xref: /linux/kernel/irq/affinity.c (revision 96ac6d435100450f0565708d9b885ea2a7400e0a)
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(const struct irq_affinity *affd,
98 				      unsigned int startvec,
99 				      unsigned int numvecs,
100 				      unsigned int firstvec,
101 				      cpumask_var_t *node_to_cpumask,
102 				      const struct cpumask *cpu_mask,
103 				      struct cpumask *nmsk,
104 				      struct irq_affinity_desc *masks)
105 {
106 	unsigned int n, nodes, cpus_per_vec, extra_vecs, done = 0;
107 	unsigned int last_affv = firstvec + numvecs;
108 	unsigned int curvec = startvec;
109 	nodemask_t nodemsk = NODE_MASK_NONE;
110 
111 	if (!cpumask_weight(cpu_mask))
112 		return 0;
113 
114 	nodes = get_nodes_in_cpumask(node_to_cpumask, cpu_mask, &nodemsk);
115 
116 	/*
117 	 * If the number of nodes in the mask is greater than or equal the
118 	 * number of vectors we just spread the vectors across the nodes.
119 	 */
120 	if (numvecs <= nodes) {
121 		for_each_node_mask(n, nodemsk) {
122 			cpumask_or(&masks[curvec].mask, &masks[curvec].mask,
123 				   node_to_cpumask[n]);
124 			if (++curvec == last_affv)
125 				curvec = firstvec;
126 		}
127 		return numvecs;
128 	}
129 
130 	for_each_node_mask(n, nodemsk) {
131 		unsigned int ncpus, v, vecs_to_assign, vecs_per_node;
132 
133 		/* Spread the vectors per node */
134 		vecs_per_node = (numvecs - (curvec - firstvec)) / nodes;
135 
136 		/* Get the cpus on this node which are in the mask */
137 		cpumask_and(nmsk, cpu_mask, node_to_cpumask[n]);
138 
139 		/* Calculate the number of cpus per vector */
140 		ncpus = cpumask_weight(nmsk);
141 		vecs_to_assign = min(vecs_per_node, ncpus);
142 
143 		/* Account for rounding errors */
144 		extra_vecs = ncpus - vecs_to_assign * (ncpus / vecs_to_assign);
145 
146 		for (v = 0; curvec < last_affv && v < vecs_to_assign;
147 		     curvec++, v++) {
148 			cpus_per_vec = ncpus / vecs_to_assign;
149 
150 			/* Account for extra vectors to compensate rounding errors */
151 			if (extra_vecs) {
152 				cpus_per_vec++;
153 				--extra_vecs;
154 			}
155 			irq_spread_init_one(&masks[curvec].mask, nmsk,
156 						cpus_per_vec);
157 		}
158 
159 		done += v;
160 		if (done >= numvecs)
161 			break;
162 		if (curvec >= last_affv)
163 			curvec = firstvec;
164 		--nodes;
165 	}
166 	return done;
167 }
168 
169 /*
170  * build affinity in two stages:
171  *	1) spread present CPU on these vectors
172  *	2) spread other possible CPUs on these vectors
173  */
174 static int irq_build_affinity_masks(const struct irq_affinity *affd,
175 				    unsigned int startvec, unsigned int numvecs,
176 				    unsigned int firstvec,
177 				    struct irq_affinity_desc *masks)
178 {
179 	unsigned int curvec = startvec, nr_present, nr_others;
180 	cpumask_var_t *node_to_cpumask;
181 	cpumask_var_t nmsk, npresmsk;
182 	int ret = -ENOMEM;
183 
184 	if (!zalloc_cpumask_var(&nmsk, GFP_KERNEL))
185 		return ret;
186 
187 	if (!zalloc_cpumask_var(&npresmsk, GFP_KERNEL))
188 		goto fail_nmsk;
189 
190 	node_to_cpumask = alloc_node_to_cpumask();
191 	if (!node_to_cpumask)
192 		goto fail_npresmsk;
193 
194 	ret = 0;
195 	/* Stabilize the cpumasks */
196 	get_online_cpus();
197 	build_node_to_cpumask(node_to_cpumask);
198 
199 	/* Spread on present CPUs starting from affd->pre_vectors */
200 	nr_present = __irq_build_affinity_masks(affd, curvec, numvecs,
201 						firstvec, node_to_cpumask,
202 						cpu_present_mask, nmsk, masks);
203 
204 	/*
205 	 * Spread on non present CPUs starting from the next vector to be
206 	 * handled. If the spreading of present CPUs already exhausted the
207 	 * vector space, assign the non present CPUs to the already spread
208 	 * out vectors.
