xref: /linux/kernel/irq/affinity.c (revision c358f53871605a1a8d7ed6e544a05ea00e9c80cb)
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 #include <linux/sort.h>
11 
12 static void irq_spread_init_one(struct cpumask *irqmsk, struct cpumask *nmsk,
13 				unsigned int cpus_per_vec)
14 {
15 	const struct cpumask *siblmsk;
16 	int cpu, sibl;
17 
18 	for ( ; cpus_per_vec > 0; ) {
19 		cpu = cpumask_first(nmsk);
20 
21 		/* Should not happen, but I'm too lazy to think about it */
22 		if (cpu >= nr_cpu_ids)
23 			return;
24 
25 		cpumask_clear_cpu(cpu, nmsk);
26 		cpumask_set_cpu(cpu, irqmsk);
27 		cpus_per_vec--;
28 
29 		/* If the cpu has siblings, use them first */
30 		siblmsk = topology_sibling_cpumask(cpu);
31 		for (sibl = -1; cpus_per_vec > 0; ) {
32 			sibl = cpumask_next(sibl, siblmsk);
33 			if (sibl >= nr_cpu_ids)
34 				break;
35 			if (!cpumask_test_and_clear_cpu(sibl, nmsk))
36 				continue;
37 			cpumask_set_cpu(sibl, irqmsk);
38 			cpus_per_vec--;
39 		}
40 	}
41 }
42 
43 static cpumask_var_t *alloc_node_to_cpumask(void)
44 {
45 	cpumask_var_t *masks;
46 	int node;
47 
48 	masks = kcalloc(nr_node_ids, sizeof(cpumask_var_t), GFP_KERNEL);
49 	if (!masks)
50 		return NULL;
51 
52 	for (node = 0; node < nr_node_ids; node++) {
53 		if (!zalloc_cpumask_var(&masks[node], GFP_KERNEL))
54 			goto out_unwind;
55 	}
56 
57 	return masks;
58 
59 out_unwind:
60 	while (--node >= 0)
61 		free_cpumask_var(masks[node]);
62 	kfree(masks);
63 	return NULL;
64 }
65 
66 static void free_node_to_cpumask(cpumask_var_t *masks)
67 {
68 	int node;
69 
70 	for (node = 0; node < nr_node_ids; node++)
71 		free_cpumask_var(masks[node]);
72 	kfree(masks);
73 }
74 
75 static void build_node_to_cpumask(cpumask_var_t *masks)
76 {
77 	int cpu;
78 
79 	for_each_possible_cpu(cpu)
80 		cpumask_set_cpu(cpu, masks[cpu_to_node(cpu)]);
81 }
82 
83 static int get_nodes_in_cpumask(cpumask_var_t *node_to_cpumask,
84 				const struct cpumask *mask, nodemask_t *nodemsk)
85 {
86 	int n, nodes = 0;
87 
88 	/* Calculate the number of nodes in the supplied affinity mask */
89 	for_each_node(n) {
90 		if (cpumask_intersects(mask, node_to_cpumask[n])) {
91 			node_set(n, *nodemsk);
92 			nodes++;
93 		}
94 	}
95 	return nodes;
96 }
97 
98 struct node_vectors {
99 	unsigned id;
100 
101 	union {
102 		unsigned nvectors;
103 		unsigned ncpus;
104 	};
105 };
106 
107 static int ncpus_cmp_func(const void *l, const void *r)
108 {
109 	const struct node_vectors *ln = l;
110 	const struct node_vectors *rn = r;
111 
112 	return ln->ncpus - rn->ncpus;
113 }
114 
115 /*
116  * Allocate vector number for each node, so that for each node:
117  *
118  * 1) the allocated number is >= 1
119  *
120  * 2) the allocated numbver is <= active CPU number of this node
121  *
122  * The actual allocated total vectors may be less than @numvecs when
123  * active total CPU number is less than @numvecs.
124  *
125  * Active CPUs means the CPUs in '@cpu_mask AND @node_to_cpumask[]'
126  * for each node.
