1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (C) 2016 Thomas Gleixner.
4 * Copyright (C) 2016-2017 Christoph Hellwig.
6 #include <linux/kernel.h>
7 #include <linux/slab.h>
9 #include <linux/sort.h>
10 #include <linux/group_cpus.h>
14 static void grp_spread_init_one(struct cpumask *irqmsk, struct cpumask *nmsk,
15 unsigned int cpus_per_grp)
17 const struct cpumask *siblmsk;
20 for ( ; cpus_per_grp > 0; ) {
21 cpu = cpumask_first(nmsk);
23 /* Should not happen, but I'm too lazy to think about it */
24 if (cpu >= nr_cpu_ids)
27 cpumask_clear_cpu(cpu, nmsk);
28 cpumask_set_cpu(cpu, irqmsk);
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)
37 if (!cpumask_test_and_clear_cpu(sibl, nmsk))
39 cpumask_set_cpu(sibl, irqmsk);
45 static cpumask_var_t *alloc_node_to_cpumask(void)
50 masks = kcalloc(nr_node_ids, sizeof(cpumask_var_t), GFP_KERNEL);
54 for (node = 0; node < nr_node_ids; node++) {
55 if (!zalloc_cpumask_var(&masks[node], GFP_KERNEL))
63 free_cpumask_var(masks[node]);
68 static void free_node_to_cpumask(cpumask_var_t *masks)
72 for (node = 0; node < nr_node_ids; node++)
73 free_cpumask_var(masks[node]);
77 static void build_node_to_cpumask(cpumask_var_t *masks)
81 for_each_possible_cpu(cpu)
82 cpumask_set_cpu(cpu, masks[cpu_to_node(cpu)]);
85 static int get_nodes_in_cpumask(cpumask_var_t *node_to_cpumask,
86 const struct cpumask *mask, nodemask_t *nodemsk)
90 /* Calculate the number of nodes in the supplied affinity mask */
92 if (cpumask_intersects(mask, node_to_cpumask[n])) {
93 node_set(n, *nodemsk);
109 static int ncpus_cmp_func(const void *l, const void *r)
111 const struct node_groups *ln = l;
112 const struct node_groups *rn = r;
114 return ln->ncpus - rn->ncpus;
118 * Allocate group number for each node, so that for each node:
120 * 1) the allocated number is >= 1
122 * 2) the allocated number is <= active CPU number of this node
124 * The actual allocated total groups may be less than @numgrps when
125 * active total CPU number is less than @numgrps.
127 * Active CPUs means the CPUs in '@cpu_mask AND @node_to_cpumask[]'
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)
137 unsigned n, remaining_ncpus = 0;
139 for (n = 0; n < nr_node_ids; n++) {
140 node_groups[n].id = n;
141 node_groups[n].ncpus = UINT_MAX;
144 for_each_node_mask(n, nodemsk) {
147 cpumask_and(nmsk, cpu_mask, node_to_cpumask[n]);
148 ncpus = cpumask_weight(nmsk);
152 remaining_ncpus += ncpus;
153 node_groups[n].ncpus = ncpus;
156 numgrps = min_t(unsigned, remaining_ncpus, numgrps);
158 sort(node_groups, nr_node_ids, sizeof(node_groups[0]),
159 ncpus_cmp_func, NULL);
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.
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.
173 * One perfect invariant is that number of allocated groups for
174 * each node is <= CPU count of this node:
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
182 * ncpu(A) + ncpu(B) = N
183 * grps(A) + grps(B) = G
185 * grps(A) = max(1, round_down(G * ncpu(A) / N))
186 * grps(B) = G - grps(A)
188 * both N and G are integer, and 2 <= G <= N, suppose
189 * G = N - delta, and 0 <= delta <= N - 2
191 * 2) obviously grps(A) <= ncpu(A) because:
193 * if grps(A) is 1, then grps(A) <= ncpu(A) given
197 * grps(A) <= G * ncpu(A) / N <= ncpu(A), given G <= N
199 * 3) prove how grps(B) <= ncpu(B):
201 * if round_down(G * ncpu(A) / N) == 0, vecs(B) won't be
202 * over-allocated, so grps(B) <= ncpu(B),
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) =
215 * grps(A) - G >= ncpu(A) - delta - G
217 * G - grps(A) <= G + delta - ncpu(A)
219 * grps(B) <= N - ncpu(A)
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).
