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> |
Ingo Molnar | 105ab3d | 2017-02-01 16:36:40 +0100 | [diff] [blame] | 23 | #include <linux/sched/topology.h> |
Ingo Molnar | 174cd4b | 2017-02-02 19:15:33 +0100 | [diff] [blame] | 24 | #include <linux/sched/signal.h> |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 25 | #include <linux/delay.h> |
Jens Axboe | aedcd72 | 2014-09-17 08:27:03 -0600 | [diff] [blame] | 26 | #include <linux/crash_dump.h> |
Jens Axboe | 88c7b2b | 2016-08-25 08:07:30 -0600 | [diff] [blame] | 27 | #include <linux/prefetch.h> |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 28 | |
| 29 | #include <trace/events/block.h> |
| 30 | |
| 31 | #include <linux/blk-mq.h> |
| 32 | #include "blk.h" |
| 33 | #include "blk-mq.h" |
Omar Sandoval | 9c1051a | 2017-05-04 08:17:21 -0600 | [diff] [blame] | 34 | #include "blk-mq-debugfs.h" |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 35 | #include "blk-mq-tag.h" |
Jens Axboe | cf43e6b | 2016-11-07 21:32:37 -0700 | [diff] [blame] | 36 | #include "blk-stat.h" |
Jens Axboe | 87760e5 | 2016-11-09 12:38:14 -0700 | [diff] [blame] | 37 | #include "blk-wbt.h" |
Jens Axboe | bd166ef | 2017-01-17 06:03:22 -0700 | [diff] [blame] | 38 | #include "blk-mq-sched.h" |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 39 | |
| 40 | static DEFINE_MUTEX(all_q_mutex); |
| 41 | static LIST_HEAD(all_q_list); |
| 42 | |
Omar Sandoval | 34dbad5 | 2017-03-21 08:56:08 -0700 | [diff] [blame] | 43 | static void blk_mq_poll_stats_start(struct request_queue *q); |
| 44 | static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb); |
Jens Axboe | 2719aa2 | 2017-05-03 11:08:14 -0600 | [diff] [blame] | 45 | static void __blk_mq_stop_hw_queues(struct request_queue *q, bool sync); |
Omar Sandoval | 34dbad5 | 2017-03-21 08:56:08 -0700 | [diff] [blame] | 46 | |
Stephen Bates | 720b8cc | 2017-04-07 06:24:03 -0600 | [diff] [blame] | 47 | static int blk_mq_poll_stats_bkt(const struct request *rq) |
| 48 | { |
| 49 | int ddir, bytes, bucket; |
| 50 | |
Jens Axboe | 99c749a | 2017-04-21 07:55:42 -0600 | [diff] [blame] | 51 | ddir = rq_data_dir(rq); |
Stephen Bates | 720b8cc | 2017-04-07 06:24:03 -0600 | [diff] [blame] | 52 | bytes = blk_rq_bytes(rq); |
| 53 | |
| 54 | bucket = ddir + 2*(ilog2(bytes) - 9); |
| 55 | |
| 56 | if (bucket < 0) |
| 57 | return -1; |
| 58 | else if (bucket >= BLK_MQ_POLL_STATS_BKTS) |
| 59 | return ddir + BLK_MQ_POLL_STATS_BKTS - 2; |
| 60 | |
| 61 | return bucket; |
| 62 | } |
| 63 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 64 | /* |
| 65 | * Check if any of the ctx's have pending work in this hardware queue |
| 66 | */ |
Jens Axboe | 50e1dab | 2017-01-26 14:42:34 -0700 | [diff] [blame] | 67 | bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 68 | { |
Jens Axboe | bd166ef | 2017-01-17 06:03:22 -0700 | [diff] [blame] | 69 | return sbitmap_any_bit_set(&hctx->ctx_map) || |
| 70 | !list_empty_careful(&hctx->dispatch) || |
| 71 | blk_mq_sched_has_work(hctx); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 72 | } |
| 73 | |
| 74 | /* |
| 75 | * Mark this ctx as having pending work in this hardware queue |
| 76 | */ |
| 77 | static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx, |
| 78 | struct blk_mq_ctx *ctx) |
| 79 | { |
Omar Sandoval | 8845964 | 2016-09-17 08:38:44 -0600 | [diff] [blame] | 80 | if (!sbitmap_test_bit(&hctx->ctx_map, ctx->index_hw)) |
| 81 | sbitmap_set_bit(&hctx->ctx_map, ctx->index_hw); |
Jens Axboe | 1429d7c | 2014-05-19 09:23:55 -0600 | [diff] [blame] | 82 | } |
| 83 | |
| 84 | static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx, |
| 85 | struct blk_mq_ctx *ctx) |
| 86 | { |
Omar Sandoval | 8845964 | 2016-09-17 08:38:44 -0600 | [diff] [blame] | 87 | sbitmap_clear_bit(&hctx->ctx_map, ctx->index_hw); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 88 | } |
| 89 | |
Ming Lei | 1671d52 | 2017-03-27 20:06:57 +0800 | [diff] [blame] | 90 | void blk_freeze_queue_start(struct request_queue *q) |
Ming Lei | 43a5e4e | 2013-12-26 21:31:35 +0800 | [diff] [blame] | 91 | { |
Christoph Hellwig | 4ecd4fe | 2015-05-07 09:38:13 +0200 | [diff] [blame] | 92 | int freeze_depth; |
Tejun Heo | cddd5d1 | 2014-08-16 08:02:24 -0400 | [diff] [blame] | 93 | |
Christoph Hellwig | 4ecd4fe | 2015-05-07 09:38:13 +0200 | [diff] [blame] | 94 | freeze_depth = atomic_inc_return(&q->mq_freeze_depth); |
| 95 | if (freeze_depth == 1) { |
Dan Williams | 3ef28e8 | 2015-10-21 13:20:12 -0400 | [diff] [blame] | 96 | percpu_ref_kill(&q->q_usage_counter); |
Mike Snitzer | b94ec29 | 2015-03-11 23:56:38 -0400 | [diff] [blame] | 97 | blk_mq_run_hw_queues(q, false); |
Tejun Heo | cddd5d1 | 2014-08-16 08:02:24 -0400 | [diff] [blame] | 98 | } |
Tejun Heo | f3af020 | 2014-11-04 13:52:27 -0500 | [diff] [blame] | 99 | } |
Ming Lei | 1671d52 | 2017-03-27 20:06:57 +0800 | [diff] [blame] | 100 | EXPORT_SYMBOL_GPL(blk_freeze_queue_start); |
Tejun Heo | f3af020 | 2014-11-04 13:52:27 -0500 | [diff] [blame] | 101 | |
Keith Busch | 6bae363 | 2017-03-01 14:22:10 -0500 | [diff] [blame] | 102 | void blk_mq_freeze_queue_wait(struct request_queue *q) |
Tejun Heo | f3af020 | 2014-11-04 13:52:27 -0500 | [diff] [blame] | 103 | { |
Dan Williams | 3ef28e8 | 2015-10-21 13:20:12 -0400 | [diff] [blame] | 104 | 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] | 105 | } |
Keith Busch | 6bae363 | 2017-03-01 14:22:10 -0500 | [diff] [blame] | 106 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait); |
Ming Lei | 43a5e4e | 2013-12-26 21:31:35 +0800 | [diff] [blame] | 107 | |
Keith Busch | f91328c | 2017-03-01 14:22:11 -0500 | [diff] [blame] | 108 | int blk_mq_freeze_queue_wait_timeout(struct request_queue *q, |
| 109 | unsigned long timeout) |
| 110 | { |
| 111 | return wait_event_timeout(q->mq_freeze_wq, |
| 112 | percpu_ref_is_zero(&q->q_usage_counter), |
| 113 | timeout); |
| 114 | } |
| 115 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 116 | |
Tejun Heo | f3af020 | 2014-11-04 13:52:27 -0500 | [diff] [blame] | 117 | /* |
| 118 | * Guarantee no request is in use, so we can change any data structure of |
| 119 | * the queue afterward. |
| 120 | */ |
Dan Williams | 3ef28e8 | 2015-10-21 13:20:12 -0400 | [diff] [blame] | 121 | void blk_freeze_queue(struct request_queue *q) |
Tejun Heo | f3af020 | 2014-11-04 13:52:27 -0500 | [diff] [blame] | 122 | { |
Dan Williams | 3ef28e8 | 2015-10-21 13:20:12 -0400 | [diff] [blame] | 123 | /* |
| 124 | * In the !blk_mq case we are only calling this to kill the |
| 125 | * q_usage_counter, otherwise this increases the freeze depth |
| 126 | * and waits for it to return to zero. For this reason there is |
| 127 | * no blk_unfreeze_queue(), and blk_freeze_queue() is not |
| 128 | * exported to drivers as the only user for unfreeze is blk_mq. |
| 129 | */ |
Ming Lei | 1671d52 | 2017-03-27 20:06:57 +0800 | [diff] [blame] | 130 | blk_freeze_queue_start(q); |
Tejun Heo | f3af020 | 2014-11-04 13:52:27 -0500 | [diff] [blame] | 131 | blk_mq_freeze_queue_wait(q); |
| 132 | } |
Dan Williams | 3ef28e8 | 2015-10-21 13:20:12 -0400 | [diff] [blame] | 133 | |
| 134 | void blk_mq_freeze_queue(struct request_queue *q) |
| 135 | { |
| 136 | /* |
| 137 | * ...just an alias to keep freeze and unfreeze actions balanced |
| 138 | * in the blk_mq_* namespace |
| 139 | */ |
| 140 | blk_freeze_queue(q); |
| 141 | } |
Jens Axboe | c761d96 | 2015-01-02 15:05:12 -0700 | [diff] [blame] | 142 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue); |
Tejun Heo | f3af020 | 2014-11-04 13:52:27 -0500 | [diff] [blame] | 143 | |
Keith Busch | b4c6a02 | 2014-12-19 17:54:14 -0700 | [diff] [blame] | 144 | void blk_mq_unfreeze_queue(struct request_queue *q) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 145 | { |
Christoph Hellwig | 4ecd4fe | 2015-05-07 09:38:13 +0200 | [diff] [blame] | 146 | int freeze_depth; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 147 | |
Christoph Hellwig | 4ecd4fe | 2015-05-07 09:38:13 +0200 | [diff] [blame] | 148 | freeze_depth = atomic_dec_return(&q->mq_freeze_depth); |
| 149 | WARN_ON_ONCE(freeze_depth < 0); |
| 150 | if (!freeze_depth) { |
Dan Williams | 3ef28e8 | 2015-10-21 13:20:12 -0400 | [diff] [blame] | 151 | percpu_ref_reinit(&q->q_usage_counter); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 152 | wake_up_all(&q->mq_freeze_wq); |
Tejun Heo | add703f | 2014-07-01 10:34:38 -0600 | [diff] [blame] | 153 | } |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 154 | } |
Keith Busch | b4c6a02 | 2014-12-19 17:54:14 -0700 | [diff] [blame] | 155 | EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 156 | |
Bart Van Assche | 6a83e74 | 2016-11-02 10:09:51 -0600 | [diff] [blame] | 157 | /** |
| 158 | * blk_mq_quiesce_queue() - wait until all ongoing queue_rq calls have finished |
| 159 | * @q: request queue. |
| 160 | * |
| 161 | * Note: this function does not prevent that the struct request end_io() |
| 162 | * callback function is invoked. Additionally, it is not prevented that |
| 163 | * new queue_rq() calls occur unless the queue has been stopped first. |
| 164 | */ |
| 165 | void blk_mq_quiesce_queue(struct request_queue *q) |
| 166 | { |
| 167 | struct blk_mq_hw_ctx *hctx; |
| 168 | unsigned int i; |
| 169 | bool rcu = false; |
| 170 | |
Jens Axboe | 2719aa2 | 2017-05-03 11:08:14 -0600 | [diff] [blame] | 171 | __blk_mq_stop_hw_queues(q, true); |
Bart Van Assche | 6a83e74 | 2016-11-02 10:09:51 -0600 | [diff] [blame] | 172 | |
| 173 | queue_for_each_hw_ctx(q, hctx, i) { |
| 174 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
| 175 | synchronize_srcu(&hctx->queue_rq_srcu); |
| 176 | else |
| 177 | rcu = true; |
| 178 | } |
| 179 | if (rcu) |
| 180 | synchronize_rcu(); |
| 181 | } |
| 182 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue); |
| 183 | |
Jens Axboe | aed3ea9 | 2014-12-22 14:04:42 -0700 | [diff] [blame] | 184 | void blk_mq_wake_waiters(struct request_queue *q) |
| 185 | { |
| 186 | struct blk_mq_hw_ctx *hctx; |
| 187 | unsigned int i; |
| 188 | |
| 189 | queue_for_each_hw_ctx(q, hctx, i) |
| 190 | if (blk_mq_hw_queue_mapped(hctx)) |
| 191 | blk_mq_tag_wakeup_all(hctx->tags, true); |
Keith Busch | 3fd5940 | 2015-01-08 08:53:56 -0700 | [diff] [blame] | 192 | |
| 193 | /* |
| 194 | * If we are called because the queue has now been marked as |
| 195 | * dying, we need to ensure that processes currently waiting on |
| 196 | * the queue are notified as well. |
| 197 | */ |
| 198 | wake_up_all(&q->mq_freeze_wq); |
Jens Axboe | aed3ea9 | 2014-12-22 14:04:42 -0700 | [diff] [blame] | 199 | } |
| 200 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 201 | bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx) |
| 202 | { |
| 203 | return blk_mq_has_free_tags(hctx->tags); |
| 204 | } |
| 205 | EXPORT_SYMBOL(blk_mq_can_queue); |
| 206 | |
Jens Axboe | 2c3ad66 | 2016-12-14 14:34:47 -0700 | [diff] [blame] | 207 | void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx, |
| 208 | struct request *rq, unsigned int op) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 209 | { |
Christoph Hellwig | af76e55 | 2014-05-06 12:12:45 +0200 | [diff] [blame] | 210 | INIT_LIST_HEAD(&rq->queuelist); |
| 211 | /* csd/requeue_work/fifo_time is initialized before use */ |
| 212 | rq->q = q; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 213 | rq->mq_ctx = ctx; |
Christoph Hellwig | ef295ec | 2016-10-28 08:48:16 -0600 | [diff] [blame] | 214 | rq->cmd_flags = op; |
Christoph Hellwig | e806402 | 2016-10-20 15:12:13 +0200 | [diff] [blame] | 215 | if (blk_queue_io_stat(q)) |
| 216 | rq->rq_flags |= RQF_IO_STAT; |
Christoph Hellwig | af76e55 | 2014-05-06 12:12:45 +0200 | [diff] [blame] | 217 | /* do not touch atomic flags, it needs atomic ops against the timer */ |
| 218 | rq->cpu = -1; |
Christoph Hellwig | af76e55 | 2014-05-06 12:12:45 +0200 | [diff] [blame] | 219 | INIT_HLIST_NODE(&rq->hash); |
| 220 | RB_CLEAR_NODE(&rq->rb_node); |
Christoph Hellwig | af76e55 | 2014-05-06 12:12:45 +0200 | [diff] [blame] | 221 | rq->rq_disk = NULL; |
| 222 | rq->part = NULL; |
Jens Axboe | 3ee3237 | 2014-06-09 09:36:53 -0600 | [diff] [blame] | 223 | rq->start_time = jiffies; |
Christoph Hellwig | af76e55 | 2014-05-06 12:12:45 +0200 | [diff] [blame] | 224 | #ifdef CONFIG_BLK_CGROUP |
| 225 | rq->rl = NULL; |
Ming Lei | 0fec08b | 2014-01-03 10:00:08 -0700 | [diff] [blame] | 226 | set_start_time_ns(rq); |
Christoph Hellwig | af76e55 | 2014-05-06 12:12:45 +0200 | [diff] [blame] | 227 | rq->io_start_time_ns = 0; |
| 228 | #endif |
| 229 | rq->nr_phys_segments = 0; |
| 230 | #if defined(CONFIG_BLK_DEV_INTEGRITY) |
| 231 | rq->nr_integrity_segments = 0; |
| 232 | #endif |
Christoph Hellwig | af76e55 | 2014-05-06 12:12:45 +0200 | [diff] [blame] | 233 | rq->special = NULL; |
| 234 | /* tag was already set */ |
Christoph Hellwig | af76e55 | 2014-05-06 12:12:45 +0200 | [diff] [blame] | 235 | rq->extra_len = 0; |
Christoph Hellwig | af76e55 | 2014-05-06 12:12:45 +0200 | [diff] [blame] | 236 | |
Christoph Hellwig | af76e55 | 2014-05-06 12:12:45 +0200 | [diff] [blame] | 237 | INIT_LIST_HEAD(&rq->timeout_list); |
Jens Axboe | f6be4fb | 2014-06-06 11:03:48 -0600 | [diff] [blame] | 238 | rq->timeout = 0; |
| 239 | |
Christoph Hellwig | af76e55 | 2014-05-06 12:12:45 +0200 | [diff] [blame] | 240 | rq->end_io = NULL; |
| 241 | rq->end_io_data = NULL; |
| 242 | rq->next_rq = NULL; |
| 243 | |
Christoph Hellwig | ef295ec | 2016-10-28 08:48:16 -0600 | [diff] [blame] | 244 | ctx->rq_dispatched[op_is_sync(op)]++; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 245 | } |
Jens Axboe | 2c3ad66 | 2016-12-14 14:34:47 -0700 | [diff] [blame] | 246 | EXPORT_SYMBOL_GPL(blk_mq_rq_ctx_init); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 247 | |
Jens Axboe | 2c3ad66 | 2016-12-14 14:34:47 -0700 | [diff] [blame] | 248 | struct request *__blk_mq_alloc_request(struct blk_mq_alloc_data *data, |
| 249 | unsigned int op) |
Christoph Hellwig | 5dee857 | 2014-05-27 20:59:47 +0200 | [diff] [blame] | 250 | { |
| 251 | struct request *rq; |
| 252 | unsigned int tag; |
| 253 | |
Ming Lei | cb96a42 | 2014-06-01 00:43:37 +0800 | [diff] [blame] | 254 | tag = blk_mq_get_tag(data); |
Christoph Hellwig | 5dee857 | 2014-05-27 20:59:47 +0200 | [diff] [blame] | 255 | if (tag != BLK_MQ_TAG_FAIL) { |
Jens Axboe | bd166ef | 2017-01-17 06:03:22 -0700 | [diff] [blame] | 256 | struct blk_mq_tags *tags = blk_mq_tags_from_data(data); |
| 257 | |
| 258 | rq = tags->static_rqs[tag]; |
Christoph Hellwig | 5dee857 | 2014-05-27 20:59:47 +0200 | [diff] [blame] | 259 | |
Jens Axboe | bd166ef | 2017-01-17 06:03:22 -0700 | [diff] [blame] | 260 | if (data->flags & BLK_MQ_REQ_INTERNAL) { |
| 261 | rq->tag = -1; |
| 262 | rq->internal_tag = tag; |
| 263 | } else { |
Jens Axboe | 200e86b | 2017-01-25 08:11:38 -0700 | [diff] [blame] | 264 | if (blk_mq_tag_busy(data->hctx)) { |
| 265 | rq->rq_flags = RQF_MQ_INFLIGHT; |
| 266 | atomic_inc(&data->hctx->nr_active); |
| 267 | } |
Jens Axboe | bd166ef | 2017-01-17 06:03:22 -0700 | [diff] [blame] | 268 | rq->tag = tag; |
| 269 | rq->internal_tag = -1; |
Omar Sandoval | 562bef4 | 2017-02-27 09:47:55 -0800 | [diff] [blame] | 270 | data->hctx->tags->rqs[rq->tag] = rq; |
Jens Axboe | bd166ef | 2017-01-17 06:03:22 -0700 | [diff] [blame] | 271 | } |
| 272 | |
Christoph Hellwig | ef295ec | 2016-10-28 08:48:16 -0600 | [diff] [blame] | 273 | blk_mq_rq_ctx_init(data->q, data->ctx, rq, op); |
Christoph Hellwig | 5dee857 | 2014-05-27 20:59:47 +0200 | [diff] [blame] | 274 | return rq; |
| 275 | } |
| 276 | |
| 277 | return NULL; |
| 278 | } |
Jens Axboe | 2c3ad66 | 2016-12-14 14:34:47 -0700 | [diff] [blame] | 279 | EXPORT_SYMBOL_GPL(__blk_mq_alloc_request); |
Christoph Hellwig | 5dee857 | 2014-05-27 20:59:47 +0200 | [diff] [blame] | 280 | |
Christoph Hellwig | 6f3b0e8 | 2015-11-26 09:13:05 +0100 | [diff] [blame] | 281 | struct request *blk_mq_alloc_request(struct request_queue *q, int rw, |
| 282 | unsigned int flags) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 283 | { |
Jens Axboe | 5a797e0 | 2017-01-26 12:22:11 -0700 | [diff] [blame] | 284 | struct blk_mq_alloc_data alloc_data = { .flags = flags }; |
Jens Axboe | bd166ef | 2017-01-17 06:03:22 -0700 | [diff] [blame] | 285 | struct request *rq; |
Joe Lawrence | a492f07 | 2014-08-28 08:15:21 -0600 | [diff] [blame] | 286 | int ret; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 287 | |
Christoph Hellwig | 6f3b0e8 | 2015-11-26 09:13:05 +0100 | [diff] [blame] | 288 | ret = blk_queue_enter(q, flags & BLK_MQ_REQ_NOWAIT); |
Joe Lawrence | a492f07 | 2014-08-28 08:15:21 -0600 | [diff] [blame] | 289 | if (ret) |
| 290 | return ERR_PTR(ret); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 291 | |
Jens Axboe | bd166ef | 2017-01-17 06:03:22 -0700 | [diff] [blame] | 292 | rq = blk_mq_sched_get_request(q, NULL, rw, &alloc_data); |
Jens Axboe | 841bac2 | 2016-09-21 10:08:43 -0600 | [diff] [blame] | 293 | |
Jens Axboe | bd166ef | 2017-01-17 06:03:22 -0700 | [diff] [blame] | 294 | blk_mq_put_ctx(alloc_data.ctx); |
| 295 | blk_queue_exit(q); |
| 296 | |
| 297 | if (!rq) |
Joe Lawrence | a492f07 | 2014-08-28 08:15:21 -0600 | [diff] [blame] | 298 | return ERR_PTR(-EWOULDBLOCK); |
Christoph Hellwig | 0c4de0f | 2016-07-19 11:31:50 +0200 | [diff] [blame] | 299 | |
| 300 | rq->__data_len = 0; |
| 301 | rq->__sector = (sector_t) -1; |
| 302 | rq->bio = rq->biotail = NULL; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 303 | return rq; |
| 304 | } |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 305 | EXPORT_SYMBOL(blk_mq_alloc_request); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 306 | |
Ming Lin | 1f5bd33 | 2016-06-13 16:45:21 +0200 | [diff] [blame] | 307 | struct request *blk_mq_alloc_request_hctx(struct request_queue *q, int rw, |
| 308 | unsigned int flags, unsigned int hctx_idx) |
| 309 | { |
Omar Sandoval | 6d2809d | 2017-02-27 10:28:27 -0800 | [diff] [blame] | 310 | struct blk_mq_alloc_data alloc_data = { .flags = flags }; |
Ming Lin | 1f5bd33 | 2016-06-13 16:45:21 +0200 | [diff] [blame] | 311 | struct request *rq; |
Omar Sandoval | 6d2809d | 2017-02-27 10:28:27 -0800 | [diff] [blame] | 312 | unsigned int cpu; |
Ming Lin | 1f5bd33 | 2016-06-13 16:45:21 +0200 | [diff] [blame] | 313 | int ret; |
| 314 | |
| 315 | /* |
| 316 | * If the tag allocator sleeps we could get an allocation for a |
| 317 | * different hardware context. No need to complicate the low level |
| 318 | * allocator for this for the rare use case of a command tied to |
| 319 | * a specific queue. |
| 320 | */ |
| 321 | if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT))) |
| 322 | return ERR_PTR(-EINVAL); |
| 323 | |
| 324 | if (hctx_idx >= q->nr_hw_queues) |
| 325 | return ERR_PTR(-EIO); |
| 326 | |
| 327 | ret = blk_queue_enter(q, true); |
| 328 | if (ret) |
| 329 | return ERR_PTR(ret); |
| 330 | |
Christoph Hellwig | c8712c6 | 2016-09-23 10:25:48 -0600 | [diff] [blame] | 331 | /* |
| 332 | * Check if the hardware context is actually mapped to anything. |
| 333 | * If not tell the caller that it should skip this queue. |
| 334 | */ |
Omar Sandoval | 6d2809d | 2017-02-27 10:28:27 -0800 | [diff] [blame] | 335 | alloc_data.hctx = q->queue_hw_ctx[hctx_idx]; |
| 336 | if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) { |
| 337 | blk_queue_exit(q); |
| 338 | return ERR_PTR(-EXDEV); |
Christoph Hellwig | c8712c6 | 2016-09-23 10:25:48 -0600 | [diff] [blame] | 339 | } |
Omar Sandoval | 6d2809d | 2017-02-27 10:28:27 -0800 | [diff] [blame] | 340 | cpu = cpumask_first(alloc_data.hctx->cpumask); |
| 341 | alloc_data.ctx = __blk_mq_get_ctx(q, cpu); |
Ming Lin | 1f5bd33 | 2016-06-13 16:45:21 +0200 | [diff] [blame] | 342 | |
Omar Sandoval | 6d2809d | 2017-02-27 10:28:27 -0800 | [diff] [blame] | 343 | rq = blk_mq_sched_get_request(q, NULL, rw, &alloc_data); |
| 344 | |
Omar Sandoval | 6d2809d | 2017-02-27 10:28:27 -0800 | [diff] [blame] | 345 | blk_queue_exit(q); |
| 346 | |
| 347 | if (!rq) |
| 348 | return ERR_PTR(-EWOULDBLOCK); |
Ming Lin | 1f5bd33 | 2016-06-13 16:45:21 +0200 | [diff] [blame] | 349 | |
| 350 | return rq; |
| 351 | } |
| 352 | EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx); |
| 353 | |
Jens Axboe | bd166ef | 2017-01-17 06:03:22 -0700 | [diff] [blame] | 354 | void __blk_mq_finish_request(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx, |
| 355 | struct request *rq) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 356 | { |
Jens Axboe | bd166ef | 2017-01-17 06:03:22 -0700 | [diff] [blame] | 357 | const int sched_tag = rq->internal_tag; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 358 | struct request_queue *q = rq->q; |
| 359 | |
Christoph Hellwig | e806402 | 2016-10-20 15:12:13 +0200 | [diff] [blame] | 360 | if (rq->rq_flags & RQF_MQ_INFLIGHT) |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 361 | atomic_dec(&hctx->nr_active); |
Jens Axboe | 87760e5 | 2016-11-09 12:38:14 -0700 | [diff] [blame] | 362 | |
