Jens Axboe | 75bb462 | 2014-05-28 10:15:41 -0600 | [diff] [blame] | 1 | /* |
| 2 | * Block multiqueue core code |
| 3 | * |
| 4 | * Copyright (C) 2013-2014 Jens Axboe |
| 5 | * Copyright (C) 2013-2014 Christoph Hellwig |
| 6 | */ |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 7 | #include <linux/kernel.h> |
| 8 | #include <linux/module.h> |
| 9 | #include <linux/backing-dev.h> |
| 10 | #include <linux/bio.h> |
| 11 | #include <linux/blkdev.h> |
Catalin Marinas | f75782e | 2015-09-14 18:16:02 +0100 | [diff] [blame] | 12 | #include <linux/kmemleak.h> |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 13 | #include <linux/mm.h> |
| 14 | #include <linux/init.h> |
| 15 | #include <linux/slab.h> |
| 16 | #include <linux/workqueue.h> |
| 17 | #include <linux/smp.h> |
| 18 | #include <linux/llist.h> |
| 19 | #include <linux/list_sort.h> |
| 20 | #include <linux/cpu.h> |
| 21 | #include <linux/cache.h> |
| 22 | #include <linux/sched/sysctl.h> |
| 23 | #include <linux/delay.h> |
Jens Axboe | aedcd72 | 2014-09-17 08:27:03 -0600 | [diff] [blame] | 24 | #include <linux/crash_dump.h> |
Jens Axboe | 88c7b2b | 2016-08-25 08:07:30 -0600 | [diff] [blame] | 25 | #include <linux/prefetch.h> |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 26 | |
| 27 | #include <trace/events/block.h> |
| 28 | |
| 29 | #include <linux/blk-mq.h> |
| 30 | #include "blk.h" |
| 31 | #include "blk-mq.h" |
| 32 | #include "blk-mq-tag.h" |
Jens Axboe | cf43e6b | 2016-11-07 21:32:37 -0700 | [diff] [blame] | 33 | #include "blk-stat.h" |
Jens Axboe | 87760e5 | 2016-11-09 12:38:14 -0700 | [diff] [blame] | 34 | #include "blk-wbt.h" |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 35 | |
| 36 | static DEFINE_MUTEX(all_q_mutex); |
| 37 | static LIST_HEAD(all_q_list); |
| 38 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 39 | /* |
| 40 | * Check if any of the ctx's have pending work in this hardware queue |
| 41 | */ |
| 42 | static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx) |
| 43 | { |
Omar Sandoval | 8845964 | 2016-09-17 08:38:44 -0600 | [diff] [blame] | 44 | return sbitmap_any_bit_set(&hctx->ctx_map); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 45 | } |
| 46 | |
| 47 | /* |
| 48 | * Mark this ctx as having pending work in this hardware queue |
| 49 | */ |
| 50 | static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx, |
| 51 | struct blk_mq_ctx *ctx) |
| 52 | { |
Omar Sandoval | 8845964 | 2016-09-17 08:38:44 -0600 | [diff] [blame] | 53 | if (!sbitmap_test_bit(&hctx->ctx_map, ctx->index_hw)) |
| 54 | sbitmap_set_bit(&hctx->ctx_map, ctx->index_hw); |
Jens Axboe | 1429d7c | 2014-05-19 09:23:55 -0600 | [diff] [blame] | 55 | } |
| 56 | |
| 57 | static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx, |
| 58 | struct blk_mq_ctx *ctx) |
| 59 | { |
Omar Sandoval | 8845964 | 2016-09-17 08:38:44 -0600 | [diff] [blame] | 60 | sbitmap_clear_bit(&hctx->ctx_map, ctx->index_hw); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 61 | } |
| 62 | |
Keith Busch | b4c6a02 | 2014-12-19 17:54:14 -0700 | [diff] [blame] | 63 | void blk_mq_freeze_queue_start(struct request_queue *q) |
Ming Lei | 43a5e4e | 2013-12-26 21:31:35 +0800 | [diff] [blame] | 64 | { |
Christoph Hellwig | 4ecd4fe | 2015-05-07 09:38:13 +0200 | [diff] [blame] | 65 | int freeze_depth; |
Tejun Heo | cddd5d1 | 2014-08-16 08:02:24 -0400 | [diff] [blame] | 66 | |
Christoph Hellwig | 4ecd4fe | 2015-05-07 09:38:13 +0200 | [diff] [blame] | 67 | freeze_depth = atomic_inc_return(&q->mq_freeze_depth); |
| 68 | if (freeze_depth == 1) { |
Dan Williams | 3ef28e8 | 2015-10-21 13:20:12 -0400 | [diff] [blame] | 69 | percpu_ref_kill(&q->q_usage_counter); |
Mike Snitzer | b94ec29 | 2015-03-11 23:56:38 -0400 | [diff] [blame] | 70 | blk_mq_run_hw_queues(q, false); |
Tejun Heo | cddd5d1 | 2014-08-16 08:02:24 -0400 | [diff] [blame] | 71 | } |
Tejun Heo | f3af020 | 2014-11-04 13:52:27 -0500 | [diff] [blame] | 72 | } |
Keith Busch | b4c6a02 | 2014-12-19 17:54:14 -0700 | [diff] [blame] | 73 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_start); |
Tejun Heo | f3af020 | 2014-11-04 13:52:27 -0500 | [diff] [blame] | 74 | |
| 75 | static void blk_mq_freeze_queue_wait(struct request_queue *q) |
| 76 | { |
Dan Williams | 3ef28e8 | 2015-10-21 13:20:12 -0400 | [diff] [blame] | 77 | wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter)); |
Ming Lei | 43a5e4e | 2013-12-26 21:31:35 +0800 | [diff] [blame] | 78 | } |
| 79 | |
Tejun Heo | f3af020 | 2014-11-04 13:52:27 -0500 | [diff] [blame] | 80 | /* |
| 81 | * Guarantee no request is in use, so we can change any data structure of |
| 82 | * the queue afterward. |
| 83 | */ |
Dan Williams | 3ef28e8 | 2015-10-21 13:20:12 -0400 | [diff] [blame] | 84 | void blk_freeze_queue(struct request_queue *q) |
Tejun Heo | f3af020 | 2014-11-04 13:52:27 -0500 | [diff] [blame] | 85 | { |
Dan Williams | 3ef28e8 | 2015-10-21 13:20:12 -0400 | [diff] [blame] | 86 | /* |
| 87 | * In the !blk_mq case we are only calling this to kill the |
| 88 | * q_usage_counter, otherwise this increases the freeze depth |
| 89 | * and waits for it to return to zero. For this reason there is |
| 90 | * no blk_unfreeze_queue(), and blk_freeze_queue() is not |
| 91 | * exported to drivers as the only user for unfreeze is blk_mq. |
| 92 | */ |
Tejun Heo | f3af020 | 2014-11-04 13:52:27 -0500 | [diff] [blame] | 93 | blk_mq_freeze_queue_start(q); |
| 94 | blk_mq_freeze_queue_wait(q); |
| 95 | } |
Dan Williams | 3ef28e8 | 2015-10-21 13:20:12 -0400 | [diff] [blame] | 96 | |
| 97 | void blk_mq_freeze_queue(struct request_queue *q) |
| 98 | { |
| 99 | /* |
| 100 | * ...just an alias to keep freeze and unfreeze actions balanced |
| 101 | * in the blk_mq_* namespace |
| 102 | */ |
| 103 | blk_freeze_queue(q); |
| 104 | } |
Jens Axboe | c761d96 | 2015-01-02 15:05:12 -0700 | [diff] [blame] | 105 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue); |
Tejun Heo | f3af020 | 2014-11-04 13:52:27 -0500 | [diff] [blame] | 106 | |
Keith Busch | b4c6a02 | 2014-12-19 17:54:14 -0700 | [diff] [blame] | 107 | void blk_mq_unfreeze_queue(struct request_queue *q) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 108 | { |
Christoph Hellwig | 4ecd4fe | 2015-05-07 09:38:13 +0200 | [diff] [blame] | 109 | int freeze_depth; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 110 | |
Christoph Hellwig | 4ecd4fe | 2015-05-07 09:38:13 +0200 | [diff] [blame] | 111 | freeze_depth = atomic_dec_return(&q->mq_freeze_depth); |
| 112 | WARN_ON_ONCE(freeze_depth < 0); |
| 113 | if (!freeze_depth) { |
Dan Williams | 3ef28e8 | 2015-10-21 13:20:12 -0400 | [diff] [blame] | 114 | percpu_ref_reinit(&q->q_usage_counter); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 115 | wake_up_all(&q->mq_freeze_wq); |
Tejun Heo | add703f | 2014-07-01 10:34:38 -0600 | [diff] [blame] | 116 | } |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 117 | } |
Keith Busch | b4c6a02 | 2014-12-19 17:54:14 -0700 | [diff] [blame] | 118 | EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 119 | |
Bart Van Assche | 6a83e74 | 2016-11-02 10:09:51 -0600 | [diff] [blame] | 120 | /** |
| 121 | * blk_mq_quiesce_queue() - wait until all ongoing queue_rq calls have finished |
| 122 | * @q: request queue. |
| 123 | * |
| 124 | * Note: this function does not prevent that the struct request end_io() |
| 125 | * callback function is invoked. Additionally, it is not prevented that |
| 126 | * new queue_rq() calls occur unless the queue has been stopped first. |
| 127 | */ |
| 128 | void blk_mq_quiesce_queue(struct request_queue *q) |
| 129 | { |
| 130 | struct blk_mq_hw_ctx *hctx; |
| 131 | unsigned int i; |
| 132 | bool rcu = false; |
| 133 | |
| 134 | blk_mq_stop_hw_queues(q); |
| 135 | |
| 136 | queue_for_each_hw_ctx(q, hctx, i) { |
| 137 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
| 138 | synchronize_srcu(&hctx->queue_rq_srcu); |
| 139 | else |
| 140 | rcu = true; |
| 141 | } |
| 142 | if (rcu) |
| 143 | synchronize_rcu(); |
| 144 | } |
| 145 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue); |
| 146 | |
Jens Axboe | aed3ea9 | 2014-12-22 14:04:42 -0700 | [diff] [blame] | 147 | void blk_mq_wake_waiters(struct request_queue *q) |
| 148 | { |
| 149 | struct blk_mq_hw_ctx *hctx; |
| 150 | unsigned int i; |
| 151 | |
| 152 | queue_for_each_hw_ctx(q, hctx, i) |
| 153 | if (blk_mq_hw_queue_mapped(hctx)) |
| 154 | blk_mq_tag_wakeup_all(hctx->tags, true); |
Keith Busch | 3fd5940 | 2015-01-08 08:53:56 -0700 | [diff] [blame] | 155 | |
| 156 | /* |
| 157 | * If we are called because the queue has now been marked as |
| 158 | * dying, we need to ensure that processes currently waiting on |
| 159 | * the queue are notified as well. |
| 160 | */ |
| 161 | wake_up_all(&q->mq_freeze_wq); |
Jens Axboe | aed3ea9 | 2014-12-22 14:04:42 -0700 | [diff] [blame] | 162 | } |
| 163 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 164 | bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx) |
| 165 | { |
| 166 | return blk_mq_has_free_tags(hctx->tags); |
| 167 | } |
| 168 | EXPORT_SYMBOL(blk_mq_can_queue); |
| 169 | |
Jens Axboe | 2c3ad66 | 2016-12-14 14:34:47 -0700 | [diff] [blame] | 170 | void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx, |
| 171 | struct request *rq, unsigned int op) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 172 | { |
Christoph Hellwig | af76e55 | 2014-05-06 12:12:45 +0200 | [diff] [blame] | 173 | INIT_LIST_HEAD(&rq->queuelist); |
| 174 | /* csd/requeue_work/fifo_time is initialized before use */ |
| 175 | rq->q = q; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 176 | rq->mq_ctx = ctx; |
Christoph Hellwig | ef295ec | 2016-10-28 08:48:16 -0600 | [diff] [blame] | 177 | rq->cmd_flags = op; |
Christoph Hellwig | e806402 | 2016-10-20 15:12:13 +0200 | [diff] [blame] | 178 | if (blk_queue_io_stat(q)) |
| 179 | rq->rq_flags |= RQF_IO_STAT; |
Christoph Hellwig | af76e55 | 2014-05-06 12:12:45 +0200 | [diff] [blame] | 180 | /* do not touch atomic flags, it needs atomic ops against the timer */ |
| 181 | rq->cpu = -1; |
Christoph Hellwig | af76e55 | 2014-05-06 12:12:45 +0200 | [diff] [blame] | 182 | INIT_HLIST_NODE(&rq->hash); |
| 183 | RB_CLEAR_NODE(&rq->rb_node); |
Christoph Hellwig | af76e55 | 2014-05-06 12:12:45 +0200 | [diff] [blame] | 184 | rq->rq_disk = NULL; |
| 185 | rq->part = NULL; |
Jens Axboe | 3ee3237 | 2014-06-09 09:36:53 -0600 | [diff] [blame] | 186 | rq->start_time = jiffies; |
Christoph Hellwig | af76e55 | 2014-05-06 12:12:45 +0200 | [diff] [blame] | 187 | #ifdef CONFIG_BLK_CGROUP |
| 188 | rq->rl = NULL; |
Ming Lei | 0fec08b | 2014-01-03 10:00:08 -0700 | [diff] [blame] | 189 | set_start_time_ns(rq); |
Christoph Hellwig | af76e55 | 2014-05-06 12:12:45 +0200 | [diff] [blame] | 190 | rq->io_start_time_ns = 0; |
| 191 | #endif |
| 192 | rq->nr_phys_segments = 0; |
| 193 | #if defined(CONFIG_BLK_DEV_INTEGRITY) |
| 194 | rq->nr_integrity_segments = 0; |
| 195 | #endif |
Christoph Hellwig | af76e55 | 2014-05-06 12:12:45 +0200 | [diff] [blame] | 196 | rq->special = NULL; |
| 197 | /* tag was already set */ |
| 198 | rq->errors = 0; |
Christoph Hellwig | af76e55 | 2014-05-06 12:12:45 +0200 | [diff] [blame] | 199 | |
Tony Battersby | 6f4a1626 | 2014-08-22 15:53:39 -0400 | [diff] [blame] | 200 | rq->cmd = rq->__cmd; |
| 201 | |
Christoph Hellwig | af76e55 | 2014-05-06 12:12:45 +0200 | [diff] [blame] | 202 | rq->extra_len = 0; |
| 203 | rq->sense_len = 0; |
| 204 | rq->resid_len = 0; |
| 205 | rq->sense = NULL; |
| 206 | |
Christoph Hellwig | af76e55 | 2014-05-06 12:12:45 +0200 | [diff] [blame] | 207 | INIT_LIST_HEAD(&rq->timeout_list); |
Jens Axboe | f6be4fb | 2014-06-06 11:03:48 -0600 | [diff] [blame] | 208 | rq->timeout = 0; |
| 209 | |
Christoph Hellwig | af76e55 | 2014-05-06 12:12:45 +0200 | [diff] [blame] | 210 | rq->end_io = NULL; |
| 211 | rq->end_io_data = NULL; |
| 212 | rq->next_rq = NULL; |
| 213 | |
Christoph Hellwig | ef295ec | 2016-10-28 08:48:16 -0600 | [diff] [blame] | 214 | ctx->rq_dispatched[op_is_sync(op)]++; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 215 | } |
Jens Axboe | 2c3ad66 | 2016-12-14 14:34:47 -0700 | [diff] [blame] | 216 | EXPORT_SYMBOL_GPL(blk_mq_rq_ctx_init); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 217 | |
Jens Axboe | 2c3ad66 | 2016-12-14 14:34:47 -0700 | [diff] [blame] | 218 | struct request *__blk_mq_alloc_request(struct blk_mq_alloc_data *data, |
| 219 | unsigned int op) |
Christoph Hellwig | 5dee857 | 2014-05-27 20:59:47 +0200 | [diff] [blame] | 220 | { |
| 221 | struct request *rq; |
| 222 | unsigned int tag; |
| 223 | |
Ming Lei | cb96a42 | 2014-06-01 00:43:37 +0800 | [diff] [blame] | 224 | tag = blk_mq_get_tag(data); |
Christoph Hellwig | 5dee857 | 2014-05-27 20:59:47 +0200 | [diff] [blame] | 225 | if (tag != BLK_MQ_TAG_FAIL) { |
Ming Lei | cb96a42 | 2014-06-01 00:43:37 +0800 | [diff] [blame] | 226 | rq = data->hctx->tags->rqs[tag]; |
Christoph Hellwig | 5dee857 | 2014-05-27 20:59:47 +0200 | [diff] [blame] | 227 | |
Ming Lei | cb96a42 | 2014-06-01 00:43:37 +0800 | [diff] [blame] | 228 | if (blk_mq_tag_busy(data->hctx)) { |
Christoph Hellwig | e806402 | 2016-10-20 15:12:13 +0200 | [diff] [blame] | 229 | rq->rq_flags = RQF_MQ_INFLIGHT; |
Ming Lei | cb96a42 | 2014-06-01 00:43:37 +0800 | [diff] [blame] | 230 | atomic_inc(&data->hctx->nr_active); |
Christoph Hellwig | 5dee857 | 2014-05-27 20:59:47 +0200 | [diff] [blame] | 231 | } |
| 232 | |
| 233 | rq->tag = tag; |
Christoph Hellwig | ef295ec | 2016-10-28 08:48:16 -0600 | [diff] [blame] | 234 | blk_mq_rq_ctx_init(data->q, data->ctx, rq, op); |
Christoph Hellwig | 5dee857 | 2014-05-27 20:59:47 +0200 | [diff] [blame] | 235 | return rq; |
| 236 | } |
| 237 | |
| 238 | return NULL; |
| 239 | } |
Jens Axboe | 2c3ad66 | 2016-12-14 14:34:47 -0700 | [diff] [blame] | 240 | EXPORT_SYMBOL_GPL(__blk_mq_alloc_request); |
Christoph Hellwig | 5dee857 | 2014-05-27 20:59:47 +0200 | [diff] [blame] | 241 | |
Christoph Hellwig | 6f3b0e8 | 2015-11-26 09:13:05 +0100 | [diff] [blame] | 242 | struct request *blk_mq_alloc_request(struct request_queue *q, int rw, |
| 243 | unsigned int flags) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 244 | { |
Christoph Hellwig | d852564 | 2014-05-27 20:59:50 +0200 | [diff] [blame] | 245 | struct blk_mq_ctx *ctx; |
| 246 | struct blk_mq_hw_ctx *hctx; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 247 | struct request *rq; |
Ming Lei | cb96a42 | 2014-06-01 00:43:37 +0800 | [diff] [blame] | 248 | struct blk_mq_alloc_data alloc_data; |
Joe Lawrence | a492f07 | 2014-08-28 08:15:21 -0600 | [diff] [blame] | 249 | int ret; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 250 | |
Christoph Hellwig | 6f3b0e8 | 2015-11-26 09:13:05 +0100 | [diff] [blame] | 251 | ret = blk_queue_enter(q, flags & BLK_MQ_REQ_NOWAIT); |
Joe Lawrence | a492f07 | 2014-08-28 08:15:21 -0600 | [diff] [blame] | 252 | if (ret) |
| 253 | return ERR_PTR(ret); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 254 | |
Christoph Hellwig | d852564 | 2014-05-27 20:59:50 +0200 | [diff] [blame] | 255 | ctx = blk_mq_get_ctx(q); |
Christoph Hellwig | 7d7e0f9 | 2016-09-14 16:18:54 +0200 | [diff] [blame] | 256 | hctx = blk_mq_map_queue(q, ctx->cpu); |
Christoph Hellwig | 6f3b0e8 | 2015-11-26 09:13:05 +0100 | [diff] [blame] | 257 | blk_mq_set_alloc_data(&alloc_data, q, flags, ctx, hctx); |
Christoph Hellwig | ef295ec | 2016-10-28 08:48:16 -0600 | [diff] [blame] | 258 | rq = __blk_mq_alloc_request(&alloc_data, rw); |
Christoph Hellwig | d852564 | 2014-05-27 20:59:50 +0200 | [diff] [blame] | 259 | blk_mq_put_ctx(ctx); |
Jens Axboe | 841bac2 | 2016-09-21 10:08:43 -0600 | [diff] [blame] | 260 | |
Keith Busch | c76541a | 2014-12-19 17:54:13 -0700 | [diff] [blame] | 261 | if (!rq) { |
Dan Williams | 3ef28e8 | 2015-10-21 13:20:12 -0400 | [diff] [blame] | 262 | blk_queue_exit(q); |
Joe Lawrence | a492f07 | 2014-08-28 08:15:21 -0600 | [diff] [blame] | 263 | return ERR_PTR(-EWOULDBLOCK); |
Keith Busch | c76541a | 2014-12-19 17:54:13 -0700 | [diff] [blame] | 264 | } |
Christoph Hellwig | 0c4de0f | 2016-07-19 11:31:50 +0200 | [diff] [blame] | 265 | |
| 266 | rq->__data_len = 0; |
| 267 | rq->__sector = (sector_t) -1; |
| 268 | rq->bio = rq->biotail = NULL; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 269 | return rq; |
| 270 | } |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 271 | EXPORT_SYMBOL(blk_mq_alloc_request); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 272 | |
Ming Lin | 1f5bd33 | 2016-06-13 16:45:21 +0200 | [diff] [blame] | 273 | struct request *blk_mq_alloc_request_hctx(struct request_queue *q, int rw, |
| 274 | unsigned int flags, unsigned int hctx_idx) |
| 275 | { |
| 276 | struct blk_mq_hw_ctx *hctx; |
| 277 | struct blk_mq_ctx *ctx; |
| 278 | struct request *rq; |
| 279 | struct blk_mq_alloc_data alloc_data; |
| 280 | int ret; |
| 281 | |
| 282 | /* |
| 283 | * If the tag allocator sleeps we could get an allocation for a |
| 284 | * different hardware context. No need to complicate the low level |
| 285 | * allocator for this for the rare use case of a command tied to |
| 286 | * a specific queue. |
| 287 | */ |
| 288 | if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT))) |
| 289 | return ERR_PTR(-EINVAL); |
| 290 | |
| 291 | if (hctx_idx >= q->nr_hw_queues) |
| 292 | return ERR_PTR(-EIO); |
| 293 | |
| 294 | ret = blk_queue_enter(q, true); |
| 295 | if (ret) |
| 296 | return ERR_PTR(ret); |
| 297 | |
Christoph Hellwig | c8712c6 | 2016-09-23 10:25:48 -0600 | [diff] [blame] | 298 | /* |
| 299 | * Check if the hardware context is actually mapped to anything. |
| 300 | * If not tell the caller that it should skip this queue. |
| 301 | */ |
Ming Lin | 1f5bd33 | 2016-06-13 16:45:21 +0200 | [diff] [blame] | 302 | hctx = q->queue_hw_ctx[hctx_idx]; |
Christoph Hellwig | c8712c6 | 2016-09-23 10:25:48 -0600 | [diff] [blame] | 303 | if (!blk_mq_hw_queue_mapped(hctx)) { |
| 304 | ret = -EXDEV; |
| 305 | goto out_queue_exit; |
| 306 | } |
Ming Lin | 1f5bd33 | 2016-06-13 16:45:21 +0200 | [diff] [blame] | 307 | ctx = __blk_mq_get_ctx(q, cpumask_first(hctx->cpumask)); |
| 308 | |
| 309 | blk_mq_set_alloc_data(&alloc_data, q, flags, ctx, hctx); |
Christoph Hellwig | ef295ec | 2016-10-28 08:48:16 -0600 | [diff] [blame] | 310 | rq = __blk_mq_alloc_request(&alloc_data, rw); |
Ming Lin | 1f5bd33 | 2016-06-13 16:45:21 +0200 | [diff] [blame] | 311 | if (!rq) { |
Christoph Hellwig | c8712c6 | 2016-09-23 10:25:48 -0600 | [diff] [blame] | 312 | ret = -EWOULDBLOCK; |
| 313 | goto out_queue_exit; |
Ming Lin | 1f5bd33 | 2016-06-13 16:45:21 +0200 | [diff] [blame] | 314 | } |
| 315 | |
| 316 | return rq; |
Christoph Hellwig | c8712c6 | 2016-09-23 10:25:48 -0600 | [diff] [blame] | 317 | |
| 318 | out_queue_exit: |
| 319 | blk_queue_exit(q); |