127  */
128 static void alloc_nodes_vectors(unsigned int numvecs,
129 				cpumask_var_t *node_to_cpumask,
130 				const struct cpumask *cpu_mask,
131 				const nodemask_t nodemsk,
132 				struct cpumask *nmsk,
133 				struct node_vectors *node_vectors)
134 {
135 	unsigned n, remaining_ncpus = 0;
136 
137 	for (n = 0; n < nr_node_ids; n++) {
138 		node_vectors[n].id = n;
139 		node_vectors[n].ncpus = UINT_MAX;
140 	}
141 
142 	for_each_node_mask(n, nodemsk) {
143 		unsigned ncpus;
144 
145 		cpumask_and(nmsk, cpu_mask, node_to_cpumask[n]);
146 		ncpus = cpumask_weight(nmsk);
147 
148 		if (!ncpus)
149 			continue;
150 		remaining_ncpus += ncpus;
151 		node_vectors[n].ncpus = ncpus;
152 	}
153 
154 	numvecs = min_t(unsigned, remaining_ncpus, numvecs);
155 
156 	sort(node_vectors, nr_node_ids, sizeof(node_vectors[0]),
157 	     ncpus_cmp_func, NULL);
158 
159 	/*
160 	 * Allocate vectors for each node according to the ratio of this
161 	 * node's nr_cpus to remaining un-assigned ncpus. 'numvecs' is
162 	 * bigger than number of active numa nodes. Always start the
163 	 * allocation from the node with minimized nr_cpus.
164 	 *
165 	 * This way guarantees that each active node gets allocated at
166 	 * least one vector, and the theory is simple: over-allocation
167 	 * is only done when this node is assigned by one vector, so
168 	 * other nodes will be allocated >= 1 vector, since 'numvecs' is
169 	 * bigger than number of numa nodes.
170 	 *
171 	 * One perfect invariant is that number of allocated vectors for
172 	 * each node is <= CPU count of this node:
173 	 *
174 	 * 1) suppose there are two nodes: A and B
175 	 * 	ncpu(X) is CPU count of node X
176 	 * 	vecs(X) is the vector count allocated to node X via this
177 	 * 	algorithm
178 	 *
179 	 * 	ncpu(A) <= ncpu(B)
180 	 * 	ncpu(A) + ncpu(B) = N
181 	 * 	vecs(A) + vecs(B) = V
182 	 *
183 	 * 	vecs(A) = max(1, round_down(V * ncpu(A) / N))
184 	 * 	vecs(B) = V - vecs(A)
185 	 *
186 	 * 	both N and V are integer, and 2 <= V <= N, suppose
187 	 * 	V = N - delta, and 0 <= delta <= N - 2
188 	 *
189 	 * 2) obviously vecs(A) <= ncpu(A) because:
190 	 *
191 	 * 	if vecs(A) is 1, then vecs(A) <= ncpu(A) given
192 	 * 	ncpu(A) >= 1
193 	 *
194 	 * 	otherwise,
195 	 * 		vecs(A) <= V * ncpu(A) / N <= ncpu(A), given V <= N
196 	 *
197 	 * 3) prove how vecs(B) <= ncpu(B):
198 	 *
199 	 * 	if round_down(V * ncpu(A) / N) == 0, vecs(B) won't be
200 	 * 	over-allocated, so vecs(B) <= ncpu(B),
201 	 *
202 	 * 	otherwise:
203 	 *
204 	 * 	vecs(A) =
205 	 * 		round_down(V * ncpu(A) / N) =
206 	 * 		round_down((N - delta) * ncpu(A) / N) =
207 	 * 		round_down((N * ncpu(A) - delta * ncpu(A)) / N)	 >=
208 	 * 		round_down((N * ncpu(A) - delta * N) / N)	 =
209 	 * 		cpu(A) - delta
210 	 *
211 	 * 	then:
212 	 *
213 	 * 	vecs(A) - V >= ncpu(A) - delta - V
214 	 * 	=>
215 	 * 	V - vecs(A) <= V + delta - ncpu(A)
216 	 * 	=>
217 	 * 	vecs(B) <= N - ncpu(A)
218 	 * 	=>
219 	 * 	vecs(B) <= cpu(B)
220 	 *
221 	 * For nodes >= 3, it can be thought as one node and another big
222 	 * node given that is exactly what this algorithm is implemented,
223 	 * and we always re-calculate 'remaining_ncpus' & 'numvecs', and
224 	 * finally for each node X: vecs(X) <= ncpu(X).
