aboutsummaryrefslogtreecommitdiffstats
path: root/block
diff options
context:
space:
mode:
authorJens Axboe <axboe@kernel.dk>2013-11-08 11:08:12 -0500
committerJens Axboe <axboe@kernel.dk>2013-11-08 11:08:12 -0500
commite37459b8e2c7db6735e39e019e448b76e5e77647 (patch)
treea3f0944db87a8ae0d41e5acbbbabc1e7ef534d1b /block
parentc7d1ba417c7cb7297d14dd47a390ec90ce548d5c (diff)
parente7e245000110a7794de8f925b9edc06a9c852f80 (diff)
Merge branch 'blk-mq/core' into for-3.13/core
Signed-off-by: Jens Axboe <axboe@kernel.dk> Conflicts: block/blk-timeout.c
Diffstat (limited to 'block')
-rw-r--r--block/Makefile5
-rw-r--r--block/blk-core.c157
-rw-r--r--block/blk-exec.c14
-rw-r--r--block/blk-flush.c154
-rw-r--r--block/blk-merge.c17
-rw-r--r--block/blk-mq-cpu.c93
-rw-r--r--block/blk-mq-cpumap.c108
-rw-r--r--block/blk-mq-sysfs.c384
-rw-r--r--block/blk-mq-tag.c204
-rw-r--r--block/blk-mq-tag.h27
-rw-r--r--block/blk-mq.c1500
-rw-r--r--block/blk-mq.h52
-rw-r--r--block/blk-sysfs.c13
-rw-r--r--block/blk-timeout.c74
-rw-r--r--block/blk.h17
15 files changed, 2701 insertions, 118 deletions
diff --git a/block/Makefile b/block/Makefile
index 671a83d063a5..20645e88fb57 100644
--- a/block/Makefile
+++ b/block/Makefile
@@ -5,8 +5,9 @@
5obj-$(CONFIG_BLOCK) := elevator.o blk-core.o blk-tag.o blk-sysfs.o \ 5obj-$(CONFIG_BLOCK) := elevator.o blk-core.o blk-tag.o blk-sysfs.o \
6 blk-flush.o blk-settings.o blk-ioc.o blk-map.o \ 6 blk-flush.o blk-settings.o blk-ioc.o blk-map.o \
7 blk-exec.o blk-merge.o blk-softirq.o blk-timeout.o \ 7 blk-exec.o blk-merge.o blk-softirq.o blk-timeout.o \
8 blk-iopoll.o blk-lib.o ioctl.o genhd.o scsi_ioctl.o \ 8 blk-iopoll.o blk-lib.o blk-mq.o blk-mq-tag.o \
9 partition-generic.o partitions/ 9 blk-mq-sysfs.o blk-mq-cpu.o blk-mq-cpumap.o ioctl.o \
10 genhd.o scsi_ioctl.o partition-generic.o partitions/
10 11
11obj-$(CONFIG_BLK_DEV_BSG) += bsg.o 12obj-$(CONFIG_BLK_DEV_BSG) += bsg.o
12obj-$(CONFIG_BLK_DEV_BSGLIB) += bsg-lib.o 13obj-$(CONFIG_BLK_DEV_BSGLIB) += bsg-lib.o
diff --git a/block/blk-core.c b/block/blk-core.c
index 25f13479f552..8bdd0121212a 100644
--- a/block/blk-core.c
+++ b/block/blk-core.c
@@ -16,6 +16,7 @@
16#include <linux/backing-dev.h> 16#include <linux/backing-dev.h>
17#include <linux/bio.h> 17#include <linux/bio.h>
18#include <linux/blkdev.h> 18#include <linux/blkdev.h>
19#include <linux/blk-mq.h>
19#include <linux/highmem.h> 20#include <linux/highmem.h>
20#include <linux/mm.h> 21#include <linux/mm.h>
21#include <linux/kernel_stat.h> 22#include <linux/kernel_stat.h>
@@ -48,7 +49,7 @@ DEFINE_IDA(blk_queue_ida);
48/* 49/*
49 * For the allocated request tables 50 * For the allocated request tables
50 */ 51 */
51static struct kmem_cache *request_cachep; 52struct kmem_cache *request_cachep = NULL;
52 53
53/* 54/*
54 * For queue allocation 55 * For queue allocation
@@ -60,42 +61,6 @@ struct kmem_cache *blk_requestq_cachep;
60 */ 61 */
61static struct workqueue_struct *kblockd_workqueue; 62static struct workqueue_struct *kblockd_workqueue;
62 63
63static void drive_stat_acct(struct request *rq, int new_io)
64{
65 struct hd_struct *part;
66 int rw = rq_data_dir(rq);
67 int cpu;
68
69 if (!blk_do_io_stat(rq))
70 return;
71
72 cpu = part_stat_lock();
73
74 if (!new_io) {
75 part = rq->part;
76 part_stat_inc(cpu, part, merges[rw]);
77 } else {
78 part = disk_map_sector_rcu(rq->rq_disk, blk_rq_pos(rq));
79 if (!hd_struct_try_get(part)) {
80 /*
81 * The partition is already being removed,
82 * the request will be accounted on the disk only
83 *
84 * We take a reference on disk->part0 although that
85 * partition will never be deleted, so we can treat
86 * it as any other partition.
87 */
88 part = &rq->rq_disk->part0;
89 hd_struct_get(part);
90 }
91 part_round_stats(cpu, part);
92 part_inc_in_flight(part, rw);
93 rq->part = part;
94 }
95
96 part_stat_unlock();
97}
98
99void blk_queue_congestion_threshold(struct request_queue *q) 64void blk_queue_congestion_threshold(struct request_queue *q)
100{ 65{
101 int nr; 66 int nr;
@@ -145,7 +110,6 @@ void blk_rq_init(struct request_queue *q, struct request *rq)
145 rq->cmd = rq->__cmd; 110 rq->cmd = rq->__cmd;
146 rq->cmd_len = BLK_MAX_CDB; 111 rq->cmd_len = BLK_MAX_CDB;
147 rq->tag = -1; 112 rq->tag = -1;
148 rq->ref_count = 1;
149 rq->start_time = jiffies; 113 rq->start_time = jiffies;
150 set_start_time_ns(rq); 114 set_start_time_ns(rq);
151 rq->part = NULL; 115 rq->part = NULL;
@@ -174,9 +138,9 @@ void blk_dump_rq_flags(struct request *rq, char *msg)
174{ 138{
175 int bit; 139 int bit;
176 140
177 printk(KERN_INFO "%s: dev %s: type=%x, flags=%x\n", msg, 141 printk(KERN_INFO "%s: dev %s: type=%x, flags=%llx\n", msg,
178 rq->rq_disk ? rq->rq_disk->disk_name : "?", rq->cmd_type, 142 rq->rq_disk ? rq->rq_disk->disk_name : "?", rq->cmd_type,
179 rq->cmd_flags); 143 (unsigned long long) rq->cmd_flags);
180 144
181 printk(KERN_INFO " sector %llu, nr/cnr %u/%u\n", 145 printk(KERN_INFO " sector %llu, nr/cnr %u/%u\n",
182 (unsigned long long)blk_rq_pos(rq), 146 (unsigned long long)blk_rq_pos(rq),
@@ -595,9 +559,12 @@ struct request_queue *blk_alloc_queue_node(gfp_t gfp_mask, int node_id)
595 if (!q) 559 if (!q)
596 return NULL; 560 return NULL;
597 561
562 if (percpu_counter_init(&q->mq_usage_counter, 0))
563 goto fail_q;
564
598 q->id = ida_simple_get(&blk_queue_ida, 0, 0, gfp_mask); 565 q->id = ida_simple_get(&blk_queue_ida, 0, 0, gfp_mask);
599 if (q->id < 0) 566 if (q->id < 0)
600 goto fail_q; 567 goto fail_c;
601 568
602 q->backing_dev_info.ra_pages = 569 q->backing_dev_info.ra_pages =
603 (VM_MAX_READAHEAD * 1024) / PAGE_CACHE_SIZE; 570 (VM_MAX_READAHEAD * 1024) / PAGE_CACHE_SIZE;
@@ -644,6 +611,8 @@ struct request_queue *blk_alloc_queue_node(gfp_t gfp_mask, int node_id)
644 q->bypass_depth = 1; 611 q->bypass_depth = 1;
645 __set_bit(QUEUE_FLAG_BYPASS, &q->queue_flags); 612 __set_bit(QUEUE_FLAG_BYPASS, &q->queue_flags);
646 613
614 init_waitqueue_head(&q->mq_freeze_wq);
615
647 if (blkcg_init_queue(q)) 616 if (blkcg_init_queue(q))
648 goto fail_bdi; 617 goto fail_bdi;
649 618
@@ -653,6 +622,8 @@ fail_bdi:
653 bdi_destroy(&q->backing_dev_info); 622 bdi_destroy(&q->backing_dev_info);
654fail_id: 623fail_id:
655 ida_simple_remove(&blk_queue_ida, q->id); 624 ida_simple_remove(&blk_queue_ida, q->id);
625fail_c:
626 percpu_counter_destroy(&q->mq_usage_counter);
656fail_q: 627fail_q:
657 kmem_cache_free(blk_requestq_cachep, q); 628 kmem_cache_free(blk_requestq_cachep, q);
658 return NULL; 629 return NULL;
@@ -1119,7 +1090,8 @@ retry:
1119 goto retry; 1090 goto retry;
1120} 1091}
1121 1092
1122struct request *blk_get_request(struct request_queue *q, int rw, gfp_t gfp_mask) 1093static struct request *blk_old_get_request(struct request_queue *q, int rw,
1094 gfp_t gfp_mask)
1123{ 1095{
1124 struct request *rq; 1096 struct request *rq;
1125 1097
@@ -1136,6 +1108,14 @@ struct request *blk_get_request(struct request_queue *q, int rw, gfp_t gfp_mask)
1136 1108
1137 return rq; 1109 return rq;
1138} 1110}
1111
1112struct request *blk_get_request(struct request_queue *q, int rw, gfp_t gfp_mask)
1113{
1114 if (q->mq_ops)
1115 return blk_mq_alloc_request(q, rw, gfp_mask, false);
1116 else
1117 return blk_old_get_request(q, rw, gfp_mask);
1118}
1139EXPORT_SYMBOL(blk_get_request); 1119EXPORT_SYMBOL(blk_get_request);
1140 1120
1141/** 1121/**
@@ -1221,7 +1201,7 @@ EXPORT_SYMBOL(blk_requeue_request);
1221static void add_acct_request(struct request_queue *q, struct request *rq, 1201static void add_acct_request(struct request_queue *q, struct request *rq,
1222 int where) 1202 int where)
1223{ 1203{
1224 drive_stat_acct(rq, 1); 1204 blk_account_io_start(rq, true);
1225 __elv_add_request(q, rq, where); 1205 __elv_add_request(q, rq, where);
1226} 1206}
1227 1207
@@ -1282,8 +1262,6 @@ void __blk_put_request(struct request_queue *q, struct request *req)
1282{ 1262{
1283 if (unlikely(!q)) 1263 if (unlikely(!q))
1284 return; 1264 return;
1285 if (unlikely(--req->ref_count))
1286 return;
1287 1265
1288 blk_pm_put_request(req); 1266 blk_pm_put_request(req);
1289 1267
@@ -1312,12 +1290,17 @@ EXPORT_SYMBOL_GPL(__blk_put_request);
1312 1290
1313void blk_put_request(struct request *req) 1291void blk_put_request(struct request *req)
1314{ 1292{
1315 unsigned long flags;
1316 struct request_queue *q = req->q; 1293 struct request_queue *q = req->q;
1317 1294
1318 spin_lock_irqsave(q->queue_lock, flags); 1295 if (q->mq_ops)
1319 __blk_put_request(q, req); 1296 blk_mq_free_request(req);
1320 spin_unlock_irqrestore(q->queue_lock, flags); 1297 else {
1298 unsigned long flags;
1299
1300 spin_lock_irqsave(q->queue_lock, flags);
1301 __blk_put_request(q, req);
1302 spin_unlock_irqrestore(q->queue_lock, flags);
1303 }
1321} 1304}
1322EXPORT_SYMBOL(blk_put_request); 1305EXPORT_SYMBOL(blk_put_request);
1323 1306
@@ -1353,8 +1336,8 @@ void blk_add_request_payload(struct request *rq, struct page *page,
1353} 1336}
1354EXPORT_SYMBOL_GPL(blk_add_request_payload); 1337EXPORT_SYMBOL_GPL(blk_add_request_payload);
1355 1338
1356static bool bio_attempt_back_merge(struct request_queue *q, struct request *req, 1339bool bio_attempt_back_merge(struct request_queue *q, struct request *req,
1357 struct bio *bio) 1340 struct bio *bio)
1358{ 1341{
1359 const int ff = bio->bi_rw & REQ_FAILFAST_MASK; 1342 const int ff = bio->bi_rw & REQ_FAILFAST_MASK;
1360 1343
@@ -1371,12 +1354,12 @@ static bool bio_attempt_back_merge(struct request_queue *q, struct request *req,
1371 req->__data_len += bio->bi_size; 1354 req->__data_len += bio->bi_size;
1372 req->ioprio = ioprio_best(req->ioprio, bio_prio(bio)); 1355 req->ioprio = ioprio_best(req->ioprio, bio_prio(bio));
1373 1356
1374 drive_stat_acct(req, 0); 1357 blk_account_io_start(req, false);
1375 return true; 1358 return true;
1376} 1359}
1377 1360
1378static bool bio_attempt_front_merge(struct request_queue *q, 1361bool bio_attempt_front_merge(struct request_queue *q, struct request *req,
1379 struct request *req, struct bio *bio) 1362 struct bio *bio)
1380{ 1363{
1381 const int ff = bio->bi_rw & REQ_FAILFAST_MASK; 1364 const int ff = bio->bi_rw & REQ_FAILFAST_MASK;
1382 1365
@@ -1401,12 +1384,12 @@ static bool bio_attempt_front_merge(struct request_queue *q,
1401 req->__data_len += bio->bi_size; 1384 req->__data_len += bio->bi_size;
1402 req->ioprio = ioprio_best(req->ioprio, bio_prio(bio)); 1385 req->ioprio = ioprio_best(req->ioprio, bio_prio(bio));
1403 1386
1404 drive_stat_acct(req, 0); 1387 blk_account_io_start(req, false);
1405 return true; 1388 return true;
1406} 1389}
1407 1390
1408/** 1391/**
1409 * attempt_plug_merge - try to merge with %current's plugged list 1392 * blk_attempt_plug_merge - try to merge with %current's plugged list
1410 * @q: request_queue new bio is being queued at 1393 * @q: request_queue new bio is being queued at
1411 * @bio: new bio being queued 1394 * @bio: new bio being queued
1412 * @request_count: out parameter for number of traversed plugged requests 1395 * @request_count: out parameter for number of traversed plugged requests
@@ -1422,12 +1405,13 @@ static bool bio_attempt_front_merge(struct request_queue *q,
1422 * reliable access to the elevator outside queue lock. Only check basic 1405 * reliable access to the elevator outside queue lock. Only check basic
1423 * merging parameters without querying the elevator. 1406 * merging parameters without querying the elevator.
1424 */ 1407 */
1425static bool attempt_plug_merge(struct request_queue *q, struct bio *bio, 1408bool blk_attempt_plug_merge(struct request_queue *q, struct bio *bio,
1426 unsigned int *request_count) 1409 unsigned int *request_count)
1427{ 1410{
1428 struct blk_plug *plug; 1411 struct blk_plug *plug;
1429 struct request *rq; 1412 struct request *rq;
1430 bool ret = false; 1413 bool ret = false;
1414 struct list_head *plug_list;
1431 1415
1432 if (blk_queue_nomerges(q)) 1416 if (blk_queue_nomerges(q))
1433 goto out; 1417 goto out;
@@ -1437,7 +1421,12 @@ static bool attempt_plug_merge(struct request_queue *q, struct bio *bio,
1437 goto out; 1421 goto out;
1438 *request_count = 0; 1422 *request_count = 0;
1439 1423
1440 list_for_each_entry_reverse(rq, &plug->list, queuelist) { 1424 if (q->mq_ops)
1425 plug_list = &plug->mq_list;
1426 else
1427 plug_list = &plug->list;
1428
1429 list_for_each_entry_reverse(rq, plug_list, queuelist) {
1441 int el_ret; 1430 int el_ret;
1442 1431
1443 if (rq->q == q) 1432 if (rq->q == q)
@@ -1505,7 +1494,7 @@ void blk_queue_bio(struct request_queue *q, struct bio *bio)
1505 * Check if we can merge with the plugged list before grabbing 1494 * Check if we can merge with the plugged list before grabbing
1506 * any locks. 1495 * any locks.
