| Jens Axboe | 320ae51 | 2013-10-24 09:20:05 +0100 | [diff] [blame] | 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 | |
| 12 | static 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 | |
| 21 | static 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 | |
| 27 | static 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 | |
| 38 | int 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 | |
| 93 | unsigned 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 | } |