | list_for_each_entry_safe(rq, next, &rq_list, queuelist) { |
| 462 | if (!(rq->cmd_flags & REQ_SOFTBARRIER)) |
| 463 | continue; |
| 464 | |
| 465 | rq->cmd_flags &= ~REQ_SOFTBARRIER; |
| 466 | list_del_init(&rq->queuelist); |
| 467 | blk_mq_insert_request(rq, true, false, false); |
| 468 | } |
| 469 | |
| 470 | while (!list_empty(&rq_list)) { |
| 471 | rq = list_entry(rq_list.next, struct request, queuelist); |
| 472 | list_del_init(&rq->queuelist); |
| 473 | blk_mq_insert_request(rq, false, false, false); |
| 474 | } |
| 475 | |
Jens Axboe | 8b95741 | 2014-09-19 13:10:29 -0600 | [diff] [blame] | 476 | /* |
| 477 | * Use the start variant of queue running here, so that running |
| 478 | * the requeue work will kick stopped queues. |
| 479 | */ |
| 480 | blk_mq_start_hw_queues(q); |
Christoph Hellwig | 6fca6a6 | 2014-05-28 08:08:02 -0600 | [diff] [blame] | 481 | } |
| 482 | |
| 483 | void blk_mq_add_to_requeue_list(struct request *rq, bool at_head) |
| 484 | { |
| 485 | struct request_queue *q = rq->q; |
| 486 | unsigned long flags; |
| 487 | |
| 488 | /* |
| 489 | * We abuse this flag that is otherwise used by the I/O scheduler to |
| 490 | * request head insertation from the workqueue. |
| 491 | */ |
| 492 | BUG_ON(rq->cmd_flags & REQ_SOFTBARRIER); |
| 493 | |
| 494 | spin_lock_irqsave(&q->requeue_lock, flags); |
| 495 | if (at_head) { |
| 496 | rq->cmd_flags |= REQ_SOFTBARRIER; |
| 497 | list_add(&rq->queuelist, &q->requeue_list); |
| 498 | } else { |
| 499 | list_add_tail(&rq->queuelist, &q->requeue_list); |
| 500 | } |
| 501 | spin_unlock_irqrestore(&q->requeue_lock, flags); |
| 502 | } |
| 503 | EXPORT_SYMBOL(blk_mq_add_to_requeue_list); |
| 504 | |
| 505 | void blk_mq_kick_requeue_list(struct request_queue *q) |
| 506 | { |
| 507 | kblockd_schedule_work(&q->requeue_work); |
| 508 | } |
| 509 | EXPORT_SYMBOL(blk_mq_kick_requeue_list); |
| 510 | |
Ming Lei | 7c94e1c | 2014-09-25 23:23:43 +0800 | [diff] [blame] | 511 | static inline bool is_flush_request(struct request *rq, |
| 512 | struct blk_flush_queue *fq, unsigned int tag) |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 513 | { |
Jens Axboe | 0e62f51 | 2014-06-04 10:23:49 -0600 | [diff] [blame] | 514 | return ((rq->cmd_flags & REQ_FLUSH_SEQ) && |
Ming Lei | 7c94e1c | 2014-09-25 23:23:43 +0800 | [diff] [blame] | 515 | fq->flush_rq->tag == tag); |
Jens Axboe | 0e62f51 | 2014-06-04 10:23:49 -0600 | [diff] [blame] | 516 | } |
Shaohua Li | 2230237 | 2014-05-30 08:06:42 -0600 | [diff] [blame] | 517 | |
Jens Axboe | 0e62f51 | 2014-06-04 10:23:49 -0600 | [diff] [blame] | 518 | struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag) |
| 519 | { |
| 520 | struct request *rq = tags->rqs[tag]; |
Ming Lei | e97c293 | 2014-09-25 23:23:46 +0800 | [diff] [blame^] | 521 | /* mq_ctx of flush rq is always cloned from the corresponding req */ |
| 522 | struct blk_flush_queue *fq = blk_get_flush_queue(rq->q, rq->mq_ctx); |
Shaohua Li | 2230237 | 2014-05-30 08:06:42 -0600 | [diff] [blame] | 523 | |
Ming Lei | 7c94e1c | 2014-09-25 23:23:43 +0800 | [diff] [blame] | 524 | if (!is_flush_request(rq, fq, tag)) |
Jens Axboe | 0e62f51 | 2014-06-04 10:23:49 -0600 | [diff] [blame] | 525 | return rq; |
| 526 | |
Ming Lei | 7c94e1c | 2014-09-25 23:23:43 +0800 | [diff] [blame] | 527 | return fq->flush_rq; |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 528 | } |
| 529 | EXPORT_SYMBOL(blk_mq_tag_to_rq); |
| 530 | |
Christoph Hellwig | 46f92d4 | 2014-09-13 16:40:12 -0700 | [diff] [blame] | 531 | struct blk_mq_timeout_data { |
| 532 | unsigned long next; |
| 533 | unsigned int next_set; |
| 534 | }; |
| 535 | |
Christoph Hellwig | 9041583 | 2014-09-22 10:21:48 -0600 | [diff] [blame] | 536 | void blk_mq_rq_timed_out(struct request *req, bool reserved) |
Jens Axboe | 87ee7b1 | 2014-04-24 08:51:47 -0600 | [diff] [blame] | 537 | { |
Christoph Hellwig | 46f92d4 | 2014-09-13 16:40:12 -0700 | [diff] [blame] | 538 | struct blk_mq_ops *ops = req->q->mq_ops; |
| 539 | enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER; |
Jens Axboe | 87ee7b1 | 2014-04-24 08:51:47 -0600 | [diff] [blame] | 540 | |
| 541 | /* |
| 542 | * We know that complete is set at this point. If STARTED isn't set |
| 543 | * anymore, then the request isn't active and the "timeout" should |
| 544 | * just be ignored. This can happen due to the bitflag ordering. |
| 545 | * Timeout first checks if STARTED is set, and if it is, assumes |
| 546 | * the request is active. But if we race with completion, then |
| 547 | * we both flags will get cleared. So check here again, and ignore |
| 548 | * a timeout event with a request that isn't active. |
| 549 | */ |
Christoph Hellwig | 46f92d4 | 2014-09-13 16:40:12 -0700 | [diff] [blame] | 550 | if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags)) |
| 551 | return; |
Jens Axboe | 87ee7b1 | 2014-04-24 08:51:47 -0600 | [diff] [blame] | 552 | |
Christoph Hellwig | 46f92d4 | 2014-09-13 16:40:12 -0700 | [diff] [blame] | 553 | if (ops->timeout) |
Christoph Hellwig | 0152fb6 | 2014-09-13 16:40:13 -0700 | [diff] [blame] | 554 | ret = ops->timeout(req, reserved); |
Jens Axboe | 87ee7b1 | 2014-04-24 08:51:47 -0600 | [diff] [blame] | 555 | |
Christoph Hellwig | 46f92d4 | 2014-09-13 16:40:12 -0700 | [diff] [blame] | 556 | switch (ret) { |
| 557 | case BLK_EH_HANDLED: |
| 558 | __blk_mq_complete_request(req); |
| 559 | break; |
| 560 | case BLK_EH_RESET_TIMER: |
| 561 | blk_add_timer(req); |
| 562 | blk_clear_rq_complete(req); |
| 563 | break; |
| 564 | case BLK_EH_NOT_HANDLED: |
| 565 | break; |
| 566 | default: |
| 567 | printk(KERN_ERR "block: bad eh return: %d\n", ret); |
| 568 | break; |
| 569 | } |
Jens Axboe | 87ee7b1 | 2014-04-24 08:51:47 -0600 | [diff] [blame] | 570 | } |
Christoph Hellwig | 81481eb | 2014-09-13 16:40:11 -0700 | [diff] [blame] | 571 | |
Christoph Hellwig | 81481eb | 2014-09-13 16:40:11 -0700 | [diff] [blame] | 572 | static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx, |
| 573 | struct request *rq, void *priv, bool reserved) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 574 | { |
Christoph Hellwig | 81481eb | 2014-09-13 16:40:11 -0700 | [diff] [blame] | 575 | struct blk_mq_timeout_data *data = priv; |
| 576 | |
Christoph Hellwig | 46f92d4 | 2014-09-13 16:40:12 -0700 | [diff] [blame] | 577 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) |
| 578 | return; |
| 579 | |
| 580 | if (time_after_eq(jiffies, rq->deadline)) { |
| 581 | if (!blk_mark_rq_complete(rq)) |
Christoph Hellwig | 0152fb6 | 2014-09-13 16:40:13 -0700 | [diff] [blame] | 582 | blk_mq_rq_timed_out(rq, reserved); |
Christoph Hellwig | 46f92d4 | 2014-09-13 16:40:12 -0700 | [diff] [blame] | 583 | } else if (!data->next_set || time_after(data->next, rq->deadline)) { |
| 584 | data->next = rq->deadline; |
| 585 | data->next_set = 1; |
| 586 | } |
Christoph Hellwig | 81481eb | 2014-09-13 16:40:11 -0700 | [diff] [blame] | 587 | } |
| 588 | |
| 589 | static void blk_mq_rq_timer(unsigned long priv) |
| 590 | { |
| 591 | struct request_queue *q = (struct request_queue *)priv; |
| 592 | struct blk_mq_timeout_data data = { |
| 593 | .next = 0, |
| 594 | .next_set = 0, |
| 595 | }; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 596 | struct blk_mq_hw_ctx *hctx; |
Christoph Hellwig | 81481eb | 2014-09-13 16:40:11 -0700 | [diff] [blame] | 597 | int i; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 598 | |
Jens Axboe | 484b406 | 2014-05-21 14:01:15 -0600 | [diff] [blame] | 599 | queue_for_each_hw_ctx(q, hctx, i) { |
| 600 | /* |
| 601 | * If not software queues are currently mapped to this |
| 602 | * hardware queue, there's nothing to check |
| 603 | */ |
| 604 | if (!hctx->nr_ctx || !hctx->tags) |
| 605 | continue; |
| 606 | |
Christoph Hellwig | 81481eb | 2014-09-13 16:40:11 -0700 | [diff] [blame] | 607 | blk_mq_tag_busy_iter(hctx, blk_mq_check_expired, &data); |
Jens Axboe | 484b406 | 2014-05-21 14:01:15 -0600 | [diff] [blame] | 608 | } |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 609 | |
Christoph Hellwig | 81481eb | 2014-09-13 16:40:11 -0700 | [diff] [blame] | 610 | if (data.next_set) { |
| 611 | data.next = blk_rq_timeout(round_jiffies_up(data.next)); |
| 612 | mod_timer(&q->timeout, data.next); |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 613 | } else { |
| 614 | queue_for_each_hw_ctx(q, hctx, i) |
| 615 | blk_mq_tag_idle(hctx); |
| 616 | } |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 617 | } |
| 618 | |
| 619 | /* |
| 620 | * Reverse check our software queue for entries that we could potentially |
| 621 | * merge with. Currently includes a hand-wavy stop count of 8, to not spend |
| 622 | * too much time checking for merges. |
| 623 | */ |
| 624 | static bool blk_mq_attempt_merge(struct request_queue *q, |
| 625 | struct blk_mq_ctx *ctx, struct bio *bio) |
| 626 | { |
| 627 | struct request *rq; |
| 628 | int checked = 8; |
| 629 | |
| 630 | list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) { |
| 631 | int el_ret; |
| 632 | |
| 633 | if (!checked--) |
| 634 | break; |
| 635 | |
| 636 | if (!blk_rq_merge_ok(rq, bio)) |
| 637 | continue; |
| 638 | |
| 639 | el_ret = blk_try_merge(rq, bio); |
| 640 | if (el_ret == ELEVATOR_BACK_MERGE) { |
| 641 | if (bio_attempt_back_merge(q, rq, bio)) { |
| 642 | ctx->rq_merged++; |
| 643 | return true; |
| 644 | } |
| 645 | break; |
| 646 | } else if (el_ret == ELEVATOR_FRONT_MERGE) { |
| 647 | if (bio_attempt_front_merge(q, rq, bio)) { |
| 648 | ctx->rq_merged++; |