| 320 | return ERR_PTR(ret); |
Ming Lin | 1f5bd33 | 2016-06-13 16:45:21 +0200 | [diff] [blame] | 321 | } |
| 322 | EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx); |
| 323 | |
Jens Axboe | 2c3ad66 | 2016-12-14 14:34:47 -0700 | [diff] [blame] | 324 | void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx, |
| 325 | struct request *rq) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 326 | { |
| 327 | const int tag = rq->tag; |
| 328 | struct request_queue *q = rq->q; |
| 329 | |
Christoph Hellwig | e806402 | 2016-10-20 15:12:13 +0200 | [diff] [blame] | 330 | if (rq->rq_flags & RQF_MQ_INFLIGHT) |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 331 | atomic_dec(&hctx->nr_active); |
Jens Axboe | 87760e5 | 2016-11-09 12:38:14 -0700 | [diff] [blame] | 332 | |
| 333 | wbt_done(q->rq_wb, &rq->issue_stat); |
Christoph Hellwig | e806402 | 2016-10-20 15:12:13 +0200 | [diff] [blame] | 334 | rq->rq_flags = 0; |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 335 | |
Christoph Hellwig | af76e55 | 2014-05-06 12:12:45 +0200 | [diff] [blame] | 336 | clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags); |
Jens Axboe | 06426ad | 2016-11-14 13:01:59 -0700 | [diff] [blame] | 337 | clear_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags); |
Jens Axboe | 4941115 | 2017-01-13 08:09:05 -0700 | [diff] [blame] | 338 | blk_mq_put_tag(hctx, hctx->tags, ctx, tag); |
Dan Williams | 3ef28e8 | 2015-10-21 13:20:12 -0400 | [diff] [blame] | 339 | blk_queue_exit(q); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 340 | } |
| 341 | |
Jens Axboe | 16a3c2a | 2016-12-15 14:27:46 -0700 | [diff] [blame] | 342 | static void blk_mq_free_hctx_request(struct blk_mq_hw_ctx *hctx, |
| 343 | struct request *rq) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 344 | { |
| 345 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
Jens Axboe | 7c7f2f2 | 2014-11-17 10:41:57 -0700 | [diff] [blame] | 346 | |
| 347 | ctx->rq_completed[rq_is_sync(rq)]++; |
| 348 | __blk_mq_free_request(hctx, ctx, rq); |
Jens Axboe | 7c7f2f2 | 2014-11-17 10:41:57 -0700 | [diff] [blame] | 349 | } |
Jens Axboe | 7c7f2f2 | 2014-11-17 10:41:57 -0700 | [diff] [blame] | 350 | |
| 351 | void blk_mq_free_request(struct request *rq) |
| 352 | { |
Christoph Hellwig | 7d7e0f9 | 2016-09-14 16:18:54 +0200 | [diff] [blame] | 353 | blk_mq_free_hctx_request(blk_mq_map_queue(rq->q, rq->mq_ctx->cpu), rq); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 354 | } |
Jens Axboe | 1a3b595 | 2014-11-17 10:40:48 -0700 | [diff] [blame] | 355 | EXPORT_SYMBOL_GPL(blk_mq_free_request); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 356 | |
Christoph Hellwig | c8a446a | 2014-09-13 16:40:10 -0700 | [diff] [blame] | 357 | inline void __blk_mq_end_request(struct request *rq, int error) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 358 | { |
Ming Lei | 0d11e6a | 2013-12-05 10:50:39 -0700 | [diff] [blame] | 359 | blk_account_io_done(rq); |
| 360 | |
Christoph Hellwig | 91b6363 | 2014-04-16 09:44:53 +0200 | [diff] [blame] | 361 | if (rq->end_io) { |
Jens Axboe | 87760e5 | 2016-11-09 12:38:14 -0700 | [diff] [blame] | 362 | wbt_done(rq->q->rq_wb, &rq->issue_stat); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 363 | rq->end_io(rq, error); |
Christoph Hellwig | 91b6363 | 2014-04-16 09:44:53 +0200 | [diff] [blame] | 364 | } else { |
| 365 | if (unlikely(blk_bidi_rq(rq))) |
| 366 | blk_mq_free_request(rq->next_rq); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 367 | blk_mq_free_request(rq); |
Christoph Hellwig | 91b6363 | 2014-04-16 09:44:53 +0200 | [diff] [blame] | 368 | } |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 369 | } |
Christoph Hellwig | c8a446a | 2014-09-13 16:40:10 -0700 | [diff] [blame] | 370 | EXPORT_SYMBOL(__blk_mq_end_request); |
Christoph Hellwig | 63151a4 | 2014-04-16 09:44:52 +0200 | [diff] [blame] | 371 | |
Christoph Hellwig | c8a446a | 2014-09-13 16:40:10 -0700 | [diff] [blame] | 372 | void blk_mq_end_request(struct request *rq, int error) |
Christoph Hellwig | 63151a4 | 2014-04-16 09:44:52 +0200 | [diff] [blame] | 373 | { |
| 374 | if (blk_update_request(rq, error, blk_rq_bytes(rq))) |
| 375 | BUG(); |
Christoph Hellwig | c8a446a | 2014-09-13 16:40:10 -0700 | [diff] [blame] | 376 | __blk_mq_end_request(rq, error); |
Christoph Hellwig | 63151a4 | 2014-04-16 09:44:52 +0200 | [diff] [blame] | 377 | } |
Christoph Hellwig | c8a446a | 2014-09-13 16:40:10 -0700 | [diff] [blame] | 378 | EXPORT_SYMBOL(blk_mq_end_request); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 379 | |
Christoph Hellwig | 30a91cb | 2014-02-10 03:24:38 -0800 | [diff] [blame] | 380 | static void __blk_mq_complete_request_remote(void *data) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 381 | { |
Christoph Hellwig | 3d6efbf | 2014-01-08 09:33:37 -0800 | [diff] [blame] | 382 | struct request *rq = data; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 383 | |
Christoph Hellwig | 30a91cb | 2014-02-10 03:24:38 -0800 | [diff] [blame] | 384 | rq->q->softirq_done_fn(rq); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 385 | } |
| 386 | |
Jens Axboe | ed85186 | 2014-05-30 21:20:50 -0600 | [diff] [blame] | 387 | static void blk_mq_ipi_complete_request(struct request *rq) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 388 | { |
| 389 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
Christoph Hellwig | 3853520 | 2014-04-25 02:32:53 -0700 | [diff] [blame] | 390 | bool shared = false; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 391 | int cpu; |
| 392 | |
Christoph Hellwig | 3853520 | 2014-04-25 02:32:53 -0700 | [diff] [blame] | 393 | if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) { |
Christoph Hellwig | 30a91cb | 2014-02-10 03:24:38 -0800 | [diff] [blame] | 394 | rq->q->softirq_done_fn(rq); |
| 395 | return; |
| 396 | } |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 397 | |
| 398 | cpu = get_cpu(); |
Christoph Hellwig | 3853520 | 2014-04-25 02:32:53 -0700 | [diff] [blame] | 399 | if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags)) |
| 400 | shared = cpus_share_cache(cpu, ctx->cpu); |
| 401 | |
| 402 | if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) { |
Christoph Hellwig | 30a91cb | 2014-02-10 03:24:38 -0800 | [diff] [blame] | 403 | rq->csd.func = __blk_mq_complete_request_remote; |
Christoph Hellwig | 3d6efbf | 2014-01-08 09:33:37 -0800 | [diff] [blame] | 404 | rq->csd.info = rq; |
| 405 | rq->csd.flags = 0; |
Frederic Weisbecker | c46fff2 | 2014-02-24 16:40:02 +0100 | [diff] [blame] | 406 | smp_call_function_single_async(ctx->cpu, &rq->csd); |
Christoph Hellwig | 3d6efbf | 2014-01-08 09:33:37 -0800 | [diff] [blame] | 407 | } else { |
Christoph Hellwig | 30a91cb | 2014-02-10 03:24:38 -0800 | [diff] [blame] | 408 | rq->q->softirq_done_fn(rq); |
Christoph Hellwig | 3d6efbf | 2014-01-08 09:33:37 -0800 | [diff] [blame] | 409 | } |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 410 | put_cpu(); |
| 411 | } |
Christoph Hellwig | 30a91cb | 2014-02-10 03:24:38 -0800 | [diff] [blame] | 412 | |
Jens Axboe | cf43e6b | 2016-11-07 21:32:37 -0700 | [diff] [blame] | 413 | static void blk_mq_stat_add(struct request *rq) |
| 414 | { |
| 415 | if (rq->rq_flags & RQF_STATS) { |
| 416 | /* |
| 417 | * We could rq->mq_ctx here, but there's less of a risk |
| 418 | * of races if we have the completion event add the stats |
| 419 | * to the local software queue. |
| 420 | */ |
| 421 | struct blk_mq_ctx *ctx; |
| 422 | |
| 423 | ctx = __blk_mq_get_ctx(rq->q, raw_smp_processor_id()); |
| 424 | blk_stat_add(&ctx->stat[rq_data_dir(rq)], rq); |
| 425 | } |
| 426 | } |
| 427 | |
Jens Axboe | 1fa8cc5 | 2015-11-05 14:32:55 -0700 | [diff] [blame] | 428 | static void __blk_mq_complete_request(struct request *rq) |
Jens Axboe | ed85186 | 2014-05-30 21:20:50 -0600 | [diff] [blame] | 429 | { |
| 430 | struct request_queue *q = rq->q; |
| 431 | |
Jens Axboe | cf43e6b | 2016-11-07 21:32:37 -0700 | [diff] [blame] | 432 | blk_mq_stat_add(rq); |
| 433 | |
Jens Axboe | ed85186 | 2014-05-30 21:20:50 -0600 | [diff] [blame] | 434 | if (!q->softirq_done_fn) |
Christoph Hellwig | c8a446a | 2014-09-13 16:40:10 -0700 | [diff] [blame] | 435 | blk_mq_end_request(rq, rq->errors); |
Jens Axboe | ed85186 | 2014-05-30 21:20:50 -0600 | [diff] [blame] | 436 | else |
| 437 | blk_mq_ipi_complete_request(rq); |
| 438 | } |
| 439 | |
Christoph Hellwig | 30a91cb | 2014-02-10 03:24:38 -0800 | [diff] [blame] | 440 | /** |
| 441 | * blk_mq_complete_request - end I/O on a request |
| 442 | * @rq: the request being processed |
| 443 | * |
| 444 | * Description: |
| 445 | * Ends all I/O on a request. It does not handle partial completions. |
| 446 | * The actual completion happens out-of-order, through a IPI handler. |
| 447 | **/ |
Christoph Hellwig | f4829a9 | 2015-09-27 21:01:50 +0200 | [diff] [blame] | 448 | void blk_mq_complete_request(struct request *rq, int error) |
Christoph Hellwig | 30a91cb | 2014-02-10 03:24:38 -0800 | [diff] [blame] | 449 | { |
Jens Axboe | 95f0968 | 2014-05-27 17:46:48 -0600 | [diff] [blame] | 450 | struct request_queue *q = rq->q; |
| 451 | |
| 452 | if (unlikely(blk_should_fake_timeout(q))) |
Christoph Hellwig | 30a91cb | 2014-02-10 03:24:38 -0800 | [diff] [blame] | 453 | return; |
Christoph Hellwig | f4829a9 | 2015-09-27 21:01:50 +0200 | [diff] [blame] | 454 | if (!blk_mark_rq_complete(rq)) { |
| 455 | rq->errors = error; |
Jens Axboe | ed85186 | 2014-05-30 21:20:50 -0600 | [diff] [blame] | 456 | __blk_mq_complete_request(rq); |
Christoph Hellwig | f4829a9 | 2015-09-27 21:01:50 +0200 | [diff] [blame] | 457 | } |
Christoph Hellwig | 30a91cb | 2014-02-10 03:24:38 -0800 | [diff] [blame] | 458 | } |
| 459 | EXPORT_SYMBOL(blk_mq_complete_request); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 460 | |
Keith Busch | 973c019 | 2015-01-07 18:55:43 -0700 | [diff] [blame] | 461 | int blk_mq_request_started(struct request *rq) |
| 462 | { |
| 463 | return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags); |
| 464 | } |
| 465 | EXPORT_SYMBOL_GPL(blk_mq_request_started); |
| 466 | |
Christoph Hellwig | e249007 | 2014-09-13 16:40:09 -0700 | [diff] [blame] | 467 | void blk_mq_start_request(struct request *rq) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 468 | { |
| 469 | struct request_queue *q = rq->q; |
| 470 | |
| 471 | trace_block_rq_issue(q, rq); |
| 472 | |
Christoph Hellwig | 742ee69 | 2014-04-14 10:30:06 +0200 | [diff] [blame] | 473 | rq->resid_len = blk_rq_bytes(rq); |
Christoph Hellwig | 91b6363 | 2014-04-16 09:44:53 +0200 | [diff] [blame] | 474 | if (unlikely(blk_bidi_rq(rq))) |
| 475 | rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq); |
Christoph Hellwig | 742ee69 | 2014-04-14 10:30:06 +0200 | [diff] [blame] | 476 | |
Jens Axboe | cf43e6b | 2016-11-07 21:32:37 -0700 | [diff] [blame] | 477 | if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) { |
| 478 | blk_stat_set_issue_time(&rq->issue_stat); |
| 479 | rq->rq_flags |= RQF_STATS; |
Jens Axboe | 87760e5 | 2016-11-09 12:38:14 -0700 | [diff] [blame] | 480 | wbt_issue(q->rq_wb, &rq->issue_stat); |
Jens Axboe | cf43e6b | 2016-11-07 21:32:37 -0700 | [diff] [blame] | 481 | } |
| 482 | |
Ming Lei | 2b8393b | 2014-06-10 00:16:41 +0800 | [diff] [blame] | 483 | blk_add_timer(rq); |
Jens Axboe | 87ee7b1 | 2014-04-24 08:51:47 -0600 | [diff] [blame] | 484 | |
| 485 | /* |
Jens Axboe | 538b753 | 2014-09-16 10:37:37 -0600 | [diff] [blame] | 486 | * Ensure that ->deadline is visible before set the started |
| 487 | * flag and clear the completed flag. |
| 488 | */ |
| 489 | smp_mb__before_atomic(); |
| 490 | |
| 491 | /* |
Jens Axboe | 87ee7b1 | 2014-04-24 08:51:47 -0600 | [diff] [blame] | 492 | * Mark us as started and clear complete. Complete might have been |
| 493 | * set if requeue raced with timeout, which then marked it as |
| 494 | * complete. So be sure to clear complete again when we start |
| 495 | * the request, otherwise we'll ignore the completion event. |
| 496 | */ |
Jens Axboe | 4b57052 | 2014-05-29 11:00:11 -0600 | [diff] [blame] | 497 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) |
| 498 | set_bit(REQ_ATOM_STARTED, &rq->atomic_flags); |
| 499 | if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags)) |
| 500 | clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags); |
Christoph Hellwig | 49f5baa | 2014-02-11 08:27:14 -0800 | [diff] [blame] | 501 | |
| 502 | if (q->dma_drain_size && blk_rq_bytes(rq)) { |
| 503 | /* |
| 504 | * Make sure space for the drain appears. We know we can do |
| 505 | * this because max_hw_segments has been adjusted to be one |
| 506 | * fewer than the device can handle. |
| 507 | */ |
| 508 | rq->nr_phys_segments++; |
| 509 | } |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 510 | } |
Christoph Hellwig | e249007 | 2014-09-13 16:40:09 -0700 | [diff] [blame] | 511 | EXPORT_SYMBOL(blk_mq_start_request); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 512 | |
Christoph Hellwig | ed0791b | 2014-04-16 09:44:57 +0200 | [diff] [blame] | 513 | static void __blk_mq_requeue_request(struct request *rq) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 514 | { |
| 515 | struct request_queue *q = rq->q; |
| 516 | |
| 517 | trace_block_rq_requeue(q, rq); |
Jens Axboe | 87760e5 | 2016-11-09 12:38:14 -0700 | [diff] [blame] | 518 | wbt_requeue(q->rq_wb, &rq->issue_stat); |
Christoph Hellwig | 49f5baa | 2014-02-11 08:27:14 -0800 | [diff] [blame] | 519 | |
Christoph Hellwig | e249007 | 2014-09-13 16:40:09 -0700 | [diff] [blame] | 520 | if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) { |
| 521 | if (q->dma_drain_size && blk_rq_bytes(rq)) |
| 522 | rq->nr_phys_segments--; |
| 523 | } |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 524 | } |
| 525 | |
Bart Van Assche | 2b053ac | 2016-10-28 17:21:41 -0700 | [diff] [blame] | 526 | void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list) |
Christoph Hellwig | ed0791b | 2014-04-16 09:44:57 +0200 | [diff] [blame] | 527 | { |
Christoph Hellwig | ed0791b | 2014-04-16 09:44:57 +0200 | [diff] [blame] | 528 | __blk_mq_requeue_request(rq); |
Christoph Hellwig | ed0791b | 2014-04-16 09:44:57 +0200 | [diff] [blame] | 529 | |
Christoph Hellwig | ed0791b | 2014-04-16 09:44:57 +0200 | [diff] [blame] | 530 | BUG_ON(blk_queued_rq(rq)); |
Bart Van Assche | 2b053ac | 2016-10-28 17:21:41 -0700 | [diff] [blame] | 531 | blk_mq_add_to_requeue_list(rq, true, kick_requeue_list); |
Christoph Hellwig | ed0791b | 2014-04-16 09:44:57 +0200 | [diff] [blame] | 532 | } |
| 533 | EXPORT_SYMBOL(blk_mq_requeue_request); |
| 534 | |
Christoph Hellwig | 6fca6a6 | 2014-05-28 08:08:02 -0600 | [diff] [blame] | 535 | static void blk_mq_requeue_work(struct work_struct *work) |
| 536 | { |
| 537 | struct request_queue *q = |
Mike Snitzer | 2849450 | 2016-09-14 13:28:30 -0400 | [diff] [blame] | 538 | container_of(work, struct request_queue, requeue_work.work); |
Christoph Hellwig | 6fca6a6 | 2014-05-28 08:08:02 -0600 | [diff] [blame] | 539 | LIST_HEAD(rq_list); |
| 540 | struct request *rq, *next; |
| 541 | unsigned long flags; |
| 542 | |
| 543 | spin_lock_irqsave(&q->requeue_lock, flags); |
| 544 | list_splice_init(&q->requeue_list, &rq_list); |
| 545 | spin_unlock_irqrestore(&q->requeue_lock, flags); |
| 546 | |
| 547 | list_for_each_entry_safe(rq, next, &rq_list, queuelist) { |
Christoph Hellwig | e806402 | 2016-10-20 15:12:13 +0200 | [diff] [blame] | 548 | if (!(rq->rq_flags & RQF_SOFTBARRIER)) |
Christoph Hellwig | 6fca6a6 | 2014-05-28 08:08:02 -0600 | [diff] [blame] | 549 | continue; |
| 550 | |
Christoph Hellwig | e806402 | 2016-10-20 15:12:13 +0200 | [diff] [blame] | 551 | rq->rq_flags &= ~RQF_SOFTBARRIER; |
Christoph Hellwig | 6fca6a6 | 2014-05-28 08:08:02 -0600 | [diff] [blame] | 552 | list_del_init(&rq->queuelist); |
| 553 | blk_mq_insert_request(rq, true, false, false); |
| 554 | } |
| 555 | |
| 556 | while (!list_empty(&rq_list)) { |
| 557 | rq = list_entry(rq_list.next, struct request, queuelist); |
| 558 | list_del_init(&rq->queuelist); |
| 559 | blk_mq_insert_request(rq, false, false, false); |
| 560 | } |
| 561 | |
Bart Van Assche | 52d7f1b | 2016-10-28 17:20:32 -0700 | [diff] [blame] | 562 | blk_mq_run_hw_queues(q, false); |
Christoph Hellwig | 6fca6a6 | 2014-05-28 08:08:02 -0600 | [diff] [blame] | 563 | } |
| 564 | |
Bart Van Assche | 2b053ac | 2016-10-28 17:21:41 -0700 | [diff] [blame] | 565 | void blk_mq_add_to_requeue_list(struct request *rq, bool at_head, |
| 566 | bool kick_requeue_list) |
Christoph Hellwig | 6fca6a6 | 2014-05-28 08:08:02 -0600 | [diff] [blame] | 567 | { |
| 568 | struct request_queue *q = rq->q; |
| 569 | unsigned long flags; |
| 570 | |
| 571 | /* |
| 572 | * We abuse this flag that is otherwise used by the I/O scheduler to |
| 573 | * request head insertation from the workqueue. |
| 574 | */ |
Christoph Hellwig | e806402 | 2016-10-20 15:12:13 +0200 | [diff] [blame] | 575 | BUG_ON(rq->rq_flags & RQF_SOFTBARRIER); |
Christoph Hellwig | 6fca6a6 | 2014-05-28 08:08:02 -0600 | [diff] [blame] | 576 | |
| 577 | spin_lock_irqsave(&q->requeue_lock, flags); |
| 578 | if (at_head) { |
Christoph Hellwig | e806402 | 2016-10-20 15:12:13 +0200 | [diff] [blame] | 579 | rq->rq_flags |= RQF_SOFTBARRIER; |
Christoph Hellwig | 6fca6a6 | 2014-05-28 08:08:02 -0600 | [diff] [blame] | 580 | list_add(&rq->queuelist, &q->requeue_list); |
| 581 | } else { |
| 582 | list_add_tail(&rq->queuelist, &q->requeue_list); |
| 583 | } |
| 584 | spin_unlock_irqrestore(&q->requeue_lock, flags); |
Bart Van Assche | 2b053ac | 2016-10-28 17:21:41 -0700 | [diff] [blame] | 585 | |
| 586 | if (kick_requeue_list) |
| 587 | blk_mq_kick_requeue_list(q); |
Christoph Hellwig | 6fca6a6 | 2014-05-28 08:08:02 -0600 | [diff] [blame] | 588 | } |
| 589 | EXPORT_SYMBOL(blk_mq_add_to_requeue_list); |
| 590 | |
| 591 | void blk_mq_kick_requeue_list(struct request_queue *q) |
| 592 | { |
Mike Snitzer | 2849450 | 2016-09-14 13:28:30 -0400 | [diff] [blame] | 593 | kblockd_schedule_delayed_work(&q->requeue_work, 0); |
Christoph Hellwig | 6fca6a6 | 2014-05-28 08:08:02 -0600 | [diff] [blame] | 594 | } |
| 595 | EXPORT_SYMBOL(blk_mq_kick_requeue_list); |
| 596 | |
Mike Snitzer | 2849450 | 2016-09-14 13:28:30 -0400 | [diff] [blame] | 597 | void blk_mq_delay_kick_requeue_list(struct request_queue *q, |
| 598 | unsigned long msecs) |
| 599 | { |
| 600 | kblockd_schedule_delayed_work(&q->requeue_work, |
| 601 | msecs_to_jiffies(msecs)); |
| 602 | } |
| 603 | EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list); |
| 604 | |
Jens Axboe | 1885b24 | 2015-01-07 18:55:45 -0700 | [diff] [blame] | 605 | void blk_mq_abort_requeue_list(struct request_queue *q) |
| 606 | { |
| 607 | unsigned long flags; |
| 608 | LIST_HEAD(rq_list); |
