xref: /linux/lib/group_cpus.c (revision 4b132aacb0768ac1e652cf517097ea6f237214b9)
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/kernel.h>
7 #include <linux/slab.h>
8 #include <linux/cpu.h>
9 #include <linux/sort.h>
10 #include <linux/group_cpus.h>
11 
12 #ifdef CONFIG_SMP
13 
14 static void grp_spread_init_one(struct cpumask *irqmsk, struct cpumask *nmsk,
15 				unsigned int cpus_per_grp)
16 {
17 	const struct cpumask *siblmsk;
18 	int cpu, sibl;
19 
20 	for ( ; cpus_per_grp > 0; ) {
21 		cpu = cpumask_first(nmsk);
22 
23 		/* Should not happen, but I'm too lazy to think about it */
24 		if (cpu >= nr_cpu_ids)
25 			return;
26 
27 		cpumask_clear_cpu(cpu, nmsk);
28 		cpumask_set_cpu(cpu, irqmsk);
29 		cpus_per_grp--;
30 
31 		/* If the cpu has siblings, use them first */
32 		siblmsk = topology_sibling_cpumask(cpu);
33 		for (sibl = -1; cpus_per_grp > 0; ) {
34 			sibl = cpumask_next(sibl, siblmsk);
35 			if (sibl >= nr_cpu_ids)
36 				break;
37 			if (!cpumask_test_and_clear_cpu(sibl, nmsk))
38 				continue;
39 			cpumask_set_cpu(sibl, irqmsk);
40 			cpus_per_grp--;
41 		}
42 	}
43 }
44 
45 static cpumask_var_t *alloc_node_to_cpumask(void)
46 {
47 	cpumask_var_t *masks;
48 	int node;
49 
50 	masks = kcalloc(nr_node_ids, sizeof(cpumask_var_t), GFP_KERNEL);
51 	if (!masks)
52 		return NULL;
53 
54 	for (node = 0; node < nr_node_ids; node++) {
55 		if (!zalloc_cpumask_var(&masks[node], GFP_KERNEL))
56 			goto out_unwind;
57 	}
58 
59 	return masks;
60 
61 out_unwind:
62 	while (--node >= 0)
63 		free_cpumask_var(masks[node]);
64 	kfree(masks);
65 	return NULL;
66 }
67 
68 static void free_node_to_cpumask(cpumask_var_t *masks)
69 {
70 	int node;
71 
72 	for (node = 0; node < nr_node_ids; node++)
73 		free_cpumask_var(masks[node]);
74 	kfree(masks);
75 }
76 
77 static void build_node_to_cpumask(cpumask_var_t *masks)
78 {
79 	int cpu;
80 
81 	for_each_possible_cpu(cpu)
82 		cpumask_set_cpu(cpu, masks[cpu_to_node(cpu)]);
83 }
84 
85 static int get_nodes_in_cpumask(cpumask_var_t *node_to_cpumask,
86 				const struct cpumask *mask, nodemask_t *nodemsk)
87 {
88 	int n, nodes = 0;
89 
90 	/* Calculate the number of nodes in the supplied affinity mask */
91 	for_each_node(n) {
92 		if (cpumask_intersects(mask, node_to_cpumask[n])) {
93 			node_set(n, *nodemsk);
94 			nodes++;
95 		}
96 	}
97 	return nodes;
98 }
99 
100 struct node_groups {
101 	unsigned id;
102 
103 	union {
104 		unsigned ngroups;
105 		unsigned ncpus;
106 	};
107 };
108 
109 static int ncpus_cmp_func(const void *l, const void *r)
110 {
111 	const struct node_groups *ln = l;
112 	const struct node_groups *rn = r;
113 
114 	return ln->ncpus - rn->ncpus;
115 }
116 
117 /*
118  * Allocate group number for each node, so that for each node:
119  *
120  * 1) the allocated number is >= 1
121  *
122  * 2) the allocated number is <= active CPU number of this node
123  *
124  * The actual allocated total groups may be less than @numgrps when
125  * active total CPU number is less than @numgrps.
126  *
127  * Active CPUs means the CPUs in '@cpu_mask AND @node_to_cpumask[]'
128  * for each node.