| 649 | return true; |
| 650 | } |
| 651 | break; |
| 652 | } |
| 653 | } |
| 654 | |
| 655 | return false; |
| 656 | } |
| 657 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 658 | /* |
Jens Axboe | 1429d7c | 2014-05-19 09:23:55 -0600 | [diff] [blame] | 659 | * Process software queues that have been marked busy, splicing them |
| 660 | * to the for-dispatch |
| 661 | */ |
| 662 | static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list) |
| 663 | { |
| 664 | struct blk_mq_ctx *ctx; |
| 665 | int i; |
| 666 | |
| 667 | for (i = 0; i < hctx->ctx_map.map_size; i++) { |
| 668 | struct blk_align_bitmap *bm = &hctx->ctx_map.map[i]; |
| 669 | unsigned int off, bit; |
| 670 | |
| 671 | if (!bm->word) |
| 672 | continue; |
| 673 | |
| 674 | bit = 0; |
| 675 | off = i * hctx->ctx_map.bits_per_word; |
| 676 | do { |
| 677 | bit = find_next_bit(&bm->word, bm->depth, bit); |
| 678 | if (bit >= bm->depth) |
| 679 | break; |
| 680 | |
| 681 | ctx = hctx->ctxs[bit + off]; |
| 682 | clear_bit(bit, &bm->word); |
| 683 | spin_lock(&ctx->lock); |
| 684 | list_splice_tail_init(&ctx->rq_list, list); |
| 685 | spin_unlock(&ctx->lock); |
| 686 | |
| 687 | bit++; |
| 688 | } while (1); |
| 689 | } |
| 690 | } |
| 691 | |
| 692 | /* |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 693 | * Run this hardware queue, pulling any software queues mapped to it in. |
| 694 | * Note that this function currently has various problems around ordering |
| 695 | * of IO. In particular, we'd like FIFO behaviour on handling existing |
| 696 | * items on the hctx->dispatch list. Ignore that for now. |
| 697 | */ |
| 698 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx) |
| 699 | { |
| 700 | struct request_queue *q = hctx->queue; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 701 | struct request *rq; |
| 702 | LIST_HEAD(rq_list); |
Jens Axboe | 1429d7c | 2014-05-19 09:23:55 -0600 | [diff] [blame] | 703 | int queued; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 704 | |
Jens Axboe | fd1270d | 2014-04-16 09:23:48 -0600 | [diff] [blame] | 705 | WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask)); |
Jens Axboe | e4043dc | 2014-04-09 10:18:23 -0600 | [diff] [blame] | 706 | |
Jens Axboe | 5d12f90 | 2014-03-19 15:25:02 -0600 | [diff] [blame] | 707 | if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state))) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 708 | return; |
| 709 | |
| 710 | hctx->run++; |
| 711 | |
| 712 | /* |
| 713 | * Touch any software queue that has pending entries. |
| 714 | */ |
Jens Axboe | 1429d7c | 2014-05-19 09:23:55 -0600 | [diff] [blame] | 715 | flush_busy_ctxs(hctx, &rq_list); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 716 | |
| 717 | /* |
| 718 | * If we have previous entries on our dispatch list, grab them |
| 719 | * and stuff them at the front for more fair dispatch. |
| 720 | */ |
| 721 | if (!list_empty_careful(&hctx->dispatch)) { |
| 722 | spin_lock(&hctx->lock); |
| 723 | if (!list_empty(&hctx->dispatch)) |
| 724 | list_splice_init(&hctx->dispatch, &rq_list); |
| 725 | spin_unlock(&hctx->lock); |
| 726 | } |
| 727 | |
| 728 | /* |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 729 | * Now process all the entries, sending them to the driver. |
| 730 | */ |
Jens Axboe | 1429d7c | 2014-05-19 09:23:55 -0600 | [diff] [blame] | 731 | queued = 0; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 732 | while (!list_empty(&rq_list)) { |
| 733 | int ret; |
| 734 | |
| 735 | rq = list_first_entry(&rq_list, struct request, queuelist); |
| 736 | list_del_init(&rq->queuelist); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 737 | |
Christoph Hellwig | bf57229 | 2014-09-13 16:40:08 -0700 | [diff] [blame] | 738 | ret = q->mq_ops->queue_rq(hctx, rq, list_empty(&rq_list)); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 739 | switch (ret) { |
| 740 | case BLK_MQ_RQ_QUEUE_OK: |
| 741 | queued++; |
| 742 | continue; |
| 743 | case BLK_MQ_RQ_QUEUE_BUSY: |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 744 | list_add(&rq->queuelist, &rq_list); |
Christoph Hellwig | ed0791b | 2014-04-16 09:44:57 +0200 | [diff] [blame] | 745 | __blk_mq_requeue_request(rq); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 746 | break; |
| 747 | default: |
| 748 | pr_err("blk-mq: bad return on queue: %d\n", ret); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 749 | case BLK_MQ_RQ_QUEUE_ERROR: |
Christoph Hellwig | 1e93b8c | 2014-02-11 08:27:13 -0800 | [diff] [blame] | 750 | rq->errors = -EIO; |
Christoph Hellwig | c8a446a | 2014-09-13 16:40:10 -0700 | [diff] [blame] | 751 | blk_mq_end_request(rq, rq->errors); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 752 | break; |
| 753 | } |
| 754 | |
| 755 | if (ret == BLK_MQ_RQ_QUEUE_BUSY) |
| 756 | break; |
| 757 | } |
| 758 | |
| 759 | if (!queued) |
| 760 | hctx->dispatched[0]++; |
| 761 | else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1))) |
| 762 | hctx->dispatched[ilog2(queued) + 1]++; |
| 763 | |
| 764 | /* |
| 765 | * Any items that need requeuing? Stuff them into hctx->dispatch, |
| 766 | * that is where we will continue on next queue run. |
| 767 | */ |
| 768 | if (!list_empty(&rq_list)) { |
| 769 | spin_lock(&hctx->lock); |
| 770 | list_splice(&rq_list, &hctx->dispatch); |
| 771 | spin_unlock(&hctx->lock); |
| 772 | } |
| 773 | } |
| 774 | |
Jens Axboe | 506e931 | 2014-05-07 10:26:44 -0600 | [diff] [blame] | 775 | /* |
| 776 | * It'd be great if the workqueue API had a way to pass |
| 777 | * in a mask and had some smarts for more clever placement. |
| 778 | * For now we just round-robin here, switching for every |
| 779 | * BLK_MQ_CPU_WORK_BATCH queued items. |
| 780 | */ |
| 781 | static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx) |
| 782 | { |
| 783 | int cpu = hctx->next_cpu; |
| 784 | |
| 785 | if (--hctx->next_cpu_batch <= 0) { |
| 786 | int next_cpu; |
| 787 | |
| 788 | next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask); |
| 789 | if (next_cpu >= nr_cpu_ids) |
| 790 | next_cpu = cpumask_first(hctx->cpumask); |
| 791 | |
| 792 | hctx->next_cpu = next_cpu; |
| 793 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; |
| 794 | } |
| 795 | |
| 796 | return cpu; |
| 797 | } |
| 798 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 799 | void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
| 800 | { |
Jens Axboe | 5d12f90 | 2014-03-19 15:25:02 -0600 | [diff] [blame] | 801 | if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state))) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 802 | return; |
| 803 | |
Jens Axboe | e4043dc | 2014-04-09 10:18:23 -0600 | [diff] [blame] | 804 | if (!async && cpumask_test_cpu(smp_processor_id(), hctx->cpumask)) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 805 | __blk_mq_run_hw_queue(hctx); |
Jens Axboe | e4043dc | 2014-04-09 10:18:23 -0600 | [diff] [blame] | 806 | else if (hctx->queue->nr_hw_queues == 1) |
Christoph Hellwig | 70f4db6 | 2014-04-16 10:48:08 -0600 | [diff] [blame] | 807 | kblockd_schedule_delayed_work(&hctx->run_work, 0); |
Jens Axboe | e4043dc | 2014-04-09 10:18:23 -0600 | [diff] [blame] | 808 | else { |
| 809 | unsigned int cpu; |
| 810 | |
Jens Axboe | 506e931 | 2014-05-07 10:26:44 -0600 | [diff] [blame] | 811 | cpu = blk_mq_hctx_next_cpu(hctx); |
Christoph Hellwig | 70f4db6 | 2014-04-16 10:48:08 -0600 | [diff] [blame] | 812 | kblockd_schedule_delayed_work_on(cpu, &hctx->run_work, 0); |
Jens Axboe | e4043dc | 2014-04-09 10:18:23 -0600 | [diff] [blame] | 813 | } |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 814 | } |
| 815 | |
| 816 | void blk_mq_run_queues(struct request_queue *q, bool async) |
| 817 | { |
| 818 | struct blk_mq_hw_ctx *hctx; |
| 819 | int i; |
| 820 | |
| 821 | queue_for_each_hw_ctx(q, hctx, i) { |
| 822 | if ((!blk_mq_hctx_has_pending(hctx) && |
| 823 | list_empty_careful(&hctx->dispatch)) || |
Jens Axboe | 5d12f90 | 2014-03-19 15:25:02 -0600 | [diff] [blame] | 824 | test_bit(BLK_MQ_S_STOPPED, &hctx->state)) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 825 | continue; |
| 826 | |
Jens Axboe | e4043dc | 2014-04-09 10:18:23 -0600 | [diff] [blame] | 827 | preempt_disable(); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 828 | blk_mq_run_hw_queue(hctx, async); |
Jens Axboe | e4043dc | 2014-04-09 10:18:23 -0600 | [diff] [blame] | 829 | preempt_enable(); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 830 | } |
| 831 | } |
| 832 | EXPORT_SYMBOL(blk_mq_run_queues); |
| 833 | |
| 834 | void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) |
| 835 | { |
Christoph Hellwig | 70f4db6 | 2014-04-16 10:48:08 -0600 | [diff] [blame] | 836 | cancel_delayed_work(&hctx->run_work); |
| 837 | cancel_delayed_work(&hctx->delay_work); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 838 | set_bit(BLK_MQ_S_STOPPED, &hctx->state); |
| 839 | } |
| 840 | EXPORT_SYMBOL(blk_mq_stop_hw_queue); |
| 841 | |
Christoph Hellwig | 280d45f | 2013-10-25 14:45:58 +0100 | [diff] [blame] | 842 | void blk_mq_stop_hw_queues(struct request_queue *q) |
| 843 | { |
| 844 | struct blk_mq_hw_ctx *hctx; |
| 845 | int i; |
| 846 | |
| 847 | queue_for_each_hw_ctx(q, hctx, i) |
| 848 | blk_mq_stop_hw_queue(hctx); |
| 849 | } |
| 850 | EXPORT_SYMBOL(blk_mq_stop_hw_queues); |
| 851 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 852 | void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) |
| 853 | { |
| 854 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); |