1507 */ 1496 */
1508 if (attempt_plug_merge(q, bio, &request_count)) 1497 if (blk_attempt_plug_merge(q, bio, &request_count))
1509 return; 1498 return;
1510 1499
1511 spin_lock_irq(q->queue_lock); 1500 spin_lock_irq(q->queue_lock);
@@ -1573,7 +1562,7 @@ get_rq:
1573 } 1562 }
1574 } 1563 }
1575 list_add_tail(&req->queuelist, &plug->list); 1564 list_add_tail(&req->queuelist, &plug->list);
1576 drive_stat_acct(req, 1); 1565 blk_account_io_start(req, true);
1577 } else { 1566 } else {
1578 spin_lock_irq(q->queue_lock); 1567 spin_lock_irq(q->queue_lock);
1579 add_acct_request(q, req, where); 1568 add_acct_request(q, req, where);
@@ -2027,7 +2016,7 @@ unsigned int blk_rq_err_bytes(const struct request *rq)
2027} 2016}
2028EXPORT_SYMBOL_GPL(blk_rq_err_bytes); 2017EXPORT_SYMBOL_GPL(blk_rq_err_bytes);
2029 2018
2030static void blk_account_io_completion(struct request *req, unsigned int bytes) 2019void blk_account_io_completion(struct request *req, unsigned int bytes)
2031{ 2020{
2032 if (blk_do_io_stat(req)) { 2021 if (blk_do_io_stat(req)) {
2033 const int rw = rq_data_dir(req); 2022 const int rw = rq_data_dir(req);
@@ -2041,7 +2030,7 @@ static void blk_account_io_completion(struct request *req, unsigned int bytes)
2041 } 2030 }
2042} 2031}
2043 2032
2044static void blk_account_io_done(struct request *req) 2033void blk_account_io_done(struct request *req)
2045{ 2034{
2046 /* 2035 /*
2047 * Account IO completion. flush_rq isn't accounted as a 2036 * Account IO completion. flush_rq isn't accounted as a
@@ -2089,6 +2078,42 @@ static inline struct request *blk_pm_peek_request(struct request_queue *q,
2089} 2078}
2090#endif 2079#endif
2091 2080
2081void blk_account_io_start(struct request *rq, bool new_io)
2082{
2083 struct hd_struct *part;
2084 int rw = rq_data_dir(rq);
2085 int cpu;
2086
2087 if (!blk_do_io_stat(rq))
2088 return;
2089
2090 cpu = part_stat_lock();
2091
2092 if (!new_io) {
2093 part = rq->part;
2094 part_stat_inc(cpu, part, merges[rw]);
2095 } else {
2096 part = disk_map_sector_rcu(rq->rq_disk, blk_rq_pos(rq));
2097 if (!hd_struct_try_get(part)) {
2098 /*
2099 * The partition is already being removed,
2100 * the request will be accounted on the disk only
2101 *
2102 * We take a reference on disk->part0 although that
2103 * partition will never be deleted, so we can treat
2104 * it as any other partition.
2105 */
2106 part = &rq->rq_disk->part0;
2107 hd_struct_get(part);
2108 }
2109 part_round_stats(cpu, part);
2110 part_inc_in_flight(part, rw);
2111 rq->part = part;
2112 }
2113
2114 part_stat_unlock();
2115}
2116
2092/** 2117/**
2093 * blk_peek_request - peek at the top of a request queue 2118 * blk_peek_request - peek at the top of a request queue
2094 * @q: request queue to peek at 2119 * @q: request queue to peek at
@@ -2465,7 +2490,6 @@ static void blk_finish_request(struct request *req, int error)
2465 if (req->cmd_flags & REQ_DONTPREP) 2490 if (req->cmd_flags & REQ_DONTPREP)
2466 blk_unprep_request(req); 2491 blk_unprep_request(req);
2467 2492
2468
2469 blk_account_io_done(req); 2493 blk_account_io_done(req);
2470 2494
2471 if (req->end_io) 2495 if (req->end_io)
@@ -2887,6 +2911,7 @@ void blk_start_plug(struct blk_plug *plug)
2887 2911
2888 plug->magic = PLUG_MAGIC; 2912 plug->magic = PLUG_MAGIC;
2889 INIT_LIST_HEAD(&plug->list); 2913 INIT_LIST_HEAD(&plug->list);
2914 INIT_LIST_HEAD(&plug->mq_list);
2890 INIT_LIST_HEAD(&plug->cb_list); 2915 INIT_LIST_HEAD(&plug->cb_list);
2891 2916
2892 /* 2917 /*
@@ -2984,6 +3009,10 @@ void blk_flush_plug_list(struct blk_plug *plug, bool from_schedule)
2984 BUG_ON(plug->magic != PLUG_MAGIC); 3009 BUG_ON(plug->magic != PLUG_MAGIC);
2985 3010
2986 flush_plug_callbacks(plug, from_schedule); 3011 flush_plug_callbacks(plug, from_schedule);
3012
3013 if (!list_empty(&plug->mq_list))
3014 blk_mq_flush_plug_list(plug, from_schedule);
3015
2987 if (list_empty(&plug->list)) 3016 if (list_empty(&plug->list))
2988 return; 3017 return;
2989 3018
diff --git a/block/blk-exec.c b/block/blk-exec.c
index ae4f27d7944e..c3edf9dff566 100644
--- a/block/blk-exec.c
+++ b/block/blk-exec.c
@@ -5,6 +5,7 @@
5#include <linux/module.h> 5#include <linux/module.h>
6#include <linux/bio.h> 6#include <linux/bio.h>
7#include <linux/blkdev.h> 7#include <linux/blkdev.h>
8#include <linux/blk-mq.h>
8#include <linux/sched/sysctl.h> 9#include <linux/sched/sysctl.h>
9 10
10#include "blk.h" 11#include "blk.h"
@@ -24,7 +25,6 @@ static void blk_end_sync_rq(struct request *rq, int error)
24 struct completion *waiting = rq->end_io_data; 25 struct completion *waiting = rq->end_io_data;
25 26
26 rq->end_io_data = NULL; 27 rq->end_io_data = NULL;
27 __blk_put_request(rq->q, rq);
28 28
29 /* 29 /*
30 * complete last, if this is a stack request the process (and thus 30 * complete last, if this is a stack request the process (and thus
@@ -59,6 +59,12 @@ void blk_execute_rq_nowait(struct request_queue *q, struct gendisk *bd_disk,
59 59
60 rq->rq_disk = bd_disk; 60 rq->rq_disk = bd_disk;
61 rq->end_io = done; 61 rq->end_io = done;
62
63 if (q->mq_ops) {
64 blk_mq_insert_request(q, rq, true);
65 return;
66 }
67
62 /* 68 /*
63 * need to check this before __blk_run_queue(), because rq can 69 * need to check this before __blk_run_queue(), because rq can
64 * be freed before that returns. 70 * be freed before that returns.
@@ -103,12 +109,6 @@ int blk_execute_rq(struct request_queue *q, struct gendisk *bd_disk,
103 int err = 0; 109 int err = 0;
104 unsigned long hang_check; 110 unsigned long hang_check;
105 111
106 /*
107 * we need an extra reference to the request, so we can look at
108 * it after io completion
109 */
110 rq->ref_count++;
111
112 if (!rq->sense) { 112 if (!rq->sense) {
113 memset(sense, 0, sizeof(sense)); 113 memset(sense, 0, sizeof(sense));
114 rq->sense = sense; 114 rq->sense = sense;
diff --git a/block/blk-flush.c b/block/blk-flush.c
index cc2b827a853c..331e627301ea 100644
--- a/block/blk-flush.c
+++ b/block/blk-flush.c
@@ -69,8 +69,10 @@
69#include <linux/bio.h> 69#include <linux/bio.h>
70#include <linux/blkdev.h> 70#include <linux/blkdev.h>
71#include <linux/gfp.h> 71#include <linux/gfp.h>
72#include <linux/blk-mq.h>
72 73
73#include "blk.h" 74#include "blk.h"
75#include "blk-mq.h"
74 76
75/* FLUSH/FUA sequences */ 77/* FLUSH/FUA sequences */
76enum { 78enum {
@@ -124,6 +126,24 @@ static void blk_flush_restore_request(struct request *rq)
124 /* make @rq a normal request */ 126 /* make @rq a normal request */
125 rq->cmd_flags &= ~REQ_FLUSH_SEQ; 127 rq->cmd_flags &= ~REQ_FLUSH_SEQ;
126 rq->end_io = rq->flush.saved_end_io; 128 rq->end_io = rq->flush.saved_end_io;
129
130 blk_clear_rq_complete(rq);
131}
132
133static void mq_flush_data_run(struct work_struct *work)
134{
135 struct request *rq;
136
137 rq = container_of(work, struct request, mq_flush_data);
138
139 memset(&rq->csd, 0, sizeof(rq->csd));
140 blk_mq_run_request(rq, true, false);
141}
142
143static void blk_mq_flush_data_insert(struct request *rq)
144{
145 INIT_WORK(&rq->mq_flush_data, mq_flush_data_run);
146 kblockd_schedule_work(rq->q, &rq->mq_flush_data);
127} 147}
128 148
129/** 149/**
@@ -136,7 +156,7 @@ static void blk_flush_restore_request(struct request *rq)
136 * completion and trigger the next step. 156 * completion and trigger the next step.
137 * 157 *
138 * CONTEXT: 158 * CONTEXT:
139 * spin_lock_irq(q->queue_lock) 159 * spin_lock_irq(q->queue_lock or q->mq_flush_lock)
140 * 160 *
141 * RETURNS: 161 * RETURNS:
142 * %true if requests were added to the dispatch queue, %false otherwise. 162 * %true if requests were added to the dispatch queue, %false otherwise.
@@ -146,7 +166,7 @@ static bool blk_flush_complete_seq(struct request *rq, unsigned int seq,
146{ 166{
147 struct request_queue *q = rq->q; 167 struct request_queue *q = rq->q;
148 struct list_head *pending = &q->flush_queue[q->flush_pending_idx]; 168 struct list_head *pending = &q->flush_queue[q->flush_pending_idx];
149 bool queued = false; 169 bool queued = false, kicked;
150 170
151 BUG_ON(rq->flush.seq & seq); 171 BUG_ON(rq->flush.seq & seq);
152 rq->flush.seq |= seq; 172 rq->flush.seq |= seq;
@@ -167,8 +187,12 @@ static bool blk_flush_complete_seq(struct request *rq, unsigned int seq,
167 187
168 case REQ_FSEQ_DATA: 188 case REQ_FSEQ_DATA:
169 list_move_tail(&rq->flush.list, &q->flush_data_in_flight); 189 list_move_tail(&rq->flush.list, &q->flush_data_in_flight);
170 list_add(&rq->queuelist, &q->queue_head); 190 if (q->mq_ops)
171 queued = true; 191 blk_mq_flush_data_insert(rq);
192 else {
193 list_add(&rq->queuelist, &q->queue_head);
194 queued = true;
195 }
172 break; 196 break;
173 197
174 case REQ_FSEQ_DONE: 198 case REQ_FSEQ_DONE:
@@ -181,28 +205,43 @@ static bool blk_flush_complete_seq(struct request *rq, unsigned int seq,
181 BUG_ON(!list_empty(&rq->queuelist)); 205 BUG_ON(!list_empty(&rq->queuelist));
182 list_del_init(&rq->flush.list); 206 list_del_init(&rq->flush.list);
183 blk_flush_restore_request(rq); 207 blk_flush_restore_request(rq);
184 __blk_end_request_all(rq, error); 208 if (q->mq_ops)
209 blk_mq_end_io(rq, error);
210 else
211 __blk_end_request_all(rq, error);
185 break; 212 break;
186 213
187 default: 214 default:
188 BUG(); 215 BUG();
189 } 216 }
190 217
191 return blk_kick_flush(q) | queued; 218 kicked = blk_kick_flush(q);
219 /* blk_mq_run_flush will run queue */
220 if (q->mq_ops)
221 return queued;
222 return kicked | queued;
192} 223}
193 224
194static void flush_end_io(struct request *flush_rq, int error) 225static void flush_end_io(struct request *flush_rq, int error)
195{ 226{
196 struct request_queue *q = flush_rq->q; 227 struct request_queue *q = flush_rq->q;
197 struct list_head *running = &q->flush_queue[q->flush_running_idx]; 228 struct list_head *running;
198 bool queued = false; 229 bool queued = false;
199 struct request *rq, *n; 230 struct request *rq, *n;
231 unsigned long flags = 0;
200 232
233 if (q->mq_ops) {
234 blk_mq_free_request(flush_rq);
235 spin_lock_irqsave(&q->mq_flush_lock, flags);
236 }
237 running = &q->flush_queue[q->flush_running_idx];
201 BUG_ON(q->flush_pending_idx == q->flush_running_idx); 238 BUG_ON(q->flush_pending_idx == q->flush_running_idx);
202 239
203 /* account completion of the flush request */ 240 /* account completion of the flush request */
204 q->flush_running_idx ^= 1; 241 q->flush_running_idx ^= 1;
205 elv_completed_request(q, flush_rq); 242
243 if (!q->mq_ops)
244 elv_completed_request(q, flush_rq);
206 245
207 /* and push the waiting requests to the next stage */ 246 /* and push the waiting requests to the next stage */
208 list_for_each_entry_safe(rq, n, running, flush.list) { 247 list_for_each_entry_safe(rq, n, running, flush.list) {
@@ -223,9 +262,48 @@ static void flush_end_io(struct request *flush_rq, int error)
223 * directly into request_fn may confuse the driver. Always use 262 * directly into request_fn may confuse the driver. Always use
224 * kblockd. 263 * kblockd.
225 */ 264 */
226 if (queued || q->flush_queue_delayed) 265 if (queued || q->flush_queue_delayed) {
227 blk_run_queue_async(q); 266 if (!q->mq_ops)
267 blk_run_queue_async(q);
268 else
269 /*
270 * This can be optimized to only run queues with requests
271 * queued if necessary.
272 */
273 blk_mq_run_queues(q, true);
274 }
228 q->flush_queue_delayed = 0; 275 q->flush_queue_delayed = 0;
276 if (q->mq_ops)
277 spin_unlock_irqrestore(&q->mq_flush_lock, flags);
278}
279
280static void mq_flush_work(struct work_struct *work)
281{
282 struct request_queue *q;
283 struct request *rq;
284
285 q = container_of(work, struct request_queue, mq_flush_work);
286
287 /* We don't need set REQ_FLUSH_SEQ, it's for consistency */
288 rq = blk_mq_alloc_request(q, WRITE_FLUSH|REQ_FLUSH_SEQ,
289 __GFP_WAIT|GFP_ATOMIC, true);
290 rq->cmd_type = REQ_TYPE_FS;
291 rq->end_io = flush_end_io;
292
293 blk_mq_run_request(rq, true, false);
294}
295
296/*
297 * We can't directly use q->flush_rq, because it doesn't have tag and is not in
298 * hctx->rqs[]. so we must allocate a new request, since we can't sleep here,
299 * so offload the work to workqueue.
300 *
301 * Note: we assume a flush request finished in any hardware queue will flush
302 * the whole disk cache.
303 */
304static void mq_run_flush(struct request_queue *q)
305{
306 kblockd_schedule_work(q, &q->mq_flush_work);
229} 307}
230 308
231/** 309/**
@@ -236,7 +314,7 @@ static void flush_end_io(struct request *flush_rq, int error)
236 * Please read the comment at the top of this file for more info. 314 * Please read the comment at the top of this file for more info.
237 * 315 *
238 * CONTEXT: 316 * CONTEXT:
239 * spin_lock_irq(q->queue_lock) 317 * spin_lock_irq(q->queue_lock or q->mq_flush_lock)
240 * 318 *
241 * RETURNS: 319 * RETURNS:
242 * %true if flush was issued, %false otherwise. 320 * %true if flush was issued, %false otherwise.
@@ -261,13 +339,18 @@ static bool blk_kick_flush(struct request_queue *q)
261 * Issue flush and toggle pending_idx. This makes pending_idx 339 * Issue flush and toggle pending_idx. This makes pending_idx
262 * different from running_idx, which means flush is in flight. 340 * different from running_idx, which means flush is in flight.