129  */
130 static void alloc_nodes_groups(unsigned int numgrps,
131 			       cpumask_var_t *node_to_cpumask,
132 			       const struct cpumask *cpu_mask,
133 			       const nodemask_t nodemsk,
134 			       struct cpumask *nmsk,
135 			       struct node_groups *node_groups)
136 {
137 	unsigned n, remaining_ncpus = 0;
138 
139 	for (n = 0; n < nr_node_ids; n++) {
140 		node_groups[n].id = n;
141 		node_groups[n].ncpus = UINT_MAX;
142 	}
143 
144 	for_each_node_mask(n, nodemsk) {
145 		unsigned ncpus;
146 
147 		cpumask_and(nmsk, cpu_mask, node_to_cpumask[n]);
148 		ncpus = cpumask_weight(nmsk);
149 
150 		if (!ncpus)
151 			continue;
152 		remaining_ncpus += ncpus;
153 		node_groups[n].ncpus = ncpus;
154 	}
155 
156 	numgrps = min_t(unsigned, remaining_ncpus, numgrps);
157 
158 	sort(node_groups, nr_node_ids, sizeof(node_groups[0]),
159 	     ncpus_cmp_func, NULL);
160 
161 	/*
162 	 * Allocate groups for each node according to the ratio of this
163 	 * node's nr_cpus to remaining un-assigned ncpus. 'numgrps' is
164 	 * bigger than number of active numa nodes. Always start the
165 	 * allocation from the node with minimized nr_cpus.
166 	 *
167 	 * This way guarantees that each active node gets allocated at
168 	 * least one group, and the theory is simple: over-allocation
169 	 * is only done when this node is assigned by one group, so
170 	 * other nodes will be allocated >= 1 groups, since 'numgrps' is
171 	 * bigger than number of numa nodes.
172 	 *
173 	 * One perfect invariant is that number of allocated groups for
174 	 * each node is <= CPU count of this node:
175 	 *
176 	 * 1) suppose there are two nodes: A and B
177 	 * 	ncpu(X) is CPU count of node X
178 	 * 	grps(X) is the group count allocated to node X via this
179 	 * 	algorithm
180 	 *
181 	 * 	ncpu(A) <= ncpu(B)
182 	 * 	ncpu(A) + ncpu(B) = N
183 	 * 	grps(A) + grps(B) = G
184 	 *
185 	 * 	grps(A) = max(1, round_down(G * ncpu(A) / N))
186 	 * 	grps(B) = G - grps(A)
187 	 *
188 	 * 	both N and G are integer, and 2 <= G <= N, suppose
189 	 * 	G = N - delta, and 0 <= delta <= N - 2
190 	 *
191 	 * 2) obviously grps(A) <= ncpu(A) because:
192 	 *
193 	 * 	if grps(A) is 1, then grps(A) <= ncpu(A) given
194 	 * 	ncpu(A) >= 1
195 	 *
196 	 * 	otherwise,
197 	 * 		grps(A) <= G * ncpu(A) / N <= ncpu(A), given G <= N
198 	 *
199 	 * 3) prove how grps(B) <= ncpu(B):
200 	 *
201 	 * 	if round_down(G * ncpu(A) / N) == 0, vecs(B) won't be
202 	 * 	over-allocated, so grps(B) <= ncpu(B),
203 	 *
204 	 * 	otherwise:
205 	 *
206 	 * 	grps(A) =
207 	 * 		round_down(G * ncpu(A) / N) =
208 	 * 		round_down((N - delta) * ncpu(A) / N) =
209 	 * 		round_down((N * ncpu(A) - delta * ncpu(A)) / N)	 >=
210 	 * 		round_down((N * ncpu(A) - delta * N) / N)	 =
211 	 * 		cpu(A) - delta
212 	 *
213 	 * 	then:
214 	 *
215 	 * 	grps(A) - G >= ncpu(A) - delta - G
216 	 * 	=>
217 	 * 	G - grps(A) <= G + delta - ncpu(A)
218 	 * 	=>
219 	 * 	grps(B) <= N - ncpu(A)
220 	 * 	=>
221 	 * 	grps(B) <= cpu(B)
222 	 *
223 	 * For nodes >= 3, it can be thought as one node and another big
224 	 * node given that is exactly what this algorithm is implemented,
225 	 * and we always re-calculate 'remaining_ncpus' & 'numgrps', and
226 	 * finally for each node X: grps(X) <= ncpu(X).