Jens Axboe | e4043dc | 2014-04-09 10:18:23 -0600 | [diff] [blame] | 855 | |
| 856 | preempt_disable(); |
Jens Axboe | 0ffbce8 | 2014-06-25 08:22:34 -0600 | [diff] [blame] | 857 | blk_mq_run_hw_queue(hctx, false); |
Jens Axboe | e4043dc | 2014-04-09 10:18:23 -0600 | [diff] [blame] | 858 | preempt_enable(); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 859 | } |
| 860 | EXPORT_SYMBOL(blk_mq_start_hw_queue); |
| 861 | |
Christoph Hellwig | 2f26855 | 2014-04-16 09:44:56 +0200 | [diff] [blame] | 862 | void blk_mq_start_hw_queues(struct request_queue *q) |
| 863 | { |
| 864 | struct blk_mq_hw_ctx *hctx; |
| 865 | int i; |
| 866 | |
| 867 | queue_for_each_hw_ctx(q, hctx, i) |
| 868 | blk_mq_start_hw_queue(hctx); |
| 869 | } |
| 870 | EXPORT_SYMBOL(blk_mq_start_hw_queues); |
| 871 | |
| 872 | |
Christoph Hellwig | 1b4a325 | 2014-04-16 09:44:54 +0200 | [diff] [blame] | 873 | 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] | 874 | { |
| 875 | struct blk_mq_hw_ctx *hctx; |
| 876 | int i; |
| 877 | |
| 878 | queue_for_each_hw_ctx(q, hctx, i) { |
| 879 | if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state)) |
| 880 | continue; |
| 881 | |
| 882 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); |
Jens Axboe | e4043dc | 2014-04-09 10:18:23 -0600 | [diff] [blame] | 883 | preempt_disable(); |
Christoph Hellwig | 1b4a325 | 2014-04-16 09:44:54 +0200 | [diff] [blame] | 884 | blk_mq_run_hw_queue(hctx, async); |
Jens Axboe | e4043dc | 2014-04-09 10:18:23 -0600 | [diff] [blame] | 885 | preempt_enable(); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 886 | } |
| 887 | } |
| 888 | EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); |
| 889 | |
Christoph Hellwig | 70f4db6 | 2014-04-16 10:48:08 -0600 | [diff] [blame] | 890 | static void blk_mq_run_work_fn(struct work_struct *work) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 891 | { |
| 892 | struct blk_mq_hw_ctx *hctx; |
| 893 | |
Christoph Hellwig | 70f4db6 | 2014-04-16 10:48:08 -0600 | [diff] [blame] | 894 | hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work); |
Jens Axboe | e4043dc | 2014-04-09 10:18:23 -0600 | [diff] [blame] | 895 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 896 | __blk_mq_run_hw_queue(hctx); |
| 897 | } |
| 898 | |
Christoph Hellwig | 70f4db6 | 2014-04-16 10:48:08 -0600 | [diff] [blame] | 899 | static void blk_mq_delay_work_fn(struct work_struct *work) |
| 900 | { |
| 901 | struct blk_mq_hw_ctx *hctx; |
| 902 | |
| 903 | hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work); |
| 904 | |
| 905 | if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state)) |
| 906 | __blk_mq_run_hw_queue(hctx); |
| 907 | } |
| 908 | |
| 909 | void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) |
| 910 | { |
| 911 | unsigned long tmo = msecs_to_jiffies(msecs); |
| 912 | |
| 913 | if (hctx->queue->nr_hw_queues == 1) |
| 914 | kblockd_schedule_delayed_work(&hctx->delay_work, tmo); |
| 915 | else { |
| 916 | unsigned int cpu; |
| 917 | |
Jens Axboe | 506e931 | 2014-05-07 10:26:44 -0600 | [diff] [blame] | 918 | cpu = blk_mq_hctx_next_cpu(hctx); |
Christoph Hellwig | 70f4db6 | 2014-04-16 10:48:08 -0600 | [diff] [blame] | 919 | kblockd_schedule_delayed_work_on(cpu, &hctx->delay_work, tmo); |
| 920 | } |
| 921 | } |
| 922 | EXPORT_SYMBOL(blk_mq_delay_queue); |
| 923 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 924 | static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, |
Christoph Hellwig | 72a0a36 | 2014-02-07 10:22:36 -0800 | [diff] [blame] | 925 | struct request *rq, bool at_head) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 926 | { |
| 927 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
| 928 | |
Jens Axboe | 01b983c | 2013-11-19 18:59:10 -0700 | [diff] [blame] | 929 | trace_block_rq_insert(hctx->queue, rq); |
| 930 | |
Christoph Hellwig | 72a0a36 | 2014-02-07 10:22:36 -0800 | [diff] [blame] | 931 | if (at_head) |
| 932 | list_add(&rq->queuelist, &ctx->rq_list); |
| 933 | else |
| 934 | list_add_tail(&rq->queuelist, &ctx->rq_list); |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 935 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 936 | blk_mq_hctx_mark_pending(hctx, ctx); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 937 | } |
| 938 | |
Christoph Hellwig | eeabc85 | 2014-03-21 08:57:37 -0600 | [diff] [blame] | 939 | void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue, |
| 940 | bool async) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 941 | { |
| 942 | struct request_queue *q = rq->q; |
| 943 | struct blk_mq_hw_ctx *hctx; |
Christoph Hellwig | eeabc85 | 2014-03-21 08:57:37 -0600 | [diff] [blame] | 944 | struct blk_mq_ctx *ctx = rq->mq_ctx, *current_ctx; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 945 | |
| 946 | current_ctx = blk_mq_get_ctx(q); |
Christoph Hellwig | eeabc85 | 2014-03-21 08:57:37 -0600 | [diff] [blame] | 947 | if (!cpu_online(ctx->cpu)) |
| 948 | rq->mq_ctx = ctx = current_ctx; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 949 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 950 | hctx = q->mq_ops->map_queue(q, ctx->cpu); |
| 951 | |
Christoph Hellwig | a57a178 | 2014-09-16 14:44:07 -0700 | [diff] [blame] | 952 | spin_lock(&ctx->lock); |
| 953 | __blk_mq_insert_request(hctx, rq, at_head); |
| 954 | spin_unlock(&ctx->lock); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 955 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 956 | if (run_queue) |
| 957 | blk_mq_run_hw_queue(hctx, async); |
Jens Axboe | e4043dc | 2014-04-09 10:18:23 -0600 | [diff] [blame] | 958 | |
| 959 | blk_mq_put_ctx(current_ctx); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 960 | } |
| 961 | |
| 962 | static void blk_mq_insert_requests(struct request_queue *q, |
| 963 | struct blk_mq_ctx *ctx, |
| 964 | struct list_head *list, |
| 965 | int depth, |
| 966 | bool from_schedule) |
| 967 | |
| 968 | { |
| 969 | struct blk_mq_hw_ctx *hctx; |
| 970 | struct blk_mq_ctx *current_ctx; |
| 971 | |
| 972 | trace_block_unplug(q, depth, !from_schedule); |
| 973 | |
| 974 | current_ctx = blk_mq_get_ctx(q); |
| 975 | |
| 976 | if (!cpu_online(ctx->cpu)) |
| 977 | ctx = current_ctx; |
| 978 | hctx = q->mq_ops->map_queue(q, ctx->cpu); |
| 979 | |
| 980 | /* |
| 981 | * preemption doesn't flush plug list, so it's possible ctx->cpu is |
| 982 | * offline now |
| 983 | */ |
| 984 | spin_lock(&ctx->lock); |
| 985 | while (!list_empty(list)) { |
| 986 | struct request *rq; |
| 987 | |
| 988 | rq = list_first_entry(list, struct request, queuelist); |
| 989 | list_del_init(&rq->queuelist); |
| 990 | rq->mq_ctx = ctx; |
Christoph Hellwig | 72a0a36 | 2014-02-07 10:22:36 -0800 | [diff] [blame] | 991 | __blk_mq_insert_request(hctx, rq, false); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 992 | } |
| 993 | spin_unlock(&ctx->lock); |
| 994 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 995 | blk_mq_run_hw_queue(hctx, from_schedule); |
Jens Axboe | e4043dc | 2014-04-09 10:18:23 -0600 | [diff] [blame] | 996 | blk_mq_put_ctx(current_ctx); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 997 | } |
| 998 | |
| 999 | static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b) |
| 1000 | { |
| 1001 | struct request *rqa = container_of(a, struct request, queuelist); |
| 1002 | struct request *rqb = container_of(b, struct request, queuelist); |
| 1003 | |
| 1004 | return !(rqa->mq_ctx < rqb->mq_ctx || |
| 1005 | (rqa->mq_ctx == rqb->mq_ctx && |
| 1006 | blk_rq_pos(rqa) < blk_rq_pos(rqb))); |
| 1007 | } |
| 1008 | |
| 1009 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) |
| 1010 | { |
| 1011 | struct blk_mq_ctx *this_ctx; |
| 1012 | struct request_queue *this_q; |
| 1013 | struct request *rq; |
| 1014 | LIST_HEAD(list); |
| 1015 | LIST_HEAD(ctx_list); |
| 1016 | unsigned int depth; |
| 1017 | |
| 1018 | list_splice_init(&plug->mq_list, &list); |
| 1019 | |
| 1020 | list_sort(NULL, &list, plug_ctx_cmp); |
| 1021 | |
| 1022 | this_q = NULL; |
| 1023 | this_ctx = NULL; |
| 1024 | depth = 0; |
| 1025 | |
| 1026 | while (!list_empty(&list)) { |
| 1027 | rq = list_entry_rq(list.next); |
| 1028 | list_del_init(&rq->queuelist); |
| 1029 | BUG_ON(!rq->q); |
| 1030 | if (rq->mq_ctx != this_ctx) { |
| 1031 | if (this_ctx) { |
| 1032 | blk_mq_insert_requests(this_q, this_ctx, |
| 1033 | &ctx_list, depth, |
| 1034 | from_schedule); |
| 1035 | } |
| 1036 | |
| 1037 | this_ctx = rq->mq_ctx; |
| 1038 | this_q = rq->q; |
| 1039 | depth = 0; |
| 1040 | } |
| 1041 | |
| 1042 | depth++; |
| 1043 | list_add_tail(&rq->queuelist, &ctx_list); |
| 1044 | } |
| 1045 | |
| 1046 | /* |
| 1047 | * If 'this_ctx' is set, we know we have entries to complete |
| 1048 | * on 'ctx_list'. Do those. |
| 1049 | */ |
| 1050 | if (this_ctx) { |
| 1051 | blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth, |
| 1052 | from_schedule); |
| 1053 | } |
| 1054 | } |
| 1055 | |
| 1056 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio) |
| 1057 | { |
| 1058 | init_request_from_bio(rq, bio); |
Jens Axboe | 4b57052 | 2014-05-29 11:00:11 -0600 | [diff] [blame] | 1059 | |
Jens Axboe | 3ee3237 | 2014-06-09 09:36:53 -0600 | [diff] [blame] | 1060 | if (blk_do_io_stat(rq)) |
Jens Axboe | 4b57052 | 2014-05-29 11:00:11 -0600 | [diff] [blame] | 1061 | blk_account_io_start(rq, 1); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1062 | } |
| 1063 | |
Jens Axboe | 274a584 | 2014-08-15 12:44:08 -0600 | [diff] [blame] | 1064 | static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx) |