| 363 | wbt_done(q->rq_wb, &rq->issue_stat); |
Christoph Hellwig | e806402 | 2016-10-20 15:12:13 +0200 | [diff] [blame] | 364 | rq->rq_flags = 0; |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 365 | |
Christoph Hellwig | af76e55 | 2014-05-06 12:12:45 +0200 | [diff] [blame] | 366 | clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags); |
Jens Axboe | 06426ad | 2016-11-14 13:01:59 -0700 | [diff] [blame] | 367 | clear_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags); |
Jens Axboe | bd166ef | 2017-01-17 06:03:22 -0700 | [diff] [blame] | 368 | if (rq->tag != -1) |
| 369 | blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag); |
| 370 | if (sched_tag != -1) |
Omar Sandoval | c05f852 | 2017-04-14 01:00:01 -0700 | [diff] [blame] | 371 | blk_mq_put_tag(hctx, hctx->sched_tags, ctx, sched_tag); |
Bart Van Assche | 6d8c6c0 | 2017-04-07 12:40:09 -0600 | [diff] [blame] | 372 | blk_mq_sched_restart(hctx); |
Dan Williams | 3ef28e8 | 2015-10-21 13:20:12 -0400 | [diff] [blame] | 373 | blk_queue_exit(q); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 374 | } |
| 375 | |
Jens Axboe | bd166ef | 2017-01-17 06:03:22 -0700 | [diff] [blame] | 376 | static void blk_mq_finish_hctx_request(struct blk_mq_hw_ctx *hctx, |
Jens Axboe | 16a3c2a | 2016-12-15 14:27:46 -0700 | [diff] [blame] | 377 | struct request *rq) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 378 | { |
| 379 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
Jens Axboe | 7c7f2f2 | 2014-11-17 10:41:57 -0700 | [diff] [blame] | 380 | |
| 381 | ctx->rq_completed[rq_is_sync(rq)]++; |
Jens Axboe | bd166ef | 2017-01-17 06:03:22 -0700 | [diff] [blame] | 382 | __blk_mq_finish_request(hctx, ctx, rq); |
| 383 | } |
| 384 | |
| 385 | void blk_mq_finish_request(struct request *rq) |
| 386 | { |
| 387 | blk_mq_finish_hctx_request(blk_mq_map_queue(rq->q, rq->mq_ctx->cpu), rq); |
Jens Axboe | 7c7f2f2 | 2014-11-17 10:41:57 -0700 | [diff] [blame] | 388 | } |
Omar Sandoval | 5b72727 | 2017-04-14 01:00:00 -0700 | [diff] [blame] | 389 | EXPORT_SYMBOL_GPL(blk_mq_finish_request); |
Jens Axboe | 7c7f2f2 | 2014-11-17 10:41:57 -0700 | [diff] [blame] | 390 | |
| 391 | void blk_mq_free_request(struct request *rq) |
| 392 | { |
Jens Axboe | bd166ef | 2017-01-17 06:03:22 -0700 | [diff] [blame] | 393 | blk_mq_sched_put_request(rq); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 394 | } |
Jens Axboe | 1a3b595 | 2014-11-17 10:40:48 -0700 | [diff] [blame] | 395 | EXPORT_SYMBOL_GPL(blk_mq_free_request); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 396 | |
Christoph Hellwig | c8a446a | 2014-09-13 16:40:10 -0700 | [diff] [blame] | 397 | inline void __blk_mq_end_request(struct request *rq, int error) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 398 | { |
Ming Lei | 0d11e6a | 2013-12-05 10:50:39 -0700 | [diff] [blame] | 399 | blk_account_io_done(rq); |
| 400 | |
Christoph Hellwig | 91b6363 | 2014-04-16 09:44:53 +0200 | [diff] [blame] | 401 | if (rq->end_io) { |
Jens Axboe | 87760e5 | 2016-11-09 12:38:14 -0700 | [diff] [blame] | 402 | wbt_done(rq->q->rq_wb, &rq->issue_stat); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 403 | rq->end_io(rq, error); |
Christoph Hellwig | 91b6363 | 2014-04-16 09:44:53 +0200 | [diff] [blame] | 404 | } else { |
| 405 | if (unlikely(blk_bidi_rq(rq))) |
| 406 | blk_mq_free_request(rq->next_rq); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 407 | blk_mq_free_request(rq); |
Christoph Hellwig | 91b6363 | 2014-04-16 09:44:53 +0200 | [diff] [blame] | 408 | } |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 409 | } |
Christoph Hellwig | c8a446a | 2014-09-13 16:40:10 -0700 | [diff] [blame] | 410 | EXPORT_SYMBOL(__blk_mq_end_request); |
Christoph Hellwig | 63151a4 | 2014-04-16 09:44:52 +0200 | [diff] [blame] | 411 | |
Christoph Hellwig | c8a446a | 2014-09-13 16:40:10 -0700 | [diff] [blame] | 412 | void blk_mq_end_request(struct request *rq, int error) |
Christoph Hellwig | 63151a4 | 2014-04-16 09:44:52 +0200 | [diff] [blame] | 413 | { |
| 414 | if (blk_update_request(rq, error, blk_rq_bytes(rq))) |
| 415 | BUG(); |
Christoph Hellwig | c8a446a | 2014-09-13 16:40:10 -0700 | [diff] [blame] | 416 | __blk_mq_end_request(rq, error); |
Christoph Hellwig | 63151a4 | 2014-04-16 09:44:52 +0200 | [diff] [blame] | 417 | } |
Christoph Hellwig | c8a446a | 2014-09-13 16:40:10 -0700 | [diff] [blame] | 418 | EXPORT_SYMBOL(blk_mq_end_request); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 419 | |
Christoph Hellwig | 30a91cb | 2014-02-10 03:24:38 -0800 | [diff] [blame] | 420 | static void __blk_mq_complete_request_remote(void *data) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 421 | { |
Christoph Hellwig | 3d6efbf | 2014-01-08 09:33:37 -0800 | [diff] [blame] | 422 | struct request *rq = data; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 423 | |
Christoph Hellwig | 30a91cb | 2014-02-10 03:24:38 -0800 | [diff] [blame] | 424 | rq->q->softirq_done_fn(rq); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 425 | } |
| 426 | |
Christoph Hellwig | 453f834 | 2017-04-20 16:03:10 +0200 | [diff] [blame] | 427 | static void __blk_mq_complete_request(struct request *rq) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 428 | { |
| 429 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
Christoph Hellwig | 3853520 | 2014-04-25 02:32:53 -0700 | [diff] [blame] | 430 | bool shared = false; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 431 | int cpu; |
| 432 | |
Christoph Hellwig | 453f834 | 2017-04-20 16:03:10 +0200 | [diff] [blame] | 433 | if (rq->internal_tag != -1) |
| 434 | blk_mq_sched_completed_request(rq); |
| 435 | if (rq->rq_flags & RQF_STATS) { |
| 436 | blk_mq_poll_stats_start(rq->q); |
| 437 | blk_stat_add(rq); |
| 438 | } |
| 439 | |
Christoph Hellwig | 3853520 | 2014-04-25 02:32:53 -0700 | [diff] [blame] | 440 | if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) { |
Christoph Hellwig | 30a91cb | 2014-02-10 03:24:38 -0800 | [diff] [blame] | 441 | rq->q->softirq_done_fn(rq); |
| 442 | return; |
| 443 | } |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 444 | |
| 445 | cpu = get_cpu(); |
Christoph Hellwig | 3853520 | 2014-04-25 02:32:53 -0700 | [diff] [blame] | 446 | if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags)) |
| 447 | shared = cpus_share_cache(cpu, ctx->cpu); |
| 448 | |
| 449 | if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) { |
Christoph Hellwig | 30a91cb | 2014-02-10 03:24:38 -0800 | [diff] [blame] | 450 | rq->csd.func = __blk_mq_complete_request_remote; |
Christoph Hellwig | 3d6efbf | 2014-01-08 09:33:37 -0800 | [diff] [blame] | 451 | rq->csd.info = rq; |
| 452 | rq->csd.flags = 0; |
Frederic Weisbecker | c46fff2 | 2014-02-24 16:40:02 +0100 | [diff] [blame] | 453 | smp_call_function_single_async(ctx->cpu, &rq->csd); |
Christoph Hellwig | 3d6efbf | 2014-01-08 09:33:37 -0800 | [diff] [blame] | 454 | } else { |
Christoph Hellwig | 30a91cb | 2014-02-10 03:24:38 -0800 | [diff] [blame] | 455 | rq->q->softirq_done_fn(rq); |
Christoph Hellwig | 3d6efbf | 2014-01-08 09:33:37 -0800 | [diff] [blame] | 456 | } |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 457 | put_cpu(); |
| 458 | } |
Christoph Hellwig | 30a91cb | 2014-02-10 03:24:38 -0800 | [diff] [blame] | 459 | |
| 460 | /** |
| 461 | * blk_mq_complete_request - end I/O on a request |
| 462 | * @rq: the request being processed |
| 463 | * |
| 464 | * Description: |
| 465 | * Ends all I/O on a request. It does not handle partial completions. |
| 466 | * The actual completion happens out-of-order, through a IPI handler. |
| 467 | **/ |
Christoph Hellwig | 08e0029 | 2017-04-20 16:03:09 +0200 | [diff] [blame] | 468 | void blk_mq_complete_request(struct request *rq) |
Christoph Hellwig | 30a91cb | 2014-02-10 03:24:38 -0800 | [diff] [blame] | 469 | { |
Jens Axboe | 95f0968 | 2014-05-27 17:46:48 -0600 | [diff] [blame] | 470 | struct request_queue *q = rq->q; |
| 471 | |
| 472 | if (unlikely(blk_should_fake_timeout(q))) |
Christoph Hellwig | 30a91cb | 2014-02-10 03:24:38 -0800 | [diff] [blame] | 473 | return; |
Christoph Hellwig | 08e0029 | 2017-04-20 16:03:09 +0200 | [diff] [blame] | 474 | if (!blk_mark_rq_complete(rq)) |
Jens Axboe | ed85186 | 2014-05-30 21:20:50 -0600 | [diff] [blame] | 475 | __blk_mq_complete_request(rq); |
Christoph Hellwig | 30a91cb | 2014-02-10 03:24:38 -0800 | [diff] [blame] | 476 | } |
| 477 | EXPORT_SYMBOL(blk_mq_complete_request); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 478 | |
Keith Busch | 973c019 | 2015-01-07 18:55:43 -0700 | [diff] [blame] | 479 | int blk_mq_request_started(struct request *rq) |
| 480 | { |
| 481 | return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags); |
| 482 | } |
| 483 | EXPORT_SYMBOL_GPL(blk_mq_request_started); |
| 484 | |
Christoph Hellwig | e249007 | 2014-09-13 16:40:09 -0700 | [diff] [blame] | 485 | void blk_mq_start_request(struct request *rq) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 486 | { |
| 487 | struct request_queue *q = rq->q; |
| 488 | |
Jens Axboe | bd166ef | 2017-01-17 06:03:22 -0700 | [diff] [blame] | 489 | blk_mq_sched_started_request(rq); |
| 490 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 491 | trace_block_rq_issue(q, rq); |
| 492 | |
Jens Axboe | cf43e6b | 2016-11-07 21:32:37 -0700 | [diff] [blame] | 493 | if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) { |
Shaohua Li | 88eeca4 | 2017-03-27 15:19:41 -0700 | [diff] [blame] | 494 | blk_stat_set_issue(&rq->issue_stat, blk_rq_sectors(rq)); |
Jens Axboe | cf43e6b | 2016-11-07 21:32:37 -0700 | [diff] [blame] | 495 | rq->rq_flags |= RQF_STATS; |
Jens Axboe | 87760e5 | 2016-11-09 12:38:14 -0700 | [diff] [blame] | 496 | wbt_issue(q->rq_wb, &rq->issue_stat); |
Jens Axboe | cf43e6b | 2016-11-07 21:32:37 -0700 | [diff] [blame] | 497 | } |
| 498 | |
Ming Lei | 2b8393b | 2014-06-10 00:16:41 +0800 | [diff] [blame] | 499 | blk_add_timer(rq); |
Jens Axboe | 87ee7b1 | 2014-04-24 08:51:47 -0600 | [diff] [blame] | 500 | |
| 501 | /* |
Jens Axboe | 538b753 | 2014-09-16 10:37:37 -0600 | [diff] [blame] | 502 | * Ensure that ->deadline is visible before set the started |
| 503 | * flag and clear the completed flag. |
| 504 | */ |
| 505 | smp_mb__before_atomic(); |
| 506 | |
| 507 | /* |
Jens Axboe | 87ee7b1 | 2014-04-24 08:51:47 -0600 | [diff] [blame] | 508 | * Mark us as started and clear complete. Complete might have been |
| 509 | * set if requeue raced with timeout, which then marked it as |
| 510 | * complete. So be sure to clear complete again when we start |
| 511 | * the request, otherwise we'll ignore the completion event. |
| 512 | */ |
Jens Axboe | 4b57052 | 2014-05-29 11:00:11 -0600 | [diff] [blame] | 513 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) |
| 514 | set_bit(REQ_ATOM_STARTED, &rq->atomic_flags); |
| 515 | if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags)) |
| 516 | clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags); |
Christoph Hellwig | 49f5baa | 2014-02-11 08:27:14 -0800 | [diff] [blame] | 517 | |
| 518 | if (q->dma_drain_size && blk_rq_bytes(rq)) { |
| 519 | /* |
| 520 | * Make sure space for the drain appears. We know we can do |
| 521 | * this because max_hw_segments has been adjusted to be one |
| 522 | * fewer than the device can handle. |
| 523 | */ |
| 524 | rq->nr_phys_segments++; |
| 525 | } |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 526 | } |
Christoph Hellwig | e249007 | 2014-09-13 16:40:09 -0700 | [diff] [blame] | 527 | EXPORT_SYMBOL(blk_mq_start_request); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 528 | |
Ming Lei | d9d149a | 2017-03-27 20:06:55 +0800 | [diff] [blame] | 529 | /* |
| 530 | * When we reach here because queue is busy, REQ_ATOM_COMPLETE |
Jens Axboe | 48b99c9 | 2017-03-29 11:10:34 -0600 | [diff] [blame] | 531 | * flag isn't set yet, so there may be race with timeout handler, |
Ming Lei | d9d149a | 2017-03-27 20:06:55 +0800 | [diff] [blame] | 532 | * but given rq->deadline is just set in .queue_rq() under |
| 533 | * this situation, the race won't be possible in reality because |
| 534 | * rq->timeout should be set as big enough to cover the window |
| 535 | * between blk_mq_start_request() called from .queue_rq() and |
| 536 | * clearing REQ_ATOM_STARTED here. |
| 537 | */ |
Christoph Hellwig | ed0791b | 2014-04-16 09:44:57 +0200 | [diff] [blame] | 538 | static void __blk_mq_requeue_request(struct request *rq) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 539 | { |
| 540 | struct request_queue *q = rq->q; |
| 541 | |
| 542 | trace_block_rq_requeue(q, rq); |
Jens Axboe | 87760e5 | 2016-11-09 12:38:14 -0700 | [diff] [blame] | 543 | wbt_requeue(q->rq_wb, &rq->issue_stat); |
Jens Axboe | bd166ef | 2017-01-17 06:03:22 -0700 | [diff] [blame] | 544 | blk_mq_sched_requeue_request(rq); |
Christoph Hellwig | 49f5baa | 2014-02-11 08:27:14 -0800 | [diff] [blame] | 545 | |
Christoph Hellwig | e249007 | 2014-09-13 16:40:09 -0700 | [diff] [blame] | 546 | if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) { |
| 547 | if (q->dma_drain_size && blk_rq_bytes(rq)) |
| 548 | rq->nr_phys_segments--; |
| 549 | } |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 550 | } |
| 551 | |
Bart Van Assche | 2b053ac | 2016-10-28 17:21:41 -0700 | [diff] [blame] | 552 | void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list) |
Christoph Hellwig | ed0791b | 2014-04-16 09:44:57 +0200 | [diff] [blame] | 553 | { |
Christoph Hellwig | ed0791b | 2014-04-16 09:44:57 +0200 | [diff] [blame] | 554 | __blk_mq_requeue_request(rq); |
Christoph Hellwig | ed0791b | 2014-04-16 09:44:57 +0200 | [diff] [blame] | 555 | |
Christoph Hellwig | ed0791b | 2014-04-16 09:44:57 +0200 | [diff] [blame] | 556 | BUG_ON(blk_queued_rq(rq)); |
Bart Van Assche | 2b053ac | 2016-10-28 17:21:41 -0700 | [diff] [blame] | 557 | blk_mq_add_to_requeue_list(rq, true, kick_requeue_list); |
Christoph Hellwig | ed0791b | 2014-04-16 09:44:57 +0200 | [diff] [blame] | 558 | } |
| 559 | EXPORT_SYMBOL(blk_mq_requeue_request); |
| 560 | |
Christoph Hellwig | 6fca6a6 | 2014-05-28 08:08:02 -0600 | [diff] [blame] | 561 | static void blk_mq_requeue_work(struct work_struct *work) |
| 562 | { |
| 563 | struct request_queue *q = |
Mike Snitzer | 2849450 | 2016-09-14 13:28:30 -0400 | [diff] [blame] | 564 | container_of(work, struct request_queue, requeue_work.work); |
Christoph Hellwig | 6fca6a6 | 2014-05-28 08:08:02 -0600 | [diff] [blame] | 565 | LIST_HEAD(rq_list); |
| 566 | struct request *rq, *next; |
| 567 | unsigned long flags; |
| 568 | |
| 569 | spin_lock_irqsave(&q->requeue_lock, flags); |
| 570 | list_splice_init(&q->requeue_list, &rq_list); |
| 571 | spin_unlock_irqrestore(&q->requeue_lock, flags); |
| 572 | |
| 573 | list_for_each_entry_safe(rq, next, &rq_list, queuelist) { |
Christoph Hellwig | e806402 | 2016-10-20 15:12:13 +0200 | [diff] [blame] | 574 | if (!(rq->rq_flags & RQF_SOFTBARRIER)) |
Christoph Hellwig | 6fca6a6 | 2014-05-28 08:08:02 -0600 | [diff] [blame] | 575 | continue; |
| 576 | |
Christoph Hellwig | e806402 | 2016-10-20 15:12:13 +0200 | [diff] [blame] | 577 | rq->rq_flags &= ~RQF_SOFTBARRIER; |
Christoph Hellwig | 6fca6a6 | 2014-05-28 08:08:02 -0600 | [diff] [blame] | 578 | list_del_init(&rq->queuelist); |
Jens Axboe | bd6737f | 2017-01-27 01:00:47 -0700 | [diff] [blame] | 579 | blk_mq_sched_insert_request(rq, true, false, false, true); |
Christoph Hellwig | 6fca6a6 | 2014-05-28 08:08:02 -0600 | [diff] [blame] | 580 | } |
| 581 | |
| 582 | while (!list_empty(&rq_list)) { |
| 583 | rq = list_entry(rq_list.next, struct request, queuelist); |
| 584 | list_del_init(&rq->queuelist); |
Jens Axboe | bd6737f | 2017-01-27 01:00:47 -0700 | [diff] [blame] | 585 | blk_mq_sched_insert_request(rq, false, false, false, true); |
Christoph Hellwig | 6fca6a6 | 2014-05-28 08:08:02 -0600 | [diff] [blame] | 586 | } |
| 587 | |
Bart Van Assche | 52d7f1b | 2016-10-28 17:20:32 -0700 | [diff] [blame] | 588 | blk_mq_run_hw_queues(q, false); |
Christoph Hellwig | 6fca6a6 | 2014-05-28 08:08:02 -0600 | [diff] [blame] | 589 | } |
| 590 | |
Bart Van Assche | 2b053ac | 2016-10-28 17:21:41 -0700 | [diff] [blame] | 591 | void blk_mq_add_to_requeue_list(struct request *rq, bool at_head, |
| 592 | bool kick_requeue_list) |
Christoph Hellwig | 6fca6a6 | 2014-05-28 08:08:02 -0600 | [diff] [blame] | 593 | { |
| 594 | struct request_queue *q = rq->q; |
| 595 | unsigned long flags; |
| 596 | |
| 597 | /* |
| 598 | * We abuse this flag that is otherwise used by the I/O scheduler to |
| 599 | * request head insertation from the workqueue. |
| 600 | */ |
Christoph Hellwig | e806402 | 2016-10-20 15:12:13 +0200 | [diff] [blame] | 601 | BUG_ON(rq->rq_flags & RQF_SOFTBARRIER); |
Christoph Hellwig | 6fca6a6 | 2014-05-28 08:08:02 -0600 | [diff] [blame] | 602 | |
| 603 | spin_lock_irqsave(&q->requeue_lock, flags); |
| 604 | if (at_head) { |
Christoph Hellwig | e806402 | 2016-10-20 15:12:13 +0200 | [diff] [blame] | 605 | rq->rq_flags |= RQF_SOFTBARRIER; |
Christoph Hellwig | 6fca6a6 | 2014-05-28 08:08:02 -0600 | [diff] [blame] | 606 | list_add(&rq->queuelist, &q->requeue_list); |
| 607 | } else { |
| 608 | list_add_tail(&rq->queuelist, &q->requeue_list); |
| 609 | } |
| 610 | spin_unlock_irqrestore(&q->requeue_lock, flags); |
Bart Van Assche | 2b053ac | 2016-10-28 17:21:41 -0700 | [diff] [blame] | 611 | |
| 612 | if (kick_requeue_list) |
| 613 | blk_mq_kick_requeue_list(q); |
Christoph Hellwig | 6fca6a6 | 2014-05-28 08:08:02 -0600 | [diff] [blame] | 614 | } |
| 615 | EXPORT_SYMBOL(blk_mq_add_to_requeue_list); |
| 616 | |
| 617 | void blk_mq_kick_requeue_list(struct request_queue *q) |
| 618 | { |
Mike Snitzer | 2849450 | 2016-09-14 13:28:30 -0400 | [diff] [blame] | 619 | kblockd_schedule_delayed_work(&q->requeue_work, 0); |
Christoph Hellwig | 6fca6a6 | 2014-05-28 08:08:02 -0600 | [diff] [blame] | 620 | } |
| 621 | EXPORT_SYMBOL(blk_mq_kick_requeue_list); |
| 622 | |
Mike Snitzer | 2849450 | 2016-09-14 13:28:30 -0400 | [diff] [blame] | 623 | void blk_mq_delay_kick_requeue_list(struct request_queue *q, |
| 624 | unsigned long msecs) |
| 625 | { |
| 626 | kblockd_schedule_delayed_work(&q->requeue_work, |
| 627 | msecs_to_jiffies(msecs)); |
| 628 | } |
| 629 | EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list); |
| 630 | |
Jens Axboe | 0e62f51 | 2014-06-04 10:23:49 -0600 | [diff] [blame] | 631 | struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag) |
| 632 | { |
Jens Axboe | 88c7b2b | 2016-08-25 08:07:30 -0600 | [diff] [blame] | 633 | if (tag < tags->nr_tags) { |
| 634 | prefetch(tags->rqs[tag]); |
Hannes Reinecke | 4ee86ba | 2016-03-15 12:03:28 -0700 | [diff] [blame] | 635 | return tags->rqs[tag]; |
Jens Axboe | 88c7b2b | 2016-08-25 08:07:30 -0600 | [diff] [blame] | 636 | } |
Hannes Reinecke | 4ee86ba | 2016-03-15 12:03:28 -0700 | [diff] [blame] | 637 | |
| 638 | return NULL; |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 639 | } |
| 640 | EXPORT_SYMBOL(blk_mq_tag_to_rq); |
| 641 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 642 | struct blk_mq_timeout_data { |
Christoph Hellwig | 46f92d4 | 2014-09-13 16:40:12 -0700 | [diff] [blame] | 643 | unsigned long next; |
| 644 | unsigned int next_set; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 645 | }; |
| 646 | |
Christoph Hellwig | 9041583 | 2014-09-22 10:21:48 -0600 | [diff] [blame] | 647 | void blk_mq_rq_timed_out(struct request *req, bool reserved) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 648 | { |
Jens Axboe | f8a5b12 | 2016-12-13 09:24:51 -0700 | [diff] [blame] | 649 | const struct blk_mq_ops *ops = req->q->mq_ops; |
Christoph Hellwig | 46f92d4 | 2014-09-13 16:40:12 -0700 | [diff] [blame] | 650 | enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER; |
Jens Axboe | 87ee7b1 | 2014-04-24 08:51:47 -0600 | [diff] [blame] | 651 | |
| 652 | /* |
| 653 | * We know that complete is set at this point. If STARTED isn't set |
| 654 | * anymore, then the request isn't active and the "timeout" should |
| 655 | * just be ignored. This can happen due to the bitflag ordering. |
| 656 | * Timeout first checks if STARTED is set, and if it is, assumes |
| 657 | * the request is active. But if we race with completion, then |
Jens Axboe | 48b99c9 | 2017-03-29 11:10:34 -0600 | [diff] [blame] | 658 | * both flags will get cleared. So check here again, and ignore |
Jens Axboe | 87ee7b1 | 2014-04-24 08:51:47 -0600 | [diff] [blame] | 659 | * a timeout event with a request that isn't active. |
| 660 | */ |
Christoph Hellwig | 46f92d4 | 2014-09-13 16:40:12 -0700 | [diff] [blame] | 661 | if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags)) |
| 662 | return; |
Jens Axboe | 87ee7b1 | 2014-04-24 08:51:47 -0600 | [diff] [blame] | 663 | |
Christoph Hellwig | 46f92d4 | 2014-09-13 16:40:12 -0700 | [diff] [blame] | 664 | if (ops->timeout) |
Christoph Hellwig | 0152fb6 | 2014-09-13 16:40:13 -0700 | [diff] [blame] | 665 | ret = ops->timeout(req, reserved); |
