drivers/base/cacheinfo.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * cacheinfo support - processor cache information via sysfs
   4  *
   5  * Based on arch/x86/kernel/cpu/intel_cacheinfo.c
   6  * Author: Sudeep Holla <sudeep.holla@arm.com>
   7  */
   8 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   9
  10 #include <linux/acpi.h>
  11 #include <linux/bitops.h>
  12 #include <linux/cacheinfo.h>
  13 #include <linux/compiler.h>
  14 #include <linux/cpu.h>
  15 #include <linux/device.h>
  16 #include <linux/init.h>
  17 #include <linux/of.h>
  18 #include <linux/sched.h>
  19 #include <linux/slab.h>
  20 #include <linux/smp.h>
  21 #include <linux/sysfs.h>
  22
  23 /* pointer to per cpu cacheinfo */
  24 static DEFINE_PER_CPU(struct cpu_cacheinfo, ci_cpu_cacheinfo);
  25 #define ci_cacheinfo(cpu)       (&per_cpu(ci_cpu_cacheinfo, cpu))
  26 #define cache_leaves(cpu)       (ci_cacheinfo(cpu)->num_leaves)
  27 #define per_cpu_cacheinfo(cpu)  (ci_cacheinfo(cpu)->info_list)
  28 #define per_cpu_cacheinfo_idx(cpu, idx)         \
  29                                 (per_cpu_cacheinfo(cpu) + (idx))
  30
  31 /* Set if no cache information is found in DT/ACPI. */
  32 static bool use_arch_info;
  33
  34 struct cpu_cacheinfo *get_cpu_cacheinfo(unsigned int cpu)
  35 {
  36         return ci_cacheinfo(cpu);
  37 }
  38
  39 static inline bool cache_leaves_are_shared(struct cacheinfo *this_leaf,
  40                                            struct cacheinfo *sib_leaf)
  41 {
  42         /*
  43          * For non DT/ACPI systems, assume unique level 1 caches,
  44          * system-wide shared caches for all other levels.
  45          */
  46         if (!(IS_ENABLED(CONFIG_OF) || IS_ENABLED(CONFIG_ACPI)) ||
  47             use_arch_info)
  48                 return (this_leaf->level != 1) && (sib_leaf->level != 1);
  49
  50         if ((sib_leaf->attributes & CACHE_ID) &&
  51             (this_leaf->attributes & CACHE_ID))
  52                 return sib_leaf->id == this_leaf->id;
  53
  54         return sib_leaf->fw_token == this_leaf->fw_token;
  55 }
  56
  57 bool last_level_cache_is_valid(unsigned int cpu)
  58 {
  59         struct cacheinfo *llc;
  60
  61         if (!cache_leaves(cpu))
  62                 return false;
  63
  64         llc = per_cpu_cacheinfo_idx(cpu, cache_leaves(cpu) - 1);
  65
  66         return (llc->attributes & CACHE_ID) || !!llc->fw_token;
  67
  68 }
  69
  70 bool last_level_cache_is_shared(unsigned int cpu_x, unsigned int cpu_y)
  71 {
  72         struct cacheinfo *llc_x, *llc_y;
  73
  74         if (!last_level_cache_is_valid(cpu_x) ||
  75             !last_level_cache_is_valid(cpu_y))
  76                 return false;
  77
  78         llc_x = per_cpu_cacheinfo_idx(cpu_x, cache_leaves(cpu_x) - 1);
  79         llc_y = per_cpu_cacheinfo_idx(cpu_y, cache_leaves(cpu_y) - 1);
  80
  81         return cache_leaves_are_shared(llc_x, llc_y);
  82 }
  83
  84 #ifdef CONFIG_OF
  85
  86 static bool of_check_cache_nodes(struct device_node *np);
  87
  88 /* OF properties to query for a given cache type */
  89 struct cache_type_info {
  90         const char *size_prop;
  91         const char *line_size_props[2];
  92         const char *nr_sets_prop;
  93 };
  94
  95 static const struct cache_type_info cache_type_info[] = {
  96         {
  97                 .size_prop       = "cache-size",
  98                 .line_size_props = { "cache-line-size",
  99                                      "cache-block-size", },
 100                 .nr_sets_prop    = "cache-sets",
 101         }, {
 102                 .size_prop       = "i-cache-size",
 103                 .line_size_props = { "i-cache-line-size",
 104                                      "i-cache-block-size", },
 105                 .nr_sets_prop    = "i-cache-sets",
 106         }, {
 107                 .size_prop       = "d-cache-size",
 108                 .line_size_props = { "d-cache-line-size",
 109                                      "d-cache-block-size", },
 110                 .nr_sets_prop    = "d-cache-sets",
 111         },
 112 };
 113
 114 static inline int get_cacheinfo_idx(enum cache_type type)
 115 {
 116         if (type == CACHE_TYPE_UNIFIED)
 117                 return 0;
 118         return type;
 119 }
 120
 121 static void cache_size(struct cacheinfo *this_leaf, struct device_node *np)
 122 {
 123         const char *propname;
 124         int ct_idx;
 125
 126         ct_idx = get_cacheinfo_idx(this_leaf->type);
 127         propname = cache_type_info[ct_idx].size_prop;
 128
 129         of_property_read_u32(np, propname, &this_leaf->size);
 130 }
 131
 132 /* not cache_line_size() because that's a macro in include/linux/cache.h */
 133 static void cache_get_line_size(struct cacheinfo *this_leaf,
 134                                 struct device_node *np)
 135 {
