xref: /openbmc/linux/lib/cpumask.c (revision 2aa71b4b)
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  * Returns >= 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  * Returns TRUE if memory allocation succeeded, FALSE otherwise.
53  *
54  * In addition, mask will be NULL if this fails.  Note that gcc is
55  * usually smart enough to know that mask can never be NULL if
56  * CONFIG_CPUMASK_OFFSTACK=n, so does code elimination in that case
57  * too.
58  */
59 bool alloc_cpumask_var_node(cpumask_var_t *mask, gfp_t flags, int node)
60 {
61 	*mask = kmalloc_node(cpumask_size(), flags, node);
62 
63 #ifdef CONFIG_DEBUG_PER_CPU_MAPS
64 	if (!*mask) {
65 		printk(KERN_ERR "=> alloc_cpumask_var: failed!\n");
66 		dump_stack();
67 	}
68 #endif
69 
70 	return *mask != NULL;
71 }
72 EXPORT_SYMBOL(alloc_cpumask_var_node);
73 
74 /**
75  * alloc_bootmem_cpumask_var - allocate a struct cpumask from the bootmem arena.
76  * @mask: pointer to cpumask_var_t where the cpumask is returned
77  *
78  * Only defined when CONFIG_CPUMASK_OFFSTACK=y, otherwise is
79  * a nop (in <linux/cpumask.h>).
80  * Either returns an allocated (zero-filled) cpumask, or causes the
81  * system to panic.
82  */
83 void __init alloc_bootmem_cpumask_var(cpumask_var_t *mask)
84 {
85 	*mask = memblock_alloc(cpumask_size(), SMP_CACHE_BYTES);
86 	if (!*mask)
87 		panic("%s: Failed to allocate %u bytes\n", __func__,
88 		      cpumask_size());
89 }
90 
91 /**
92  * free_cpumask_var - frees memory allocated for a struct cpumask.
93  * @mask: cpumask to free
94  *
95  * This is safe on a NULL mask.
96  */
97 void free_cpumask_var(cpumask_var_t mask)
98 {
99 	kfree(mask);
100 }
101 EXPORT_SYMBOL(free_cpumask_var);
102 
103 /**
104  * free_bootmem_cpumask_var - frees result of alloc_bootmem_cpumask_var
105  * @mask: cpumask to free
106  */
107 void __init free_bootmem_cpumask_var(cpumask_var_t mask)
108 {
109 	memblock_free(mask, cpumask_size());
110 }
111 #endif
112 
113 /**
114  * cpumask_local_spread - select the i'th cpu based on NUMA distances
115  * @i: index number
116  * @node: local numa_node
117  *
118  * Returns online CPU according to a numa aware policy; local cpus are returned
119  * first, followed by non-local ones, then it wraps around.
120  *
121  * For those who wants to enumerate all CPUs based on their NUMA distances,
122  * i.e. call this function in a loop, like:
123  *
124  * for (i = 0; i < num_online_cpus(); i++) {
125  *	cpu = cpumask_local_spread(i, node);
126  *	do_something(cpu);
127  * }
128  *
129  * There's a better alternative based on for_each()-like iterators:
130  *
131  *	for_each_numa_hop_mask(mask, node) {
132  *		for_each_cpu_andnot(cpu, mask, prev)
133  *			do_something(cpu);
134  *		prev = mask;
135  *	}
136  *
137  * It's simpler and more verbose than above. Complexity of iterator-based
138  * enumeration is O(sched_domains_numa_levels * nr_cpu_ids), while
139  * cpumask_local_spread() when called for each cpu is
140  * O(sched_domains_numa_levels * nr_cpu_ids * log(nr_cpu_ids)).
141  */
142 unsigned int cpumask_local_spread(unsigned int i, int node)
143 {
144 	unsigned int cpu;
145 
146 	/* Wrap: we always want a cpu. */
147 	i %= num_online_cpus();
148 
149 	cpu = (node == NUMA_NO_NODE) ?
150 		cpumask_nth(i, cpu_online_mask) :
151 		sched_numa_find_nth_cpu(cpu_online_mask, i, node);
152 
153 	WARN_ON(cpu >= nr_cpu_ids);
154 	return cpu;
155 }
156 EXPORT_SYMBOL(cpumask_local_spread);
157 
158 static DEFINE_PER_CPU(int, distribute_cpu_mask_prev);
159 
160 /**
161  * cpumask_any_and_distribute - Return an arbitrary cpu within src1p & src2p.
162  * @src1p: first &cpumask for intersection
163  * @src2p: second &cpumask for intersection
164  *
165  * Iterated calls using the same srcp1 and srcp2 will be distributed within
166  * their intersection.
167  *
168  * Returns >= nr_cpu_ids if the intersection is empty.
169  */
170 unsigned int cpumask_any_and_distribute(const struct cpumask *src1p,
171 			       const struct cpumask *src2p)
172 {
173 	unsigned int next, prev;
174 
175 	/* NOTE: our first selection will skip 0. */
176 	prev = __this_cpu_read(distribute_cpu_mask_prev);
177 
178 	next = find_next_and_bit_wrap(cpumask_bits(src1p), cpumask_bits(src2p),
179 					nr_cpumask_bits, prev + 1);
180 	if (next < nr_cpu_ids)
181 		__this_cpu_write(distribute_cpu_mask_prev, next);
182 
183 	return next;
184 }
185 EXPORT_SYMBOL(cpumask_any_and_distribute);
186 
187 unsigned int cpumask_any_distribute(const struct cpumask *srcp)
188 {
189 	unsigned int next, prev;
190 
191 	/* NOTE: our first selection will skip 0. */
192 	prev = __this_cpu_read(distribute_cpu_mask_prev);
193 	next = find_next_bit_wrap(cpumask_bits(srcp), nr_cpumask_bits, prev + 1);
194 	if (next < nr_cpu_ids)
195 		__this_cpu_write(distribute_cpu_mask_prev, next);
196 
197 	return next;
198 }
199 EXPORT_SYMBOL(cpumask_any_distribute);
200