| 1065 | { |
| 1066 | return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) && |
| 1067 | !blk_queue_nomerges(hctx->queue); |
| 1068 | } |
| 1069 | |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1070 | static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx, |
| 1071 | struct blk_mq_ctx *ctx, |
| 1072 | struct request *rq, struct bio *bio) |
| 1073 | { |
Jens Axboe | 274a584 | 2014-08-15 12:44:08 -0600 | [diff] [blame] | 1074 | if (!hctx_allow_merges(hctx)) { |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1075 | blk_mq_bio_to_request(rq, bio); |
| 1076 | spin_lock(&ctx->lock); |
| 1077 | insert_rq: |
| 1078 | __blk_mq_insert_request(hctx, rq, false); |
| 1079 | spin_unlock(&ctx->lock); |
| 1080 | return false; |
| 1081 | } else { |
Jens Axboe | 274a584 | 2014-08-15 12:44:08 -0600 | [diff] [blame] | 1082 | struct request_queue *q = hctx->queue; |
| 1083 | |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1084 | spin_lock(&ctx->lock); |
| 1085 | if (!blk_mq_attempt_merge(q, ctx, bio)) { |
| 1086 | blk_mq_bio_to_request(rq, bio); |
| 1087 | goto insert_rq; |
| 1088 | } |
| 1089 | |
| 1090 | spin_unlock(&ctx->lock); |
| 1091 | __blk_mq_free_request(hctx, ctx, rq); |
| 1092 | return true; |
| 1093 | } |
| 1094 | } |
| 1095 | |
| 1096 | struct blk_map_ctx { |
| 1097 | struct blk_mq_hw_ctx *hctx; |
| 1098 | struct blk_mq_ctx *ctx; |
| 1099 | }; |
| 1100 | |
| 1101 | static struct request *blk_mq_map_request(struct request_queue *q, |
| 1102 | struct bio *bio, |
| 1103 | struct blk_map_ctx *data) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1104 | { |
| 1105 | struct blk_mq_hw_ctx *hctx; |
| 1106 | struct blk_mq_ctx *ctx; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1107 | struct request *rq; |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1108 | int rw = bio_data_dir(bio); |
Ming Lei | cb96a42 | 2014-06-01 00:43:37 +0800 | [diff] [blame] | 1109 | struct blk_mq_alloc_data alloc_data; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1110 | |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1111 | if (unlikely(blk_mq_queue_enter(q))) { |
Nicholas Bellinger | 14ec77f | 2014-02-07 13:45:39 -0700 | [diff] [blame] | 1112 | bio_endio(bio, -EIO); |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1113 | return NULL; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1114 | } |
| 1115 | |
| 1116 | ctx = blk_mq_get_ctx(q); |
| 1117 | hctx = q->mq_ops->map_queue(q, ctx->cpu); |
| 1118 | |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1119 | if (rw_is_sync(bio->bi_rw)) |
Shaohua Li | 27fbf4e | 2014-02-19 20:20:21 +0800 | [diff] [blame] | 1120 | rw |= REQ_SYNC; |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1121 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1122 | trace_block_getrq(q, bio, rw); |
Ming Lei | cb96a42 | 2014-06-01 00:43:37 +0800 | [diff] [blame] | 1123 | blk_mq_set_alloc_data(&alloc_data, q, GFP_ATOMIC, false, ctx, |
| 1124 | hctx); |
| 1125 | rq = __blk_mq_alloc_request(&alloc_data, rw); |
Christoph Hellwig | 5dee857 | 2014-05-27 20:59:47 +0200 | [diff] [blame] | 1126 | if (unlikely(!rq)) { |
Christoph Hellwig | 793597a | 2014-05-27 20:59:49 +0200 | [diff] [blame] | 1127 | __blk_mq_run_hw_queue(hctx); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1128 | blk_mq_put_ctx(ctx); |
| 1129 | trace_block_sleeprq(q, bio, rw); |
Christoph Hellwig | 793597a | 2014-05-27 20:59:49 +0200 | [diff] [blame] | 1130 | |
| 1131 | ctx = blk_mq_get_ctx(q); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1132 | hctx = q->mq_ops->map_queue(q, ctx->cpu); |
Ming Lei | cb96a42 | 2014-06-01 00:43:37 +0800 | [diff] [blame] | 1133 | blk_mq_set_alloc_data(&alloc_data, q, |
| 1134 | __GFP_WAIT|GFP_ATOMIC, false, ctx, hctx); |
| 1135 | rq = __blk_mq_alloc_request(&alloc_data, rw); |
| 1136 | ctx = alloc_data.ctx; |
| 1137 | hctx = alloc_data.hctx; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1138 | } |
| 1139 | |
| 1140 | hctx->queued++; |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1141 | data->hctx = hctx; |
| 1142 | data->ctx = ctx; |
| 1143 | return rq; |
| 1144 | } |
| 1145 | |
| 1146 | /* |
| 1147 | * Multiple hardware queue variant. This will not use per-process plugs, |
| 1148 | * but will attempt to bypass the hctx queueing if we can go straight to |
| 1149 | * hardware for SYNC IO. |
| 1150 | */ |
| 1151 | static void blk_mq_make_request(struct request_queue *q, struct bio *bio) |
| 1152 | { |
| 1153 | const int is_sync = rw_is_sync(bio->bi_rw); |
| 1154 | const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA); |
| 1155 | struct blk_map_ctx data; |
| 1156 | struct request *rq; |
| 1157 | |
| 1158 | blk_queue_bounce(q, &bio); |
| 1159 | |
| 1160 | if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) { |
| 1161 | bio_endio(bio, -EIO); |
| 1162 | return; |
| 1163 | } |
| 1164 | |
| 1165 | rq = blk_mq_map_request(q, bio, &data); |
| 1166 | if (unlikely(!rq)) |
| 1167 | return; |
| 1168 | |
| 1169 | if (unlikely(is_flush_fua)) { |
| 1170 | blk_mq_bio_to_request(rq, bio); |
| 1171 | blk_insert_flush(rq); |
| 1172 | goto run_queue; |
| 1173 | } |
| 1174 | |
| 1175 | if (is_sync) { |
| 1176 | int ret; |
| 1177 | |
| 1178 | blk_mq_bio_to_request(rq, bio); |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1179 | |
| 1180 | /* |
| 1181 | * For OK queue, we are done. For error, kill it. Any other |
| 1182 | * error (busy), just add it to our list as we previously |
| 1183 | * would have done |
| 1184 | */ |
Christoph Hellwig | bf57229 | 2014-09-13 16:40:08 -0700 | [diff] [blame] | 1185 | ret = q->mq_ops->queue_rq(data.hctx, rq, true); |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1186 | if (ret == BLK_MQ_RQ_QUEUE_OK) |
| 1187 | goto done; |
| 1188 | else { |
| 1189 | __blk_mq_requeue_request(rq); |
| 1190 | |
| 1191 | if (ret == BLK_MQ_RQ_QUEUE_ERROR) { |
| 1192 | rq->errors = -EIO; |
Christoph Hellwig | c8a446a | 2014-09-13 16:40:10 -0700 | [diff] [blame] | 1193 | blk_mq_end_request(rq, rq->errors); |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1194 | goto done; |
| 1195 | } |
| 1196 | } |
| 1197 | } |
| 1198 | |
| 1199 | if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) { |
| 1200 | /* |
| 1201 | * For a SYNC request, send it to the hardware immediately. For |
| 1202 | * an ASYNC request, just ensure that we run it later on. The |
| 1203 | * latter allows for merging opportunities and more efficient |
| 1204 | * dispatching. |
| 1205 | */ |
| 1206 | run_queue: |
| 1207 | blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua); |
| 1208 | } |
| 1209 | done: |
| 1210 | blk_mq_put_ctx(data.ctx); |
| 1211 | } |
| 1212 | |
| 1213 | /* |
| 1214 | * Single hardware queue variant. This will attempt to use any per-process |
| 1215 | * plug for merging and IO deferral. |
| 1216 | */ |
| 1217 | static void blk_sq_make_request(struct request_queue *q, struct bio *bio) |
| 1218 | { |
| 1219 | const int is_sync = rw_is_sync(bio->bi_rw); |
| 1220 | const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA); |
| 1221 | unsigned int use_plug, request_count = 0; |
| 1222 | struct blk_map_ctx data; |
| 1223 | struct request *rq; |
| 1224 | |
| 1225 | /* |
| 1226 | * If we have multiple hardware queues, just go directly to |
| 1227 | * one of those for sync IO. |
| 1228 | */ |
| 1229 | use_plug = !is_flush_fua && !is_sync; |
| 1230 | |
| 1231 | blk_queue_bounce(q, &bio); |
| 1232 | |
| 1233 | if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) { |
| 1234 | bio_endio(bio, -EIO); |
| 1235 | return; |
| 1236 | } |
| 1237 | |
| 1238 | if (use_plug && !blk_queue_nomerges(q) && |
| 1239 | blk_attempt_plug_merge(q, bio, &request_count)) |
| 1240 | return; |
| 1241 | |
| 1242 | rq = blk_mq_map_request(q, bio, &data); |
Jens Axboe | ff87bce | 2014-06-03 11:59:49 -0600 | [diff] [blame] | 1243 | if (unlikely(!rq)) |
| 1244 | return; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1245 | |
| 1246 | if (unlikely(is_flush_fua)) { |
| 1247 | blk_mq_bio_to_request(rq, bio); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1248 | blk_insert_flush(rq); |
| 1249 | goto run_queue; |
| 1250 | } |
| 1251 | |
| 1252 | /* |
| 1253 | * A task plug currently exists. Since this is completely lockless, |
| 1254 | * utilize that to temporarily store requests until the task is |
| 1255 | * either done or scheduled away. |
| 1256 | */ |
| 1257 | if (use_plug) { |
| 1258 | struct blk_plug *plug = current->plug; |
| 1259 | |
| 1260 | if (plug) { |
| 1261 | blk_mq_bio_to_request(rq, bio); |
Shaohua Li | 92f399c | 2013-10-29 12:01:03 -0600 | [diff] [blame] | 1262 | if (list_empty(&plug->mq_list)) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1263 | trace_block_plug(q); |
| 1264 | else if (request_count >= BLK_MAX_REQUEST_COUNT) { |
| 1265 | blk_flush_plug_list(plug, false); |
| 1266 | trace_block_plug(q); |
| 1267 | } |
| 1268 | list_add_tail(&rq->queuelist, &plug->mq_list); |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1269 | blk_mq_put_ctx(data.ctx); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1270 | return; |
| 1271 | } |
| 1272 | } |
| 1273 | |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1274 | if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) { |
| 1275 | /* |
| 1276 | * For a SYNC request, send it to the hardware immediately. For |
| 1277 | * an ASYNC request, just ensure that we run it later on. The |
| 1278 | * latter allows for merging opportunities and more efficient |