 136         int i, lim, ct_idx;
 137
 138         ct_idx = get_cacheinfo_idx(this_leaf->type);
 139         lim = ARRAY_SIZE(cache_type_info[ct_idx].line_size_props);
 140
 141         for (i = 0; i < lim; i++) {
 142                 int ret;
 143                 u32 line_size;
 144                 const char *propname;
 145
 146                 propname = cache_type_info[ct_idx].line_size_props[i];
 147                 ret = of_property_read_u32(np, propname, &line_size);
 148                 if (!ret) {
 149                         this_leaf->coherency_line_size = line_size;
 150                         break;
 151                 }
 152         }
 153 }
 154
 155 static void cache_nr_sets(struct cacheinfo *this_leaf, struct device_node *np)
 156 {
 157         const char *propname;
 158         int ct_idx;
 159
 160         ct_idx = get_cacheinfo_idx(this_leaf->type);
 161         propname = cache_type_info[ct_idx].nr_sets_prop;
 162
 163         of_property_read_u32(np, propname, &this_leaf->number_of_sets);
 164 }
 165
 166 static void cache_associativity(struct cacheinfo *this_leaf)
 167 {
 168         unsigned int line_size = this_leaf->coherency_line_size;
 169         unsigned int nr_sets = this_leaf->number_of_sets;
 170         unsigned int size = this_leaf->size;
 171
 172         /*
 173          * If the cache is fully associative, there is no need to
 174          * check the other properties.
 175          */
 176         if (!(nr_sets == 1) && (nr_sets > 0 && size > 0 && line_size > 0))
 177                 this_leaf->ways_of_associativity = (size / nr_sets) / line_size;
 178 }
 179
 180 static bool cache_node_is_unified(struct cacheinfo *this_leaf,
 181                                   struct device_node *np)
 182 {
 183         return of_property_read_bool(np, "cache-unified");
 184 }
 185
 186 static void cache_of_set_props(struct cacheinfo *this_leaf,
 187                                struct device_node *np)
 188 {
 189         /*
 190          * init_cache_level must setup the cache level correctly
 191          * overriding the architecturally specified levels, so
 192          * if type is NONE at this stage, it should be unified
 193          */
 194         if (this_leaf->type == CACHE_TYPE_NOCACHE &&
 195             cache_node_is_unified(this_leaf, np))
 196                 this_leaf->type = CACHE_TYPE_UNIFIED;
 197         cache_size(this_leaf, np);
 198         cache_get_line_size(this_leaf, np);
 199         cache_nr_sets(this_leaf, np);
 200         cache_associativity(this_leaf);
 201 }
 202
 203 static int cache_setup_of_node(unsigned int cpu)
 204 {
 205         struct cacheinfo *this_leaf;
 206         unsigned int index = 0;
 207
 208         struct device_node *np __free(device_node) = of_cpu_device_node_get(cpu);
 209         if (!np) {
 210                 pr_err("Failed to find cpu%d device node\n", cpu);
 211                 return -ENOENT;
 212         }
 213
 214         if (!of_check_cache_nodes(np)) {
 215                 return -ENOENT;
 216         }
 217
 218         while (index < cache_leaves(cpu)) {
 219                 this_leaf = per_cpu_cacheinfo_idx(cpu, index);
 220                 if (this_leaf->level != 1) {
 221                         struct device_node *prev __free(device_node) = np;
 222                         np = of_find_next_cache_node(np);
 223                         if (!np)
 224                                 break;
 225                 }
 226                 cache_of_set_props(this_leaf, np);
 227                 this_leaf->fw_token = np;
 228                 index++;
 229         }
 230
 231         if (index != cache_leaves(cpu)) /* not all OF nodes populated */
 232                 return -ENOENT;
 233
 234         return 0;
 235 }
 236
 237 static bool of_check_cache_nodes(struct device_node *np)
 238 {
 239         if (of_property_present(np, "cache-size")   ||
 240             of_property_present(np, "i-cache-size") ||
 241             of_property_present(np, "d-cache-size") ||
 242             of_property_present(np, "cache-unified"))
 243                 return true;
 244
 245         struct device_node *next __free(device_node) = of_find_next_cache_node(np);
 246         if (next) {
 247                 return true;
 248         }
 249
 250         return false;
 251 }
 252
 253 static int of_count_cache_leaves(struct device_node *np)
 254 {
 255         unsigned int leaves = 0;
 256
 257         if (of_property_read_bool(np, "cache-size"))
 258                 ++leaves;
 259         if (of_property_read_bool(np, "i-cache-size"))
 260                 ++leaves;
 261         if (of_property_read_bool(np, "d-cache-size"))
 262                 ++leaves;
 263
 264         if (!leaves) {
 265                 /* The '[i-|d-|]cache-size' property is required, but
 266                  * if absent, fallback on the 'cache-unified' property.