263 */ 341 */
342 q->flush_pending_idx ^= 1;
343 if (q->mq_ops) {
344 mq_run_flush(q);
345 return true;
346 }
347
264 blk_rq_init(q, &q->flush_rq); 348 blk_rq_init(q, &q->flush_rq);
265 q->flush_rq.cmd_type = REQ_TYPE_FS; 349 q->flush_rq.cmd_type = REQ_TYPE_FS;
266 q->flush_rq.cmd_flags = WRITE_FLUSH | REQ_FLUSH_SEQ; 350 q->flush_rq.cmd_flags = WRITE_FLUSH | REQ_FLUSH_SEQ;
267 q->flush_rq.rq_disk = first_rq->rq_disk; 351 q->flush_rq.rq_disk = first_rq->rq_disk;
268 q->flush_rq.end_io = flush_end_io; 352 q->flush_rq.end_io = flush_end_io;
269 353
270 q->flush_pending_idx ^= 1;
271 list_add_tail(&q->flush_rq.queuelist, &q->queue_head); 354 list_add_tail(&q->flush_rq.queuelist, &q->queue_head);
272 return true; 355 return true;
273} 356}
@@ -284,16 +367,37 @@ static void flush_data_end_io(struct request *rq, int error)
284 blk_run_queue_async(q); 367 blk_run_queue_async(q);
285} 368}
286 369
370static void mq_flush_data_end_io(struct request *rq, int error)
371{
372 struct request_queue *q = rq->q;
373 struct blk_mq_hw_ctx *hctx;
374 struct blk_mq_ctx *ctx;
375 unsigned long flags;
376
377 ctx = rq->mq_ctx;
378 hctx = q->mq_ops->map_queue(q, ctx->cpu);
379
380 /*
381 * After populating an empty queue, kick it to avoid stall. Read
382 * the comment in flush_end_io().
383 */
384 spin_lock_irqsave(&q->mq_flush_lock, flags);
385 if (blk_flush_complete_seq(rq, REQ_FSEQ_DATA, error))
386 blk_mq_run_hw_queue(hctx, true);
387 spin_unlock_irqrestore(&q->mq_flush_lock, flags);
388}
389
287/** 390/**
288 * blk_insert_flush - insert a new FLUSH/FUA request 391 * blk_insert_flush - insert a new FLUSH/FUA request
289 * @rq: request to insert 392 * @rq: request to insert
290 * 393 *
291 * To be called from __elv_add_request() for %ELEVATOR_INSERT_FLUSH insertions. 394 * To be called from __elv_add_request() for %ELEVATOR_INSERT_FLUSH insertions.
395 * or __blk_mq_run_hw_queue() to dispatch request.
292 * @rq is being submitted. Analyze what needs to be done and put it on the 396 * @rq is being submitted. Analyze what needs to be done and put it on the
293 * right queue. 397 * right queue.
294 * 398 *
295 * CONTEXT: 399 * CONTEXT:
296 * spin_lock_irq(q->queue_lock) 400 * spin_lock_irq(q->queue_lock) in !mq case
297 */ 401 */
298void blk_insert_flush(struct request *rq) 402void blk_insert_flush(struct request *rq)
299{ 403{
@@ -316,7 +420,10 @@ void blk_insert_flush(struct request *rq)
316 * complete the request. 420 * complete the request.
317 */ 421 */
318 if (!policy) { 422 if (!policy) {
319 __blk_end_bidi_request(rq, 0, 0, 0); 423 if (q->mq_ops)
424 blk_mq_end_io(rq, 0);
425 else
426 __blk_end_bidi_request(rq, 0, 0, 0);
320 return; 427 return;
321 } 428 }
322 429
@@ -329,7 +436,10 @@ void blk_insert_flush(struct request *rq)
329 */ 436 */
330 if ((policy & REQ_FSEQ_DATA) && 437 if ((policy & REQ_FSEQ_DATA) &&
331 !(policy & (REQ_FSEQ_PREFLUSH | REQ_FSEQ_POSTFLUSH))) { 438 !(policy & (REQ_FSEQ_PREFLUSH | REQ_FSEQ_POSTFLUSH))) {
332 list_add_tail(&rq->queuelist, &q->queue_head); 439 if (q->mq_ops) {
440 blk_mq_run_request(rq, false, true);
441 } else
442 list_add_tail(&rq->queuelist, &q->queue_head);
333 return; 443 return;
334 } 444 }
335 445
@@ -341,6 +451,14 @@ void blk_insert_flush(struct request *rq)
341 INIT_LIST_HEAD(&rq->flush.list); 451 INIT_LIST_HEAD(&rq->flush.list);
342 rq->cmd_flags |= REQ_FLUSH_SEQ; 452 rq->cmd_flags |= REQ_FLUSH_SEQ;
343 rq->flush.saved_end_io = rq->end_io; /* Usually NULL */ 453 rq->flush.saved_end_io = rq->end_io; /* Usually NULL */
454 if (q->mq_ops) {
455 rq->end_io = mq_flush_data_end_io;
456
457 spin_lock_irq(&q->mq_flush_lock);
458 blk_flush_complete_seq(rq, REQ_FSEQ_ACTIONS & ~policy, 0);
459 spin_unlock_irq(&q->mq_flush_lock);
460 return;
461 }
344 rq->end_io = flush_data_end_io; 462 rq->end_io = flush_data_end_io;
345 463
346 blk_flush_complete_seq(rq, REQ_FSEQ_ACTIONS & ~policy, 0); 464 blk_flush_complete_seq(rq, REQ_FSEQ_ACTIONS & ~policy, 0);
@@ -453,3 +571,9 @@ int blkdev_issue_flush(struct block_device *bdev, gfp_t gfp_mask,
453 return ret; 571 return ret;
454} 572}
455EXPORT_SYMBOL(blkdev_issue_flush); 573EXPORT_SYMBOL(blkdev_issue_flush);
574
575void blk_mq_init_flush(struct request_queue *q)
576{
577 spin_lock_init(&q->mq_flush_lock);
578 INIT_WORK(&q->mq_flush_work, mq_flush_work);
579}
diff --git a/block/blk-merge.c b/block/blk-merge.c
index 5f2448253797..1ffc58977835 100644
--- a/block/blk-merge.c
+++ b/block/blk-merge.c
@@ -308,6 +308,17 @@ int ll_front_merge_fn(struct request_queue *q, struct request *req,
308 return ll_new_hw_segment(q, req, bio); 308 return ll_new_hw_segment(q, req, bio);
309} 309}
310 310
311/*
312 * blk-mq uses req->special to carry normal driver per-request payload, it
313 * does not indicate a prepared command that we cannot merge with.
314 */
315static bool req_no_special_merge(struct request *req)
316{
317 struct request_queue *q = req->q;
318
319 return !q->mq_ops && req->special;
320}
321
311static int ll_merge_requests_fn(struct request_queue *q, struct request *req, 322static int ll_merge_requests_fn(struct request_queue *q, struct request *req,
312 struct request *next) 323 struct request *next)
313{ 324{
@@ -319,7 +330,7 @@ static int ll_merge_requests_fn(struct request_queue *q, struct request *req,
319 * First check if the either of the requests are re-queued 330 * First check if the either of the requests are re-queued
320 * requests. Can't merge them if they are. 331 * requests. Can't merge them if they are.
321 */ 332 */
322 if (req->special || next->special) 333 if (req_no_special_merge(req) || req_no_special_merge(next))
323 return 0; 334 return 0;
324 335
325 /* 336 /*
@@ -416,7 +427,7 @@ static int attempt_merge(struct request_queue *q, struct request *req,
416 427
417 if (rq_data_dir(req) != rq_data_dir(next) 428 if (rq_data_dir(req) != rq_data_dir(next)
418 || req->rq_disk != next->rq_disk 429 || req->rq_disk != next->rq_disk
419 || next->special) 430 || req_no_special_merge(next))
420 return 0; 431 return 0;
421 432
422 if (req->cmd_flags & REQ_WRITE_SAME && 433 if (req->cmd_flags & REQ_WRITE_SAME &&
@@ -515,7 +526,7 @@ bool blk_rq_merge_ok(struct request *rq, struct bio *bio)
515 return false; 526 return false;
516 527
517 /* must be same device and not a special request */ 528 /* must be same device and not a special request */
518 if (rq->rq_disk != bio->bi_bdev->bd_disk || rq->special) 529 if (rq->rq_disk != bio->bi_bdev->bd_disk || req_no_special_merge(rq))
519 return false; 530 return false;
520 531
521 /* only merge integrity protected bio into ditto rq */ 532 /* only merge integrity protected bio into ditto rq */
diff --git a/block/blk-mq-cpu.c b/block/blk-mq-cpu.c
new file mode 100644
index 000000000000..f8ea39d7ae54
--- /dev/null
+++ b/block/blk-mq-cpu.c
@@ -0,0 +1,93 @@
1#include <linux/kernel.h>
2#include <linux/module.h>
3#include <linux/init.h>
4#include <linux/blkdev.h>
5#include <linux/list.h>
6#include <linux/llist.h>
7#include <linux/smp.h>
8#include <linux/cpu.h>
9
10#include <linux/blk-mq.h>
11#include "blk-mq.h"
12
13static LIST_HEAD(blk_mq_cpu_notify_list);
14static DEFINE_SPINLOCK(blk_mq_cpu_notify_lock);
15
16static int __cpuinit blk_mq_main_cpu_notify(struct notifier_block *self,
17 unsigned long action, void *hcpu)
18{
19 unsigned int cpu = (unsigned long) hcpu;
20 struct blk_mq_cpu_notifier *notify;
21
22 spin_lock(&blk_mq_cpu_notify_lock);
23
24 list_for_each_entry(notify, &blk_mq_cpu_notify_list, list)
25 notify->notify(notify->data, action, cpu);
26
27 spin_unlock(&blk_mq_cpu_notify_lock);
28 return NOTIFY_OK;
29}
30
31static void __cpuinit blk_mq_cpu_notify(void *data, unsigned long action,
32 unsigned int cpu)
33{
34 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
35 /*
36 * If the CPU goes away, ensure that we run any pending
37 * completions.
38 */
39 struct llist_node *node;
40 struct request *rq;
41
42 local_irq_disable();
43
44 node = llist_del_all(&per_cpu(ipi_lists, cpu));
45 while (node) {
46 struct llist_node *next = node->next;
47
48 rq = llist_entry(node, struct request, ll_list);
49 __blk_mq_end_io(rq, rq->errors);
50 node = next;
51 }
52
53 local_irq_enable();
54 }
55}
56
57static struct notifier_block __cpuinitdata blk_mq_main_cpu_notifier = {
58 .notifier_call = blk_mq_main_cpu_notify,
59};
60
61void blk_mq_register_cpu_notifier(struct blk_mq_cpu_notifier *notifier)
62{
63 BUG_ON(!notifier->notify);
64
65 spin_lock(&blk_mq_cpu_notify_lock);
66 list_add_tail(&notifier->list, &blk_mq_cpu_notify_list);
67 spin_unlock(&blk_mq_cpu_notify_lock);
68}
69
70void blk_mq_unregister_cpu_notifier(struct blk_mq_cpu_notifier *notifier)
71{
72 spin_lock(&blk_mq_cpu_notify_lock);
73 list_del(&notifier->list);
74 spin_unlock(&blk_mq_cpu_notify_lock);
75}
76
77void blk_mq_init_cpu_notifier(struct blk_mq_cpu_notifier *notifier,
78 void (*fn)(void *, unsigned long, unsigned int),
79 void *data)
80{
81 notifier->notify = fn;
82 notifier->data = data;
83}
84
85static struct blk_mq_cpu_notifier __cpuinitdata cpu_notifier = {
86 .notify = blk_mq_cpu_notify,
87};
88
89void __init blk_mq_cpu_init(void)
90{
91 register_hotcpu_notifier(&blk_mq_main_cpu_notifier);
92 blk_mq_register_cpu_notifier(&cpu_notifier);
93}
diff --git a/block/blk-mq-cpumap.c b/block/blk-mq-cpumap.c
new file mode 100644
index 000000000000..f8721278601c
--- /dev/null
+++ b/block/blk-mq-cpumap.c
@@ -0,0 +1,108 @@
1#include <linux/kernel.h>
2#include <linux/threads.h>
3#include <linux/module.h>
4#include <linux/mm.h>
5#include <linux/smp.h>
6#include <linux/cpu.h>
7
8#include <linux/blk-mq.h>
9#include "blk.h"
10#include "blk-mq.h"
11
12static void show_map(unsigned int *map, unsigned int nr)
13{
14 int i;
15
16 pr_info("blk-mq: CPU -> queue map\n");
17 for_each_online_cpu(i)
18 pr_info(" CPU%2u -> Queue %u\n", i, map[i]);
19}
20
21static int cpu_to_queue_index(unsigned int nr_cpus, unsigned int nr_queues,
22 const int cpu)
23{
24 return cpu / ((nr_cpus + nr_queues - 1) / nr_queues);
25}
26
27static int get_first_sibling(unsigned int cpu)
28{
29 unsigned int ret;
30
31 ret = cpumask_first(topology_thread_cpumask(cpu));
32 if (ret < nr_cpu_ids)
33 return ret;
34
35 return cpu;
36}
37
38int blk_mq_update_queue_map(unsigned int *map, unsigned int nr_queues)
39{
40 unsigned int i, nr_cpus, nr_uniq_cpus, queue, first_sibling;
41 cpumask_var_t cpus;
42
43 if (!alloc_cpumask_var(&cpus, GFP_ATOMIC))
44 return 1;
45
46 cpumask_clear(cpus);
47 nr_cpus = nr_uniq_cpus = 0;
48 for_each_online_cpu(i) {
49 nr_cpus++;
50 first_sibling = get_first_sibling(i);
51 if (!cpumask_test_cpu(first_sibling, cpus))
52 nr_uniq_cpus++;
53 cpumask_set_cpu(i, cpus);
54 }
55
56 queue = 0;
57 for_each_possible_cpu(i) {
58 if (!cpu_online(i)) {
59 map[i] = 0;
60 continue;
61 }
62
63 /*
64 * Easy case - we have equal or more hardware queues. Or
65 * there are no thread siblings to take into account. Do
66 * 1:1 if enough, or sequential mapping if less.
67 */
68 if (nr_queues >= nr_cpus || nr_cpus == nr_uniq_cpus) {
69 map[i] = cpu_to_queue_index(nr_cpus, nr_queues, queue);
70 queue++;
71 continue;
72 }
73
74 /*
75 * Less then nr_cpus queues, and we have some number of
76 * threads per cores. Map sibling threads to the same
77 * queue.
