kernel/irq/affinity.c

   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                                 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                                     int startvec, int numvecs,
  99                                     cpumask_var_t *node_to_cpumask,
 100                                     const struct cpumask *cpu_mask,
 101                                     struct cpumask *nmsk,
 102                                     struct cpumask *masks)
 103 {
 104         int n, nodes, cpus_per_vec, extra_vecs, done = 0;
 105         int last_affv = affd->pre_vectors + numvecs;
 106         int curvec = startvec;
 107         nodemask_t nodemsk = NODE_MASK_NONE;
 108
 109         if (!cpumask_weight(cpu_mask))
 110                 return 0;
 111
 112         nodes = get_nodes_in_cpumask(node_to_cpumask, cpu_mask, &nodemsk);
 113
 114         /*
 115          * If the number of nodes in the mask is greater than or equal the
 116          * number of vectors we just spread the vectors across the nodes.
 117          */
 118         if (numvecs <= nodes) {
 119                 for_each_node_mask(n, nodemsk) {
 120                         cpumask_or(masks + curvec, masks + curvec, node_to_cpumask[n]);
 121                         if (++curvec == last_affv)
 122                                 curvec = affd->pre_vectors;
 123                 }
 124                 done = numvecs;
 125                 goto out;
 126         }
 127
 128         for_each_node_mask(n, nodemsk) {
 129                 int ncpus, v, vecs_to_assign, vecs_per_node;
 130
 131                 /* Spread the vectors per node */
 132                 vecs_per_node = (numvecs - (curvec - affd->pre_vectors)) / nodes;
 133
 134                 /* Get the cpus on this node which are in the mask */
 135                 cpumask_and(nmsk, cpu_mask, node_to_cpumask[n]);
 136
 137                 /* Calculate the number of cpus per vector */
 138                 ncpus = cpumask_weight(nmsk);
 139                 vecs_to_assign = min(vecs_per_node, ncpus);
 140
 141                 /* Account for rounding errors */
 142                 extra_vecs = ncpus - vecs_to_assign * (ncpus / vecs_to_assign);
 143
 144                 for (v = 0; curvec < last_affv && v < vecs_to_assign;
 145                      curvec++, v++) {
 146                         cpus_per_vec = ncpus / vecs_to_assign;
 147
 148                         /* Account for extra vectors to compensate rounding errors */
 149                         if (extra_vecs) {
 150                                 cpus_per_vec++;
 151                                 --extra_vecs;
 152                         }
 153                         irq_spread_init_one(masks + curvec, nmsk, cpus_per_vec);
 154                 }
 155
 156                 done += v;
 157                 if (done >= numvecs)
 158                         break;
 159                 if (curvec >= last_affv)
 160                         curvec = affd->pre_vectors;
 161                 --nodes;
 162         }
 163
 164 out:
 165         return done;
 166 }
 167
 168 /**
 169  * irq_create_affinity_masks - Create affinity masks for multiqueue spreading
 170  * @nvecs:      The total number of vectors
 171  * @affd:       Description of the affinity requirements
 172  *
 173  * Returns the masks pointer or NULL if allocation failed.
 174  */
 175 struct cpumask *
 176 irq_create_affinity_masks(int nvecs, const struct irq_affinity *affd)
 177 {
 178         int affvecs = nvecs - affd->pre_vectors - affd->post_vectors;
 179         int curvec, usedvecs;
 180         cpumask_var_t nmsk, npresmsk, *node_to_cpumask;
 181         struct cpumask *masks = NULL;
 182
 183         /*
 184          * If there aren't any vectors left after applying the pre/post
 185          * vectors don't bother with assigning affinity.
 186          */
 187         if (nvecs == affd->pre_vectors + affd->post_vectors)
 188                 return NULL;
 189
 190         if (!zalloc_cpumask_var(&nmsk, GFP_KERNEL))
 191                 return NULL;
 192
 193         if (!zalloc_cpumask_var(&npresmsk, GFP_KERNEL))
 194                 goto outcpumsk;
 195
 196         node_to_cpumask = alloc_node_to_cpumask();
 197         if (!node_to_cpumask)
 198                 goto outnpresmsk;
 199
 200         masks = kcalloc(nvecs, sizeof(*masks), GFP_KERNEL);
 201         if (!masks)
 202                 goto outnodemsk;
 203
 204         /* Fill out vectors at the beginning that don't need affinity */
 205         for (curvec = 0; curvec < affd->pre_vectors; curvec++)
 206                 cpumask_copy(masks + curvec, irq_default_affinity);
 207
 208         /* Stabilize the cpumasks */
 209         get_online_cpus();
 210         build_node_to_cpumask(node_to_cpumask);
 211
 212         /* Spread on present CPUs starting from affd->pre_vectors */
 213         usedvecs = irq_build_affinity_masks(affd, curvec, affvecs,
 214                                             node_to_cpumask, cpu_present_mask,
 215                                             nmsk, masks);
 216
 217         /*
 218          * Spread on non present CPUs starting from the next vector to be
 219          * handled. If the spreading of present CPUs already exhausted the
 220          * vector space, assign the non present CPUs to the already spread
 221          * out vectors.
 222          */
 223         if (usedvecs >= affvecs)
 224                 curvec = affd->pre_vectors;
 225         else
 226                 curvec = affd->pre_vectors + usedvecs;
 227         cpumask_andnot(npresmsk, cpu_possible_mask, cpu_present_mask);
 228         usedvecs += irq_build_affinity_masks(affd, curvec, affvecs,
 229                                              node_to_cpumask, npresmsk,
 230                                              nmsk, masks);
 231         put_online_cpus();
 232
 233         /* Fill out vectors at the end that don't need affinity */
 234         if (usedvecs >= affvecs)
 235                 curvec = affd->pre_vectors + affvecs;
 236         else
 237                 curvec = affd->pre_vectors + usedvecs;
 238         for (; curvec < nvecs; curvec++)
 239                 cpumask_copy(masks + curvec, irq_default_affinity);
 240
 241 outnodemsk:
 242         free_node_to_cpumask(node_to_cpumask);
 243 outnpresmsk:
 244         free_cpumask_var(npresmsk);
 245 outcpumsk:
 246         free_cpumask_var(nmsk);
 247         return masks;
 248 }
 249
 250 /**
 251  * irq_calc_affinity_vectors - Calculate the optimal number of vectors
 252  * @minvec:     The minimum number of vectors available
 253  * @maxvec:     The maximum number of vectors available
 254  * @affd:       Description of the affinity requirements
 255  */
 256 int irq_calc_affinity_vectors(int minvec, int maxvec, const struct irq_affinity *affd)
 257 {
 258         int resv = affd->pre_vectors + affd->post_vectors;
 259         int vecs = maxvec - resv;
 260         int ret;
 261
 262         if (resv > minvec)
 263                 return 0;
 264
 265         get_online_cpus();
 266         ret = min_t(int, cpumask_weight(cpu_possible_mask), vecs) + resv;
 267         put_online_cpus();
 268         return ret;
 269 }