 267                  */
 268                 if (of_property_read_bool(np, "cache-unified"))
 269                         return 1;
 270                 else
 271                         return 2;
 272         }
 273
 274         return leaves;
 275 }
 276
 277 int init_of_cache_level(unsigned int cpu)
 278 {
 279         struct cpu_cacheinfo *this_cpu_ci = get_cpu_cacheinfo(cpu);
 280         struct device_node *np __free(device_node) = of_cpu_device_node_get(cpu);
 281         unsigned int levels = 0, leaves, level;
 282
 283         if (!of_check_cache_nodes(np)) {
 284                 return -ENOENT;
 285         }
 286
 287         leaves = of_count_cache_leaves(np);
 288         if (leaves > 0)
 289                 levels = 1;
 290
 291         while (1) {
 292                 struct device_node *prev __free(device_node) = np;
 293                 np = of_find_next_cache_node(np);
 294                 if (!np)
 295                         break;
 296
 297                 if (!of_device_is_compatible(np, "cache"))
 298                         return -EINVAL;
 299                 if (of_property_read_u32(np, "cache-level", &level))
 300                         return -EINVAL;
 301                 if (level <= levels)
 302                         return -EINVAL;
 303
 304                 leaves += of_count_cache_leaves(np);
 305                 levels = level;
 306         }
 307
 308         this_cpu_ci->num_levels = levels;
 309         this_cpu_ci->num_leaves = leaves;
 310
 311         return 0;
 312 }
 313
 314 #else
 315 static inline int cache_setup_of_node(unsigned int cpu) { return 0; }
 316 int init_of_cache_level(unsigned int cpu) { return 0; }
 317 #endif
 318
 319 int __weak cache_setup_acpi(unsigned int cpu)
 320 {
 321         return -ENOTSUPP;
 322 }
 323
 324 unsigned int coherency_max_size;
 325
 326 static int cache_setup_properties(unsigned int cpu)
 327 {
 328         int ret = 0;
 329
 330         if (of_have_populated_dt())
 331                 ret = cache_setup_of_node(cpu);
 332         else if (!acpi_disabled)
 333                 ret = cache_setup_acpi(cpu);
 334
 335         // Assume there is no cache information available in DT/ACPI from now.
 336         if (ret && use_arch_cache_info())
 337                 use_arch_info = true;
 338
 339         return ret;
 340 }
 341
 342 static int cache_shared_cpu_map_setup(unsigned int cpu)
 343 {
 344         struct cpu_cacheinfo *this_cpu_ci = get_cpu_cacheinfo(cpu);
 345         struct cacheinfo *this_leaf, *sib_leaf;
 346         unsigned int index, sib_index;
 347         int ret = 0;
 348
 349         if (this_cpu_ci->cpu_map_populated)
 350                 return 0;
 351
 352         /*
 353          * skip setting up cache properties if LLC is valid, just need
 354          * to update the shared cpu_map if the cache attributes were
 355          * populated early before all the cpus are brought online
 356          */
 357         if (!last_level_cache_is_valid(cpu) && !use_arch_info) {
 358                 ret = cache_setup_properties(cpu);
 359                 if (ret)
 360                         return ret;
 361         }
 362
 363         for (index = 0; index < cache_leaves(cpu); index++) {
 364                 unsigned int i;
 365
 366                 this_leaf = per_cpu_cacheinfo_idx(cpu, index);
 367
 368                 cpumask_set_cpu(cpu, &this_leaf->shared_cpu_map);
 369                 for_each_online_cpu(i) {
 370                         struct cpu_cacheinfo *sib_cpu_ci = get_cpu_cacheinfo(i);
 371
 372                         if (i == cpu || !sib_cpu_ci->info_list)
 373                                 continue;/* skip if itself or no cacheinfo */
 374                         for (sib_index = 0; sib_index < cache_leaves(i); sib_index++) {
 375                                 sib_leaf = per_cpu_cacheinfo_idx(i, sib_index);
 376
 377                                 /*
 378                                  * Comparing cache IDs only makes sense if the leaves
 379                                  * belong to the same cache level of same type. Skip
 380                                  * the check if level and type do not match.
