Jens Axboe | 75bb462 | 2014-05-28 10:15:41 -0600 | [diff] [blame] | 1 | /* |
| 2 | * CPU <-> hardware queue mapping helpers |
| 3 | * |
| 4 | * Copyright (C) 2013-2014 Jens Axboe |
| 5 | */ |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 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 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 17 | static int cpu_to_queue_index(unsigned int nr_cpus, unsigned int nr_queues, |
| 18 | const int cpu) |
| 19 | { |
| 20 | return cpu / ((nr_cpus + nr_queues - 1) / nr_queues); |
| 21 | } |
| 22 | |
| 23 | static int get_first_sibling(unsigned int cpu) |
| 24 | { |
| 25 | unsigned int ret; |
| 26 | |
| 27 | ret = cpumask_first(topology_thread_cpumask(cpu)); |
| 28 | if (ret < nr_cpu_ids) |
| 29 | return ret; |
| 30 | |
| 31 | return cpu; |
| 32 | } |
| 33 | |
| 34 | int blk_mq_update_queue_map(unsigned int *map, unsigned int nr_queues) |
| 35 | { |
| 36 | unsigned int i, nr_cpus, nr_uniq_cpus, queue, first_sibling; |
| 37 | cpumask_var_t cpus; |
| 38 | |
| 39 | if (!alloc_cpumask_var(&cpus, GFP_ATOMIC)) |
| 40 | return 1; |
| 41 | |
| 42 | cpumask_clear(cpus); |
| 43 | nr_cpus = nr_uniq_cpus = 0; |
| 44 | for_each_online_cpu(i) { |
| 45 | nr_cpus++; |
| 46 | first_sibling = get_first_sibling(i); |
| 47 | if (!cpumask_test_cpu(first_sibling, cpus)) |
| 48 | nr_uniq_cpus++; |
| 49 | cpumask_set_cpu(i, cpus); |
| 50 | } |
| 51 | |
| 52 | queue = 0; |
| 53 | for_each_possible_cpu(i) { |
| 54 | if (!cpu_online(i)) { |
| 55 | map[i] = 0; |
| 56 | continue; |
| 57 | } |
| 58 | |
| 59 | /* |
| 60 | * Easy case - we have equal or more hardware queues. Or |
| 61 | * there are no thread siblings to take into account. Do |
| 62 | * 1:1 if enough, or sequential mapping if less. |
| 63 | */ |
| 64 | if (nr_queues >= nr_cpus || nr_cpus == nr_uniq_cpus) { |
| 65 | map[i] = cpu_to_queue_index(nr_cpus, nr_queues, queue); |
| 66 | queue++; |
| 67 | continue; |
| 68 | } |
| 69 | |
| 70 | /* |
| 71 | * Less then nr_cpus queues, and we have some number of |
| 72 | * threads per cores. Map sibling threads to the same |
| 73 | * queue. |
| 74 | */ |
| 75 | first_sibling = get_first_sibling(i); |
| 76 | if (first_sibling == i) { |
| 77 | map[i] = cpu_to_queue_index(nr_uniq_cpus, nr_queues, |
| 78 | queue); |
| 79 | queue++; |
| 80 | } else |
| 81 | map[i] = map[first_sibling]; |
| 82 | } |
| 83 | |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 84 | free_cpumask_var(cpus); |
| 85 | return 0; |
| 86 | } |
| 87 | |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 88 | unsigned int *blk_mq_make_queue_map(struct blk_mq_tag_set *set) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 89 | { |
| 90 | unsigned int *map; |
| 91 | |
| 92 | /* If cpus are offline, map them to first hctx */ |
| 93 | map = kzalloc_node(sizeof(*map) * num_possible_cpus(), GFP_KERNEL, |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 94 | set->numa_node); |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 95 | if (!map) |
| 96 | return NULL; |
| 97 | |
Christoph Hellwig | 24d2f90 | 2014-04-15 14:14:00 -0600 | [diff] [blame] | 98 | if (!blk_mq_update_queue_map(map, set->nr_hw_queues)) |
Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 99 | return map; |
| 100 | |
| 101 | kfree(map); |
| 102 | return NULL; |
| 103 | } |
Jens Axboe | f14bbe7 | 2014-05-27 12:06:53 -0600 | [diff] [blame] | 104 | |
| 105 | /* |
| 106 | * We have no quick way of doing reverse lookups. This is only used at |
| 107 | * queue init time, so runtime isn't important. |
| 108 | */ |
| 109 | int blk_mq_hw_queue_to_node(unsigned int *mq_map, unsigned int index) |
| 110 | { |
| 111 | int i; |
| 112 | |
| 113 | for_each_possible_cpu(i) { |
| 114 | if (index == mq_map[i]) |
| 115 | return cpu_to_node(i); |
| 116 | } |
| 117 | |
| 118 | return NUMA_NO_NODE; |
| 119 | } |