| 1279 | * dispatching. |
| 1280 | */ |
| 1281 | run_queue: |
| 1282 | blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1283 | } |
| 1284 | |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1285 | blk_mq_put_ctx(data.ctx); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1286 | } |
| 1287 | |
| 1288 | /* |
| 1289 | * Default mapping to a software queue, since we use one per CPU. |
| 1290 | */ |
| 1291 | struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu) |
| 1292 | { |
| 1293 | return q->queue_hw_ctx[q->mq_map[cpu]]; |
| 1294 | } |
| 1295 | EXPORT_SYMBOL(blk_mq_map_queue); |
| 1296 | |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1297 | static void blk_mq_free_rq_map(struct blk_mq_tag_set *set, |
| 1298 | struct blk_mq_tags *tags, unsigned int hctx_idx) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1299 | { |
| 1300 | struct page *page; |
| 1301 | |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1302 | if (tags->rqs && set->ops->exit_request) { |
Christoph Hellwig | e9b267d | 2014-04-15 13:59:10 -0600 | [diff] [blame] | 1303 | int i; |
| 1304 | |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1305 | for (i = 0; i < tags->nr_tags; i++) { |
| 1306 | if (!tags->rqs[i]) |
Christoph Hellwig | e9b267d | 2014-04-15 13:59:10 -0600 | [diff] [blame] | 1307 | continue; |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1308 | set->ops->exit_request(set->driver_data, tags->rqs[i], |
| 1309 | hctx_idx, i); |
Jens Axboe | a516440 | 2014-09-10 09:02:03 -0600 | [diff] [blame] | 1310 | tags->rqs[i] = NULL; |
Christoph Hellwig | e9b267d | 2014-04-15 13:59:10 -0600 | [diff] [blame] | 1311 | } |
| 1312 | } |
| 1313 | |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1314 | while (!list_empty(&tags->page_list)) { |
| 1315 | page = list_first_entry(&tags->page_list, struct page, lru); |
Dave Hansen | 6753471 | 2014-01-08 20:17:46 -0700 | [diff] [blame] | 1316 | list_del_init(&page->lru); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1317 | __free_pages(page, page->private); |
| 1318 | } |
| 1319 | |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1320 | kfree(tags->rqs); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1321 | |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1322 | blk_mq_free_tags(tags); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1323 | } |
| 1324 | |
| 1325 | static size_t order_to_size(unsigned int order) |
| 1326 | { |
Ming Lei | 4ca0850 | 2014-04-19 18:00:18 +0800 | [diff] [blame] | 1327 | return (size_t)PAGE_SIZE << order; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1328 | } |
| 1329 | |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1330 | static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set, |
| 1331 | unsigned int hctx_idx) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1332 | { |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1333 | struct blk_mq_tags *tags; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1334 | unsigned int i, j, entries_per_page, max_order = 4; |
| 1335 | size_t rq_size, left; |
| 1336 | |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1337 | tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags, |
| 1338 | set->numa_node); |
| 1339 | if (!tags) |
| 1340 | return NULL; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1341 | |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1342 | INIT_LIST_HEAD(&tags->page_list); |
| 1343 | |
Jens Axboe | a516440 | 2014-09-10 09:02:03 -0600 | [diff] [blame] | 1344 | tags->rqs = kzalloc_node(set->queue_depth * sizeof(struct request *), |
| 1345 | GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY, |
| 1346 | set->numa_node); |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1347 | if (!tags->rqs) { |
| 1348 | blk_mq_free_tags(tags); |
| 1349 | return NULL; |
| 1350 | } |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1351 | |
| 1352 | /* |
| 1353 | * rq_size is the size of the request plus driver payload, rounded |
| 1354 | * to the cacheline size |
| 1355 | */ |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1356 | rq_size = round_up(sizeof(struct request) + set->cmd_size, |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1357 | cache_line_size()); |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1358 | left = rq_size * set->queue_depth; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1359 | |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1360 | for (i = 0; i < set->queue_depth; ) { |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1361 | int this_order = max_order; |
| 1362 | struct page *page; |
| 1363 | int to_do; |
| 1364 | void *p; |
| 1365 | |
| 1366 | while (left < order_to_size(this_order - 1) && this_order) |
| 1367 | this_order--; |
| 1368 | |
| 1369 | do { |
Jens Axboe | a516440 | 2014-09-10 09:02:03 -0600 | [diff] [blame] | 1370 | page = alloc_pages_node(set->numa_node, |
| 1371 | GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY, |
| 1372 | this_order); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1373 | if (page) |
| 1374 | break; |
| 1375 | if (!this_order--) |
| 1376 | break; |
| 1377 | if (order_to_size(this_order) < rq_size) |
| 1378 | break; |
| 1379 | } while (1); |
| 1380 | |
| 1381 | if (!page) |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1382 | goto fail; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1383 | |
| 1384 | page->private = this_order; |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1385 | list_add_tail(&page->lru, &tags->page_list); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1386 | |
| 1387 | p = page_address(page); |
| 1388 | entries_per_page = order_to_size(this_order) / rq_size; |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1389 | to_do = min(entries_per_page, set->queue_depth - i); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1390 | left -= to_do * rq_size; |
| 1391 | for (j = 0; j < to_do; j++) { |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1392 | tags->rqs[i] = p; |
David Hildenbrand | 683d0e1 | 2014-09-18 11:04:31 +0200 | [diff] [blame] | 1393 | tags->rqs[i]->atomic_flags = 0; |
| 1394 | tags->rqs[i]->cmd_flags = 0; |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1395 | if (set->ops->init_request) { |
| 1396 | if (set->ops->init_request(set->driver_data, |
| 1397 | tags->rqs[i], hctx_idx, i, |
Jens Axboe | a516440 | 2014-09-10 09:02:03 -0600 | [diff] [blame] | 1398 | set->numa_node)) { |
| 1399 | tags->rqs[i] = NULL; |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1400 | goto fail; |
Jens Axboe | a516440 | 2014-09-10 09:02:03 -0600 | [diff] [blame] | 1401 | } |
Christoph Hellwig | e9b267d | 2014-04-15 13:59:10 -0600 | [diff] [blame] | 1402 | } |
| 1403 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1404 | p += rq_size; |
| 1405 | i++; |
| 1406 | } |
| 1407 | } |
| 1408 | |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1409 | return tags; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1410 | |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1411 | fail: |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1412 | blk_mq_free_rq_map(set, tags, hctx_idx); |
| 1413 | return NULL; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1414 | } |
| 1415 | |
Jens Axboe | 1429d7c | 2014-05-19 09:23:55 -0600 | [diff] [blame] | 1416 | static void blk_mq_free_bitmap(struct blk_mq_ctxmap *bitmap) |
| 1417 | { |
| 1418 | kfree(bitmap->map); |
| 1419 | } |
| 1420 | |
| 1421 | static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap *bitmap, int node) |
| 1422 | { |
| 1423 | unsigned int bpw = 8, total, num_maps, i; |
| 1424 | |
| 1425 | bitmap->bits_per_word = bpw; |
| 1426 | |
| 1427 | num_maps = ALIGN(nr_cpu_ids, bpw) / bpw; |
| 1428 | bitmap->map = kzalloc_node(num_maps * sizeof(struct blk_align_bitmap), |
| 1429 | GFP_KERNEL, node); |
| 1430 | if (!bitmap->map) |
| 1431 | return -ENOMEM; |
| 1432 | |
| 1433 | bitmap->map_size = num_maps; |
| 1434 | |
| 1435 | total = nr_cpu_ids; |
| 1436 | for (i = 0; i < num_maps; i++) { |
| 1437 | bitmap->map[i].depth = min(total, bitmap->bits_per_word); |
| 1438 | total -= bitmap->map[i].depth; |
| 1439 | } |
| 1440 | |
| 1441 | return 0; |
| 1442 | } |
| 1443 | |
Jens Axboe | 484b406 | 2014-05-21 14:01:15 -0600 | [diff] [blame] | 1444 | static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx *hctx, int cpu) |
| 1445 | { |
| 1446 | struct request_queue *q = hctx->queue; |
| 1447 | struct blk_mq_ctx *ctx; |
| 1448 | LIST_HEAD(tmp); |
| 1449 | |
| 1450 | /* |
| 1451 | * Move ctx entries to new CPU, if this one is going away. |
| 1452 | */ |
| 1453 | ctx = __blk_mq_get_ctx(q, cpu); |
| 1454 | |
| 1455 | spin_lock(&ctx->lock); |
| 1456 | if (!list_empty(&ctx->rq_list)) { |
| 1457 | list_splice_init(&ctx->rq_list, &tmp); |
| 1458 | blk_mq_hctx_clear_pending(hctx, ctx); |
| 1459 | } |
| 1460 | spin_unlock(&ctx->lock); |
| 1461 | |
| 1462 | if (list_empty(&tmp)) |
| 1463 | return NOTIFY_OK; |
| 1464 | |
| 1465 | ctx = blk_mq_get_ctx(q); |
| 1466 | spin_lock(&ctx->lock); |
| 1467 | |
| 1468 | while (!list_empty(&tmp)) { |
| 1469 | struct request *rq; |
| 1470 | |
| 1471 | rq = list_first_entry(&tmp, struct request, queuelist); |
| 1472 | rq->mq_ctx = ctx; |
| 1473 | list_move_tail(&rq->queuelist, &ctx->rq_list); |
| 1474 | } |
| 1475 | |
| 1476 | hctx = q->mq_ops->map_queue(q, ctx->cpu); |
| 1477 | blk_mq_hctx_mark_pending(hctx, ctx); |
| 1478 | |
| 1479 | spin_unlock(&ctx->lock); |
| 1480 | |