| 609 | |
| 610 | spin_lock_irqsave(&q->requeue_lock, flags); |
| 611 | list_splice_init(&q->requeue_list, &rq_list); |
| 612 | spin_unlock_irqrestore(&q->requeue_lock, flags); |
| 613 | |
| 614 | while (!list_empty(&rq_list)) { |
| 615 | struct request *rq; |
| 616 | |
| 617 | rq = list_first_entry(&rq_list, struct request, queuelist); |
| 618 | list_del_init(&rq->queuelist); |
| 619 | rq->errors = -EIO; |
| 620 | blk_mq_end_request(rq, rq->errors); |
| 621 | } |
| 622 | } |
| 623 | EXPORT_SYMBOL(blk_mq_abort_requeue_list); |
| 624 | |
Jens Axboe | 0e62f51 | 2014-06-04 10:23:49 -0600 | [diff] [blame] | 625 | struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag) |
| 626 | { |
Jens Axboe | 88c7b2b | 2016-08-25 08:07:30 -0600 | [diff] [blame] | 627 | if (tag < tags->nr_tags) { |
| 628 | prefetch(tags->rqs[tag]); |
Hannes Reinecke | 4ee86ba | 2016-03-15 12:03:28 -0700 | [diff] [blame] | 629 | return tags->rqs[tag]; |
Jens Axboe | 88c7b2b | 2016-08-25 08:07:30 -0600 | [diff] [blame] | 630 | } |
Hannes Reinecke | 4ee86ba | 2016-03-15 12:03:28 -0700 | [diff] [blame] | 631 | |
| 632 | return NULL; |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 633 | } |
| 634 | EXPORT_SYMBOL(blk_mq_tag_to_rq); |
| 635 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 636 | struct blk_mq_timeout_data { |
Christoph Hellwig | 46f92d4 | 2014-09-13 16:40:12 -0700 | [diff] [blame] | 637 | unsigned long next; |
| 638 | unsigned int next_set; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 639 | }; |
| 640 | |
Christoph Hellwig | 9041583 | 2014-09-22 10:21:48 -0600 | [diff] [blame] | 641 | void blk_mq_rq_timed_out(struct request *req, bool reserved) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 642 | { |
Jens Axboe | f8a5b12 | 2016-12-13 09:24:51 -0700 | [diff] [blame] | 643 | const struct blk_mq_ops *ops = req->q->mq_ops; |
Christoph Hellwig | 46f92d4 | 2014-09-13 16:40:12 -0700 | [diff] [blame] | 644 | enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER; |
Jens Axboe | 87ee7b1 | 2014-04-24 08:51:47 -0600 | [diff] [blame] | 645 | |
| 646 | /* |
| 647 | * We know that complete is set at this point. If STARTED isn't set |
| 648 | * anymore, then the request isn't active and the "timeout" should |
| 649 | * just be ignored. This can happen due to the bitflag ordering. |
| 650 | * Timeout first checks if STARTED is set, and if it is, assumes |
| 651 | * the request is active. But if we race with completion, then |
| 652 | * we both flags will get cleared. So check here again, and ignore |
| 653 | * a timeout event with a request that isn't active. |
| 654 | */ |
Christoph Hellwig | 46f92d4 | 2014-09-13 16:40:12 -0700 | [diff] [blame] | 655 | if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags)) |
| 656 | return; |
Jens Axboe | 87ee7b1 | 2014-04-24 08:51:47 -0600 | [diff] [blame] | 657 | |
Christoph Hellwig | 46f92d4 | 2014-09-13 16:40:12 -0700 | [diff] [blame] | 658 | if (ops->timeout) |
Christoph Hellwig | 0152fb6 | 2014-09-13 16:40:13 -0700 | [diff] [blame] | 659 | ret = ops->timeout(req, reserved); |
Jens Axboe | 87ee7b1 | 2014-04-24 08:51:47 -0600 | [diff] [blame] | 660 | |
Christoph Hellwig | 46f92d4 | 2014-09-13 16:40:12 -0700 | [diff] [blame] | 661 | switch (ret) { |
| 662 | case BLK_EH_HANDLED: |
| 663 | __blk_mq_complete_request(req); |
| 664 | break; |
| 665 | case BLK_EH_RESET_TIMER: |
| 666 | blk_add_timer(req); |
| 667 | blk_clear_rq_complete(req); |
| 668 | break; |
| 669 | case BLK_EH_NOT_HANDLED: |
| 670 | break; |
| 671 | default: |
| 672 | printk(KERN_ERR "block: bad eh return: %d\n", ret); |
| 673 | break; |
| 674 | } |
Jens Axboe | 87ee7b1 | 2014-04-24 08:51:47 -0600 | [diff] [blame] | 675 | } |
Keith Busch | 5b3f25f | 2015-01-07 18:55:46 -0700 | [diff] [blame] | 676 | |
Christoph Hellwig | 81481eb | 2014-09-13 16:40:11 -0700 | [diff] [blame] | 677 | static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx, |
| 678 | struct request *rq, void *priv, bool reserved) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 679 | { |
Christoph Hellwig | 81481eb | 2014-09-13 16:40:11 -0700 | [diff] [blame] | 680 | struct blk_mq_timeout_data *data = priv; |
| 681 | |
Keith Busch | eb130db | 2015-01-08 08:59:53 -0700 | [diff] [blame] | 682 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) { |
| 683 | /* |
| 684 | * If a request wasn't started before the queue was |
| 685 | * marked dying, kill it here or it'll go unnoticed. |
| 686 | */ |
Keith Busch | a59e0f5 | 2016-02-11 13:05:38 -0700 | [diff] [blame] | 687 | if (unlikely(blk_queue_dying(rq->q))) { |
| 688 | rq->errors = -EIO; |
| 689 | blk_mq_end_request(rq, rq->errors); |
| 690 | } |
Christoph Hellwig | 46f92d4 | 2014-09-13 16:40:12 -0700 | [diff] [blame] | 691 | return; |
Keith Busch | eb130db | 2015-01-08 08:59:53 -0700 | [diff] [blame] | 692 | } |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 693 | |
Christoph Hellwig | 46f92d4 | 2014-09-13 16:40:12 -0700 | [diff] [blame] | 694 | if (time_after_eq(jiffies, rq->deadline)) { |
| 695 | if (!blk_mark_rq_complete(rq)) |
Christoph Hellwig | 0152fb6 | 2014-09-13 16:40:13 -0700 | [diff] [blame] | 696 | blk_mq_rq_timed_out(rq, reserved); |
Christoph Hellwig | 46f92d4 | 2014-09-13 16:40:12 -0700 | [diff] [blame] | 697 | } else if (!data->next_set || time_after(data->next, rq->deadline)) { |
| 698 | data->next = rq->deadline; |
| 699 | data->next_set = 1; |
| 700 | } |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 701 | } |
| 702 | |
Christoph Hellwig | 287922eb | 2015-10-30 20:57:30 +0800 | [diff] [blame] | 703 | static void blk_mq_timeout_work(struct work_struct *work) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 704 | { |
Christoph Hellwig | 287922eb | 2015-10-30 20:57:30 +0800 | [diff] [blame] | 705 | struct request_queue *q = |
| 706 | container_of(work, struct request_queue, timeout_work); |
Christoph Hellwig | 81481eb | 2014-09-13 16:40:11 -0700 | [diff] [blame] | 707 | struct blk_mq_timeout_data data = { |
| 708 | .next = 0, |
| 709 | .next_set = 0, |
| 710 | }; |
Christoph Hellwig | 81481eb | 2014-09-13 16:40:11 -0700 | [diff] [blame] | 711 | int i; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 712 | |
Gabriel Krisman Bertazi | 71f79fb | 2016-08-01 08:23:39 -0600 | [diff] [blame] | 713 | /* A deadlock might occur if a request is stuck requiring a |
| 714 | * timeout at the same time a queue freeze is waiting |
| 715 | * completion, since the timeout code would not be able to |
| 716 | * acquire the queue reference here. |
| 717 | * |
| 718 | * That's why we don't use blk_queue_enter here; instead, we use |
| 719 | * percpu_ref_tryget directly, because we need to be able to |
| 720 | * obtain a reference even in the short window between the queue |
| 721 | * starting to freeze, by dropping the first reference in |
| 722 | * blk_mq_freeze_queue_start, and the moment the last request is |
| 723 | * consumed, marked by the instant q_usage_counter reaches |
| 724 | * zero. |
| 725 | */ |
| 726 | if (!percpu_ref_tryget(&q->q_usage_counter)) |
Christoph Hellwig | 287922eb | 2015-10-30 20:57:30 +0800 | [diff] [blame] | 727 | return; |
| 728 | |
Christoph Hellwig | 0bf6cd5 | 2015-09-27 21:01:51 +0200 | [diff] [blame] | 729 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 730 | |
Christoph Hellwig | 81481eb | 2014-09-13 16:40:11 -0700 | [diff] [blame] | 731 | if (data.next_set) { |
| 732 | data.next = blk_rq_timeout(round_jiffies_up(data.next)); |
| 733 | mod_timer(&q->timeout, data.next); |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 734 | } else { |
Christoph Hellwig | 0bf6cd5 | 2015-09-27 21:01:51 +0200 | [diff] [blame] | 735 | struct blk_mq_hw_ctx *hctx; |
| 736 | |
Ming Lei | f054b56 | 2015-04-21 10:00:19 +0800 | [diff] [blame] | 737 | queue_for_each_hw_ctx(q, hctx, i) { |
| 738 | /* the hctx may be unmapped, so check it here */ |
| 739 | if (blk_mq_hw_queue_mapped(hctx)) |
| 740 | blk_mq_tag_idle(hctx); |
| 741 | } |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 742 | } |
Christoph Hellwig | 287922eb | 2015-10-30 20:57:30 +0800 | [diff] [blame] | 743 | blk_queue_exit(q); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 744 | } |
| 745 | |
| 746 | /* |
| 747 | * Reverse check our software queue for entries that we could potentially |
| 748 | * merge with. Currently includes a hand-wavy stop count of 8, to not spend |
| 749 | * too much time checking for merges. |
| 750 | */ |
| 751 | static bool blk_mq_attempt_merge(struct request_queue *q, |
| 752 | struct blk_mq_ctx *ctx, struct bio *bio) |
| 753 | { |
| 754 | struct request *rq; |
| 755 | int checked = 8; |
| 756 | |
| 757 | list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) { |
| 758 | int el_ret; |
| 759 | |
| 760 | if (!checked--) |
| 761 | break; |
| 762 | |
| 763 | if (!blk_rq_merge_ok(rq, bio)) |
| 764 | continue; |
| 765 | |
| 766 | el_ret = blk_try_merge(rq, bio); |
| 767 | if (el_ret == ELEVATOR_BACK_MERGE) { |
| 768 | if (bio_attempt_back_merge(q, rq, bio)) { |
| 769 | ctx->rq_merged++; |
| 770 | return true; |
| 771 | } |
| 772 | break; |
| 773 | } else if (el_ret == ELEVATOR_FRONT_MERGE) { |
| 774 | if (bio_attempt_front_merge(q, rq, bio)) { |
| 775 | ctx->rq_merged++; |
| 776 | return true; |
| 777 | } |
| 778 | break; |
| 779 | } |
| 780 | } |
| 781 | |
| 782 | return false; |
| 783 | } |
| 784 | |
Omar Sandoval | 8845964 | 2016-09-17 08:38:44 -0600 | [diff] [blame] | 785 | struct flush_busy_ctx_data { |
| 786 | struct blk_mq_hw_ctx *hctx; |
| 787 | struct list_head *list; |
| 788 | }; |
| 789 | |
| 790 | static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data) |
| 791 | { |
| 792 | struct flush_busy_ctx_data *flush_data = data; |
| 793 | struct blk_mq_hw_ctx *hctx = flush_data->hctx; |
| 794 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; |
| 795 | |
| 796 | sbitmap_clear_bit(sb, bitnr); |
| 797 | spin_lock(&ctx->lock); |
| 798 | list_splice_tail_init(&ctx->rq_list, flush_data->list); |
| 799 | spin_unlock(&ctx->lock); |
| 800 | return true; |
| 801 | } |
| 802 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 803 | /* |
Jens Axboe | 1429d7c | 2014-05-19 09:23:55 -0600 | [diff] [blame] | 804 | * Process software queues that have been marked busy, splicing them |
| 805 | * to the for-dispatch |
| 806 | */ |
Jens Axboe | 2c3ad66 | 2016-12-14 14:34:47 -0700 | [diff] [blame] | 807 | void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list) |
Jens Axboe | 1429d7c | 2014-05-19 09:23:55 -0600 | [diff] [blame] | 808 | { |
Omar Sandoval | 8845964 | 2016-09-17 08:38:44 -0600 | [diff] [blame] | 809 | struct flush_busy_ctx_data data = { |
| 810 | .hctx = hctx, |
| 811 | .list = list, |
| 812 | }; |
Jens Axboe | 1429d7c | 2014-05-19 09:23:55 -0600 | [diff] [blame] | 813 | |
Omar Sandoval | 8845964 | 2016-09-17 08:38:44 -0600 | [diff] [blame] | 814 | sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data); |
Jens Axboe | 1429d7c | 2014-05-19 09:23:55 -0600 | [diff] [blame] | 815 | } |
Jens Axboe | 2c3ad66 | 2016-12-14 14:34:47 -0700 | [diff] [blame] | 816 | EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs); |
Jens Axboe | 1429d7c | 2014-05-19 09:23:55 -0600 | [diff] [blame] | 817 | |
Jens Axboe | 703fd1c | 2016-09-16 13:59:14 -0600 | [diff] [blame] | 818 | static inline unsigned int queued_to_index(unsigned int queued) |
| 819 | { |
| 820 | if (!queued) |
| 821 | return 0; |
Jens Axboe | 1429d7c | 2014-05-19 09:23:55 -0600 | [diff] [blame] | 822 | |
Jens Axboe | 703fd1c | 2016-09-16 13:59:14 -0600 | [diff] [blame] | 823 | return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1); |
Jens Axboe | 1429d7c | 2014-05-19 09:23:55 -0600 | [diff] [blame] | 824 | } |
| 825 | |
Jens Axboe | f04c3df | 2016-12-07 08:41:17 -0700 | [diff] [blame] | 826 | bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list) |
| 827 | { |
| 828 | struct request_queue *q = hctx->queue; |
| 829 | struct request *rq; |
| 830 | LIST_HEAD(driver_list); |
| 831 | struct list_head *dptr; |
| 832 | int queued, ret = BLK_MQ_RQ_QUEUE_OK; |
| 833 | |
| 834 | /* |
| 835 | * Start off with dptr being NULL, so we start the first request |
| 836 | * immediately, even if we have more pending. |
| 837 | */ |
| 838 | dptr = NULL; |
| 839 | |
| 840 | /* |
| 841 | * Now process all the entries, sending them to the driver. |
| 842 | */ |
| 843 | queued = 0; |
| 844 | while (!list_empty(list)) { |
| 845 | struct blk_mq_queue_data bd; |
| 846 | |
| 847 | rq = list_first_entry(list, struct request, queuelist); |
| 848 | list_del_init(&rq->queuelist); |
| 849 | |
| 850 | bd.rq = rq; |
| 851 | bd.list = dptr; |
| 852 | bd.last = list_empty(list); |
| 853 | |
| 854 | ret = q->mq_ops->queue_rq(hctx, &bd); |
| 855 | switch (ret) { |
| 856 | case BLK_MQ_RQ_QUEUE_OK: |
| 857 | queued++; |
| 858 | break; |
| 859 | case BLK_MQ_RQ_QUEUE_BUSY: |
| 860 | list_add(&rq->queuelist, list); |
| 861 | __blk_mq_requeue_request(rq); |
| 862 | break; |
| 863 | default: |
| 864 | pr_err("blk-mq: bad return on queue: %d\n", ret); |
| 865 | case BLK_MQ_RQ_QUEUE_ERROR: |
| 866 | rq->errors = -EIO; |
| 867 | blk_mq_end_request(rq, rq->errors); |
| 868 | break; |
| 869 | } |
| 870 | |
| 871 | if (ret == BLK_MQ_RQ_QUEUE_BUSY) |
| 872 | break; |
| 873 | |
| 874 | /* |
| 875 | * We've done the first request. If we have more than 1 |
| 876 | * left in the list, set dptr to defer issue. |
| 877 | */ |
| 878 | if (!dptr && list->next != list->prev) |
| 879 | dptr = &driver_list; |
| 880 | } |
| 881 | |
| 882 | hctx->dispatched[queued_to_index(queued)]++; |
| 883 | |
| 884 | /* |
| 885 | * Any items that need requeuing? Stuff them into hctx->dispatch, |
| 886 | * that is where we will continue on next queue run. |
| 887 | */ |
| 888 | if (!list_empty(list)) { |
| 889 | spin_lock(&hctx->lock); |
| 890 | list_splice(list, &hctx->dispatch); |
| 891 | spin_unlock(&hctx->lock); |
| 892 | |
| 893 | /* |
| 894 | * the queue is expected stopped with BLK_MQ_RQ_QUEUE_BUSY, but |
| 895 | * it's possible the queue is stopped and restarted again |
| 896 | * before this. Queue restart will dispatch requests. And since |
| 897 | * requests in rq_list aren't added into hctx->dispatch yet, |
| 898 | * the requests in rq_list might get lost. |
| 899 | * |
| 900 | * blk_mq_run_hw_queue() already checks the STOPPED bit |
| 901 | **/ |
| 902 | blk_mq_run_hw_queue(hctx, true); |
| 903 | } |
| 904 | |
| 905 | return ret != BLK_MQ_RQ_QUEUE_BUSY; |
| 906 | } |
| 907 | |
Jens Axboe | 1429d7c | 2014-05-19 09:23:55 -0600 | [diff] [blame] | 908 | /* |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 909 | * Run this hardware queue, pulling any software queues mapped to it in. |
| 910 | * Note that this function currently has various problems around ordering |
| 911 | * of IO. In particular, we'd like FIFO behaviour on handling existing |
| 912 | * items on the hctx->dispatch list. Ignore that for now. |
| 913 | */ |
Bart Van Assche | 6a83e74 | 2016-11-02 10:09:51 -0600 | [diff] [blame] | 914 | static void blk_mq_process_rq_list(struct blk_mq_hw_ctx *hctx) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 915 | { |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 916 | LIST_HEAD(rq_list); |
Jens Axboe | 74c4505 | 2014-10-29 11:14:52 -0600 | [diff] [blame] | 917 | LIST_HEAD(driver_list); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 918 | |
Bart Van Assche | 5d1b25c | 2016-10-28 17:19:15 -0700 | [diff] [blame] | 919 | if (unlikely(blk_mq_hctx_stopped(hctx))) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 920 | return; |
| 921 | |
| 922 | hctx->run++; |
| 923 | |
| 924 | /* |
| 925 | * Touch any software queue that has pending entries. |
| 926 | */ |
Jens Axboe | 2c3ad66 | 2016-12-14 14:34:47 -0700 | [diff] [blame] | 927 | blk_mq_flush_busy_ctxs(hctx, &rq_list); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 928 | |
| 929 | /* |
| 930 | * If we have previous entries on our dispatch list, grab them |
| 931 | * and stuff them at the front for more fair dispatch. |
| 932 | */ |
| 933 | if (!list_empty_careful(&hctx->dispatch)) { |
| 934 | spin_lock(&hctx->lock); |
| 935 | if (!list_empty(&hctx->dispatch)) |
| 936 | list_splice_init(&hctx->dispatch, &rq_list); |
| 937 | spin_unlock(&hctx->lock); |
| 938 | } |
| 939 | |
Jens Axboe | f04c3df | 2016-12-07 08:41:17 -0700 | [diff] [blame] | 940 | blk_mq_dispatch_rq_list(hctx, &rq_list); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 941 | } |
| 942 | |
Bart Van Assche | 6a83e74 | 2016-11-02 10:09:51 -0600 | [diff] [blame] | 943 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx) |
| 944 | { |
| 945 | int srcu_idx; |
| 946 | |
| 947 | WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) && |
| 948 | cpu_online(hctx->next_cpu)); |
| 949 | |
| 950 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) { |
| 951 | rcu_read_lock(); |
| 952 | blk_mq_process_rq_list(hctx); |
| 953 | rcu_read_unlock(); |
| 954 | } else { |
| 955 | srcu_idx = srcu_read_lock(&hctx->queue_rq_srcu); |
| 956 | blk_mq_process_rq_list(hctx); |
| 957 | srcu_read_unlock(&hctx->queue_rq_srcu, srcu_idx); |
| 958 | } |
| 959 | } |
| 960 | |
Jens Axboe | 506e931 | 2014-05-07 10:26:44 -0600 | [diff] [blame] | 961 | /* |
| 962 | * It'd be great if the workqueue API had a way to pass |
| 963 | * in a mask and had some smarts for more clever placement. |
| 964 | * For now we just round-robin here, switching for every |
| 965 | * BLK_MQ_CPU_WORK_BATCH queued items. |
| 966 | */ |
| 967 | static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx) |
| 968 | { |
Christoph Hellwig | b657d7e | 2014-11-24 09:27:23 +0100 | [diff] [blame] | 969 | if (hctx->queue->nr_hw_queues == 1) |
| 970 | return WORK_CPU_UNBOUND; |
Jens Axboe | 506e931 | 2014-05-07 10:26:44 -0600 | [diff] [blame] | 971 | |
| 972 | if (--hctx->next_cpu_batch <= 0) { |
Gabriel Krisman Bertazi | c02ebfd | 2016-09-28 00:24:24 -0300 | [diff] [blame] | 973 | int next_cpu; |
Jens Axboe | 506e931 | 2014-05-07 10:26:44 -0600 | [diff] [blame] | 974 | |
| 975 | next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask); |
| 976 | if (next_cpu >= nr_cpu_ids) |
| 977 | next_cpu = cpumask_first(hctx->cpumask); |
| 978 | |
| 979 | hctx->next_cpu = next_cpu; |
| 980 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; |
| 981 | } |
| 982 | |
Christoph Hellwig | b657d7e | 2014-11-24 09:27:23 +0100 | [diff] [blame] | 983 | return hctx->next_cpu; |
Jens Axboe | 506e931 | 2014-05-07 10:26:44 -0600 | [diff] [blame] | 984 | } |
| 985 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 986 | void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
| 987 | { |
Bart Van Assche | 5d1b25c | 2016-10-28 17:19:15 -0700 | [diff] [blame] | 988 | if (unlikely(blk_mq_hctx_stopped(hctx) || |