225 	 *
226 	 */
227 	for (n = 0; n < nr_node_ids; n++) {
228 		unsigned nvectors, ncpus;
229 
230 		if (node_vectors[n].ncpus == UINT_MAX)
231 			continue;
232 
233 		WARN_ON_ONCE(numvecs == 0);
234 
235 		ncpus = node_vectors[n].ncpus;
236 		nvectors = max_t(unsigned, 1,
237 				 numvecs * ncpus / remaining_ncpus);
238 		WARN_ON_ONCE(nvectors > ncpus);
239 
240 		node_vectors[n].nvectors = nvectors;
241 
242 		remaining_ncpus -= ncpus;
243 		numvecs -= nvectors;
244 	}
245 }
246 
247 static int __irq_build_affinity_masks(unsigned int startvec,
248 				      unsigned int numvecs,
249 				      unsigned int firstvec,
250 				      cpumask_var_t *node_to_cpumask,
251 				      const struct cpumask *cpu_mask,
252 				      struct cpumask *nmsk,
253 				      struct irq_affinity_desc *masks)
254 {
255 	unsigned int i, n, nodes, cpus_per_vec, extra_vecs, done = 0;
256 	unsigned int last_affv = firstvec + numvecs;
257 	unsigned int curvec = startvec;
258 	nodemask_t nodemsk = NODE_MASK_NONE;
259 	struct node_vectors *node_vectors;
260 
261 	if (cpumask_empty(cpu_mask))
262 		return 0;
263 
264 	nodes = get_nodes_in_cpumask(node_to_cpumask, cpu_mask, &nodemsk);
265 
266 	/*
267 	 * If the number of nodes in the mask is greater than or equal the
268 	 * number of vectors we just spread the vectors across the nodes.
269 	 */
270 	if (numvecs <= nodes) {
271 		for_each_node_mask(n, nodemsk) {
272 			/* Ensure that only CPUs which are in both masks are set */
273 			cpumask_and(nmsk, cpu_mask, node_to_cpumask[n]);
274 			cpumask_or(&masks[curvec].mask, &masks[curvec].mask, nmsk);
275 			if (++curvec == last_affv)
276 				curvec = firstvec;
277 		}
278 		return numvecs;
279 	}
280 
281 	node_vectors = kcalloc(nr_node_ids,
282 			       sizeof(struct node_vectors),
283 			       GFP_KERNEL);
284 	if (!node_vectors)
285 		return -ENOMEM;
286 
287 	/* allocate vector number for each node */
288 	alloc_nodes_vectors(numvecs, node_to_cpumask, cpu_mask,
289 			    nodemsk, nmsk, node_vectors);
290 
291 	for (i = 0; i < nr_node_ids; i++) {
292 		unsigned int ncpus, v;
293 		struct node_vectors *nv = &node_vectors[i];
294 
295 		if (nv->nvectors == UINT_MAX)
296 			continue;
297 
298 		/* Get the cpus on this node which are in the mask */
299 		cpumask_and(nmsk, cpu_mask, node_to_cpumask[nv->id]);
300 		ncpus = cpumask_weight(nmsk);
301 		if (!ncpus)
302 			continue;
303 
304 		WARN_ON_ONCE(nv->nvectors > ncpus);
305 
306 		/* Account for rounding errors */
307 		extra_vecs = ncpus - nv->nvectors * (ncpus / nv->nvectors);
308 
309 		/* Spread allocated vectors on CPUs of the current node */
310 		for (v = 0; v < nv->nvectors; v++, curvec++) {
311 			cpus_per_vec = ncpus / nv->nvectors;
312 
313 			/* Account for extra vectors to compensate rounding errors */
314 			if (extra_vecs) {
315 				cpus_per_vec++;
316 				--extra_vecs;