227 	 *
228 	 */
229 	for (n = 0; n < nr_node_ids; n++) {
230 		unsigned ngroups, ncpus;
231 
232 		if (node_groups[n].ncpus == UINT_MAX)
233 			continue;
234 
235 		WARN_ON_ONCE(numgrps == 0);
236 
237 		ncpus = node_groups[n].ncpus;
238 		ngroups = max_t(unsigned, 1,
239 				 numgrps * ncpus / remaining_ncpus);
240 		WARN_ON_ONCE(ngroups > ncpus);
241 
242 		node_groups[n].ngroups = ngroups;
243 
244 		remaining_ncpus -= ncpus;
245 		numgrps -= ngroups;
246 	}
247 }
248 
249 static int __group_cpus_evenly(unsigned int startgrp, unsigned int numgrps,
250 			       cpumask_var_t *node_to_cpumask,
251 			       const struct cpumask *cpu_mask,
252 			       struct cpumask *nmsk, struct cpumask *masks)
253 {
254 	unsigned int i, n, nodes, cpus_per_grp, extra_grps, done = 0;
255 	unsigned int last_grp = numgrps;
256 	unsigned int curgrp = startgrp;
257 	nodemask_t nodemsk = NODE_MASK_NONE;
258 	struct node_groups *node_groups;
259 
260 	if (cpumask_empty(cpu_mask))
261 		return 0;
262 
263 	nodes = get_nodes_in_cpumask(node_to_cpumask, cpu_mask, &nodemsk);
264 
265 	/*
266 	 * If the number of nodes in the mask is greater than or equal the
267 	 * number of groups we just spread the groups across the nodes.
268 	 */
269 	if (numgrps <= nodes) {
270 		for_each_node_mask(n, nodemsk) {
271 			/* Ensure that only CPUs which are in both masks are set */
272 			cpumask_and(nmsk, cpu_mask, node_to_cpumask[n]);
273 			cpumask_or(&masks[curgrp], &masks[curgrp], nmsk);
274 			if (++curgrp == last_grp)
275 				curgrp = 0;
276 		}
277 		return numgrps;
278 	}
279 
280 	node_groups = kcalloc(nr_node_ids,
281 			       sizeof(struct node_groups),
282 			       GFP_KERNEL);
283 	if (!node_groups)
284 		return -ENOMEM;
285 
286 	/* allocate group number for each node */
287 	alloc_nodes_groups(numgrps, node_to_cpumask, cpu_mask,
288 			   nodemsk, nmsk, node_groups);
289 	for (i = 0; i < nr_node_ids; i++) {
290 		unsigned int ncpus, v;
291 		struct node_groups *nv = &node_groups[i];
292 
293 		if (nv->ngroups == UINT_MAX)
294 			continue;
295 
296 		/* Get the cpus on this node which are in the mask */
297 		cpumask_and(nmsk, cpu_mask, node_to_cpumask[nv->id]);
298 		ncpus = cpumask_weight(nmsk);
299 		if (!ncpus)
300 			continue;
301 
302 		WARN_ON_ONCE(nv->ngroups > ncpus);
303 
304 		/* Account for rounding errors */
305 		extra_grps = ncpus - nv->ngroups * (ncpus / nv->ngroups);
306 
307 		/* Spread allocated groups on CPUs of the current node */
308 		for (v = 0; v < nv->ngroups; v++, curgrp++) {
309 			cpus_per_grp = ncpus / nv->ngroups;
310 
311 			/* Account for extra groups to compensate rounding errors */
312 			if (extra_grps) {
313 				cpus_per_grp++;
314 				--extra_grps;
315 			}
316 
317 			/*
318 			 * wrapping has to be considered given 'startgrp'
319 			 * may start anywhere
320 			 */
321 			if (curgrp >= last_grp)
322 				curgrp = 0;
323 			grp_spread_init_one(&masks[curgrp], nmsk,
324 						cpus_per_grp);
325 		}
326 		done += nv->ngroups;
327 	}
328 	kfree(node_groups);
329 	return done;
330 }
331 
332 /**
333  * group_cpus_evenly - Group all CPUs evenly per NUMA/CPU locality
334  * @numgrps: number of groups
335  *
336  * Return: cpumask array if successful, NULL otherwise. And each element