 381                                  */
 382                                 if (sib_leaf->level != this_leaf->level ||
 383                                     sib_leaf->type != this_leaf->type)
 384                                         continue;
 385
 386                                 if (cache_leaves_are_shared(this_leaf, sib_leaf)) {
 387                                         cpumask_set_cpu(cpu, &sib_leaf->shared_cpu_map);
 388                                         cpumask_set_cpu(i, &this_leaf->shared_cpu_map);
 389                                         break;
 390                                 }
 391                         }
 392                 }
 393                 /* record the maximum cache line size */
 394                 if (this_leaf->coherency_line_size > coherency_max_size)
 395                         coherency_max_size = this_leaf->coherency_line_size;
 396         }
 397
 398         /* shared_cpu_map is now populated for the cpu */
 399         this_cpu_ci->cpu_map_populated = true;
 400         return 0;
 401 }
 402
 403 static void cache_shared_cpu_map_remove(unsigned int cpu)
 404 {
 405         struct cpu_cacheinfo *this_cpu_ci = get_cpu_cacheinfo(cpu);
 406         struct cacheinfo *this_leaf, *sib_leaf;
 407         unsigned int sibling, index, sib_index;
 408
 409         for (index = 0; index < cache_leaves(cpu); index++) {
 410                 this_leaf = per_cpu_cacheinfo_idx(cpu, index);
 411                 for_each_cpu(sibling, &this_leaf->shared_cpu_map) {
 412                         struct cpu_cacheinfo *sib_cpu_ci =
 413                                                 get_cpu_cacheinfo(sibling);
 414
 415                         if (sibling == cpu || !sib_cpu_ci->info_list)
 416                                 continue;/* skip if itself or no cacheinfo */
 417
 418                         for (sib_index = 0; sib_index < cache_leaves(sibling); sib_index++) {
 419                                 sib_leaf = per_cpu_cacheinfo_idx(sibling, sib_index);
 420
 421                                 /*
 422                                  * Comparing cache IDs only makes sense if the leaves
 423                                  * belong to the same cache level of same type. Skip
 424                                  * the check if level and type do not match.
 425                                  */
 426                                 if (sib_leaf->level != this_leaf->level ||
 427                                     sib_leaf->type != this_leaf->type)
 428                                         continue;
 429
 430                                 if (cache_leaves_are_shared(this_leaf, sib_leaf)) {
 431                                         cpumask_clear_cpu(cpu, &sib_leaf->shared_cpu_map);
 432                                         cpumask_clear_cpu(sibling, &this_leaf->shared_cpu_map);
 433                                         break;
 434                                 }
 435                         }
 436                 }
 437         }
 438
 439         /* cpu is no longer populated in the shared map */
 440         this_cpu_ci->cpu_map_populated = false;
 441 }
 442
 443 static void free_cache_attributes(unsigned int cpu)
 444 {
 445         if (!per_cpu_cacheinfo(cpu))
 446                 return;
 447
 448         cache_shared_cpu_map_remove(cpu);
 449 }
 450
 451 int __weak early_cache_level(unsigned int cpu)
 452 {
 453         return -ENOENT;
 454 }
 455
 456 int __weak init_cache_level(unsigned int cpu)
 457 {
 458         return -ENOENT;
 459 }
 460
 461 int __weak populate_cache_leaves(unsigned int cpu)
 462 {
 463         return -ENOENT;
 464 }
 465
 466 static inline
 467 int allocate_cache_info(int cpu)
 468 {
 469         per_cpu_cacheinfo(cpu) = kcalloc(cache_leaves(cpu),
 470                                          sizeof(struct cacheinfo), GFP_ATOMIC);
 471         if (!per_cpu_cacheinfo(cpu)) {
 472                 cache_leaves(cpu) = 0;
 473                 return -ENOMEM;
 474         }
 475
 476         return 0;
 477 }
 478
 479 int fetch_cache_info(unsigned int cpu)
 480 {
 481         struct cpu_cacheinfo *this_cpu_ci = get_cpu_cacheinfo(cpu);
 482         unsigned int levels = 0, split_levels = 0;
 483         int ret;
 484
 485         if (acpi_disabled) {
 486                 ret = init_of_cache_level(cpu);
 487         } else {
 488                 ret = acpi_get_cache_info(cpu, &levels, &split_levels);
 489                 if (!ret) {
 490                         this_cpu_ci->num_levels = levels;
 491                         /*
 492                          * This assumes that:
 493                          * - there cannot be any split caches (data/instruction)
 494                          *   above a unified cache
 495                          * - data/instruction caches come by pair
 496                          */
 497                         this_cpu_ci->num_leaves = levels + split_levels;
 498                 }
 499         }
 500
 501         if (ret || !cache_leaves(cpu)) {
 502                 ret = early_cache_level(cpu);
 503                 if (ret)
 504                         return ret;
 505
 506                 if (!cache_leaves(cpu))
 507                         return -ENOENT;
 508
 509                 this_cpu_ci->early_ci_levels = true;
 510         }
 511
 512         return allocate_cache_info(cpu);
 513 }
 514
 515 static inline int init_level_allocate_ci(unsigned int cpu)
 516 {
 517         unsigned int early_leaves = cache_leaves(cpu);
 518
 519         /* Since early initialization/allocation of the cacheinfo is allowed
 520          * via fetch_cache_info() and this also gets called as CPU hotplug
 521          * callbacks via cacheinfo_cpu_online, the init/alloc can be skipped
 522          * as it will happen only once (the cacheinfo memory is never freed).