| 989 | !blk_mq_hw_queue_mapped(hctx))) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 990 | return; |
| 991 | |
Jens Axboe | 1b792f2 | 2016-09-21 10:12:13 -0600 | [diff] [blame] | 992 | if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) { |
Paolo Bonzini | 2a90d4a | 2014-11-07 23:04:00 +0100 | [diff] [blame] | 993 | int cpu = get_cpu(); |
| 994 | if (cpumask_test_cpu(cpu, hctx->cpumask)) { |
Paolo Bonzini | 398205b | 2014-11-07 23:03:59 +0100 | [diff] [blame] | 995 | __blk_mq_run_hw_queue(hctx); |
Paolo Bonzini | 2a90d4a | 2014-11-07 23:04:00 +0100 | [diff] [blame] | 996 | put_cpu(); |
Paolo Bonzini | 398205b | 2014-11-07 23:03:59 +0100 | [diff] [blame] | 997 | return; |
| 998 | } |
Jens Axboe | e4043dc | 2014-04-09 10:18:23 -0600 | [diff] [blame] | 999 | |
Paolo Bonzini | 2a90d4a | 2014-11-07 23:04:00 +0100 | [diff] [blame] | 1000 | put_cpu(); |
Jens Axboe | e4043dc | 2014-04-09 10:18:23 -0600 | [diff] [blame] | 1001 | } |
Paolo Bonzini | 398205b | 2014-11-07 23:03:59 +0100 | [diff] [blame] | 1002 | |
Jens Axboe | 27489a3 | 2016-08-24 15:54:25 -0600 | [diff] [blame] | 1003 | kblockd_schedule_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1004 | } |
| 1005 | |
Mike Snitzer | b94ec29 | 2015-03-11 23:56:38 -0400 | [diff] [blame] | 1006 | void blk_mq_run_hw_queues(struct request_queue *q, bool async) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1007 | { |
| 1008 | struct blk_mq_hw_ctx *hctx; |
| 1009 | int i; |
| 1010 | |
| 1011 | queue_for_each_hw_ctx(q, hctx, i) { |
| 1012 | if ((!blk_mq_hctx_has_pending(hctx) && |
| 1013 | list_empty_careful(&hctx->dispatch)) || |
Bart Van Assche | 5d1b25c | 2016-10-28 17:19:15 -0700 | [diff] [blame] | 1014 | blk_mq_hctx_stopped(hctx)) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1015 | continue; |
| 1016 | |
Mike Snitzer | b94ec29 | 2015-03-11 23:56:38 -0400 | [diff] [blame] | 1017 | blk_mq_run_hw_queue(hctx, async); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1018 | } |
| 1019 | } |
Mike Snitzer | b94ec29 | 2015-03-11 23:56:38 -0400 | [diff] [blame] | 1020 | EXPORT_SYMBOL(blk_mq_run_hw_queues); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1021 | |
Bart Van Assche | fd00144 | 2016-10-28 17:19:37 -0700 | [diff] [blame] | 1022 | /** |
| 1023 | * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped |
| 1024 | * @q: request queue. |
| 1025 | * |
| 1026 | * The caller is responsible for serializing this function against |
| 1027 | * blk_mq_{start,stop}_hw_queue(). |
| 1028 | */ |
| 1029 | bool blk_mq_queue_stopped(struct request_queue *q) |
| 1030 | { |
| 1031 | struct blk_mq_hw_ctx *hctx; |
| 1032 | int i; |
| 1033 | |
| 1034 | queue_for_each_hw_ctx(q, hctx, i) |
| 1035 | if (blk_mq_hctx_stopped(hctx)) |
| 1036 | return true; |
| 1037 | |
| 1038 | return false; |
| 1039 | } |
| 1040 | EXPORT_SYMBOL(blk_mq_queue_stopped); |
| 1041 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1042 | void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) |
| 1043 | { |
Jens Axboe | 27489a3 | 2016-08-24 15:54:25 -0600 | [diff] [blame] | 1044 | cancel_work(&hctx->run_work); |
Christoph Hellwig | 70f4db6 | 2014-04-16 10:48:08 -0600 | [diff] [blame] | 1045 | cancel_delayed_work(&hctx->delay_work); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1046 | set_bit(BLK_MQ_S_STOPPED, &hctx->state); |
| 1047 | } |
| 1048 | EXPORT_SYMBOL(blk_mq_stop_hw_queue); |
| 1049 | |
Christoph Hellwig | 280d45f | 2013-10-25 14:45:58 +0100 | [diff] [blame] | 1050 | void blk_mq_stop_hw_queues(struct request_queue *q) |
| 1051 | { |
| 1052 | struct blk_mq_hw_ctx *hctx; |
| 1053 | int i; |
| 1054 | |
| 1055 | queue_for_each_hw_ctx(q, hctx, i) |
| 1056 | blk_mq_stop_hw_queue(hctx); |
| 1057 | } |
| 1058 | EXPORT_SYMBOL(blk_mq_stop_hw_queues); |
| 1059 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1060 | void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) |
| 1061 | { |
| 1062 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); |
Jens Axboe | e4043dc | 2014-04-09 10:18:23 -0600 | [diff] [blame] | 1063 | |
Jens Axboe | 0ffbce8 | 2014-06-25 08:22:34 -0600 | [diff] [blame] | 1064 | blk_mq_run_hw_queue(hctx, false); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1065 | } |
| 1066 | EXPORT_SYMBOL(blk_mq_start_hw_queue); |
| 1067 | |
Christoph Hellwig | 2f26855 | 2014-04-16 09:44:56 +0200 | [diff] [blame] | 1068 | void blk_mq_start_hw_queues(struct request_queue *q) |
| 1069 | { |
| 1070 | struct blk_mq_hw_ctx *hctx; |
| 1071 | int i; |
| 1072 | |
| 1073 | queue_for_each_hw_ctx(q, hctx, i) |
| 1074 | blk_mq_start_hw_queue(hctx); |
| 1075 | } |
| 1076 | EXPORT_SYMBOL(blk_mq_start_hw_queues); |
| 1077 | |
Jens Axboe | ae911c5 | 2016-12-08 13:19:30 -0700 | [diff] [blame] | 1078 | void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
| 1079 | { |
| 1080 | if (!blk_mq_hctx_stopped(hctx)) |
| 1081 | return; |
| 1082 | |
| 1083 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); |
| 1084 | blk_mq_run_hw_queue(hctx, async); |
| 1085 | } |
| 1086 | EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue); |
| 1087 | |
Christoph Hellwig | 1b4a325 | 2014-04-16 09:44:54 +0200 | [diff] [blame] | 1088 | 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] | 1089 | { |
| 1090 | struct blk_mq_hw_ctx *hctx; |
| 1091 | int i; |
| 1092 | |
Jens Axboe | ae911c5 | 2016-12-08 13:19:30 -0700 | [diff] [blame] | 1093 | queue_for_each_hw_ctx(q, hctx, i) |
| 1094 | blk_mq_start_stopped_hw_queue(hctx, async); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1095 | } |
| 1096 | EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); |
| 1097 | |
Christoph Hellwig | 70f4db6 | 2014-04-16 10:48:08 -0600 | [diff] [blame] | 1098 | static void blk_mq_run_work_fn(struct work_struct *work) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1099 | { |
| 1100 | struct blk_mq_hw_ctx *hctx; |
| 1101 | |
Jens Axboe | 27489a3 | 2016-08-24 15:54:25 -0600 | [diff] [blame] | 1102 | hctx = container_of(work, struct blk_mq_hw_ctx, run_work); |
Jens Axboe | e4043dc | 2014-04-09 10:18:23 -0600 | [diff] [blame] | 1103 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1104 | __blk_mq_run_hw_queue(hctx); |
| 1105 | } |
| 1106 | |
Christoph Hellwig | 70f4db6 | 2014-04-16 10:48:08 -0600 | [diff] [blame] | 1107 | static void blk_mq_delay_work_fn(struct work_struct *work) |
| 1108 | { |
| 1109 | struct blk_mq_hw_ctx *hctx; |
| 1110 | |
| 1111 | hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work); |
| 1112 | |
| 1113 | if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state)) |
| 1114 | __blk_mq_run_hw_queue(hctx); |
| 1115 | } |
| 1116 | |
| 1117 | void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) |
| 1118 | { |
Ming Lei | 19c66e5 | 2014-12-03 19:38:04 +0800 | [diff] [blame] | 1119 | if (unlikely(!blk_mq_hw_queue_mapped(hctx))) |
| 1120 | return; |
Christoph Hellwig | 70f4db6 | 2014-04-16 10:48:08 -0600 | [diff] [blame] | 1121 | |
Christoph Hellwig | b657d7e | 2014-11-24 09:27:23 +0100 | [diff] [blame] | 1122 | kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx), |
| 1123 | &hctx->delay_work, msecs_to_jiffies(msecs)); |
Christoph Hellwig | 70f4db6 | 2014-04-16 10:48:08 -0600 | [diff] [blame] | 1124 | } |
| 1125 | EXPORT_SYMBOL(blk_mq_delay_queue); |
| 1126 | |
Ming Lei | cfd0c55 | 2015-10-20 23:13:57 +0800 | [diff] [blame] | 1127 | static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx, |
Ming Lei | cfd0c55 | 2015-10-20 23:13:57 +0800 | [diff] [blame] | 1128 | struct request *rq, |
| 1129 | bool at_head) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1130 | { |
Jens Axboe | e57690f | 2016-08-24 15:34:35 -0600 | [diff] [blame] | 1131 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
| 1132 | |
Jens Axboe | 01b983c | 2013-11-19 18:59:10 -0700 | [diff] [blame] | 1133 | trace_block_rq_insert(hctx->queue, rq); |
| 1134 | |
Christoph Hellwig | 72a0a36 | 2014-02-07 10:22:36 -0800 | [diff] [blame] | 1135 | if (at_head) |
| 1136 | list_add(&rq->queuelist, &ctx->rq_list); |
| 1137 | else |
| 1138 | list_add_tail(&rq->queuelist, &ctx->rq_list); |
Ming Lei | cfd0c55 | 2015-10-20 23:13:57 +0800 | [diff] [blame] | 1139 | } |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 1140 | |
Jens Axboe | 2c3ad66 | 2016-12-14 14:34:47 -0700 | [diff] [blame] | 1141 | void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq, |
| 1142 | bool at_head) |
Ming Lei | cfd0c55 | 2015-10-20 23:13:57 +0800 | [diff] [blame] | 1143 | { |
| 1144 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
| 1145 | |
Jens Axboe | e57690f | 2016-08-24 15:34:35 -0600 | [diff] [blame] | 1146 | __blk_mq_insert_req_list(hctx, rq, at_head); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1147 | blk_mq_hctx_mark_pending(hctx, ctx); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1148 | } |
| 1149 | |
Christoph Hellwig | eeabc85 | 2014-03-21 08:57:37 -0600 | [diff] [blame] | 1150 | void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue, |
Jens Axboe | e57690f | 2016-08-24 15:34:35 -0600 | [diff] [blame] | 1151 | bool async) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1152 | { |
Jens Axboe | e57690f | 2016-08-24 15:34:35 -0600 | [diff] [blame] | 1153 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1154 | struct request_queue *q = rq->q; |
Christoph Hellwig | 7d7e0f9 | 2016-09-14 16:18:54 +0200 | [diff] [blame] | 1155 | struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1156 | |
Christoph Hellwig | a57a178 | 2014-09-16 14:44:07 -0700 | [diff] [blame] | 1157 | spin_lock(&ctx->lock); |
| 1158 | __blk_mq_insert_request(hctx, rq, at_head); |
| 1159 | spin_unlock(&ctx->lock); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1160 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1161 | if (run_queue) |
| 1162 | blk_mq_run_hw_queue(hctx, async); |
| 1163 | } |
| 1164 | |
| 1165 | static void blk_mq_insert_requests(struct request_queue *q, |
| 1166 | struct blk_mq_ctx *ctx, |
| 1167 | struct list_head *list, |
| 1168 | int depth, |
| 1169 | bool from_schedule) |
| 1170 | |
| 1171 | { |
Christoph Hellwig | 7d7e0f9 | 2016-09-14 16:18:54 +0200 | [diff] [blame] | 1172 | struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1173 | |
| 1174 | trace_block_unplug(q, depth, !from_schedule); |
| 1175 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1176 | /* |
| 1177 | * preemption doesn't flush plug list, so it's possible ctx->cpu is |
| 1178 | * offline now |
| 1179 | */ |
| 1180 | spin_lock(&ctx->lock); |
| 1181 | while (!list_empty(list)) { |
| 1182 | struct request *rq; |
| 1183 | |
| 1184 | rq = list_first_entry(list, struct request, queuelist); |
Jens Axboe | e57690f | 2016-08-24 15:34:35 -0600 | [diff] [blame] | 1185 | BUG_ON(rq->mq_ctx != ctx); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1186 | list_del_init(&rq->queuelist); |
Jens Axboe | e57690f | 2016-08-24 15:34:35 -0600 | [diff] [blame] | 1187 | __blk_mq_insert_req_list(hctx, rq, false); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1188 | } |
Ming Lei | cfd0c55 | 2015-10-20 23:13:57 +0800 | [diff] [blame] | 1189 | blk_mq_hctx_mark_pending(hctx, ctx); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1190 | spin_unlock(&ctx->lock); |
| 1191 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1192 | blk_mq_run_hw_queue(hctx, from_schedule); |
| 1193 | } |
| 1194 | |
| 1195 | static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b) |
| 1196 | { |
| 1197 | struct request *rqa = container_of(a, struct request, queuelist); |
| 1198 | struct request *rqb = container_of(b, struct request, queuelist); |
| 1199 | |
| 1200 | return !(rqa->mq_ctx < rqb->mq_ctx || |
| 1201 | (rqa->mq_ctx == rqb->mq_ctx && |
| 1202 | blk_rq_pos(rqa) < blk_rq_pos(rqb))); |
| 1203 | } |
| 1204 | |
| 1205 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) |
| 1206 | { |
| 1207 | struct blk_mq_ctx *this_ctx; |
| 1208 | struct request_queue *this_q; |
| 1209 | struct request *rq; |
| 1210 | LIST_HEAD(list); |
| 1211 | LIST_HEAD(ctx_list); |
| 1212 | unsigned int depth; |
| 1213 | |
| 1214 | list_splice_init(&plug->mq_list, &list); |
| 1215 | |
| 1216 | list_sort(NULL, &list, plug_ctx_cmp); |
| 1217 | |
| 1218 | this_q = NULL; |
| 1219 | this_ctx = NULL; |
| 1220 | depth = 0; |
| 1221 | |
| 1222 | while (!list_empty(&list)) { |
| 1223 | rq = list_entry_rq(list.next); |
| 1224 | list_del_init(&rq->queuelist); |
| 1225 | BUG_ON(!rq->q); |
| 1226 | if (rq->mq_ctx != this_ctx) { |
| 1227 | if (this_ctx) { |
| 1228 | blk_mq_insert_requests(this_q, this_ctx, |
| 1229 | &ctx_list, depth, |
| 1230 | from_schedule); |
| 1231 | } |
| 1232 | |
| 1233 | this_ctx = rq->mq_ctx; |
| 1234 | this_q = rq->q; |
| 1235 | depth = 0; |
| 1236 | } |
| 1237 | |
| 1238 | depth++; |
| 1239 | list_add_tail(&rq->queuelist, &ctx_list); |
| 1240 | } |
| 1241 | |
| 1242 | /* |
| 1243 | * If 'this_ctx' is set, we know we have entries to complete |
| 1244 | * on 'ctx_list'. Do those. |
| 1245 | */ |
| 1246 | if (this_ctx) { |
| 1247 | blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth, |
| 1248 | from_schedule); |
| 1249 | } |
| 1250 | } |
| 1251 | |
| 1252 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio) |
| 1253 | { |
| 1254 | init_request_from_bio(rq, bio); |
Jens Axboe | 4b57052 | 2014-05-29 11:00:11 -0600 | [diff] [blame] | 1255 | |
Jens Axboe | 6e85eaf | 2016-12-02 20:00:14 -0700 | [diff] [blame] | 1256 | blk_account_io_start(rq, true); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1257 | } |
| 1258 | |
Jens Axboe | 274a584 | 2014-08-15 12:44:08 -0600 | [diff] [blame] | 1259 | static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx) |
| 1260 | { |
| 1261 | return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) && |
| 1262 | !blk_queue_nomerges(hctx->queue); |
| 1263 | } |
| 1264 | |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1265 | static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx, |
| 1266 | struct blk_mq_ctx *ctx, |
| 1267 | struct request *rq, struct bio *bio) |
| 1268 | { |
Ming Lei | e18378a | 2015-10-20 23:13:54 +0800 | [diff] [blame] | 1269 | if (!hctx_allow_merges(hctx) || !bio_mergeable(bio)) { |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1270 | blk_mq_bio_to_request(rq, bio); |
| 1271 | spin_lock(&ctx->lock); |
| 1272 | insert_rq: |
| 1273 | __blk_mq_insert_request(hctx, rq, false); |
| 1274 | spin_unlock(&ctx->lock); |
| 1275 | return false; |
| 1276 | } else { |
Jens Axboe | 274a584 | 2014-08-15 12:44:08 -0600 | [diff] [blame] | 1277 | struct request_queue *q = hctx->queue; |
| 1278 | |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1279 | spin_lock(&ctx->lock); |
| 1280 | if (!blk_mq_attempt_merge(q, ctx, bio)) { |
| 1281 | blk_mq_bio_to_request(rq, bio); |
| 1282 | goto insert_rq; |
| 1283 | } |
| 1284 | |
| 1285 | spin_unlock(&ctx->lock); |
| 1286 | __blk_mq_free_request(hctx, ctx, rq); |
| 1287 | return true; |
| 1288 | } |
| 1289 | } |
| 1290 | |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1291 | static struct request *blk_mq_map_request(struct request_queue *q, |
| 1292 | struct bio *bio, |
Jens Axboe | 2552e3f | 2016-10-27 19:03:55 -0600 | [diff] [blame] | 1293 | struct blk_mq_alloc_data *data) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1294 | { |
| 1295 | struct blk_mq_hw_ctx *hctx; |
| 1296 | struct blk_mq_ctx *ctx; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1297 | struct request *rq; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1298 | |
Dan Williams | 3ef28e8 | 2015-10-21 13:20:12 -0400 | [diff] [blame] | 1299 | blk_queue_enter_live(q); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1300 | ctx = blk_mq_get_ctx(q); |
Christoph Hellwig | 7d7e0f9 | 2016-09-14 16:18:54 +0200 | [diff] [blame] | 1301 | hctx = blk_mq_map_queue(q, ctx->cpu); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1302 | |
Christoph Hellwig | ef295ec | 2016-10-28 08:48:16 -0600 | [diff] [blame] | 1303 | trace_block_getrq(q, bio, bio->bi_opf); |
Jens Axboe | 2552e3f | 2016-10-27 19:03:55 -0600 | [diff] [blame] | 1304 | blk_mq_set_alloc_data(data, q, 0, ctx, hctx); |
Christoph Hellwig | ef295ec | 2016-10-28 08:48:16 -0600 | [diff] [blame] | 1305 | rq = __blk_mq_alloc_request(data, bio->bi_opf); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1306 | |
Jens Axboe | 7dd2fb6 | 2016-10-27 09:49:19 -0600 | [diff] [blame] | 1307 | data->hctx->queued++; |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1308 | return rq; |
| 1309 | } |
| 1310 | |
Jens Axboe | fd2d332 | 2017-01-12 10:04:45 -0700 | [diff] [blame^] | 1311 | static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq) |
| 1312 | { |
| 1313 | return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false); |
| 1314 | } |
| 1315 | |
Jens Axboe | 066a4a7 | 2016-11-11 12:24:46 -0700 | [diff] [blame] | 1316 | static void blk_mq_try_issue_directly(struct request *rq, blk_qc_t *cookie) |
Shaohua Li | f984df1 | 2015-05-08 10:51:32 -0700 | [diff] [blame] | 1317 | { |
| 1318 | int ret; |
| 1319 | struct request_queue *q = rq->q; |
Jens Axboe | 066a4a7 | 2016-11-11 12:24:46 -0700 | [diff] [blame] | 1320 | struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, rq->mq_ctx->cpu); |
Shaohua Li | f984df1 | 2015-05-08 10:51:32 -0700 | [diff] [blame] | 1321 | struct blk_mq_queue_data bd = { |
| 1322 | .rq = rq, |
| 1323 | .list = NULL, |
| 1324 | .last = 1 |
| 1325 | }; |
Jens Axboe | fd2d332 | 2017-01-12 10:04:45 -0700 | [diff] [blame^] | 1326 | blk_qc_t new_cookie = request_to_qc_t(hctx, rq); |
Shaohua Li | f984df1 | 2015-05-08 10:51:32 -0700 | [diff] [blame] | 1327 | |
Bart Van Assche | 2253efc | 2016-10-28 17:20:02 -0700 | [diff] [blame] | 1328 | if (blk_mq_hctx_stopped(hctx)) |
| 1329 | goto insert; |
| 1330 | |
Shaohua Li | f984df1 | 2015-05-08 10:51:32 -0700 | [diff] [blame] | 1331 | /* |
| 1332 | * For OK queue, we are done. For error, kill it. Any other |
| 1333 | * error (busy), just add it to our list as we previously |
| 1334 | * would have done |
| 1335 | */ |
| 1336 | ret = q->mq_ops->queue_rq(hctx, &bd); |
Jens Axboe | 7b37163 | 2015-11-05 10:41:40 -0700 | [diff] [blame] | 1337 | if (ret == BLK_MQ_RQ_QUEUE_OK) { |
| 1338 | *cookie = new_cookie; |
Bart Van Assche | 2253efc | 2016-10-28 17:20:02 -0700 | [diff] [blame] | 1339 | return; |
Shaohua Li | f984df1 | 2015-05-08 10:51:32 -0700 | [diff] [blame] | 1340 | } |
Jens Axboe | 7b37163 | 2015-11-05 10:41:40 -0700 | [diff] [blame] | 1341 | |
| 1342 | __blk_mq_requeue_request(rq); |
| 1343 | |
| 1344 | if (ret == BLK_MQ_RQ_QUEUE_ERROR) { |
| 1345 | *cookie = BLK_QC_T_NONE; |
| 1346 | rq->errors = -EIO; |
| 1347 | blk_mq_end_request(rq, rq->errors); |
Bart Van Assche | 2253efc | 2016-10-28 17:20:02 -0700 | [diff] [blame] | 1348 | return; |