209 	 */
210 	if (nr_present >= numvecs)
211 		curvec = firstvec;
212 	else
213 		curvec = firstvec + nr_present;
214 	cpumask_andnot(npresmsk, cpu_possible_mask, cpu_present_mask);
215 	nr_others = __irq_build_affinity_masks(affd, curvec, numvecs,
216 					       firstvec, node_to_cpumask,
217 					       npresmsk, nmsk, masks);
218 	put_online_cpus();
219 
220 	if (nr_present < numvecs)
221 		WARN_ON(nr_present + nr_others < numvecs);
222 
223 	free_node_to_cpumask(node_to_cpumask);
224 
225  fail_npresmsk:
226 	free_cpumask_var(npresmsk);
227 
228  fail_nmsk:
229 	free_cpumask_var(nmsk);
230 	return ret;
231 }
232 
233 static void default_calc_sets(struct irq_affinity *affd, unsigned int affvecs)
234 {
235 	affd->nr_sets = 1;
236 	affd->set_size[0] = affvecs;
237 }
238 
239 /**
240  * irq_create_affinity_masks - Create affinity masks for multiqueue spreading
241  * @nvecs:	The total number of vectors
242  * @affd:	Description of the affinity requirements
243  *
244  * Returns the irq_affinity_desc pointer or NULL if allocation failed.
245  */
246 struct irq_affinity_desc *
247 irq_create_affinity_masks(unsigned int nvecs, struct irq_affinity *affd)
248 {
249 	unsigned int affvecs, curvec, usedvecs, i;
250 	struct irq_affinity_desc *masks = NULL;
251 
252 	/*
253 	 * Determine the number of vectors which need interrupt affinities
254 	 * assigned. If the pre/post request exhausts the available vectors
255 	 * then nothing to do here except for invoking the calc_sets()
256 	 * callback so the device driver can adjust to the situation. If there
257 	 * is only a single vector, then managing the queue is pointless as
258 	 * well.
259 	 */
260 	if (nvecs > 1 && nvecs > affd->pre_vectors + affd->post_vectors)
261 		affvecs = nvecs - affd->pre_vectors - affd->post_vectors;
262 	else
263 		affvecs = 0;
264 
265 	/*
266 	 * Simple invocations do not provide a calc_sets() callback. Install
267 	 * the generic one.
268 	 */
269 	if (!affd->calc_sets)
270 		affd->calc_sets = default_calc_sets;
271 
272 	/* Recalculate the sets */
273 	affd->calc_sets(affd, affvecs);
274 
275 	if (WARN_ON_ONCE(affd->nr_sets > IRQ_AFFINITY_MAX_SETS))
276 		return NULL;
277 
278 	/* Nothing to assign? */
279 	if (!affvecs)
280 		return NULL;
281 
282 	masks = kcalloc(nvecs, sizeof(*masks), GFP_KERNEL);
283 	if (!masks)
284 		return NULL;
285 
286 	/* Fill out vectors at the beginning that don't need affinity */
287 	for (curvec = 0; curvec < affd->pre_vectors; curvec++)
288 		cpumask_copy(&masks[curvec].mask, irq_default_affinity);
289 
290 	/*
291 	 * Spread on present CPUs starting from affd->pre_vectors. If we
292 	 * have multiple sets, build each sets affinity mask separately.
293 	 */
294 	for (i = 0, usedvecs = 0; i < affd->nr_sets; i++) {
295 		unsigned int this_vecs = affd->set_size[i];
296 		int ret;
297 
298 		ret = irq_build_affinity_masks(affd, curvec, this_vecs,
299 					       curvec, masks);
300 		if (ret) {
301 			kfree(masks);
302 			return NULL;
303 		}
304 		curvec += this_vecs;
305 		usedvecs += this_vecs;
306 	}
307 
308 	/* Fill out vectors at the end that don't need affinity */
309 	if (usedvecs >= affvecs)
310 		curvec = affd->pre_vectors + affvecs;
311 	else
312 		curvec = affd->pre_vectors + usedvecs;
313 	for (; curvec < nvecs; curvec++)
314 		cpumask_copy(&masks[curvec].mask, irq_default_affinity);
315 
316 	/* Mark the managed interrupts */
317 	for (i = affd->pre_vectors; i < nvecs - affd->post_vectors; i++)
318 		masks[i].is_managed = 1;
319 
320 	return masks;
321 }
322 
323 /**
324  * irq_calc_affinity_vectors - Calculate the optimal number of vectors
325  * @minvec:	The minimum number of vectors available
326  * @maxvec:	The maximum number of vectors available
327  * @affd:	Description of the affinity requirements
328  */
329 unsigned int irq_calc_affinity_vectors(unsigned int minvec, unsigned int maxvec,
330 				       const struct irq_affinity *affd)
331 {
332 	unsigned int resv = affd->pre_vectors + affd->post_vectors;
333 	unsigned int set_vecs;
334 
335 	if (resv > minvec)
336 		return 0;
337 
338 	if (affd->calc_sets) {
339 		set_vecs = maxvec - resv;
340 	} else {
341 		get_online_cpus();
342 		set_vecs = cpumask_weight(cpu_possible_mask);
343 		put_online_cpus();
344 	}
345 
346 	return resv + min(set_vecs, maxvec - resv);
347 }
348