229 for (n = 0; n < nr_node_ids; n++) {
230 unsigned ngroups, ncpus;
232 if (node_groups[n].ncpus == UINT_MAX)
235 WARN_ON_ONCE(numgrps == 0);
237 ncpus = node_groups[n].ncpus;
238 ngroups = max_t(unsigned, 1,
239 numgrps * ncpus / remaining_ncpus);
240 WARN_ON_ONCE(ngroups > ncpus);
242 node_groups[n].ngroups = ngroups;
244 remaining_ncpus -= ncpus;
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)
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;
260 if (cpumask_empty(cpu_mask))
263 nodes = get_nodes_in_cpumask(node_to_cpumask, cpu_mask, &nodemsk);
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.
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)
280 node_groups = kcalloc(nr_node_ids,
281 sizeof(struct node_groups),
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];
293 if (nv->ngroups == UINT_MAX)
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);
302 WARN_ON_ONCE(nv->ngroups > ncpus);
304 /* Account for rounding errors */
305 extra_grps = ncpus - nv->ngroups * (ncpus / nv->ngroups);
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;
311 /* Account for extra groups to compensate rounding errors */
318 * wrapping has to be considered given 'startgrp'
321 if (curgrp >= last_grp)
323 grp_spread_init_one(&masks[curgrp], nmsk,
333 * group_cpus_evenly - Group all CPUs evenly per NUMA/CPU locality
334 * @numgrps: number of groups
336 * Return: cpumask array if successful, NULL otherwise. And each element
337 * includes CPUs assigned to this group
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
344 * We guarantee in the resulted grouping that all CPUs are covered, and
345 * no same CPU is assigned to multiple groups
347 struct cpumask *group_cpus_evenly(unsigned int numgrps)
349 unsigned int curgrp = 0, nr_present = 0, nr_others = 0;
350 cpumask_var_t *node_to_cpumask;
351 cpumask_var_t nmsk, npresmsk;
353 struct cpumask *masks = NULL;
355 if (!zalloc_cpumask_var(&nmsk, GFP_KERNEL))
358 if (!zalloc_cpumask_var(&npresmsk, GFP_KERNEL))
361 node_to_cpumask = alloc_node_to_cpumask();
362 if (!node_to_cpumask)
365 masks = kcalloc(numgrps, sizeof(*masks), GFP_KERNEL);
367 goto fail_node_to_cpumask;
369 /* Stabilize the cpumasks */
371 build_node_to_cpumask(node_to_cpumask);
373 /* grouping present CPUs first */
374 ret = __group_cpus_evenly(curgrp, numgrps, node_to_cpumask,
375 cpu_present_mask, nmsk, masks);
377 goto fail_build_affinity;
381 * Allocate non present CPUs starting from the next group to be
382 * handled. If the grouping of present CPUs already exhausted the
383 * group space, assign the non present CPUs to the already
384 * allocated out groups.
386 if (nr_present >= numgrps)
390 cpumask_andnot(npresmsk, cpu_possible_mask, cpu_present_mask);
391 ret = __group_cpus_evenly(curgrp, numgrps, node_to_cpumask,
392 npresmsk, nmsk, masks);
400 WARN_ON(nr_present + nr_others < numgrps);
402 fail_node_to_cpumask:
403 free_node_to_cpumask(node_to_cpumask);
406 free_cpumask_var(npresmsk);
409 free_cpumask_var(nmsk);
416 #else /* CONFIG_SMP */
417 struct cpumask *group_cpus_evenly(unsigned int numgrps)
419 struct cpumask *masks = kcalloc(numgrps, sizeof(*masks), GFP_KERNEL);
424 /* assign all CPUs(cpu 0) to the 1st group only */
425 cpumask_copy(&masks[0], cpu_possible_mask);
428 #endif /* CONFIG_SMP */