Jens Axboe | 7b37163 | 2015-11-05 10:41:40 -0700 | [diff] [blame] | 1349 | } |
| 1350 | |
Bart Van Assche | 2253efc | 2016-10-28 17:20:02 -0700 | [diff] [blame] | 1351 | insert: |
| 1352 | blk_mq_insert_request(rq, false, true, true); |
Shaohua Li | f984df1 | 2015-05-08 10:51:32 -0700 | [diff] [blame] | 1353 | } |
| 1354 | |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1355 | /* |
| 1356 | * Multiple hardware queue variant. This will not use per-process plugs, |
| 1357 | * but will attempt to bypass the hctx queueing if we can go straight to |
| 1358 | * hardware for SYNC IO. |
| 1359 | */ |
Jens Axboe | dece163 | 2015-11-05 10:41:16 -0700 | [diff] [blame] | 1360 | static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio) |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1361 | { |
Christoph Hellwig | ef295ec | 2016-10-28 08:48:16 -0600 | [diff] [blame] | 1362 | const int is_sync = op_is_sync(bio->bi_opf); |
Jens Axboe | 1eff9d3 | 2016-08-05 15:35:16 -0600 | [diff] [blame] | 1363 | const int is_flush_fua = bio->bi_opf & (REQ_PREFLUSH | REQ_FUA); |
Jens Axboe | 2552e3f | 2016-10-27 19:03:55 -0600 | [diff] [blame] | 1364 | struct blk_mq_alloc_data data; |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1365 | struct request *rq; |
Bart Van Assche | 6a83e74 | 2016-11-02 10:09:51 -0600 | [diff] [blame] | 1366 | unsigned int request_count = 0, srcu_idx; |
Shaohua Li | f984df1 | 2015-05-08 10:51:32 -0700 | [diff] [blame] | 1367 | struct blk_plug *plug; |
Shaohua Li | 5b3f341 | 2015-05-08 10:51:33 -0700 | [diff] [blame] | 1368 | struct request *same_queue_rq = NULL; |
Jens Axboe | 7b37163 | 2015-11-05 10:41:40 -0700 | [diff] [blame] | 1369 | blk_qc_t cookie; |
Jens Axboe | 87760e5 | 2016-11-09 12:38:14 -0700 | [diff] [blame] | 1370 | unsigned int wb_acct; |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1371 | |
| 1372 | blk_queue_bounce(q, &bio); |
| 1373 | |
| 1374 | if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) { |
Christoph Hellwig | 4246a0b | 2015-07-20 15:29:37 +0200 | [diff] [blame] | 1375 | bio_io_error(bio); |
Jens Axboe | dece163 | 2015-11-05 10:41:16 -0700 | [diff] [blame] | 1376 | return BLK_QC_T_NONE; |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1377 | } |
| 1378 | |
Kent Overstreet | 54efd50 | 2015-04-23 22:37:18 -0700 | [diff] [blame] | 1379 | blk_queue_split(q, &bio, q->bio_split); |
| 1380 | |
Omar Sandoval | 87c279e | 2016-06-01 22:18:48 -0700 | [diff] [blame] | 1381 | if (!is_flush_fua && !blk_queue_nomerges(q) && |
| 1382 | blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq)) |
| 1383 | return BLK_QC_T_NONE; |
Shaohua Li | f984df1 | 2015-05-08 10:51:32 -0700 | [diff] [blame] | 1384 | |
Jens Axboe | 87760e5 | 2016-11-09 12:38:14 -0700 | [diff] [blame] | 1385 | wb_acct = wbt_wait(q->rq_wb, bio, NULL); |
| 1386 | |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1387 | rq = blk_mq_map_request(q, bio, &data); |
Jens Axboe | 87760e5 | 2016-11-09 12:38:14 -0700 | [diff] [blame] | 1388 | if (unlikely(!rq)) { |
| 1389 | __wbt_done(q->rq_wb, wb_acct); |
Jens Axboe | dece163 | 2015-11-05 10:41:16 -0700 | [diff] [blame] | 1390 | return BLK_QC_T_NONE; |
Jens Axboe | 87760e5 | 2016-11-09 12:38:14 -0700 | [diff] [blame] | 1391 | } |
| 1392 | |
| 1393 | wbt_track(&rq->issue_stat, wb_acct); |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1394 | |
Jens Axboe | fd2d332 | 2017-01-12 10:04:45 -0700 | [diff] [blame^] | 1395 | cookie = request_to_qc_t(data.hctx, rq); |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1396 | |
| 1397 | if (unlikely(is_flush_fua)) { |
| 1398 | blk_mq_bio_to_request(rq, bio); |
| 1399 | blk_insert_flush(rq); |
| 1400 | goto run_queue; |
| 1401 | } |
| 1402 | |
Shaohua Li | f984df1 | 2015-05-08 10:51:32 -0700 | [diff] [blame] | 1403 | plug = current->plug; |
Jens Axboe | e167dfb | 2014-10-29 11:18:26 -0600 | [diff] [blame] | 1404 | /* |
| 1405 | * If the driver supports defer issued based on 'last', then |
| 1406 | * queue it up like normal since we can potentially save some |
| 1407 | * CPU this way. |
| 1408 | */ |
Shaohua Li | f984df1 | 2015-05-08 10:51:32 -0700 | [diff] [blame] | 1409 | if (((plug && !blk_queue_nomerges(q)) || is_sync) && |
| 1410 | !(data.hctx->flags & BLK_MQ_F_DEFER_ISSUE)) { |
| 1411 | struct request *old_rq = NULL; |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1412 | |
| 1413 | blk_mq_bio_to_request(rq, bio); |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1414 | |
| 1415 | /* |
Bart Van Assche | 6a83e74 | 2016-11-02 10:09:51 -0600 | [diff] [blame] | 1416 | * We do limited plugging. If the bio can be merged, do that. |
Shaohua Li | f984df1 | 2015-05-08 10:51:32 -0700 | [diff] [blame] | 1417 | * Otherwise the existing request in the plug list will be |
| 1418 | * issued. So the plug list will have one request at most |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1419 | */ |
Shaohua Li | f984df1 | 2015-05-08 10:51:32 -0700 | [diff] [blame] | 1420 | if (plug) { |
Shaohua Li | 5b3f341 | 2015-05-08 10:51:33 -0700 | [diff] [blame] | 1421 | /* |
| 1422 | * The plug list might get flushed before this. If that |
Jens Axboe | b094f89 | 2015-11-20 20:29:45 -0700 | [diff] [blame] | 1423 | * happens, same_queue_rq is invalid and plug list is |
| 1424 | * empty |
| 1425 | */ |
Shaohua Li | 5b3f341 | 2015-05-08 10:51:33 -0700 | [diff] [blame] | 1426 | if (same_queue_rq && !list_empty(&plug->mq_list)) { |
| 1427 | old_rq = same_queue_rq; |
Shaohua Li | f984df1 | 2015-05-08 10:51:32 -0700 | [diff] [blame] | 1428 | list_del_init(&old_rq->queuelist); |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1429 | } |
Shaohua Li | f984df1 | 2015-05-08 10:51:32 -0700 | [diff] [blame] | 1430 | list_add_tail(&rq->queuelist, &plug->mq_list); |
| 1431 | } else /* is_sync */ |
| 1432 | old_rq = rq; |
| 1433 | blk_mq_put_ctx(data.ctx); |
| 1434 | if (!old_rq) |
Jens Axboe | 7b37163 | 2015-11-05 10:41:40 -0700 | [diff] [blame] | 1435 | goto done; |
Bart Van Assche | 6a83e74 | 2016-11-02 10:09:51 -0600 | [diff] [blame] | 1436 | |
| 1437 | if (!(data.hctx->flags & BLK_MQ_F_BLOCKING)) { |
| 1438 | rcu_read_lock(); |
Jens Axboe | 066a4a7 | 2016-11-11 12:24:46 -0700 | [diff] [blame] | 1439 | blk_mq_try_issue_directly(old_rq, &cookie); |
Bart Van Assche | 6a83e74 | 2016-11-02 10:09:51 -0600 | [diff] [blame] | 1440 | rcu_read_unlock(); |
| 1441 | } else { |
| 1442 | srcu_idx = srcu_read_lock(&data.hctx->queue_rq_srcu); |
Jens Axboe | 066a4a7 | 2016-11-11 12:24:46 -0700 | [diff] [blame] | 1443 | blk_mq_try_issue_directly(old_rq, &cookie); |
Bart Van Assche | 6a83e74 | 2016-11-02 10:09:51 -0600 | [diff] [blame] | 1444 | srcu_read_unlock(&data.hctx->queue_rq_srcu, srcu_idx); |
| 1445 | } |
Jens Axboe | 7b37163 | 2015-11-05 10:41:40 -0700 | [diff] [blame] | 1446 | goto done; |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1447 | } |
| 1448 | |
| 1449 | if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) { |
| 1450 | /* |
| 1451 | * For a SYNC request, send it to the hardware immediately. For |
| 1452 | * an ASYNC request, just ensure that we run it later on. The |
| 1453 | * latter allows for merging opportunities and more efficient |
| 1454 | * dispatching. |
| 1455 | */ |
| 1456 | run_queue: |
| 1457 | blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua); |
| 1458 | } |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1459 | blk_mq_put_ctx(data.ctx); |
Jens Axboe | 7b37163 | 2015-11-05 10:41:40 -0700 | [diff] [blame] | 1460 | done: |
| 1461 | return cookie; |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1462 | } |
| 1463 | |
| 1464 | /* |
| 1465 | * Single hardware queue variant. This will attempt to use any per-process |
| 1466 | * plug for merging and IO deferral. |
| 1467 | */ |
Jens Axboe | dece163 | 2015-11-05 10:41:16 -0700 | [diff] [blame] | 1468 | static blk_qc_t blk_sq_make_request(struct request_queue *q, struct bio *bio) |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1469 | { |
Christoph Hellwig | ef295ec | 2016-10-28 08:48:16 -0600 | [diff] [blame] | 1470 | const int is_sync = op_is_sync(bio->bi_opf); |
Jens Axboe | 1eff9d3 | 2016-08-05 15:35:16 -0600 | [diff] [blame] | 1471 | const int is_flush_fua = bio->bi_opf & (REQ_PREFLUSH | REQ_FUA); |
Jeff Moyer | e6c4438 | 2015-05-08 10:51:30 -0700 | [diff] [blame] | 1472 | struct blk_plug *plug; |
| 1473 | unsigned int request_count = 0; |
Jens Axboe | 2552e3f | 2016-10-27 19:03:55 -0600 | [diff] [blame] | 1474 | struct blk_mq_alloc_data data; |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1475 | struct request *rq; |
Jens Axboe | 7b37163 | 2015-11-05 10:41:40 -0700 | [diff] [blame] | 1476 | blk_qc_t cookie; |
Jens Axboe | 87760e5 | 2016-11-09 12:38:14 -0700 | [diff] [blame] | 1477 | unsigned int wb_acct; |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1478 | |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1479 | blk_queue_bounce(q, &bio); |
| 1480 | |
| 1481 | if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) { |
Christoph Hellwig | 4246a0b | 2015-07-20 15:29:37 +0200 | [diff] [blame] | 1482 | bio_io_error(bio); |
Jens Axboe | dece163 | 2015-11-05 10:41:16 -0700 | [diff] [blame] | 1483 | return BLK_QC_T_NONE; |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1484 | } |
| 1485 | |
Kent Overstreet | 54efd50 | 2015-04-23 22:37:18 -0700 | [diff] [blame] | 1486 | blk_queue_split(q, &bio, q->bio_split); |
| 1487 | |
Omar Sandoval | 87c279e | 2016-06-01 22:18:48 -0700 | [diff] [blame] | 1488 | if (!is_flush_fua && !blk_queue_nomerges(q)) { |
| 1489 | if (blk_attempt_plug_merge(q, bio, &request_count, NULL)) |
| 1490 | return BLK_QC_T_NONE; |
| 1491 | } else |
| 1492 | request_count = blk_plug_queued_count(q); |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1493 | |
Jens Axboe | 87760e5 | 2016-11-09 12:38:14 -0700 | [diff] [blame] | 1494 | wb_acct = wbt_wait(q->rq_wb, bio, NULL); |
| 1495 | |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1496 | rq = blk_mq_map_request(q, bio, &data); |
Jens Axboe | 87760e5 | 2016-11-09 12:38:14 -0700 | [diff] [blame] | 1497 | if (unlikely(!rq)) { |
| 1498 | __wbt_done(q->rq_wb, wb_acct); |
Jens Axboe | dece163 | 2015-11-05 10:41:16 -0700 | [diff] [blame] | 1499 | return BLK_QC_T_NONE; |
Jens Axboe | 87760e5 | 2016-11-09 12:38:14 -0700 | [diff] [blame] | 1500 | } |
| 1501 | |
| 1502 | wbt_track(&rq->issue_stat, wb_acct); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1503 | |
Jens Axboe | fd2d332 | 2017-01-12 10:04:45 -0700 | [diff] [blame^] | 1504 | cookie = request_to_qc_t(data.hctx, rq); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1505 | |
| 1506 | if (unlikely(is_flush_fua)) { |
| 1507 | blk_mq_bio_to_request(rq, bio); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1508 | blk_insert_flush(rq); |
| 1509 | goto run_queue; |
| 1510 | } |
| 1511 | |
| 1512 | /* |
| 1513 | * A task plug currently exists. Since this is completely lockless, |
| 1514 | * utilize that to temporarily store requests until the task is |
| 1515 | * either done or scheduled away. |
| 1516 | */ |
Jeff Moyer | e6c4438 | 2015-05-08 10:51:30 -0700 | [diff] [blame] | 1517 | plug = current->plug; |
| 1518 | if (plug) { |
Shaohua Li | 600271d | 2016-11-03 17:03:54 -0700 | [diff] [blame] | 1519 | struct request *last = NULL; |
| 1520 | |
Jeff Moyer | e6c4438 | 2015-05-08 10:51:30 -0700 | [diff] [blame] | 1521 | blk_mq_bio_to_request(rq, bio); |
Ming Lei | 0a6219a | 2016-11-16 18:07:05 +0800 | [diff] [blame] | 1522 | |
| 1523 | /* |
| 1524 | * @request_count may become stale because of schedule |
| 1525 | * out, so check the list again. |
| 1526 | */ |
| 1527 | if (list_empty(&plug->mq_list)) |
| 1528 | request_count = 0; |
Ming Lei | 676d060 | 2015-10-20 23:13:56 +0800 | [diff] [blame] | 1529 | if (!request_count) |
Jeff Moyer | e6c4438 | 2015-05-08 10:51:30 -0700 | [diff] [blame] | 1530 | trace_block_plug(q); |
Shaohua Li | 600271d | 2016-11-03 17:03:54 -0700 | [diff] [blame] | 1531 | else |
| 1532 | last = list_entry_rq(plug->mq_list.prev); |
Jens Axboe | b094f89 | 2015-11-20 20:29:45 -0700 | [diff] [blame] | 1533 | |
| 1534 | blk_mq_put_ctx(data.ctx); |
| 1535 | |
Shaohua Li | 600271d | 2016-11-03 17:03:54 -0700 | [diff] [blame] | 1536 | if (request_count >= BLK_MAX_REQUEST_COUNT || (last && |
| 1537 | blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) { |
Jeff Moyer | e6c4438 | 2015-05-08 10:51:30 -0700 | [diff] [blame] | 1538 | blk_flush_plug_list(plug, false); |
| 1539 | trace_block_plug(q); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1540 | } |
Jens Axboe | b094f89 | 2015-11-20 20:29:45 -0700 | [diff] [blame] | 1541 | |
Jeff Moyer | e6c4438 | 2015-05-08 10:51:30 -0700 | [diff] [blame] | 1542 | list_add_tail(&rq->queuelist, &plug->mq_list); |
Jens Axboe | 7b37163 | 2015-11-05 10:41:40 -0700 | [diff] [blame] | 1543 | return cookie; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1544 | } |
| 1545 | |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1546 | if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) { |
| 1547 | /* |
| 1548 | * For a SYNC request, send it to the hardware immediately. For |
| 1549 | * an ASYNC request, just ensure that we run it later on. The |
| 1550 | * latter allows for merging opportunities and more efficient |
| 1551 | * dispatching. |
| 1552 | */ |
| 1553 | run_queue: |
| 1554 | blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1555 | } |
| 1556 | |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1557 | blk_mq_put_ctx(data.ctx); |
Jens Axboe | 7b37163 | 2015-11-05 10:41:40 -0700 | [diff] [blame] | 1558 | return cookie; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1559 | } |
| 1560 | |
Jens Axboe | cc71a6f | 2017-01-11 14:29:56 -0700 | [diff] [blame] | 1561 | void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, |
| 1562 | unsigned int hctx_idx) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1563 | { |
| 1564 | struct page *page; |
| 1565 | |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1566 | if (tags->rqs && set->ops->exit_request) { |
Christoph Hellwig | e9b267d | 2014-04-15 13:59:10 -0600 | [diff] [blame] | 1567 | int i; |
| 1568 | |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1569 | for (i = 0; i < tags->nr_tags; i++) { |
| 1570 | if (!tags->rqs[i]) |
Christoph Hellwig | e9b267d | 2014-04-15 13:59:10 -0600 | [diff] [blame] | 1571 | continue; |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1572 | set->ops->exit_request(set->driver_data, tags->rqs[i], |
| 1573 | hctx_idx, i); |
Jens Axboe | a516440 | 2014-09-10 09:02:03 -0600 | [diff] [blame] | 1574 | tags->rqs[i] = NULL; |
Christoph Hellwig | e9b267d | 2014-04-15 13:59:10 -0600 | [diff] [blame] | 1575 | } |
| 1576 | } |
| 1577 | |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1578 | while (!list_empty(&tags->page_list)) { |
| 1579 | page = list_first_entry(&tags->page_list, struct page, lru); |
Dave Hansen | 6753471 | 2014-01-08 20:17:46 -0700 | [diff] [blame] | 1580 | list_del_init(&page->lru); |
Catalin Marinas | f75782e | 2015-09-14 18:16:02 +0100 | [diff] [blame] | 1581 | /* |
| 1582 | * Remove kmemleak object previously allocated in |
| 1583 | * blk_mq_init_rq_map(). |
| 1584 | */ |
| 1585 | kmemleak_free(page_address(page)); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1586 | __free_pages(page, page->private); |
| 1587 | } |
Jens Axboe | cc71a6f | 2017-01-11 14:29:56 -0700 | [diff] [blame] | 1588 | } |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1589 | |
Jens Axboe | cc71a6f | 2017-01-11 14:29:56 -0700 | [diff] [blame] | 1590 | void blk_mq_free_rq_map(struct blk_mq_tags *tags) |
| 1591 | { |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1592 | kfree(tags->rqs); |
Jens Axboe | cc71a6f | 2017-01-11 14:29:56 -0700 | [diff] [blame] | 1593 | tags->rqs = NULL; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1594 | |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1595 | blk_mq_free_tags(tags); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1596 | } |
| 1597 | |
Jens Axboe | cc71a6f | 2017-01-11 14:29:56 -0700 | [diff] [blame] | 1598 | struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, |
| 1599 | unsigned int hctx_idx, |
| 1600 | unsigned int nr_tags, |
| 1601 | unsigned int reserved_tags) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1602 | { |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1603 | struct blk_mq_tags *tags; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1604 | |
Jens Axboe | cc71a6f | 2017-01-11 14:29:56 -0700 | [diff] [blame] | 1605 | tags = blk_mq_init_tags(nr_tags, reserved_tags, |
Shaohua Li | 24391c0 | 2015-01-23 14:18:00 -0700 | [diff] [blame] | 1606 | set->numa_node, |
| 1607 | BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags)); |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1608 | if (!tags) |
| 1609 | return NULL; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1610 | |
Jens Axboe | cc71a6f | 2017-01-11 14:29:56 -0700 | [diff] [blame] | 1611 | tags->rqs = kzalloc_node(nr_tags * sizeof(struct request *), |
Gabriel Krisman Bertazi | 36e1f3d1 | 2016-12-06 13:31:44 -0200 | [diff] [blame] | 1612 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, |
Jens Axboe | a516440 | 2014-09-10 09:02:03 -0600 | [diff] [blame] | 1613 | set->numa_node); |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1614 | if (!tags->rqs) { |
| 1615 | blk_mq_free_tags(tags); |
| 1616 | return NULL; |
| 1617 | } |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1618 | |
Jens Axboe | cc71a6f | 2017-01-11 14:29:56 -0700 | [diff] [blame] | 1619 | return tags; |
| 1620 | } |
| 1621 | |
| 1622 | static size_t order_to_size(unsigned int order) |
| 1623 | { |
| 1624 | return (size_t)PAGE_SIZE << order; |
| 1625 | } |
| 1626 | |
| 1627 | int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, |
| 1628 | unsigned int hctx_idx, unsigned int depth) |
| 1629 | { |
| 1630 | unsigned int i, j, entries_per_page, max_order = 4; |
| 1631 | size_t rq_size, left; |
| 1632 | |
| 1633 | INIT_LIST_HEAD(&tags->page_list); |
| 1634 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1635 | /* |
| 1636 | * rq_size is the size of the request plus driver payload, rounded |
| 1637 | * to the cacheline size |
| 1638 | */ |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1639 | rq_size = round_up(sizeof(struct request) + set->cmd_size, |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1640 | cache_line_size()); |
Jens Axboe | cc71a6f | 2017-01-11 14:29:56 -0700 | [diff] [blame] | 1641 | left = rq_size * depth; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1642 | |
Jens Axboe | cc71a6f | 2017-01-11 14:29:56 -0700 | [diff] [blame] | 1643 | for (i = 0; i < depth; ) { |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1644 | int this_order = max_order; |