78 */
79 first_sibling = get_first_sibling(i);
80 if (first_sibling == i) {
81 map[i] = cpu_to_queue_index(nr_uniq_cpus, nr_queues,
82 queue);
83 queue++;
84 } else
85 map[i] = map[first_sibling];
86 }
87
88 show_map(map, nr_cpus);
89 free_cpumask_var(cpus);
90 return 0;
91}
92
93unsigned int *blk_mq_make_queue_map(struct blk_mq_reg *reg)
94{
95 unsigned int *map;
96
97 /* If cpus are offline, map them to first hctx */
98 map = kzalloc_node(sizeof(*map) * num_possible_cpus(), GFP_KERNEL,
99 reg->numa_node);
100 if (!map)
101 return NULL;
102
103 if (!blk_mq_update_queue_map(map, reg->nr_hw_queues))
104 return map;
105
106 kfree(map);
107 return NULL;
108}
diff --git a/block/blk-mq-sysfs.c b/block/blk-mq-sysfs.c
new file mode 100644
index 000000000000..ba6cf8e9aa0a
--- /dev/null
+++ b/block/blk-mq-sysfs.c
@@ -0,0 +1,384 @@
1#include <linux/kernel.h>
2#include <linux/module.h>
3#include <linux/backing-dev.h>
4#include <linux/bio.h>
5#include <linux/blkdev.h>
6#include <linux/mm.h>
7#include <linux/init.h>
8#include <linux/slab.h>
9#include <linux/workqueue.h>
10#include <linux/smp.h>
11
12#include <linux/blk-mq.h>
13#include "blk-mq.h"
14#include "blk-mq-tag.h"
15
16static void blk_mq_sysfs_release(struct kobject *kobj)
17{
18}
19
20struct blk_mq_ctx_sysfs_entry {
21 struct attribute attr;
22 ssize_t (*show)(struct blk_mq_ctx *, char *);
23 ssize_t (*store)(struct blk_mq_ctx *, const char *, size_t);
24};
25
26struct blk_mq_hw_ctx_sysfs_entry {
27 struct attribute attr;
28 ssize_t (*show)(struct blk_mq_hw_ctx *, char *);
29 ssize_t (*store)(struct blk_mq_hw_ctx *, const char *, size_t);
30};
31
32static ssize_t blk_mq_sysfs_show(struct kobject *kobj, struct attribute *attr,
33 char *page)
34{
35 struct blk_mq_ctx_sysfs_entry *entry;
36 struct blk_mq_ctx *ctx;
37 struct request_queue *q;
38 ssize_t res;
39
40 entry = container_of(attr, struct blk_mq_ctx_sysfs_entry, attr);
41 ctx = container_of(kobj, struct blk_mq_ctx, kobj);
42 q = ctx->queue;
43
44 if (!entry->show)
45 return -EIO;
46
47 res = -ENOENT;
48 mutex_lock(&q->sysfs_lock);
49 if (!blk_queue_dying(q))
50 res = entry->show(ctx, page);
51 mutex_unlock(&q->sysfs_lock);
52 return res;
53}
54
55static ssize_t blk_mq_sysfs_store(struct kobject *kobj, struct attribute *attr,
56 const char *page, size_t length)
57{
58 struct blk_mq_ctx_sysfs_entry *entry;
59 struct blk_mq_ctx *ctx;
60 struct request_queue *q;
61 ssize_t res;
62
63 entry = container_of(attr, struct blk_mq_ctx_sysfs_entry, attr);
64 ctx = container_of(kobj, struct blk_mq_ctx, kobj);
65 q = ctx->queue;
66
67 if (!entry->store)
68 return -EIO;
69
70 res = -ENOENT;
71 mutex_lock(&q->sysfs_lock);
72 if (!blk_queue_dying(q))
73 res = entry->store(ctx, page, length);
74 mutex_unlock(&q->sysfs_lock);
75 return res;
76}
77
78static ssize_t blk_mq_hw_sysfs_show(struct kobject *kobj,
79 struct attribute *attr, char *page)
80{
81 struct blk_mq_hw_ctx_sysfs_entry *entry;
82 struct blk_mq_hw_ctx *hctx;
83 struct request_queue *q;
84 ssize_t res;
85
86 entry = container_of(attr, struct blk_mq_hw_ctx_sysfs_entry, attr);
87 hctx = container_of(kobj, struct blk_mq_hw_ctx, kobj);
88 q = hctx->queue;
89
90 if (!entry->show)
91 return -EIO;
92
93 res = -ENOENT;
94 mutex_lock(&q->sysfs_lock);
95 if (!blk_queue_dying(q))
96 res = entry->show(hctx, page);
97 mutex_unlock(&q->sysfs_lock);
98 return res;
99}
100
101static ssize_t blk_mq_hw_sysfs_store(struct kobject *kobj,
102 struct attribute *attr, const char *page,
103 size_t length)
104{
105 struct blk_mq_hw_ctx_sysfs_entry *entry;
106 struct blk_mq_hw_ctx *hctx;
107 struct request_queue *q;
108 ssize_t res;
109
110 entry = container_of(attr, struct blk_mq_hw_ctx_sysfs_entry, attr);
111 hctx = container_of(kobj, struct blk_mq_hw_ctx, kobj);
112 q = hctx->queue;
113
114 if (!entry->store)
115 return -EIO;
116
117 res = -ENOENT;
118 mutex_lock(&q->sysfs_lock);
119 if (!blk_queue_dying(q))
120 res = entry->store(hctx, page, length);
121 mutex_unlock(&q->sysfs_lock);
122 return res;
123}
124
125static ssize_t blk_mq_sysfs_dispatched_show(struct blk_mq_ctx *ctx, char *page)
126{
127 return sprintf(page, "%lu %lu\n", ctx->rq_dispatched[1],
128 ctx->rq_dispatched[0]);
129}
130
131static ssize_t blk_mq_sysfs_merged_show(struct blk_mq_ctx *ctx, char *page)
132{
133 return sprintf(page, "%lu\n", ctx->rq_merged);
134}
135
136static ssize_t blk_mq_sysfs_completed_show(struct blk_mq_ctx *ctx, char *page)
137{
138 return sprintf(page, "%lu %lu\n", ctx->rq_completed[1],
139 ctx->rq_completed[0]);
140}
141
142static ssize_t sysfs_list_show(char *page, struct list_head *list, char *msg)
143{
144 char *start_page = page;
145 struct request *rq;
146
147 page += sprintf(page, "%s:\n", msg);
148
149 list_for_each_entry(rq, list, queuelist)
150 page += sprintf(page, "\t%p\n", rq);
151
152 return page - start_page;
153}
154
155static ssize_t blk_mq_sysfs_rq_list_show(struct blk_mq_ctx *ctx, char *page)
156{
157 ssize_t ret;
158
159 spin_lock(&ctx->lock);
160 ret = sysfs_list_show(page, &ctx->rq_list, "CTX pending");
161 spin_unlock(&ctx->lock);
162
163 return ret;
164}
165
166static ssize_t blk_mq_hw_sysfs_queued_show(struct blk_mq_hw_ctx *hctx,
167 char *page)
168{
169 return sprintf(page, "%lu\n", hctx->queued);
170}
171
172static ssize_t blk_mq_hw_sysfs_run_show(struct blk_mq_hw_ctx *hctx, char *page)
173{
174 return sprintf(page, "%lu\n", hctx->run);
175}
176
177static ssize_t blk_mq_hw_sysfs_dispatched_show(struct blk_mq_hw_ctx *hctx,
178 char *page)
179{
180 char *start_page = page;
181 int i;
182
183 page += sprintf(page, "%8u\t%lu\n", 0U, hctx->dispatched[0]);
184
185 for (i = 1; i < BLK_MQ_MAX_DISPATCH_ORDER; i++) {
186 unsigned long d = 1U << (i - 1);
187
188 page += sprintf(page, "%8lu\t%lu\n", d, hctx->dispatched[i]);
189 }
190
191 return page - start_page;
192}
193
194static ssize_t blk_mq_hw_sysfs_rq_list_show(struct blk_mq_hw_ctx *hctx,
195 char *page)
196{
197 ssize_t ret;
198
199 spin_lock(&hctx->lock);
200 ret = sysfs_list_show(page, &hctx->dispatch, "HCTX pending");
201 spin_unlock(&hctx->lock);
202
203 return ret;
204}
205
206static ssize_t blk_mq_hw_sysfs_ipi_show(struct blk_mq_hw_ctx *hctx, char *page)
207{
208 ssize_t ret;
209
210 spin_lock(&hctx->lock);
211 ret = sprintf(page, "%u\n", !!(hctx->flags & BLK_MQ_F_SHOULD_IPI));
212 spin_unlock(&hctx->lock);
213
214 return ret;
215}
216
217static ssize_t blk_mq_hw_sysfs_ipi_store(struct blk_mq_hw_ctx *hctx,
218 const char *page, size_t len)
219{
220 struct blk_mq_ctx *ctx;
221 unsigned long ret;
222 unsigned int i;
223
224 if (kstrtoul(page, 10, &ret)) {
225 pr_err("blk-mq-sysfs: invalid input '%s'\n", page);
226 return -EINVAL;
227 }
228
229 spin_lock(&hctx->lock);
230 if (ret)
231 hctx->flags |= BLK_MQ_F_SHOULD_IPI;
232 else
233 hctx->flags &= ~BLK_MQ_F_SHOULD_IPI;
234 spin_unlock(&hctx->lock);
235
236 hctx_for_each_ctx(hctx, ctx, i)
237 ctx->ipi_redirect = !!ret;
238
239 return len;
240}
241
242static ssize_t blk_mq_hw_sysfs_tags_show(struct blk_mq_hw_ctx *hctx, char *page)
243{
244 return blk_mq_tag_sysfs_show(hctx->tags, page);
245}
246
247static struct blk_mq_ctx_sysfs_entry blk_mq_sysfs_dispatched = {
248 .attr = {.name = "dispatched", .mode = S_IRUGO },
249 .show = blk_mq_sysfs_dispatched_show,
250};
251static struct blk_mq_ctx_sysfs_entry blk_mq_sysfs_merged = {
252 .attr = {.name = "merged", .mode = S_IRUGO },
253 .show = blk_mq_sysfs_merged_show,
254};
255static struct blk_mq_ctx_sysfs_entry blk_mq_sysfs_completed = {
256 .attr = {.name = "completed", .mode = S_IRUGO },
257 .show = blk_mq_sysfs_completed_show,
258};
259static struct blk_mq_ctx_sysfs_entry blk_mq_sysfs_rq_list = {
260 .attr = {.name = "rq_list", .mode = S_IRUGO },
261 .show = blk_mq_sysfs_rq_list_show,
262};
263
264static struct attribute *default_ctx_attrs[] = {
265 &blk_mq_sysfs_dispatched.attr,
266 &blk_mq_sysfs_merged.attr,
267 &blk_mq_sysfs_completed.attr,
268 &blk_mq_sysfs_rq_list.attr,
269 NULL,
270};
271
272static struct blk_mq_hw_ctx_sysfs_entry blk_mq_hw_sysfs_queued = {
273 .attr = {.name = "queued", .mode = S_IRUGO },
274 .show = blk_mq_hw_sysfs_queued_show,
275};
276static struct blk_mq_hw_ctx_sysfs_entry blk_mq_hw_sysfs_run = {
277 .attr = {.name = "run", .mode = S_IRUGO },
278 .show = blk_mq_hw_sysfs_run_show,
279};
280static struct blk_mq_hw_ctx_sysfs_entry blk_mq_hw_sysfs_dispatched = {
281 .attr = {.name = "dispatched", .mode = S_IRUGO },
282 .show = blk_mq_hw_sysfs_dispatched_show,
283};
284static struct blk_mq_hw_ctx_sysfs_entry blk_mq_hw_sysfs_pending = {
285 .attr = {.name = "pending", .mode = S_IRUGO },
286 .show = blk_mq_hw_sysfs_rq_list_show,
287};
288static struct blk_mq_hw_ctx_sysfs_entry blk_mq_hw_sysfs_ipi = {
289 .attr = {.name = "ipi_redirect", .mode = S_IRUGO | S_IWUSR},
290 .show = blk_mq_hw_sysfs_ipi_show,
291 .store = blk_mq_hw_sysfs_ipi_store,
292};
293static struct blk_mq_hw_ctx_sysfs_entry blk_mq_hw_sysfs_tags = {
294 .attr = {.name = "tags", .mode = S_IRUGO },
295 .show = blk_mq_hw_sysfs_tags_show,
296};
297
298static struct attribute *default_hw_ctx_attrs[] = {
299 &blk_mq_hw_sysfs_queued.attr,
300 &blk_mq_hw_sysfs_run.attr,
301 &blk_mq_hw_sysfs_dispatched.attr,
302 &blk_mq_hw_sysfs_pending.attr,
303 &blk_mq_hw_sysfs_ipi.attr,
304 &blk_mq_hw_sysfs_tags.attr,
305 NULL,
306};
307
308static const struct sysfs_ops blk_mq_sysfs_ops = {
309 .show = blk_mq_sysfs_show,
310 .store = blk_mq_sysfs_store,
311};
312
313static const struct sysfs_ops blk_mq_hw_sysfs_ops = {
314 .show = blk_mq_hw_sysfs_show,
315 .store = blk_mq_hw_sysfs_store,
316};
317
318static struct kobj_type blk_mq_ktype = {
319 .sysfs_ops = &blk_mq_sysfs_ops,
320 .release = blk_mq_sysfs_release,
321};
322
323static struct kobj_type blk_mq_ctx_ktype = {
324 .sysfs_ops = &blk_mq_sysfs_ops,
325 .default_attrs = default_ctx_attrs,
326 .release = blk_mq_sysfs_release,
327};
328
329static struct kobj_type blk_mq_hw_ktype = {
330 .sysfs_ops = &blk_mq_hw_sysfs_ops,
331 .default_attrs = default_hw_ctx_attrs,
332 .release = blk_mq_sysfs_release,
333};
334
335void blk_mq_unregister_disk(struct gendisk *disk)
336{
337 struct request_queue *q = disk->queue;
338
339 kobject_uevent(&q->mq_kobj, KOBJ_REMOVE);
340 kobject_del(&q->mq_kobj);
341
342 kobject_put(&disk_to_dev(disk)->kobj);
343}
344
345int blk_mq_register_disk(struct gendisk *disk)
346{
347 struct device *dev = disk_to_dev(disk);
348 struct request_queue *q = disk->queue;
349 struct blk_mq_hw_ctx *hctx;
350 struct blk_mq_ctx *ctx;
351 int ret, i, j;
352
353 kobject_init(&q->mq_kobj, &blk_mq_ktype);
354
355 ret = kobject_add(&q->mq_kobj, kobject_get(&dev->kobj), "%s", "mq");
356 if (ret < 0)
357 return ret;
358
359 kobject_uevent(&q->mq_kobj, KOBJ_ADD);
360
361 queue_for_each_hw_ctx(q, hctx, i) {
362 kobject_init(&hctx->kobj, &blk_mq_hw_ktype);
363 ret = kobject_add(&hctx->kobj, &q->mq_kobj, "%u", i);
364 if (ret)
365 break;
366
367 if (!hctx->nr_ctx)
368 continue;
369
370 hctx_for_each_ctx(hctx, ctx, j) {
371 kobject_init(&ctx->kobj, &blk_mq_ctx_ktype);
372 ret = kobject_add(&ctx->kobj, &hctx->kobj, "cpu%u", ctx->cpu);
373 if (ret)
374 break;
375 }
376 }
377
378 if (ret) {
379 blk_mq_unregister_disk(disk);
380 return ret;
381 }
382
383 return 0;
384}
diff --git a/block/blk-mq-tag.c b/block/blk-mq-tag.c
new file mode 100644
index 000000000000..d64a02fb1f73
--- /dev/null
+++ b/block/blk-mq-tag.c
@@ -0,0 +1,204 @@
1#include <linux/kernel.h>
2#include <linux/module.h>
3#include <linux/percpu_ida.h>
4
5#include <linux/blk-mq.h>
6#include "blk.h"
7#include "blk-mq.h"
8#include "blk-mq-tag.h"
9
10/*
11 * Per tagged queue (tag address space) map
12 */
13struct blk_mq_tags {
14 unsigned int nr_tags;
15 unsigned int nr_reserved_tags;
16 unsigned int nr_batch_move;
17 unsigned int nr_max_cache;
18
19 struct percpu_ida free_tags;
20 struct percpu_ida reserved_tags;
21};
22
23void blk_mq_wait_for_tags(struct blk_mq_tags *tags)
24{
25 int tag = blk_mq_get_tag(tags, __GFP_WAIT, false);
26 blk_mq_put_tag(tags, tag);
27}
28
29bool blk_mq_has_free_tags(struct blk_mq_tags *tags)
30{
31 return !tags ||
32 percpu_ida_free_tags(&tags->free_tags, nr_cpu_ids) != 0;
33}
34
35static unsigned int __blk_mq_get_tag(struct blk_mq_tags *tags, gfp_t gfp)
36{
37 int tag;
38
39 tag = percpu_ida_alloc(&tags->free_tags, gfp);
40 if (tag < 0)
41 return BLK_MQ_TAG_FAIL;
42 return tag + tags->nr_reserved_tags;
43}
44
45static unsigned int __blk_mq_get_reserved_tag(struct blk_mq_tags *tags,
46 gfp_t gfp)
47{
48 int tag;
49
50 if (unlikely(!tags->nr_reserved_tags)) {
51 WARN_ON_ONCE(1);
52 return BLK_MQ_TAG_FAIL;
53 }
54
55 tag = percpu_ida_alloc(&tags->reserved_tags, gfp);
56 if (tag < 0)
57 return BLK_MQ_TAG_FAIL;
58 return tag;
59}
60
61unsigned int blk_mq_get_tag(struct blk_mq_tags *tags, gfp_t gfp, bool reserved)
62{
63 if (!reserved)
64 return __blk_mq_get_tag(tags, gfp);
65
66 return __blk_mq_get_reserved_tag(tags, gfp);
67}
68
69static void __blk_mq_put_tag(struct blk_mq_tags *tags, unsigned int tag)
70{
71 BUG_ON(tag >= tags->nr_tags);
72
73 percpu_ida_free(&tags->free_tags, tag - tags->nr_reserved_tags);
74}
75
76static void __blk_mq_put_reserved_tag(struct blk_mq_tags *tags,
77 unsigned int tag)
78{
79 BUG_ON(tag >= tags->nr_reserved_tags);
80
81 percpu_ida_free(&tags->reserved_tags, tag);
82}
83
84void blk_mq_put_tag(struct blk_mq_tags *tags, unsigned int tag)
85{
86 if (tag >= tags->nr_reserved_tags)
87 __blk_mq_put_tag(tags, tag);
88 else
89 __blk_mq_put_reserved_tag(tags, tag);
90}
91
92static int __blk_mq_tag_iter(unsigned id, void *data)
93{
94 unsigned long *tag_map = data;
95 __set_bit(id, tag_map);
96 return 0;
97}
98
99void blk_mq_tag_busy_iter(struct blk_mq_tags *tags,
100 void (*fn)(void *, unsigned long *), void *data)
101{
102 unsigned long *tag_map;
103 size_t map_size;
104
105 map_size = ALIGN(tags->nr_tags, BITS_PER_LONG) / BITS_PER_LONG;
106 tag_map = kzalloc(map_size * sizeof(unsigned long), GFP_ATOMIC);
107 if (!tag_map)
108 return;
109
110 percpu_ida_for_each_free(&tags->free_tags, __blk_mq_tag_iter, tag_map);
111 if (tags->nr_reserved_tags)
112 percpu_ida_for_each_free(&tags->reserved_tags, __blk_mq_tag_iter,
113 tag_map);
114
115 fn(data, tag_map);
116 kfree(tag_map);
117}
118
119struct blk_mq_tags *blk_mq_init_tags(unsigned int total_tags,
120 unsigned int reserved_tags, int node)
121{
122 unsigned int nr_tags, nr_cache;
123 struct blk_mq_tags *tags;
124 int ret;
125
126 if (total_tags > BLK_MQ_TAG_MAX) {
127 pr_err("blk-mq: tag depth too large\n");
128 return NULL;
129 }
130
131 tags = kzalloc_node(sizeof(*tags), GFP_KERNEL, node);
132 if (!tags)
133 return NULL;
134
135 nr_tags = total_tags - reserved_tags;
136 nr_cache = nr_tags / num_possible_cpus();
137
138 if (nr_cache < BLK_MQ_TAG_CACHE_MIN)
139 nr_cache = BLK_MQ_TAG_CACHE_MIN;
140 else if (nr_cache > BLK_MQ_TAG_CACHE_MAX)
141 nr_cache = BLK_MQ_TAG_CACHE_MAX;
142
143 tags->nr_tags = total_tags;
144 tags->nr_reserved_tags = reserved_tags;
145 tags->nr_max_cache = nr_cache;
146 tags->nr_batch_move = max(1u, nr_cache / 2);
147
148 ret = __percpu_ida_init(&tags->free_tags, tags->nr_tags -
149 tags->nr_reserved_tags,
150 tags->nr_max_cache,
151 tags->nr_batch_move);
152 if (ret)
153 goto err_free_tags;
154
155 if (reserved_tags) {
156 /*
157 * With max_cahe and batch set to 1, the allocator fallbacks to
158 * no cached. It's fine reserved tags allocation is slow.
