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