| 1645 | struct page *page; |
| 1646 | int to_do; |
| 1647 | void *p; |
| 1648 | |
Bartlomiej Zolnierkiewicz | b3a834b | 2016-05-16 09:54:47 -0600 | [diff] [blame] | 1649 | while (this_order && left < order_to_size(this_order - 1)) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1650 | this_order--; |
| 1651 | |
| 1652 | do { |
Jens Axboe | a516440 | 2014-09-10 09:02:03 -0600 | [diff] [blame] | 1653 | page = alloc_pages_node(set->numa_node, |
Gabriel Krisman Bertazi | 36e1f3d1 | 2016-12-06 13:31:44 -0200 | [diff] [blame] | 1654 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO, |
Jens Axboe | a516440 | 2014-09-10 09:02:03 -0600 | [diff] [blame] | 1655 | this_order); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1656 | if (page) |
| 1657 | break; |
| 1658 | if (!this_order--) |
| 1659 | break; |
| 1660 | if (order_to_size(this_order) < rq_size) |
| 1661 | break; |
| 1662 | } while (1); |
| 1663 | |
| 1664 | if (!page) |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1665 | goto fail; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1666 | |
| 1667 | page->private = this_order; |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1668 | list_add_tail(&page->lru, &tags->page_list); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1669 | |
| 1670 | p = page_address(page); |
Catalin Marinas | f75782e | 2015-09-14 18:16:02 +0100 | [diff] [blame] | 1671 | /* |
| 1672 | * Allow kmemleak to scan these pages as they contain pointers |
| 1673 | * to additional allocations like via ops->init_request(). |
| 1674 | */ |
Gabriel Krisman Bertazi | 36e1f3d1 | 2016-12-06 13:31:44 -0200 | [diff] [blame] | 1675 | kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1676 | entries_per_page = order_to_size(this_order) / rq_size; |
Jens Axboe | cc71a6f | 2017-01-11 14:29:56 -0700 | [diff] [blame] | 1677 | to_do = min(entries_per_page, depth - i); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1678 | left -= to_do * rq_size; |
| 1679 | for (j = 0; j < to_do; j++) { |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1680 | tags->rqs[i] = p; |
| 1681 | if (set->ops->init_request) { |
| 1682 | if (set->ops->init_request(set->driver_data, |
| 1683 | tags->rqs[i], hctx_idx, i, |
Jens Axboe | a516440 | 2014-09-10 09:02:03 -0600 | [diff] [blame] | 1684 | set->numa_node)) { |
| 1685 | tags->rqs[i] = NULL; |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1686 | goto fail; |
Jens Axboe | a516440 | 2014-09-10 09:02:03 -0600 | [diff] [blame] | 1687 | } |
Christoph Hellwig | e9b267d | 2014-04-15 13:59:10 -0600 | [diff] [blame] | 1688 | } |
| 1689 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1690 | p += rq_size; |
| 1691 | i++; |
| 1692 | } |
| 1693 | } |
Jens Axboe | cc71a6f | 2017-01-11 14:29:56 -0700 | [diff] [blame] | 1694 | return 0; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1695 | |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1696 | fail: |
Jens Axboe | cc71a6f | 2017-01-11 14:29:56 -0700 | [diff] [blame] | 1697 | blk_mq_free_rqs(set, tags, hctx_idx); |
| 1698 | return -ENOMEM; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1699 | } |
| 1700 | |
Jens Axboe | e57690f | 2016-08-24 15:34:35 -0600 | [diff] [blame] | 1701 | /* |
| 1702 | * 'cpu' is going away. splice any existing rq_list entries from this |
| 1703 | * software queue to the hw queue dispatch list, and ensure that it |
| 1704 | * gets run. |
| 1705 | */ |
Thomas Gleixner | 9467f85 | 2016-09-22 08:05:17 -0600 | [diff] [blame] | 1706 | static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node) |
Jens Axboe | 484b406 | 2014-05-21 14:01:15 -0600 | [diff] [blame] | 1707 | { |
Thomas Gleixner | 9467f85 | 2016-09-22 08:05:17 -0600 | [diff] [blame] | 1708 | struct blk_mq_hw_ctx *hctx; |
Jens Axboe | 484b406 | 2014-05-21 14:01:15 -0600 | [diff] [blame] | 1709 | struct blk_mq_ctx *ctx; |
| 1710 | LIST_HEAD(tmp); |
| 1711 | |
Thomas Gleixner | 9467f85 | 2016-09-22 08:05:17 -0600 | [diff] [blame] | 1712 | hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead); |
Jens Axboe | e57690f | 2016-08-24 15:34:35 -0600 | [diff] [blame] | 1713 | ctx = __blk_mq_get_ctx(hctx->queue, cpu); |
Jens Axboe | 484b406 | 2014-05-21 14:01:15 -0600 | [diff] [blame] | 1714 | |
| 1715 | spin_lock(&ctx->lock); |
| 1716 | if (!list_empty(&ctx->rq_list)) { |
| 1717 | list_splice_init(&ctx->rq_list, &tmp); |
| 1718 | blk_mq_hctx_clear_pending(hctx, ctx); |
| 1719 | } |
| 1720 | spin_unlock(&ctx->lock); |
| 1721 | |
| 1722 | if (list_empty(&tmp)) |
Thomas Gleixner | 9467f85 | 2016-09-22 08:05:17 -0600 | [diff] [blame] | 1723 | return 0; |
Jens Axboe | 484b406 | 2014-05-21 14:01:15 -0600 | [diff] [blame] | 1724 | |
Jens Axboe | e57690f | 2016-08-24 15:34:35 -0600 | [diff] [blame] | 1725 | spin_lock(&hctx->lock); |
| 1726 | list_splice_tail_init(&tmp, &hctx->dispatch); |
| 1727 | spin_unlock(&hctx->lock); |
Jens Axboe | 484b406 | 2014-05-21 14:01:15 -0600 | [diff] [blame] | 1728 | |
| 1729 | blk_mq_run_hw_queue(hctx, true); |
Thomas Gleixner | 9467f85 | 2016-09-22 08:05:17 -0600 | [diff] [blame] | 1730 | return 0; |
Jens Axboe | 484b406 | 2014-05-21 14:01:15 -0600 | [diff] [blame] | 1731 | } |
| 1732 | |
Thomas Gleixner | 9467f85 | 2016-09-22 08:05:17 -0600 | [diff] [blame] | 1733 | static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx) |
Jens Axboe | 484b406 | 2014-05-21 14:01:15 -0600 | [diff] [blame] | 1734 | { |
Thomas Gleixner | 9467f85 | 2016-09-22 08:05:17 -0600 | [diff] [blame] | 1735 | cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD, |
| 1736 | &hctx->cpuhp_dead); |
Jens Axboe | 484b406 | 2014-05-21 14:01:15 -0600 | [diff] [blame] | 1737 | } |
| 1738 | |
Ming Lei | c3b4afc | 2015-06-04 22:25:04 +0800 | [diff] [blame] | 1739 | /* hctx->ctxs will be freed in queue's release handler */ |
Ming Lei | 08e98fc | 2014-09-25 23:23:38 +0800 | [diff] [blame] | 1740 | static void blk_mq_exit_hctx(struct request_queue *q, |
| 1741 | struct blk_mq_tag_set *set, |
| 1742 | struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) |
| 1743 | { |
Ming Lei | f70ced0 | 2014-09-25 23:23:47 +0800 | [diff] [blame] | 1744 | unsigned flush_start_tag = set->queue_depth; |
| 1745 | |
Ming Lei | 08e98fc | 2014-09-25 23:23:38 +0800 | [diff] [blame] | 1746 | blk_mq_tag_idle(hctx); |
| 1747 | |
Ming Lei | f70ced0 | 2014-09-25 23:23:47 +0800 | [diff] [blame] | 1748 | if (set->ops->exit_request) |
| 1749 | set->ops->exit_request(set->driver_data, |
| 1750 | hctx->fq->flush_rq, hctx_idx, |
| 1751 | flush_start_tag + hctx_idx); |
| 1752 | |
Ming Lei | 08e98fc | 2014-09-25 23:23:38 +0800 | [diff] [blame] | 1753 | if (set->ops->exit_hctx) |
| 1754 | set->ops->exit_hctx(hctx, hctx_idx); |
| 1755 | |
Bart Van Assche | 6a83e74 | 2016-11-02 10:09:51 -0600 | [diff] [blame] | 1756 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
| 1757 | cleanup_srcu_struct(&hctx->queue_rq_srcu); |
| 1758 | |
Thomas Gleixner | 9467f85 | 2016-09-22 08:05:17 -0600 | [diff] [blame] | 1759 | blk_mq_remove_cpuhp(hctx); |
Ming Lei | f70ced0 | 2014-09-25 23:23:47 +0800 | [diff] [blame] | 1760 | blk_free_flush_queue(hctx->fq); |
Omar Sandoval | 8845964 | 2016-09-17 08:38:44 -0600 | [diff] [blame] | 1761 | sbitmap_free(&hctx->ctx_map); |
Ming Lei | 08e98fc | 2014-09-25 23:23:38 +0800 | [diff] [blame] | 1762 | } |
| 1763 | |
Ming Lei | 624dbe4 | 2014-05-27 23:35:13 +0800 | [diff] [blame] | 1764 | static void blk_mq_exit_hw_queues(struct request_queue *q, |
| 1765 | struct blk_mq_tag_set *set, int nr_queue) |
| 1766 | { |
| 1767 | struct blk_mq_hw_ctx *hctx; |
| 1768 | unsigned int i; |
| 1769 | |
| 1770 | queue_for_each_hw_ctx(q, hctx, i) { |
| 1771 | if (i == nr_queue) |
| 1772 | break; |
Ming Lei | 08e98fc | 2014-09-25 23:23:38 +0800 | [diff] [blame] | 1773 | blk_mq_exit_hctx(q, set, hctx, i); |
Ming Lei | 624dbe4 | 2014-05-27 23:35:13 +0800 | [diff] [blame] | 1774 | } |
Ming Lei | 624dbe4 | 2014-05-27 23:35:13 +0800 | [diff] [blame] | 1775 | } |
| 1776 | |
| 1777 | static void blk_mq_free_hw_queues(struct request_queue *q, |
| 1778 | struct blk_mq_tag_set *set) |
| 1779 | { |
| 1780 | struct blk_mq_hw_ctx *hctx; |
| 1781 | unsigned int i; |
| 1782 | |
Ming Lei | e09aae7 | 2015-01-29 20:17:27 +0800 | [diff] [blame] | 1783 | queue_for_each_hw_ctx(q, hctx, i) |
Ming Lei | 624dbe4 | 2014-05-27 23:35:13 +0800 | [diff] [blame] | 1784 | free_cpumask_var(hctx->cpumask); |
Ming Lei | 624dbe4 | 2014-05-27 23:35:13 +0800 | [diff] [blame] | 1785 | } |
| 1786 | |
Ming Lei | 08e98fc | 2014-09-25 23:23:38 +0800 | [diff] [blame] | 1787 | static int blk_mq_init_hctx(struct request_queue *q, |
| 1788 | struct blk_mq_tag_set *set, |
| 1789 | struct blk_mq_hw_ctx *hctx, unsigned hctx_idx) |
| 1790 | { |
| 1791 | int node; |
Ming Lei | f70ced0 | 2014-09-25 23:23:47 +0800 | [diff] [blame] | 1792 | unsigned flush_start_tag = set->queue_depth; |
Ming Lei | 08e98fc | 2014-09-25 23:23:38 +0800 | [diff] [blame] | 1793 | |
| 1794 | node = hctx->numa_node; |
| 1795 | if (node == NUMA_NO_NODE) |
| 1796 | node = hctx->numa_node = set->numa_node; |
| 1797 | |
Jens Axboe | 27489a3 | 2016-08-24 15:54:25 -0600 | [diff] [blame] | 1798 | INIT_WORK(&hctx->run_work, blk_mq_run_work_fn); |
Ming Lei | 08e98fc | 2014-09-25 23:23:38 +0800 | [diff] [blame] | 1799 | INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn); |
| 1800 | spin_lock_init(&hctx->lock); |
| 1801 | INIT_LIST_HEAD(&hctx->dispatch); |
| 1802 | hctx->queue = q; |
| 1803 | hctx->queue_num = hctx_idx; |
Jeff Moyer | 2404e60 | 2015-11-03 10:40:06 -0500 | [diff] [blame] | 1804 | hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED; |
Ming Lei | 08e98fc | 2014-09-25 23:23:38 +0800 | [diff] [blame] | 1805 | |
Thomas Gleixner | 9467f85 | 2016-09-22 08:05:17 -0600 | [diff] [blame] | 1806 | cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead); |
Ming Lei | 08e98fc | 2014-09-25 23:23:38 +0800 | [diff] [blame] | 1807 | |
| 1808 | hctx->tags = set->tags[hctx_idx]; |
| 1809 | |
| 1810 | /* |
| 1811 | * Allocate space for all possible cpus to avoid allocation at |
| 1812 | * runtime |
| 1813 | */ |
| 1814 | hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *), |
| 1815 | GFP_KERNEL, node); |
| 1816 | if (!hctx->ctxs) |
| 1817 | goto unregister_cpu_notifier; |
| 1818 | |
Omar Sandoval | 8845964 | 2016-09-17 08:38:44 -0600 | [diff] [blame] | 1819 | if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL, |
| 1820 | node)) |
Ming Lei | 08e98fc | 2014-09-25 23:23:38 +0800 | [diff] [blame] | 1821 | goto free_ctxs; |
| 1822 | |
| 1823 | hctx->nr_ctx = 0; |
| 1824 | |
| 1825 | if (set->ops->init_hctx && |
| 1826 | set->ops->init_hctx(hctx, set->driver_data, hctx_idx)) |
| 1827 | goto free_bitmap; |
| 1828 | |
Ming Lei | f70ced0 | 2014-09-25 23:23:47 +0800 | [diff] [blame] | 1829 | hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size); |
| 1830 | if (!hctx->fq) |
| 1831 | goto exit_hctx; |
| 1832 | |
| 1833 | if (set->ops->init_request && |
| 1834 | set->ops->init_request(set->driver_data, |
| 1835 | hctx->fq->flush_rq, hctx_idx, |
| 1836 | flush_start_tag + hctx_idx, node)) |
| 1837 | goto free_fq; |
| 1838 | |
Bart Van Assche | 6a83e74 | 2016-11-02 10:09:51 -0600 | [diff] [blame] | 1839 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
| 1840 | init_srcu_struct(&hctx->queue_rq_srcu); |
| 1841 | |
Ming Lei | 08e98fc | 2014-09-25 23:23:38 +0800 | [diff] [blame] | 1842 | return 0; |
| 1843 | |
Ming Lei | f70ced0 | 2014-09-25 23:23:47 +0800 | [diff] [blame] | 1844 | free_fq: |
| 1845 | kfree(hctx->fq); |
| 1846 | exit_hctx: |
| 1847 | if (set->ops->exit_hctx) |
| 1848 | set->ops->exit_hctx(hctx, hctx_idx); |
Ming Lei | 08e98fc | 2014-09-25 23:23:38 +0800 | [diff] [blame] | 1849 | free_bitmap: |
Omar Sandoval | 8845964 | 2016-09-17 08:38:44 -0600 | [diff] [blame] | 1850 | sbitmap_free(&hctx->ctx_map); |
Ming Lei | 08e98fc | 2014-09-25 23:23:38 +0800 | [diff] [blame] | 1851 | free_ctxs: |
| 1852 | kfree(hctx->ctxs); |
| 1853 | unregister_cpu_notifier: |
Thomas Gleixner | 9467f85 | 2016-09-22 08:05:17 -0600 | [diff] [blame] | 1854 | blk_mq_remove_cpuhp(hctx); |
Ming Lei | 08e98fc | 2014-09-25 23:23:38 +0800 | [diff] [blame] | 1855 | return -1; |
| 1856 | } |
| 1857 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1858 | static void blk_mq_init_cpu_queues(struct request_queue *q, |
| 1859 | unsigned int nr_hw_queues) |
| 1860 | { |
| 1861 | unsigned int i; |
| 1862 | |
| 1863 | for_each_possible_cpu(i) { |
| 1864 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); |
| 1865 | struct blk_mq_hw_ctx *hctx; |
| 1866 | |
| 1867 | memset(__ctx, 0, sizeof(*__ctx)); |
| 1868 | __ctx->cpu = i; |
| 1869 | spin_lock_init(&__ctx->lock); |
| 1870 | INIT_LIST_HEAD(&__ctx->rq_list); |
| 1871 | __ctx->queue = q; |
Jens Axboe | cf43e6b | 2016-11-07 21:32:37 -0700 | [diff] [blame] | 1872 | blk_stat_init(&__ctx->stat[BLK_STAT_READ]); |
| 1873 | blk_stat_init(&__ctx->stat[BLK_STAT_WRITE]); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1874 | |
| 1875 | /* 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] | 1876 | if (!cpu_online(i)) |
| 1877 | continue; |
| 1878 | |
Christoph Hellwig | 7d7e0f9 | 2016-09-14 16:18:54 +0200 | [diff] [blame] | 1879 | hctx = blk_mq_map_queue(q, i); |
Jens Axboe | e4043dc | 2014-04-09 10:18:23 -0600 | [diff] [blame] | 1880 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1881 | /* |
| 1882 | * Set local node, IFF we have more than one hw queue. If |
| 1883 | * not, we remain on the home node of the device |
| 1884 | */ |
| 1885 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) |
Raghavendra K T | bffed45 | 2015-12-02 16:59:05 +0530 | [diff] [blame] | 1886 | hctx->numa_node = local_memory_node(cpu_to_node(i)); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1887 | } |
| 1888 | } |
| 1889 | |
Jens Axboe | cc71a6f | 2017-01-11 14:29:56 -0700 | [diff] [blame] | 1890 | static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx) |
| 1891 | { |
| 1892 | int ret = 0; |
| 1893 | |
| 1894 | set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx, |
| 1895 | set->queue_depth, set->reserved_tags); |
| 1896 | if (!set->tags[hctx_idx]) |
| 1897 | return false; |
| 1898 | |
| 1899 | ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx, |
| 1900 | set->queue_depth); |
| 1901 | if (!ret) |
| 1902 | return true; |
| 1903 | |
| 1904 | blk_mq_free_rq_map(set->tags[hctx_idx]); |
| 1905 | set->tags[hctx_idx] = NULL; |
| 1906 | return false; |
| 1907 | } |
| 1908 | |
| 1909 | static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set, |
| 1910 | unsigned int hctx_idx) |
| 1911 | { |
| 1912 | blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx); |
| 1913 | blk_mq_free_rq_map(set->tags[hctx_idx]); |
| 1914 | set->tags[hctx_idx] = NULL; |
| 1915 | } |
| 1916 | |
Akinobu Mita | 5778322 | 2015-09-27 02:09:23 +0900 | [diff] [blame] | 1917 | static void blk_mq_map_swqueue(struct request_queue *q, |
| 1918 | const struct cpumask *online_mask) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1919 | { |
Gabriel Krisman Bertazi | d1b1cea | 2016-12-14 18:48:36 -0200 | [diff] [blame] | 1920 | unsigned int i, hctx_idx; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1921 | struct blk_mq_hw_ctx *hctx; |
| 1922 | struct blk_mq_ctx *ctx; |
Ming Lei | 2a34c08 | 2015-04-21 10:00:20 +0800 | [diff] [blame] | 1923 | struct blk_mq_tag_set *set = q->tag_set; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1924 | |
Akinobu Mita | 60de074 | 2015-09-27 02:09:25 +0900 | [diff] [blame] | 1925 | /* |
| 1926 | * Avoid others reading imcomplete hctx->cpumask through sysfs |
| 1927 | */ |
| 1928 | mutex_lock(&q->sysfs_lock); |
| 1929 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1930 | queue_for_each_hw_ctx(q, hctx, i) { |
Jens Axboe | e4043dc | 2014-04-09 10:18:23 -0600 | [diff] [blame] | 1931 | cpumask_clear(hctx->cpumask); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1932 | hctx->nr_ctx = 0; |
| 1933 | } |
| 1934 | |
| 1935 | /* |
| 1936 | * Map software to hardware queues |
| 1937 | */ |
Thomas Gleixner | 897bb0c | 2016-03-19 11:30:33 +0100 | [diff] [blame] | 1938 | for_each_possible_cpu(i) { |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1939 | /* If the cpu isn't online, the cpu is mapped to first hctx */ |
Akinobu Mita | 5778322 | 2015-09-27 02:09:23 +0900 | [diff] [blame] | 1940 | if (!cpumask_test_cpu(i, online_mask)) |
Jens Axboe | e4043dc | 2014-04-09 10:18:23 -0600 | [diff] [blame] | 1941 | continue; |
| 1942 | |
Gabriel Krisman Bertazi | d1b1cea | 2016-12-14 18:48:36 -0200 | [diff] [blame] | 1943 | hctx_idx = q->mq_map[i]; |
| 1944 | /* unmapped hw queue can be remapped after CPU topo changed */ |
Jens Axboe | cc71a6f | 2017-01-11 14:29:56 -0700 | [diff] [blame] | 1945 | if (!set->tags[hctx_idx] && |
| 1946 | !__blk_mq_alloc_rq_map(set, hctx_idx)) { |
Gabriel Krisman Bertazi | d1b1cea | 2016-12-14 18:48:36 -0200 | [diff] [blame] | 1947 | /* |
| 1948 | * If tags initialization fail for some hctx, |
| 1949 | * that hctx won't be brought online. In this |
| 1950 | * case, remap the current ctx to hctx[0] which |
| 1951 | * is guaranteed to always have tags allocated |
| 1952 | */ |
Jens Axboe | cc71a6f | 2017-01-11 14:29:56 -0700 | [diff] [blame] | 1953 | q->mq_map[i] = 0; |
Gabriel Krisman Bertazi | d1b1cea | 2016-12-14 18:48:36 -0200 | [diff] [blame] | 1954 | } |
| 1955 | |
Thomas Gleixner | 897bb0c | 2016-03-19 11:30:33 +0100 | [diff] [blame] | 1956 | ctx = per_cpu_ptr(q->queue_ctx, i); |
Christoph Hellwig | 7d7e0f9 | 2016-09-14 16:18:54 +0200 | [diff] [blame] | 1957 | hctx = blk_mq_map_queue(q, i); |
Keith Busch | 868f2f0 | 2015-12-17 17:08:14 -0700 | [diff] [blame] | 1958 | |
Jens Axboe | e4043dc | 2014-04-09 10:18:23 -0600 | [diff] [blame] | 1959 | cpumask_set_cpu(i, hctx->cpumask); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1960 | ctx->index_hw = hctx->nr_ctx; |
| 1961 | hctx->ctxs[hctx->nr_ctx++] = ctx; |
| 1962 | } |
Jens Axboe | 506e931 | 2014-05-07 10:26:44 -0600 | [diff] [blame] | 1963 | |
Akinobu Mita | 60de074 | 2015-09-27 02:09:25 +0900 | [diff] [blame] | 1964 | mutex_unlock(&q->sysfs_lock); |
| 1965 | |
Jens Axboe | 506e931 | 2014-05-07 10:26:44 -0600 | [diff] [blame] | 1966 | queue_for_each_hw_ctx(q, hctx, i) { |
Jens Axboe | 484b406 | 2014-05-21 14:01:15 -0600 | [diff] [blame] | 1967 | /* |
Jens Axboe | a68aafa | 2014-08-15 13:19:15 -0600 | [diff] [blame] | 1968 | * If no software queues are mapped to this hardware queue, |
| 1969 | * disable it and free the request entries. |
Jens Axboe | 484b406 | 2014-05-21 14:01:15 -0600 | [diff] [blame] | 1970 | */ |
| 1971 | if (!hctx->nr_ctx) { |
Gabriel Krisman Bertazi | d1b1cea | 2016-12-14 18:48:36 -0200 | [diff] [blame] | 1972 | /* Never unmap queue 0. We need it as a |