 523          * Just populate the cacheinfo. However, if the cacheinfo has been
 524          * allocated early through the arch-specific early_cache_level() call,
 525          * there is a chance the info is wrong (this can happen on arm64). In
 526          * that case, call init_cache_level() anyway to give the arch-specific
 527          * code a chance to make things right.
 528          */
 529         if (per_cpu_cacheinfo(cpu) && !ci_cacheinfo(cpu)->early_ci_levels)
 530                 return 0;
 531
 532         if (init_cache_level(cpu) || !cache_leaves(cpu))
 533                 return -ENOENT;
 534
 535         /*
 536          * Now that we have properly initialized the cache level info, make
 537          * sure we don't try to do that again the next time we are called
 538          * (e.g. as CPU hotplug callbacks).
 539          */
 540         ci_cacheinfo(cpu)->early_ci_levels = false;
 541
 542         if (cache_leaves(cpu) <= early_leaves)
 543                 return 0;
 544
 545         kfree(per_cpu_cacheinfo(cpu));
 546         return allocate_cache_info(cpu);
 547 }
 548
 549 int detect_cache_attributes(unsigned int cpu)
 550 {
 551         int ret;
 552
 553         ret = init_level_allocate_ci(cpu);
 554         if (ret)
 555                 return ret;
 556
 557         /*
 558          * If LLC is valid the cache leaves were already populated so just go to
 559          * update the cpu map.
 560          */
 561         if (!last_level_cache_is_valid(cpu)) {
 562                 /*
 563                  * populate_cache_leaves() may completely setup the cache leaves and
 564                  * shared_cpu_map or it may leave it partially setup.
 565                  */
 566                 ret = populate_cache_leaves(cpu);
 567                 if (ret)
 568                         goto free_ci;
 569         }
 570
 571         /*
 572          * For systems using DT for cache hierarchy, fw_token
 573          * and shared_cpu_map will be set up here only if they are
 574          * not populated already
 575          */
 576         ret = cache_shared_cpu_map_setup(cpu);
 577         if (ret) {
 578                 pr_warn("Unable to detect cache hierarchy for CPU %d\n", cpu);
 579                 goto free_ci;
 580         }
 581
 582         return 0;
 583
 584 free_ci:
 585         free_cache_attributes(cpu);
 586         return ret;
 587 }
 588
 589 /* pointer to cpuX/cache device */
 590 static DEFINE_PER_CPU(struct device *, ci_cache_dev);
 591 #define per_cpu_cache_dev(cpu)  (per_cpu(ci_cache_dev, cpu))
 592
 593 static cpumask_t cache_dev_map;
 594
 595 /* pointer to array of devices for cpuX/cache/indexY */
 596 static DEFINE_PER_CPU(struct device **, ci_index_dev);
 597 #define per_cpu_index_dev(cpu)  (per_cpu(ci_index_dev, cpu))
 598 #define per_cache_index_dev(cpu, idx)   ((per_cpu_index_dev(cpu))[idx])
 599
 600 #define show_one(file_name, object)                             \
 601 static ssize_t file_name##_show(struct device *dev,             \
 602                 struct device_attribute *attr, char *buf)       \
 603 {                                                               \
 604         struct cacheinfo *this_leaf = dev_get_drvdata(dev);     \
 605         return sysfs_emit(buf, "%u\n", this_leaf->object);      \
 606 }
 607
 608 show_one(id, id);
 609 show_one(level, level);
 610 show_one(coherency_line_size, coherency_line_size);
 611 show_one(number_of_sets, number_of_sets);
 612 show_one(physical_line_partition, physical_line_partition);
 613 show_one(ways_of_associativity, ways_of_associativity);
 614
 615 static ssize_t size_show(struct device *dev,
 616                          struct device_attribute *attr, char *buf)
 617 {
 618         struct cacheinfo *this_leaf = dev_get_drvdata(dev);
 619
 620         return sysfs_emit(buf, "%uK\n", this_leaf->size >> 10);
 621 }
 622
 623 static ssize_t shared_cpu_map_show(struct device *dev,
 624                                    struct device_attribute *attr, char *buf)
 625 {
 626         struct cacheinfo *this_leaf = dev_get_drvdata(dev);
 627         const struct cpumask *mask = &this_leaf->shared_cpu_map;
 628
 629         return sysfs_emit(buf, "%*pb\n", nr_cpu_ids, mask);
 630 }
 631
 632 static ssize_t shared_cpu_list_show(struct device *dev,
 633                                     struct device_attribute *attr, char *buf)
 634 {
 635         struct cacheinfo *this_leaf = dev_get_drvdata(dev);
 636         const struct cpumask *mask = &this_leaf->shared_cpu_map;
