xref: /linux/lib/cpumask.c (revision b58b13f156c00c2457035b7071eaaac105fe6836)
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 /**
11  * cpumask_next_wrap - helper to implement for_each_cpu_wrap
12  * @n: the cpu prior to the place to search
13  * @mask: the cpumask pointer
14  * @start: the start point of the iteration
15  * @wrap: assume @n crossing @start terminates the iteration
16  *
17  * Return: >= nr_cpu_ids on completion
18  *
19  * Note: the @wrap argument is required for the start condition when
20  * we cannot assume @start is set in @mask.
21  */
22 unsigned int cpumask_next_wrap(int n, const struct cpumask *mask, int start, bool wrap)
23 {
24 	unsigned int next;
25 
26 again:
27 	next = cpumask_next(n, mask);
28 
29 	if (wrap && n < start && next >= start) {
30 		return nr_cpumask_bits;
31 
32 	} else if (next >= nr_cpumask_bits) {
33 		wrap = true;
34 		n = -1;
35 		goto again;
36 	}
37 
38 	return next;
39 }
40 EXPORT_SYMBOL(cpumask_next_wrap);
41 
42 /* These are not inline because of header tangles. */
43 #ifdef CONFIG_CPUMASK_OFFSTACK
44 /**
45  * alloc_cpumask_var_node - allocate a struct cpumask on a given node
46  * @mask: pointer to cpumask_var_t where the cpumask is returned
47  * @flags: GFP_ flags
48  * @node: memory node from which to allocate or %NUMA_NO_NODE
49  *
50  * Only defined when CONFIG_CPUMASK_OFFSTACK=y, otherwise is
51  * a nop returning a constant 1 (in <linux/cpumask.h>).
52  *
53  * Return: TRUE if memory allocation succeeded, FALSE otherwise.
54  *
55  * In addition, mask will be NULL if this fails.  Note that gcc is
56  * usually smart enough to know that mask can never be NULL if
57  * CONFIG_CPUMASK_OFFSTACK=n, so does code elimination in that case
58  * too.
59  */
60 bool alloc_cpumask_var_node(cpumask_var_t *mask, gfp_t flags, int node)
61 {
62 	*mask = kmalloc_node(cpumask_size(), flags, node);
63 
64 #ifdef CONFIG_DEBUG_PER_CPU_MAPS
65 	if (!*mask) {
66 		printk(KERN_ERR "=> alloc_cpumask_var: failed!\n");
67 		dump_stack();
68 	}
69 #endif
70 
71 	return *mask != NULL;
72 }
73 EXPORT_SYMBOL(alloc_cpumask_var_node);
74 
75 /**
76  * alloc_bootmem_cpumask_var - allocate a struct cpumask from the bootmem arena.
77  * @mask: pointer to cpumask_var_t where the cpumask is returned
78  *
79  * Only defined when CONFIG_CPUMASK_OFFSTACK=y, otherwise is
80  * a nop (in <linux/cpumask.h>).
81  * Either returns an allocated (zero-filled) cpumask, or causes the
82  * system to panic.
83  */
84 void __init alloc_bootmem_cpumask_var(cpumask_var_t *mask)
85 {
86 	*mask = memblock_alloc(cpumask_size(), SMP_CACHE_BYTES);
87 	if (!*mask)
88 		panic("%s: Failed to allocate %u bytes\n", __func__,
89 		      cpumask_size());
90 }
91 
92 /**
93  * free_cpumask_var - frees memory allocated for a struct cpumask.
94  * @mask: cpumask to free
95  *
96  * This is safe on a NULL mask.
