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 */
 22unsigned int cpumask_next_wrap(int n, const struct cpumask *mask, int start, bool wrap)
 23{
 24	unsigned int next;
 25
 26again:
 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}
 40EXPORT_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 */
 60bool 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}
 73EXPORT_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 */
 84void __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 */
 98void free_cpumask_var(cpumask_var_t mask)
 99{
100	kfree(mask);
101}
102EXPORT_SYMBOL(free_cpumask_var);
103
104/**
105 * free_bootmem_cpumask_var - frees result of alloc_bootmem_cpumask_var
106 * @mask: cpumask to free
107 */
108void __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 */
143unsigned 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}
155EXPORT_SYMBOL(cpumask_local_spread);
156
157static 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 */
169unsigned 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}
184EXPORT_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 */
192unsigned 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}
204EXPORT_SYMBOL(cpumask_any_distribute);

  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 - get the next cpu in a cpumask
 12 * @n: the cpu prior to the place to search (ie. return will be > @n)
 13 * @srcp: the cpumask pointer
 14 *
 15 * Returns >= nr_cpu_ids if no further cpus set.
 16 */
 17unsigned int cpumask_next(int n, const struct cpumask *srcp)
 18{
 19	/* -1 is a legal arg here. */
 20	if (n != -1)
 21		cpumask_check(n);
 22	return find_next_bit(cpumask_bits(srcp), nr_cpumask_bits, n + 1);
 23}
 24EXPORT_SYMBOL(cpumask_next);
 25
 26/**
 27 * cpumask_next_and - get the next cpu in *src1p & *src2p
 28 * @n: the cpu prior to the place to search (ie. return will be > @n)
 29 * @src1p: the first cpumask pointer
 30 * @src2p: the second cpumask pointer
 31 *
 32 * Returns >= nr_cpu_ids if no further cpus set in both.
 33 */
 34int cpumask_next_and(int n, const struct cpumask *src1p,
 35		     const struct cpumask *src2p)
 36{
 37	/* -1 is a legal arg here. */
 38	if (n != -1)
 39		cpumask_check(n);
 40	return find_next_and_bit(cpumask_bits(src1p), cpumask_bits(src2p),
 41		nr_cpumask_bits, n + 1);
 42}
 43EXPORT_SYMBOL(cpumask_next_and);
 44
 45/**
 46 * cpumask_any_but - return a "random" in a cpumask, but not this one.
 47 * @mask: the cpumask to search
 48 * @cpu: the cpu to ignore.
 49 *
 50 * Often used to find any cpu but smp_processor_id() in a mask.
 51 * Returns >= nr_cpu_ids if no cpus set.
 52 */
 53int cpumask_any_but(const struct cpumask *mask, unsigned int cpu)
 54{
 55	unsigned int i;
 56
 57	cpumask_check(cpu);
 58	for_each_cpu(i, mask)
 59		if (i != cpu)
 60			break;
 61	return i;
 62}
 63EXPORT_SYMBOL(cpumask_any_but);
 64
 65/**
 66 * cpumask_next_wrap - helper to implement for_each_cpu_wrap
 67 * @n: the cpu prior to the place to search
 68 * @mask: the cpumask pointer
 69 * @start: the start point of the iteration
 70 * @wrap: assume @n crossing @start terminates the iteration
 71 *
 72 * Returns >= nr_cpu_ids on completion
 73 *
 74 * Note: the @wrap argument is required for the start condition when
 75 * we cannot assume @start is set in @mask.
 76 */
 77int cpumask_next_wrap(int n, const struct cpumask *mask, int start, bool wrap)
 78{
 79	int next;
 80
 81again:
 82	next = cpumask_next(n, mask);
 83
 84	if (wrap && n < start && next >= start) {
 85		return nr_cpumask_bits;
 86
 87	} else if (next >= nr_cpumask_bits) {
 88		wrap = true;
 89		n = -1;
 90		goto again;
 91	}
 92
 93	return next;
 94}
 95EXPORT_SYMBOL(cpumask_next_wrap);
 96
 97/* These are not inline because of header tangles. */
 98#ifdef CONFIG_CPUMASK_OFFSTACK
 99/**
100 * alloc_cpumask_var_node - allocate a struct cpumask on a given node
101 * @mask: pointer to cpumask_var_t where the cpumask is returned
102 * @flags: GFP_ flags
 
103 *
104 * Only defined when CONFIG_CPUMASK_OFFSTACK=y, otherwise is
105 * a nop returning a constant 1 (in <linux/cpumask.h>)
106 * Returns TRUE if memory allocation succeeded, FALSE otherwise.
 