| 1973 | * fallback in case of a new remap fails |
| 1974 | * allocation |
| 1975 | */ |
Jens Axboe | cc71a6f | 2017-01-11 14:29:56 -0700 | [diff] [blame] | 1976 | if (i && set->tags[i]) |
| 1977 | blk_mq_free_map_and_requests(set, i); |
| 1978 | |
Ming Lei | 2a34c08 | 2015-04-21 10:00:20 +0800 | [diff] [blame] | 1979 | hctx->tags = NULL; |
Jens Axboe | 484b406 | 2014-05-21 14:01:15 -0600 | [diff] [blame] | 1980 | continue; |
| 1981 | } |
| 1982 | |
Ming Lei | 2a34c08 | 2015-04-21 10:00:20 +0800 | [diff] [blame] | 1983 | hctx->tags = set->tags[i]; |
| 1984 | WARN_ON(!hctx->tags); |
| 1985 | |
Jens Axboe | 484b406 | 2014-05-21 14:01:15 -0600 | [diff] [blame] | 1986 | /* |
Chong Yuan | 889fa31 | 2015-04-15 11:39:29 -0600 | [diff] [blame] | 1987 | * Set the map size to the number of mapped software queues. |
| 1988 | * This is more accurate and more efficient than looping |
| 1989 | * over all possibly mapped software queues. |
| 1990 | */ |
Omar Sandoval | 8845964 | 2016-09-17 08:38:44 -0600 | [diff] [blame] | 1991 | sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx); |
Chong Yuan | 889fa31 | 2015-04-15 11:39:29 -0600 | [diff] [blame] | 1992 | |
| 1993 | /* |
Jens Axboe | 484b406 | 2014-05-21 14:01:15 -0600 | [diff] [blame] | 1994 | * Initialize batch roundrobin counts |
| 1995 | */ |
Jens Axboe | 506e931 | 2014-05-07 10:26:44 -0600 | [diff] [blame] | 1996 | hctx->next_cpu = cpumask_first(hctx->cpumask); |
| 1997 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; |
| 1998 | } |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1999 | } |
| 2000 | |
Jeff Moyer | 2404e60 | 2015-11-03 10:40:06 -0500 | [diff] [blame] | 2001 | static void queue_set_hctx_shared(struct request_queue *q, bool shared) |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 2002 | { |
| 2003 | struct blk_mq_hw_ctx *hctx; |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 2004 | int i; |
| 2005 | |
Jeff Moyer | 2404e60 | 2015-11-03 10:40:06 -0500 | [diff] [blame] | 2006 | queue_for_each_hw_ctx(q, hctx, i) { |
| 2007 | if (shared) |
| 2008 | hctx->flags |= BLK_MQ_F_TAG_SHARED; |
| 2009 | else |
| 2010 | hctx->flags &= ~BLK_MQ_F_TAG_SHARED; |
| 2011 | } |
| 2012 | } |
| 2013 | |
| 2014 | static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, bool shared) |
| 2015 | { |
| 2016 | struct request_queue *q; |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 2017 | |
| 2018 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
| 2019 | blk_mq_freeze_queue(q); |
Jeff Moyer | 2404e60 | 2015-11-03 10:40:06 -0500 | [diff] [blame] | 2020 | queue_set_hctx_shared(q, shared); |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 2021 | blk_mq_unfreeze_queue(q); |
| 2022 | } |
| 2023 | } |
| 2024 | |
| 2025 | static void blk_mq_del_queue_tag_set(struct request_queue *q) |
| 2026 | { |
| 2027 | struct blk_mq_tag_set *set = q->tag_set; |
| 2028 | |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 2029 | mutex_lock(&set->tag_list_lock); |
| 2030 | list_del_init(&q->tag_set_list); |
Jeff Moyer | 2404e60 | 2015-11-03 10:40:06 -0500 | [diff] [blame] | 2031 | if (list_is_singular(&set->tag_list)) { |
| 2032 | /* just transitioned to unshared */ |
| 2033 | set->flags &= ~BLK_MQ_F_TAG_SHARED; |
| 2034 | /* update existing queue */ |
| 2035 | blk_mq_update_tag_set_depth(set, false); |
| 2036 | } |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 2037 | mutex_unlock(&set->tag_list_lock); |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 2038 | } |
| 2039 | |
| 2040 | static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, |
| 2041 | struct request_queue *q) |
| 2042 | { |
| 2043 | q->tag_set = set; |
| 2044 | |
| 2045 | mutex_lock(&set->tag_list_lock); |
Jeff Moyer | 2404e60 | 2015-11-03 10:40:06 -0500 | [diff] [blame] | 2046 | |
| 2047 | /* Check to see if we're transitioning to shared (from 1 to 2 queues). */ |
| 2048 | if (!list_empty(&set->tag_list) && !(set->flags & BLK_MQ_F_TAG_SHARED)) { |
| 2049 | set->flags |= BLK_MQ_F_TAG_SHARED; |
| 2050 | /* update existing queue */ |
| 2051 | blk_mq_update_tag_set_depth(set, true); |
| 2052 | } |
| 2053 | if (set->flags & BLK_MQ_F_TAG_SHARED) |
| 2054 | queue_set_hctx_shared(q, true); |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 2055 | list_add_tail(&q->tag_set_list, &set->tag_list); |
Jeff Moyer | 2404e60 | 2015-11-03 10:40:06 -0500 | [diff] [blame] | 2056 | |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 2057 | mutex_unlock(&set->tag_list_lock); |
| 2058 | } |
| 2059 | |
Ming Lei | e09aae7 | 2015-01-29 20:17:27 +0800 | [diff] [blame] | 2060 | /* |
| 2061 | * It is the actual release handler for mq, but we do it from |
| 2062 | * request queue's release handler for avoiding use-after-free |
| 2063 | * and headache because q->mq_kobj shouldn't have been introduced, |
| 2064 | * but we can't group ctx/kctx kobj without it. |
| 2065 | */ |
| 2066 | void blk_mq_release(struct request_queue *q) |
| 2067 | { |
| 2068 | struct blk_mq_hw_ctx *hctx; |
| 2069 | unsigned int i; |
| 2070 | |
| 2071 | /* hctx kobj stays in hctx */ |
Ming Lei | c3b4afc | 2015-06-04 22:25:04 +0800 | [diff] [blame] | 2072 | queue_for_each_hw_ctx(q, hctx, i) { |
| 2073 | if (!hctx) |
| 2074 | continue; |
| 2075 | kfree(hctx->ctxs); |
Ming Lei | e09aae7 | 2015-01-29 20:17:27 +0800 | [diff] [blame] | 2076 | kfree(hctx); |
Ming Lei | c3b4afc | 2015-06-04 22:25:04 +0800 | [diff] [blame] | 2077 | } |
Ming Lei | e09aae7 | 2015-01-29 20:17:27 +0800 | [diff] [blame] | 2078 | |
Akinobu Mita | a723bab | 2015-09-27 02:09:21 +0900 | [diff] [blame] | 2079 | q->mq_map = NULL; |
| 2080 | |
Ming Lei | e09aae7 | 2015-01-29 20:17:27 +0800 | [diff] [blame] | 2081 | kfree(q->queue_hw_ctx); |
| 2082 | |
| 2083 | /* ctx kobj stays in queue_ctx */ |
| 2084 | free_percpu(q->queue_ctx); |
| 2085 | } |
| 2086 | |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 2087 | 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] | 2088 | { |
Mike Snitzer | b62c21b | 2015-03-12 23:56:02 -0400 | [diff] [blame] | 2089 | struct request_queue *uninit_q, *q; |
| 2090 | |
| 2091 | uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node); |
| 2092 | if (!uninit_q) |
| 2093 | return ERR_PTR(-ENOMEM); |
| 2094 | |
| 2095 | q = blk_mq_init_allocated_queue(set, uninit_q); |
| 2096 | if (IS_ERR(q)) |
| 2097 | blk_cleanup_queue(uninit_q); |
| 2098 | |
| 2099 | return q; |
| 2100 | } |
| 2101 | EXPORT_SYMBOL(blk_mq_init_queue); |
| 2102 | |
Keith Busch | 868f2f0 | 2015-12-17 17:08:14 -0700 | [diff] [blame] | 2103 | static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set, |
| 2104 | struct request_queue *q) |
Mike Snitzer | b62c21b | 2015-03-12 23:56:02 -0400 | [diff] [blame] | 2105 | { |
Keith Busch | 868f2f0 | 2015-12-17 17:08:14 -0700 | [diff] [blame] | 2106 | int i, j; |
| 2107 | struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2108 | |
Keith Busch | 868f2f0 | 2015-12-17 17:08:14 -0700 | [diff] [blame] | 2109 | blk_mq_sysfs_unregister(q); |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 2110 | for (i = 0; i < set->nr_hw_queues; i++) { |
Keith Busch | 868f2f0 | 2015-12-17 17:08:14 -0700 | [diff] [blame] | 2111 | int node; |
Jens Axboe | f14bbe7 | 2014-05-27 12:06:53 -0600 | [diff] [blame] | 2112 | |
Keith Busch | 868f2f0 | 2015-12-17 17:08:14 -0700 | [diff] [blame] | 2113 | if (hctxs[i]) |
| 2114 | continue; |
| 2115 | |
| 2116 | node = blk_mq_hw_queue_to_node(q->mq_map, i); |
Christoph Hellwig | cdef54d | 2014-05-28 18:11:06 +0200 | [diff] [blame] | 2117 | hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx), |
| 2118 | GFP_KERNEL, node); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2119 | if (!hctxs[i]) |
Keith Busch | 868f2f0 | 2015-12-17 17:08:14 -0700 | [diff] [blame] | 2120 | break; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2121 | |
Jens Axboe | a86073e | 2014-10-13 15:41:54 -0600 | [diff] [blame] | 2122 | if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL, |
Keith Busch | 868f2f0 | 2015-12-17 17:08:14 -0700 | [diff] [blame] | 2123 | node)) { |
| 2124 | kfree(hctxs[i]); |
| 2125 | hctxs[i] = NULL; |
| 2126 | break; |
| 2127 | } |
Jens Axboe | e4043dc | 2014-04-09 10:18:23 -0600 | [diff] [blame] | 2128 | |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 2129 | atomic_set(&hctxs[i]->nr_active, 0); |
Jens Axboe | f14bbe7 | 2014-05-27 12:06:53 -0600 | [diff] [blame] | 2130 | hctxs[i]->numa_node = node; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2131 | hctxs[i]->queue_num = i; |
Keith Busch | 868f2f0 | 2015-12-17 17:08:14 -0700 | [diff] [blame] | 2132 | |
| 2133 | if (blk_mq_init_hctx(q, set, hctxs[i], i)) { |
| 2134 | free_cpumask_var(hctxs[i]->cpumask); |
| 2135 | kfree(hctxs[i]); |
| 2136 | hctxs[i] = NULL; |
| 2137 | break; |
| 2138 | } |
| 2139 | blk_mq_hctx_kobj_init(hctxs[i]); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2140 | } |
Keith Busch | 868f2f0 | 2015-12-17 17:08:14 -0700 | [diff] [blame] | 2141 | for (j = i; j < q->nr_hw_queues; j++) { |
| 2142 | struct blk_mq_hw_ctx *hctx = hctxs[j]; |
| 2143 | |
| 2144 | if (hctx) { |
Jens Axboe | cc71a6f | 2017-01-11 14:29:56 -0700 | [diff] [blame] | 2145 | if (hctx->tags) |
| 2146 | blk_mq_free_map_and_requests(set, j); |
Keith Busch | 868f2f0 | 2015-12-17 17:08:14 -0700 | [diff] [blame] | 2147 | blk_mq_exit_hctx(q, set, hctx, j); |
| 2148 | free_cpumask_var(hctx->cpumask); |
| 2149 | kobject_put(&hctx->kobj); |
| 2150 | kfree(hctx->ctxs); |
| 2151 | kfree(hctx); |
| 2152 | hctxs[j] = NULL; |
| 2153 | |
| 2154 | } |
| 2155 | } |
| 2156 | q->nr_hw_queues = i; |
| 2157 | blk_mq_sysfs_register(q); |
| 2158 | } |
| 2159 | |
| 2160 | struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set, |
| 2161 | struct request_queue *q) |
| 2162 | { |
Ming Lei | 6684167 | 2016-02-12 15:27:00 +0800 | [diff] [blame] | 2163 | /* mark the queue as mq asap */ |
| 2164 | q->mq_ops = set->ops; |
| 2165 | |
Keith Busch | 868f2f0 | 2015-12-17 17:08:14 -0700 | [diff] [blame] | 2166 | q->queue_ctx = alloc_percpu(struct blk_mq_ctx); |
| 2167 | if (!q->queue_ctx) |
Ming Lin | c7de572 | 2016-05-25 23:23:27 -0700 | [diff] [blame] | 2168 | goto err_exit; |
Keith Busch | 868f2f0 | 2015-12-17 17:08:14 -0700 | [diff] [blame] | 2169 | |
| 2170 | q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)), |
| 2171 | GFP_KERNEL, set->numa_node); |
| 2172 | if (!q->queue_hw_ctx) |
| 2173 | goto err_percpu; |
| 2174 | |
Christoph Hellwig | bdd17e7 | 2016-09-14 16:18:53 +0200 | [diff] [blame] | 2175 | q->mq_map = set->mq_map; |
Keith Busch | 868f2f0 | 2015-12-17 17:08:14 -0700 | [diff] [blame] | 2176 | |
| 2177 | blk_mq_realloc_hw_ctxs(set, q); |
| 2178 | if (!q->nr_hw_queues) |
| 2179 | goto err_hctxs; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2180 | |
Christoph Hellwig | 287922eb | 2015-10-30 20:57:30 +0800 | [diff] [blame] | 2181 | INIT_WORK(&q->timeout_work, blk_mq_timeout_work); |
Ming Lei | e56f698 | 2015-07-16 19:53:22 +0800 | [diff] [blame] | 2182 | blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2183 | |
| 2184 | q->nr_queues = nr_cpu_ids; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2185 | |
Jens Axboe | 94eddfb | 2013-11-19 09:25:07 -0700 | [diff] [blame] | 2186 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2187 | |
Jens Axboe | 05f1dd5 | 2014-05-29 09:53:32 -0600 | [diff] [blame] | 2188 | if (!(set->flags & BLK_MQ_F_SG_MERGE)) |
| 2189 | q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE; |
| 2190 | |
Christoph Hellwig | 1be036e | 2014-02-07 10:22:39 -0800 | [diff] [blame] | 2191 | q->sg_reserved_size = INT_MAX; |
| 2192 | |
Mike Snitzer | 2849450 | 2016-09-14 13:28:30 -0400 | [diff] [blame] | 2193 | INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work); |
Christoph Hellwig | 6fca6a6 | 2014-05-28 08:08:02 -0600 | [diff] [blame] | 2194 | INIT_LIST_HEAD(&q->requeue_list); |
| 2195 | spin_lock_init(&q->requeue_lock); |
| 2196 | |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 2197 | if (q->nr_hw_queues > 1) |
| 2198 | blk_queue_make_request(q, blk_mq_make_request); |
| 2199 | else |
| 2200 | blk_queue_make_request(q, blk_sq_make_request); |
| 2201 | |
Jens Axboe | eba7176 | 2014-05-20 15:17:27 -0600 | [diff] [blame] | 2202 | /* |
| 2203 | * Do this after blk_queue_make_request() overrides it... |
| 2204 | */ |
| 2205 | q->nr_requests = set->queue_depth; |
| 2206 | |
Jens Axboe | 64f1c21 | 2016-11-14 13:03:03 -0700 | [diff] [blame] | 2207 | /* |
| 2208 | * Default to classic polling |
| 2209 | */ |
| 2210 | q->poll_nsec = -1; |
| 2211 | |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 2212 | if (set->ops->complete) |
| 2213 | blk_queue_softirq_done(q, set->ops->complete); |
Christoph Hellwig | 30a91cb | 2014-02-10 03:24:38 -0800 | [diff] [blame] | 2214 | |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 2215 | blk_mq_init_cpu_queues(q, set->nr_hw_queues); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2216 | |
Akinobu Mita | 5778322 | 2015-09-27 02:09:23 +0900 | [diff] [blame] | 2217 | get_online_cpus(); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2218 | mutex_lock(&all_q_mutex); |
Akinobu Mita | 4593fdb | 2015-09-27 02:09:20 +0900 | [diff] [blame] | 2219 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2220 | list_add_tail(&q->all_q_node, &all_q_list); |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 2221 | blk_mq_add_queue_tag_set(set, q); |
Akinobu Mita | 5778322 | 2015-09-27 02:09:23 +0900 | [diff] [blame] | 2222 | blk_mq_map_swqueue(q, cpu_online_mask); |
Jens Axboe | 484b406 | 2014-05-21 14:01:15 -0600 | [diff] [blame] | 2223 | |
Akinobu Mita | 4593fdb | 2015-09-27 02:09:20 +0900 | [diff] [blame] | 2224 | mutex_unlock(&all_q_mutex); |
Akinobu Mita | 5778322 | 2015-09-27 02:09:23 +0900 | [diff] [blame] | 2225 | put_online_cpus(); |
Akinobu Mita | 4593fdb | 2015-09-27 02:09:20 +0900 | [diff] [blame] | 2226 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2227 | return q; |
Christoph Hellwig | 1874198 | 2014-02-10 09:29:00 -0700 | [diff] [blame] | 2228 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2229 | err_hctxs: |
Keith Busch | 868f2f0 | 2015-12-17 17:08:14 -0700 | [diff] [blame] | 2230 | kfree(q->queue_hw_ctx); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2231 | err_percpu: |
Keith Busch | 868f2f0 | 2015-12-17 17:08:14 -0700 | [diff] [blame] | 2232 | free_percpu(q->queue_ctx); |
Ming Lin | c7de572 | 2016-05-25 23:23:27 -0700 | [diff] [blame] | 2233 | err_exit: |
| 2234 | q->mq_ops = NULL; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2235 | return ERR_PTR(-ENOMEM); |
| 2236 | } |
Mike Snitzer | b62c21b | 2015-03-12 23:56:02 -0400 | [diff] [blame] | 2237 | EXPORT_SYMBOL(blk_mq_init_allocated_queue); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2238 | |
| 2239 | void blk_mq_free_queue(struct request_queue *q) |
| 2240 | { |
Ming Lei | 624dbe4 | 2014-05-27 23:35:13 +0800 | [diff] [blame] | 2241 | struct blk_mq_tag_set *set = q->tag_set; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2242 | |
Akinobu Mita | 0e62636 | 2015-09-27 02:09:22 +0900 | [diff] [blame] | 2243 | mutex_lock(&all_q_mutex); |
| 2244 | list_del_init(&q->all_q_node); |
| 2245 | mutex_unlock(&all_q_mutex); |
| 2246 | |
Jens Axboe | 87760e5 | 2016-11-09 12:38:14 -0700 | [diff] [blame] | 2247 | wbt_exit(q); |
| 2248 | |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 2249 | blk_mq_del_queue_tag_set(q); |
| 2250 | |
Ming Lei | 624dbe4 | 2014-05-27 23:35:13 +0800 | [diff] [blame] | 2251 | blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); |
| 2252 | blk_mq_free_hw_queues(q, set); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2253 | } |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2254 | |
| 2255 | /* Basically redo blk_mq_init_queue with queue frozen */ |
Akinobu Mita | 5778322 | 2015-09-27 02:09:23 +0900 | [diff] [blame] | 2256 | static void blk_mq_queue_reinit(struct request_queue *q, |
| 2257 | const struct cpumask *online_mask) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2258 | { |
Christoph Hellwig | 4ecd4fe | 2015-05-07 09:38:13 +0200 | [diff] [blame] | 2259 | WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth)); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2260 | |
Jens Axboe | 67aec14 | 2014-05-30 08:25:36 -0600 | [diff] [blame] | 2261 | blk_mq_sysfs_unregister(q); |
| 2262 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2263 | /* |
| 2264 | * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe |
| 2265 | * we should change hctx numa_node according to new topology (this |
| 2266 | * involves free and re-allocate memory, worthy doing?) |
| 2267 | */ |
| 2268 | |
Akinobu Mita | 5778322 | 2015-09-27 02:09:23 +0900 | [diff] [blame] | 2269 | blk_mq_map_swqueue(q, online_mask); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2270 | |
Jens Axboe | 67aec14 | 2014-05-30 08:25:36 -0600 | [diff] [blame] | 2271 | blk_mq_sysfs_register(q); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2272 | } |
| 2273 | |
Sebastian Andrzej Siewior | 65d5291 | 2016-09-22 08:05:19 -0600 | [diff] [blame] | 2274 | /* |
| 2275 | * New online cpumask which is going to be set in this hotplug event. |
| 2276 | * Declare this cpumasks as global as cpu-hotplug operation is invoked |
| 2277 | * one-by-one and dynamically allocating this could result in a failure. |
| 2278 | */ |
| 2279 | static struct cpumask cpuhp_online_new; |
| 2280 | |
| 2281 | static void blk_mq_queue_reinit_work(void) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2282 | { |
| 2283 | struct request_queue *q; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2284 | |
| 2285 | mutex_lock(&all_q_mutex); |
Tejun Heo | f3af020 | 2014-11-04 13:52:27 -0500 | [diff] [blame] | 2286 | /* |
| 2287 | * We need to freeze and reinit all existing queues. Freezing |
| 2288 | * involves synchronous wait for an RCU grace period and doing it |
| 2289 | * one by one may take a long time. Start freezing all queues in |
| 2290 | * one swoop and then wait for the completions so that freezing can |
| 2291 | * take place in parallel. |
| 2292 | */ |
| 2293 | list_for_each_entry(q, &all_q_list, all_q_node) |
| 2294 | blk_mq_freeze_queue_start(q); |
Gabriel Krisman Bertazi | 415d3da | 2016-11-28 15:01:48 -0200 | [diff] [blame] | 2295 | list_for_each_entry(q, &all_q_list, all_q_node) |
Tejun Heo | f3af020 | 2014-11-04 13:52:27 -0500 | [diff] [blame] | 2296 | blk_mq_freeze_queue_wait(q); |
| 2297 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2298 | list_for_each_entry(q, &all_q_list, all_q_node) |
Sebastian Andrzej Siewior | 65d5291 | 2016-09-22 08:05:19 -0600 | [diff] [blame] | 2299 | blk_mq_queue_reinit(q, &cpuhp_online_new); |