159 */
160 ret = __percpu_ida_init(&tags->reserved_tags, reserved_tags,
161 1, 1);
162 if (ret)
163 goto err_reserved_tags;
164 }
165
166 return tags;
167
168err_reserved_tags:
169 percpu_ida_destroy(&tags->free_tags);
170err_free_tags:
171 kfree(tags);
172 return NULL;
173}
174
175void blk_mq_free_tags(struct blk_mq_tags *tags)
176{
177 percpu_ida_destroy(&tags->free_tags);
178 percpu_ida_destroy(&tags->reserved_tags);
179 kfree(tags);
180}
181
182ssize_t blk_mq_tag_sysfs_show(struct blk_mq_tags *tags, char *page)
183{
184 char *orig_page = page;
185 int cpu;
186
187 if (!tags)
188 return 0;
189
190 page += sprintf(page, "nr_tags=%u, reserved_tags=%u, batch_move=%u,"
191 " max_cache=%u\n", tags->nr_tags, tags->nr_reserved_tags,
192 tags->nr_batch_move, tags->nr_max_cache);
193
194 page += sprintf(page, "nr_free=%u, nr_reserved=%u\n",
195 percpu_ida_free_tags(&tags->free_tags, nr_cpu_ids),
196 percpu_ida_free_tags(&tags->reserved_tags, nr_cpu_ids));
197
198 for_each_possible_cpu(cpu) {
199 page += sprintf(page, " cpu%02u: nr_free=%u\n", cpu,
200 percpu_ida_free_tags(&tags->free_tags, cpu));
201 }
202
203 return page - orig_page;
204}
diff --git a/block/blk-mq-tag.h b/block/blk-mq-tag.h
new file mode 100644
index 000000000000..947ba2c6148e
--- /dev/null
+++ b/block/blk-mq-tag.h
@@ -0,0 +1,27 @@
1#ifndef INT_BLK_MQ_TAG_H
2#define INT_BLK_MQ_TAG_H
3
4struct blk_mq_tags;
5
6extern struct blk_mq_tags *blk_mq_init_tags(unsigned int nr_tags, unsigned int reserved_tags, int node);
7extern void blk_mq_free_tags(struct blk_mq_tags *tags);
8
9extern unsigned int blk_mq_get_tag(struct blk_mq_tags *tags, gfp_t gfp, bool reserved);
10extern void blk_mq_wait_for_tags(struct blk_mq_tags *tags);
11extern void blk_mq_put_tag(struct blk_mq_tags *tags, unsigned int tag);
12extern void blk_mq_tag_busy_iter(struct blk_mq_tags *tags, void (*fn)(void *data, unsigned long *), void *data);
13extern bool blk_mq_has_free_tags(struct blk_mq_tags *tags);
14extern ssize_t blk_mq_tag_sysfs_show(struct blk_mq_tags *tags, char *page);
15
16enum {
17 BLK_MQ_TAG_CACHE_MIN = 1,
18 BLK_MQ_TAG_CACHE_MAX = 64,
19};
20
21enum {
22 BLK_MQ_TAG_FAIL = -1U,
23 BLK_MQ_TAG_MIN = BLK_MQ_TAG_CACHE_MIN,
24 BLK_MQ_TAG_MAX = BLK_MQ_TAG_FAIL - 1,
25};
26
27#endif
diff --git a/block/blk-mq.c b/block/blk-mq.c
new file mode 100644
index 000000000000..88d4e864d4c0
--- /dev/null
+++ b/block/blk-mq.c
@@ -0,0 +1,1500 @@
1#include <linux/kernel.h>
2#include <linux/module.h>
3#include <linux/backing-dev.h>
4#include <linux/bio.h>
5#include <linux/blkdev.h>
6#include <linux/mm.h>
7#include <linux/init.h>
8#include <linux/slab.h>
9#include <linux/workqueue.h>
10#include <linux/smp.h>
11#include <linux/llist.h>
12#include <linux/list_sort.h>
13#include <linux/cpu.h>
14#include <linux/cache.h>
15#include <linux/sched/sysctl.h>
16#include <linux/delay.h>
17
18#include <trace/events/block.h>
19
20#include <linux/blk-mq.h>
21#include "blk.h"
22#include "blk-mq.h"
23#include "blk-mq-tag.h"
24
25static DEFINE_MUTEX(all_q_mutex);
26static LIST_HEAD(all_q_list);
27
28static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx);
29
30DEFINE_PER_CPU(struct llist_head, ipi_lists);
31
32static struct blk_mq_ctx *__blk_mq_get_ctx(struct request_queue *q,
33 unsigned int cpu)
34{
35 return per_cpu_ptr(q->queue_ctx, cpu);
36}
37
38/*
39 * This assumes per-cpu software queueing queues. They could be per-node
40 * as well, for instance. For now this is hardcoded as-is. Note that we don't
41 * care about preemption, since we know the ctx's are persistent. This does
42 * mean that we can't rely on ctx always matching the currently running CPU.
43 */
44static struct blk_mq_ctx *blk_mq_get_ctx(struct request_queue *q)
45{
46 return __blk_mq_get_ctx(q, get_cpu());
47}
48
49static void blk_mq_put_ctx(struct blk_mq_ctx *ctx)
50{
51 put_cpu();
52}
53
54/*
55 * Check if any of the ctx's have pending work in this hardware queue
56 */
57static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
58{
59 unsigned int i;
60
61 for (i = 0; i < hctx->nr_ctx_map; i++)
62 if (hctx->ctx_map[i])
63 return true;
64
65 return false;
66}
67
68/*
69 * Mark this ctx as having pending work in this hardware queue
70 */
71static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
72 struct blk_mq_ctx *ctx)
73{
74 if (!test_bit(ctx->index_hw, hctx->ctx_map))
75 set_bit(ctx->index_hw, hctx->ctx_map);
76}
77
78static struct request *blk_mq_alloc_rq(struct blk_mq_hw_ctx *hctx, gfp_t gfp,
79 bool reserved)
80{
81 struct request *rq;
82 unsigned int tag;
83
84 tag = blk_mq_get_tag(hctx->tags, gfp, reserved);
85 if (tag != BLK_MQ_TAG_FAIL) {
86 rq = hctx->rqs[tag];
87 rq->tag = tag;
88
89 return rq;
90 }
91
92 return NULL;
93}
94
95static int blk_mq_queue_enter(struct request_queue *q)
96{
97 int ret;
98
99 __percpu_counter_add(&q->mq_usage_counter, 1, 1000000);
100 smp_wmb();
101 /* we have problems to freeze the queue if it's initializing */
102 if (!blk_queue_bypass(q) || !blk_queue_init_done(q))
103 return 0;
104
105 __percpu_counter_add(&q->mq_usage_counter, -1, 1000000);
106
107 spin_lock_irq(q->queue_lock);
108 ret = wait_event_interruptible_lock_irq(q->mq_freeze_wq,
109 !blk_queue_bypass(q), *q->queue_lock);
110 /* inc usage with lock hold to avoid freeze_queue runs here */
111 if (!ret)
112 __percpu_counter_add(&q->mq_usage_counter, 1, 1000000);
113 spin_unlock_irq(q->queue_lock);
114
115 return ret;
116}
117
118static void blk_mq_queue_exit(struct request_queue *q)
119{
120 __percpu_counter_add(&q->mq_usage_counter, -1, 1000000);
121}
122
123/*
124 * Guarantee no request is in use, so we can change any data structure of
125 * the queue afterward.
126 */
127static void blk_mq_freeze_queue(struct request_queue *q)
128{
129 bool drain;
130
131 spin_lock_irq(q->queue_lock);
132 drain = !q->bypass_depth++;
133 queue_flag_set(QUEUE_FLAG_BYPASS, q);
134 spin_unlock_irq(q->queue_lock);
135
136 if (!drain)
137 return;
138
139 while (true) {
140 s64 count;
141
142 spin_lock_irq(q->queue_lock);
143 count = percpu_counter_sum(&q->mq_usage_counter);
144 spin_unlock_irq(q->queue_lock);
145
146 if (count == 0)
147 break;
148 blk_mq_run_queues(q, false);
149 msleep(10);
150 }
151}
152
153static void blk_mq_unfreeze_queue(struct request_queue *q)
154{
155 bool wake = false;
156
157 spin_lock_irq(q->queue_lock);
158 if (!--q->bypass_depth) {
159 queue_flag_clear(QUEUE_FLAG_BYPASS, q);
160 wake = true;
161 }
162 WARN_ON_ONCE(q->bypass_depth < 0);
163 spin_unlock_irq(q->queue_lock);
164 if (wake)
165 wake_up_all(&q->mq_freeze_wq);
166}
167
168bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
169{
170 return blk_mq_has_free_tags(hctx->tags);
171}
172EXPORT_SYMBOL(blk_mq_can_queue);
173
174static void blk_mq_rq_ctx_init(struct blk_mq_ctx *ctx, struct request *rq,
175 unsigned int rw_flags)
176{
177 rq->mq_ctx = ctx;
178 rq->cmd_flags = rw_flags;
179 ctx->rq_dispatched[rw_is_sync(rw_flags)]++;
180}
181
182static struct request *__blk_mq_alloc_request(struct blk_mq_hw_ctx *hctx,
183 gfp_t gfp, bool reserved)
184{
185 return blk_mq_alloc_rq(hctx, gfp, reserved);
186}
187
188static struct request *blk_mq_alloc_request_pinned(struct request_queue *q,
189 int rw, gfp_t gfp,
190 bool reserved)
191{
192 struct request *rq;
193
194 do {
195 struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
196 struct blk_mq_hw_ctx *hctx = q->mq_ops->map_queue(q, ctx->cpu);
197
198 rq = __blk_mq_alloc_request(hctx, gfp & ~__GFP_WAIT, reserved);
199 if (rq) {
200 blk_mq_rq_ctx_init(ctx, rq, rw);
201 break;
202 } else if (!(gfp & __GFP_WAIT))
203 break;
204
205 blk_mq_put_ctx(ctx);
206 __blk_mq_run_hw_queue(hctx);
207 blk_mq_wait_for_tags(hctx->tags);
208 } while (1);
209
210 return rq;
211}
212
213struct request *blk_mq_alloc_request(struct request_queue *q, int rw,
214 gfp_t gfp, bool reserved)
215{
216 struct request *rq;
217
218 if (blk_mq_queue_enter(q))
219 return NULL;
220
221 rq = blk_mq_alloc_request_pinned(q, rw, gfp, reserved);
222 blk_mq_put_ctx(rq->mq_ctx);
223 return rq;
224}
225
226struct request *blk_mq_alloc_reserved_request(struct request_queue *q, int rw,
227 gfp_t gfp)
228{
229 struct request *rq;
230
231 if (blk_mq_queue_enter(q))
232 return NULL;
233
234 rq = blk_mq_alloc_request_pinned(q, rw, gfp, true);
235 blk_mq_put_ctx(rq->mq_ctx);
236 return rq;
237}
238EXPORT_SYMBOL(blk_mq_alloc_reserved_request);
239
240/*
241 * Re-init and set pdu, if we have it
242 */
243static void blk_mq_rq_init(struct blk_mq_hw_ctx *hctx, struct request *rq)
244{
245 blk_rq_init(hctx->queue, rq);
246
247 if (hctx->cmd_size)
248 rq->special = blk_mq_rq_to_pdu(rq);
249}
250
251static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx,
252 struct blk_mq_ctx *ctx, struct request *rq)
253{
254 const int tag = rq->tag;
255 struct request_queue *q = rq->q;
256
257 blk_mq_rq_init(hctx, rq);
258 blk_mq_put_tag(hctx->tags, tag);
259
260 blk_mq_queue_exit(q);
261}
262
263void blk_mq_free_request(struct request *rq)
264{
265 struct blk_mq_ctx *ctx = rq->mq_ctx;
266 struct blk_mq_hw_ctx *hctx;
267 struct request_queue *q = rq->q;
268
269 ctx->rq_completed[rq_is_sync(rq)]++;
270
271 hctx = q->mq_ops->map_queue(q, ctx->cpu);
272 __blk_mq_free_request(hctx, ctx, rq);
273}
274
275static void blk_mq_bio_endio(struct request *rq, struct bio *bio, int error)
276{
277 if (error)
278 clear_bit(BIO_UPTODATE, &bio->bi_flags);
279 else if (!test_bit(BIO_UPTODATE, &bio->bi_flags))
280 error = -EIO;
281
282 if (unlikely(rq->cmd_flags & REQ_QUIET))
283 set_bit(BIO_QUIET, &bio->bi_flags);
284
285 /* don't actually finish bio if it's part of flush sequence */
286 if (!(rq->cmd_flags & REQ_FLUSH_SEQ))
287 bio_endio(bio, error);
288}
289
290void blk_mq_complete_request(struct request *rq, int error)
291{
292 struct bio *bio = rq->bio;
293 unsigned int bytes = 0;
294
295 trace_block_rq_complete(rq->q, rq);
296
297 while (bio) {
298 struct bio *next = bio->bi_next;
299
300 bio->bi_next = NULL;
301 bytes += bio->bi_size;
302 blk_mq_bio_endio(rq, bio, error);
303 bio = next;
304 }
305
306 blk_account_io_completion(rq, bytes);
307
308 if (rq->end_io)
309 rq->end_io(rq, error);
310 else
311 blk_mq_free_request(rq);
312
313 blk_account_io_done(rq);
314}
315
316void __blk_mq_end_io(struct request *rq, int error)
317{
318 if (!blk_mark_rq_complete(rq))
319 blk_mq_complete_request(rq, error);
320}
321
322#if defined(CONFIG_SMP) && defined(CONFIG_USE_GENERIC_SMP_HELPERS)
323
324/*
325 * Called with interrupts disabled.
326 */
327static void ipi_end_io(void *data)
328{
329 struct llist_head *list = &per_cpu(ipi_lists, smp_processor_id());
330 struct llist_node *entry, *next;
331 struct request *rq;
332
333 entry = llist_del_all(list);
334
335 while (entry) {
336 next = entry->next;
337 rq = llist_entry(entry, struct request, ll_list);
338 __blk_mq_end_io(rq, rq->errors);
339 entry = next;
340 }
341}
342
343static int ipi_remote_cpu(struct blk_mq_ctx *ctx, const int cpu,
344 struct request *rq, const int error)
345{
346 struct call_single_data *data = &rq->csd;
347
348 rq->errors = error;
349 rq->ll_list.next = NULL;
350
351 /*
352 * If the list is non-empty, an existing IPI must already
353 * be "in flight". If that is the case, we need not schedule
354 * a new one.