337  * includes CPUs assigned to this group
338  *
339  * Try to put close CPUs from viewpoint of CPU and NUMA locality into
340  * same group, and run two-stage grouping:
341  *	1) allocate present CPUs on these groups evenly first
342  *	2) allocate other possible CPUs on these groups evenly
343  *
344  * We guarantee in the resulted grouping that all CPUs are covered, and
345  * no same CPU is assigned to multiple groups
346  */
347 struct cpumask *group_cpus_evenly(unsigned int numgrps)
348 {
349 	unsigned int curgrp = 0, nr_present = 0, nr_others = 0;
350 	cpumask_var_t *node_to_cpumask;
351 	cpumask_var_t nmsk, npresmsk;
352 	int ret = -ENOMEM;
353 	struct cpumask *masks = NULL;
354 
355 	if (!zalloc_cpumask_var(&nmsk, GFP_KERNEL))
356 		return NULL;
357 
358 	if (!zalloc_cpumask_var(&npresmsk, GFP_KERNEL))
359 		goto fail_nmsk;
360 
361 	node_to_cpumask = alloc_node_to_cpumask();
362 	if (!node_to_cpumask)
363 		goto fail_npresmsk;
364 
365 	masks = kcalloc(numgrps, sizeof(*masks), GFP_KERNEL);
366 	if (!masks)
367 		goto fail_node_to_cpumask;
368 
369 	build_node_to_cpumask(node_to_cpumask);
370 
371 	/*
372 	 * Make a local cache of 'cpu_present_mask', so the two stages
373 	 * spread can observe consistent 'cpu_present_mask' without holding
374 	 * cpu hotplug lock, then we can reduce deadlock risk with cpu
375 	 * hotplug code.
376 	 *
377 	 * Here CPU hotplug may happen when reading `cpu_present_mask`, and
378 	 * we can live with the case because it only affects that hotplug
379 	 * CPU is handled in the 1st or 2nd stage, and either way is correct
380 	 * from API user viewpoint since 2-stage spread is sort of
381 	 * optimization.
382 	 */
383 	cpumask_copy(npresmsk, data_race(cpu_present_mask));
384 
385 	/* grouping present CPUs first */
386 	ret = __group_cpus_evenly(curgrp, numgrps, node_to_cpumask,
387 				  npresmsk, nmsk, masks);
388 	if (ret < 0)
389 		goto fail_build_affinity;
390 	nr_present = ret;
391 
392 	/*
393 	 * Allocate non present CPUs starting from the next group to be
394 	 * handled. If the grouping of present CPUs already exhausted the
395 	 * group space, assign the non present CPUs to the already
396 	 * allocated out groups.
397 	 */
398 	if (nr_present >= numgrps)
399 		curgrp = 0;
400 	else
401 		curgrp = nr_present;
402 	cpumask_andnot(npresmsk, cpu_possible_mask, npresmsk);
403 	ret = __group_cpus_evenly(curgrp, numgrps, node_to_cpumask,
404 				  npresmsk, nmsk, masks);
405 	if (ret >= 0)
406 		nr_others = ret;
407 
408  fail_build_affinity:
409 	if (ret >= 0)
410 		WARN_ON(nr_present + nr_others < numgrps);
411 
412  fail_node_to_cpumask:
413 	free_node_to_cpumask(node_to_cpumask);
414 
415  fail_npresmsk:
416 	free_cpumask_var(npresmsk);
417 
418  fail_nmsk:
419 	free_cpumask_var(nmsk);
420 	if (ret < 0) {
421 		kfree(masks);
422 		return NULL;
423 	}
424 	return masks;
425 }
426 #else /* CONFIG_SMP */
427 struct cpumask *group_cpus_evenly(unsigned int numgrps)
428 {
429 	struct cpumask *masks = kcalloc(numgrps, sizeof(*masks), GFP_KERNEL);
430 
431 	if (!masks)
432 		return NULL;
433 
434 	/* assign all CPUs(cpu 0) to the 1st group only */
435 	cpumask_copy(&masks[0], cpu_possible_mask);
436 	return masks;
437 }
438 #endif /* CONFIG_SMP */
439 EXPORT_SYMBOL_GPL(group_cpus_evenly);
440