| 1481 | blk_mq_run_hw_queue(hctx, true); |
| 1482 | blk_mq_put_ctx(ctx); |
| 1483 | return NOTIFY_OK; |
| 1484 | } |
| 1485 | |
| 1486 | static int blk_mq_hctx_cpu_online(struct blk_mq_hw_ctx *hctx, int cpu) |
| 1487 | { |
| 1488 | struct request_queue *q = hctx->queue; |
| 1489 | struct blk_mq_tag_set *set = q->tag_set; |
| 1490 | |
| 1491 | if (set->tags[hctx->queue_num]) |
| 1492 | return NOTIFY_OK; |
| 1493 | |
| 1494 | set->tags[hctx->queue_num] = blk_mq_init_rq_map(set, hctx->queue_num); |
| 1495 | if (!set->tags[hctx->queue_num]) |
| 1496 | return NOTIFY_STOP; |
| 1497 | |
| 1498 | hctx->tags = set->tags[hctx->queue_num]; |
| 1499 | return NOTIFY_OK; |
| 1500 | } |
| 1501 | |
| 1502 | static int blk_mq_hctx_notify(void *data, unsigned long action, |
| 1503 | unsigned int cpu) |
| 1504 | { |
| 1505 | struct blk_mq_hw_ctx *hctx = data; |
| 1506 | |
| 1507 | if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) |
| 1508 | return blk_mq_hctx_cpu_offline(hctx, cpu); |
| 1509 | else if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) |
| 1510 | return blk_mq_hctx_cpu_online(hctx, cpu); |
| 1511 | |
| 1512 | return NOTIFY_OK; |
| 1513 | } |
| 1514 | |
Ming Lei | 08e98fc | 2014-09-25 23:23:38 +0800 | [diff] [blame] | 1515 | static void blk_mq_exit_hctx(struct request_queue *q, |
| 1516 | struct blk_mq_tag_set *set, |
| 1517 | struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) |
| 1518 | { |
| 1519 | blk_mq_tag_idle(hctx); |
| 1520 | |
| 1521 | if (set->ops->exit_hctx) |
| 1522 | set->ops->exit_hctx(hctx, hctx_idx); |
| 1523 | |
| 1524 | blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier); |
| 1525 | kfree(hctx->ctxs); |
| 1526 | blk_mq_free_bitmap(&hctx->ctx_map); |
| 1527 | } |
| 1528 | |
Ming Lei | 624dbe4 | 2014-05-27 23:35:13 +0800 | [diff] [blame] | 1529 | static void blk_mq_exit_hw_queues(struct request_queue *q, |
| 1530 | struct blk_mq_tag_set *set, int nr_queue) |
| 1531 | { |
| 1532 | struct blk_mq_hw_ctx *hctx; |
| 1533 | unsigned int i; |
| 1534 | |
| 1535 | queue_for_each_hw_ctx(q, hctx, i) { |
| 1536 | if (i == nr_queue) |
| 1537 | break; |
Ming Lei | 08e98fc | 2014-09-25 23:23:38 +0800 | [diff] [blame] | 1538 | blk_mq_exit_hctx(q, set, hctx, i); |
Ming Lei | 624dbe4 | 2014-05-27 23:35:13 +0800 | [diff] [blame] | 1539 | } |
Ming Lei | 624dbe4 | 2014-05-27 23:35:13 +0800 | [diff] [blame] | 1540 | } |
| 1541 | |
| 1542 | static void blk_mq_free_hw_queues(struct request_queue *q, |
| 1543 | struct blk_mq_tag_set *set) |
| 1544 | { |
| 1545 | struct blk_mq_hw_ctx *hctx; |
| 1546 | unsigned int i; |
| 1547 | |
| 1548 | queue_for_each_hw_ctx(q, hctx, i) { |
| 1549 | free_cpumask_var(hctx->cpumask); |
Christoph Hellwig | cdef54d | 2014-05-28 18:11:06 +0200 | [diff] [blame] | 1550 | kfree(hctx); |
Ming Lei | 624dbe4 | 2014-05-27 23:35:13 +0800 | [diff] [blame] | 1551 | } |
| 1552 | } |
| 1553 | |
Ming Lei | 08e98fc | 2014-09-25 23:23:38 +0800 | [diff] [blame] | 1554 | static int blk_mq_init_hctx(struct request_queue *q, |
| 1555 | struct blk_mq_tag_set *set, |
| 1556 | struct blk_mq_hw_ctx *hctx, unsigned hctx_idx) |
| 1557 | { |
| 1558 | int node; |
| 1559 | |
| 1560 | node = hctx->numa_node; |
| 1561 | if (node == NUMA_NO_NODE) |
| 1562 | node = hctx->numa_node = set->numa_node; |
| 1563 | |
| 1564 | INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn); |
| 1565 | INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn); |
| 1566 | spin_lock_init(&hctx->lock); |
| 1567 | INIT_LIST_HEAD(&hctx->dispatch); |
| 1568 | hctx->queue = q; |
| 1569 | hctx->queue_num = hctx_idx; |
| 1570 | hctx->flags = set->flags; |
| 1571 | hctx->cmd_size = set->cmd_size; |
| 1572 | |
| 1573 | blk_mq_init_cpu_notifier(&hctx->cpu_notifier, |
| 1574 | blk_mq_hctx_notify, hctx); |
| 1575 | blk_mq_register_cpu_notifier(&hctx->cpu_notifier); |
| 1576 | |
| 1577 | hctx->tags = set->tags[hctx_idx]; |
| 1578 | |
| 1579 | /* |
| 1580 | * Allocate space for all possible cpus to avoid allocation at |
| 1581 | * runtime |
| 1582 | */ |
| 1583 | hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *), |
| 1584 | GFP_KERNEL, node); |
| 1585 | if (!hctx->ctxs) |
| 1586 | goto unregister_cpu_notifier; |
| 1587 | |
| 1588 | if (blk_mq_alloc_bitmap(&hctx->ctx_map, node)) |
| 1589 | goto free_ctxs; |
| 1590 | |
| 1591 | hctx->nr_ctx = 0; |
| 1592 | |
| 1593 | if (set->ops->init_hctx && |
| 1594 | set->ops->init_hctx(hctx, set->driver_data, hctx_idx)) |
| 1595 | goto free_bitmap; |
| 1596 | |
| 1597 | return 0; |
| 1598 | |
| 1599 | free_bitmap: |
| 1600 | blk_mq_free_bitmap(&hctx->ctx_map); |
| 1601 | free_ctxs: |
| 1602 | kfree(hctx->ctxs); |
| 1603 | unregister_cpu_notifier: |
| 1604 | blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier); |
| 1605 | |
| 1606 | return -1; |
| 1607 | } |
| 1608 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1609 | static int blk_mq_init_hw_queues(struct request_queue *q, |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1610 | struct blk_mq_tag_set *set) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1611 | { |
| 1612 | struct blk_mq_hw_ctx *hctx; |
Ming Lei | 624dbe4 | 2014-05-27 23:35:13 +0800 | [diff] [blame] | 1613 | unsigned int i; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1614 | |
| 1615 | /* |
| 1616 | * Initialize hardware queues |
| 1617 | */ |
| 1618 | queue_for_each_hw_ctx(q, hctx, i) { |
Ming Lei | 08e98fc | 2014-09-25 23:23:38 +0800 | [diff] [blame] | 1619 | if (blk_mq_init_hctx(q, set, hctx, i)) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1620 | break; |
| 1621 | } |
| 1622 | |
| 1623 | if (i == q->nr_hw_queues) |
| 1624 | return 0; |
| 1625 | |
| 1626 | /* |
| 1627 | * Init failed |
| 1628 | */ |
Ming Lei | 624dbe4 | 2014-05-27 23:35:13 +0800 | [diff] [blame] | 1629 | blk_mq_exit_hw_queues(q, set, i); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1630 | |
| 1631 | return 1; |
| 1632 | } |
| 1633 | |
| 1634 | static void blk_mq_init_cpu_queues(struct request_queue *q, |
| 1635 | unsigned int nr_hw_queues) |
| 1636 | { |
| 1637 | unsigned int i; |
| 1638 | |
| 1639 | for_each_possible_cpu(i) { |
| 1640 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); |
| 1641 | struct blk_mq_hw_ctx *hctx; |
| 1642 | |
| 1643 | memset(__ctx, 0, sizeof(*__ctx)); |
| 1644 | __ctx->cpu = i; |
| 1645 | spin_lock_init(&__ctx->lock); |
| 1646 | INIT_LIST_HEAD(&__ctx->rq_list); |
| 1647 | __ctx->queue = q; |
| 1648 | |
| 1649 | /* 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] | 1650 | if (!cpu_online(i)) |
| 1651 | continue; |
| 1652 | |
Jens Axboe | e4043dc | 2014-04-09 10:18:23 -0600 | [diff] [blame] | 1653 | hctx = q->mq_ops->map_queue(q, i); |
| 1654 | cpumask_set_cpu(i, hctx->cpumask); |
| 1655 | hctx->nr_ctx++; |
| 1656 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1657 | /* |
| 1658 | * Set local node, IFF we have more than one hw queue. If |
| 1659 | * not, we remain on the home node of the device |
| 1660 | */ |
| 1661 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) |
| 1662 | hctx->numa_node = cpu_to_node(i); |
| 1663 | } |
| 1664 | } |
| 1665 | |
| 1666 | static void blk_mq_map_swqueue(struct request_queue *q) |
| 1667 | { |
| 1668 | unsigned int i; |
| 1669 | struct blk_mq_hw_ctx *hctx; |
| 1670 | struct blk_mq_ctx *ctx; |
| 1671 | |
| 1672 | queue_for_each_hw_ctx(q, hctx, i) { |
Jens Axboe | e4043dc | 2014-04-09 10:18:23 -0600 | [diff] [blame] | 1673 | cpumask_clear(hctx->cpumask); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1674 | hctx->nr_ctx = 0; |
| 1675 | } |
| 1676 | |
| 1677 | /* |
| 1678 | * Map software to hardware queues |
| 1679 | */ |
| 1680 | queue_for_each_ctx(q, ctx, i) { |
| 1681 | /* If the cpu isn't online, the cpu is mapped to first hctx */ |
Jens Axboe | e4043dc | 2014-04-09 10:18:23 -0600 | [diff] [blame] | 1682 | if (!cpu_online(i)) |
| 1683 | continue; |
| 1684 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1685 | hctx = q->mq_ops->map_queue(q, i); |
Jens Axboe | e4043dc | 2014-04-09 10:18:23 -0600 | [diff] [blame] | 1686 | cpumask_set_cpu(i, hctx->cpumask); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1687 | ctx->index_hw = hctx->nr_ctx; |
| 1688 | hctx->ctxs[hctx->nr_ctx++] = ctx; |
| 1689 | } |
Jens Axboe | 506e931 | 2014-05-07 10:26:44 -0600 | [diff] [blame] | 1690 | |
| 1691 | queue_for_each_hw_ctx(q, hctx, i) { |
Jens Axboe | 484b406 | 2014-05-21 14:01:15 -0600 | [diff] [blame] | 1692 | /* |
Jens Axboe | a68aafa | 2014-08-15 13:19:15 -0600 | [diff] [blame] | 1693 | * If no software queues are mapped to this hardware queue, |
| 1694 | * disable it and free the request entries. |
Jens Axboe | 484b406 | 2014-05-21 14:01:15 -0600 | [diff] [blame] | 1695 | */ |
| 1696 | if (!hctx->nr_ctx) { |
| 1697 | struct blk_mq_tag_set *set = q->tag_set; |
| 1698 | |
| 1699 | if (set->tags[i]) { |
| 1700 | blk_mq_free_rq_map(set, set->tags[i], i); |
| 1701 | set->tags[i] = NULL; |
| 1702 | hctx->tags = NULL; |
| 1703 | } |
| 1704 | continue; |
| 1705 | } |
| 1706 | |
| 1707 | /* |
| 1708 | * Initialize batch roundrobin counts |
| 1709 | */ |
Jens Axboe | 506e931 | 2014-05-07 10:26:44 -0600 | [diff] [blame] | 1710 | hctx->next_cpu = cpumask_first(hctx->cpumask); |
| 1711 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; |
| 1712 | } |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1713 | } |
| 1714 | |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 1715 | static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set) |
| 1716 | { |
| 1717 | struct blk_mq_hw_ctx *hctx; |