355 */
356 if (llist_add(&rq->ll_list, &per_cpu(ipi_lists, ctx->cpu))) {
357 data->func = ipi_end_io;
358 data->flags = 0;
359 __smp_call_function_single(ctx->cpu, data, 0);
360 }
361
362 return true;
363}
364#else /* CONFIG_SMP && CONFIG_USE_GENERIC_SMP_HELPERS */
365static int ipi_remote_cpu(struct blk_mq_ctx *ctx, const int cpu,
366 struct request *rq, const int error)
367{
368 return false;
369}
370#endif
371
372/*
373 * End IO on this request on a multiqueue enabled driver. We'll either do
374 * it directly inline, or punt to a local IPI handler on the matching
375 * remote CPU.
376 */
377void blk_mq_end_io(struct request *rq, int error)
378{
379 struct blk_mq_ctx *ctx = rq->mq_ctx;
380 int cpu;
381
382 if (!ctx->ipi_redirect)
383 return __blk_mq_end_io(rq, error);
384
385 cpu = get_cpu();
386
387 if (cpu == ctx->cpu || !cpu_online(ctx->cpu) ||
388 !ipi_remote_cpu(ctx, cpu, rq, error))
389 __blk_mq_end_io(rq, error);
390
391 put_cpu();
392}
393EXPORT_SYMBOL(blk_mq_end_io);
394
395static void blk_mq_start_request(struct request *rq)
396{
397 struct request_queue *q = rq->q;
398
399 trace_block_rq_issue(q, rq);
400
401 /*
402 * Just mark start time and set the started bit. Due to memory
403 * ordering, we know we'll see the correct deadline as long as
404 * REQ_ATOMIC_STARTED is seen.
405 */
406 rq->deadline = jiffies + q->rq_timeout;
407 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
408}
409
410static void blk_mq_requeue_request(struct request *rq)
411{
412 struct request_queue *q = rq->q;
413
414 trace_block_rq_requeue(q, rq);
415 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
416}
417
418struct blk_mq_timeout_data {
419 struct blk_mq_hw_ctx *hctx;
420 unsigned long *next;
421 unsigned int *next_set;
422};
423
424static void blk_mq_timeout_check(void *__data, unsigned long *free_tags)
425{
426 struct blk_mq_timeout_data *data = __data;
427 struct blk_mq_hw_ctx *hctx = data->hctx;
428 unsigned int tag;
429
430 /* It may not be in flight yet (this is where
431 * the REQ_ATOMIC_STARTED flag comes in). The requests are
432 * statically allocated, so we know it's always safe to access the
433 * memory associated with a bit offset into ->rqs[].
434 */
435 tag = 0;
436 do {
437 struct request *rq;
438
439 tag = find_next_zero_bit(free_tags, hctx->queue_depth, tag);
440 if (tag >= hctx->queue_depth)
441 break;
442
443 rq = hctx->rqs[tag++];
444
445 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
446 continue;
447
448 blk_rq_check_expired(rq, data->next, data->next_set);
449 } while (1);
450}
451
452static void blk_mq_hw_ctx_check_timeout(struct blk_mq_hw_ctx *hctx,
453 unsigned long *next,
454 unsigned int *next_set)
455{
456 struct blk_mq_timeout_data data = {
457 .hctx = hctx,
458 .next = next,
459 .next_set = next_set,
460 };
461
462 /*
463 * Ask the tagging code to iterate busy requests, so we can
464 * check them for timeout.
465 */
466 blk_mq_tag_busy_iter(hctx->tags, blk_mq_timeout_check, &data);
467}
468
469static void blk_mq_rq_timer(unsigned long data)
470{
471 struct request_queue *q = (struct request_queue *) data;
472 struct blk_mq_hw_ctx *hctx;
473 unsigned long next = 0;
474 int i, next_set = 0;
475
476 queue_for_each_hw_ctx(q, hctx, i)
477 blk_mq_hw_ctx_check_timeout(hctx, &next, &next_set);
478
479 if (next_set)
480 mod_timer(&q->timeout, round_jiffies_up(next));
481}
482
483/*
484 * Reverse check our software queue for entries that we could potentially
485 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
486 * too much time checking for merges.
487 */
488static bool blk_mq_attempt_merge(struct request_queue *q,
489 struct blk_mq_ctx *ctx, struct bio *bio)
490{
491 struct request *rq;
492 int checked = 8;
493
494 list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
495 int el_ret;
496
497 if (!checked--)
498 break;
499
500 if (!blk_rq_merge_ok(rq, bio))
501 continue;
502
503 el_ret = blk_try_merge(rq, bio);
504 if (el_ret == ELEVATOR_BACK_MERGE) {
505 if (bio_attempt_back_merge(q, rq, bio)) {
506 ctx->rq_merged++;
507 return true;
508 }
509 break;
510 } else if (el_ret == ELEVATOR_FRONT_MERGE) {
511 if (bio_attempt_front_merge(q, rq, bio)) {
512 ctx->rq_merged++;
513 return true;
514 }
515 break;
516 }
517 }
518
519 return false;
520}
521
522void blk_mq_add_timer(struct request *rq)
523{
524 __blk_add_timer(rq, NULL);
525}
526
527/*
528 * Run this hardware queue, pulling any software queues mapped to it in.
529 * Note that this function currently has various problems around ordering
530 * of IO. In particular, we'd like FIFO behaviour on handling existing
531 * items on the hctx->dispatch list. Ignore that for now.
532 */
533static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
534{
535 struct request_queue *q = hctx->queue;
536 struct blk_mq_ctx *ctx;
537 struct request *rq;
538 LIST_HEAD(rq_list);
539 int bit, queued;
540
541 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->flags)))
542 return;
543
544 hctx->run++;
545
546 /*
547 * Touch any software queue that has pending entries.
548 */
549 for_each_set_bit(bit, hctx->ctx_map, hctx->nr_ctx) {
550 clear_bit(bit, hctx->ctx_map);
551 ctx = hctx->ctxs[bit];
552 BUG_ON(bit != ctx->index_hw);
553
554 spin_lock(&ctx->lock);
555 list_splice_tail_init(&ctx->rq_list, &rq_list);
556 spin_unlock(&ctx->lock);
557 }
558
559 /*
560 * If we have previous entries on our dispatch list, grab them
561 * and stuff them at the front for more fair dispatch.
562 */
563 if (!list_empty_careful(&hctx->dispatch)) {
564 spin_lock(&hctx->lock);
565 if (!list_empty(&hctx->dispatch))
566 list_splice_init(&hctx->dispatch, &rq_list);
567 spin_unlock(&hctx->lock);
568 }
569
570 /*
571 * Delete and return all entries from our dispatch list
572 */
573 queued = 0;
574
575 /*
576 * Now process all the entries, sending them to the driver.
577 */
578 while (!list_empty(&rq_list)) {
579 int ret;
580
581 rq = list_first_entry(&rq_list, struct request, queuelist);
582 list_del_init(&rq->queuelist);
583 blk_mq_start_request(rq);
584
585 /*
586 * Last request in the series. Flag it as such, this
587 * enables drivers to know when IO should be kicked off,
588 * if they don't do it on a per-request basis.
589 *
590 * Note: the flag isn't the only condition drivers
591 * should do kick off. If drive is busy, the last
592 * request might not have the bit set.
593 */
594 if (list_empty(&rq_list))
595 rq->cmd_flags |= REQ_END;
596
597 ret = q->mq_ops->queue_rq(hctx, rq);
598 switch (ret) {
599 case BLK_MQ_RQ_QUEUE_OK:
600 queued++;
601 continue;
602 case BLK_MQ_RQ_QUEUE_BUSY:
603 /*
604 * FIXME: we should have a mechanism to stop the queue
605 * like blk_stop_queue, otherwise we will waste cpu
606 * time
607 */
608 list_add(&rq->queuelist, &rq_list);
609 blk_mq_requeue_request(rq);
610 break;
611 default:
612 pr_err("blk-mq: bad return on queue: %d\n", ret);
613 rq->errors = -EIO;
614 case BLK_MQ_RQ_QUEUE_ERROR:
615 blk_mq_end_io(rq, rq->errors);
616 break;
617 }
618
619 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
620 break;
621 }
622
623 if (!queued)
624 hctx->dispatched[0]++;
625 else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1)))
626 hctx->dispatched[ilog2(queued) + 1]++;
627
628 /*
629 * Any items that need requeuing? Stuff them into hctx->dispatch,
630 * that is where we will continue on next queue run.
631 */
632 if (!list_empty(&rq_list)) {
633 spin_lock(&hctx->lock);
634 list_splice(&rq_list, &hctx->dispatch);
635 spin_unlock(&hctx->lock);
636 }
637}
638
639void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
640{
641 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->flags)))
642 return;
643
644 if (!async)
645 __blk_mq_run_hw_queue(hctx);
646 else {
647 struct request_queue *q = hctx->queue;
648
649 kblockd_schedule_delayed_work(q, &hctx->delayed_work, 0);
650 }
651}
652
653void blk_mq_run_queues(struct request_queue *q, bool async)
654{
655 struct blk_mq_hw_ctx *hctx;
656 int i;
657
658 queue_for_each_hw_ctx(q, hctx, i) {
659 if ((!blk_mq_hctx_has_pending(hctx) &&
660 list_empty_careful(&hctx->dispatch)) ||
661 test_bit(BLK_MQ_S_STOPPED, &hctx->flags))
662 continue;
663
664 blk_mq_run_hw_queue(hctx, async);
665 }
666}
667EXPORT_SYMBOL(blk_mq_run_queues);
668
669void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
670{
671 cancel_delayed_work(&hctx->delayed_work);
672 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
673}
674EXPORT_SYMBOL(blk_mq_stop_hw_queue);
675
676void blk_mq_stop_hw_queues(struct request_queue *q)
677{
678 struct blk_mq_hw_ctx *hctx;
679 int i;
680
681 queue_for_each_hw_ctx(q, hctx, i)
682 blk_mq_stop_hw_queue(hctx);
683}
684EXPORT_SYMBOL(blk_mq_stop_hw_queues);
685
686void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
687{
688 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
689 __blk_mq_run_hw_queue(hctx);
690}
691EXPORT_SYMBOL(blk_mq_start_hw_queue);
692
693void blk_mq_start_stopped_hw_queues(struct request_queue *q)
694{
695 struct blk_mq_hw_ctx *hctx;
696 int i;
697
698 queue_for_each_hw_ctx(q, hctx, i) {
699 if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state))
700 continue;
701
702 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
703 blk_mq_run_hw_queue(hctx, true);
704 }
705}
706EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
707
708static void blk_mq_work_fn(struct work_struct *work)
709{
710 struct blk_mq_hw_ctx *hctx;
711
712 hctx = container_of(work, struct blk_mq_hw_ctx, delayed_work.work);
713 __blk_mq_run_hw_queue(hctx);
714}
715
716static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx,
717 struct request *rq)
718{
719 struct blk_mq_ctx *ctx = rq->mq_ctx;
720
721 list_add_tail(&rq->queuelist, &ctx->rq_list);
722 blk_mq_hctx_mark_pending(hctx, ctx);
723
724 /*
725 * We do this early, to ensure we are on the right CPU.
726 */
727 blk_mq_add_timer(rq);
728}
729
730void blk_mq_insert_request(struct request_queue *q, struct request *rq,
731 bool run_queue)
732{
733 struct blk_mq_hw_ctx *hctx;
734 struct blk_mq_ctx *ctx, *current_ctx;
735
736 ctx = rq->mq_ctx;
737 hctx = q->mq_ops->map_queue(q, ctx->cpu);
738
739 if (rq->cmd_flags & (REQ_FLUSH | REQ_FUA)) {
740 blk_insert_flush(rq);
741 } else {
742 current_ctx = blk_mq_get_ctx(q);
743
744 if (!cpu_online(ctx->cpu)) {
745 ctx = current_ctx;
746 hctx = q->mq_ops->map_queue(q, ctx->cpu);
747 rq->mq_ctx = ctx;
748 }
749 spin_lock(&ctx->lock);
750 __blk_mq_insert_request(hctx, rq);
751 spin_unlock(&ctx->lock);
752
753 blk_mq_put_ctx(current_ctx);
754 }
755
756 if (run_queue)
757 __blk_mq_run_hw_queue(hctx);
758}
759EXPORT_SYMBOL(blk_mq_insert_request);
760
761/*
762 * This is a special version of blk_mq_insert_request to bypass FLUSH request
763 * check. Should only be used internally.
764 */
765void blk_mq_run_request(struct request *rq, bool run_queue, bool async)
766{
767 struct request_queue *q = rq->q;
768 struct blk_mq_hw_ctx *hctx;
769 struct blk_mq_ctx *ctx, *current_ctx;
770
771 current_ctx = blk_mq_get_ctx(q);
772
773 ctx = rq->mq_ctx;
774 if (!cpu_online(ctx->cpu)) {
775 ctx = current_ctx;
776 rq->mq_ctx = ctx;
777 }
778 hctx = q->mq_ops->map_queue(q, ctx->cpu);
779
780 /* ctx->cpu might be offline */
781 spin_lock(&ctx->lock);
782 __blk_mq_insert_request(hctx, rq);
783 spin_unlock(&ctx->lock);
784
785 blk_mq_put_ctx(current_ctx);
786
787 if (run_queue)
788 blk_mq_run_hw_queue(hctx, async);
789}
790
791static void blk_mq_insert_requests(struct request_queue *q,
792 struct blk_mq_ctx *ctx,
793 struct list_head *list,
794 int depth,
795 bool from_schedule)
796
797{
798 struct blk_mq_hw_ctx *hctx;
799 struct blk_mq_ctx *current_ctx;
800
801 trace_block_unplug(q, depth, !from_schedule);
802
803 current_ctx = blk_mq_get_ctx(q);
804
805 if (!cpu_online(ctx->cpu))
806 ctx = current_ctx;
807 hctx = q->mq_ops->map_queue(q, ctx->cpu);
808
809 /*
810 * preemption doesn't flush plug list, so it's possible ctx->cpu is
811 * offline now
812 */
813 spin_lock(&ctx->lock);
814 while (!list_empty(list)) {
815 struct request *rq;
816
817 rq = list_first_entry(list, struct request, queuelist);
818 list_del_init(&rq->queuelist);
819 rq->mq_ctx = ctx;
820 __blk_mq_insert_request(hctx, rq);
821 }
822 spin_unlock(&ctx->lock);
823
824 blk_mq_put_ctx(current_ctx);
825
826 blk_mq_run_hw_queue(hctx, from_schedule);
827}
828
829static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
830{
831 struct request *rqa = container_of(a, struct request, queuelist);
832 struct request *rqb = container_of(b, struct request, queuelist);
833
834 return !(rqa->mq_ctx < rqb->mq_ctx ||
835 (rqa->mq_ctx == rqb->mq_ctx &&
836 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
837}
838
839void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
840{
841 struct blk_mq_ctx *this_ctx;
842 struct request_queue *this_q;
843 struct request *rq;
844 LIST_HEAD(list);
845 LIST_HEAD(ctx_list);
846 unsigned int depth;
847
848 list_splice_init(&plug->mq_list, &list);
849
850 list_sort(NULL, &list, plug_ctx_cmp);
851
852 this_q = NULL;
853 this_ctx = NULL;
854 depth = 0;
855
856 while (!list_empty(&list)) {
857 rq = list_entry_rq(list.next);
858 list_del_init(&rq->queuelist);
859 BUG_ON(!rq->q);
860 if (rq->mq_ctx != this_ctx) {
861 if (this_ctx) {
862 blk_mq_insert_requests(this_q, this_ctx,
863 &ctx_list, depth,
864 from_schedule);
865 }
866
867 this_ctx = rq->mq_ctx;
868 this_q = rq->q;
869 depth = 0;
870 }
871
872 depth++;
873 list_add_tail(&rq->queuelist, &ctx_list);
874 }
875
876 /*
877 * If 'this_ctx' is set, we know we have entries to complete
878 * on 'ctx_list'. Do those.
879 */
880 if (this_ctx) {
881 blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
882 from_schedule);
883 }
884}
885
886static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
887{
888 init_request_from_bio(rq, bio);
889 blk_account_io_start(rq, 1);
890}
891
892static void blk_mq_make_request(struct request_queue *q, struct bio *bio)
893{
894 struct blk_mq_hw_ctx *hctx;
895 struct blk_mq_ctx *ctx;
896 const int is_sync = rw_is_sync(bio->bi_rw);
897 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
898 int rw = bio_data_dir(bio);
899 struct request *rq;
900 unsigned int use_plug, request_count = 0;
901
902 /*
903 * If we have multiple hardware queues, just go directly to
904 * one of those for sync IO.
