xref: /openbmc/linux/kernel/irq/affinity.c (revision e00a844a)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright (C) 2016 Thomas Gleixner.
4  * Copyright (C) 2016-2017 Christoph Hellwig.
5  */
6 #include <linux/interrupt.h>
7 #include <linux/kernel.h>
8 #include <linux/slab.h>
9 #include <linux/cpu.h>
10 
11 static void irq_spread_init_one(struct cpumask *irqmsk, struct cpumask *nmsk,
12 				int cpus_per_vec)
13 {
14 	const struct cpumask *siblmsk;
15 	int cpu, sibl;
16 
17 	for ( ; cpus_per_vec > 0; ) {
18 		cpu = cpumask_first(nmsk);
19 
20 		/* Should not happen, but I'm too lazy to think about it */
21 		if (cpu >= nr_cpu_ids)
22 			return;
23 
24 		cpumask_clear_cpu(cpu, nmsk);
25 		cpumask_set_cpu(cpu, irqmsk);
26 		cpus_per_vec--;
27 
28 		/* If the cpu has siblings, use them first */
29 		siblmsk = topology_sibling_cpumask(cpu);
30 		for (sibl = -1; cpus_per_vec > 0; ) {
31 			sibl = cpumask_next(sibl, siblmsk);
32 			if (sibl >= nr_cpu_ids)
33 				break;
34 			if (!cpumask_test_and_clear_cpu(sibl, nmsk))
35 				continue;
36 			cpumask_set_cpu(sibl, irqmsk);
37 			cpus_per_vec--;
38 		}
39 	}
40 }
41 
42 static cpumask_var_t *alloc_node_to_present_cpumask(void)
43 {
44 	cpumask_var_t *masks;
45 	int node;
46 
47 	masks = kcalloc(nr_node_ids, sizeof(cpumask_var_t), GFP_KERNEL);
48 	if (!masks)
49 		return NULL;
50 
51 	for (node = 0; node < nr_node_ids; node++) {
52 		if (!zalloc_cpumask_var(&masks[node], GFP_KERNEL))
53 			goto out_unwind;
54 	}
55 
56 	return masks;
57 
58 out_unwind:
59 	while (--node >= 0)
60 		free_cpumask_var(masks[node]);
61 	kfree(masks);
62 	return NULL;
63 }
64 
65 static void free_node_to_present_cpumask(cpumask_var_t *masks)
66 {
67 	int node;
68 
69 	for (node = 0; node < nr_node_ids; node++)
70 		free_cpumask_var(masks[node]);
71 	kfree(masks);
72 }
73 
74 static void build_node_to_present_cpumask(cpumask_var_t *masks)
75 {
76 	int cpu;
77 
78 	for_each_present_cpu(cpu)
79 		cpumask_set_cpu(cpu, masks[cpu_to_node(cpu)]);
80 }
81 
82 static int get_nodes_in_cpumask(cpumask_var_t *node_to_present_cpumask,
83 				const struct cpumask *mask, nodemask_t *nodemsk)
84 {
85 	int n, nodes = 0;
86 
87 	/* Calculate the number of nodes in the supplied affinity mask */
88 	for_each_node(n) {
89 		if (cpumask_intersects(mask, node_to_present_cpumask[n])) {
90 			node_set(n, *nodemsk);
91 			nodes++;
92 		}
93 	}
94 	return nodes;
95 }
96 
97 /**
98  * irq_create_affinity_masks - Create affinity masks for multiqueue spreading
99  * @nvecs:	The total number of vectors
100  * @affd:	Description of the affinity requirements
101  *
102  * Returns the masks pointer or NULL if allocation failed.
103  */
104 struct cpumask *
105 irq_create_affinity_masks(int nvecs, const struct irq_affinity *affd)
106 {
107 	int n, nodes, cpus_per_vec, extra_vecs, curvec;
108 	int affv = nvecs - affd->pre_vectors - affd->post_vectors;
109 	int last_affv = affv + affd->pre_vectors;
110 	nodemask_t nodemsk = NODE_MASK_NONE;
111 	struct cpumask *masks;
112 	cpumask_var_t nmsk, *node_to_present_cpumask;
113 
114 	/*
115 	 * If there aren't any vectors left after applying the pre/post
116 	 * vectors don't bother with assigning affinity.