Tejun Heo | f3af020 | 2014-11-04 13:52:27 -0500 | [diff] [blame] | 2300 | |
| 2301 | list_for_each_entry(q, &all_q_list, all_q_node) |
| 2302 | blk_mq_unfreeze_queue(q); |
| 2303 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2304 | mutex_unlock(&all_q_mutex); |
Sebastian Andrzej Siewior | 65d5291 | 2016-09-22 08:05:19 -0600 | [diff] [blame] | 2305 | } |
| 2306 | |
| 2307 | static int blk_mq_queue_reinit_dead(unsigned int cpu) |
| 2308 | { |
Sebastian Andrzej Siewior | 97a3286 | 2016-09-23 15:02:38 +0200 | [diff] [blame] | 2309 | cpumask_copy(&cpuhp_online_new, cpu_online_mask); |
Sebastian Andrzej Siewior | 65d5291 | 2016-09-22 08:05:19 -0600 | [diff] [blame] | 2310 | blk_mq_queue_reinit_work(); |
| 2311 | return 0; |
| 2312 | } |
| 2313 | |
| 2314 | /* |
| 2315 | * Before hotadded cpu starts handling requests, new mappings must be |
| 2316 | * established. Otherwise, these requests in hw queue might never be |
| 2317 | * dispatched. |
| 2318 | * |
| 2319 | * For example, there is a single hw queue (hctx) and two CPU queues (ctx0 |
| 2320 | * for CPU0, and ctx1 for CPU1). |
| 2321 | * |
| 2322 | * Now CPU1 is just onlined and a request is inserted into ctx1->rq_list |
| 2323 | * and set bit0 in pending bitmap as ctx1->index_hw is still zero. |
| 2324 | * |
Jens Axboe | 2c3ad66 | 2016-12-14 14:34:47 -0700 | [diff] [blame] | 2325 | * And then while running hw queue, blk_mq_flush_busy_ctxs() finds bit0 is set |
| 2326 | * in pending bitmap and tries to retrieve requests in hctx->ctxs[0]->rq_list. |
| 2327 | * But htx->ctxs[0] is a pointer to ctx0, so the request in ctx1->rq_list is |
| 2328 | * ignored. |
Sebastian Andrzej Siewior | 65d5291 | 2016-09-22 08:05:19 -0600 | [diff] [blame] | 2329 | */ |
| 2330 | static int blk_mq_queue_reinit_prepare(unsigned int cpu) |
| 2331 | { |
| 2332 | cpumask_copy(&cpuhp_online_new, cpu_online_mask); |
| 2333 | cpumask_set_cpu(cpu, &cpuhp_online_new); |
| 2334 | blk_mq_queue_reinit_work(); |
| 2335 | return 0; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2336 | } |
| 2337 | |
Jens Axboe | a516440 | 2014-09-10 09:02:03 -0600 | [diff] [blame] | 2338 | static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) |
| 2339 | { |
| 2340 | int i; |
| 2341 | |
Jens Axboe | cc71a6f | 2017-01-11 14:29:56 -0700 | [diff] [blame] | 2342 | for (i = 0; i < set->nr_hw_queues; i++) |
| 2343 | if (!__blk_mq_alloc_rq_map(set, i)) |
Jens Axboe | a516440 | 2014-09-10 09:02:03 -0600 | [diff] [blame] | 2344 | goto out_unwind; |
Jens Axboe | a516440 | 2014-09-10 09:02:03 -0600 | [diff] [blame] | 2345 | |
| 2346 | return 0; |
| 2347 | |
| 2348 | out_unwind: |
| 2349 | while (--i >= 0) |
Jens Axboe | cc71a6f | 2017-01-11 14:29:56 -0700 | [diff] [blame] | 2350 | blk_mq_free_rq_map(set->tags[i]); |
Jens Axboe | a516440 | 2014-09-10 09:02:03 -0600 | [diff] [blame] | 2351 | |
Jens Axboe | a516440 | 2014-09-10 09:02:03 -0600 | [diff] [blame] | 2352 | return -ENOMEM; |
| 2353 | } |
| 2354 | |
| 2355 | /* |
| 2356 | * Allocate the request maps associated with this tag_set. Note that this |
| 2357 | * may reduce the depth asked for, if memory is tight. set->queue_depth |
| 2358 | * will be updated to reflect the allocated depth. |
| 2359 | */ |
| 2360 | static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) |
| 2361 | { |
| 2362 | unsigned int depth; |
| 2363 | int err; |
| 2364 | |
| 2365 | depth = set->queue_depth; |
| 2366 | do { |
| 2367 | err = __blk_mq_alloc_rq_maps(set); |
| 2368 | if (!err) |
| 2369 | break; |
| 2370 | |
| 2371 | set->queue_depth >>= 1; |
| 2372 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) { |
| 2373 | err = -ENOMEM; |
| 2374 | break; |
| 2375 | } |
| 2376 | } while (set->queue_depth); |
| 2377 | |
| 2378 | if (!set->queue_depth || err) { |
| 2379 | pr_err("blk-mq: failed to allocate request map\n"); |
| 2380 | return -ENOMEM; |
| 2381 | } |
| 2382 | |
| 2383 | if (depth != set->queue_depth) |
| 2384 | pr_info("blk-mq: reduced tag depth (%u -> %u)\n", |
| 2385 | depth, set->queue_depth); |
| 2386 | |
| 2387 | return 0; |
| 2388 | } |
| 2389 | |
Jens Axboe | a4391c6 | 2014-06-05 15:21:56 -0600 | [diff] [blame] | 2390 | /* |
| 2391 | * Alloc a tag set to be associated with one or more request queues. |
| 2392 | * May fail with EINVAL for various error conditions. May adjust the |
| 2393 | * requested depth down, if if it too large. In that case, the set |
| 2394 | * value will be stored in set->queue_depth. |
| 2395 | */ |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 2396 | int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) |
| 2397 | { |
Christoph Hellwig | da695ba | 2016-09-14 16:18:55 +0200 | [diff] [blame] | 2398 | int ret; |
| 2399 | |
Bart Van Assche | 205fb5f | 2014-10-30 14:45:11 +0100 | [diff] [blame] | 2400 | BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS); |
| 2401 | |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 2402 | if (!set->nr_hw_queues) |
| 2403 | return -EINVAL; |
Jens Axboe | a4391c6 | 2014-06-05 15:21:56 -0600 | [diff] [blame] | 2404 | if (!set->queue_depth) |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 2405 | return -EINVAL; |
| 2406 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) |
| 2407 | return -EINVAL; |
| 2408 | |
Christoph Hellwig | 7d7e0f9 | 2016-09-14 16:18:54 +0200 | [diff] [blame] | 2409 | if (!set->ops->queue_rq) |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 2410 | return -EINVAL; |
| 2411 | |
Jens Axboe | a4391c6 | 2014-06-05 15:21:56 -0600 | [diff] [blame] | 2412 | if (set->queue_depth > BLK_MQ_MAX_DEPTH) { |
| 2413 | pr_info("blk-mq: reduced tag depth to %u\n", |
| 2414 | BLK_MQ_MAX_DEPTH); |
| 2415 | set->queue_depth = BLK_MQ_MAX_DEPTH; |
| 2416 | } |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 2417 | |
Shaohua Li | 6637fad | 2014-11-30 16:00:58 -0800 | [diff] [blame] | 2418 | /* |
| 2419 | * If a crashdump is active, then we are potentially in a very |
| 2420 | * memory constrained environment. Limit us to 1 queue and |
| 2421 | * 64 tags to prevent using too much memory. |
| 2422 | */ |
| 2423 | if (is_kdump_kernel()) { |
| 2424 | set->nr_hw_queues = 1; |
| 2425 | set->queue_depth = min(64U, set->queue_depth); |
| 2426 | } |
Keith Busch | 868f2f0 | 2015-12-17 17:08:14 -0700 | [diff] [blame] | 2427 | /* |
| 2428 | * There is no use for more h/w queues than cpus. |
| 2429 | */ |
| 2430 | if (set->nr_hw_queues > nr_cpu_ids) |
| 2431 | set->nr_hw_queues = nr_cpu_ids; |
Shaohua Li | 6637fad | 2014-11-30 16:00:58 -0800 | [diff] [blame] | 2432 | |
Keith Busch | 868f2f0 | 2015-12-17 17:08:14 -0700 | [diff] [blame] | 2433 | set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *), |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 2434 | GFP_KERNEL, set->numa_node); |
| 2435 | if (!set->tags) |
Jens Axboe | a516440 | 2014-09-10 09:02:03 -0600 | [diff] [blame] | 2436 | return -ENOMEM; |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 2437 | |
Christoph Hellwig | da695ba | 2016-09-14 16:18:55 +0200 | [diff] [blame] | 2438 | ret = -ENOMEM; |
| 2439 | set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids, |
| 2440 | GFP_KERNEL, set->numa_node); |
Christoph Hellwig | bdd17e7 | 2016-09-14 16:18:53 +0200 | [diff] [blame] | 2441 | if (!set->mq_map) |
| 2442 | goto out_free_tags; |
| 2443 | |
Christoph Hellwig | da695ba | 2016-09-14 16:18:55 +0200 | [diff] [blame] | 2444 | if (set->ops->map_queues) |
| 2445 | ret = set->ops->map_queues(set); |
| 2446 | else |
| 2447 | ret = blk_mq_map_queues(set); |
| 2448 | if (ret) |
| 2449 | goto out_free_mq_map; |
| 2450 | |
| 2451 | ret = blk_mq_alloc_rq_maps(set); |
| 2452 | if (ret) |
Christoph Hellwig | bdd17e7 | 2016-09-14 16:18:53 +0200 | [diff] [blame] | 2453 | goto out_free_mq_map; |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 2454 | |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 2455 | mutex_init(&set->tag_list_lock); |
| 2456 | INIT_LIST_HEAD(&set->tag_list); |
| 2457 | |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 2458 | return 0; |
Christoph Hellwig | bdd17e7 | 2016-09-14 16:18:53 +0200 | [diff] [blame] | 2459 | |
| 2460 | out_free_mq_map: |
| 2461 | kfree(set->mq_map); |
| 2462 | set->mq_map = NULL; |
| 2463 | out_free_tags: |
Robert Elliott | 5676e7b | 2014-09-02 11:38:44 -0500 | [diff] [blame] | 2464 | kfree(set->tags); |
| 2465 | set->tags = NULL; |
Christoph Hellwig | da695ba | 2016-09-14 16:18:55 +0200 | [diff] [blame] | 2466 | return ret; |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 2467 | } |
| 2468 | EXPORT_SYMBOL(blk_mq_alloc_tag_set); |
| 2469 | |
| 2470 | void blk_mq_free_tag_set(struct blk_mq_tag_set *set) |
| 2471 | { |
| 2472 | int i; |
| 2473 | |
Jens Axboe | cc71a6f | 2017-01-11 14:29:56 -0700 | [diff] [blame] | 2474 | for (i = 0; i < nr_cpu_ids; i++) |
| 2475 | blk_mq_free_map_and_requests(set, i); |
Jens Axboe | 484b406 | 2014-05-21 14:01:15 -0600 | [diff] [blame] | 2476 | |
Christoph Hellwig | bdd17e7 | 2016-09-14 16:18:53 +0200 | [diff] [blame] | 2477 | kfree(set->mq_map); |
| 2478 | set->mq_map = NULL; |
| 2479 | |
Ming Lei | 981bd18 | 2014-04-24 00:07:34 +0800 | [diff] [blame] | 2480 | kfree(set->tags); |
Robert Elliott | 5676e7b | 2014-09-02 11:38:44 -0500 | [diff] [blame] | 2481 | set->tags = NULL; |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 2482 | } |
| 2483 | EXPORT_SYMBOL(blk_mq_free_tag_set); |
| 2484 | |
Jens Axboe | e3a2b3f | 2014-05-20 11:49:02 -0600 | [diff] [blame] | 2485 | int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr) |
| 2486 | { |
| 2487 | struct blk_mq_tag_set *set = q->tag_set; |
| 2488 | struct blk_mq_hw_ctx *hctx; |
| 2489 | int i, ret; |
| 2490 | |
| 2491 | if (!set || nr > set->queue_depth) |
| 2492 | return -EINVAL; |
| 2493 | |
| 2494 | ret = 0; |
| 2495 | queue_for_each_hw_ctx(q, hctx, i) { |
Keith Busch | e9137d4 | 2016-02-18 14:56:35 -0700 | [diff] [blame] | 2496 | if (!hctx->tags) |
| 2497 | continue; |
Jens Axboe | e3a2b3f | 2014-05-20 11:49:02 -0600 | [diff] [blame] | 2498 | ret = blk_mq_tag_update_depth(hctx->tags, nr); |
| 2499 | if (ret) |
| 2500 | break; |
| 2501 | } |
| 2502 | |
| 2503 | if (!ret) |
| 2504 | q->nr_requests = nr; |
| 2505 | |
| 2506 | return ret; |
| 2507 | } |
| 2508 | |
Keith Busch | 868f2f0 | 2015-12-17 17:08:14 -0700 | [diff] [blame] | 2509 | void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues) |
| 2510 | { |
| 2511 | struct request_queue *q; |
| 2512 | |
| 2513 | if (nr_hw_queues > nr_cpu_ids) |
| 2514 | nr_hw_queues = nr_cpu_ids; |
| 2515 | if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues) |
| 2516 | return; |
| 2517 | |
| 2518 | list_for_each_entry(q, &set->tag_list, tag_set_list) |
| 2519 | blk_mq_freeze_queue(q); |
| 2520 | |
| 2521 | set->nr_hw_queues = nr_hw_queues; |
| 2522 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
| 2523 | blk_mq_realloc_hw_ctxs(set, q); |
| 2524 | |
| 2525 | if (q->nr_hw_queues > 1) |
| 2526 | blk_queue_make_request(q, blk_mq_make_request); |
| 2527 | else |
| 2528 | blk_queue_make_request(q, blk_sq_make_request); |
| 2529 | |
| 2530 | blk_mq_queue_reinit(q, cpu_online_mask); |
| 2531 | } |
| 2532 | |
| 2533 | list_for_each_entry(q, &set->tag_list, tag_set_list) |
| 2534 | blk_mq_unfreeze_queue(q); |
| 2535 | } |
| 2536 | EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues); |
| 2537 | |
Jens Axboe | 64f1c21 | 2016-11-14 13:03:03 -0700 | [diff] [blame] | 2538 | static unsigned long blk_mq_poll_nsecs(struct request_queue *q, |
| 2539 | struct blk_mq_hw_ctx *hctx, |
| 2540 | struct request *rq) |
| 2541 | { |
| 2542 | struct blk_rq_stat stat[2]; |
| 2543 | unsigned long ret = 0; |
| 2544 | |
| 2545 | /* |
| 2546 | * If stats collection isn't on, don't sleep but turn it on for |
| 2547 | * future users |
| 2548 | */ |
| 2549 | if (!blk_stat_enable(q)) |
| 2550 | return 0; |
| 2551 | |
| 2552 | /* |
| 2553 | * We don't have to do this once per IO, should optimize this |
| 2554 | * to just use the current window of stats until it changes |
| 2555 | */ |
| 2556 | memset(&stat, 0, sizeof(stat)); |
| 2557 | blk_hctx_stat_get(hctx, stat); |
| 2558 | |
| 2559 | /* |
| 2560 | * As an optimistic guess, use half of the mean service time |
| 2561 | * for this type of request. We can (and should) make this smarter. |
| 2562 | * For instance, if the completion latencies are tight, we can |
| 2563 | * get closer than just half the mean. This is especially |
| 2564 | * important on devices where the completion latencies are longer |
| 2565 | * than ~10 usec. |
| 2566 | */ |
| 2567 | if (req_op(rq) == REQ_OP_READ && stat[BLK_STAT_READ].nr_samples) |
| 2568 | ret = (stat[BLK_STAT_READ].mean + 1) / 2; |
| 2569 | else if (req_op(rq) == REQ_OP_WRITE && stat[BLK_STAT_WRITE].nr_samples) |
| 2570 | ret = (stat[BLK_STAT_WRITE].mean + 1) / 2; |
| 2571 | |
| 2572 | return ret; |
| 2573 | } |
| 2574 | |
Jens Axboe | 06426ad | 2016-11-14 13:01:59 -0700 | [diff] [blame] | 2575 | static bool blk_mq_poll_hybrid_sleep(struct request_queue *q, |
Jens Axboe | 64f1c21 | 2016-11-14 13:03:03 -0700 | [diff] [blame] | 2576 | struct blk_mq_hw_ctx *hctx, |
Jens Axboe | 06426ad | 2016-11-14 13:01:59 -0700 | [diff] [blame] | 2577 | struct request *rq) |
| 2578 | { |
| 2579 | struct hrtimer_sleeper hs; |
| 2580 | enum hrtimer_mode mode; |
Jens Axboe | 64f1c21 | 2016-11-14 13:03:03 -0700 | [diff] [blame] | 2581 | unsigned int nsecs; |
Jens Axboe | 06426ad | 2016-11-14 13:01:59 -0700 | [diff] [blame] | 2582 | ktime_t kt; |
| 2583 | |
Jens Axboe | 64f1c21 | 2016-11-14 13:03:03 -0700 | [diff] [blame] | 2584 | if (test_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags)) |
| 2585 | return false; |
| 2586 | |
| 2587 | /* |
| 2588 | * poll_nsec can be: |
| 2589 | * |
| 2590 | * -1: don't ever hybrid sleep |
| 2591 | * 0: use half of prev avg |
| 2592 | * >0: use this specific value |
| 2593 | */ |
| 2594 | if (q->poll_nsec == -1) |
| 2595 | return false; |
| 2596 | else if (q->poll_nsec > 0) |
| 2597 | nsecs = q->poll_nsec; |
| 2598 | else |
| 2599 | nsecs = blk_mq_poll_nsecs(q, hctx, rq); |
| 2600 | |
| 2601 | if (!nsecs) |
Jens Axboe | 06426ad | 2016-11-14 13:01:59 -0700 | [diff] [blame] | 2602 | return false; |
| 2603 | |
| 2604 | set_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags); |
| 2605 | |
| 2606 | /* |
| 2607 | * This will be replaced with the stats tracking code, using |
| 2608 | * 'avg_completion_time / 2' as the pre-sleep target. |
| 2609 | */ |
Thomas Gleixner | 8b0e195 | 2016-12-25 12:30:41 +0100 | [diff] [blame] | 2610 | kt = nsecs; |
Jens Axboe | 06426ad | 2016-11-14 13:01:59 -0700 | [diff] [blame] | 2611 | |
| 2612 | mode = HRTIMER_MODE_REL; |
| 2613 | hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode); |
| 2614 | hrtimer_set_expires(&hs.timer, kt); |
| 2615 | |
| 2616 | hrtimer_init_sleeper(&hs, current); |
| 2617 | do { |
| 2618 | if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags)) |
| 2619 | break; |
| 2620 | set_current_state(TASK_UNINTERRUPTIBLE); |
| 2621 | hrtimer_start_expires(&hs.timer, mode); |
| 2622 | if (hs.task) |
| 2623 | io_schedule(); |
| 2624 | hrtimer_cancel(&hs.timer); |
| 2625 | mode = HRTIMER_MODE_ABS; |
| 2626 | } while (hs.task && !signal_pending(current)); |
| 2627 | |
| 2628 | __set_current_state(TASK_RUNNING); |
| 2629 | destroy_hrtimer_on_stack(&hs.timer); |
| 2630 | return true; |
| 2631 | } |
| 2632 | |
Jens Axboe | bbd7bb7 | 2016-11-04 09:34:34 -0600 | [diff] [blame] | 2633 | static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq) |
| 2634 | { |
| 2635 | struct request_queue *q = hctx->queue; |
| 2636 | long state; |
| 2637 | |
Jens Axboe | 06426ad | 2016-11-14 13:01:59 -0700 | [diff] [blame] | 2638 | /* |
| 2639 | * If we sleep, have the caller restart the poll loop to reset |
| 2640 | * the state. Like for the other success return cases, the |
| 2641 | * caller is responsible for checking if the IO completed. If |
| 2642 | * the IO isn't complete, we'll get called again and will go |
| 2643 | * straight to the busy poll loop. |
| 2644 | */ |
Jens Axboe | 64f1c21 | 2016-11-14 13:03:03 -0700 | [diff] [blame] | 2645 | if (blk_mq_poll_hybrid_sleep(q, hctx, rq)) |
Jens Axboe | 06426ad | 2016-11-14 13:01:59 -0700 | [diff] [blame] | 2646 | return true; |
| 2647 | |
Jens Axboe | bbd7bb7 | 2016-11-04 09:34:34 -0600 | [diff] [blame] | 2648 | hctx->poll_considered++; |
| 2649 | |
| 2650 | state = current->state; |
| 2651 | while (!need_resched()) { |
| 2652 | int ret; |
| 2653 | |
| 2654 | hctx->poll_invoked++; |
| 2655 | |
| 2656 | ret = q->mq_ops->poll(hctx, rq->tag); |
| 2657 | if (ret > 0) { |
| 2658 | hctx->poll_success++; |
| 2659 | set_current_state(TASK_RUNNING); |
| 2660 | return true; |
| 2661 | } |
| 2662 | |
| 2663 | if (signal_pending_state(state, current)) |
| 2664 | set_current_state(TASK_RUNNING); |
| 2665 | |
| 2666 | if (current->state == TASK_RUNNING) |
| 2667 | return true; |
| 2668 | if (ret < 0) |
| 2669 | break; |
| 2670 | cpu_relax(); |
| 2671 | } |
| 2672 | |
| 2673 | return false; |
| 2674 | } |
| 2675 | |
| 2676 | bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie) |
| 2677 | { |
| 2678 | struct blk_mq_hw_ctx *hctx; |
| 2679 | struct blk_plug *plug; |
| 2680 | struct request *rq; |
| 2681 | |
| 2682 | if (!q->mq_ops || !q->mq_ops->poll || !blk_qc_t_valid(cookie) || |
| 2683 | !test_bit(QUEUE_FLAG_POLL, &q->queue_flags)) |
| 2684 | return false; |
| 2685 | |
| 2686 | plug = current->plug; |
| 2687 | if (plug) |
| 2688 | blk_flush_plug_list(plug, false); |
| 2689 | |
| 2690 | hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)]; |
| 2691 | rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie)); |
| 2692 | |
| 2693 | return __blk_mq_poll(hctx, rq); |
| 2694 | } |
| 2695 | EXPORT_SYMBOL_GPL(blk_mq_poll); |
| 2696 | |
Jens Axboe | 676141e | 2014-03-20 13:29:18 -0600 | [diff] [blame] | 2697 | void blk_mq_disable_hotplug(void) |
| 2698 | { |
| 2699 | mutex_lock(&all_q_mutex); |
| 2700 | } |
| 2701 | |
| 2702 | void blk_mq_enable_hotplug(void) |
| 2703 | { |
| 2704 | mutex_unlock(&all_q_mutex); |
| 2705 | } |
| 2706 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2707 | static int __init blk_mq_init(void) |
| 2708 | { |
Thomas Gleixner | 9467f85 | 2016-09-22 08:05:17 -0600 | [diff] [blame] | 2709 | cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL, |
| 2710 | blk_mq_hctx_notify_dead); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2711 | |
Sebastian Andrzej Siewior | 65d5291 | 2016-09-22 08:05:19 -0600 | [diff] [blame] | 2712 | cpuhp_setup_state_nocalls(CPUHP_BLK_MQ_PREPARE, "block/mq:prepare", |
| 2713 | blk_mq_queue_reinit_prepare, |
| 2714 | blk_mq_queue_reinit_dead); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2715 | return 0; |
| 2716 | } |
| 2717 | subsys_initcall(blk_mq_init); |