Jens Axboe | 87ee7b1 | 2014-04-24 08:51:47 -0600 | [diff] [blame] | 666 | |
Christoph Hellwig | 46f92d4 | 2014-09-13 16:40:12 -0700 | [diff] [blame] | 667 | switch (ret) { |
| 668 | case BLK_EH_HANDLED: |
| 669 | __blk_mq_complete_request(req); |
| 670 | break; |
| 671 | case BLK_EH_RESET_TIMER: |
| 672 | blk_add_timer(req); |
| 673 | blk_clear_rq_complete(req); |
| 674 | break; |
| 675 | case BLK_EH_NOT_HANDLED: |
| 676 | break; |
| 677 | default: |
| 678 | printk(KERN_ERR "block: bad eh return: %d\n", ret); |
| 679 | break; |
| 680 | } |
Jens Axboe | 87ee7b1 | 2014-04-24 08:51:47 -0600 | [diff] [blame] | 681 | } |
Keith Busch | 5b3f25f | 2015-01-07 18:55:46 -0700 | [diff] [blame] | 682 | |
Christoph Hellwig | 81481eb | 2014-09-13 16:40:11 -0700 | [diff] [blame] | 683 | static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx, |
| 684 | struct request *rq, void *priv, bool reserved) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 685 | { |
Christoph Hellwig | 81481eb | 2014-09-13 16:40:11 -0700 | [diff] [blame] | 686 | struct blk_mq_timeout_data *data = priv; |
| 687 | |
Ming Lei | 95a4960 | 2017-03-22 10:14:43 +0800 | [diff] [blame] | 688 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) |
Christoph Hellwig | 46f92d4 | 2014-09-13 16:40:12 -0700 | [diff] [blame] | 689 | return; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 690 | |
Ming Lei | d9d149a | 2017-03-27 20:06:55 +0800 | [diff] [blame] | 691 | /* |
| 692 | * The rq being checked may have been freed and reallocated |
| 693 | * out already here, we avoid this race by checking rq->deadline |
| 694 | * and REQ_ATOM_COMPLETE flag together: |
| 695 | * |
| 696 | * - if rq->deadline is observed as new value because of |
| 697 | * reusing, the rq won't be timed out because of timing. |
| 698 | * - if rq->deadline is observed as previous value, |
| 699 | * REQ_ATOM_COMPLETE flag won't be cleared in reuse path |
| 700 | * because we put a barrier between setting rq->deadline |
| 701 | * and clearing the flag in blk_mq_start_request(), so |
| 702 | * this rq won't be timed out too. |
| 703 | */ |
Christoph Hellwig | 46f92d4 | 2014-09-13 16:40:12 -0700 | [diff] [blame] | 704 | if (time_after_eq(jiffies, rq->deadline)) { |
| 705 | if (!blk_mark_rq_complete(rq)) |
Christoph Hellwig | 0152fb6 | 2014-09-13 16:40:13 -0700 | [diff] [blame] | 706 | blk_mq_rq_timed_out(rq, reserved); |
Christoph Hellwig | 46f92d4 | 2014-09-13 16:40:12 -0700 | [diff] [blame] | 707 | } else if (!data->next_set || time_after(data->next, rq->deadline)) { |
| 708 | data->next = rq->deadline; |
| 709 | data->next_set = 1; |
| 710 | } |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 711 | } |
| 712 | |
Christoph Hellwig | 287922eb | 2015-10-30 20:57:30 +0800 | [diff] [blame] | 713 | static void blk_mq_timeout_work(struct work_struct *work) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 714 | { |
Christoph Hellwig | 287922eb | 2015-10-30 20:57:30 +0800 | [diff] [blame] | 715 | struct request_queue *q = |
| 716 | container_of(work, struct request_queue, timeout_work); |
Christoph Hellwig | 81481eb | 2014-09-13 16:40:11 -0700 | [diff] [blame] | 717 | struct blk_mq_timeout_data data = { |
| 718 | .next = 0, |
| 719 | .next_set = 0, |
| 720 | }; |
Christoph Hellwig | 81481eb | 2014-09-13 16:40:11 -0700 | [diff] [blame] | 721 | int i; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 722 | |
Gabriel Krisman Bertazi | 71f79fb | 2016-08-01 08:23:39 -0600 | [diff] [blame] | 723 | /* A deadlock might occur if a request is stuck requiring a |
| 724 | * timeout at the same time a queue freeze is waiting |
| 725 | * completion, since the timeout code would not be able to |
| 726 | * acquire the queue reference here. |
| 727 | * |
| 728 | * That's why we don't use blk_queue_enter here; instead, we use |
| 729 | * percpu_ref_tryget directly, because we need to be able to |
| 730 | * obtain a reference even in the short window between the queue |
| 731 | * starting to freeze, by dropping the first reference in |
Ming Lei | 1671d52 | 2017-03-27 20:06:57 +0800 | [diff] [blame] | 732 | * blk_freeze_queue_start, and the moment the last request is |
Gabriel Krisman Bertazi | 71f79fb | 2016-08-01 08:23:39 -0600 | [diff] [blame] | 733 | * consumed, marked by the instant q_usage_counter reaches |
| 734 | * zero. |
| 735 | */ |
| 736 | if (!percpu_ref_tryget(&q->q_usage_counter)) |
Christoph Hellwig | 287922eb | 2015-10-30 20:57:30 +0800 | [diff] [blame] | 737 | return; |
| 738 | |
Christoph Hellwig | 0bf6cd5 | 2015-09-27 21:01:51 +0200 | [diff] [blame] | 739 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 740 | |
Christoph Hellwig | 81481eb | 2014-09-13 16:40:11 -0700 | [diff] [blame] | 741 | if (data.next_set) { |
| 742 | data.next = blk_rq_timeout(round_jiffies_up(data.next)); |
| 743 | mod_timer(&q->timeout, data.next); |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 744 | } else { |
Christoph Hellwig | 0bf6cd5 | 2015-09-27 21:01:51 +0200 | [diff] [blame] | 745 | struct blk_mq_hw_ctx *hctx; |
| 746 | |
Ming Lei | f054b56 | 2015-04-21 10:00:19 +0800 | [diff] [blame] | 747 | queue_for_each_hw_ctx(q, hctx, i) { |
| 748 | /* the hctx may be unmapped, so check it here */ |
| 749 | if (blk_mq_hw_queue_mapped(hctx)) |
| 750 | blk_mq_tag_idle(hctx); |
| 751 | } |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 752 | } |
Christoph Hellwig | 287922eb | 2015-10-30 20:57:30 +0800 | [diff] [blame] | 753 | blk_queue_exit(q); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 754 | } |
| 755 | |
| 756 | /* |
| 757 | * Reverse check our software queue for entries that we could potentially |
| 758 | * merge with. Currently includes a hand-wavy stop count of 8, to not spend |
| 759 | * too much time checking for merges. |
| 760 | */ |
| 761 | static bool blk_mq_attempt_merge(struct request_queue *q, |
| 762 | struct blk_mq_ctx *ctx, struct bio *bio) |
| 763 | { |
| 764 | struct request *rq; |
| 765 | int checked = 8; |
| 766 | |
| 767 | list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) { |
Christoph Hellwig | 34fe7c0 | 2017-02-08 14:46:48 +0100 | [diff] [blame] | 768 | bool merged = false; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 769 | |
| 770 | if (!checked--) |
| 771 | break; |
| 772 | |
| 773 | if (!blk_rq_merge_ok(rq, bio)) |
| 774 | continue; |
| 775 | |
Christoph Hellwig | 34fe7c0 | 2017-02-08 14:46:48 +0100 | [diff] [blame] | 776 | switch (blk_try_merge(rq, bio)) { |
| 777 | case ELEVATOR_BACK_MERGE: |
| 778 | if (blk_mq_sched_allow_merge(q, rq, bio)) |
| 779 | merged = bio_attempt_back_merge(q, rq, bio); |
| 780 | break; |
| 781 | case ELEVATOR_FRONT_MERGE: |
| 782 | if (blk_mq_sched_allow_merge(q, rq, bio)) |
| 783 | merged = bio_attempt_front_merge(q, rq, bio); |
| 784 | break; |
Christoph Hellwig | 1e73973 | 2017-02-08 14:46:49 +0100 | [diff] [blame] | 785 | case ELEVATOR_DISCARD_MERGE: |
| 786 | merged = bio_attempt_discard_merge(q, rq, bio); |
| 787 | break; |
Christoph Hellwig | 34fe7c0 | 2017-02-08 14:46:48 +0100 | [diff] [blame] | 788 | default: |
Jens Axboe | bd166ef | 2017-01-17 06:03:22 -0700 | [diff] [blame] | 789 | continue; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 790 | } |
Christoph Hellwig | 34fe7c0 | 2017-02-08 14:46:48 +0100 | [diff] [blame] | 791 | |
| 792 | if (merged) |
| 793 | ctx->rq_merged++; |
| 794 | return merged; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 795 | } |
| 796 | |
| 797 | return false; |
| 798 | } |
| 799 | |
Omar Sandoval | 8845964 | 2016-09-17 08:38:44 -0600 | [diff] [blame] | 800 | struct flush_busy_ctx_data { |
| 801 | struct blk_mq_hw_ctx *hctx; |
| 802 | struct list_head *list; |
| 803 | }; |
| 804 | |
| 805 | static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data) |
| 806 | { |
| 807 | struct flush_busy_ctx_data *flush_data = data; |
| 808 | struct blk_mq_hw_ctx *hctx = flush_data->hctx; |
| 809 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; |
| 810 | |
| 811 | sbitmap_clear_bit(sb, bitnr); |
| 812 | spin_lock(&ctx->lock); |
| 813 | list_splice_tail_init(&ctx->rq_list, flush_data->list); |
| 814 | spin_unlock(&ctx->lock); |
| 815 | return true; |
| 816 | } |
| 817 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 818 | /* |
Jens Axboe | 1429d7c | 2014-05-19 09:23:55 -0600 | [diff] [blame] | 819 | * Process software queues that have been marked busy, splicing them |
| 820 | * to the for-dispatch |
| 821 | */ |
Jens Axboe | 2c3ad66 | 2016-12-14 14:34:47 -0700 | [diff] [blame] | 822 | 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] | 823 | { |
Omar Sandoval | 8845964 | 2016-09-17 08:38:44 -0600 | [diff] [blame] | 824 | struct flush_busy_ctx_data data = { |
| 825 | .hctx = hctx, |
| 826 | .list = list, |
| 827 | }; |
Jens Axboe | 1429d7c | 2014-05-19 09:23:55 -0600 | [diff] [blame] | 828 | |
Omar Sandoval | 8845964 | 2016-09-17 08:38:44 -0600 | [diff] [blame] | 829 | sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data); |
Jens Axboe | 1429d7c | 2014-05-19 09:23:55 -0600 | [diff] [blame] | 830 | } |
Jens Axboe | 2c3ad66 | 2016-12-14 14:34:47 -0700 | [diff] [blame] | 831 | EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs); |
Jens Axboe | 1429d7c | 2014-05-19 09:23:55 -0600 | [diff] [blame] | 832 | |
Jens Axboe | 703fd1c | 2016-09-16 13:59:14 -0600 | [diff] [blame] | 833 | static inline unsigned int queued_to_index(unsigned int queued) |
| 834 | { |
| 835 | if (!queued) |
| 836 | return 0; |
Jens Axboe | 1429d7c | 2014-05-19 09:23:55 -0600 | [diff] [blame] | 837 | |
Jens Axboe | 703fd1c | 2016-09-16 13:59:14 -0600 | [diff] [blame] | 838 | return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1); |
Jens Axboe | 1429d7c | 2014-05-19 09:23:55 -0600 | [diff] [blame] | 839 | } |
| 840 | |
Jens Axboe | bd6737f | 2017-01-27 01:00:47 -0700 | [diff] [blame] | 841 | bool blk_mq_get_driver_tag(struct request *rq, struct blk_mq_hw_ctx **hctx, |
| 842 | bool wait) |
Jens Axboe | bd166ef | 2017-01-17 06:03:22 -0700 | [diff] [blame] | 843 | { |
| 844 | struct blk_mq_alloc_data data = { |
| 845 | .q = rq->q, |
Jens Axboe | bd166ef | 2017-01-17 06:03:22 -0700 | [diff] [blame] | 846 | .hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu), |
| 847 | .flags = wait ? 0 : BLK_MQ_REQ_NOWAIT, |
| 848 | }; |
| 849 | |
Jens Axboe | 5feeacd | 2017-04-20 17:23:13 -0600 | [diff] [blame] | 850 | might_sleep_if(wait); |
| 851 | |
Omar Sandoval | 81380ca | 2017-04-07 08:56:26 -0600 | [diff] [blame] | 852 | if (rq->tag != -1) |
| 853 | goto done; |
Jens Axboe | bd166ef | 2017-01-17 06:03:22 -0700 | [diff] [blame] | 854 | |
Sagi Grimberg | 415b806 | 2017-02-27 10:04:39 -0700 | [diff] [blame] | 855 | if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag)) |
| 856 | data.flags |= BLK_MQ_REQ_RESERVED; |
| 857 | |
Jens Axboe | bd166ef | 2017-01-17 06:03:22 -0700 | [diff] [blame] | 858 | rq->tag = blk_mq_get_tag(&data); |
| 859 | if (rq->tag >= 0) { |
Jens Axboe | 200e86b | 2017-01-25 08:11:38 -0700 | [diff] [blame] | 860 | if (blk_mq_tag_busy(data.hctx)) { |
| 861 | rq->rq_flags |= RQF_MQ_INFLIGHT; |
| 862 | atomic_inc(&data.hctx->nr_active); |
| 863 | } |
Jens Axboe | bd166ef | 2017-01-17 06:03:22 -0700 | [diff] [blame] | 864 | data.hctx->tags->rqs[rq->tag] = rq; |
Jens Axboe | bd166ef | 2017-01-17 06:03:22 -0700 | [diff] [blame] | 865 | } |
| 866 | |
Omar Sandoval | 81380ca | 2017-04-07 08:56:26 -0600 | [diff] [blame] | 867 | done: |
| 868 | if (hctx) |
| 869 | *hctx = data.hctx; |
| 870 | return rq->tag != -1; |
Jens Axboe | bd166ef | 2017-01-17 06:03:22 -0700 | [diff] [blame] | 871 | } |
| 872 | |
Jens Axboe | 113285b | 2017-03-02 13:26:04 -0700 | [diff] [blame] | 873 | static void __blk_mq_put_driver_tag(struct blk_mq_hw_ctx *hctx, |
| 874 | struct request *rq) |
Jens Axboe | 99cf1dc | 2017-01-26 12:32:32 -0700 | [diff] [blame] | 875 | { |
Jens Axboe | 99cf1dc | 2017-01-26 12:32:32 -0700 | [diff] [blame] | 876 | blk_mq_put_tag(hctx, hctx->tags, rq->mq_ctx, rq->tag); |
| 877 | rq->tag = -1; |
| 878 | |
| 879 | if (rq->rq_flags & RQF_MQ_INFLIGHT) { |
| 880 | rq->rq_flags &= ~RQF_MQ_INFLIGHT; |
| 881 | atomic_dec(&hctx->nr_active); |
| 882 | } |
| 883 | } |
| 884 | |
Jens Axboe | 113285b | 2017-03-02 13:26:04 -0700 | [diff] [blame] | 885 | static void blk_mq_put_driver_tag_hctx(struct blk_mq_hw_ctx *hctx, |
| 886 | struct request *rq) |
| 887 | { |
| 888 | if (rq->tag == -1 || rq->internal_tag == -1) |
| 889 | return; |
| 890 | |
| 891 | __blk_mq_put_driver_tag(hctx, rq); |
| 892 | } |
| 893 | |
| 894 | static void blk_mq_put_driver_tag(struct request *rq) |
| 895 | { |
| 896 | struct blk_mq_hw_ctx *hctx; |
| 897 | |
| 898 | if (rq->tag == -1 || rq->internal_tag == -1) |
| 899 | return; |
| 900 | |
| 901 | hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu); |
| 902 | __blk_mq_put_driver_tag(hctx, rq); |
| 903 | } |
| 904 | |
Jens Axboe | bd166ef | 2017-01-17 06:03:22 -0700 | [diff] [blame] | 905 | /* |
| 906 | * If we fail getting a driver tag because all the driver tags are already |
| 907 | * assigned and on the dispatch list, BUT the first entry does not have a |
| 908 | * tag, then we could deadlock. For that case, move entries with assigned |
| 909 | * driver tags to the front, leaving the set of tagged requests in the |
| 910 | * same order, and the untagged set in the same order. |
| 911 | */ |
| 912 | static bool reorder_tags_to_front(struct list_head *list) |
| 913 | { |
| 914 | struct request *rq, *tmp, *first = NULL; |
| 915 | |
| 916 | list_for_each_entry_safe_reverse(rq, tmp, list, queuelist) { |
| 917 | if (rq == first) |
| 918 | break; |
| 919 | if (rq->tag != -1) { |
| 920 | list_move(&rq->queuelist, list); |
| 921 | if (!first) |
| 922 | first = rq; |
| 923 | } |
| 924 | } |
| 925 | |
| 926 | return first != NULL; |
| 927 | } |
| 928 | |
Omar Sandoval | da55f2c | 2017-02-22 10:58:29 -0800 | [diff] [blame] | 929 | static int blk_mq_dispatch_wake(wait_queue_t *wait, unsigned mode, int flags, |
| 930 | void *key) |
| 931 | { |
| 932 | struct blk_mq_hw_ctx *hctx; |
| 933 | |
| 934 | hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait); |
| 935 | |
| 936 | list_del(&wait->task_list); |
| 937 | clear_bit_unlock(BLK_MQ_S_TAG_WAITING, &hctx->state); |
| 938 | blk_mq_run_hw_queue(hctx, true); |
| 939 | return 1; |
| 940 | } |
| 941 | |
| 942 | static bool blk_mq_dispatch_wait_add(struct blk_mq_hw_ctx *hctx) |
| 943 | { |
| 944 | struct sbq_wait_state *ws; |
| 945 | |
| 946 | /* |
| 947 | * The TAG_WAITING bit serves as a lock protecting hctx->dispatch_wait. |
| 948 | * The thread which wins the race to grab this bit adds the hardware |
| 949 | * queue to the wait queue. |
| 950 | */ |
| 951 | if (test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state) || |
| 952 | test_and_set_bit_lock(BLK_MQ_S_TAG_WAITING, &hctx->state)) |
| 953 | return false; |
| 954 | |
| 955 | init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake); |
| 956 | ws = bt_wait_ptr(&hctx->tags->bitmap_tags, hctx); |
| 957 | |
| 958 | /* |
| 959 | * As soon as this returns, it's no longer safe to fiddle with |
| 960 | * hctx->dispatch_wait, since a completion can wake up the wait queue |
| 961 | * and unlock the bit. |
| 962 | */ |
| 963 | add_wait_queue(&ws->wait, &hctx->dispatch_wait); |
| 964 | return true; |
| 965 | } |
| 966 | |
Omar Sandoval | 81380ca | 2017-04-07 08:56:26 -0600 | [diff] [blame] | 967 | bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list) |
Jens Axboe | f04c3df | 2016-12-07 08:41:17 -0700 | [diff] [blame] | 968 | { |
Omar Sandoval | 81380ca | 2017-04-07 08:56:26 -0600 | [diff] [blame] | 969 | struct blk_mq_hw_ctx *hctx; |
Jens Axboe | f04c3df | 2016-12-07 08:41:17 -0700 | [diff] [blame] | 970 | struct request *rq; |
Jens Axboe | 93efe98 | 2017-03-24 12:04:19 -0600 | [diff] [blame] | 971 | int errors, queued, ret = BLK_MQ_RQ_QUEUE_OK; |
Jens Axboe | f04c3df | 2016-12-07 08:41:17 -0700 | [diff] [blame] | 972 | |
Omar Sandoval | 81380ca | 2017-04-07 08:56:26 -0600 | [diff] [blame] | 973 | if (list_empty(list)) |
| 974 | return false; |
| 975 | |
Jens Axboe | f04c3df | 2016-12-07 08:41:17 -0700 | [diff] [blame] | 976 | /* |
Jens Axboe | f04c3df | 2016-12-07 08:41:17 -0700 | [diff] [blame] | 977 | * Now process all the entries, sending them to the driver. |
| 978 | */ |
Jens Axboe | 93efe98 | 2017-03-24 12:04:19 -0600 | [diff] [blame] | 979 | errors = queued = 0; |
Omar Sandoval | 81380ca | 2017-04-07 08:56:26 -0600 | [diff] [blame] | 980 | do { |
Jens Axboe | f04c3df | 2016-12-07 08:41:17 -0700 | [diff] [blame] | 981 | struct blk_mq_queue_data bd; |
| 982 | |
| 983 | rq = list_first_entry(list, struct request, queuelist); |
Jens Axboe | bd166ef | 2017-01-17 06:03:22 -0700 | [diff] [blame] | 984 | if (!blk_mq_get_driver_tag(rq, &hctx, false)) { |
| 985 | if (!queued && reorder_tags_to_front(list)) |
| 986 | continue; |
Jens Axboe | 3c782d6 | 2017-01-26 12:50:36 -0700 | [diff] [blame] | 987 | |
| 988 | /* |
Omar Sandoval | da55f2c | 2017-02-22 10:58:29 -0800 | [diff] [blame] | 989 | * The initial allocation attempt failed, so we need to |
| 990 | * rerun the hardware queue when a tag is freed. |
Jens Axboe | 3c782d6 | 2017-01-26 12:50:36 -0700 | [diff] [blame] | 991 | */ |
Omar Sandoval | 807b104 | 2017-04-05 12:01:35 -0700 | [diff] [blame] | 992 | if (!blk_mq_dispatch_wait_add(hctx)) |
Jens Axboe | 3c782d6 | 2017-01-26 12:50:36 -0700 | [diff] [blame] | 993 | break; |
Omar Sandoval | 807b104 | 2017-04-05 12:01:35 -0700 | [diff] [blame] | 994 | |
| 995 | /* |
| 996 | * It's possible that a tag was freed in the window |
| 997 | * between the allocation failure and adding the |
| 998 | * hardware queue to the wait queue. |
| 999 | */ |
| 1000 | if (!blk_mq_get_driver_tag(rq, &hctx, false)) |
| 1001 | break; |
Jens Axboe | bd166ef | 2017-01-17 06:03:22 -0700 | [diff] [blame] | 1002 | } |
Omar Sandoval | da55f2c | 2017-02-22 10:58:29 -0800 | [diff] [blame] | 1003 | |
Jens Axboe | f04c3df | 2016-12-07 08:41:17 -0700 | [diff] [blame] | 1004 | list_del_init(&rq->queuelist); |
| 1005 | |
| 1006 | bd.rq = rq; |
Jens Axboe | 113285b | 2017-03-02 13:26:04 -0700 | [diff] [blame] | 1007 | |
| 1008 | /* |
| 1009 | * Flag last if we have no more requests, or if we have more |
| 1010 | * but can't assign a driver tag to it. |
| 1011 | */ |
| 1012 | if (list_empty(list)) |
| 1013 | bd.last = true; |
| 1014 | else { |
| 1015 | struct request *nxt; |
| 1016 | |
| 1017 | nxt = list_first_entry(list, struct request, queuelist); |
| 1018 | bd.last = !blk_mq_get_driver_tag(nxt, NULL, false); |
| 1019 | } |
Jens Axboe | f04c3df | 2016-12-07 08:41:17 -0700 | [diff] [blame] | 1020 | |
| 1021 | ret = q->mq_ops->queue_rq(hctx, &bd); |
| 1022 | switch (ret) { |
| 1023 | case BLK_MQ_RQ_QUEUE_OK: |
| 1024 | queued++; |
| 1025 | break; |
| 1026 | case BLK_MQ_RQ_QUEUE_BUSY: |
Jens Axboe | 113285b | 2017-03-02 13:26:04 -0700 | [diff] [blame] | 1027 | blk_mq_put_driver_tag_hctx(hctx, rq); |
Jens Axboe | f04c3df | 2016-12-07 08:41:17 -0700 | [diff] [blame] | 1028 | list_add(&rq->queuelist, list); |
| 1029 | __blk_mq_requeue_request(rq); |
| 1030 | break; |
| 1031 | default: |
| 1032 | pr_err("blk-mq: bad return on queue: %d\n", ret); |
| 1033 | case BLK_MQ_RQ_QUEUE_ERROR: |
Jens Axboe | 93efe98 | 2017-03-24 12:04:19 -0600 | [diff] [blame] | 1034 | errors++; |
Christoph Hellwig | caf7df1 | 2017-04-20 16:03:16 +0200 | [diff] [blame] | 1035 | blk_mq_end_request(rq, -EIO); |
Jens Axboe | f04c3df | 2016-12-07 08:41:17 -0700 | [diff] [blame] | 1036 | break; |
| 1037 | } |
| 1038 | |
| 1039 | if (ret == BLK_MQ_RQ_QUEUE_BUSY) |
| 1040 | break; |
Omar Sandoval | 81380ca | 2017-04-07 08:56:26 -0600 | [diff] [blame] | 1041 | } while (!list_empty(list)); |
Jens Axboe | f04c3df | 2016-12-07 08:41:17 -0700 | [diff] [blame] | 1042 | |
| 1043 | hctx->dispatched[queued_to_index(queued)]++; |
| 1044 | |
| 1045 | /* |
| 1046 | * Any items that need requeuing? Stuff them into hctx->dispatch, |
| 1047 | * that is where we will continue on next queue run. |
| 1048 | */ |
| 1049 | if (!list_empty(list)) { |
Jens Axboe | 113285b | 2017-03-02 13:26:04 -0700 | [diff] [blame] | 1050 | /* |
Bart Van Assche | 710c785 | 2017-04-07 11:16:51 -0700 | [diff] [blame] | 1051 | * If an I/O scheduler has been configured and we got a driver |
| 1052 | * tag for the next request already, free it again. |
Jens Axboe | 113285b | 2017-03-02 13:26:04 -0700 | [diff] [blame] | 1053 | */ |
| 1054 | rq = list_first_entry(list, struct request, queuelist); |
| 1055 | blk_mq_put_driver_tag(rq); |
| 1056 | |
Jens Axboe | f04c3df | 2016-12-07 08:41:17 -0700 | [diff] [blame] | 1057 | spin_lock(&hctx->lock); |
Jens Axboe | c13660a | 2017-01-26 12:40:07 -0700 | [diff] [blame] | 1058 | list_splice_init(list, &hctx->dispatch); |
Jens Axboe | f04c3df | 2016-12-07 08:41:17 -0700 | [diff] [blame] | 1059 | spin_unlock(&hctx->lock); |
| 1060 | |
| 1061 | /* |