 637
 638         return sysfs_emit(buf, "%*pbl\n", nr_cpu_ids, mask);
 639 }
 640
 641 static ssize_t type_show(struct device *dev,
 642                          struct device_attribute *attr, char *buf)
 643 {
 644         struct cacheinfo *this_leaf = dev_get_drvdata(dev);
 645         const char *output;
 646
 647         switch (this_leaf->type) {
 648         case CACHE_TYPE_DATA:
 649                 output = "Data";
 650                 break;
 651         case CACHE_TYPE_INST:
 652                 output = "Instruction";
 653                 break;
 654         case CACHE_TYPE_UNIFIED:
 655                 output = "Unified";
 656                 break;
 657         default:
 658                 return -EINVAL;
 659         }
 660
 661         return sysfs_emit(buf, "%s\n", output);
 662 }
 663
 664 static ssize_t allocation_policy_show(struct device *dev,
 665                                       struct device_attribute *attr, char *buf)
 666 {
 667         struct cacheinfo *this_leaf = dev_get_drvdata(dev);
 668         unsigned int ci_attr = this_leaf->attributes;
 669         const char *output;
 670
 671         if ((ci_attr & CACHE_READ_ALLOCATE) && (ci_attr & CACHE_WRITE_ALLOCATE))
 672                 output = "ReadWriteAllocate";
 673         else if (ci_attr & CACHE_READ_ALLOCATE)
 674                 output = "ReadAllocate";
 675         else if (ci_attr & CACHE_WRITE_ALLOCATE)
 676                 output = "WriteAllocate";
 677         else
 678                 return 0;
 679
 680         return sysfs_emit(buf, "%s\n", output);
 681 }
 682
 683 static ssize_t write_policy_show(struct device *dev,
 684                                  struct device_attribute *attr, char *buf)
 685 {
 686         struct cacheinfo *this_leaf = dev_get_drvdata(dev);
 687         unsigned int ci_attr = this_leaf->attributes;
 688         int n = 0;
 689
 690         if (ci_attr & CACHE_WRITE_THROUGH)
 691                 n = sysfs_emit(buf, "WriteThrough\n");
 692         else if (ci_attr & CACHE_WRITE_BACK)
 693                 n = sysfs_emit(buf, "WriteBack\n");
 694         return n;
 695 }
 696
 697 static DEVICE_ATTR_RO(id);
 698 static DEVICE_ATTR_RO(level);
 699 static DEVICE_ATTR_RO(type);
 700 static DEVICE_ATTR_RO(coherency_line_size);
 701 static DEVICE_ATTR_RO(ways_of_associativity);
 702 static DEVICE_ATTR_RO(number_of_sets);
 703 static DEVICE_ATTR_RO(size);
 704 static DEVICE_ATTR_RO(allocation_policy);
 705 static DEVICE_ATTR_RO(write_policy);
 706 static DEVICE_ATTR_RO(shared_cpu_map);
 707 static DEVICE_ATTR_RO(shared_cpu_list);
 708 static DEVICE_ATTR_RO(physical_line_partition);
 709
 710 static struct attribute *cache_default_attrs[] = {
 711         &dev_attr_id.attr,
 712         &dev_attr_type.attr,
 713         &dev_attr_level.attr,
 714         &dev_attr_shared_cpu_map.attr,
 715         &dev_attr_shared_cpu_list.attr,
 716         &dev_attr_coherency_line_size.attr,
 717         &dev_attr_ways_of_associativity.attr,
 718         &dev_attr_number_of_sets.attr,
 719         &dev_attr_size.attr,
 720         &dev_attr_allocation_policy.attr,
 721         &dev_attr_write_policy.attr,
 722         &dev_attr_physical_line_partition.attr,
 723         NULL
 724 };
 725
 726 static umode_t
 727 cache_default_attrs_is_visible(struct kobject *kobj,
 728                                struct attribute *attr, int unused)
 729 {
 730         struct device *dev = kobj_to_dev(kobj);
 731         struct cacheinfo *this_leaf = dev_get_drvdata(dev);
 732         const struct cpumask *mask = &this_leaf->shared_cpu_map;
 733         umode_t mode = attr->mode;
 734
 735         if ((attr == &dev_attr_id.attr) && (this_leaf->attributes & CACHE_ID))
 736                 return mode;
 737         if ((attr == &dev_attr_type.attr) && this_leaf->type)
 738                 return mode;
 739         if ((attr == &dev_attr_level.attr) && this_leaf->level)
 740                 return mode;
 741         if ((attr == &dev_attr_shared_cpu_map.attr) && !cpumask_empty(mask))
 742                 return mode;
 743         if ((attr == &dev_attr_shared_cpu_list.attr) && !cpumask_empty(mask))
 744                 return mode;
 745         if ((attr == &dev_attr_coherency_line_size.attr) &&
 746             this_leaf->coherency_line_size)
 747                 return mode;
 748         if ((attr == &dev_attr_ways_of_associativity.attr) &&
 749             this_leaf->size) /* allow 0 = full associativity */
 750                 return mode;