107 *
108 * In addition, mask will be NULL if this fails.  Note that gcc is
109 * usually smart enough to know that mask can never be NULL if
110 * CONFIG_CPUMASK_OFFSTACK=n, so does code elimination in that case
111 * too.
112 */
113bool alloc_cpumask_var_node(cpumask_var_t *mask, gfp_t flags, int node)
114{
115	*mask = kmalloc_node(cpumask_size(), flags, node);
116
117#ifdef CONFIG_DEBUG_PER_CPU_MAPS
118	if (!*mask) {
119		printk(KERN_ERR "=> alloc_cpumask_var: failed!\n");
120		dump_stack();
121	}
122#endif
123
124	return *mask != NULL;
125}
126EXPORT_SYMBOL(alloc_cpumask_var_node);
127
128bool zalloc_cpumask_var_node(cpumask_var_t *mask, gfp_t flags, int node)
129{
130	return alloc_cpumask_var_node(mask, flags | __GFP_ZERO, node);
131}
132EXPORT_SYMBOL(zalloc_cpumask_var_node);
133
134/**
135 * alloc_cpumask_var - allocate a struct cpumask
136 * @mask: pointer to cpumask_var_t where the cpumask is returned
137 * @flags: GFP_ flags
138 *
139 * Only defined when CONFIG_CPUMASK_OFFSTACK=y, otherwise is
140 * a nop returning a constant 1 (in <linux/cpumask.h>).
141 *
142 * See alloc_cpumask_var_node.
143 */
144bool alloc_cpumask_var(cpumask_var_t *mask, gfp_t flags)
145{
146	return alloc_cpumask_var_node(mask, flags, NUMA_NO_NODE);
147}
148EXPORT_SYMBOL(alloc_cpumask_var);
149
150bool zalloc_cpumask_var(cpumask_var_t *mask, gfp_t flags)
151{
152	return alloc_cpumask_var(mask, flags | __GFP_ZERO);
153}
154EXPORT_SYMBOL(zalloc_cpumask_var);
155
156/**
157 * alloc_bootmem_cpumask_var - allocate a struct cpumask from the bootmem arena.
158 * @mask: pointer to cpumask_var_t where the cpumask is returned
159 *
160 * Only defined when CONFIG_CPUMASK_OFFSTACK=y, otherwise is
161 * a nop (in <linux/cpumask.h>).
162 * Either returns an allocated (zero-filled) cpumask, or causes the
163 * system to panic.
164 */
165void __init alloc_bootmem_cpumask_var(cpumask_var_t *mask)
166{
167	*mask = memblock_alloc(cpumask_size(), SMP_CACHE_BYTES);
168	if (!*mask)
169		panic("%s: Failed to allocate %u bytes\n", __func__,
170		      cpumask_size());
171}
172
173/**
174 * free_cpumask_var - frees memory allocated for a struct cpumask.
175 * @mask: cpumask to free
176 *
177 * This is safe on a NULL mask.
178 */
179void free_cpumask_var(cpumask_var_t mask)
180{
181	kfree(mask);
182}
183EXPORT_SYMBOL(free_cpumask_var);
184
185/**
186 * free_bootmem_cpumask_var - frees result of alloc_bootmem_cpumask_var
187 * @mask: cpumask to free
188 */
189void __init free_bootmem_cpumask_var(cpumask_var_t mask)
190{
191	memblock_free_early(__pa(mask), cpumask_size());
192}
193#endif
194
195/**
196 * cpumask_local_spread - select the i'th cpu with local numa cpu's first
197 * @i: index number
198 * @node: local numa_node
199 *
200 * This function selects an online CPU according to a numa aware policy;
201 * local cpus are returned first, followed by non-local ones, then it
202 * wraps around.
 
 
203 *
204 * It's not very efficient, but useful for setup.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
205 */
206unsigned int cpumask_local_spread(unsigned int i, int node)
207{
208	int cpu;
209
210	/* Wrap: we always want a cpu. */
211	i %= num_online_cpus();
212
213	if (node == NUMA_NO_NODE) {
214		for_each_cpu(cpu, cpu_online_mask)
215			if (i-- == 0)
216				return cpu;
217	} else {
218		/* NUMA first. */
219		for_each_cpu_and(cpu, cpumask_of_node(node), cpu_online_mask)
220			if (i-- == 0)
221				return cpu;
222
223		for_each_cpu(cpu, cpu_online_mask) {
224			/* Skip NUMA nodes, done above. */
225			if (cpumask_test_cpu(cpu, cpumask_of_node(node)))
226				continue;
227
228			if (i-- == 0)
229				return cpu;
230		}
231	}
232	BUG();
233}
234EXPORT_SYMBOL(cpumask_local_spread);