97  */
98 void free_cpumask_var(cpumask_var_t mask)
99 {
100 	kfree(mask);
101 }
102 EXPORT_SYMBOL(free_cpumask_var);
103 
104 /**
105  * free_bootmem_cpumask_var - frees result of alloc_bootmem_cpumask_var
106  * @mask: cpumask to free
107  */
108 void __init free_bootmem_cpumask_var(cpumask_var_t mask)
109 {
110 	memblock_free(mask, cpumask_size());
111 }
112 #endif
113 
114 /**
115  * cpumask_local_spread - select the i'th cpu based on NUMA distances
116  * @i: index number
117  * @node: local numa_node
118  *
119  * Return: online CPU according to a numa aware policy; local cpus are returned
120  * first, followed by non-local ones, then it wraps around.
121  *
122  * For those who wants to enumerate all CPUs based on their NUMA distances,
123  * i.e. call this function in a loop, like:
124  *
125  * for (i = 0; i < num_online_cpus(); i++) {
126  *	cpu = cpumask_local_spread(i, node);
127  *	do_something(cpu);
128  * }
129  *
130  * There's a better alternative based on for_each()-like iterators:
131  *
132  *	for_each_numa_hop_mask(mask, node) {
133  *		for_each_cpu_andnot(cpu, mask, prev)
134  *			do_something(cpu);
135  *		prev = mask;
136  *	}
137  *
138  * It's simpler and more verbose than above. Complexity of iterator-based
139  * enumeration is O(sched_domains_numa_levels * nr_cpu_ids), while
140  * cpumask_local_spread() when called for each cpu is
141  * O(sched_domains_numa_levels * nr_cpu_ids * log(nr_cpu_ids)).
142  */
143 unsigned int cpumask_local_spread(unsigned int i, int node)
144 {
145 	unsigned int cpu;
146 
147 	/* Wrap: we always want a cpu. */
148 	i %= num_online_cpus();
149 
150 	cpu = sched_numa_find_nth_cpu(cpu_online_mask, i, node);
151 
152 	WARN_ON(cpu >= nr_cpu_ids);
153 	return cpu;
154 }
155 EXPORT_SYMBOL(cpumask_local_spread);
156 
157 static DEFINE_PER_CPU(int, distribute_cpu_mask_prev);
158 
159 /**
160  * cpumask_any_and_distribute - Return an arbitrary cpu within src1p & src2p.
161  * @src1p: first &cpumask for intersection
162  * @src2p: second &cpumask for intersection
163  *
164  * Iterated calls using the same srcp1 and srcp2 will be distributed within
165  * their intersection.
166  *
167  * Return: >= nr_cpu_ids if the intersection is empty.
168  */
169 unsigned int cpumask_any_and_distribute(const struct cpumask *src1p,
170 			       const struct cpumask *src2p)
171 {
172 	unsigned int next, prev;
173 
174 	/* NOTE: our first selection will skip 0. */
175 	prev = __this_cpu_read(distribute_cpu_mask_prev);
176 
177 	next = find_next_and_bit_wrap(cpumask_bits(src1p), cpumask_bits(src2p),
178 					nr_cpumask_bits, prev + 1);
179 	if (next < nr_cpu_ids)
180 		__this_cpu_write(distribute_cpu_mask_prev, next);
181 
182 	return next;
183 }
184 EXPORT_SYMBOL(cpumask_any_and_distribute);
185 
186 /**
187  * cpumask_any_distribute - Return an arbitrary cpu from srcp
188  * @srcp: &cpumask for selection
189  *
190  * Return: >= nr_cpu_ids if the intersection is empty.
191  */
192 unsigned int cpumask_any_distribute(const struct cpumask *srcp)
193 {
194 	unsigned int next, prev;
195 
196 	/* NOTE: our first selection will skip 0. */
197 	prev = __this_cpu_read(distribute_cpu_mask_prev);
198 	next = find_next_bit_wrap(cpumask_bits(srcp), nr_cpumask_bits, prev + 1);
199 	if (next < nr_cpu_ids)
200 		__this_cpu_write(distribute_cpu_mask_prev, next);
201 
202 	return next;
203 }
204 EXPORT_SYMBOL(cpumask_any_distribute);
205