 751         if ((attr == &dev_attr_number_of_sets.attr) &&
 752             this_leaf->number_of_sets)
 753                 return mode;
 754         if ((attr == &dev_attr_size.attr) && this_leaf->size)
 755                 return mode;
 756         if ((attr == &dev_attr_write_policy.attr) &&
 757             (this_leaf->attributes & CACHE_WRITE_POLICY_MASK))
 758                 return mode;
 759         if ((attr == &dev_attr_allocation_policy.attr) &&
 760             (this_leaf->attributes & CACHE_ALLOCATE_POLICY_MASK))
 761                 return mode;
 762         if ((attr == &dev_attr_physical_line_partition.attr) &&
 763             this_leaf->physical_line_partition)
 764                 return mode;
 765
 766         return 0;
 767 }
 768
 769 static const struct attribute_group cache_default_group = {
 770         .attrs = cache_default_attrs,
 771         .is_visible = cache_default_attrs_is_visible,
 772 };
 773
 774 static const struct attribute_group *cache_default_groups[] = {
 775         &cache_default_group,
 776         NULL,
 777 };
 778
 779 static const struct attribute_group *cache_private_groups[] = {
 780         &cache_default_group,
 781         NULL, /* Place holder for private group */
 782         NULL,
 783 };
 784
 785 const struct attribute_group *
 786 __weak cache_get_priv_group(struct cacheinfo *this_leaf)
 787 {
 788         return NULL;
 789 }
 790
 791 static const struct attribute_group **
 792 cache_get_attribute_groups(struct cacheinfo *this_leaf)
 793 {
 794         const struct attribute_group *priv_group =
 795                         cache_get_priv_group(this_leaf);
 796
 797         if (!priv_group)
 798                 return cache_default_groups;
 799
 800         if (!cache_private_groups[1])
 801                 cache_private_groups[1] = priv_group;
 802
 803         return cache_private_groups;
 804 }
 805
 806 /* Add/Remove cache interface for CPU device */
 807 static void cpu_cache_sysfs_exit(unsigned int cpu)
 808 {
 809         int i;
 810         struct device *ci_dev;
 811
 812         if (per_cpu_index_dev(cpu)) {
 813                 for (i = 0; i < cache_leaves(cpu); i++) {
 814                         ci_dev = per_cache_index_dev(cpu, i);
 815                         if (!ci_dev)
 816                                 continue;
 817                         device_unregister(ci_dev);
 818                 }
 819                 kfree(per_cpu_index_dev(cpu));
 820                 per_cpu_index_dev(cpu) = NULL;
 821         }
 822         device_unregister(per_cpu_cache_dev(cpu));
 823         per_cpu_cache_dev(cpu) = NULL;
 824 }
 825
 826 static int cpu_cache_sysfs_init(unsigned int cpu)
 827 {
 828         struct device *dev = get_cpu_device(cpu);
 829
 830         if (per_cpu_cacheinfo(cpu) == NULL)
 831                 return -ENOENT;
 832
 833         per_cpu_cache_dev(cpu) = cpu_device_create(dev, NULL, NULL, "cache");
 834         if (IS_ERR(per_cpu_cache_dev(cpu)))
 835                 return PTR_ERR(per_cpu_cache_dev(cpu));
 836
 837         /* Allocate all required memory */
 838         per_cpu_index_dev(cpu) = kcalloc(cache_leaves(cpu),
 839                                          sizeof(struct device *), GFP_KERNEL);
 840         if (unlikely(per_cpu_index_dev(cpu) == NULL))
 841                 goto err_out;
 842
 843         return 0;
 844
 845 err_out:
 846         cpu_cache_sysfs_exit(cpu);
 847         return -ENOMEM;
 848 }
 849
 850 static int cache_add_dev(unsigned int cpu)
 851 {
 852         unsigned int i;
 853         int rc;
 854         struct device *ci_dev, *parent;
 855         struct cacheinfo *this_leaf;
 856         const struct attribute_group **cache_groups;
 857
 858         rc = cpu_cache_sysfs_init(cpu);
 859         if (unlikely(rc < 0))
 860                 return rc;
 861
 862         parent = per_cpu_cache_dev(cpu);
 863         for (i = 0; i < cache_leaves(cpu); i++) {
 864                 this_leaf = per_cpu_cacheinfo_idx(cpu, i);
 865                 if (this_leaf->disable_sysfs)
 866                         continue;
 867                 if (this_leaf->type == CACHE_TYPE_NOCACHE)
 868                         break;
 869                 cache_groups = cache_get_attribute_groups(this_leaf);
 870                 ci_dev = cpu_device_create(parent, this_leaf, cache_groups,
 871                                            "index%1u", i);
 872                 if (IS_ERR(ci_dev)) {
 873                         rc = PTR_ERR(ci_dev);
