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
 17static 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
 23static 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
 34int blk_mq_update_queue_map(unsigned int *map, unsigned int nr_queues,
 35			    const struct cpumask *online_mask)
 36{
 37	unsigned int i, nr_cpus, nr_uniq_cpus, queue, first_sibling;
 38	cpumask_var_t cpus;
 39
 40	if (!alloc_cpumask_var(&cpus, GFP_ATOMIC))
 41		return 1;
 42
 43	cpumask_clear(cpus);
 44	nr_cpus = nr_uniq_cpus = 0;
 45	for_each_cpu(i, online_mask) {
 46		nr_cpus++;
 47		first_sibling = get_first_sibling(i);
 48		if (!cpumask_test_cpu(first_sibling, cpus))
 49			nr_uniq_cpus++;
 50		cpumask_set_cpu(i, cpus);
 
 51	}
 52
 53	queue = 0;
 54	for_each_possible_cpu(i) {
 55		if (!cpumask_test_cpu(i, online_mask)) {
 56			map[i] = 0;
 57			continue;
 58		}
 59
 60		/*
 61		 * Easy case - we have equal or more hardware queues. Or
 62		 * there are no thread siblings to take into account. Do
 63		 * 1:1 if enough, or sequential mapping if less.
 
 64		 */
 65		if (nr_queues >= nr_cpus || nr_cpus == nr_uniq_cpus) {
 66			map[i] = cpu_to_queue_index(nr_cpus, nr_queues, queue);
 67			queue++;
 68			continue;
 
 
 
 
 69		}
 70
 71		/*
 72		 * Less then nr_cpus queues, and we have some number of
 73		 * threads per cores. Map sibling threads to the same
 74		 * queue.
 75		 */
 76		first_sibling = get_first_sibling(i);
 77		if (first_sibling == i) {
 78			map[i] = cpu_to_queue_index(nr_uniq_cpus, nr_queues,
 79							queue);
 80			queue++;
 81		} else
 82			map[i] = map[first_sibling];
 83	}
 84
 85	free_cpumask_var(cpus);
 86	return 0;
 87}
 
 88
 89unsigned int *blk_mq_make_queue_map(struct blk_mq_tag_set *set)
 90{
 91	unsigned int *map;
 92
 93	/* If cpus are offline, map them to first hctx */
 94	map = kzalloc_node(sizeof(*map) * nr_cpu_ids, GFP_KERNEL,
 95				set->numa_node);
 96	if (!map)
 97		return NULL;
 98
 99	if (!blk_mq_update_queue_map(map, set->nr_hw_queues, cpu_online_mask))
100		return map;
101
102	kfree(map);
103	return NULL;
104}
105
106/*
107 * We have no quick way of doing reverse lookups. This is only used at
108 * queue init time, so runtime isn't important.
109 */
110int blk_mq_hw_queue_to_node(unsigned int *mq_map, unsigned int index)
111{
112	int i;
113
114	for_each_possible_cpu(i) {
115		if (index == mq_map[i])
116			return local_memory_node(cpu_to_node(i));
117	}
118
119	return NUMA_NO_NODE;
120}

 1// SPDX-License-Identifier: GPL-2.0
 2/*
 3 * CPU <-> hardware queue mapping helpers
 4 *
 5 * Copyright (C) 2013-2014 Jens Axboe
 6 */
 7#include <linux/kernel.h>
 8#include <linux/threads.h>
 9#include <linux/module.h>
10#include <linux/mm.h>
11#include <linux/smp.h>
12#include <linux/cpu.h>
13
14#include <linux/blk-mq.h>
15#include "blk.h"
16#include "blk-mq.h"
17
18static int queue_index(struct blk_mq_queue_map *qmap,
19		       unsigned int nr_queues, const int q)
20{
21	return qmap->queue_offset + (q % nr_queues);
22}
23
24static int get_first_sibling(unsigned int cpu)
25{
26	unsigned int ret;
27
28	ret = cpumask_first(topology_sibling_cpumask(cpu));
29	if (ret < nr_cpu_ids)
30		return ret;
31
32	return cpu;
33}
34
35int blk_mq_map_queues(struct blk_mq_queue_map *qmap)
 
36{
37	unsigned int *map = qmap->mq_map;
38	unsigned int nr_queues = qmap->nr_queues;
39	unsigned int cpu, first_sibling, q = 0;
40
41	for_each_possible_cpu(cpu)
42		map[cpu] = -1;
43
44	/*
45	 * Spread queues among present CPUs first for minimizing
46	 * count of dead queues which are mapped by all un-present CPUs
47	 */
48	for_each_present_cpu(cpu) {
49		if (q >= nr_queues)
50			break;
51		map[cpu] = queue_index(qmap, nr_queues, q++);
52	}
53
54	for_each_possible_cpu(cpu) {
55		if (map[cpu] != -1)
 
 
56			continue;
 
 
57		/*
58		 * First do sequential mapping between CPUs and queues.
59		 * In case we still have CPUs to map, and we have some number of
60		 * threads per cores then map sibling threads to the same queue
61		 * for performance optimizations.
62		 */
63		if (q < nr_queues) {
64			map[cpu] = queue_index(qmap, nr_queues, q++);
65		} else {
66			first_sibling = get_first_sibling(cpu);
67			if (first_sibling == cpu)
68				map[cpu] = queue_index(qmap, nr_queues, q++);
69			else
70				map[cpu] = map[first_sibling];
71		}
 
 
 
 
 
 
 
 
 
 
 
 
 
72	}
73
 
74	return 0;
75}
76EXPORT_SYMBOL_GPL(blk_mq_map_queues);
77
78/**
79 * blk_mq_hw_queue_to_node - Look up the memory node for a hardware queue index
80 * @qmap: CPU to hardware queue map.
81 * @index: hardware queue index.
82 *
 
 
 
 
 
 
 
 
 
 
 
 
 
83 * We have no quick way of doing reverse lookups. This is only used at
84 * queue init time, so runtime isn't important.
85 */
86int blk_mq_hw_queue_to_node(struct blk_mq_queue_map *qmap, unsigned int index)
87{
88	int i;
89
90	for_each_possible_cpu(i) {
91		if (index == qmap->mq_map[i])
92			return local_memory_node(cpu_to_node(i));
93	}
94
95	return NUMA_NO_NODE;
96}