| 1718 | struct request_queue *q; |
| 1719 | bool shared; |
| 1720 | int i; |
| 1721 | |
| 1722 | if (set->tag_list.next == set->tag_list.prev) |
| 1723 | shared = false; |
| 1724 | else |
| 1725 | shared = true; |
| 1726 | |
| 1727 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
| 1728 | blk_mq_freeze_queue(q); |
| 1729 | |
| 1730 | queue_for_each_hw_ctx(q, hctx, i) { |
| 1731 | if (shared) |
| 1732 | hctx->flags |= BLK_MQ_F_TAG_SHARED; |
| 1733 | else |
| 1734 | hctx->flags &= ~BLK_MQ_F_TAG_SHARED; |
| 1735 | } |
| 1736 | blk_mq_unfreeze_queue(q); |
| 1737 | } |
| 1738 | } |
| 1739 | |
| 1740 | static void blk_mq_del_queue_tag_set(struct request_queue *q) |
| 1741 | { |
| 1742 | struct blk_mq_tag_set *set = q->tag_set; |
| 1743 | |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 1744 | mutex_lock(&set->tag_list_lock); |
| 1745 | list_del_init(&q->tag_set_list); |
| 1746 | blk_mq_update_tag_set_depth(set); |
| 1747 | mutex_unlock(&set->tag_list_lock); |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 1748 | } |
| 1749 | |
| 1750 | static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, |
| 1751 | struct request_queue *q) |
| 1752 | { |
| 1753 | q->tag_set = set; |
| 1754 | |
| 1755 | mutex_lock(&set->tag_list_lock); |
| 1756 | list_add_tail(&q->tag_set_list, &set->tag_list); |
| 1757 | blk_mq_update_tag_set_depth(set); |
| 1758 | mutex_unlock(&set->tag_list_lock); |
| 1759 | } |
| 1760 | |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1761 | 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] | 1762 | { |
| 1763 | struct blk_mq_hw_ctx **hctxs; |
Ming Lei | e6cdb09 | 2014-06-03 11:24:06 +0800 | [diff] [blame] | 1764 | struct blk_mq_ctx __percpu *ctx; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1765 | struct request_queue *q; |
Jens Axboe | f14bbe7 | 2014-05-27 12:06:53 -0600 | [diff] [blame] | 1766 | unsigned int *map; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1767 | int i; |
| 1768 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1769 | ctx = alloc_percpu(struct blk_mq_ctx); |
| 1770 | if (!ctx) |
| 1771 | return ERR_PTR(-ENOMEM); |
| 1772 | |
Jens Axboe | aedcd72 | 2014-09-17 08:27:03 -0600 | [diff] [blame] | 1773 | /* |
| 1774 | * If a crashdump is active, then we are potentially in a very |
| 1775 | * memory constrained environment. Limit us to 1 queue and |
| 1776 | * 64 tags to prevent using too much memory. |
| 1777 | */ |
| 1778 | if (is_kdump_kernel()) { |
| 1779 | set->nr_hw_queues = 1; |
| 1780 | set->queue_depth = min(64U, set->queue_depth); |
| 1781 | } |
| 1782 | |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1783 | hctxs = kmalloc_node(set->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL, |
| 1784 | set->numa_node); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1785 | |
| 1786 | if (!hctxs) |
| 1787 | goto err_percpu; |
| 1788 | |
Jens Axboe | f14bbe7 | 2014-05-27 12:06:53 -0600 | [diff] [blame] | 1789 | map = blk_mq_make_queue_map(set); |
| 1790 | if (!map) |
| 1791 | goto err_map; |
| 1792 | |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1793 | for (i = 0; i < set->nr_hw_queues; i++) { |
Jens Axboe | f14bbe7 | 2014-05-27 12:06:53 -0600 | [diff] [blame] | 1794 | int node = blk_mq_hw_queue_to_node(map, i); |
| 1795 | |
Christoph Hellwig | cdef54d | 2014-05-28 18:11:06 +0200 | [diff] [blame] | 1796 | hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx), |
| 1797 | GFP_KERNEL, node); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1798 | if (!hctxs[i]) |
| 1799 | goto err_hctxs; |
| 1800 | |
Jens Axboe | e4043dc | 2014-04-09 10:18:23 -0600 | [diff] [blame] | 1801 | if (!zalloc_cpumask_var(&hctxs[i]->cpumask, GFP_KERNEL)) |
| 1802 | goto err_hctxs; |
| 1803 | |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 1804 | atomic_set(&hctxs[i]->nr_active, 0); |
Jens Axboe | f14bbe7 | 2014-05-27 12:06:53 -0600 | [diff] [blame] | 1805 | hctxs[i]->numa_node = node; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1806 | hctxs[i]->queue_num = i; |
| 1807 | } |
| 1808 | |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1809 | q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1810 | if (!q) |
| 1811 | goto err_hctxs; |
| 1812 | |
Tejun Heo | add703f | 2014-07-01 10:34:38 -0600 | [diff] [blame] | 1813 | if (percpu_ref_init(&q->mq_usage_counter, blk_mq_usage_counter_release)) |
Ming Lei | 3d2936f | 2014-05-27 23:35:14 +0800 | [diff] [blame] | 1814 | goto err_map; |
| 1815 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1816 | setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q); |
| 1817 | blk_queue_rq_timeout(q, 30000); |
| 1818 | |
| 1819 | q->nr_queues = nr_cpu_ids; |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1820 | q->nr_hw_queues = set->nr_hw_queues; |
Jens Axboe | f14bbe7 | 2014-05-27 12:06:53 -0600 | [diff] [blame] | 1821 | q->mq_map = map; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1822 | |
| 1823 | q->queue_ctx = ctx; |
| 1824 | q->queue_hw_ctx = hctxs; |
| 1825 | |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1826 | q->mq_ops = set->ops; |
Jens Axboe | 94eddfb | 2013-11-19 09:25:07 -0700 | [diff] [blame] | 1827 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1828 | |
Jens Axboe | 05f1dd5 | 2014-05-29 09:53:32 -0600 | [diff] [blame] | 1829 | if (!(set->flags & BLK_MQ_F_SG_MERGE)) |
| 1830 | q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE; |
| 1831 | |
Christoph Hellwig | 1be036e | 2014-02-07 10:22:39 -0800 | [diff] [blame] | 1832 | q->sg_reserved_size = INT_MAX; |
| 1833 | |
Christoph Hellwig | 6fca6a6 | 2014-05-28 08:08:02 -0600 | [diff] [blame] | 1834 | INIT_WORK(&q->requeue_work, blk_mq_requeue_work); |
| 1835 | INIT_LIST_HEAD(&q->requeue_list); |
| 1836 | spin_lock_init(&q->requeue_lock); |
| 1837 | |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1838 | if (q->nr_hw_queues > 1) |
| 1839 | blk_queue_make_request(q, blk_mq_make_request); |
| 1840 | else |
| 1841 | blk_queue_make_request(q, blk_sq_make_request); |
| 1842 | |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1843 | if (set->timeout) |
| 1844 | blk_queue_rq_timeout(q, set->timeout); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1845 | |
Jens Axboe | eba7176 | 2014-05-20 15:17:27 -0600 | [diff] [blame] | 1846 | /* |
| 1847 | * Do this after blk_queue_make_request() overrides it... |
| 1848 | */ |
| 1849 | q->nr_requests = set->queue_depth; |
| 1850 | |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1851 | if (set->ops->complete) |
| 1852 | blk_queue_softirq_done(q, set->ops->complete); |
Christoph Hellwig | 30a91cb | 2014-02-10 03:24:38 -0800 | [diff] [blame] | 1853 | |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1854 | blk_mq_init_cpu_queues(q, set->nr_hw_queues); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1855 | |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1856 | if (blk_mq_init_hw_queues(q, set)) |
Ming Lei | 1bcb1ea | 2014-09-25 23:23:39 +0800 | [diff] [blame] | 1857 | goto err_hw; |
Christoph Hellwig | 1874198 | 2014-02-10 09:29:00 -0700 | [diff] [blame] | 1858 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1859 | mutex_lock(&all_q_mutex); |
| 1860 | list_add_tail(&q->all_q_node, &all_q_list); |
| 1861 | mutex_unlock(&all_q_mutex); |
| 1862 | |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 1863 | blk_mq_add_queue_tag_set(set, q); |
| 1864 | |
Ming Lei | ba48338 | 2014-09-25 23:23:44 +0800 | [diff] [blame] | 1865 | q->fq = blk_alloc_flush_queue(q); |
| 1866 | if (!q->fq) |
Ming Lei | 1bcb1ea | 2014-09-25 23:23:39 +0800 | [diff] [blame] | 1867 | goto err_hw_queues; |
| 1868 | |
Jens Axboe | 484b406 | 2014-05-21 14:01:15 -0600 | [diff] [blame] | 1869 | blk_mq_map_swqueue(q); |
| 1870 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1871 | return q; |
Christoph Hellwig | 1874198 | 2014-02-10 09:29:00 -0700 | [diff] [blame] | 1872 | |
Ming Lei | 1bcb1ea | 2014-09-25 23:23:39 +0800 | [diff] [blame] | 1873 | err_hw_queues: |
| 1874 | blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1875 | err_hw: |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1876 | blk_cleanup_queue(q); |
| 1877 | err_hctxs: |
Jens Axboe | f14bbe7 | 2014-05-27 12:06:53 -0600 | [diff] [blame] | 1878 | kfree(map); |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1879 | for (i = 0; i < set->nr_hw_queues; i++) { |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1880 | if (!hctxs[i]) |
| 1881 | break; |
Jens Axboe | e4043dc | 2014-04-09 10:18:23 -0600 | [diff] [blame] | 1882 | free_cpumask_var(hctxs[i]->cpumask); |
Christoph Hellwig | cdef54d | 2014-05-28 18:11:06 +0200 | [diff] [blame] | 1883 | kfree(hctxs[i]); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1884 | } |
Jens Axboe | f14bbe7 | 2014-05-27 12:06:53 -0600 | [diff] [blame] | 1885 | err_map: |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1886 | kfree(hctxs); |
| 1887 | err_percpu: |
| 1888 | free_percpu(ctx); |
| 1889 | return ERR_PTR(-ENOMEM); |
| 1890 | } |
| 1891 | EXPORT_SYMBOL(blk_mq_init_queue); |
| 1892 | |
| 1893 | void blk_mq_free_queue(struct request_queue *q) |
| 1894 | { |
Ming Lei | 624dbe4 | 2014-05-27 23:35:13 +0800 | [diff] [blame] | 1895 | struct blk_mq_tag_set *set = q->tag_set; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1896 | |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 1897 | blk_mq_del_queue_tag_set(q); |
| 1898 | |
Ming Lei | 624dbe4 | 2014-05-27 23:35:13 +0800 | [diff] [blame] | 1899 | blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); |