317 			}
318 
319 			/*
320 			 * wrapping has to be considered given 'startvec'
321 			 * may start anywhere
322 			 */
323 			if (curvec >= last_affv)
324 				curvec = firstvec;
325 			irq_spread_init_one(&masks[curvec].mask, nmsk,
326 						cpus_per_vec);
327 		}
328 		done += nv->nvectors;
329 	}
330 	kfree(node_vectors);
331 	return done;
332 }
333 
334 /*
335  * build affinity in two stages:
336  *	1) spread present CPU on these vectors
337  *	2) spread other possible CPUs on these vectors
338  */
339 static int irq_build_affinity_masks(unsigned int startvec, unsigned int numvecs,
340 				    unsigned int firstvec,
341 				    struct irq_affinity_desc *masks)
342 {
343 	unsigned int curvec = startvec, nr_present = 0, nr_others = 0;
344 	cpumask_var_t *node_to_cpumask;
345 	cpumask_var_t nmsk, npresmsk;
346 	int ret = -ENOMEM;
347 
348 	if (!zalloc_cpumask_var(&nmsk, GFP_KERNEL))
349 		return ret;
350 
351 	if (!zalloc_cpumask_var(&npresmsk, GFP_KERNEL))
352 		goto fail_nmsk;
353 
354 	node_to_cpumask = alloc_node_to_cpumask();
355 	if (!node_to_cpumask)
356 		goto fail_npresmsk;
357 
358 	/* Stabilize the cpumasks */
359 	cpus_read_lock();
360 	build_node_to_cpumask(node_to_cpumask);
361 
362 	/* Spread on present CPUs starting from affd->pre_vectors */
363 	ret = __irq_build_affinity_masks(curvec, numvecs, firstvec,
364 					 node_to_cpumask, cpu_present_mask,
365 					 nmsk, masks);
366 	if (ret < 0)
367 		goto fail_build_affinity;
368 	nr_present = ret;
369 
370 	/*
371 	 * Spread on non present CPUs starting from the next vector to be
372 	 * handled. If the spreading of present CPUs already exhausted the
373 	 * vector space, assign the non present CPUs to the already spread
374 	 * out vectors.
375 	 */
376 	if (nr_present >= numvecs)
377 		curvec = firstvec;
378 	else
379 		curvec = firstvec + nr_present;
380 	cpumask_andnot(npresmsk, cpu_possible_mask, cpu_present_mask);
381 	ret = __irq_build_affinity_masks(curvec, numvecs, firstvec,
382 					 node_to_cpumask, npresmsk, nmsk,
383 					 masks);
384 	if (ret >= 0)
385 		nr_others = ret;
386 
387  fail_build_affinity:
388 	cpus_read_unlock();
389 
390 	if (ret >= 0)
391 		WARN_ON(nr_present + nr_others < numvecs);
392 
393 	free_node_to_cpumask(node_to_cpumask);
394 
395  fail_npresmsk:
396 	free_cpumask_var(npresmsk);
397 
398  fail_nmsk:
399 	free_cpumask_var(nmsk);
400 	return ret < 0 ? ret : 0;
401 }
402 
403 static void default_calc_sets(struct irq_affinity *affd, unsigned int affvecs)
404 {
405 	affd->nr_sets = 1;
406 	affd->set_size[0] = affvecs;
407 }
408 
409 /**
410  * irq_create_affinity_masks - Create affinity masks for multiqueue spreading
411  * @nvecs:	The total number of vectors
412  * @affd:	Description of the affinity requirements
413  *
414  * Returns the irq_affinity_desc pointer or NULL if allocation failed.