 874                         goto err;
 875                 }
 876                 per_cache_index_dev(cpu, i) = ci_dev;
 877         }
 878         cpumask_set_cpu(cpu, &cache_dev_map);
 879
 880         return 0;
 881 err:
 882         cpu_cache_sysfs_exit(cpu);
 883         return rc;
 884 }
 885
 886 static unsigned int cpu_map_shared_cache(bool online, unsigned int cpu,
 887                                          cpumask_t **map)
 888 {
 889         struct cacheinfo *llc, *sib_llc;
 890         unsigned int sibling;
 891
 892         if (!last_level_cache_is_valid(cpu))
 893                 return 0;
 894
 895         llc = per_cpu_cacheinfo_idx(cpu, cache_leaves(cpu) - 1);
 896
 897         if (llc->type != CACHE_TYPE_DATA && llc->type != CACHE_TYPE_UNIFIED)
 898                 return 0;
 899
 900         if (online) {
 901                 *map = &llc->shared_cpu_map;
 902                 return cpumask_weight(*map);
 903         }
 904
 905         /* shared_cpu_map of offlined CPU will be cleared, so use sibling map */
 906         for_each_cpu(sibling, &llc->shared_cpu_map) {
 907                 if (sibling == cpu || !last_level_cache_is_valid(sibling))
 908                         continue;
 909                 sib_llc = per_cpu_cacheinfo_idx(sibling, cache_leaves(sibling) - 1);
 910                 *map = &sib_llc->shared_cpu_map;
 911                 return cpumask_weight(*map);
 912         }
 913
 914         return 0;
 915 }
 916
 917 /*
 918  * Calculate the size of the per-CPU data cache slice.  This can be
 919  * used to estimate the size of the data cache slice that can be used
 920  * by one CPU under ideal circumstances.  UNIFIED caches are counted
 921  * in addition to DATA caches.  So, please consider code cache usage
 922  * when use the result.
 923  *
 924  * Because the cache inclusive/non-inclusive information isn't
 925  * available, we just use the size of the per-CPU slice of LLC to make
 926  * the result more predictable across architectures.
 927  */
 928 static void update_per_cpu_data_slice_size_cpu(unsigned int cpu)
 929 {
 930         struct cpu_cacheinfo *ci;
 931         struct cacheinfo *llc;
 932         unsigned int nr_shared;
 933
 934         if (!last_level_cache_is_valid(cpu))
 935                 return;
 936
 937         ci = ci_cacheinfo(cpu);
 938         llc = per_cpu_cacheinfo_idx(cpu, cache_leaves(cpu) - 1);
 939
 940         if (llc->type != CACHE_TYPE_DATA && llc->type != CACHE_TYPE_UNIFIED)
 941                 return;
 942
 943         nr_shared = cpumask_weight(&llc->shared_cpu_map);
 944         if (nr_shared)
 945                 ci->per_cpu_data_slice_size = llc->size / nr_shared;
 946 }
 947
 948 static void update_per_cpu_data_slice_size(bool cpu_online, unsigned int cpu,
 949                                            cpumask_t *cpu_map)
 950 {
 951         unsigned int icpu;
 952
 953         for_each_cpu(icpu, cpu_map) {
 954                 if (!cpu_online && icpu == cpu)
 955                         continue;
 956                 update_per_cpu_data_slice_size_cpu(icpu);
 957                 setup_pcp_cacheinfo(icpu);
 958         }
 959 }
 960
 961 static int cacheinfo_cpu_online(unsigned int cpu)
 962 {
 963         int rc = detect_cache_attributes(cpu);
 964         cpumask_t *cpu_map;
 965
 966         if (rc)
 967                 return rc;
 968         rc = cache_add_dev(cpu);
 969         if (rc)
 970                 goto err;
 971         if (cpu_map_shared_cache(true, cpu, &cpu_map))
 972                 update_per_cpu_data_slice_size(true, cpu, cpu_map);
 973         return 0;
 974 err:
 975         free_cache_attributes(cpu);
 976         return rc;
 977 }
 978
 979 static int cacheinfo_cpu_pre_down(unsigned int cpu)
 980 {
 981         cpumask_t *cpu_map;
 982         unsigned int nr_shared;
 983
 984         nr_shared = cpu_map_shared_cache(false, cpu, &cpu_map);
 985         if (cpumask_test_and_clear_cpu(cpu, &cache_dev_map))
 986                 cpu_cache_sysfs_exit(cpu);
 987
 988         free_cache_attributes(cpu);
 989         if (nr_shared > 1)
 990                 update_per_cpu_data_slice_size(false, cpu, cpu_map);
 991         return 0;
 992 }
 993
 994 static int __init cacheinfo_sysfs_init(void)
 995 {
 996         return cpuhp_setup_state(CPUHP_AP_BASE_CACHEINFO_ONLINE,
 997                                  "base/cacheinfo:online",
 998                                  cacheinfo_cpu_online, cacheinfo_cpu_pre_down);
 999 }
1000 device_initcall(cacheinfo_sysfs_init);