Bart Van Assche | 710c785 | 2017-04-07 11:16:51 -0700 | [diff] [blame] | 1062 | * If SCHED_RESTART was set by the caller of this function and |
| 1063 | * it is no longer set that means that it was cleared by another |
| 1064 | * thread and hence that a queue rerun is needed. |
Jens Axboe | f04c3df | 2016-12-07 08:41:17 -0700 | [diff] [blame] | 1065 | * |
Bart Van Assche | 710c785 | 2017-04-07 11:16:51 -0700 | [diff] [blame] | 1066 | * If TAG_WAITING is set that means that an I/O scheduler has |
| 1067 | * been configured and another thread is waiting for a driver |
| 1068 | * tag. To guarantee fairness, do not rerun this hardware queue |
| 1069 | * but let the other thread grab the driver tag. |
Jens Axboe | bd166ef | 2017-01-17 06:03:22 -0700 | [diff] [blame] | 1070 | * |
Bart Van Assche | 710c785 | 2017-04-07 11:16:51 -0700 | [diff] [blame] | 1071 | * If no I/O scheduler has been configured it is possible that |
| 1072 | * the hardware queue got stopped and restarted before requests |
| 1073 | * were pushed back onto the dispatch list. Rerun the queue to |
| 1074 | * avoid starvation. Notes: |
| 1075 | * - blk_mq_run_hw_queue() checks whether or not a queue has |
| 1076 | * been stopped before rerunning a queue. |
| 1077 | * - Some but not all block drivers stop a queue before |
| 1078 | * returning BLK_MQ_RQ_QUEUE_BUSY. Two exceptions are scsi-mq |
| 1079 | * and dm-rq. |
Jens Axboe | bd166ef | 2017-01-17 06:03:22 -0700 | [diff] [blame] | 1080 | */ |
Omar Sandoval | da55f2c | 2017-02-22 10:58:29 -0800 | [diff] [blame] | 1081 | if (!blk_mq_sched_needs_restart(hctx) && |
| 1082 | !test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state)) |
Jens Axboe | bd166ef | 2017-01-17 06:03:22 -0700 | [diff] [blame] | 1083 | blk_mq_run_hw_queue(hctx, true); |
Jens Axboe | f04c3df | 2016-12-07 08:41:17 -0700 | [diff] [blame] | 1084 | } |
| 1085 | |
Jens Axboe | 93efe98 | 2017-03-24 12:04:19 -0600 | [diff] [blame] | 1086 | return (queued + errors) != 0; |
Jens Axboe | f04c3df | 2016-12-07 08:41:17 -0700 | [diff] [blame] | 1087 | } |
| 1088 | |
Bart Van Assche | 6a83e74 | 2016-11-02 10:09:51 -0600 | [diff] [blame] | 1089 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx) |
| 1090 | { |
| 1091 | int srcu_idx; |
| 1092 | |
| 1093 | WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) && |
| 1094 | cpu_online(hctx->next_cpu)); |
| 1095 | |
| 1096 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) { |
| 1097 | rcu_read_lock(); |
Jens Axboe | bd166ef | 2017-01-17 06:03:22 -0700 | [diff] [blame] | 1098 | blk_mq_sched_dispatch_requests(hctx); |
Bart Van Assche | 6a83e74 | 2016-11-02 10:09:51 -0600 | [diff] [blame] | 1099 | rcu_read_unlock(); |
| 1100 | } else { |
Jens Axboe | bf4907c | 2017-03-30 12:30:39 -0600 | [diff] [blame] | 1101 | might_sleep(); |
| 1102 | |
Bart Van Assche | 6a83e74 | 2016-11-02 10:09:51 -0600 | [diff] [blame] | 1103 | srcu_idx = srcu_read_lock(&hctx->queue_rq_srcu); |
Jens Axboe | bd166ef | 2017-01-17 06:03:22 -0700 | [diff] [blame] | 1104 | blk_mq_sched_dispatch_requests(hctx); |
Bart Van Assche | 6a83e74 | 2016-11-02 10:09:51 -0600 | [diff] [blame] | 1105 | srcu_read_unlock(&hctx->queue_rq_srcu, srcu_idx); |
| 1106 | } |
| 1107 | } |
| 1108 | |
Jens Axboe | 506e931 | 2014-05-07 10:26:44 -0600 | [diff] [blame] | 1109 | /* |
| 1110 | * It'd be great if the workqueue API had a way to pass |
| 1111 | * in a mask and had some smarts for more clever placement. |
| 1112 | * For now we just round-robin here, switching for every |
| 1113 | * BLK_MQ_CPU_WORK_BATCH queued items. |
| 1114 | */ |
| 1115 | static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx) |
| 1116 | { |
Christoph Hellwig | b657d7e | 2014-11-24 09:27:23 +0100 | [diff] [blame] | 1117 | if (hctx->queue->nr_hw_queues == 1) |
| 1118 | return WORK_CPU_UNBOUND; |
Jens Axboe | 506e931 | 2014-05-07 10:26:44 -0600 | [diff] [blame] | 1119 | |
| 1120 | if (--hctx->next_cpu_batch <= 0) { |
Gabriel Krisman Bertazi | c02ebfd | 2016-09-28 00:24:24 -0300 | [diff] [blame] | 1121 | int next_cpu; |
Jens Axboe | 506e931 | 2014-05-07 10:26:44 -0600 | [diff] [blame] | 1122 | |
| 1123 | next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask); |
| 1124 | if (next_cpu >= nr_cpu_ids) |
| 1125 | next_cpu = cpumask_first(hctx->cpumask); |
| 1126 | |
| 1127 | hctx->next_cpu = next_cpu; |
| 1128 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; |
| 1129 | } |
| 1130 | |
Christoph Hellwig | b657d7e | 2014-11-24 09:27:23 +0100 | [diff] [blame] | 1131 | return hctx->next_cpu; |
Jens Axboe | 506e931 | 2014-05-07 10:26:44 -0600 | [diff] [blame] | 1132 | } |
| 1133 | |
Bart Van Assche | 7587a5a | 2017-04-07 11:16:52 -0700 | [diff] [blame] | 1134 | static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async, |
| 1135 | unsigned long msecs) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1136 | { |
Bart Van Assche | 5d1b25c | 2016-10-28 17:19:15 -0700 | [diff] [blame] | 1137 | if (unlikely(blk_mq_hctx_stopped(hctx) || |
| 1138 | !blk_mq_hw_queue_mapped(hctx))) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1139 | return; |
| 1140 | |
Jens Axboe | 1b792f2 | 2016-09-21 10:12:13 -0600 | [diff] [blame] | 1141 | if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) { |
Paolo Bonzini | 2a90d4a | 2014-11-07 23:04:00 +0100 | [diff] [blame] | 1142 | int cpu = get_cpu(); |
| 1143 | if (cpumask_test_cpu(cpu, hctx->cpumask)) { |
Paolo Bonzini | 398205b | 2014-11-07 23:03:59 +0100 | [diff] [blame] | 1144 | __blk_mq_run_hw_queue(hctx); |
Paolo Bonzini | 2a90d4a | 2014-11-07 23:04:00 +0100 | [diff] [blame] | 1145 | put_cpu(); |
Paolo Bonzini | 398205b | 2014-11-07 23:03:59 +0100 | [diff] [blame] | 1146 | return; |
| 1147 | } |
Jens Axboe | e4043dc | 2014-04-09 10:18:23 -0600 | [diff] [blame] | 1148 | |
Paolo Bonzini | 2a90d4a | 2014-11-07 23:04:00 +0100 | [diff] [blame] | 1149 | put_cpu(); |
Jens Axboe | e4043dc | 2014-04-09 10:18:23 -0600 | [diff] [blame] | 1150 | } |
Paolo Bonzini | 398205b | 2014-11-07 23:03:59 +0100 | [diff] [blame] | 1151 | |
Jens Axboe | 9f99373 | 2017-04-10 09:54:54 -0600 | [diff] [blame] | 1152 | kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx), |
| 1153 | &hctx->run_work, |
| 1154 | msecs_to_jiffies(msecs)); |
Bart Van Assche | 7587a5a | 2017-04-07 11:16:52 -0700 | [diff] [blame] | 1155 | } |
| 1156 | |
| 1157 | void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) |
| 1158 | { |
| 1159 | __blk_mq_delay_run_hw_queue(hctx, true, msecs); |
| 1160 | } |
| 1161 | EXPORT_SYMBOL(blk_mq_delay_run_hw_queue); |
| 1162 | |
| 1163 | void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
| 1164 | { |
| 1165 | __blk_mq_delay_run_hw_queue(hctx, async, 0); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1166 | } |
Omar Sandoval | 5b72727 | 2017-04-14 01:00:00 -0700 | [diff] [blame] | 1167 | EXPORT_SYMBOL(blk_mq_run_hw_queue); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1168 | |
Mike Snitzer | b94ec29 | 2015-03-11 23:56:38 -0400 | [diff] [blame] | 1169 | void blk_mq_run_hw_queues(struct request_queue *q, bool async) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1170 | { |
| 1171 | struct blk_mq_hw_ctx *hctx; |
| 1172 | int i; |
| 1173 | |
| 1174 | queue_for_each_hw_ctx(q, hctx, i) { |
Jens Axboe | bd166ef | 2017-01-17 06:03:22 -0700 | [diff] [blame] | 1175 | if (!blk_mq_hctx_has_pending(hctx) || |
Bart Van Assche | 5d1b25c | 2016-10-28 17:19:15 -0700 | [diff] [blame] | 1176 | blk_mq_hctx_stopped(hctx)) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1177 | continue; |
| 1178 | |
Mike Snitzer | b94ec29 | 2015-03-11 23:56:38 -0400 | [diff] [blame] | 1179 | blk_mq_run_hw_queue(hctx, async); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1180 | } |
| 1181 | } |
Mike Snitzer | b94ec29 | 2015-03-11 23:56:38 -0400 | [diff] [blame] | 1182 | EXPORT_SYMBOL(blk_mq_run_hw_queues); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1183 | |
Bart Van Assche | fd00144 | 2016-10-28 17:19:37 -0700 | [diff] [blame] | 1184 | /** |
| 1185 | * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped |
| 1186 | * @q: request queue. |
| 1187 | * |
| 1188 | * The caller is responsible for serializing this function against |
| 1189 | * blk_mq_{start,stop}_hw_queue(). |
| 1190 | */ |
| 1191 | bool blk_mq_queue_stopped(struct request_queue *q) |
| 1192 | { |
| 1193 | struct blk_mq_hw_ctx *hctx; |
| 1194 | int i; |
| 1195 | |
| 1196 | queue_for_each_hw_ctx(q, hctx, i) |
| 1197 | if (blk_mq_hctx_stopped(hctx)) |
| 1198 | return true; |
| 1199 | |
| 1200 | return false; |
| 1201 | } |
| 1202 | EXPORT_SYMBOL(blk_mq_queue_stopped); |
| 1203 | |
Jens Axboe | 2719aa2 | 2017-05-03 11:08:14 -0600 | [diff] [blame] | 1204 | static void __blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx, bool sync) |
| 1205 | { |
| 1206 | if (sync) |
| 1207 | cancel_delayed_work_sync(&hctx->run_work); |
| 1208 | else |
| 1209 | cancel_delayed_work(&hctx->run_work); |
| 1210 | |
| 1211 | set_bit(BLK_MQ_S_STOPPED, &hctx->state); |
| 1212 | } |
| 1213 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1214 | void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) |
| 1215 | { |
Jens Axboe | 2719aa2 | 2017-05-03 11:08:14 -0600 | [diff] [blame] | 1216 | __blk_mq_stop_hw_queue(hctx, false); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1217 | } |
| 1218 | EXPORT_SYMBOL(blk_mq_stop_hw_queue); |
| 1219 | |
Colin Ian King | ebd7685 | 2017-05-08 11:44:40 +0100 | [diff] [blame] | 1220 | static void __blk_mq_stop_hw_queues(struct request_queue *q, bool sync) |
Christoph Hellwig | 280d45f | 2013-10-25 14:45:58 +0100 | [diff] [blame] | 1221 | { |
| 1222 | struct blk_mq_hw_ctx *hctx; |
| 1223 | int i; |
| 1224 | |
| 1225 | queue_for_each_hw_ctx(q, hctx, i) |
Jens Axboe | 2719aa2 | 2017-05-03 11:08:14 -0600 | [diff] [blame] | 1226 | __blk_mq_stop_hw_queue(hctx, sync); |
| 1227 | } |
| 1228 | |
| 1229 | void blk_mq_stop_hw_queues(struct request_queue *q) |
| 1230 | { |
| 1231 | __blk_mq_stop_hw_queues(q, false); |
Christoph Hellwig | 280d45f | 2013-10-25 14:45:58 +0100 | [diff] [blame] | 1232 | } |
| 1233 | EXPORT_SYMBOL(blk_mq_stop_hw_queues); |
| 1234 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1235 | void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) |
| 1236 | { |
| 1237 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); |
Jens Axboe | e4043dc | 2014-04-09 10:18:23 -0600 | [diff] [blame] | 1238 | |
Jens Axboe | 0ffbce8 | 2014-06-25 08:22:34 -0600 | [diff] [blame] | 1239 | blk_mq_run_hw_queue(hctx, false); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1240 | } |
| 1241 | EXPORT_SYMBOL(blk_mq_start_hw_queue); |
| 1242 | |
Christoph Hellwig | 2f26855 | 2014-04-16 09:44:56 +0200 | [diff] [blame] | 1243 | void blk_mq_start_hw_queues(struct request_queue *q) |
| 1244 | { |
| 1245 | struct blk_mq_hw_ctx *hctx; |
| 1246 | int i; |
| 1247 | |
| 1248 | queue_for_each_hw_ctx(q, hctx, i) |
| 1249 | blk_mq_start_hw_queue(hctx); |
| 1250 | } |
| 1251 | EXPORT_SYMBOL(blk_mq_start_hw_queues); |
| 1252 | |
Jens Axboe | ae911c5 | 2016-12-08 13:19:30 -0700 | [diff] [blame] | 1253 | void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
| 1254 | { |
| 1255 | if (!blk_mq_hctx_stopped(hctx)) |
| 1256 | return; |
| 1257 | |
| 1258 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); |
| 1259 | blk_mq_run_hw_queue(hctx, async); |
| 1260 | } |
| 1261 | EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue); |
| 1262 | |
Christoph Hellwig | 1b4a325 | 2014-04-16 09:44:54 +0200 | [diff] [blame] | 1263 | 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] | 1264 | { |
| 1265 | struct blk_mq_hw_ctx *hctx; |
| 1266 | int i; |
| 1267 | |
Jens Axboe | ae911c5 | 2016-12-08 13:19:30 -0700 | [diff] [blame] | 1268 | queue_for_each_hw_ctx(q, hctx, i) |
| 1269 | blk_mq_start_stopped_hw_queue(hctx, async); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1270 | } |
| 1271 | EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); |
| 1272 | |
Christoph Hellwig | 70f4db6 | 2014-04-16 10:48:08 -0600 | [diff] [blame] | 1273 | static void blk_mq_run_work_fn(struct work_struct *work) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1274 | { |
| 1275 | struct blk_mq_hw_ctx *hctx; |
| 1276 | |
Jens Axboe | 9f99373 | 2017-04-10 09:54:54 -0600 | [diff] [blame] | 1277 | hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work); |
Jens Axboe | 21c6e93 | 2017-04-10 09:54:56 -0600 | [diff] [blame] | 1278 | |
| 1279 | /* |
| 1280 | * If we are stopped, don't run the queue. The exception is if |
| 1281 | * BLK_MQ_S_START_ON_RUN is set. For that case, we auto-clear |
| 1282 | * the STOPPED bit and run it. |
| 1283 | */ |
| 1284 | if (test_bit(BLK_MQ_S_STOPPED, &hctx->state)) { |
| 1285 | if (!test_bit(BLK_MQ_S_START_ON_RUN, &hctx->state)) |
| 1286 | return; |
| 1287 | |
| 1288 | clear_bit(BLK_MQ_S_START_ON_RUN, &hctx->state); |
| 1289 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); |
| 1290 | } |
Jens Axboe | e4043dc | 2014-04-09 10:18:23 -0600 | [diff] [blame] | 1291 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1292 | __blk_mq_run_hw_queue(hctx); |
| 1293 | } |
| 1294 | |
Christoph Hellwig | 70f4db6 | 2014-04-16 10:48:08 -0600 | [diff] [blame] | 1295 | |
| 1296 | void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) |
| 1297 | { |
Ming Lei | 19c66e5 | 2014-12-03 19:38:04 +0800 | [diff] [blame] | 1298 | if (unlikely(!blk_mq_hw_queue_mapped(hctx))) |
| 1299 | return; |
Christoph Hellwig | 70f4db6 | 2014-04-16 10:48:08 -0600 | [diff] [blame] | 1300 | |
Jens Axboe | 21c6e93 | 2017-04-10 09:54:56 -0600 | [diff] [blame] | 1301 | /* |
| 1302 | * Stop the hw queue, then modify currently delayed work. |
| 1303 | * This should prevent us from running the queue prematurely. |
| 1304 | * Mark the queue as auto-clearing STOPPED when it runs. |
| 1305 | */ |
Jens Axboe | 7e79dad | 2017-01-19 07:58:59 -0700 | [diff] [blame] | 1306 | blk_mq_stop_hw_queue(hctx); |
Jens Axboe | 21c6e93 | 2017-04-10 09:54:56 -0600 | [diff] [blame] | 1307 | set_bit(BLK_MQ_S_START_ON_RUN, &hctx->state); |
| 1308 | kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), |
| 1309 | &hctx->run_work, |
| 1310 | msecs_to_jiffies(msecs)); |
Christoph Hellwig | 70f4db6 | 2014-04-16 10:48:08 -0600 | [diff] [blame] | 1311 | } |
| 1312 | EXPORT_SYMBOL(blk_mq_delay_queue); |
| 1313 | |
Ming Lei | cfd0c55 | 2015-10-20 23:13:57 +0800 | [diff] [blame] | 1314 | 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] | 1315 | struct request *rq, |
| 1316 | bool at_head) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1317 | { |
Jens Axboe | e57690f | 2016-08-24 15:34:35 -0600 | [diff] [blame] | 1318 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
| 1319 | |
Jens Axboe | 01b983c | 2013-11-19 18:59:10 -0700 | [diff] [blame] | 1320 | trace_block_rq_insert(hctx->queue, rq); |
| 1321 | |
Christoph Hellwig | 72a0a36 | 2014-02-07 10:22:36 -0800 | [diff] [blame] | 1322 | if (at_head) |
| 1323 | list_add(&rq->queuelist, &ctx->rq_list); |
| 1324 | else |
| 1325 | list_add_tail(&rq->queuelist, &ctx->rq_list); |
Ming Lei | cfd0c55 | 2015-10-20 23:13:57 +0800 | [diff] [blame] | 1326 | } |
Jens Axboe | 4bb659b | 2014-05-09 09:36:49 -0600 | [diff] [blame] | 1327 | |
Jens Axboe | 2c3ad66 | 2016-12-14 14:34:47 -0700 | [diff] [blame] | 1328 | void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq, |
| 1329 | bool at_head) |
Ming Lei | cfd0c55 | 2015-10-20 23:13:57 +0800 | [diff] [blame] | 1330 | { |
| 1331 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
| 1332 | |
Jens Axboe | e57690f | 2016-08-24 15:34:35 -0600 | [diff] [blame] | 1333 | __blk_mq_insert_req_list(hctx, rq, at_head); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1334 | blk_mq_hctx_mark_pending(hctx, ctx); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1335 | } |
| 1336 | |
Jens Axboe | bd166ef | 2017-01-17 06:03:22 -0700 | [diff] [blame] | 1337 | void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx, |
| 1338 | struct list_head *list) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1339 | |
| 1340 | { |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1341 | /* |
| 1342 | * preemption doesn't flush plug list, so it's possible ctx->cpu is |
| 1343 | * offline now |
| 1344 | */ |
| 1345 | spin_lock(&ctx->lock); |
| 1346 | while (!list_empty(list)) { |
| 1347 | struct request *rq; |
| 1348 | |
| 1349 | rq = list_first_entry(list, struct request, queuelist); |
Jens Axboe | e57690f | 2016-08-24 15:34:35 -0600 | [diff] [blame] | 1350 | BUG_ON(rq->mq_ctx != ctx); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1351 | list_del_init(&rq->queuelist); |
Jens Axboe | e57690f | 2016-08-24 15:34:35 -0600 | [diff] [blame] | 1352 | __blk_mq_insert_req_list(hctx, rq, false); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1353 | } |
Ming Lei | cfd0c55 | 2015-10-20 23:13:57 +0800 | [diff] [blame] | 1354 | blk_mq_hctx_mark_pending(hctx, ctx); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1355 | spin_unlock(&ctx->lock); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1356 | } |
| 1357 | |
| 1358 | static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b) |
| 1359 | { |
| 1360 | struct request *rqa = container_of(a, struct request, queuelist); |
| 1361 | struct request *rqb = container_of(b, struct request, queuelist); |
| 1362 | |
| 1363 | return !(rqa->mq_ctx < rqb->mq_ctx || |
| 1364 | (rqa->mq_ctx == rqb->mq_ctx && |
| 1365 | blk_rq_pos(rqa) < blk_rq_pos(rqb))); |
| 1366 | } |
| 1367 | |
| 1368 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) |
| 1369 | { |
| 1370 | struct blk_mq_ctx *this_ctx; |
| 1371 | struct request_queue *this_q; |
| 1372 | struct request *rq; |
| 1373 | LIST_HEAD(list); |
| 1374 | LIST_HEAD(ctx_list); |
| 1375 | unsigned int depth; |
| 1376 | |
| 1377 | list_splice_init(&plug->mq_list, &list); |
| 1378 | |
| 1379 | list_sort(NULL, &list, plug_ctx_cmp); |
| 1380 | |
| 1381 | this_q = NULL; |
| 1382 | this_ctx = NULL; |
| 1383 | depth = 0; |
| 1384 | |
| 1385 | while (!list_empty(&list)) { |
| 1386 | rq = list_entry_rq(list.next); |
| 1387 | list_del_init(&rq->queuelist); |
| 1388 | BUG_ON(!rq->q); |
| 1389 | if (rq->mq_ctx != this_ctx) { |
| 1390 | if (this_ctx) { |
Jens Axboe | bd166ef | 2017-01-17 06:03:22 -0700 | [diff] [blame] | 1391 | trace_block_unplug(this_q, depth, from_schedule); |
| 1392 | blk_mq_sched_insert_requests(this_q, this_ctx, |
| 1393 | &ctx_list, |
| 1394 | from_schedule); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1395 | } |
| 1396 | |
| 1397 | this_ctx = rq->mq_ctx; |
| 1398 | this_q = rq->q; |
| 1399 | depth = 0; |
| 1400 | } |
| 1401 | |
| 1402 | depth++; |
| 1403 | list_add_tail(&rq->queuelist, &ctx_list); |
| 1404 | } |
| 1405 | |
| 1406 | /* |
| 1407 | * If 'this_ctx' is set, we know we have entries to complete |
| 1408 | * on 'ctx_list'. Do those. |
| 1409 | */ |
| 1410 | if (this_ctx) { |
Jens Axboe | bd166ef | 2017-01-17 06:03:22 -0700 | [diff] [blame] | 1411 | trace_block_unplug(this_q, depth, from_schedule); |
| 1412 | blk_mq_sched_insert_requests(this_q, this_ctx, &ctx_list, |
| 1413 | from_schedule); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1414 | } |
| 1415 | } |
| 1416 | |
| 1417 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio) |
| 1418 | { |
Bart Van Assche | da8d7f0 | 2017-04-19 14:01:24 -0700 | [diff] [blame] | 1419 | blk_init_request_from_bio(rq, bio); |
Jens Axboe | 4b57052 | 2014-05-29 11:00:11 -0600 | [diff] [blame] | 1420 | |
Jens Axboe | 6e85eaf | 2016-12-02 20:00:14 -0700 | [diff] [blame] | 1421 | blk_account_io_start(rq, true); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1422 | } |
| 1423 | |
Jens Axboe | 274a584 | 2014-08-15 12:44:08 -0600 | [diff] [blame] | 1424 | static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx) |
| 1425 | { |
| 1426 | return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) && |
| 1427 | !blk_queue_nomerges(hctx->queue); |
| 1428 | } |
| 1429 | |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1430 | static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx, |
| 1431 | struct blk_mq_ctx *ctx, |
| 1432 | struct request *rq, struct bio *bio) |
| 1433 | { |
Ming Lei | e18378a | 2015-10-20 23:13:54 +0800 | [diff] [blame] | 1434 | if (!hctx_allow_merges(hctx) || !bio_mergeable(bio)) { |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1435 | blk_mq_bio_to_request(rq, bio); |
| 1436 | spin_lock(&ctx->lock); |
| 1437 | insert_rq: |
| 1438 | __blk_mq_insert_request(hctx, rq, false); |
| 1439 | spin_unlock(&ctx->lock); |
| 1440 | return false; |
| 1441 | } else { |
Jens Axboe | 274a584 | 2014-08-15 12:44:08 -0600 | [diff] [blame] | 1442 | struct request_queue *q = hctx->queue; |
| 1443 | |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1444 | spin_lock(&ctx->lock); |
| 1445 | if (!blk_mq_attempt_merge(q, ctx, bio)) { |
| 1446 | blk_mq_bio_to_request(rq, bio); |
| 1447 | goto insert_rq; |
| 1448 | } |
| 1449 | |
