xref: /openbmc/linux/kernel/irq/affinity.c (revision b593bce5)
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 				unsigned 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_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_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_cpumask(cpumask_var_t *masks)
75 {
76 	int cpu;
77 
78 	for_each_possible_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_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_cpumask[n])) {
90 			node_set(n, *nodemsk);
91 			nodes++;
92 		}
93 	}
94 	return nodes;
95 }
96 
97 static int __irq_build_affinity_masks(unsigned int startvec,
98 				      unsigned int numvecs,
99 				      unsigned int firstvec,
100 				      cpumask_var_t *node_to_cpumask,
101 				      const struct cpumask *cpu_mask,
102 				      struct cpumask *nmsk,
103 				      struct irq_affinity_desc *masks)
104 {
105 	unsigned int n, nodes, cpus_per_vec, extra_vecs, done = 0;
106 	unsigned int last_affv = firstvec + numvecs;
107 	unsigned int curvec = startvec;
108 	nodemask_t nodemsk = NODE_MASK_NONE;
109 
110 	if (!cpumask_weight(cpu_mask))
111 		return 0;
112 
113 	nodes = get_nodes_in_cpumask(node_to_cpumask, cpu_mask, &nodemsk);
114 
115 	/*
116 	 * If the number of nodes in the mask is greater than or equal the
117 	 * number of vectors we just spread the vectors across the nodes.
118 	 */
119 	if (numvecs <= nodes) {
120 		for_each_node_mask(n, nodemsk) {
121 			cpumask_or(&masks[curvec].mask, &masks[curvec].mask,
122 				   node_to_cpumask[n]);
123 			if (++curvec == last_affv)
124 				curvec = firstvec;
125 		}
126 		return numvecs;
127 	}
128 
129 	for_each_node_mask(n, nodemsk) {
130 		unsigned int ncpus, v, vecs_to_assign, vecs_per_node;
131 
132 		/* Spread the vectors per node */
133 		vecs_per_node = (numvecs - (curvec - firstvec)) / nodes;
134 
135 		/* Get the cpus on this node which are in the mask */
136 		cpumask_and(nmsk, cpu_mask, node_to_cpumask[n]);
137 
138 		/* Calculate the number of cpus per vector */
139 		ncpus = cpumask_weight(nmsk);
140 		vecs_to_assign = min(vecs_per_node, ncpus);
141 
142 		/* Account for rounding errors */
143 		extra_vecs = ncpus - vecs_to_assign * (ncpus / vecs_to_assign);
144 
145 		for (v = 0; curvec < last_affv && v < vecs_to_assign;
146 		     curvec++, v++) {
147 			cpus_per_vec = ncpus / vecs_to_assign;
148 
149 			/* Account for extra vectors to compensate rounding errors */
150 			if (extra_vecs) {
151 				cpus_per_vec++;
152 				--extra_vecs;
153 			}
154 			irq_spread_init_one(&masks[curvec].mask, nmsk,
155 						cpus_per_vec);
156 		}
157 
158 		done += v;
159 		if (done >= numvecs)
160 			break;
161 		if (curvec >= last_affv)
162 			curvec = firstvec;
163 		--nodes;
164 	}
165 	return done;
166 }
167 
168 /*
169  * build affinity in two stages:
170  *	1) spread present CPU on these vectors
171  *	2) spread other possible CPUs on these vectors
172  */
173 static int irq_build_affinity_masks(unsigned int startvec, unsigned int numvecs,
174 				    unsigned int firstvec,
175 				    struct irq_affinity_desc *masks)
176 {
177 	unsigned int curvec = startvec, nr_present, nr_others;
178 	cpumask_var_t *node_to_cpumask;
179 	cpumask_var_t nmsk, npresmsk;
180 	int ret = -ENOMEM;
181 
182 	if (!zalloc_cpumask_var(&nmsk, GFP_KERNEL))
183 		return ret;
184 
185 	if (!zalloc_cpumask_var(&npresmsk, GFP_KERNEL))
186 		goto fail_nmsk;
187 
188 	node_to_cpumask = alloc_node_to_cpumask();
189 	if (!node_to_cpumask)
190 		goto fail_npresmsk;
191 
192 	ret = 0;
193 	/* Stabilize the cpumasks */
194 	get_online_cpus();
195 	build_node_to_cpumask(node_to_cpumask);
196 
197 	/* Spread on present CPUs starting from affd->pre_vectors */
198 	nr_present = __irq_build_affinity_masks(curvec, numvecs,
199 						firstvec, node_to_cpumask,
200 						cpu_present_mask, nmsk, masks);
201 
202 	/*
203 	 * Spread on non present CPUs starting from the next vector to be
204 	 * handled. If the spreading of present CPUs already exhausted the
205 	 * vector space, assign the non present CPUs to the already spread
206 	 * out vectors.