905 */
906 use_plug = !is_flush_fua && ((q->nr_hw_queues == 1) || !is_sync);
907
908 blk_queue_bounce(q, &bio);
909
910 if (use_plug && blk_attempt_plug_merge(q, bio, &request_count))
911 return;
912
913 if (blk_mq_queue_enter(q)) {
914 bio_endio(bio, -EIO);
915 return;
916 }
917
918 ctx = blk_mq_get_ctx(q);
919 hctx = q->mq_ops->map_queue(q, ctx->cpu);
920
921 trace_block_getrq(q, bio, rw);
922 rq = __blk_mq_alloc_request(hctx, GFP_ATOMIC, false);
923 if (likely(rq))
924 blk_mq_rq_ctx_init(ctx, rq, rw);
925 else {
926 blk_mq_put_ctx(ctx);
927 trace_block_sleeprq(q, bio, rw);
928 rq = blk_mq_alloc_request_pinned(q, rw, __GFP_WAIT|GFP_ATOMIC,
929 false);
930 ctx = rq->mq_ctx;
931 hctx = q->mq_ops->map_queue(q, ctx->cpu);
932 }
933
934 hctx->queued++;
935
936 if (unlikely(is_flush_fua)) {
937 blk_mq_bio_to_request(rq, bio);
938 blk_mq_put_ctx(ctx);
939 blk_insert_flush(rq);
940 goto run_queue;
941 }
942
943 /*
944 * A task plug currently exists. Since this is completely lockless,
945 * utilize that to temporarily store requests until the task is
946 * either done or scheduled away.
947 */
948 if (use_plug) {
949 struct blk_plug *plug = current->plug;
950
951 if (plug) {
952 blk_mq_bio_to_request(rq, bio);
953 if (list_empty(&plug->mq_list))
954 trace_block_plug(q);
955 else if (request_count >= BLK_MAX_REQUEST_COUNT) {
956 blk_flush_plug_list(plug, false);
957 trace_block_plug(q);
958 }
959 list_add_tail(&rq->queuelist, &plug->mq_list);
960 blk_mq_put_ctx(ctx);
961 return;
962 }
963 }
964
965 spin_lock(&ctx->lock);
966
967 if ((hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
968 blk_mq_attempt_merge(q, ctx, bio))
969 __blk_mq_free_request(hctx, ctx, rq);
970 else {
971 blk_mq_bio_to_request(rq, bio);
972 __blk_mq_insert_request(hctx, rq);
973 }
974
975 spin_unlock(&ctx->lock);
976 blk_mq_put_ctx(ctx);
977
978 /*
979 * For a SYNC request, send it to the hardware immediately. For an
980 * ASYNC request, just ensure that we run it later on. The latter
981 * allows for merging opportunities and more efficient dispatching.
982 */
983run_queue:
984 blk_mq_run_hw_queue(hctx, !is_sync || is_flush_fua);
985}
986
987/*
988 * Default mapping to a software queue, since we use one per CPU.
989 */
990struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu)
991{
992 return q->queue_hw_ctx[q->mq_map[cpu]];
993}
994EXPORT_SYMBOL(blk_mq_map_queue);
995
996struct blk_mq_hw_ctx *blk_mq_alloc_single_hw_queue(struct blk_mq_reg *reg,
997 unsigned int hctx_index)
998{
999 return kmalloc_node(sizeof(struct blk_mq_hw_ctx),
1000 GFP_KERNEL | __GFP_ZERO, reg->numa_node);
1001}
1002EXPORT_SYMBOL(blk_mq_alloc_single_hw_queue);
1003
1004void blk_mq_free_single_hw_queue(struct blk_mq_hw_ctx *hctx,
1005 unsigned int hctx_index)
1006{
1007 kfree(hctx);
1008}
1009EXPORT_SYMBOL(blk_mq_free_single_hw_queue);
1010
1011static void blk_mq_hctx_notify(void *data, unsigned long action,
1012 unsigned int cpu)
1013{
1014 struct blk_mq_hw_ctx *hctx = data;
1015 struct blk_mq_ctx *ctx;
1016 LIST_HEAD(tmp);
1017
1018 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
1019 return;
1020
1021 /*
1022 * Move ctx entries to new CPU, if this one is going away.
1023 */
1024 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
1025
1026 spin_lock(&ctx->lock);
1027 if (!list_empty(&ctx->rq_list)) {
1028 list_splice_init(&ctx->rq_list, &tmp);
1029 clear_bit(ctx->index_hw, hctx->ctx_map);
1030 }
1031 spin_unlock(&ctx->lock);
1032
1033 if (list_empty(&tmp))
1034 return;
1035
1036 ctx = blk_mq_get_ctx(hctx->queue);
1037 spin_lock(&ctx->lock);
1038
1039 while (!list_empty(&tmp)) {
1040 struct request *rq;
1041
1042 rq = list_first_entry(&tmp, struct request, queuelist);
1043 rq->mq_ctx = ctx;
1044 list_move_tail(&rq->queuelist, &ctx->rq_list);
1045 }
1046
1047 blk_mq_hctx_mark_pending(hctx, ctx);
1048
1049 spin_unlock(&ctx->lock);
1050 blk_mq_put_ctx(ctx);
1051}
1052
1053static void blk_mq_init_hw_commands(struct blk_mq_hw_ctx *hctx,
1054 void (*init)(void *, struct blk_mq_hw_ctx *,
1055 struct request *, unsigned int),
1056 void *data)
1057{
1058 unsigned int i;
1059
1060 for (i = 0; i < hctx->queue_depth; i++) {
1061 struct request *rq = hctx->rqs[i];
1062
1063 init(data, hctx, rq, i);
1064 }
1065}
1066
1067void blk_mq_init_commands(struct request_queue *q,
1068 void (*init)(void *, struct blk_mq_hw_ctx *,
1069 struct request *, unsigned int),
1070 void *data)
1071{
1072 struct blk_mq_hw_ctx *hctx;
1073 unsigned int i;
1074
1075 queue_for_each_hw_ctx(q, hctx, i)
1076 blk_mq_init_hw_commands(hctx, init, data);
1077}
1078EXPORT_SYMBOL(blk_mq_init_commands);
1079
1080static void blk_mq_free_rq_map(struct blk_mq_hw_ctx *hctx)
1081{
1082 struct page *page;
1083
1084 while (!list_empty(&hctx->page_list)) {
1085 page = list_first_entry(&hctx->page_list, struct page, list);
1086 list_del_init(&page->list);
1087 __free_pages(page, page->private);
1088 }
1089
1090 kfree(hctx->rqs);
1091
1092 if (hctx->tags)
1093 blk_mq_free_tags(hctx->tags);
1094}
1095
1096static size_t order_to_size(unsigned int order)
1097{
1098 size_t ret = PAGE_SIZE;
1099
1100 while (order--)
1101 ret *= 2;
1102
1103 return ret;
1104}
1105
1106static int blk_mq_init_rq_map(struct blk_mq_hw_ctx *hctx,
1107 unsigned int reserved_tags, int node)
1108{
1109 unsigned int i, j, entries_per_page, max_order = 4;
1110 size_t rq_size, left;
1111
1112 INIT_LIST_HEAD(&hctx->page_list);
1113
1114 hctx->rqs = kmalloc_node(hctx->queue_depth * sizeof(struct request *),
1115 GFP_KERNEL, node);
1116 if (!hctx->rqs)
1117 return -ENOMEM;
1118
1119 /*
1120 * rq_size is the size of the request plus driver payload, rounded
1121 * to the cacheline size
1122 */
1123 rq_size = round_up(sizeof(struct request) + hctx->cmd_size,
1124 cache_line_size());
1125 left = rq_size * hctx->queue_depth;
1126
1127 for (i = 0; i < hctx->queue_depth;) {
1128 int this_order = max_order;
1129 struct page *page;
1130 int to_do;
1131 void *p;
1132
1133 while (left < order_to_size(this_order - 1) && this_order)
1134 this_order--;
1135
1136 do {
1137 page = alloc_pages_node(node, GFP_KERNEL, this_order);
1138 if (page)
1139 break;
1140 if (!this_order--)
1141 break;
1142 if (order_to_size(this_order) < rq_size)
1143 break;
1144 } while (1);
1145
1146 if (!page)
1147 break;
1148
1149 page->private = this_order;
1150 list_add_tail(&page->list, &hctx->page_list);
1151
1152 p = page_address(page);
1153 entries_per_page = order_to_size(this_order) / rq_size;
1154 to_do = min(entries_per_page, hctx->queue_depth - i);
1155 left -= to_do * rq_size;
1156 for (j = 0; j < to_do; j++) {
1157 hctx->rqs[i] = p;
1158 blk_mq_rq_init(hctx, hctx->rqs[i]);
1159 p += rq_size;
1160 i++;
1161 }
1162 }
1163
1164 if (i < (reserved_tags + BLK_MQ_TAG_MIN))
1165 goto err_rq_map;
1166 else if (i != hctx->queue_depth) {
1167 hctx->queue_depth = i;
1168 pr_warn("%s: queue depth set to %u because of low memory\n",
1169 __func__, i);
1170 }
1171
1172 hctx->tags = blk_mq_init_tags(hctx->queue_depth, reserved_tags, node);
1173 if (!hctx->tags) {
1174err_rq_map:
1175 blk_mq_free_rq_map(hctx);
1176 return -ENOMEM;
1177 }
1178
1179 return 0;
1180}
1181
1182static int blk_mq_init_hw_queues(struct request_queue *q,
1183 struct blk_mq_reg *reg, void *driver_data)
1184{
1185 struct blk_mq_hw_ctx *hctx;
1186 unsigned int i, j;
1187
1188 /*
1189 * Initialize hardware queues
1190 */
1191 queue_for_each_hw_ctx(q, hctx, i) {
1192 unsigned int num_maps;
1193 int node;
1194
1195 node = hctx->numa_node;
1196 if (node == NUMA_NO_NODE)
1197 node = hctx->numa_node = reg->numa_node;
1198
1199 INIT_DELAYED_WORK(&hctx->delayed_work, blk_mq_work_fn);
1200 spin_lock_init(&hctx->lock);
1201 INIT_LIST_HEAD(&hctx->dispatch);
1202 hctx->queue = q;
1203 hctx->queue_num = i;
1204 hctx->flags = reg->flags;
1205 hctx->queue_depth = reg->queue_depth;
1206 hctx->cmd_size = reg->cmd_size;
1207
1208 blk_mq_init_cpu_notifier(&hctx->cpu_notifier,
1209 blk_mq_hctx_notify, hctx);
1210 blk_mq_register_cpu_notifier(&hctx->cpu_notifier);
1211
1212 if (blk_mq_init_rq_map(hctx, reg->reserved_tags, node))
1213 break;
1214
1215 /*
1216 * Allocate space for all possible cpus to avoid allocation in
1217 * runtime
1218 */
1219 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1220 GFP_KERNEL, node);
1221 if (!hctx->ctxs)
1222 break;
1223
1224 num_maps = ALIGN(nr_cpu_ids, BITS_PER_LONG) / BITS_PER_LONG;
1225 hctx->ctx_map = kzalloc_node(num_maps * sizeof(unsigned long),
1226 GFP_KERNEL, node);
1227 if (!hctx->ctx_map)
1228 break;
1229
1230 hctx->nr_ctx_map = num_maps;
1231 hctx->nr_ctx = 0;
1232
1233 if (reg->ops->init_hctx &&
1234 reg->ops->init_hctx(hctx, driver_data, i))
1235 break;
1236 }
1237
1238 if (i == q->nr_hw_queues)
1239 return 0;
1240
1241 /*
1242 * Init failed
1243 */
1244 queue_for_each_hw_ctx(q, hctx, j) {
1245 if (i == j)
1246 break;
1247
1248 if (reg->ops->exit_hctx)
1249 reg->ops->exit_hctx(hctx, j);
1250
1251 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1252 blk_mq_free_rq_map(hctx);
1253 kfree(hctx->ctxs);
1254 }
1255
1256 return 1;
1257}
1258
1259static void blk_mq_init_cpu_queues(struct request_queue *q,
1260 unsigned int nr_hw_queues)
1261{
1262 unsigned int i;
1263
1264 for_each_possible_cpu(i) {
1265 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1266 struct blk_mq_hw_ctx *hctx;
1267
1268 memset(__ctx, 0, sizeof(*__ctx));
1269 __ctx->cpu = i;
1270 spin_lock_init(&__ctx->lock);
1271 INIT_LIST_HEAD(&__ctx->rq_list);
1272 __ctx->queue = q;
1273
1274 /* If the cpu isn't online, the cpu is mapped to first hctx */
1275 hctx = q->mq_ops->map_queue(q, i);
1276 hctx->nr_ctx++;
1277
1278 if (!cpu_online(i))
1279 continue;
1280
1281 /*
1282 * Set local node, IFF we have more than one hw queue. If
1283 * not, we remain on the home node of the device
1284 */
1285 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1286 hctx->numa_node = cpu_to_node(i);
1287 }
1288}
1289
1290static void blk_mq_map_swqueue(struct request_queue *q)
1291{
1292 unsigned int i;
1293 struct blk_mq_hw_ctx *hctx;
1294 struct blk_mq_ctx *ctx;
1295
1296 queue_for_each_hw_ctx(q, hctx, i) {
1297 hctx->nr_ctx = 0;
1298 }
1299
1300 /*
1301 * Map software to hardware queues
1302 */
1303 queue_for_each_ctx(q, ctx, i) {
1304 /* If the cpu isn't online, the cpu is mapped to first hctx */
1305 hctx = q->mq_ops->map_queue(q, i);
1306 ctx->index_hw = hctx->nr_ctx;
1307 hctx->ctxs[hctx->nr_ctx++] = ctx;
1308 }
1309}
1310
1311struct request_queue *blk_mq_init_queue(struct blk_mq_reg *reg,
1312 void *driver_data)
1313{
1314 struct blk_mq_hw_ctx **hctxs;
1315 struct blk_mq_ctx *ctx;
1316 struct request_queue *q;
1317 int i;
1318
1319 if (!reg->nr_hw_queues ||
1320 !reg->ops->queue_rq || !reg->ops->map_queue ||
1321 !reg->ops->alloc_hctx || !reg->ops->free_hctx)
1322 return ERR_PTR(-EINVAL);
1323
1324 if (!reg->queue_depth)
1325 reg->queue_depth = BLK_MQ_MAX_DEPTH;
1326 else if (reg->queue_depth > BLK_MQ_MAX_DEPTH) {
1327 pr_err("blk-mq: queuedepth too large (%u)\n", reg->queue_depth);
1328 reg->queue_depth = BLK_MQ_MAX_DEPTH;
1329 }
1330
1331 /*
1332 * Set aside a tag for flush requests. It will only be used while
1333 * another flush request is in progress but outside the driver.