| 1450 | spin_unlock(&ctx->lock); |
Jens Axboe | bd166ef | 2017-01-17 06:03:22 -0700 | [diff] [blame] | 1451 | __blk_mq_finish_request(hctx, ctx, rq); |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1452 | return true; |
| 1453 | } |
| 1454 | } |
| 1455 | |
Jens Axboe | fd2d332 | 2017-01-12 10:04:45 -0700 | [diff] [blame] | 1456 | static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq) |
| 1457 | { |
Jens Axboe | bd166ef | 2017-01-17 06:03:22 -0700 | [diff] [blame] | 1458 | if (rq->tag != -1) |
| 1459 | return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false); |
| 1460 | |
| 1461 | return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true); |
Jens Axboe | fd2d332 | 2017-01-12 10:04:45 -0700 | [diff] [blame] | 1462 | } |
| 1463 | |
Ming Lei | d964f04 | 2017-06-06 23:22:00 +0800 | [diff] [blame] | 1464 | static void __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, |
| 1465 | struct request *rq, |
| 1466 | blk_qc_t *cookie, bool may_sleep) |
Shaohua Li | f984df1 | 2015-05-08 10:51:32 -0700 | [diff] [blame] | 1467 | { |
Shaohua Li | f984df1 | 2015-05-08 10:51:32 -0700 | [diff] [blame] | 1468 | struct request_queue *q = rq->q; |
Shaohua Li | f984df1 | 2015-05-08 10:51:32 -0700 | [diff] [blame] | 1469 | struct blk_mq_queue_data bd = { |
| 1470 | .rq = rq, |
Omar Sandoval | d945a36 | 2017-04-05 12:01:36 -0700 | [diff] [blame] | 1471 | .last = true, |
Shaohua Li | f984df1 | 2015-05-08 10:51:32 -0700 | [diff] [blame] | 1472 | }; |
Jens Axboe | bd166ef | 2017-01-17 06:03:22 -0700 | [diff] [blame] | 1473 | blk_qc_t new_cookie; |
| 1474 | int ret; |
Ming Lei | d964f04 | 2017-06-06 23:22:00 +0800 | [diff] [blame] | 1475 | bool run_queue = true; |
| 1476 | |
| 1477 | if (blk_mq_hctx_stopped(hctx)) { |
| 1478 | run_queue = false; |
| 1479 | goto insert; |
| 1480 | } |
Shaohua Li | f984df1 | 2015-05-08 10:51:32 -0700 | [diff] [blame] | 1481 | |
Jens Axboe | bd166ef | 2017-01-17 06:03:22 -0700 | [diff] [blame] | 1482 | if (q->elevator) |
Bart Van Assche | 2253efc | 2016-10-28 17:20:02 -0700 | [diff] [blame] | 1483 | goto insert; |
| 1484 | |
Ming Lei | d964f04 | 2017-06-06 23:22:00 +0800 | [diff] [blame] | 1485 | if (!blk_mq_get_driver_tag(rq, NULL, false)) |
Jens Axboe | bd166ef | 2017-01-17 06:03:22 -0700 | [diff] [blame] | 1486 | goto insert; |
| 1487 | |
| 1488 | new_cookie = request_to_qc_t(hctx, rq); |
| 1489 | |
Shaohua Li | f984df1 | 2015-05-08 10:51:32 -0700 | [diff] [blame] | 1490 | /* |
| 1491 | * For OK queue, we are done. For error, kill it. Any other |
| 1492 | * error (busy), just add it to our list as we previously |
| 1493 | * would have done |
| 1494 | */ |
| 1495 | ret = q->mq_ops->queue_rq(hctx, &bd); |
Jens Axboe | 7b37163 | 2015-11-05 10:41:40 -0700 | [diff] [blame] | 1496 | if (ret == BLK_MQ_RQ_QUEUE_OK) { |
| 1497 | *cookie = new_cookie; |
Bart Van Assche | 2253efc | 2016-10-28 17:20:02 -0700 | [diff] [blame] | 1498 | return; |
Shaohua Li | f984df1 | 2015-05-08 10:51:32 -0700 | [diff] [blame] | 1499 | } |
Jens Axboe | 7b37163 | 2015-11-05 10:41:40 -0700 | [diff] [blame] | 1500 | |
Jens Axboe | 7b37163 | 2015-11-05 10:41:40 -0700 | [diff] [blame] | 1501 | if (ret == BLK_MQ_RQ_QUEUE_ERROR) { |
| 1502 | *cookie = BLK_QC_T_NONE; |
Christoph Hellwig | caf7df1 | 2017-04-20 16:03:16 +0200 | [diff] [blame] | 1503 | blk_mq_end_request(rq, -EIO); |
Bart Van Assche | 2253efc | 2016-10-28 17:20:02 -0700 | [diff] [blame] | 1504 | return; |
Jens Axboe | 7b37163 | 2015-11-05 10:41:40 -0700 | [diff] [blame] | 1505 | } |
| 1506 | |
Josef Bacik | b58e176 | 2017-03-28 16:37:52 -0400 | [diff] [blame] | 1507 | __blk_mq_requeue_request(rq); |
Bart Van Assche | 2253efc | 2016-10-28 17:20:02 -0700 | [diff] [blame] | 1508 | insert: |
Ming Lei | d964f04 | 2017-06-06 23:22:00 +0800 | [diff] [blame] | 1509 | blk_mq_sched_insert_request(rq, false, run_queue, false, may_sleep); |
Shaohua Li | f984df1 | 2015-05-08 10:51:32 -0700 | [diff] [blame] | 1510 | } |
| 1511 | |
Christoph Hellwig | 5eb6126 | 2017-03-22 15:01:51 -0400 | [diff] [blame] | 1512 | static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, |
| 1513 | struct request *rq, blk_qc_t *cookie) |
| 1514 | { |
| 1515 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) { |
| 1516 | rcu_read_lock(); |
Ming Lei | d964f04 | 2017-06-06 23:22:00 +0800 | [diff] [blame] | 1517 | __blk_mq_try_issue_directly(hctx, rq, cookie, false); |
Christoph Hellwig | 5eb6126 | 2017-03-22 15:01:51 -0400 | [diff] [blame] | 1518 | rcu_read_unlock(); |
| 1519 | } else { |
Jens Axboe | bf4907c | 2017-03-30 12:30:39 -0600 | [diff] [blame] | 1520 | unsigned int srcu_idx; |
| 1521 | |
| 1522 | might_sleep(); |
| 1523 | |
| 1524 | srcu_idx = srcu_read_lock(&hctx->queue_rq_srcu); |
Ming Lei | d964f04 | 2017-06-06 23:22:00 +0800 | [diff] [blame] | 1525 | __blk_mq_try_issue_directly(hctx, rq, cookie, true); |
Christoph Hellwig | 5eb6126 | 2017-03-22 15:01:51 -0400 | [diff] [blame] | 1526 | srcu_read_unlock(&hctx->queue_rq_srcu, srcu_idx); |
| 1527 | } |
| 1528 | } |
| 1529 | |
Jens Axboe | dece163 | 2015-11-05 10:41:16 -0700 | [diff] [blame] | 1530 | 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] | 1531 | { |
Christoph Hellwig | ef295ec | 2016-10-28 08:48:16 -0600 | [diff] [blame] | 1532 | const int is_sync = op_is_sync(bio->bi_opf); |
Christoph Hellwig | f73f44e | 2017-01-27 08:30:47 -0700 | [diff] [blame] | 1533 | const int is_flush_fua = op_is_flush(bio->bi_opf); |
Jens Axboe | 5a797e0 | 2017-01-26 12:22:11 -0700 | [diff] [blame] | 1534 | struct blk_mq_alloc_data data = { .flags = 0 }; |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1535 | struct request *rq; |
Christoph Hellwig | 5eb6126 | 2017-03-22 15:01:51 -0400 | [diff] [blame] | 1536 | unsigned int request_count = 0; |
Shaohua Li | f984df1 | 2015-05-08 10:51:32 -0700 | [diff] [blame] | 1537 | struct blk_plug *plug; |
Shaohua Li | 5b3f341 | 2015-05-08 10:51:33 -0700 | [diff] [blame] | 1538 | struct request *same_queue_rq = NULL; |
Jens Axboe | 7b37163 | 2015-11-05 10:41:40 -0700 | [diff] [blame] | 1539 | blk_qc_t cookie; |
Jens Axboe | 87760e5 | 2016-11-09 12:38:14 -0700 | [diff] [blame] | 1540 | unsigned int wb_acct; |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1541 | |
| 1542 | blk_queue_bounce(q, &bio); |
| 1543 | |
Wen Xiong | f36ea50 | 2017-05-10 08:54:11 -0500 | [diff] [blame] | 1544 | blk_queue_split(q, &bio, q->bio_split); |
| 1545 | |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1546 | if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) { |
Christoph Hellwig | 4246a0b | 2015-07-20 15:29:37 +0200 | [diff] [blame] | 1547 | bio_io_error(bio); |
Jens Axboe | dece163 | 2015-11-05 10:41:16 -0700 | [diff] [blame] | 1548 | return BLK_QC_T_NONE; |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1549 | } |
| 1550 | |
Omar Sandoval | 87c279e | 2016-06-01 22:18:48 -0700 | [diff] [blame] | 1551 | if (!is_flush_fua && !blk_queue_nomerges(q) && |
| 1552 | blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq)) |
| 1553 | return BLK_QC_T_NONE; |
Shaohua Li | f984df1 | 2015-05-08 10:51:32 -0700 | [diff] [blame] | 1554 | |
Jens Axboe | bd166ef | 2017-01-17 06:03:22 -0700 | [diff] [blame] | 1555 | if (blk_mq_sched_bio_merge(q, bio)) |
| 1556 | return BLK_QC_T_NONE; |
| 1557 | |
Jens Axboe | 87760e5 | 2016-11-09 12:38:14 -0700 | [diff] [blame] | 1558 | wb_acct = wbt_wait(q->rq_wb, bio, NULL); |
| 1559 | |
Jens Axboe | bd166ef | 2017-01-17 06:03:22 -0700 | [diff] [blame] | 1560 | trace_block_getrq(q, bio, bio->bi_opf); |
| 1561 | |
| 1562 | rq = blk_mq_sched_get_request(q, bio, bio->bi_opf, &data); |
Jens Axboe | 87760e5 | 2016-11-09 12:38:14 -0700 | [diff] [blame] | 1563 | if (unlikely(!rq)) { |
| 1564 | __wbt_done(q->rq_wb, wb_acct); |
Jens Axboe | dece163 | 2015-11-05 10:41:16 -0700 | [diff] [blame] | 1565 | return BLK_QC_T_NONE; |
Jens Axboe | 87760e5 | 2016-11-09 12:38:14 -0700 | [diff] [blame] | 1566 | } |
| 1567 | |
| 1568 | wbt_track(&rq->issue_stat, wb_acct); |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1569 | |
Jens Axboe | fd2d332 | 2017-01-12 10:04:45 -0700 | [diff] [blame] | 1570 | cookie = request_to_qc_t(data.hctx, rq); |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 1571 | |
Shaohua Li | f984df1 | 2015-05-08 10:51:32 -0700 | [diff] [blame] | 1572 | plug = current->plug; |
Christoph Hellwig | a4d907b | 2017-03-22 15:01:53 -0400 | [diff] [blame] | 1573 | if (unlikely(is_flush_fua)) { |
Shaohua Li | f984df1 | 2015-05-08 10:51:32 -0700 | [diff] [blame] | 1574 | blk_mq_put_ctx(data.ctx); |
Christoph Hellwig | a4d907b | 2017-03-22 15:01:53 -0400 | [diff] [blame] | 1575 | blk_mq_bio_to_request(rq, bio); |
| 1576 | if (q->elevator) { |
| 1577 | blk_mq_sched_insert_request(rq, false, true, true, |
| 1578 | true); |
Bart Van Assche | 6a83e74 | 2016-11-02 10:09:51 -0600 | [diff] [blame] | 1579 | } else { |
Christoph Hellwig | a4d907b | 2017-03-22 15:01:53 -0400 | [diff] [blame] | 1580 | blk_insert_flush(rq); |
| 1581 | blk_mq_run_hw_queue(data.hctx, true); |
Bart Van Assche | 6a83e74 | 2016-11-02 10:09:51 -0600 | [diff] [blame] | 1582 | } |
Christoph Hellwig | a4d907b | 2017-03-22 15:01:53 -0400 | [diff] [blame] | 1583 | } else if (plug && q->nr_hw_queues == 1) { |
Shaohua Li | 600271d | 2016-11-03 17:03:54 -0700 | [diff] [blame] | 1584 | struct request *last = NULL; |
| 1585 | |
Jens Axboe | b00c53e | 2017-04-20 16:40:36 -0600 | [diff] [blame] | 1586 | blk_mq_put_ctx(data.ctx); |
Jeff Moyer | e6c4438 | 2015-05-08 10:51:30 -0700 | [diff] [blame] | 1587 | blk_mq_bio_to_request(rq, bio); |
Ming Lei | 0a6219a | 2016-11-16 18:07:05 +0800 | [diff] [blame] | 1588 | |
| 1589 | /* |
| 1590 | * @request_count may become stale because of schedule |
| 1591 | * out, so check the list again. |
| 1592 | */ |
| 1593 | if (list_empty(&plug->mq_list)) |
| 1594 | request_count = 0; |
Christoph Hellwig | 254d259 | 2017-03-22 15:01:50 -0400 | [diff] [blame] | 1595 | else if (blk_queue_nomerges(q)) |
| 1596 | request_count = blk_plug_queued_count(q); |
| 1597 | |
Ming Lei | 676d060 | 2015-10-20 23:13:56 +0800 | [diff] [blame] | 1598 | if (!request_count) |
Jeff Moyer | e6c4438 | 2015-05-08 10:51:30 -0700 | [diff] [blame] | 1599 | trace_block_plug(q); |
Shaohua Li | 600271d | 2016-11-03 17:03:54 -0700 | [diff] [blame] | 1600 | else |
| 1601 | last = list_entry_rq(plug->mq_list.prev); |
Jens Axboe | b094f89 | 2015-11-20 20:29:45 -0700 | [diff] [blame] | 1602 | |
Shaohua Li | 600271d | 2016-11-03 17:03:54 -0700 | [diff] [blame] | 1603 | if (request_count >= BLK_MAX_REQUEST_COUNT || (last && |
| 1604 | blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) { |
Jeff Moyer | e6c4438 | 2015-05-08 10:51:30 -0700 | [diff] [blame] | 1605 | blk_flush_plug_list(plug, false); |
| 1606 | trace_block_plug(q); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1607 | } |
Jens Axboe | b094f89 | 2015-11-20 20:29:45 -0700 | [diff] [blame] | 1608 | |
Jeff Moyer | e6c4438 | 2015-05-08 10:51:30 -0700 | [diff] [blame] | 1609 | list_add_tail(&rq->queuelist, &plug->mq_list); |
Christoph Hellwig | 2299722 | 2017-03-22 15:01:52 -0400 | [diff] [blame] | 1610 | } else if (plug && !blk_queue_nomerges(q)) { |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1611 | blk_mq_bio_to_request(rq, bio); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1612 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1613 | /* |
| 1614 | * We do limited plugging. If the bio can be merged, do that. |
| 1615 | * Otherwise the existing request in the plug list will be |
| 1616 | * issued. So the plug list will have one request at most |
Christoph Hellwig | 2299722 | 2017-03-22 15:01:52 -0400 | [diff] [blame] | 1617 | * The plug list might get flushed before this. If that happens, |
| 1618 | * the plug list is empty, and same_queue_rq is invalid. |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1619 | */ |
Christoph Hellwig | 2299722 | 2017-03-22 15:01:52 -0400 | [diff] [blame] | 1620 | if (list_empty(&plug->mq_list)) |
| 1621 | same_queue_rq = NULL; |
| 1622 | if (same_queue_rq) |
| 1623 | list_del_init(&same_queue_rq->queuelist); |
| 1624 | list_add_tail(&rq->queuelist, &plug->mq_list); |
| 1625 | |
Jens Axboe | bf4907c | 2017-03-30 12:30:39 -0600 | [diff] [blame] | 1626 | blk_mq_put_ctx(data.ctx); |
| 1627 | |
Ming Lei | dad7a3b | 2017-06-06 23:21:59 +0800 | [diff] [blame] | 1628 | if (same_queue_rq) { |
| 1629 | data.hctx = blk_mq_map_queue(q, |
| 1630 | same_queue_rq->mq_ctx->cpu); |
Christoph Hellwig | 2299722 | 2017-03-22 15:01:52 -0400 | [diff] [blame] | 1631 | blk_mq_try_issue_directly(data.hctx, same_queue_rq, |
| 1632 | &cookie); |
Ming Lei | dad7a3b | 2017-06-06 23:21:59 +0800 | [diff] [blame] | 1633 | } |
Christoph Hellwig | a4d907b | 2017-03-22 15:01:53 -0400 | [diff] [blame] | 1634 | } else if (q->nr_hw_queues > 1 && is_sync) { |
Jens Axboe | bd166ef | 2017-01-17 06:03:22 -0700 | [diff] [blame] | 1635 | blk_mq_put_ctx(data.ctx); |
| 1636 | blk_mq_bio_to_request(rq, bio); |
Christoph Hellwig | 2299722 | 2017-03-22 15:01:52 -0400 | [diff] [blame] | 1637 | blk_mq_try_issue_directly(data.hctx, rq, &cookie); |
Christoph Hellwig | a4d907b | 2017-03-22 15:01:53 -0400 | [diff] [blame] | 1638 | } else if (q->elevator) { |
Jens Axboe | b00c53e | 2017-04-20 16:40:36 -0600 | [diff] [blame] | 1639 | blk_mq_put_ctx(data.ctx); |
Christoph Hellwig | a4d907b | 2017-03-22 15:01:53 -0400 | [diff] [blame] | 1640 | blk_mq_bio_to_request(rq, bio); |
| 1641 | blk_mq_sched_insert_request(rq, false, true, true, true); |
Jens Axboe | b00c53e | 2017-04-20 16:40:36 -0600 | [diff] [blame] | 1642 | } else if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) { |
| 1643 | blk_mq_put_ctx(data.ctx); |
Christoph Hellwig | a4d907b | 2017-03-22 15:01:53 -0400 | [diff] [blame] | 1644 | blk_mq_run_hw_queue(data.hctx, true); |
Bart Van Assche | abc25a6 | 2017-04-21 12:00:40 -0600 | [diff] [blame] | 1645 | } else |
| 1646 | blk_mq_put_ctx(data.ctx); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1647 | |
Jens Axboe | 7b37163 | 2015-11-05 10:41:40 -0700 | [diff] [blame] | 1648 | return cookie; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1649 | } |
| 1650 | |
Jens Axboe | cc71a6f | 2017-01-11 14:29:56 -0700 | [diff] [blame] | 1651 | void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, |
| 1652 | unsigned int hctx_idx) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1653 | { |
| 1654 | struct page *page; |
| 1655 | |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1656 | if (tags->rqs && set->ops->exit_request) { |
Christoph Hellwig | e9b267d | 2014-04-15 13:59:10 -0600 | [diff] [blame] | 1657 | int i; |
| 1658 | |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1659 | for (i = 0; i < tags->nr_tags; i++) { |
Jens Axboe | 2af8cbe | 2017-01-13 14:39:30 -0700 | [diff] [blame] | 1660 | struct request *rq = tags->static_rqs[i]; |
| 1661 | |
| 1662 | if (!rq) |
Christoph Hellwig | e9b267d | 2014-04-15 13:59:10 -0600 | [diff] [blame] | 1663 | continue; |
Christoph Hellwig | d6296d3 | 2017-05-01 10:19:08 -0600 | [diff] [blame] | 1664 | set->ops->exit_request(set, rq, hctx_idx); |
Jens Axboe | 2af8cbe | 2017-01-13 14:39:30 -0700 | [diff] [blame] | 1665 | tags->static_rqs[i] = NULL; |
Christoph Hellwig | e9b267d | 2014-04-15 13:59:10 -0600 | [diff] [blame] | 1666 | } |
| 1667 | } |
| 1668 | |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1669 | while (!list_empty(&tags->page_list)) { |
| 1670 | page = list_first_entry(&tags->page_list, struct page, lru); |
Dave Hansen | 6753471 | 2014-01-08 20:17:46 -0700 | [diff] [blame] | 1671 | list_del_init(&page->lru); |
Catalin Marinas | f75782e | 2015-09-14 18:16:02 +0100 | [diff] [blame] | 1672 | /* |
| 1673 | * Remove kmemleak object previously allocated in |
| 1674 | * blk_mq_init_rq_map(). |
| 1675 | */ |
| 1676 | kmemleak_free(page_address(page)); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1677 | __free_pages(page, page->private); |
| 1678 | } |
Jens Axboe | cc71a6f | 2017-01-11 14:29:56 -0700 | [diff] [blame] | 1679 | } |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1680 | |
Jens Axboe | cc71a6f | 2017-01-11 14:29:56 -0700 | [diff] [blame] | 1681 | void blk_mq_free_rq_map(struct blk_mq_tags *tags) |
| 1682 | { |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1683 | kfree(tags->rqs); |
Jens Axboe | cc71a6f | 2017-01-11 14:29:56 -0700 | [diff] [blame] | 1684 | tags->rqs = NULL; |
Jens Axboe | 2af8cbe | 2017-01-13 14:39:30 -0700 | [diff] [blame] | 1685 | kfree(tags->static_rqs); |
| 1686 | tags->static_rqs = NULL; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1687 | |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1688 | blk_mq_free_tags(tags); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1689 | } |
| 1690 | |
Jens Axboe | cc71a6f | 2017-01-11 14:29:56 -0700 | [diff] [blame] | 1691 | struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, |
| 1692 | unsigned int hctx_idx, |
| 1693 | unsigned int nr_tags, |
| 1694 | unsigned int reserved_tags) |
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 | struct blk_mq_tags *tags; |
Shaohua Li | 59f082e | 2017-02-01 09:53:14 -0800 | [diff] [blame] | 1697 | int node; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1698 | |
Shaohua Li | 59f082e | 2017-02-01 09:53:14 -0800 | [diff] [blame] | 1699 | node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx); |
| 1700 | if (node == NUMA_NO_NODE) |
| 1701 | node = set->numa_node; |
| 1702 | |
| 1703 | tags = blk_mq_init_tags(nr_tags, reserved_tags, node, |
Shaohua Li | 24391c0 | 2015-01-23 14:18:00 -0700 | [diff] [blame] | 1704 | BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags)); |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1705 | if (!tags) |
| 1706 | return NULL; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1707 | |
Jens Axboe | cc71a6f | 2017-01-11 14:29:56 -0700 | [diff] [blame] | 1708 | tags->rqs = kzalloc_node(nr_tags * sizeof(struct request *), |
Gabriel Krisman Bertazi | 36e1f3d1 | 2016-12-06 13:31:44 -0200 | [diff] [blame] | 1709 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, |
Shaohua Li | 59f082e | 2017-02-01 09:53:14 -0800 | [diff] [blame] | 1710 | node); |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1711 | if (!tags->rqs) { |
| 1712 | blk_mq_free_tags(tags); |
| 1713 | return NULL; |
| 1714 | } |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1715 | |
Jens Axboe | 2af8cbe | 2017-01-13 14:39:30 -0700 | [diff] [blame] | 1716 | tags->static_rqs = kzalloc_node(nr_tags * sizeof(struct request *), |
| 1717 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, |
Shaohua Li | 59f082e | 2017-02-01 09:53:14 -0800 | [diff] [blame] | 1718 | node); |
Jens Axboe | 2af8cbe | 2017-01-13 14:39:30 -0700 | [diff] [blame] | 1719 | if (!tags->static_rqs) { |
| 1720 | kfree(tags->rqs); |
| 1721 | blk_mq_free_tags(tags); |
| 1722 | return NULL; |
| 1723 | } |
| 1724 | |
Jens Axboe | cc71a6f | 2017-01-11 14:29:56 -0700 | [diff] [blame] | 1725 | return tags; |
| 1726 | } |
| 1727 | |
| 1728 | static size_t order_to_size(unsigned int order) |
| 1729 | { |
| 1730 | return (size_t)PAGE_SIZE << order; |
| 1731 | } |
| 1732 | |
| 1733 | int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, |
| 1734 | unsigned int hctx_idx, unsigned int depth) |
| 1735 | { |
| 1736 | unsigned int i, j, entries_per_page, max_order = 4; |
| 1737 | size_t rq_size, left; |
Shaohua Li | 59f082e | 2017-02-01 09:53:14 -0800 | [diff] [blame] | 1738 | int node; |
| 1739 | |
| 1740 | node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx); |
| 1741 | if (node == NUMA_NO_NODE) |
| 1742 | node = set->numa_node; |
Jens Axboe | cc71a6f | 2017-01-11 14:29:56 -0700 | [diff] [blame] | 1743 | |
| 1744 | INIT_LIST_HEAD(&tags->page_list); |
| 1745 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1746 | /* |
| 1747 | * rq_size is the size of the request plus driver payload, rounded |
| 1748 | * to the cacheline size |
| 1749 | */ |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1750 | rq_size = round_up(sizeof(struct request) + set->cmd_size, |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1751 | cache_line_size()); |
Jens Axboe | cc71a6f | 2017-01-11 14:29:56 -0700 | [diff] [blame] | 1752 | left = rq_size * depth; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1753 | |
Jens Axboe | cc71a6f | 2017-01-11 14:29:56 -0700 | [diff] [blame] | 1754 | for (i = 0; i < depth; ) { |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1755 | int this_order = max_order; |
| 1756 | struct page *page; |
| 1757 | int to_do; |
| 1758 | void *p; |
| 1759 | |
Bartlomiej Zolnierkiewicz | b3a834b | 2016-05-16 09:54:47 -0600 | [diff] [blame] | 1760 | while (this_order && left < order_to_size(this_order - 1)) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1761 | this_order--; |
| 1762 | |
| 1763 | do { |
Shaohua Li | 59f082e | 2017-02-01 09:53:14 -0800 | [diff] [blame] | 1764 | page = alloc_pages_node(node, |
