1 // SPDX-License-Identifier: GPL-2.0
2 #include <linux/slab.h>
3 #include <linux/kernel.h>
4 #include <linux/bitops.h>
5 #include <linux/cpumask.h>
6 #include <linux/export.h>
7 #include <linux/memblock.h>
8 #include <linux/numa.h>
9
10 /* These are not inline because of header tangles. */
11 #ifdef CONFIG_CPUMASK_OFFSTACK
12 /**
13 * alloc_cpumask_var_node - allocate a struct cpumask on a given node
14 * @mask: pointer to cpumask_var_t where the cpumask is returned
15 * @flags: GFP_ flags
16 * @node: memory node from which to allocate or %NUMA_NO_NODE
17 *
18 * Only defined when CONFIG_CPUMASK_OFFSTACK=y, otherwise is
19 * a nop returning a constant 1 (in <linux/cpumask.h>).
20 *
21 * Return: TRUE if memory allocation succeeded, FALSE otherwise.
22 *
23 * In addition, mask will be NULL if this fails. Note that gcc is
24 * usually smart enough to know that mask can never be NULL if
25 * CONFIG_CPUMASK_OFFSTACK=n, so does code elimination in that case
26 * too.
27 */
alloc_cpumask_var_node(cpumask_var_t * mask,gfp_t flags,int node)28 bool alloc_cpumask_var_node(cpumask_var_t *mask, gfp_t flags, int node)
29 {
30 *mask = kmalloc_node(cpumask_size(), flags, node);
31
32 #ifdef CONFIG_DEBUG_PER_CPU_MAPS
33 if (!*mask) {
34 printk(KERN_ERR "=> alloc_cpumask_var: failed!\n");
35 dump_stack();
36 }
37 #endif
38
39 return *mask != NULL;
40 }
41 EXPORT_SYMBOL(alloc_cpumask_var_node);
42
43 /**
44 * alloc_bootmem_cpumask_var - allocate a struct cpumask from the bootmem arena.
45 * @mask: pointer to cpumask_var_t where the cpumask is returned
46 *
47 * Only defined when CONFIG_CPUMASK_OFFSTACK=y, otherwise is
48 * a nop (in <linux/cpumask.h>).
49 * Either returns an allocated (zero-filled) cpumask, or causes the
50 * system to panic.
51 */
alloc_bootmem_cpumask_var(cpumask_var_t * mask)52 void __init alloc_bootmem_cpumask_var(cpumask_var_t *mask)
53 {
54 *mask = memblock_alloc_or_panic(cpumask_size(), SMP_CACHE_BYTES);
55 }
56
57 /**
58 * free_cpumask_var - frees memory allocated for a struct cpumask.
59 * @mask: cpumask to free
60 *
61 * This is safe on a NULL mask.
62 */
free_cpumask_var(cpumask_var_t mask)63 void free_cpumask_var(cpumask_var_t mask)
64 {
65 kfree(mask);
66 }
67 EXPORT_SYMBOL(free_cpumask_var);
68
69 /**
70 * free_bootmem_cpumask_var - frees result of alloc_bootmem_cpumask_var
71 * @mask: cpumask to free
72 */
free_bootmem_cpumask_var(cpumask_var_t mask)73 void __init free_bootmem_cpumask_var(cpumask_var_t mask)
74 {
75 memblock_free(mask, cpumask_size());
76 }
77 #endif
78
79 /**
80 * cpumask_local_spread - select the i'th cpu based on NUMA distances
81 * @i: index number
82 * @node: local numa_node
83 *
84 * Return: online CPU according to a numa aware policy; local cpus are returned
85 * first, followed by non-local ones, then it wraps around.
86 *
87 * For those who wants to enumerate all CPUs based on their NUMA distances,
88 * i.e. call this function in a loop, like:
89 *
90 * for (i = 0; i < num_online_cpus(); i++) {
91 * cpu = cpumask_local_spread(i, node);
92 * do_something(cpu);
93 * }
94 *
95 * There's a better alternative based on for_each()-like iterators:
96 *
97 * for_each_numa_hop_mask(mask, node) {
98 * for_each_cpu_andnot(cpu, mask, prev)
99 * do_something(cpu);
100 * prev = mask;
101 * }
102 *
103 * It's simpler and more verbose than above. Complexity of iterator-based
104 * enumeration is O(sched_domains_numa_levels * nr_cpu_ids), while
105 * cpumask_local_spread() when called for each cpu is
106 * O(sched_domains_numa_levels * nr_cpu_ids * log(nr_cpu_ids)).
107 */
cpumask_local_spread(unsigned int i,int node)108 unsigned int cpumask_local_spread(unsigned int i, int node)
109 {
110 unsigned int cpu;
111
112 /* Wrap: we always want a cpu. */
113 i %= num_online_cpus();
114
115 cpu = sched_numa_find_nth_cpu(cpu_online_mask, i, node);
116
117 WARN_ON(cpu >= nr_cpu_ids);
118 return cpu;
119 }
120 EXPORT_SYMBOL(cpumask_local_spread);
121
122 static DEFINE_PER_CPU(int, distribute_cpu_mask_prev);
123
124 /**
125 * cpumask_any_and_distribute - Return an arbitrary cpu within src1p & src2p.
126 * @src1p: first &cpumask for intersection
127 * @src2p: second &cpumask for intersection
128 *
129 * Iterated calls using the same srcp1 and srcp2 will be distributed within
130 * their intersection.
131 *
132 * Return: >= nr_cpu_ids if the intersection is empty.
133 */
cpumask_any_and_distribute(const struct cpumask * src1p,const struct cpumask * src2p)134 unsigned int cpumask_any_and_distribute(const struct cpumask *src1p,
135 const struct cpumask *src2p)
136 {
137 unsigned int next, prev;
138
139 /* NOTE: our first selection will skip 0. */
140 prev = __this_cpu_read(distribute_cpu_mask_prev);
141
142 next = cpumask_next_and_wrap(prev, src1p, src2p);
143 if (next < nr_cpu_ids)
144 __this_cpu_write(distribute_cpu_mask_prev, next);
145
146 return next;
147 }
148 EXPORT_SYMBOL(cpumask_any_and_distribute);
149
150 /**
151 * cpumask_any_distribute - Return an arbitrary cpu from srcp
152 * @srcp: &cpumask for selection
153 *
154 * Return: >= nr_cpu_ids if the intersection is empty.
155 */
cpumask_any_distribute(const struct cpumask * srcp)156 unsigned int cpumask_any_distribute(const struct cpumask *srcp)
157 {
158 unsigned int next, prev;
159
160 /* NOTE: our first selection will skip 0. */
161 prev = __this_cpu_read(distribute_cpu_mask_prev);
162 next = cpumask_next_wrap(prev, srcp);
163 if (next < nr_cpu_ids)
164 __this_cpu_write(distribute_cpu_mask_prev, next);
165
166 return next;
167 }
168 EXPORT_SYMBOL(cpumask_any_distribute);
169