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_copy(masks + curvec, node_to_cpumask[n]);
 121                         if (++done == numvecs)
 122                                 break;
 123                         if (++curvec == last_affv)
 124                                 curvec = affd->pre_vectors;
 125                 }
 126                 goto out;
 127         }
 128
 129         for_each_node_mask(n, nodemsk) {
 130                 int ncpus, v, vecs_to_assign, vecs_per_node;
 131
 132                 /* Spread the vectors per node */
 133                 vecs_per_node = (numvecs - (curvec - affd->pre_vectors)) / nodes;
 134
 135                 /* Get the cpus on this node which are in the mask */
 136                 cpumask_and(nmsk, cpu_mask, node_to_cpumask[n]);
 137
 138                 /* Calculate the number of cpus per vector */
 139                 ncpus = cpumask_weight(nmsk);
 140                 vecs_to_assign = min(vecs_per_node, ncpus);
 141
 142                 /* Account for rounding errors */
 143                 extra_vecs = ncpus - vecs_to_assign * (ncpus / vecs_to_assign);
 144
 145                 for (v = 0; curvec < last_affv && v < vecs_to_assign;
 146                      curvec++, v++) {
 147                         cpus_per_vec = ncpus / vecs_to_assign;
 148
 149                         /* Account for extra vectors to compensate rounding errors */
 150                         if (extra_vecs) {
 151                                 cpus_per_vec++;
 152                                 --extra_vecs;
 153                         }
 154                         irq_spread_init_one(masks + curvec, nmsk, cpus_per_vec);
 155                 }
 156
 157                 done += v;
 158                 if (done >= numvecs)
 159                         break;
 160                 if (curvec >= last_affv)
 161                         curvec = affd->pre_vectors;
 162                 --nodes;
 163         }
 164
 165 out:
 166         return done;
 167 }
 168
 169 /**
 170  * irq_create_affinity_masks - Create affinity masks for multiqueue spreading
 171  * @nvecs:      The total number of vectors
 172  * @affd:       Description of the affinity requirements
 173  *
 174  * Returns the masks pointer or NULL if allocation failed.
 175  */
 176 struct cpumask *
 177 irq_create_affinity_masks(int nvecs, const struct irq_affinity *affd)
 178 {
 179         int affvecs = nvecs - affd->pre_vectors - affd->post_vectors;
 180         int curvec, usedvecs;
 181         cpumask_var_t nmsk, npresmsk, *node_to_cpumask;
 182         struct cpumask *masks = NULL;
 183
 184         /*
 185          * If there aren't any vectors left after applying the pre/post
 186          * vectors don't bother with assigning affinity.
 187          */
 188         if (nvecs == affd->pre_vectors + affd->post_vectors)
 189                 return NULL;
 190
 191         if (!zalloc_cpumask_var(&nmsk, GFP_KERNEL))
 192                 return NULL;
 193
 194         if (!zalloc_cpumask_var(&npresmsk, GFP_KERNEL))
 195                 goto outcpumsk;
 196
 197         node_to_cpumask = alloc_node_to_cpumask();
 198         if (!node_to_cpumask)
 199                 goto outnpresmsk;
 200
 201         masks = kcalloc(nvecs, sizeof(*masks), GFP_KERNEL);
 202         if (!masks)
 203                 goto outnodemsk;
 204
 205         /* Fill out vectors at the beginning that don't need affinity */
 206         for (curvec = 0; curvec < affd->pre_vectors; curvec++)
 207                 cpumask_copy(masks + curvec, irq_default_affinity);
 208
 209         /* Stabilize the cpumasks */
 210         get_online_cpus();
 211         build_node_to_cpumask(node_to_cpumask);
 212
 213         /* Spread on present CPUs starting from affd->pre_vectors */
 214         usedvecs = irq_build_affinity_masks(affd, curvec, affvecs,
 215                                             node_to_cpumask, cpu_present_mask,
 216                                             nmsk, masks);
 217
 218         /*
 219          * Spread on non present CPUs starting from the next vector to be
 220          * handled. If the spreading of present CPUs already exhausted the
 221          * vector space, assign the non present CPUs to the already spread
 222          * out vectors.
 223          */
 224         if (usedvecs >= affvecs)
 225                 curvec = affd->pre_vectors;
 226         else
 227                 curvec = affd->pre_vectors + usedvecs;
 228         cpumask_andnot(npresmsk, cpu_possible_mask, cpu_present_mask);
 229         usedvecs += irq_build_affinity_masks(affd, curvec, affvecs,
 230                                              node_to_cpumask, npresmsk,
 231                                              nmsk, masks);
 232         put_online_cpus();
 233
 234         /* Fill out vectors at the end that don't need affinity */
 235         if (usedvecs >= affvecs)
 236                 curvec = affd->pre_vectors + affvecs;
 237         else
 238                 curvec = affd->pre_vectors + usedvecs;
 239         for (; curvec < nvecs; curvec++)
 240                 cpumask_copy(masks + curvec, irq_default_affinity);
 241
 242 outnodemsk:
 243         free_node_to_cpumask(node_to_cpumask);
 244 outnpresmsk:
 245         free_cpumask_var(npresmsk);
 246 outcpumsk:
 247         free_cpumask_var(nmsk);
 248         return masks;
 249 }
 250
 251 /**
 252  * irq_calc_affinity_vectors - Calculate the optimal number of vectors
 253  * @minvec:     The minimum number of vectors available
 254  * @maxvec:     The maximum number of vectors available
 255  * @affd:       Description of the affinity requirements
 256  */
 257 int irq_calc_affinity_vectors(int minvec, int maxvec, const struct irq_affinity *affd)
 258 {
 259         int resv = affd->pre_vectors + affd->post_vectors;
 260         int vecs = maxvec - resv;
 261         int ret;
 262
 263         if (resv > minvec)
 264                 return 0;
 265
 266         get_online_cpus();
 267         ret = min_t(int, cpumask_weight(cpu_possible_mask), vecs) + resv;
 268         put_online_cpus();
 269         return ret;
 270 }