Gabriel Krisman Bertazi | 36e1f3d1 | 2016-12-06 13:31:44 -0200 | [diff] [blame] | 1765 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO, |
Jens Axboe | a516440 | 2014-09-10 09:02:03 -0600 | [diff] [blame] | 1766 | this_order); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1767 | if (page) |
| 1768 | break; |
| 1769 | if (!this_order--) |
| 1770 | break; |
| 1771 | if (order_to_size(this_order) < rq_size) |
| 1772 | break; |
| 1773 | } while (1); |
| 1774 | |
| 1775 | if (!page) |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1776 | goto fail; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1777 | |
| 1778 | page->private = this_order; |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1779 | list_add_tail(&page->lru, &tags->page_list); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1780 | |
| 1781 | p = page_address(page); |
Catalin Marinas | f75782e | 2015-09-14 18:16:02 +0100 | [diff] [blame] | 1782 | /* |
| 1783 | * Allow kmemleak to scan these pages as they contain pointers |
| 1784 | * to additional allocations like via ops->init_request(). |
| 1785 | */ |
Gabriel Krisman Bertazi | 36e1f3d1 | 2016-12-06 13:31:44 -0200 | [diff] [blame] | 1786 | kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1787 | entries_per_page = order_to_size(this_order) / rq_size; |
Jens Axboe | cc71a6f | 2017-01-11 14:29:56 -0700 | [diff] [blame] | 1788 | to_do = min(entries_per_page, depth - i); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1789 | left -= to_do * rq_size; |
| 1790 | for (j = 0; j < to_do; j++) { |
Jens Axboe | 2af8cbe | 2017-01-13 14:39:30 -0700 | [diff] [blame] | 1791 | struct request *rq = p; |
| 1792 | |
| 1793 | tags->static_rqs[i] = rq; |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1794 | if (set->ops->init_request) { |
Christoph Hellwig | d6296d3 | 2017-05-01 10:19:08 -0600 | [diff] [blame] | 1795 | if (set->ops->init_request(set, rq, hctx_idx, |
Shaohua Li | 59f082e | 2017-02-01 09:53:14 -0800 | [diff] [blame] | 1796 | node)) { |
Jens Axboe | 2af8cbe | 2017-01-13 14:39:30 -0700 | [diff] [blame] | 1797 | tags->static_rqs[i] = NULL; |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1798 | goto fail; |
Jens Axboe | a516440 | 2014-09-10 09:02:03 -0600 | [diff] [blame] | 1799 | } |
Christoph Hellwig | e9b267d | 2014-04-15 13:59:10 -0600 | [diff] [blame] | 1800 | } |
| 1801 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1802 | p += rq_size; |
| 1803 | i++; |
| 1804 | } |
| 1805 | } |
Jens Axboe | cc71a6f | 2017-01-11 14:29:56 -0700 | [diff] [blame] | 1806 | return 0; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1807 | |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 1808 | fail: |
Jens Axboe | cc71a6f | 2017-01-11 14:29:56 -0700 | [diff] [blame] | 1809 | blk_mq_free_rqs(set, tags, hctx_idx); |
| 1810 | return -ENOMEM; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1811 | } |
| 1812 | |
Jens Axboe | e57690f | 2016-08-24 15:34:35 -0600 | [diff] [blame] | 1813 | /* |
| 1814 | * 'cpu' is going away. splice any existing rq_list entries from this |
| 1815 | * software queue to the hw queue dispatch list, and ensure that it |
| 1816 | * gets run. |
| 1817 | */ |
Thomas Gleixner | 9467f85 | 2016-09-22 08:05:17 -0600 | [diff] [blame] | 1818 | 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] | 1819 | { |
Thomas Gleixner | 9467f85 | 2016-09-22 08:05:17 -0600 | [diff] [blame] | 1820 | struct blk_mq_hw_ctx *hctx; |
Jens Axboe | 484b406 | 2014-05-21 14:01:15 -0600 | [diff] [blame] | 1821 | struct blk_mq_ctx *ctx; |
| 1822 | LIST_HEAD(tmp); |
| 1823 | |
Thomas Gleixner | 9467f85 | 2016-09-22 08:05:17 -0600 | [diff] [blame] | 1824 | hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead); |
Jens Axboe | e57690f | 2016-08-24 15:34:35 -0600 | [diff] [blame] | 1825 | ctx = __blk_mq_get_ctx(hctx->queue, cpu); |
Jens Axboe | 484b406 | 2014-05-21 14:01:15 -0600 | [diff] [blame] | 1826 | |
| 1827 | spin_lock(&ctx->lock); |
| 1828 | if (!list_empty(&ctx->rq_list)) { |
| 1829 | list_splice_init(&ctx->rq_list, &tmp); |
| 1830 | blk_mq_hctx_clear_pending(hctx, ctx); |
| 1831 | } |
| 1832 | spin_unlock(&ctx->lock); |
| 1833 | |
| 1834 | if (list_empty(&tmp)) |
Thomas Gleixner | 9467f85 | 2016-09-22 08:05:17 -0600 | [diff] [blame] | 1835 | return 0; |
Jens Axboe | 484b406 | 2014-05-21 14:01:15 -0600 | [diff] [blame] | 1836 | |
Jens Axboe | e57690f | 2016-08-24 15:34:35 -0600 | [diff] [blame] | 1837 | spin_lock(&hctx->lock); |
| 1838 | list_splice_tail_init(&tmp, &hctx->dispatch); |
| 1839 | spin_unlock(&hctx->lock); |
Jens Axboe | 484b406 | 2014-05-21 14:01:15 -0600 | [diff] [blame] | 1840 | |
| 1841 | blk_mq_run_hw_queue(hctx, true); |
Thomas Gleixner | 9467f85 | 2016-09-22 08:05:17 -0600 | [diff] [blame] | 1842 | return 0; |
Jens Axboe | 484b406 | 2014-05-21 14:01:15 -0600 | [diff] [blame] | 1843 | } |
| 1844 | |
Thomas Gleixner | 9467f85 | 2016-09-22 08:05:17 -0600 | [diff] [blame] | 1845 | static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx) |
Jens Axboe | 484b406 | 2014-05-21 14:01:15 -0600 | [diff] [blame] | 1846 | { |
Thomas Gleixner | 9467f85 | 2016-09-22 08:05:17 -0600 | [diff] [blame] | 1847 | cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD, |
| 1848 | &hctx->cpuhp_dead); |
Jens Axboe | 484b406 | 2014-05-21 14:01:15 -0600 | [diff] [blame] | 1849 | } |
| 1850 | |
Ming Lei | c3b4afc | 2015-06-04 22:25:04 +0800 | [diff] [blame] | 1851 | /* hctx->ctxs will be freed in queue's release handler */ |
Ming Lei | 08e98fc | 2014-09-25 23:23:38 +0800 | [diff] [blame] | 1852 | static void blk_mq_exit_hctx(struct request_queue *q, |
| 1853 | struct blk_mq_tag_set *set, |
| 1854 | struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) |
| 1855 | { |
Omar Sandoval | 9c1051a | 2017-05-04 08:17:21 -0600 | [diff] [blame] | 1856 | blk_mq_debugfs_unregister_hctx(hctx); |
| 1857 | |
Ming Lei | 08e98fc | 2014-09-25 23:23:38 +0800 | [diff] [blame] | 1858 | blk_mq_tag_idle(hctx); |
| 1859 | |
Ming Lei | f70ced0 | 2014-09-25 23:23:47 +0800 | [diff] [blame] | 1860 | if (set->ops->exit_request) |
Christoph Hellwig | d6296d3 | 2017-05-01 10:19:08 -0600 | [diff] [blame] | 1861 | set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx); |
Ming Lei | f70ced0 | 2014-09-25 23:23:47 +0800 | [diff] [blame] | 1862 | |
Omar Sandoval | 9325263 | 2017-04-05 12:01:31 -0700 | [diff] [blame] | 1863 | blk_mq_sched_exit_hctx(q, hctx, hctx_idx); |
| 1864 | |
Ming Lei | 08e98fc | 2014-09-25 23:23:38 +0800 | [diff] [blame] | 1865 | if (set->ops->exit_hctx) |
| 1866 | set->ops->exit_hctx(hctx, hctx_idx); |
| 1867 | |
Bart Van Assche | 6a83e74 | 2016-11-02 10:09:51 -0600 | [diff] [blame] | 1868 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
| 1869 | cleanup_srcu_struct(&hctx->queue_rq_srcu); |
| 1870 | |
Thomas Gleixner | 9467f85 | 2016-09-22 08:05:17 -0600 | [diff] [blame] | 1871 | blk_mq_remove_cpuhp(hctx); |
Ming Lei | f70ced0 | 2014-09-25 23:23:47 +0800 | [diff] [blame] | 1872 | blk_free_flush_queue(hctx->fq); |
Omar Sandoval | 8845964 | 2016-09-17 08:38:44 -0600 | [diff] [blame] | 1873 | sbitmap_free(&hctx->ctx_map); |
Ming Lei | 08e98fc | 2014-09-25 23:23:38 +0800 | [diff] [blame] | 1874 | } |
| 1875 | |
Ming Lei | 624dbe4 | 2014-05-27 23:35:13 +0800 | [diff] [blame] | 1876 | static void blk_mq_exit_hw_queues(struct request_queue *q, |
| 1877 | struct blk_mq_tag_set *set, int nr_queue) |
| 1878 | { |
| 1879 | struct blk_mq_hw_ctx *hctx; |
| 1880 | unsigned int i; |
| 1881 | |
| 1882 | queue_for_each_hw_ctx(q, hctx, i) { |
| 1883 | if (i == nr_queue) |
| 1884 | break; |
Ming Lei | 08e98fc | 2014-09-25 23:23:38 +0800 | [diff] [blame] | 1885 | blk_mq_exit_hctx(q, set, hctx, i); |
Ming Lei | 624dbe4 | 2014-05-27 23:35:13 +0800 | [diff] [blame] | 1886 | } |
Ming Lei | 624dbe4 | 2014-05-27 23:35:13 +0800 | [diff] [blame] | 1887 | } |
| 1888 | |
Ming Lei | 08e98fc | 2014-09-25 23:23:38 +0800 | [diff] [blame] | 1889 | static int blk_mq_init_hctx(struct request_queue *q, |
| 1890 | struct blk_mq_tag_set *set, |
| 1891 | struct blk_mq_hw_ctx *hctx, unsigned hctx_idx) |
| 1892 | { |
| 1893 | int node; |
| 1894 | |
| 1895 | node = hctx->numa_node; |
| 1896 | if (node == NUMA_NO_NODE) |
| 1897 | node = hctx->numa_node = set->numa_node; |
| 1898 | |
Jens Axboe | 9f99373 | 2017-04-10 09:54:54 -0600 | [diff] [blame] | 1899 | INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn); |
Ming Lei | 08e98fc | 2014-09-25 23:23:38 +0800 | [diff] [blame] | 1900 | spin_lock_init(&hctx->lock); |
| 1901 | INIT_LIST_HEAD(&hctx->dispatch); |
| 1902 | hctx->queue = q; |
| 1903 | hctx->queue_num = hctx_idx; |
Jeff Moyer | 2404e60 | 2015-11-03 10:40:06 -0500 | [diff] [blame] | 1904 | hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED; |
Ming Lei | 08e98fc | 2014-09-25 23:23:38 +0800 | [diff] [blame] | 1905 | |
Thomas Gleixner | 9467f85 | 2016-09-22 08:05:17 -0600 | [diff] [blame] | 1906 | cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead); |
Ming Lei | 08e98fc | 2014-09-25 23:23:38 +0800 | [diff] [blame] | 1907 | |
| 1908 | hctx->tags = set->tags[hctx_idx]; |
| 1909 | |
| 1910 | /* |
| 1911 | * Allocate space for all possible cpus to avoid allocation at |
| 1912 | * runtime |
| 1913 | */ |
| 1914 | hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *), |
| 1915 | GFP_KERNEL, node); |
| 1916 | if (!hctx->ctxs) |
| 1917 | goto unregister_cpu_notifier; |
| 1918 | |
Omar Sandoval | 8845964 | 2016-09-17 08:38:44 -0600 | [diff] [blame] | 1919 | if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL, |
| 1920 | node)) |
Ming Lei | 08e98fc | 2014-09-25 23:23:38 +0800 | [diff] [blame] | 1921 | goto free_ctxs; |
| 1922 | |
| 1923 | hctx->nr_ctx = 0; |
| 1924 | |
| 1925 | if (set->ops->init_hctx && |
| 1926 | set->ops->init_hctx(hctx, set->driver_data, hctx_idx)) |
| 1927 | goto free_bitmap; |
| 1928 | |
Omar Sandoval | 9325263 | 2017-04-05 12:01:31 -0700 | [diff] [blame] | 1929 | if (blk_mq_sched_init_hctx(q, hctx, hctx_idx)) |
| 1930 | goto exit_hctx; |
| 1931 | |
Ming Lei | f70ced0 | 2014-09-25 23:23:47 +0800 | [diff] [blame] | 1932 | hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size); |
| 1933 | if (!hctx->fq) |
Omar Sandoval | 9325263 | 2017-04-05 12:01:31 -0700 | [diff] [blame] | 1934 | goto sched_exit_hctx; |
Ming Lei | f70ced0 | 2014-09-25 23:23:47 +0800 | [diff] [blame] | 1935 | |
| 1936 | if (set->ops->init_request && |
Christoph Hellwig | d6296d3 | 2017-05-01 10:19:08 -0600 | [diff] [blame] | 1937 | set->ops->init_request(set, hctx->fq->flush_rq, hctx_idx, |
| 1938 | node)) |
Ming Lei | f70ced0 | 2014-09-25 23:23:47 +0800 | [diff] [blame] | 1939 | goto free_fq; |
| 1940 | |
Bart Van Assche | 6a83e74 | 2016-11-02 10:09:51 -0600 | [diff] [blame] | 1941 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
| 1942 | init_srcu_struct(&hctx->queue_rq_srcu); |
| 1943 | |
Omar Sandoval | 9c1051a | 2017-05-04 08:17:21 -0600 | [diff] [blame] | 1944 | blk_mq_debugfs_register_hctx(q, hctx); |
| 1945 | |
Ming Lei | 08e98fc | 2014-09-25 23:23:38 +0800 | [diff] [blame] | 1946 | return 0; |
| 1947 | |
Ming Lei | f70ced0 | 2014-09-25 23:23:47 +0800 | [diff] [blame] | 1948 | free_fq: |
| 1949 | kfree(hctx->fq); |
Omar Sandoval | 9325263 | 2017-04-05 12:01:31 -0700 | [diff] [blame] | 1950 | sched_exit_hctx: |
| 1951 | blk_mq_sched_exit_hctx(q, hctx, hctx_idx); |
Ming Lei | f70ced0 | 2014-09-25 23:23:47 +0800 | [diff] [blame] | 1952 | exit_hctx: |
| 1953 | if (set->ops->exit_hctx) |
| 1954 | set->ops->exit_hctx(hctx, hctx_idx); |
Ming Lei | 08e98fc | 2014-09-25 23:23:38 +0800 | [diff] [blame] | 1955 | free_bitmap: |
Omar Sandoval | 8845964 | 2016-09-17 08:38:44 -0600 | [diff] [blame] | 1956 | sbitmap_free(&hctx->ctx_map); |
Ming Lei | 08e98fc | 2014-09-25 23:23:38 +0800 | [diff] [blame] | 1957 | free_ctxs: |
| 1958 | kfree(hctx->ctxs); |
| 1959 | unregister_cpu_notifier: |
Thomas Gleixner | 9467f85 | 2016-09-22 08:05:17 -0600 | [diff] [blame] | 1960 | blk_mq_remove_cpuhp(hctx); |
Ming Lei | 08e98fc | 2014-09-25 23:23:38 +0800 | [diff] [blame] | 1961 | return -1; |
| 1962 | } |
| 1963 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1964 | static void blk_mq_init_cpu_queues(struct request_queue *q, |
| 1965 | unsigned int nr_hw_queues) |
| 1966 | { |
| 1967 | unsigned int i; |
| 1968 | |
| 1969 | for_each_possible_cpu(i) { |
| 1970 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); |
| 1971 | struct blk_mq_hw_ctx *hctx; |
| 1972 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1973 | __ctx->cpu = i; |
| 1974 | spin_lock_init(&__ctx->lock); |
| 1975 | INIT_LIST_HEAD(&__ctx->rq_list); |
| 1976 | __ctx->queue = q; |
| 1977 | |
| 1978 | /* 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] | 1979 | if (!cpu_online(i)) |
| 1980 | continue; |
| 1981 | |
Christoph Hellwig | 7d7e0f9 | 2016-09-14 16:18:54 +0200 | [diff] [blame] | 1982 | hctx = blk_mq_map_queue(q, i); |
Jens Axboe | e4043dc | 2014-04-09 10:18:23 -0600 | [diff] [blame] | 1983 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1984 | /* |
| 1985 | * Set local node, IFF we have more than one hw queue. If |
| 1986 | * not, we remain on the home node of the device |
| 1987 | */ |
| 1988 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) |
Raghavendra K T | bffed45 | 2015-12-02 16:59:05 +0530 | [diff] [blame] | 1989 | hctx->numa_node = local_memory_node(cpu_to_node(i)); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 1990 | } |
| 1991 | } |
| 1992 | |
Jens Axboe | cc71a6f | 2017-01-11 14:29:56 -0700 | [diff] [blame] | 1993 | static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx) |
| 1994 | { |
| 1995 | int ret = 0; |
| 1996 | |
| 1997 | set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx, |
| 1998 | set->queue_depth, set->reserved_tags); |
| 1999 | if (!set->tags[hctx_idx]) |
| 2000 | return false; |
| 2001 | |
| 2002 | ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx, |
| 2003 | set->queue_depth); |
| 2004 | if (!ret) |
| 2005 | return true; |
| 2006 | |
| 2007 | blk_mq_free_rq_map(set->tags[hctx_idx]); |
| 2008 | set->tags[hctx_idx] = NULL; |
| 2009 | return false; |
| 2010 | } |
| 2011 | |
| 2012 | static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set, |
| 2013 | unsigned int hctx_idx) |
| 2014 | { |
Jens Axboe | bd166ef | 2017-01-17 06:03:22 -0700 | [diff] [blame] | 2015 | if (set->tags[hctx_idx]) { |
| 2016 | blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx); |
| 2017 | blk_mq_free_rq_map(set->tags[hctx_idx]); |
| 2018 | set->tags[hctx_idx] = NULL; |
| 2019 | } |
Jens Axboe | cc71a6f | 2017-01-11 14:29:56 -0700 | [diff] [blame] | 2020 | } |
| 2021 | |
Akinobu Mita | 5778322 | 2015-09-27 02:09:23 +0900 | [diff] [blame] | 2022 | static void blk_mq_map_swqueue(struct request_queue *q, |
| 2023 | const struct cpumask *online_mask) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2024 | { |
Gabriel Krisman Bertazi | d1b1cea | 2016-12-14 18:48:36 -0200 | [diff] [blame] | 2025 | unsigned int i, hctx_idx; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2026 | struct blk_mq_hw_ctx *hctx; |
| 2027 | struct blk_mq_ctx *ctx; |
Ming Lei | 2a34c08 | 2015-04-21 10:00:20 +0800 | [diff] [blame] | 2028 | struct blk_mq_tag_set *set = q->tag_set; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2029 | |
Akinobu Mita | 60de074 | 2015-09-27 02:09:25 +0900 | [diff] [blame] | 2030 | /* |
| 2031 | * Avoid others reading imcomplete hctx->cpumask through sysfs |
| 2032 | */ |
| 2033 | mutex_lock(&q->sysfs_lock); |
| 2034 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2035 | queue_for_each_hw_ctx(q, hctx, i) { |
Jens Axboe | e4043dc | 2014-04-09 10:18:23 -0600 | [diff] [blame] | 2036 | cpumask_clear(hctx->cpumask); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2037 | hctx->nr_ctx = 0; |
| 2038 | } |
| 2039 | |
| 2040 | /* |
| 2041 | * Map software to hardware queues |
| 2042 | */ |
Thomas Gleixner | 897bb0c | 2016-03-19 11:30:33 +0100 | [diff] [blame] | 2043 | for_each_possible_cpu(i) { |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2044 | /* 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] | 2045 | if (!cpumask_test_cpu(i, online_mask)) |
Jens Axboe | e4043dc | 2014-04-09 10:18:23 -0600 | [diff] [blame] | 2046 | continue; |
| 2047 | |
Gabriel Krisman Bertazi | d1b1cea | 2016-12-14 18:48:36 -0200 | [diff] [blame] | 2048 | hctx_idx = q->mq_map[i]; |
| 2049 | /* unmapped hw queue can be remapped after CPU topo changed */ |
Jens Axboe | cc71a6f | 2017-01-11 14:29:56 -0700 | [diff] [blame] | 2050 | if (!set->tags[hctx_idx] && |
| 2051 | !__blk_mq_alloc_rq_map(set, hctx_idx)) { |
Gabriel Krisman Bertazi | d1b1cea | 2016-12-14 18:48:36 -0200 | [diff] [blame] | 2052 | /* |
| 2053 | * If tags initialization fail for some hctx, |
| 2054 | * that hctx won't be brought online. In this |
| 2055 | * case, remap the current ctx to hctx[0] which |
| 2056 | * is guaranteed to always have tags allocated |
| 2057 | */ |
Jens Axboe | cc71a6f | 2017-01-11 14:29:56 -0700 | [diff] [blame] | 2058 | q->mq_map[i] = 0; |
Gabriel Krisman Bertazi | d1b1cea | 2016-12-14 18:48:36 -0200 | [diff] [blame] | 2059 | } |
| 2060 | |
Thomas Gleixner | 897bb0c | 2016-03-19 11:30:33 +0100 | [diff] [blame] | 2061 | ctx = per_cpu_ptr(q->queue_ctx, i); |
Christoph Hellwig | 7d7e0f9 | 2016-09-14 16:18:54 +0200 | [diff] [blame] | 2062 | hctx = blk_mq_map_queue(q, i); |
Keith Busch | 868f2f0 | 2015-12-17 17:08:14 -0700 | [diff] [blame] | 2063 | |
Jens Axboe | e4043dc | 2014-04-09 10:18:23 -0600 | [diff] [blame] | 2064 | cpumask_set_cpu(i, hctx->cpumask); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2065 | ctx->index_hw = hctx->nr_ctx; |
| 2066 | hctx->ctxs[hctx->nr_ctx++] = ctx; |
| 2067 | } |
Jens Axboe | 506e931 | 2014-05-07 10:26:44 -0600 | [diff] [blame] | 2068 | |
Akinobu Mita | 60de074 | 2015-09-27 02:09:25 +0900 | [diff] [blame] | 2069 | mutex_unlock(&q->sysfs_lock); |
| 2070 | |
Jens Axboe | 506e931 | 2014-05-07 10:26:44 -0600 | [diff] [blame] | 2071 | queue_for_each_hw_ctx(q, hctx, i) { |
Jens Axboe | 484b406 | 2014-05-21 14:01:15 -0600 | [diff] [blame] | 2072 | /* |
Jens Axboe | a68aafa | 2014-08-15 13:19:15 -0600 | [diff] [blame] | 2073 | * If no software queues are mapped to this hardware queue, |
| 2074 | * disable it and free the request entries. |
Jens Axboe | 484b406 | 2014-05-21 14:01:15 -0600 | [diff] [blame] | 2075 | */ |
| 2076 | if (!hctx->nr_ctx) { |
Gabriel Krisman Bertazi | d1b1cea | 2016-12-14 18:48:36 -0200 | [diff] [blame] | 2077 | /* Never unmap queue 0. We need it as a |
| 2078 | * fallback in case of a new remap fails |
| 2079 | * allocation |
| 2080 | */ |
Jens Axboe | cc71a6f | 2017-01-11 14:29:56 -0700 | [diff] [blame] | 2081 | if (i && set->tags[i]) |
| 2082 | blk_mq_free_map_and_requests(set, i); |
| 2083 | |
Ming Lei | 2a34c08 | 2015-04-21 10:00:20 +0800 | [diff] [blame] | 2084 | hctx->tags = NULL; |
Jens Axboe | 484b406 | 2014-05-21 14:01:15 -0600 | [diff] [blame] | 2085 | continue; |
| 2086 | } |
| 2087 | |
Ming Lei | 2a34c08 | 2015-04-21 10:00:20 +0800 | [diff] [blame] | 2088 | hctx->tags = set->tags[i]; |
| 2089 | WARN_ON(!hctx->tags); |
| 2090 | |
Jens Axboe | 484b406 | 2014-05-21 14:01:15 -0600 | [diff] [blame] | 2091 | /* |
Chong Yuan | 889fa31 | 2015-04-15 11:39:29 -0600 | [diff] [blame] | 2092 | * Set the map size to the number of mapped software queues. |
| 2093 | * This is more accurate and more efficient than looping |
| 2094 | * over all possibly mapped software queues. |
| 2095 | */ |
Omar Sandoval | 8845964 | 2016-09-17 08:38:44 -0600 | [diff] [blame] | 2096 | sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx); |
Chong Yuan | 889fa31 | 2015-04-15 11:39:29 -0600 | [diff] [blame] | 2097 | |
| 2098 | /* |
Jens Axboe | 484b406 | 2014-05-21 14:01:15 -0600 | [diff] [blame] | 2099 | * Initialize batch roundrobin counts |
| 2100 | */ |
Jens Axboe | 506e931 | 2014-05-07 10:26:44 -0600 | [diff] [blame] | 2101 | hctx->next_cpu = cpumask_first(hctx->cpumask); |
| 2102 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; |
| 2103 | } |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2104 | } |
| 2105 | |
Jeff Moyer | 2404e60 | 2015-11-03 10:40:06 -0500 | [diff] [blame] | 2106 | static void queue_set_hctx_shared(struct request_queue *q, bool shared) |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 2107 | { |
| 2108 | struct blk_mq_hw_ctx *hctx; |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 2109 | int i; |
| 2110 | |
Jeff Moyer | 2404e60 | 2015-11-03 10:40:06 -0500 | [diff] [blame] | 2111 | queue_for_each_hw_ctx(q, hctx, i) { |
| 2112 | if (shared) |
| 2113 | hctx->flags |= BLK_MQ_F_TAG_SHARED; |
| 2114 | else |
| 2115 | hctx->flags &= ~BLK_MQ_F_TAG_SHARED; |
| 2116 | } |
| 2117 | } |
| 2118 | |
| 2119 | static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, bool shared) |
| 2120 | { |
| 2121 | struct request_queue *q; |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 2122 | |