| 1900 | blk_mq_free_hw_queues(q, set); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1901 | |
Tejun Heo | add703f | 2014-07-01 10:34:38 -0600 | [diff] [blame] | 1902 | percpu_ref_exit(&q->mq_usage_counter); |
Ming Lei | 3d2936f | 2014-05-27 23:35:14 +0800 | [diff] [blame] | 1903 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1904 | free_percpu(q->queue_ctx); |
| 1905 | kfree(q->queue_hw_ctx); |
| 1906 | kfree(q->mq_map); |
| 1907 | |
| 1908 | q->queue_ctx = NULL; |
| 1909 | q->queue_hw_ctx = NULL; |
| 1910 | q->mq_map = NULL; |
| 1911 | |
| 1912 | mutex_lock(&all_q_mutex); |
| 1913 | list_del_init(&q->all_q_node); |
| 1914 | mutex_unlock(&all_q_mutex); |
| 1915 | } |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1916 | |
| 1917 | /* Basically redo blk_mq_init_queue with queue frozen */ |
Paul Gortmaker | f618ef7 | 2013-11-14 08:26:02 -0700 | [diff] [blame] | 1918 | static void blk_mq_queue_reinit(struct request_queue *q) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1919 | { |
| 1920 | blk_mq_freeze_queue(q); |
| 1921 | |
Jens Axboe | 67aec14 | 2014-05-30 08:25:36 -0600 | [diff] [blame] | 1922 | blk_mq_sysfs_unregister(q); |
| 1923 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1924 | blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues); |
| 1925 | |
| 1926 | /* |
| 1927 | * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe |
| 1928 | * we should change hctx numa_node according to new topology (this |
| 1929 | * involves free and re-allocate memory, worthy doing?) |
| 1930 | */ |
| 1931 | |
| 1932 | blk_mq_map_swqueue(q); |
| 1933 | |
Jens Axboe | 67aec14 | 2014-05-30 08:25:36 -0600 | [diff] [blame] | 1934 | blk_mq_sysfs_register(q); |
| 1935 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1936 | blk_mq_unfreeze_queue(q); |
| 1937 | } |
| 1938 | |
Paul Gortmaker | f618ef7 | 2013-11-14 08:26:02 -0700 | [diff] [blame] | 1939 | static int blk_mq_queue_reinit_notify(struct notifier_block *nb, |
| 1940 | unsigned long action, void *hcpu) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1941 | { |
| 1942 | struct request_queue *q; |
| 1943 | |
| 1944 | /* |
Jens Axboe | 9fccfed | 2014-05-08 14:50:19 -0600 | [diff] [blame] | 1945 | * Before new mappings are established, hotadded cpu might already |
| 1946 | * start handling requests. This doesn't break anything as we map |
| 1947 | * offline CPUs to first hardware queue. We will re-init the queue |
| 1948 | * below to get optimal settings. |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1949 | */ |
| 1950 | if (action != CPU_DEAD && action != CPU_DEAD_FROZEN && |
| 1951 | action != CPU_ONLINE && action != CPU_ONLINE_FROZEN) |
| 1952 | return NOTIFY_OK; |
| 1953 | |
| 1954 | mutex_lock(&all_q_mutex); |
| 1955 | list_for_each_entry(q, &all_q_list, all_q_node) |
| 1956 | blk_mq_queue_reinit(q); |
| 1957 | mutex_unlock(&all_q_mutex); |
| 1958 | return NOTIFY_OK; |
| 1959 | } |
| 1960 | |
Jens Axboe | a516440 | 2014-09-10 09:02:03 -0600 | [diff] [blame] | 1961 | static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) |
| 1962 | { |
| 1963 | int i; |
| 1964 | |
| 1965 | for (i = 0; i < set->nr_hw_queues; i++) { |
| 1966 | set->tags[i] = blk_mq_init_rq_map(set, i); |
| 1967 | if (!set->tags[i]) |
| 1968 | goto out_unwind; |
| 1969 | } |
| 1970 | |
| 1971 | return 0; |
| 1972 | |
| 1973 | out_unwind: |
| 1974 | while (--i >= 0) |
| 1975 | blk_mq_free_rq_map(set, set->tags[i], i); |
| 1976 | |
Jens Axboe | a516440 | 2014-09-10 09:02:03 -0600 | [diff] [blame] | 1977 | return -ENOMEM; |
| 1978 | } |
| 1979 | |
| 1980 | /* |
| 1981 | * Allocate the request maps associated with this tag_set. Note that this |
| 1982 | * may reduce the depth asked for, if memory is tight. set->queue_depth |
| 1983 | * will be updated to reflect the allocated depth. |
| 1984 | */ |
| 1985 | static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) |
| 1986 | { |
| 1987 | unsigned int depth; |
| 1988 | int err; |
| 1989 | |
| 1990 | depth = set->queue_depth; |
| 1991 | do { |
| 1992 | err = __blk_mq_alloc_rq_maps(set); |
| 1993 | if (!err) |
| 1994 | break; |
| 1995 | |
| 1996 | set->queue_depth >>= 1; |
| 1997 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) { |
| 1998 | err = -ENOMEM; |
| 1999 | break; |
| 2000 | } |
| 2001 | } while (set->queue_depth); |
| 2002 | |
| 2003 | if (!set->queue_depth || err) { |
| 2004 | pr_err("blk-mq: failed to allocate request map\n"); |
| 2005 | return -ENOMEM; |
| 2006 | } |
| 2007 | |
| 2008 | if (depth != set->queue_depth) |
| 2009 | pr_info("blk-mq: reduced tag depth (%u -> %u)\n", |
| 2010 | depth, set->queue_depth); |
| 2011 | |
| 2012 | return 0; |
| 2013 | } |
| 2014 | |
Jens Axboe | a4391c6 | 2014-06-05 15:21:56 -0600 | [diff] [blame] | 2015 | /* |
| 2016 | * Alloc a tag set to be associated with one or more request queues. |
| 2017 | * May fail with EINVAL for various error conditions. May adjust the |
| 2018 | * requested depth down, if if it too large. In that case, the set |
| 2019 | * value will be stored in set->queue_depth. |
| 2020 | */ |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 2021 | int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) |
| 2022 | { |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 2023 | if (!set->nr_hw_queues) |
| 2024 | return -EINVAL; |
Jens Axboe | a4391c6 | 2014-06-05 15:21:56 -0600 | [diff] [blame] | 2025 | if (!set->queue_depth) |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 2026 | return -EINVAL; |
| 2027 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) |
| 2028 | return -EINVAL; |
| 2029 | |
Christoph Hellwig | cdef54d | 2014-05-28 18:11:06 +0200 | [diff] [blame] | 2030 | if (!set->nr_hw_queues || !set->ops->queue_rq || !set->ops->map_queue) |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 2031 | return -EINVAL; |
| 2032 | |
Jens Axboe | a4391c6 | 2014-06-05 15:21:56 -0600 | [diff] [blame] | 2033 | if (set->queue_depth > BLK_MQ_MAX_DEPTH) { |
| 2034 | pr_info("blk-mq: reduced tag depth to %u\n", |
| 2035 | BLK_MQ_MAX_DEPTH); |
| 2036 | set->queue_depth = BLK_MQ_MAX_DEPTH; |
| 2037 | } |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 2038 | |
Ming Lei | 4847900 | 2014-04-19 18:00:17 +0800 | [diff] [blame] | 2039 | set->tags = kmalloc_node(set->nr_hw_queues * |
| 2040 | sizeof(struct blk_mq_tags *), |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 2041 | GFP_KERNEL, set->numa_node); |
| 2042 | if (!set->tags) |
Jens Axboe | a516440 | 2014-09-10 09:02:03 -0600 | [diff] [blame] | 2043 | return -ENOMEM; |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 2044 | |
Jens Axboe | a516440 | 2014-09-10 09:02:03 -0600 | [diff] [blame] | 2045 | if (blk_mq_alloc_rq_maps(set)) |
| 2046 | goto enomem; |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 2047 | |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 2048 | mutex_init(&set->tag_list_lock); |
| 2049 | INIT_LIST_HEAD(&set->tag_list); |
| 2050 | |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 2051 | return 0; |
Jens Axboe | a516440 | 2014-09-10 09:02:03 -0600 | [diff] [blame] | 2052 | enomem: |
Robert Elliott | 5676e7b | 2014-09-02 11:38:44 -0500 | [diff] [blame] | 2053 | kfree(set->tags); |
| 2054 | set->tags = NULL; |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 2055 | return -ENOMEM; |
| 2056 | } |
| 2057 | EXPORT_SYMBOL(blk_mq_alloc_tag_set); |
| 2058 | |
| 2059 | void blk_mq_free_tag_set(struct blk_mq_tag_set *set) |
| 2060 | { |
| 2061 | int i; |
| 2062 | |
Jens Axboe | 484b406 | 2014-05-21 14:01:15 -0600 | [diff] [blame] | 2063 | for (i = 0; i < set->nr_hw_queues; i++) { |
| 2064 | if (set->tags[i]) |
| 2065 | blk_mq_free_rq_map(set, set->tags[i], i); |
| 2066 | } |
| 2067 | |
Ming Lei | 981bd18 | 2014-04-24 00:07:34 +0800 | [diff] [blame] | 2068 | kfree(set->tags); |
Robert Elliott | 5676e7b | 2014-09-02 11:38:44 -0500 | [diff] [blame] | 2069 | set->tags = NULL; |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 2070 | } |
| 2071 | EXPORT_SYMBOL(blk_mq_free_tag_set); |
| 2072 | |
Jens Axboe | e3a2b3f | 2014-05-20 11:49:02 -0600 | [diff] [blame] | 2073 | int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr) |
| 2074 | { |
| 2075 | struct blk_mq_tag_set *set = q->tag_set; |
| 2076 | struct blk_mq_hw_ctx *hctx; |
| 2077 | int i, ret; |
| 2078 | |
| 2079 | if (!set || nr > set->queue_depth) |
| 2080 | return -EINVAL; |
| 2081 | |
| 2082 | ret = 0; |
| 2083 | queue_for_each_hw_ctx(q, hctx, i) { |
| 2084 | ret = blk_mq_tag_update_depth(hctx->tags, nr); |
| 2085 | if (ret) |
| 2086 | break; |
| 2087 | } |
| 2088 | |
| 2089 | if (!ret) |
| 2090 | q->nr_requests = nr; |
| 2091 | |
| 2092 | return ret; |
| 2093 | } |
| 2094 | |
Jens Axboe | 676141e | 2014-03-20 13:29:18 -0600 | [diff] [blame] | 2095 | void blk_mq_disable_hotplug(void) |
| 2096 | { |
| 2097 | mutex_lock(&all_q_mutex); |
| 2098 | } |
| 2099 | |
| 2100 | void blk_mq_enable_hotplug(void) |
| 2101 | { |
| 2102 | mutex_unlock(&all_q_mutex); |
| 2103 | } |
| 2104 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2105 | static int __init blk_mq_init(void) |
| 2106 | { |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2107 | blk_mq_cpu_init(); |
| 2108 | |
Tejun Heo | add703f | 2014-07-01 10:34:38 -0600 | [diff] [blame] | 2109 | hotcpu_notifier(blk_mq_queue_reinit_notify, 0); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2110 | |
| 2111 | return 0; |
| 2112 | } |
| 2113 | subsys_initcall(blk_mq_init); |