415  */
416 struct irq_affinity_desc *
417 irq_create_affinity_masks(unsigned int nvecs, struct irq_affinity *affd)
418 {
419 	unsigned int affvecs, curvec, usedvecs, i;
420 	struct irq_affinity_desc *masks = NULL;
421 
422 	/*
423 	 * Determine the number of vectors which need interrupt affinities
424 	 * assigned. If the pre/post request exhausts the available vectors
425 	 * then nothing to do here except for invoking the calc_sets()
426 	 * callback so the device driver can adjust to the situation.
427 	 */
428 	if (nvecs > affd->pre_vectors + affd->post_vectors)
429 		affvecs = nvecs - affd->pre_vectors - affd->post_vectors;
430 	else
431 		affvecs = 0;
432 
433 	/*
434 	 * Simple invocations do not provide a calc_sets() callback. Install
435 	 * the generic one.
436 	 */
437 	if (!affd->calc_sets)
438 		affd->calc_sets = default_calc_sets;
439 
440 	/* Recalculate the sets */
441 	affd->calc_sets(affd, affvecs);
442 
443 	if (WARN_ON_ONCE(affd->nr_sets > IRQ_AFFINITY_MAX_SETS))
444 		return NULL;
445 
446 	/* Nothing to assign? */
447 	if (!affvecs)
448 		return NULL;
449 
450 	masks = kcalloc(nvecs, sizeof(*masks), GFP_KERNEL);
451 	if (!masks)
452 		return NULL;
453 
454 	/* Fill out vectors at the beginning that don't need affinity */
455 	for (curvec = 0; curvec < affd->pre_vectors; curvec++)
456 		cpumask_copy(&masks[curvec].mask, irq_default_affinity);
457 
458 	/*
459 	 * Spread on present CPUs starting from affd->pre_vectors. If we
460 	 * have multiple sets, build each sets affinity mask separately.
461 	 */
462 	for (i = 0, usedvecs = 0; i < affd->nr_sets; i++) {
463 		unsigned int this_vecs = affd->set_size[i];
464 		int ret;
465 
466 		ret = irq_build_affinity_masks(curvec, this_vecs,
467 					       curvec, masks);
468 		if (ret) {
469 			kfree(masks);
470 			return NULL;
471 		}
472 		curvec += this_vecs;
473 		usedvecs += this_vecs;
474 	}
475 
476 	/* Fill out vectors at the end that don't need affinity */
477 	if (usedvecs >= affvecs)
478 		curvec = affd->pre_vectors + affvecs;
479 	else
480 		curvec = affd->pre_vectors + usedvecs;
481 	for (; curvec < nvecs; curvec++)
482 		cpumask_copy(&masks[curvec].mask, irq_default_affinity);
483 
484 	/* Mark the managed interrupts */
485 	for (i = affd->pre_vectors; i < nvecs - affd->post_vectors; i++)
486 		masks[i].is_managed = 1;
487 
488 	return masks;
489 }
490 
491 /**
492  * irq_calc_affinity_vectors - Calculate the optimal number of vectors
493  * @minvec:	The minimum number of vectors available
494  * @maxvec:	The maximum number of vectors available
495  * @affd:	Description of the affinity requirements
496  */
497 unsigned int irq_calc_affinity_vectors(unsigned int minvec, unsigned int maxvec,
498 				       const struct irq_affinity *affd)
499 {
500 	unsigned int resv = affd->pre_vectors + affd->post_vectors;
501 	unsigned int set_vecs;
502 
503 	if (resv > minvec)
504 		return 0;
505 
506 	if (affd->calc_sets) {
507 		set_vecs = maxvec - resv;
508 	} else {
509 		cpus_read_lock();
510 		set_vecs = cpumask_weight(cpu_possible_mask);
511 		cpus_read_unlock();
512 	}
513 
514 	return resv + min(set_vecs, maxvec - resv);
515 }
516