Bart Van Assche | 705cda9 | 2017-04-07 11:16:49 -0700 | [diff] [blame] | 2123 | lockdep_assert_held(&set->tag_list_lock); |
| 2124 | |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 2125 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
| 2126 | blk_mq_freeze_queue(q); |
Jeff Moyer | 2404e60 | 2015-11-03 10:40:06 -0500 | [diff] [blame] | 2127 | queue_set_hctx_shared(q, shared); |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 2128 | blk_mq_unfreeze_queue(q); |
| 2129 | } |
| 2130 | } |
| 2131 | |
| 2132 | static void blk_mq_del_queue_tag_set(struct request_queue *q) |
| 2133 | { |
| 2134 | struct blk_mq_tag_set *set = q->tag_set; |
| 2135 | |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 2136 | mutex_lock(&set->tag_list_lock); |
Bart Van Assche | 705cda9 | 2017-04-07 11:16:49 -0700 | [diff] [blame] | 2137 | list_del_rcu(&q->tag_set_list); |
| 2138 | INIT_LIST_HEAD(&q->tag_set_list); |
Jeff Moyer | 2404e60 | 2015-11-03 10:40:06 -0500 | [diff] [blame] | 2139 | if (list_is_singular(&set->tag_list)) { |
| 2140 | /* just transitioned to unshared */ |
| 2141 | set->flags &= ~BLK_MQ_F_TAG_SHARED; |
| 2142 | /* update existing queue */ |
| 2143 | blk_mq_update_tag_set_depth(set, false); |
| 2144 | } |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 2145 | mutex_unlock(&set->tag_list_lock); |
Bart Van Assche | 705cda9 | 2017-04-07 11:16:49 -0700 | [diff] [blame] | 2146 | |
| 2147 | synchronize_rcu(); |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 2148 | } |
| 2149 | |
| 2150 | static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, |
| 2151 | struct request_queue *q) |
| 2152 | { |
| 2153 | q->tag_set = set; |
| 2154 | |
| 2155 | mutex_lock(&set->tag_list_lock); |
Jeff Moyer | 2404e60 | 2015-11-03 10:40:06 -0500 | [diff] [blame] | 2156 | |
| 2157 | /* Check to see if we're transitioning to shared (from 1 to 2 queues). */ |
| 2158 | if (!list_empty(&set->tag_list) && !(set->flags & BLK_MQ_F_TAG_SHARED)) { |
| 2159 | set->flags |= BLK_MQ_F_TAG_SHARED; |
| 2160 | /* update existing queue */ |
| 2161 | blk_mq_update_tag_set_depth(set, true); |
| 2162 | } |
| 2163 | if (set->flags & BLK_MQ_F_TAG_SHARED) |
| 2164 | queue_set_hctx_shared(q, true); |
Bart Van Assche | 705cda9 | 2017-04-07 11:16:49 -0700 | [diff] [blame] | 2165 | list_add_tail_rcu(&q->tag_set_list, &set->tag_list); |
Jeff Moyer | 2404e60 | 2015-11-03 10:40:06 -0500 | [diff] [blame] | 2166 | |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 2167 | mutex_unlock(&set->tag_list_lock); |
| 2168 | } |
| 2169 | |
Ming Lei | e09aae7 | 2015-01-29 20:17:27 +0800 | [diff] [blame] | 2170 | /* |
| 2171 | * It is the actual release handler for mq, but we do it from |
| 2172 | * request queue's release handler for avoiding use-after-free |
| 2173 | * and headache because q->mq_kobj shouldn't have been introduced, |
| 2174 | * but we can't group ctx/kctx kobj without it. |
| 2175 | */ |
| 2176 | void blk_mq_release(struct request_queue *q) |
| 2177 | { |
| 2178 | struct blk_mq_hw_ctx *hctx; |
| 2179 | unsigned int i; |
| 2180 | |
| 2181 | /* hctx kobj stays in hctx */ |
Ming Lei | c3b4afc | 2015-06-04 22:25:04 +0800 | [diff] [blame] | 2182 | queue_for_each_hw_ctx(q, hctx, i) { |
| 2183 | if (!hctx) |
| 2184 | continue; |
Ming Lei | 6c8b232 | 2017-02-22 18:14:01 +0800 | [diff] [blame] | 2185 | kobject_put(&hctx->kobj); |
Ming Lei | c3b4afc | 2015-06-04 22:25:04 +0800 | [diff] [blame] | 2186 | } |
Ming Lei | e09aae7 | 2015-01-29 20:17:27 +0800 | [diff] [blame] | 2187 | |
Akinobu Mita | a723bab | 2015-09-27 02:09:21 +0900 | [diff] [blame] | 2188 | q->mq_map = NULL; |
| 2189 | |
Ming Lei | e09aae7 | 2015-01-29 20:17:27 +0800 | [diff] [blame] | 2190 | kfree(q->queue_hw_ctx); |
| 2191 | |
Ming Lei | 7ea5fe3 | 2017-02-22 18:14:00 +0800 | [diff] [blame] | 2192 | /* |
| 2193 | * release .mq_kobj and sw queue's kobject now because |
| 2194 | * both share lifetime with request queue. |
| 2195 | */ |
| 2196 | blk_mq_sysfs_deinit(q); |
| 2197 | |
Ming Lei | e09aae7 | 2015-01-29 20:17:27 +0800 | [diff] [blame] | 2198 | free_percpu(q->queue_ctx); |
| 2199 | } |
| 2200 | |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 2201 | 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] | 2202 | { |
Mike Snitzer | b62c21b | 2015-03-12 23:56:02 -0400 | [diff] [blame] | 2203 | struct request_queue *uninit_q, *q; |
| 2204 | |
| 2205 | uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node); |
| 2206 | if (!uninit_q) |
| 2207 | return ERR_PTR(-ENOMEM); |
| 2208 | |
| 2209 | q = blk_mq_init_allocated_queue(set, uninit_q); |
| 2210 | if (IS_ERR(q)) |
| 2211 | blk_cleanup_queue(uninit_q); |
| 2212 | |
| 2213 | return q; |
| 2214 | } |
| 2215 | EXPORT_SYMBOL(blk_mq_init_queue); |
| 2216 | |
Keith Busch | 868f2f0 | 2015-12-17 17:08:14 -0700 | [diff] [blame] | 2217 | static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set, |
| 2218 | struct request_queue *q) |
Mike Snitzer | b62c21b | 2015-03-12 23:56:02 -0400 | [diff] [blame] | 2219 | { |
Keith Busch | 868f2f0 | 2015-12-17 17:08:14 -0700 | [diff] [blame] | 2220 | int i, j; |
| 2221 | struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2222 | |
Keith Busch | 868f2f0 | 2015-12-17 17:08:14 -0700 | [diff] [blame] | 2223 | blk_mq_sysfs_unregister(q); |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 2224 | for (i = 0; i < set->nr_hw_queues; i++) { |
Keith Busch | 868f2f0 | 2015-12-17 17:08:14 -0700 | [diff] [blame] | 2225 | int node; |
Jens Axboe | f14bbe7 | 2014-05-27 12:06:53 -0600 | [diff] [blame] | 2226 | |
Keith Busch | 868f2f0 | 2015-12-17 17:08:14 -0700 | [diff] [blame] | 2227 | if (hctxs[i]) |
| 2228 | continue; |
| 2229 | |
| 2230 | node = blk_mq_hw_queue_to_node(q->mq_map, i); |
Christoph Hellwig | cdef54d | 2014-05-28 18:11:06 +0200 | [diff] [blame] | 2231 | hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx), |
| 2232 | GFP_KERNEL, node); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2233 | if (!hctxs[i]) |
Keith Busch | 868f2f0 | 2015-12-17 17:08:14 -0700 | [diff] [blame] | 2234 | break; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2235 | |
Jens Axboe | a86073e | 2014-10-13 15:41:54 -0600 | [diff] [blame] | 2236 | if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL, |
Keith Busch | 868f2f0 | 2015-12-17 17:08:14 -0700 | [diff] [blame] | 2237 | node)) { |
| 2238 | kfree(hctxs[i]); |
| 2239 | hctxs[i] = NULL; |
| 2240 | break; |
| 2241 | } |
Jens Axboe | e4043dc | 2014-04-09 10:18:23 -0600 | [diff] [blame] | 2242 | |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 2243 | atomic_set(&hctxs[i]->nr_active, 0); |
Jens Axboe | f14bbe7 | 2014-05-27 12:06:53 -0600 | [diff] [blame] | 2244 | hctxs[i]->numa_node = node; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2245 | hctxs[i]->queue_num = i; |
Keith Busch | 868f2f0 | 2015-12-17 17:08:14 -0700 | [diff] [blame] | 2246 | |
| 2247 | if (blk_mq_init_hctx(q, set, hctxs[i], i)) { |
| 2248 | free_cpumask_var(hctxs[i]->cpumask); |
| 2249 | kfree(hctxs[i]); |
| 2250 | hctxs[i] = NULL; |
| 2251 | break; |
| 2252 | } |
| 2253 | blk_mq_hctx_kobj_init(hctxs[i]); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2254 | } |
Keith Busch | 868f2f0 | 2015-12-17 17:08:14 -0700 | [diff] [blame] | 2255 | for (j = i; j < q->nr_hw_queues; j++) { |
| 2256 | struct blk_mq_hw_ctx *hctx = hctxs[j]; |
| 2257 | |
| 2258 | if (hctx) { |
Jens Axboe | cc71a6f | 2017-01-11 14:29:56 -0700 | [diff] [blame] | 2259 | if (hctx->tags) |
| 2260 | blk_mq_free_map_and_requests(set, j); |
Keith Busch | 868f2f0 | 2015-12-17 17:08:14 -0700 | [diff] [blame] | 2261 | blk_mq_exit_hctx(q, set, hctx, j); |
Keith Busch | 868f2f0 | 2015-12-17 17:08:14 -0700 | [diff] [blame] | 2262 | kobject_put(&hctx->kobj); |
Keith Busch | 868f2f0 | 2015-12-17 17:08:14 -0700 | [diff] [blame] | 2263 | hctxs[j] = NULL; |
| 2264 | |
| 2265 | } |
| 2266 | } |
| 2267 | q->nr_hw_queues = i; |
| 2268 | blk_mq_sysfs_register(q); |
| 2269 | } |
| 2270 | |
| 2271 | struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set, |
| 2272 | struct request_queue *q) |
| 2273 | { |
Ming Lei | 6684167 | 2016-02-12 15:27:00 +0800 | [diff] [blame] | 2274 | /* mark the queue as mq asap */ |
| 2275 | q->mq_ops = set->ops; |
| 2276 | |
Omar Sandoval | 34dbad5 | 2017-03-21 08:56:08 -0700 | [diff] [blame] | 2277 | q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn, |
Stephen Bates | 720b8cc | 2017-04-07 06:24:03 -0600 | [diff] [blame] | 2278 | blk_mq_poll_stats_bkt, |
| 2279 | BLK_MQ_POLL_STATS_BKTS, q); |
Omar Sandoval | 34dbad5 | 2017-03-21 08:56:08 -0700 | [diff] [blame] | 2280 | if (!q->poll_cb) |
| 2281 | goto err_exit; |
| 2282 | |
Keith Busch | 868f2f0 | 2015-12-17 17:08:14 -0700 | [diff] [blame] | 2283 | q->queue_ctx = alloc_percpu(struct blk_mq_ctx); |
| 2284 | if (!q->queue_ctx) |
Ming Lin | c7de572 | 2016-05-25 23:23:27 -0700 | [diff] [blame] | 2285 | goto err_exit; |
Keith Busch | 868f2f0 | 2015-12-17 17:08:14 -0700 | [diff] [blame] | 2286 | |
Ming Lei | 737f98c | 2017-02-22 18:13:59 +0800 | [diff] [blame] | 2287 | /* init q->mq_kobj and sw queues' kobjects */ |
| 2288 | blk_mq_sysfs_init(q); |
| 2289 | |
Keith Busch | 868f2f0 | 2015-12-17 17:08:14 -0700 | [diff] [blame] | 2290 | q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)), |
| 2291 | GFP_KERNEL, set->numa_node); |
| 2292 | if (!q->queue_hw_ctx) |
| 2293 | goto err_percpu; |
| 2294 | |
Christoph Hellwig | bdd17e7 | 2016-09-14 16:18:53 +0200 | [diff] [blame] | 2295 | q->mq_map = set->mq_map; |
Keith Busch | 868f2f0 | 2015-12-17 17:08:14 -0700 | [diff] [blame] | 2296 | |
| 2297 | blk_mq_realloc_hw_ctxs(set, q); |
| 2298 | if (!q->nr_hw_queues) |
| 2299 | goto err_hctxs; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2300 | |
Christoph Hellwig | 287922eb | 2015-10-30 20:57:30 +0800 | [diff] [blame] | 2301 | INIT_WORK(&q->timeout_work, blk_mq_timeout_work); |
Ming Lei | e56f698 | 2015-07-16 19:53:22 +0800 | [diff] [blame] | 2302 | blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2303 | |
| 2304 | q->nr_queues = nr_cpu_ids; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2305 | |
Jens Axboe | 94eddfb | 2013-11-19 09:25:07 -0700 | [diff] [blame] | 2306 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2307 | |
Jens Axboe | 05f1dd5 | 2014-05-29 09:53:32 -0600 | [diff] [blame] | 2308 | if (!(set->flags & BLK_MQ_F_SG_MERGE)) |
| 2309 | q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE; |
| 2310 | |
Christoph Hellwig | 1be036e | 2014-02-07 10:22:39 -0800 | [diff] [blame] | 2311 | q->sg_reserved_size = INT_MAX; |
| 2312 | |
Mike Snitzer | 2849450 | 2016-09-14 13:28:30 -0400 | [diff] [blame] | 2313 | INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work); |
Christoph Hellwig | 6fca6a6 | 2014-05-28 08:08:02 -0600 | [diff] [blame] | 2314 | INIT_LIST_HEAD(&q->requeue_list); |
| 2315 | spin_lock_init(&q->requeue_lock); |
| 2316 | |
Christoph Hellwig | 254d259 | 2017-03-22 15:01:50 -0400 | [diff] [blame] | 2317 | blk_queue_make_request(q, blk_mq_make_request); |
Jens Axboe | 07068d5 | 2014-05-22 10:40:51 -0600 | [diff] [blame] | 2318 | |
Jens Axboe | eba7176 | 2014-05-20 15:17:27 -0600 | [diff] [blame] | 2319 | /* |
| 2320 | * Do this after blk_queue_make_request() overrides it... |
| 2321 | */ |
| 2322 | q->nr_requests = set->queue_depth; |
| 2323 | |
Jens Axboe | 64f1c21 | 2016-11-14 13:03:03 -0700 | [diff] [blame] | 2324 | /* |
| 2325 | * Default to classic polling |
| 2326 | */ |
| 2327 | q->poll_nsec = -1; |
| 2328 | |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 2329 | if (set->ops->complete) |
| 2330 | blk_queue_softirq_done(q, set->ops->complete); |
Christoph Hellwig | 30a91cb | 2014-02-10 03:24:38 -0800 | [diff] [blame] | 2331 | |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 2332 | blk_mq_init_cpu_queues(q, set->nr_hw_queues); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2333 | |
Peter Zijlstra | eabe065 | 2017-05-04 15:05:26 +0200 | [diff] [blame] | 2334 | get_online_cpus(); |
Wanpeng Li | 51d638b | 2017-05-07 00:14:22 -0700 | [diff] [blame] | 2335 | mutex_lock(&all_q_mutex); |
Akinobu Mita | 4593fdb | 2015-09-27 02:09:20 +0900 | [diff] [blame] | 2336 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2337 | list_add_tail(&q->all_q_node, &all_q_list); |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 2338 | blk_mq_add_queue_tag_set(set, q); |
Akinobu Mita | 5778322 | 2015-09-27 02:09:23 +0900 | [diff] [blame] | 2339 | blk_mq_map_swqueue(q, cpu_online_mask); |
Jens Axboe | 484b406 | 2014-05-21 14:01:15 -0600 | [diff] [blame] | 2340 | |
Peter Zijlstra | eabe065 | 2017-05-04 15:05:26 +0200 | [diff] [blame] | 2341 | mutex_unlock(&all_q_mutex); |
Wanpeng Li | 51d638b | 2017-05-07 00:14:22 -0700 | [diff] [blame] | 2342 | put_online_cpus(); |
Akinobu Mita | 4593fdb | 2015-09-27 02:09:20 +0900 | [diff] [blame] | 2343 | |
Jens Axboe | d348499 | 2017-01-13 14:43:58 -0700 | [diff] [blame] | 2344 | if (!(set->flags & BLK_MQ_F_NO_SCHED)) { |
| 2345 | int ret; |
| 2346 | |
| 2347 | ret = blk_mq_sched_init(q); |
| 2348 | if (ret) |
| 2349 | return ERR_PTR(ret); |
| 2350 | } |
| 2351 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2352 | return q; |
Christoph Hellwig | 1874198 | 2014-02-10 09:29:00 -0700 | [diff] [blame] | 2353 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2354 | err_hctxs: |
Keith Busch | 868f2f0 | 2015-12-17 17:08:14 -0700 | [diff] [blame] | 2355 | kfree(q->queue_hw_ctx); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2356 | err_percpu: |
Keith Busch | 868f2f0 | 2015-12-17 17:08:14 -0700 | [diff] [blame] | 2357 | free_percpu(q->queue_ctx); |
Ming Lin | c7de572 | 2016-05-25 23:23:27 -0700 | [diff] [blame] | 2358 | err_exit: |
| 2359 | q->mq_ops = NULL; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2360 | return ERR_PTR(-ENOMEM); |
| 2361 | } |
Mike Snitzer | b62c21b | 2015-03-12 23:56:02 -0400 | [diff] [blame] | 2362 | EXPORT_SYMBOL(blk_mq_init_allocated_queue); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2363 | |
| 2364 | void blk_mq_free_queue(struct request_queue *q) |
| 2365 | { |
Ming Lei | 624dbe4 | 2014-05-27 23:35:13 +0800 | [diff] [blame] | 2366 | struct blk_mq_tag_set *set = q->tag_set; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2367 | |
Akinobu Mita | 0e62636 | 2015-09-27 02:09:22 +0900 | [diff] [blame] | 2368 | mutex_lock(&all_q_mutex); |
| 2369 | list_del_init(&q->all_q_node); |
| 2370 | mutex_unlock(&all_q_mutex); |
| 2371 | |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 2372 | blk_mq_del_queue_tag_set(q); |
| 2373 | |
Ming Lei | 624dbe4 | 2014-05-27 23:35:13 +0800 | [diff] [blame] | 2374 | blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2375 | } |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2376 | |
| 2377 | /* Basically redo blk_mq_init_queue with queue frozen */ |
Akinobu Mita | 5778322 | 2015-09-27 02:09:23 +0900 | [diff] [blame] | 2378 | static void blk_mq_queue_reinit(struct request_queue *q, |
| 2379 | const struct cpumask *online_mask) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2380 | { |
Christoph Hellwig | 4ecd4fe | 2015-05-07 09:38:13 +0200 | [diff] [blame] | 2381 | WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth)); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2382 | |
Omar Sandoval | 9c1051a | 2017-05-04 08:17:21 -0600 | [diff] [blame] | 2383 | blk_mq_debugfs_unregister_hctxs(q); |
Jens Axboe | 67aec14 | 2014-05-30 08:25:36 -0600 | [diff] [blame] | 2384 | blk_mq_sysfs_unregister(q); |
| 2385 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2386 | /* |
| 2387 | * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe |
| 2388 | * we should change hctx numa_node according to new topology (this |
| 2389 | * involves free and re-allocate memory, worthy doing?) |
| 2390 | */ |
| 2391 | |
Akinobu Mita | 5778322 | 2015-09-27 02:09:23 +0900 | [diff] [blame] | 2392 | blk_mq_map_swqueue(q, online_mask); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2393 | |
Jens Axboe | 67aec14 | 2014-05-30 08:25:36 -0600 | [diff] [blame] | 2394 | blk_mq_sysfs_register(q); |
Omar Sandoval | 9c1051a | 2017-05-04 08:17:21 -0600 | [diff] [blame] | 2395 | blk_mq_debugfs_register_hctxs(q); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2396 | } |
| 2397 | |
Sebastian Andrzej Siewior | 65d5291 | 2016-09-22 08:05:19 -0600 | [diff] [blame] | 2398 | /* |
| 2399 | * New online cpumask which is going to be set in this hotplug event. |
| 2400 | * Declare this cpumasks as global as cpu-hotplug operation is invoked |
| 2401 | * one-by-one and dynamically allocating this could result in a failure. |
| 2402 | */ |
| 2403 | static struct cpumask cpuhp_online_new; |
| 2404 | |
| 2405 | static void blk_mq_queue_reinit_work(void) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2406 | { |
| 2407 | struct request_queue *q; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2408 | |
| 2409 | mutex_lock(&all_q_mutex); |
Tejun Heo | f3af020 | 2014-11-04 13:52:27 -0500 | [diff] [blame] | 2410 | /* |
| 2411 | * We need to freeze and reinit all existing queues. Freezing |
| 2412 | * involves synchronous wait for an RCU grace period and doing it |
| 2413 | * one by one may take a long time. Start freezing all queues in |
| 2414 | * one swoop and then wait for the completions so that freezing can |
| 2415 | * take place in parallel. |
| 2416 | */ |
| 2417 | list_for_each_entry(q, &all_q_list, all_q_node) |
Ming Lei | 1671d52 | 2017-03-27 20:06:57 +0800 | [diff] [blame] | 2418 | blk_freeze_queue_start(q); |
Gabriel Krisman Bertazi | 415d3da | 2016-11-28 15:01:48 -0200 | [diff] [blame] | 2419 | list_for_each_entry(q, &all_q_list, all_q_node) |
Tejun Heo | f3af020 | 2014-11-04 13:52:27 -0500 | [diff] [blame] | 2420 | blk_mq_freeze_queue_wait(q); |
| 2421 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2422 | list_for_each_entry(q, &all_q_list, all_q_node) |
Sebastian Andrzej Siewior | 65d5291 | 2016-09-22 08:05:19 -0600 | [diff] [blame] | 2423 | blk_mq_queue_reinit(q, &cpuhp_online_new); |
Tejun Heo | f3af020 | 2014-11-04 13:52:27 -0500 | [diff] [blame] | 2424 | |
| 2425 | list_for_each_entry(q, &all_q_list, all_q_node) |
| 2426 | blk_mq_unfreeze_queue(q); |
| 2427 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2428 | mutex_unlock(&all_q_mutex); |
Sebastian Andrzej Siewior | 65d5291 | 2016-09-22 08:05:19 -0600 | [diff] [blame] | 2429 | } |
| 2430 | |
| 2431 | static int blk_mq_queue_reinit_dead(unsigned int cpu) |
| 2432 | { |
Sebastian Andrzej Siewior | 97a3286 | 2016-09-23 15:02:38 +0200 | [diff] [blame] | 2433 | cpumask_copy(&cpuhp_online_new, cpu_online_mask); |
Sebastian Andrzej Siewior | 65d5291 | 2016-09-22 08:05:19 -0600 | [diff] [blame] | 2434 | blk_mq_queue_reinit_work(); |
| 2435 | return 0; |
| 2436 | } |
| 2437 | |
| 2438 | /* |
| 2439 | * Before hotadded cpu starts handling requests, new mappings must be |
| 2440 | * established. Otherwise, these requests in hw queue might never be |
| 2441 | * dispatched. |
| 2442 | * |
| 2443 | * For example, there is a single hw queue (hctx) and two CPU queues (ctx0 |
| 2444 | * for CPU0, and ctx1 for CPU1). |
| 2445 | * |
| 2446 | * Now CPU1 is just onlined and a request is inserted into ctx1->rq_list |
| 2447 | * and set bit0 in pending bitmap as ctx1->index_hw is still zero. |
| 2448 | * |
Jens Axboe | 2c3ad66 | 2016-12-14 14:34:47 -0700 | [diff] [blame] | 2449 | * And then while running hw queue, blk_mq_flush_busy_ctxs() finds bit0 is set |
| 2450 | * in pending bitmap and tries to retrieve requests in hctx->ctxs[0]->rq_list. |
| 2451 | * But htx->ctxs[0] is a pointer to ctx0, so the request in ctx1->rq_list is |
| 2452 | * ignored. |
Sebastian Andrzej Siewior | 65d5291 | 2016-09-22 08:05:19 -0600 | [diff] [blame] | 2453 | */ |
| 2454 | static int blk_mq_queue_reinit_prepare(unsigned int cpu) |
| 2455 | { |
| 2456 | cpumask_copy(&cpuhp_online_new, cpu_online_mask); |
| 2457 | cpumask_set_cpu(cpu, &cpuhp_online_new); |
| 2458 | blk_mq_queue_reinit_work(); |
| 2459 | return 0; |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2460 | } |
| 2461 | |
Jens Axboe | a516440 | 2014-09-10 09:02:03 -0600 | [diff] [blame] | 2462 | static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) |
| 2463 | { |
| 2464 | int i; |
| 2465 | |
Jens Axboe | cc71a6f | 2017-01-11 14:29:56 -0700 | [diff] [blame] | 2466 | for (i = 0; i < set->nr_hw_queues; i++) |
| 2467 | if (!__blk_mq_alloc_rq_map(set, i)) |
Jens Axboe | a516440 | 2014-09-10 09:02:03 -0600 | [diff] [blame] | 2468 | goto out_unwind; |
Jens Axboe | a516440 | 2014-09-10 09:02:03 -0600 | [diff] [blame] | 2469 | |
| 2470 | return 0; |
| 2471 | |
| 2472 | out_unwind: |
| 2473 | while (--i >= 0) |
Jens Axboe | cc71a6f | 2017-01-11 14:29:56 -0700 | [diff] [blame] | 2474 | blk_mq_free_rq_map(set->tags[i]); |
Jens Axboe | a516440 | 2014-09-10 09:02:03 -0600 | [diff] [blame] | 2475 | |
Jens Axboe | a516440 | 2014-09-10 09:02:03 -0600 | [diff] [blame] | 2476 | return -ENOMEM; |