207 	 */
208 	if (nr_present >= numvecs)
209 		curvec = firstvec;
210 	else
211 		curvec = firstvec + nr_present;
212 	cpumask_andnot(npresmsk, cpu_possible_mask, cpu_present_mask);
213 	nr_others = __irq_build_affinity_masks(curvec, numvecs,
214 					       firstvec, node_to_cpumask,
215 					       npresmsk, nmsk, masks);
216 	put_online_cpus();
217 
218 	if (nr_present < numvecs)
219 		WARN_ON(nr_present + nr_others < numvecs);
220 
221 	free_node_to_cpumask(node_to_cpumask);
222 
223  fail_npresmsk:
224 	free_cpumask_var(npresmsk);
225 
226  fail_nmsk:
227 	free_cpumask_var(nmsk);
228 	return ret;
229 }
230 
231 static void default_calc_sets(struct irq_affinity *affd, unsigned int affvecs)
232 {
233 	affd->nr_sets = 1;
234 	affd->set_size[0] = affvecs;
235 }
236 
237 /**
238  * irq_create_affinity_masks - Create affinity masks for multiqueue spreading
239  * @nvecs:	The total number of vectors
240  * @affd:	Description of the affinity requirements
241  *
242  * Returns the irq_affinity_desc pointer or NULL if allocation failed.
243  */
244 struct irq_affinity_desc *
245 irq_create_affinity_masks(unsigned int nvecs, struct irq_affinity *affd)
246 {
247 	unsigned int affvecs, curvec, usedvecs, i;
248 	struct irq_affinity_desc *masks = NULL;
249 
250 	/*
251 	 * Determine the number of vectors which need interrupt affinities
252 	 * assigned. If the pre/post request exhausts the available vectors
253 	 * then nothing to do here except for invoking the calc_sets()
254 	 * callback so the device driver can adjust to the situation. If there
255 	 * is only a single vector, then managing the queue is pointless as
256 	 * well.
257 	 */
258 	if (nvecs > 1 && nvecs > affd->pre_vectors + affd->post_vectors)
259 		affvecs = nvecs - affd->pre_vectors - affd->post_vectors;
260 	else
261 		affvecs = 0;
262 
263 	/*
264 	 * Simple invocations do not provide a calc_sets() callback. Install
265 	 * the generic one.
266 	 */
267 	if (!affd->calc_sets)
268 		affd->calc_sets = default_calc_sets;
269 
270 	/* Recalculate the sets */
271 	affd->calc_sets(affd, affvecs);
272 
273 	if (WARN_ON_ONCE(affd->nr_sets > IRQ_AFFINITY_MAX_SETS))
274 		return NULL;
275 
276 	/* Nothing to assign? */
277 	if (!affvecs)
278 		return NULL;
279 
280 	masks = kcalloc(nvecs, sizeof(*masks), GFP_KERNEL);
281 	if (!masks)
282 		return NULL;
283 
284 	/* Fill out vectors at the beginning that don't need affinity */
285 	for (curvec = 0; curvec < affd->pre_vectors; curvec++)
286 		cpumask_copy(&masks[curvec].mask, irq_default_affinity);
287 
288 	/*
289 	 * Spread on present CPUs starting from affd->pre_vectors. If we
290 	 * have multiple sets, build each sets affinity mask separately.
291 	 */
292 	for (i = 0, usedvecs = 0; i < affd->nr_sets; i++) {
293 		unsigned int this_vecs = affd->set_size[i];
294 		int ret;
295 
296 		ret = irq_build_affinity_masks(curvec, this_vecs,
297 					       curvec, masks);
298 		if (ret) {
299 			kfree(masks);
300 			return NULL;
301 		}
302 		curvec += this_vecs;
303 		usedvecs += this_vecs;
304 	}
305 
306 	/* Fill out vectors at the end that don't need affinity */
307 	if (usedvecs >= affvecs)
308 		curvec = affd->pre_vectors + affvecs;
309 	else
310 		curvec = affd->pre_vectors + usedvecs;
311 	for (; curvec < nvecs; curvec++)
312 		cpumask_copy(&masks[curvec].mask, irq_default_affinity);
313 
314 	/* Mark the managed interrupts */
315 	for (i = affd->pre_vectors; i < nvecs - affd->post_vectors; i++)
316 		masks[i].is_managed = 1;
317 
318 	return masks;
319 }
320 
321 /**
322  * irq_calc_affinity_vectors - Calculate the optimal number of vectors
323  * @minvec:	The minimum number of vectors available
324  * @maxvec:	The maximum number of vectors available
325  * @affd:	Description of the affinity requirements
326  */
327 unsigned int irq_calc_affinity_vectors(unsigned int minvec, unsigned int maxvec,
328 				       const struct irq_affinity *affd)
329 {
330 	unsigned int resv = affd->pre_vectors + affd->post_vectors;
331 	unsigned int set_vecs;
332 
333 	if (resv > minvec)
334 		return 0;
335 
336 	if (affd->calc_sets) {
337 		set_vecs = maxvec - resv;
338 	} else {
339 		get_online_cpus();
340 		set_vecs = cpumask_weight(cpu_possible_mask);
341 		put_online_cpus();
342 	}
343 
344 	return resv + min(set_vecs, maxvec - resv);
345 }
346