1334 *
1335 * TODO: only allocate if flushes are supported
1336 */
1337 reg->queue_depth++;
1338 reg->reserved_tags++;
1339
1340 if (reg->queue_depth < (reg->reserved_tags + BLK_MQ_TAG_MIN))
1341 return ERR_PTR(-EINVAL);
1342
1343 ctx = alloc_percpu(struct blk_mq_ctx);
1344 if (!ctx)
1345 return ERR_PTR(-ENOMEM);
1346
1347 hctxs = kmalloc_node(reg->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL,
1348 reg->numa_node);
1349
1350 if (!hctxs)
1351 goto err_percpu;
1352
1353 for (i = 0; i < reg->nr_hw_queues; i++) {
1354 hctxs[i] = reg->ops->alloc_hctx(reg, i);
1355 if (!hctxs[i])
1356 goto err_hctxs;
1357
1358 hctxs[i]->numa_node = NUMA_NO_NODE;
1359 hctxs[i]->queue_num = i;
1360 }
1361
1362 q = blk_alloc_queue_node(GFP_KERNEL, reg->numa_node);
1363 if (!q)
1364 goto err_hctxs;
1365
1366 q->mq_map = blk_mq_make_queue_map(reg);
1367 if (!q->mq_map)
1368 goto err_map;
1369
1370 setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q);
1371 blk_queue_rq_timeout(q, 30000);
1372
1373 q->nr_queues = nr_cpu_ids;
1374 q->nr_hw_queues = reg->nr_hw_queues;
1375
1376 q->queue_ctx = ctx;
1377 q->queue_hw_ctx = hctxs;
1378
1379 q->mq_ops = reg->ops;
1380
1381 blk_queue_make_request(q, blk_mq_make_request);
1382 blk_queue_rq_timed_out(q, reg->ops->timeout);
1383 if (reg->timeout)
1384 blk_queue_rq_timeout(q, reg->timeout);
1385
1386 blk_mq_init_flush(q);
1387 blk_mq_init_cpu_queues(q, reg->nr_hw_queues);
1388
1389 if (blk_mq_init_hw_queues(q, reg, driver_data))
1390 goto err_hw;
1391
1392 blk_mq_map_swqueue(q);
1393
1394 mutex_lock(&all_q_mutex);
1395 list_add_tail(&q->all_q_node, &all_q_list);
1396 mutex_unlock(&all_q_mutex);
1397
1398 return q;
1399err_hw:
1400 kfree(q->mq_map);
1401err_map:
1402 blk_cleanup_queue(q);
1403err_hctxs:
1404 for (i = 0; i < reg->nr_hw_queues; i++) {
1405 if (!hctxs[i])
1406 break;
1407 reg->ops->free_hctx(hctxs[i], i);
1408 }
1409 kfree(hctxs);
1410err_percpu:
1411 free_percpu(ctx);
1412 return ERR_PTR(-ENOMEM);
1413}
1414EXPORT_SYMBOL(blk_mq_init_queue);
1415
1416void blk_mq_free_queue(struct request_queue *q)
1417{
1418 struct blk_mq_hw_ctx *hctx;
1419 int i;
1420
1421 queue_for_each_hw_ctx(q, hctx, i) {
1422 cancel_delayed_work_sync(&hctx->delayed_work);
1423 kfree(hctx->ctx_map);
1424 kfree(hctx->ctxs);
1425 blk_mq_free_rq_map(hctx);
1426 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1427 if (q->mq_ops->exit_hctx)
1428 q->mq_ops->exit_hctx(hctx, i);
1429 q->mq_ops->free_hctx(hctx, i);
1430 }
1431
1432 free_percpu(q->queue_ctx);
1433 kfree(q->queue_hw_ctx);
1434 kfree(q->mq_map);
1435
1436 q->queue_ctx = NULL;
1437 q->queue_hw_ctx = NULL;
1438 q->mq_map = NULL;
1439
1440 mutex_lock(&all_q_mutex);
1441 list_del_init(&q->all_q_node);
1442 mutex_unlock(&all_q_mutex);
1443}
1444EXPORT_SYMBOL(blk_mq_free_queue);
1445
1446/* Basically redo blk_mq_init_queue with queue frozen */
1447static void __cpuinit blk_mq_queue_reinit(struct request_queue *q)
1448{
1449 blk_mq_freeze_queue(q);
1450
1451 blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues);
1452
1453 /*
1454 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
1455 * we should change hctx numa_node according to new topology (this
1456 * involves free and re-allocate memory, worthy doing?)
1457 */
1458
1459 blk_mq_map_swqueue(q);
1460
1461 blk_mq_unfreeze_queue(q);
1462}
1463
1464static int __cpuinit blk_mq_queue_reinit_notify(struct notifier_block *nb,
1465 unsigned long action, void *hcpu)
1466{
1467 struct request_queue *q;
1468
1469 /*
1470 * Before new mapping is established, hotadded cpu might already start
1471 * handling requests. This doesn't break anything as we map offline
1472 * CPUs to first hardware queue. We will re-init queue below to get
1473 * optimal settings.
1474 */
1475 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN &&
1476 action != CPU_ONLINE && action != CPU_ONLINE_FROZEN)
1477 return NOTIFY_OK;
1478
1479 mutex_lock(&all_q_mutex);
1480 list_for_each_entry(q, &all_q_list, all_q_node)
1481 blk_mq_queue_reinit(q);
1482 mutex_unlock(&all_q_mutex);
1483 return NOTIFY_OK;
1484}
1485
1486static int __init blk_mq_init(void)
1487{
1488 unsigned int i;
1489
1490 for_each_possible_cpu(i)
1491 init_llist_head(&per_cpu(ipi_lists, i));
1492
1493 blk_mq_cpu_init();
1494
1495 /* Must be called after percpu_counter_hotcpu_callback() */
1496 hotcpu_notifier(blk_mq_queue_reinit_notify, -10);
1497
1498 return 0;
1499}
1500subsys_initcall(blk_mq_init);
diff --git a/block/blk-mq.h b/block/blk-mq.h
new file mode 100644
index 000000000000..52bf1f96a2c2
--- /dev/null
+++ b/block/blk-mq.h
@@ -0,0 +1,52 @@
1#ifndef INT_BLK_MQ_H
2#define INT_BLK_MQ_H
3
4struct blk_mq_ctx {
5 struct {
6 spinlock_t lock;
7 struct list_head rq_list;
8 } ____cacheline_aligned_in_smp;
9
10 unsigned int cpu;
11 unsigned int index_hw;
12 unsigned int ipi_redirect;
13
14 /* incremented at dispatch time */
15 unsigned long rq_dispatched[2];
16 unsigned long rq_merged;
17
18 /* incremented at completion time */
19 unsigned long ____cacheline_aligned_in_smp rq_completed[2];
20
21 struct request_queue *queue;
22 struct kobject kobj;
23};
24
25void __blk_mq_end_io(struct request *rq, int error);
26void blk_mq_complete_request(struct request *rq, int error);
27void blk_mq_run_request(struct request *rq, bool run_queue, bool async);
28void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async);
29void blk_mq_init_flush(struct request_queue *q);
30
31/*
32 * CPU hotplug helpers
33 */
34struct blk_mq_cpu_notifier;
35void blk_mq_init_cpu_notifier(struct blk_mq_cpu_notifier *notifier,
36 void (*fn)(void *, unsigned long, unsigned int),
37 void *data);
38void blk_mq_register_cpu_notifier(struct blk_mq_cpu_notifier *notifier);
39void blk_mq_unregister_cpu_notifier(struct blk_mq_cpu_notifier *notifier);
40void blk_mq_cpu_init(void);
41DECLARE_PER_CPU(struct llist_head, ipi_lists);
42
43/*
44 * CPU -> queue mappings
45 */
46struct blk_mq_reg;
47extern unsigned int *blk_mq_make_queue_map(struct blk_mq_reg *reg);
48extern int blk_mq_update_queue_map(unsigned int *map, unsigned int nr_queues);
49
50void blk_mq_add_timer(struct request *rq);
51
52#endif
diff --git a/block/blk-sysfs.c b/block/blk-sysfs.c
index 3aa5b195f4dd..4f8c4d90ec73 100644
--- a/block/blk-sysfs.c
+++ b/block/blk-sysfs.c
@@ -7,6 +7,7 @@
7#include <linux/bio.h> 7#include <linux/bio.h>
8#include <linux/blkdev.h> 8#include <linux/blkdev.h>
9#include <linux/blktrace_api.h> 9#include <linux/blktrace_api.h>
10#include <linux/blk-mq.h>
10 11
11#include "blk.h" 12#include "blk.h"
12#include "blk-cgroup.h" 13#include "blk-cgroup.h"
@@ -542,6 +543,11 @@ static void blk_release_queue(struct kobject *kobj)
542 if (q->queue_tags) 543 if (q->queue_tags)
543 __blk_queue_free_tags(q); 544 __blk_queue_free_tags(q);
544 545
546 percpu_counter_destroy(&q->mq_usage_counter);
547
548 if (q->mq_ops)
549 blk_mq_free_queue(q);
550
545 blk_trace_shutdown(q); 551 blk_trace_shutdown(q);
546 552
547 bdi_destroy(&q->backing_dev_info); 553 bdi_destroy(&q->backing_dev_info);
@@ -575,6 +581,7 @@ int blk_register_queue(struct gendisk *disk)
575 * bypass from queue allocation. 581 * bypass from queue allocation.
576 */ 582 */
577 blk_queue_bypass_end(q); 583 blk_queue_bypass_end(q);
584 queue_flag_set_unlocked(QUEUE_FLAG_INIT_DONE, q);
578 585
579 ret = blk_trace_init_sysfs(dev); 586 ret = blk_trace_init_sysfs(dev);
580 if (ret) 587 if (ret)
@@ -588,6 +595,9 @@ int blk_register_queue(struct gendisk *disk)
588 595
589 kobject_uevent(&q->kobj, KOBJ_ADD); 596 kobject_uevent(&q->kobj, KOBJ_ADD);
590 597
598 if (q->mq_ops)
599 blk_mq_register_disk(disk);
600
591 if (!q->request_fn) 601 if (!q->request_fn)
592 return 0; 602 return 0;
593 603
@@ -610,6 +620,9 @@ void blk_unregister_queue(struct gendisk *disk)
610 if (WARN_ON(!q)) 620 if (WARN_ON(!q))
611 return; 621 return;
612 622
623 if (q->mq_ops)
624 blk_mq_unregister_disk(disk);
625
613 if (q->request_fn) 626 if (q->request_fn)
614 elv_unregister_queue(q); 627 elv_unregister_queue(q);
615 628
diff --git a/block/blk-timeout.c b/block/blk-timeout.c
index abf725c655fc..bba81c9348e1 100644
--- a/block/blk-timeout.c
+++ b/block/blk-timeout.c
@@ -7,6 +7,7 @@
7#include <linux/fault-inject.h> 7#include <linux/fault-inject.h>
8 8
9#include "blk.h" 9#include "blk.h"
10#include "blk-mq.h"
10 11
11#ifdef CONFIG_FAIL_IO_TIMEOUT 12#ifdef CONFIG_FAIL_IO_TIMEOUT
12 13
@@ -88,10 +89,18 @@ static void blk_rq_timed_out(struct request *req)
88 ret = q->rq_timed_out_fn(req); 89 ret = q->rq_timed_out_fn(req);
89 switch (ret) { 90 switch (ret) {
90 case BLK_EH_HANDLED: 91 case BLK_EH_HANDLED:
91 __blk_complete_request(req); 92 /* Can we use req->errors here? */
93 if (q->mq_ops)
94 blk_mq_complete_request(req, req->errors);
95 else
96 __blk_complete_request(req);
92 break; 97 break;
93 case BLK_EH_RESET_TIMER: 98 case BLK_EH_RESET_TIMER:
94 blk_add_timer(req); 99 if (q->mq_ops)
100 blk_mq_add_timer(req);
101 else
102 blk_add_timer(req);
103
95 blk_clear_rq_complete(req); 104 blk_clear_rq_complete(req);
96 break; 105 break;
97 case BLK_EH_NOT_HANDLED: 106 case BLK_EH_NOT_HANDLED:
@@ -108,6 +117,23 @@ static void blk_rq_timed_out(struct request *req)
108 } 117 }
109} 118}
110 119
120void blk_rq_check_expired(struct request *rq, unsigned long *next_timeout,
121 unsigned int *next_set)
122{
123 if (time_after_eq(jiffies, rq->deadline)) {
124 list_del_init(&rq->timeout_list);
125
126 /*
127 * Check if we raced with end io completion
128 */
129 if (!blk_mark_rq_complete(rq))
130 blk_rq_timed_out(rq);
131 } else if (!*next_set || time_after(*next_timeout, rq->deadline)) {
132 *next_timeout = rq->deadline;
133 *next_set = 1;
134 }
135}
136
111void blk_rq_timed_out_timer(unsigned long data) 137void blk_rq_timed_out_timer(unsigned long data)
112{ 138{
113 struct request_queue *q = (struct request_queue *) data; 139 struct request_queue *q = (struct request_queue *) data;
@@ -117,21 +143,8 @@ void blk_rq_timed_out_timer(unsigned long data)
117 143
118 spin_lock_irqsave(q->queue_lock, flags); 144 spin_lock_irqsave(q->queue_lock, flags);
119 145
120 list_for_each_entry_safe(rq, tmp, &q->timeout_list, timeout_list) { 146 list_for_each_entry_safe(rq, tmp, &q->timeout_list, timeout_list)
121 if (time_after_eq(jiffies, rq->deadline)) { 147 blk_rq_check_expired(rq, &next, &next_set);
122 list_del_init(&rq->timeout_list);
123
124 /*
125 * Check if we raced with end io completion
126 */
127 if (blk_mark_rq_complete(rq))
128 continue;
129 blk_rq_timed_out(rq);
130 } else if (!next_set || time_after(next, rq->deadline)) {
131 next = rq->deadline;
132 next_set = 1;
133 }
134 }
135 148
136 if (next_set) 149 if (next_set)
137 mod_timer(&q->timeout, round_jiffies_up(next)); 150 mod_timer(&q->timeout, round_jiffies_up(next));
@@ -157,15 +170,7 @@ void blk_abort_request(struct request *req)
157} 170}
158EXPORT_SYMBOL_GPL(blk_abort_request); 171EXPORT_SYMBOL_GPL(blk_abort_request);
159 172
160/** 173void __blk_add_timer(struct request *req, struct list_head *timeout_list)
161 * blk_add_timer - Start timeout timer for a single request
162 * @req: request that is about to start running.
163 *
164 * Notes:
165 * Each request has its own timer, and as it is added to the queue, we
166 * set up the timer. When the request completes, we cancel the timer.
167 */
168void blk_add_timer(struct request *req)
169{ 174{
170 struct request_queue *q = req->q; 175 struct request_queue *q = req->q;
171 unsigned long expiry; 176 unsigned long expiry;
@@ -183,7 +188,8 @@ void blk_add_timer(struct request *req)
183 req->timeout = q->rq_timeout; 188 req->timeout = q->rq_timeout;
184 189
185 req->deadline = jiffies + req->timeout; 190 req->deadline = jiffies + req->timeout;
186 list_add_tail(&req->timeout_list, &q->timeout_list); 191 if (timeout_list)
192 list_add_tail(&req->timeout_list, timeout_list);
187 193
188 /* 194 /*
189 * If the timer isn't already pending or this timeout is earlier 195 * If the timer isn't already pending or this timeout is earlier
@@ -195,5 +201,19 @@ void blk_add_timer(struct request *req)
195 if (!timer_pending(&q->timeout) || 201 if (!timer_pending(&q->timeout) ||
196 time_before(expiry, q->timeout.expires)) 202 time_before(expiry, q->timeout.expires))
197 mod_timer(&q->timeout, expiry); 203 mod_timer(&q->timeout, expiry);
204
205}
206
207/**
208 * blk_add_timer - Start timeout timer for a single request
209 * @req: request that is about to start running.
210 *
211 * Notes:
212 * Each request has its own timer, and as it is added to the queue, we
213 * set up the timer. When the request completes, we cancel the timer.
214 */
215void blk_add_timer(struct request *req)
216{
217 __blk_add_timer(req, &req->q->timeout_list);
198} 218}
199 219
diff --git a/block/blk.h b/block/blk.h
index e837b8f619b7..c90e1d8f7a2b 100644
--- a/block/blk.h
+++ b/block/blk.h
@@ -10,6 +10,7 @@
10#define BLK_BATCH_REQ 32 10#define BLK_BATCH_REQ 32
11 11
12extern struct kmem_cache *blk_requestq_cachep; 12extern struct kmem_cache *blk_requestq_cachep;
13extern struct kmem_cache *request_cachep;
13extern struct kobj_type blk_queue_ktype; 14extern struct kobj_type blk_queue_ktype;
14extern struct ida blk_queue_ida; 15extern struct ida blk_queue_ida;
15 16
@@ -34,14 +35,30 @@ bool __blk_end_bidi_request(struct request *rq, int error,
34 unsigned int nr_bytes, unsigned int bidi_bytes); 35 unsigned int nr_bytes, unsigned int bidi_bytes);
35 36
36void blk_rq_timed_out_timer(unsigned long data); 37void blk_rq_timed_out_timer(unsigned long data);
38void blk_rq_check_expired(struct request *rq, unsigned long *next_timeout,
39 unsigned int *next_set);
40void __blk_add_timer(struct request *req, struct list_head *timeout_list);
37void blk_delete_timer(struct request *); 41void blk_delete_timer(struct request *);
38void blk_add_timer(struct request *); 42void blk_add_timer(struct request *);
39 43
44
45bool bio_attempt_front_merge(struct request_queue *q, struct request *req,
46 struct bio *bio);
47bool bio_attempt_back_merge(struct request_queue *q, struct request *req,
48 struct bio *bio);
49bool blk_attempt_plug_merge(struct request_queue *q, struct bio *bio,
50 unsigned int *request_count);
51
52void blk_account_io_start(struct request *req, bool new_io);
53void blk_account_io_completion(struct request *req, unsigned int bytes);
54void blk_account_io_done(struct request *req);
55
40/* 56/*
41 * Internal atomic flags for request handling 57 * Internal atomic flags for request handling
42 */ 58 */
43enum rq_atomic_flags { 59enum rq_atomic_flags {
44 REQ_ATOM_COMPLETE = 0, 60 REQ_ATOM_COMPLETE = 0,
61 REQ_ATOM_STARTED,
45}; 62};
46 63
47/* 64/*
generated by cgit v1.2.2 (git 2.25.0) at 2025-02-16 22:08:04 -0500