| 2477 | } |
| 2478 | |
| 2479 | /* |
| 2480 | * Allocate the request maps associated with this tag_set. Note that this |
| 2481 | * may reduce the depth asked for, if memory is tight. set->queue_depth |
| 2482 | * will be updated to reflect the allocated depth. |
| 2483 | */ |
| 2484 | static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) |
| 2485 | { |
| 2486 | unsigned int depth; |
| 2487 | int err; |
| 2488 | |
| 2489 | depth = set->queue_depth; |
| 2490 | do { |
| 2491 | err = __blk_mq_alloc_rq_maps(set); |
| 2492 | if (!err) |
| 2493 | break; |
| 2494 | |
| 2495 | set->queue_depth >>= 1; |
| 2496 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) { |
| 2497 | err = -ENOMEM; |
| 2498 | break; |
| 2499 | } |
| 2500 | } while (set->queue_depth); |
| 2501 | |
| 2502 | if (!set->queue_depth || err) { |
| 2503 | pr_err("blk-mq: failed to allocate request map\n"); |
| 2504 | return -ENOMEM; |
| 2505 | } |
| 2506 | |
| 2507 | if (depth != set->queue_depth) |
| 2508 | pr_info("blk-mq: reduced tag depth (%u -> %u)\n", |
| 2509 | depth, set->queue_depth); |
| 2510 | |
| 2511 | return 0; |
| 2512 | } |
| 2513 | |
Omar Sandoval | ebe8bdd | 2017-04-07 08:53:11 -0600 | [diff] [blame] | 2514 | static int blk_mq_update_queue_map(struct blk_mq_tag_set *set) |
| 2515 | { |
| 2516 | if (set->ops->map_queues) |
| 2517 | return set->ops->map_queues(set); |
| 2518 | else |
| 2519 | return blk_mq_map_queues(set); |
| 2520 | } |
| 2521 | |
Jens Axboe | a4391c6 | 2014-06-05 15:21:56 -0600 | [diff] [blame] | 2522 | /* |
| 2523 | * Alloc a tag set to be associated with one or more request queues. |
| 2524 | * May fail with EINVAL for various error conditions. May adjust the |
| 2525 | * requested depth down, if if it too large. In that case, the set |
| 2526 | * value will be stored in set->queue_depth. |
| 2527 | */ |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 2528 | int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) |
| 2529 | { |
Christoph Hellwig | da695ba | 2016-09-14 16:18:55 +0200 | [diff] [blame] | 2530 | int ret; |
| 2531 | |
Bart Van Assche | 205fb5f | 2014-10-30 14:45:11 +0100 | [diff] [blame] | 2532 | BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS); |
| 2533 | |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 2534 | if (!set->nr_hw_queues) |
| 2535 | return -EINVAL; |
Jens Axboe | a4391c6 | 2014-06-05 15:21:56 -0600 | [diff] [blame] | 2536 | if (!set->queue_depth) |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 2537 | return -EINVAL; |
| 2538 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) |
| 2539 | return -EINVAL; |
| 2540 | |
Christoph Hellwig | 7d7e0f9 | 2016-09-14 16:18:54 +0200 | [diff] [blame] | 2541 | if (!set->ops->queue_rq) |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 2542 | return -EINVAL; |
| 2543 | |
Jens Axboe | a4391c6 | 2014-06-05 15:21:56 -0600 | [diff] [blame] | 2544 | if (set->queue_depth > BLK_MQ_MAX_DEPTH) { |
| 2545 | pr_info("blk-mq: reduced tag depth to %u\n", |
| 2546 | BLK_MQ_MAX_DEPTH); |
| 2547 | set->queue_depth = BLK_MQ_MAX_DEPTH; |
| 2548 | } |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 2549 | |
Shaohua Li | 6637fad | 2014-11-30 16:00:58 -0800 | [diff] [blame] | 2550 | /* |
| 2551 | * If a crashdump is active, then we are potentially in a very |
| 2552 | * memory constrained environment. Limit us to 1 queue and |
| 2553 | * 64 tags to prevent using too much memory. |
| 2554 | */ |
| 2555 | if (is_kdump_kernel()) { |
| 2556 | set->nr_hw_queues = 1; |
| 2557 | set->queue_depth = min(64U, set->queue_depth); |
| 2558 | } |
Keith Busch | 868f2f0 | 2015-12-17 17:08:14 -0700 | [diff] [blame] | 2559 | /* |
| 2560 | * There is no use for more h/w queues than cpus. |
| 2561 | */ |
| 2562 | if (set->nr_hw_queues > nr_cpu_ids) |
| 2563 | set->nr_hw_queues = nr_cpu_ids; |
Shaohua Li | 6637fad | 2014-11-30 16:00:58 -0800 | [diff] [blame] | 2564 | |
Keith Busch | 868f2f0 | 2015-12-17 17:08:14 -0700 | [diff] [blame] | 2565 | set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *), |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 2566 | GFP_KERNEL, set->numa_node); |
| 2567 | if (!set->tags) |
Jens Axboe | a516440 | 2014-09-10 09:02:03 -0600 | [diff] [blame] | 2568 | return -ENOMEM; |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 2569 | |
Christoph Hellwig | da695ba | 2016-09-14 16:18:55 +0200 | [diff] [blame] | 2570 | ret = -ENOMEM; |
| 2571 | set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids, |
| 2572 | GFP_KERNEL, set->numa_node); |
Christoph Hellwig | bdd17e7 | 2016-09-14 16:18:53 +0200 | [diff] [blame] | 2573 | if (!set->mq_map) |
| 2574 | goto out_free_tags; |
| 2575 | |
Omar Sandoval | ebe8bdd | 2017-04-07 08:53:11 -0600 | [diff] [blame] | 2576 | ret = blk_mq_update_queue_map(set); |
Christoph Hellwig | da695ba | 2016-09-14 16:18:55 +0200 | [diff] [blame] | 2577 | if (ret) |
| 2578 | goto out_free_mq_map; |
| 2579 | |
| 2580 | ret = blk_mq_alloc_rq_maps(set); |
| 2581 | if (ret) |
Christoph Hellwig | bdd17e7 | 2016-09-14 16:18:53 +0200 | [diff] [blame] | 2582 | goto out_free_mq_map; |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 2583 | |
Jens Axboe | 0d2602c | 2014-05-13 15:10:52 -0600 | [diff] [blame] | 2584 | mutex_init(&set->tag_list_lock); |
| 2585 | INIT_LIST_HEAD(&set->tag_list); |
| 2586 | |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 2587 | return 0; |
Christoph Hellwig | bdd17e7 | 2016-09-14 16:18:53 +0200 | [diff] [blame] | 2588 | |
| 2589 | out_free_mq_map: |
| 2590 | kfree(set->mq_map); |
| 2591 | set->mq_map = NULL; |
| 2592 | out_free_tags: |
Robert Elliott | 5676e7b | 2014-09-02 11:38:44 -0500 | [diff] [blame] | 2593 | kfree(set->tags); |
| 2594 | set->tags = NULL; |
Christoph Hellwig | da695ba | 2016-09-14 16:18:55 +0200 | [diff] [blame] | 2595 | return ret; |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 2596 | } |
| 2597 | EXPORT_SYMBOL(blk_mq_alloc_tag_set); |
| 2598 | |
| 2599 | void blk_mq_free_tag_set(struct blk_mq_tag_set *set) |
| 2600 | { |
| 2601 | int i; |
| 2602 | |
Jens Axboe | cc71a6f | 2017-01-11 14:29:56 -0700 | [diff] [blame] | 2603 | for (i = 0; i < nr_cpu_ids; i++) |
| 2604 | blk_mq_free_map_and_requests(set, i); |
Jens Axboe | 484b406 | 2014-05-21 14:01:15 -0600 | [diff] [blame] | 2605 | |
Christoph Hellwig | bdd17e7 | 2016-09-14 16:18:53 +0200 | [diff] [blame] | 2606 | kfree(set->mq_map); |
| 2607 | set->mq_map = NULL; |
| 2608 | |
Ming Lei | 981bd18 | 2014-04-24 00:07:34 +0800 | [diff] [blame] | 2609 | kfree(set->tags); |
Robert Elliott | 5676e7b | 2014-09-02 11:38:44 -0500 | [diff] [blame] | 2610 | set->tags = NULL; |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 2611 | } |
| 2612 | EXPORT_SYMBOL(blk_mq_free_tag_set); |
| 2613 | |
Jens Axboe | e3a2b3f | 2014-05-20 11:49:02 -0600 | [diff] [blame] | 2614 | int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr) |
| 2615 | { |
| 2616 | struct blk_mq_tag_set *set = q->tag_set; |
| 2617 | struct blk_mq_hw_ctx *hctx; |
| 2618 | int i, ret; |
| 2619 | |
Jens Axboe | bd166ef | 2017-01-17 06:03:22 -0700 | [diff] [blame] | 2620 | if (!set) |
Jens Axboe | e3a2b3f | 2014-05-20 11:49:02 -0600 | [diff] [blame] | 2621 | return -EINVAL; |
| 2622 | |
Jens Axboe | 70f36b6 | 2017-01-19 10:59:07 -0700 | [diff] [blame] | 2623 | blk_mq_freeze_queue(q); |
Jens Axboe | 70f36b6 | 2017-01-19 10:59:07 -0700 | [diff] [blame] | 2624 | |
Jens Axboe | e3a2b3f | 2014-05-20 11:49:02 -0600 | [diff] [blame] | 2625 | ret = 0; |
| 2626 | queue_for_each_hw_ctx(q, hctx, i) { |
Keith Busch | e9137d4 | 2016-02-18 14:56:35 -0700 | [diff] [blame] | 2627 | if (!hctx->tags) |
| 2628 | continue; |
Jens Axboe | bd166ef | 2017-01-17 06:03:22 -0700 | [diff] [blame] | 2629 | /* |
| 2630 | * If we're using an MQ scheduler, just update the scheduler |
| 2631 | * queue depth. This is similar to what the old code would do. |
| 2632 | */ |
Jens Axboe | 70f36b6 | 2017-01-19 10:59:07 -0700 | [diff] [blame] | 2633 | if (!hctx->sched_tags) { |
| 2634 | ret = blk_mq_tag_update_depth(hctx, &hctx->tags, |
| 2635 | min(nr, set->queue_depth), |
| 2636 | false); |
| 2637 | } else { |
| 2638 | ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags, |
| 2639 | nr, true); |
| 2640 | } |
Jens Axboe | e3a2b3f | 2014-05-20 11:49:02 -0600 | [diff] [blame] | 2641 | if (ret) |
| 2642 | break; |
| 2643 | } |
| 2644 | |
| 2645 | if (!ret) |
| 2646 | q->nr_requests = nr; |
| 2647 | |
Jens Axboe | 70f36b6 | 2017-01-19 10:59:07 -0700 | [diff] [blame] | 2648 | blk_mq_unfreeze_queue(q); |
Jens Axboe | 70f36b6 | 2017-01-19 10:59:07 -0700 | [diff] [blame] | 2649 | |
Jens Axboe | e3a2b3f | 2014-05-20 11:49:02 -0600 | [diff] [blame] | 2650 | return ret; |
| 2651 | } |
| 2652 | |
Keith Busch | e4dc2b3 | 2017-05-30 14:39:11 -0400 | [diff] [blame] | 2653 | static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, |
| 2654 | int nr_hw_queues) |
Keith Busch | 868f2f0 | 2015-12-17 17:08:14 -0700 | [diff] [blame] | 2655 | { |
| 2656 | struct request_queue *q; |
| 2657 | |
Bart Van Assche | 705cda9 | 2017-04-07 11:16:49 -0700 | [diff] [blame] | 2658 | lockdep_assert_held(&set->tag_list_lock); |
| 2659 | |
Keith Busch | 868f2f0 | 2015-12-17 17:08:14 -0700 | [diff] [blame] | 2660 | if (nr_hw_queues > nr_cpu_ids) |
| 2661 | nr_hw_queues = nr_cpu_ids; |
| 2662 | if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues) |
| 2663 | return; |
| 2664 | |
| 2665 | list_for_each_entry(q, &set->tag_list, tag_set_list) |
| 2666 | blk_mq_freeze_queue(q); |
| 2667 | |
| 2668 | set->nr_hw_queues = nr_hw_queues; |
Omar Sandoval | ebe8bdd | 2017-04-07 08:53:11 -0600 | [diff] [blame] | 2669 | blk_mq_update_queue_map(set); |
Keith Busch | 868f2f0 | 2015-12-17 17:08:14 -0700 | [diff] [blame] | 2670 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
| 2671 | blk_mq_realloc_hw_ctxs(set, q); |
Keith Busch | 868f2f0 | 2015-12-17 17:08:14 -0700 | [diff] [blame] | 2672 | blk_mq_queue_reinit(q, cpu_online_mask); |
| 2673 | } |
| 2674 | |
| 2675 | list_for_each_entry(q, &set->tag_list, tag_set_list) |
| 2676 | blk_mq_unfreeze_queue(q); |
| 2677 | } |
Keith Busch | e4dc2b3 | 2017-05-30 14:39:11 -0400 | [diff] [blame] | 2678 | |
| 2679 | void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues) |
| 2680 | { |
| 2681 | mutex_lock(&set->tag_list_lock); |
| 2682 | __blk_mq_update_nr_hw_queues(set, nr_hw_queues); |
| 2683 | mutex_unlock(&set->tag_list_lock); |
| 2684 | } |
Keith Busch | 868f2f0 | 2015-12-17 17:08:14 -0700 | [diff] [blame] | 2685 | EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues); |
| 2686 | |
Omar Sandoval | 34dbad5 | 2017-03-21 08:56:08 -0700 | [diff] [blame] | 2687 | /* Enable polling stats and return whether they were already enabled. */ |
| 2688 | static bool blk_poll_stats_enable(struct request_queue *q) |
| 2689 | { |
| 2690 | if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) || |
| 2691 | test_and_set_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags)) |
| 2692 | return true; |
| 2693 | blk_stat_add_callback(q, q->poll_cb); |
| 2694 | return false; |
| 2695 | } |
| 2696 | |
| 2697 | static void blk_mq_poll_stats_start(struct request_queue *q) |
| 2698 | { |
| 2699 | /* |
| 2700 | * We don't arm the callback if polling stats are not enabled or the |
| 2701 | * callback is already active. |
| 2702 | */ |
| 2703 | if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) || |
| 2704 | blk_stat_is_active(q->poll_cb)) |
| 2705 | return; |
| 2706 | |
| 2707 | blk_stat_activate_msecs(q->poll_cb, 100); |
| 2708 | } |
| 2709 | |
| 2710 | static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb) |
| 2711 | { |
| 2712 | struct request_queue *q = cb->data; |
Stephen Bates | 720b8cc | 2017-04-07 06:24:03 -0600 | [diff] [blame] | 2713 | int bucket; |
Omar Sandoval | 34dbad5 | 2017-03-21 08:56:08 -0700 | [diff] [blame] | 2714 | |
Stephen Bates | 720b8cc | 2017-04-07 06:24:03 -0600 | [diff] [blame] | 2715 | for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) { |
| 2716 | if (cb->stat[bucket].nr_samples) |
| 2717 | q->poll_stat[bucket] = cb->stat[bucket]; |
| 2718 | } |
Omar Sandoval | 34dbad5 | 2017-03-21 08:56:08 -0700 | [diff] [blame] | 2719 | } |
| 2720 | |
Jens Axboe | 64f1c21 | 2016-11-14 13:03:03 -0700 | [diff] [blame] | 2721 | static unsigned long blk_mq_poll_nsecs(struct request_queue *q, |
| 2722 | struct blk_mq_hw_ctx *hctx, |
| 2723 | struct request *rq) |
| 2724 | { |
Jens Axboe | 64f1c21 | 2016-11-14 13:03:03 -0700 | [diff] [blame] | 2725 | unsigned long ret = 0; |
Stephen Bates | 720b8cc | 2017-04-07 06:24:03 -0600 | [diff] [blame] | 2726 | int bucket; |
Jens Axboe | 64f1c21 | 2016-11-14 13:03:03 -0700 | [diff] [blame] | 2727 | |
| 2728 | /* |
| 2729 | * If stats collection isn't on, don't sleep but turn it on for |
| 2730 | * future users |
| 2731 | */ |
Omar Sandoval | 34dbad5 | 2017-03-21 08:56:08 -0700 | [diff] [blame] | 2732 | if (!blk_poll_stats_enable(q)) |
Jens Axboe | 64f1c21 | 2016-11-14 13:03:03 -0700 | [diff] [blame] | 2733 | return 0; |
| 2734 | |
| 2735 | /* |
Jens Axboe | 64f1c21 | 2016-11-14 13:03:03 -0700 | [diff] [blame] | 2736 | * As an optimistic guess, use half of the mean service time |
| 2737 | * for this type of request. We can (and should) make this smarter. |
| 2738 | * For instance, if the completion latencies are tight, we can |
| 2739 | * get closer than just half the mean. This is especially |
| 2740 | * important on devices where the completion latencies are longer |
Stephen Bates | 720b8cc | 2017-04-07 06:24:03 -0600 | [diff] [blame] | 2741 | * than ~10 usec. We do use the stats for the relevant IO size |
| 2742 | * if available which does lead to better estimates. |
Jens Axboe | 64f1c21 | 2016-11-14 13:03:03 -0700 | [diff] [blame] | 2743 | */ |
Stephen Bates | 720b8cc | 2017-04-07 06:24:03 -0600 | [diff] [blame] | 2744 | bucket = blk_mq_poll_stats_bkt(rq); |
| 2745 | if (bucket < 0) |
| 2746 | return ret; |
| 2747 | |
| 2748 | if (q->poll_stat[bucket].nr_samples) |
| 2749 | ret = (q->poll_stat[bucket].mean + 1) / 2; |
Jens Axboe | 64f1c21 | 2016-11-14 13:03:03 -0700 | [diff] [blame] | 2750 | |
| 2751 | return ret; |
| 2752 | } |
| 2753 | |
Jens Axboe | 06426ad | 2016-11-14 13:01:59 -0700 | [diff] [blame] | 2754 | static bool blk_mq_poll_hybrid_sleep(struct request_queue *q, |
Jens Axboe | 64f1c21 | 2016-11-14 13:03:03 -0700 | [diff] [blame] | 2755 | struct blk_mq_hw_ctx *hctx, |
Jens Axboe | 06426ad | 2016-11-14 13:01:59 -0700 | [diff] [blame] | 2756 | struct request *rq) |
| 2757 | { |
| 2758 | struct hrtimer_sleeper hs; |
| 2759 | enum hrtimer_mode mode; |
Jens Axboe | 64f1c21 | 2016-11-14 13:03:03 -0700 | [diff] [blame] | 2760 | unsigned int nsecs; |
Jens Axboe | 06426ad | 2016-11-14 13:01:59 -0700 | [diff] [blame] | 2761 | ktime_t kt; |
| 2762 | |
Jens Axboe | 64f1c21 | 2016-11-14 13:03:03 -0700 | [diff] [blame] | 2763 | if (test_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags)) |
| 2764 | return false; |
| 2765 | |
| 2766 | /* |
| 2767 | * poll_nsec can be: |
| 2768 | * |
| 2769 | * -1: don't ever hybrid sleep |
| 2770 | * 0: use half of prev avg |
| 2771 | * >0: use this specific value |
| 2772 | */ |
| 2773 | if (q->poll_nsec == -1) |
| 2774 | return false; |
| 2775 | else if (q->poll_nsec > 0) |
| 2776 | nsecs = q->poll_nsec; |
| 2777 | else |
| 2778 | nsecs = blk_mq_poll_nsecs(q, hctx, rq); |
| 2779 | |
| 2780 | if (!nsecs) |
Jens Axboe | 06426ad | 2016-11-14 13:01:59 -0700 | [diff] [blame] | 2781 | return false; |
| 2782 | |
| 2783 | set_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags); |
| 2784 | |
| 2785 | /* |
| 2786 | * This will be replaced with the stats tracking code, using |
| 2787 | * 'avg_completion_time / 2' as the pre-sleep target. |
| 2788 | */ |
Thomas Gleixner | 8b0e195 | 2016-12-25 12:30:41 +0100 | [diff] [blame] | 2789 | kt = nsecs; |
Jens Axboe | 06426ad | 2016-11-14 13:01:59 -0700 | [diff] [blame] | 2790 | |
| 2791 | mode = HRTIMER_MODE_REL; |
| 2792 | hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode); |
| 2793 | hrtimer_set_expires(&hs.timer, kt); |
| 2794 | |
| 2795 | hrtimer_init_sleeper(&hs, current); |
| 2796 | do { |
| 2797 | if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags)) |
| 2798 | break; |
| 2799 | set_current_state(TASK_UNINTERRUPTIBLE); |
| 2800 | hrtimer_start_expires(&hs.timer, mode); |
| 2801 | if (hs.task) |
| 2802 | io_schedule(); |
| 2803 | hrtimer_cancel(&hs.timer); |
| 2804 | mode = HRTIMER_MODE_ABS; |
| 2805 | } while (hs.task && !signal_pending(current)); |
| 2806 | |
| 2807 | __set_current_state(TASK_RUNNING); |
| 2808 | destroy_hrtimer_on_stack(&hs.timer); |
| 2809 | return true; |
| 2810 | } |
| 2811 | |
Jens Axboe | bbd7bb7 | 2016-11-04 09:34:34 -0600 | [diff] [blame] | 2812 | static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq) |
| 2813 | { |
| 2814 | struct request_queue *q = hctx->queue; |
| 2815 | long state; |
| 2816 | |
Jens Axboe | 06426ad | 2016-11-14 13:01:59 -0700 | [diff] [blame] | 2817 | /* |
| 2818 | * If we sleep, have the caller restart the poll loop to reset |
| 2819 | * the state. Like for the other success return cases, the |
| 2820 | * caller is responsible for checking if the IO completed. If |
| 2821 | * the IO isn't complete, we'll get called again and will go |
| 2822 | * straight to the busy poll loop. |
| 2823 | */ |
Jens Axboe | 64f1c21 | 2016-11-14 13:03:03 -0700 | [diff] [blame] | 2824 | if (blk_mq_poll_hybrid_sleep(q, hctx, rq)) |
Jens Axboe | 06426ad | 2016-11-14 13:01:59 -0700 | [diff] [blame] | 2825 | return true; |
| 2826 | |
Jens Axboe | bbd7bb7 | 2016-11-04 09:34:34 -0600 | [diff] [blame] | 2827 | hctx->poll_considered++; |
| 2828 | |
| 2829 | state = current->state; |
| 2830 | while (!need_resched()) { |
| 2831 | int ret; |
| 2832 | |
| 2833 | hctx->poll_invoked++; |
| 2834 | |
| 2835 | ret = q->mq_ops->poll(hctx, rq->tag); |
| 2836 | if (ret > 0) { |
| 2837 | hctx->poll_success++; |
| 2838 | set_current_state(TASK_RUNNING); |
| 2839 | return true; |
| 2840 | } |
| 2841 | |
| 2842 | if (signal_pending_state(state, current)) |
| 2843 | set_current_state(TASK_RUNNING); |
| 2844 | |
| 2845 | if (current->state == TASK_RUNNING) |
| 2846 | return true; |
| 2847 | if (ret < 0) |
| 2848 | break; |
| 2849 | cpu_relax(); |
| 2850 | } |
| 2851 | |
| 2852 | return false; |
| 2853 | } |
| 2854 | |
| 2855 | bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie) |
| 2856 | { |
| 2857 | struct blk_mq_hw_ctx *hctx; |
| 2858 | struct blk_plug *plug; |
| 2859 | struct request *rq; |
| 2860 | |
| 2861 | if (!q->mq_ops || !q->mq_ops->poll || !blk_qc_t_valid(cookie) || |
| 2862 | !test_bit(QUEUE_FLAG_POLL, &q->queue_flags)) |
| 2863 | return false; |
| 2864 | |
| 2865 | plug = current->plug; |
| 2866 | if (plug) |
| 2867 | blk_flush_plug_list(plug, false); |
| 2868 | |
| 2869 | hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)]; |
Jens Axboe | bd166ef | 2017-01-17 06:03:22 -0700 | [diff] [blame] | 2870 | if (!blk_qc_t_is_internal(cookie)) |
| 2871 | rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie)); |
Jens Axboe | 3a07bb1 | 2017-04-20 14:53:28 -0600 | [diff] [blame] | 2872 | else { |
Jens Axboe | bd166ef | 2017-01-17 06:03:22 -0700 | [diff] [blame] | 2873 | rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie)); |
Jens Axboe | 3a07bb1 | 2017-04-20 14:53:28 -0600 | [diff] [blame] | 2874 | /* |
| 2875 | * With scheduling, if the request has completed, we'll |
| 2876 | * get a NULL return here, as we clear the sched tag when |
| 2877 | * that happens. The request still remains valid, like always, |
| 2878 | * so we should be safe with just the NULL check. |
| 2879 | */ |
| 2880 | if (!rq) |
| 2881 | return false; |
| 2882 | } |
Jens Axboe | bbd7bb7 | 2016-11-04 09:34:34 -0600 | [diff] [blame] | 2883 | |
| 2884 | return __blk_mq_poll(hctx, rq); |
| 2885 | } |
| 2886 | EXPORT_SYMBOL_GPL(blk_mq_poll); |
| 2887 | |
Jens Axboe | 676141e | 2014-03-20 13:29:18 -0600 | [diff] [blame] | 2888 | void blk_mq_disable_hotplug(void) |
| 2889 | { |
| 2890 | mutex_lock(&all_q_mutex); |
| 2891 | } |
| 2892 | |
| 2893 | void blk_mq_enable_hotplug(void) |
| 2894 | { |
| 2895 | mutex_unlock(&all_q_mutex); |
| 2896 | } |
| 2897 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2898 | static int __init blk_mq_init(void) |
| 2899 | { |
Thomas Gleixner | 9467f85 | 2016-09-22 08:05:17 -0600 | [diff] [blame] | 2900 | cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL, |
| 2901 | blk_mq_hctx_notify_dead); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2902 | |
Sebastian Andrzej Siewior | 65d5291 | 2016-09-22 08:05:19 -0600 | [diff] [blame] | 2903 | cpuhp_setup_state_nocalls(CPUHP_BLK_MQ_PREPARE, "block/mq:prepare", |
| 2904 | blk_mq_queue_reinit_prepare, |
| 2905 | blk_mq_queue_reinit_dead); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 2906 | return 0; |
| 2907 | } |
| 2908 | subsys_initcall(blk_mq_init); |