117 	 */
118 	if (!affv)
119 		return NULL;
120 
121 	if (!zalloc_cpumask_var(&nmsk, GFP_KERNEL))
122 		return NULL;
123 
124 	masks = kcalloc(nvecs, sizeof(*masks), GFP_KERNEL);
125 	if (!masks)
126 		goto out;
127 
128 	node_to_present_cpumask = alloc_node_to_present_cpumask();
129 	if (!node_to_present_cpumask)
130 		goto out;
131 
132 	/* Fill out vectors at the beginning that don't need affinity */
133 	for (curvec = 0; curvec < affd->pre_vectors; curvec++)
134 		cpumask_copy(masks + curvec, irq_default_affinity);
135 
136 	/* Stabilize the cpumasks */
137 	get_online_cpus();
138 	build_node_to_present_cpumask(node_to_present_cpumask);
139 	nodes = get_nodes_in_cpumask(node_to_present_cpumask, cpu_present_mask,
140 				     &nodemsk);
141 
142 	/*
143 	 * If the number of nodes in the mask is greater than or equal the
144 	 * number of vectors we just spread the vectors across the nodes.
145 	 */
146 	if (affv <= nodes) {
147 		for_each_node_mask(n, nodemsk) {
148 			cpumask_copy(masks + curvec,
149 				     node_to_present_cpumask[n]);
150 			if (++curvec == last_affv)
151 				break;
152 		}
153 		goto done;
154 	}
155 
156 	for_each_node_mask(n, nodemsk) {
157 		int ncpus, v, vecs_to_assign, vecs_per_node;
158 
159 		/* Spread the vectors per node */
160 		vecs_per_node = (affv - (curvec - affd->pre_vectors)) / nodes;
161 
162 		/* Get the cpus on this node which are in the mask */
163 		cpumask_and(nmsk, cpu_present_mask, node_to_present_cpumask[n]);
164 
165 		/* Calculate the number of cpus per vector */
166 		ncpus = cpumask_weight(nmsk);
167 		vecs_to_assign = min(vecs_per_node, ncpus);
168 
169 		/* Account for rounding errors */
170 		extra_vecs = ncpus - vecs_to_assign * (ncpus / vecs_to_assign);
171 
172 		for (v = 0; curvec < last_affv && v < vecs_to_assign;
173 		     curvec++, v++) {
174 			cpus_per_vec = ncpus / vecs_to_assign;
175 
176 			/* Account for extra vectors to compensate rounding errors */
177 			if (extra_vecs) {
178 				cpus_per_vec++;
179 				--extra_vecs;
180 			}
181 			irq_spread_init_one(masks + curvec, nmsk, cpus_per_vec);
182 		}
183 
184 		if (curvec >= last_affv)
185 			break;
186 		--nodes;
187 	}
188 
189 done:
190 	put_online_cpus();
191 
192 	/* Fill out vectors at the end that don't need affinity */
193 	for (; curvec < nvecs; curvec++)
194 		cpumask_copy(masks + curvec, irq_default_affinity);
195 	free_node_to_present_cpumask(node_to_present_cpumask);
196 out:
197 	free_cpumask_var(nmsk);
198 	return masks;
199 }
200 
201 /**
202  * irq_calc_affinity_vectors - Calculate the optimal number of vectors
203  * @minvec:	The minimum number of vectors available
204  * @maxvec:	The maximum number of vectors available
205  * @affd:	Description of the affinity requirements
206  */
207 int irq_calc_affinity_vectors(int minvec, int maxvec, const struct irq_affinity *affd)
208 {
209 	int resv = affd->pre_vectors + affd->post_vectors;
210 	int vecs = maxvec - resv;
211 	int ret;
212 
213 	if (resv > minvec)
214 		return 0;
215 
216 	get_online_cpus();
217 	ret = min_t(int, cpumask_weight(cpu_present_mask), vecs) + resv;
218 	put_online_cpus();
219 	return ret;
220 }
221