blob: 19b1d9c5f07e15107abb94241a5b8b24273ec681
1 | /* |
2 | * CPU <-> hardware queue mapping helpers |
3 | * |
4 | * Copyright (C) 2013-2014 Jens Axboe |
5 | */ |
6 | #include <linux/kernel.h> |
7 | #include <linux/threads.h> |
8 | #include <linux/module.h> |
9 | #include <linux/mm.h> |
10 | #include <linux/smp.h> |
11 | #include <linux/cpu.h> |
12 | |
13 | #include <linux/blk-mq.h> |
14 | #include "blk.h" |
15 | #include "blk-mq.h" |
16 | |
17 | static int cpu_to_queue_index(unsigned int nr_cpus, unsigned int nr_queues, |
18 | const int cpu) |
19 | { |
20 | return cpu * nr_queues / nr_cpus; |
21 | } |
22 | |
23 | static int get_first_sibling(unsigned int cpu) |
24 | { |
25 | unsigned int ret; |
26 | |
27 | ret = cpumask_first(topology_sibling_cpumask(cpu)); |
28 | if (ret < nr_cpu_ids) |
29 | return ret; |
30 | |
31 | return cpu; |
32 | } |
33 | |
34 | int blk_mq_map_queues(struct blk_mq_tag_set *set) |
35 | { |
36 | unsigned int *map = set->mq_map; |
37 | unsigned int nr_queues = set->nr_hw_queues; |
38 | const struct cpumask *online_mask = cpu_online_mask; |
39 | unsigned int i, nr_cpus, nr_uniq_cpus, queue, first_sibling; |
40 | cpumask_var_t cpus; |
41 | |
42 | if (!alloc_cpumask_var(&cpus, GFP_ATOMIC)) |
43 | return -ENOMEM; |
44 | |
45 | cpumask_clear(cpus); |
46 | nr_cpus = nr_uniq_cpus = 0; |
47 | for_each_cpu(i, online_mask) { |
48 | nr_cpus++; |
49 | first_sibling = get_first_sibling(i); |
50 | if (!cpumask_test_cpu(first_sibling, cpus)) |
51 | nr_uniq_cpus++; |
52 | cpumask_set_cpu(i, cpus); |
53 | } |
54 | |
55 | queue = 0; |
56 | for_each_possible_cpu(i) { |
57 | if (!cpumask_test_cpu(i, online_mask)) { |
58 | map[i] = 0; |
59 | continue; |
60 | } |
61 | |
62 | /* |
63 | * Easy case - we have equal or more hardware queues. Or |
64 | * there are no thread siblings to take into account. Do |
65 | * 1:1 if enough, or sequential mapping if less. |
66 | */ |
67 | if (nr_queues >= nr_cpus || nr_cpus == nr_uniq_cpus) { |
68 | map[i] = cpu_to_queue_index(nr_cpus, nr_queues, queue); |
69 | queue++; |
70 | continue; |
71 | } |
72 | |
73 | /* |
74 | * Less then nr_cpus queues, and we have some number of |
75 | * threads per cores. Map sibling threads to the same |
76 | * queue. |
77 | */ |
78 | first_sibling = get_first_sibling(i); |
79 | if (first_sibling == i) { |
80 | map[i] = cpu_to_queue_index(nr_uniq_cpus, nr_queues, |
81 | queue); |
82 | queue++; |
83 | } else |
84 | map[i] = map[first_sibling]; |
85 | } |
86 | |
87 | free_cpumask_var(cpus); |
88 | return 0; |
89 | } |
90 | |
91 | /* |
92 | * We have no quick way of doing reverse lookups. This is only used at |
93 | * queue init time, so runtime isn't important. |
94 | */ |
95 | int blk_mq_hw_queue_to_node(unsigned int *mq_map, unsigned int index) |
96 | { |
97 | int i; |
98 | |
99 | for_each_possible_cpu(i) { |
100 | if (index == mq_map[i]) |
101 | return local_memory_node(cpu_to_node(i)); |
102 | } |
103 | |
104 | return NUMA_NO_NODE; |
105 | } |
106 |