drivers/base/cpu.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * CPU subsystem support
   4  */
   5
   6 #include <linux/kernel.h>
   7 #include <linux/module.h>
   8 #include <linux/init.h>
   9 #include <linux/sched.h>
  10 #include <linux/cpu.h>
  11 #include <linux/topology.h>
  12 #include <linux/device.h>
  13 #include <linux/node.h>
  14 #include <linux/gfp.h>
  15 #include <linux/slab.h>
  16 #include <linux/percpu.h>
  17 #include <linux/acpi.h>
  18 #include <linux/of.h>
  19 #include <linux/cpufeature.h>
  20 #include <linux/tick.h>
  21 #include <linux/pm_qos.h>
  22 #include <linux/sched/isolation.h>
  23
  24 #include "base.h"
  25
  26 static DEFINE_PER_CPU(struct device *, cpu_sys_devices);
  27
  28 static int cpu_subsys_match(struct device *dev, struct device_driver *drv)
  29 {
  30         /* ACPI style match is the only one that may succeed. */
  31         if (acpi_driver_match_device(dev, drv))
  32                 return 1;
  33
  34         return 0;
  35 }
  36
  37 #ifdef CONFIG_HOTPLUG_CPU
  38 static void change_cpu_under_node(struct cpu *cpu,
  39                         unsigned int from_nid, unsigned int to_nid)
  40 {
  41         int cpuid = cpu->dev.id;
  42         unregister_cpu_under_node(cpuid, from_nid);
  43         register_cpu_under_node(cpuid, to_nid);
  44         cpu->node_id = to_nid;
  45 }
  46
  47 static int cpu_subsys_online(struct device *dev)
  48 {
  49         struct cpu *cpu = container_of(dev, struct cpu, dev);
  50         int cpuid = dev->id;
  51         int from_nid, to_nid;
  52         int ret;
  53
  54         from_nid = cpu_to_node(cpuid);
  55         if (from_nid == NUMA_NO_NODE)
  56                 return -ENODEV;
  57
  58         ret = cpu_device_up(dev);
  59         /*
  60          * When hot adding memory to memoryless node and enabling a cpu
  61          * on the node, node number of the cpu may internally change.
  62          */
  63         to_nid = cpu_to_node(cpuid);
  64         if (from_nid != to_nid)
  65                 change_cpu_under_node(cpu, from_nid, to_nid);
  66
  67         return ret;
  68 }
  69
  70 static int cpu_subsys_offline(struct device *dev)
  71 {
  72         return cpu_device_down(dev);
  73 }
  74
  75 void unregister_cpu(struct cpu *cpu)
  76 {
  77         int logical_cpu = cpu->dev.id;
  78
  79         unregister_cpu_under_node(logical_cpu, cpu_to_node(logical_cpu));
  80
  81         device_unregister(&cpu->dev);
  82         per_cpu(cpu_sys_devices, logical_cpu) = NULL;
  83         return;
  84 }
  85
  86 #ifdef CONFIG_ARCH_CPU_PROBE_RELEASE
  87 static ssize_t cpu_probe_store(struct device *dev,
  88                                struct device_attribute *attr,
  89                                const char *buf,
  90                                size_t count)
  91 {
  92         ssize_t cnt;
  93         int ret;
  94
  95         ret = lock_device_hotplug_sysfs();
  96         if (ret)
  97                 return ret;
  98
  99         cnt = arch_cpu_probe(buf, count);
 100
 101         unlock_device_hotplug();
 102         return cnt;
 103 }
 104
 105 static ssize_t cpu_release_store(struct device *dev,
 106                                  struct device_attribute *attr,
 107                                  const char *buf,
 108                                  size_t count)
 109 {
 110         ssize_t cnt;
 111         int ret;
 112
 113         ret = lock_device_hotplug_sysfs();
 114         if (ret)
 115                 return ret;
 116
 117         cnt = arch_cpu_release(buf, count);
 118
 119         unlock_device_hotplug();
 120         return cnt;
 121 }
 122
 123 static DEVICE_ATTR(probe, S_IWUSR, NULL, cpu_probe_store);
 124 static DEVICE_ATTR(release, S_IWUSR, NULL, cpu_release_store);
 125 #endif /* CONFIG_ARCH_CPU_PROBE_RELEASE */
 126 #endif /* CONFIG_HOTPLUG_CPU */
 127
 128 struct bus_type cpu_subsys = {
 129         .name = "cpu",
 130         .dev_name = "cpu",
 131         .match = cpu_subsys_match,
 132 #ifdef CONFIG_HOTPLUG_CPU
 133         .online = cpu_subsys_online,
 134         .offline = cpu_subsys_offline,
 135 #endif
 136 };
 137 EXPORT_SYMBOL_GPL(cpu_subsys);
 138
 139 #ifdef CONFIG_KEXEC
 140 #include <linux/kexec.h>
 141
 142 static ssize_t crash_notes_show(struct device *dev,
 143                                 struct device_attribute *attr,
 144                                 char *buf)
 145 {
 146         struct cpu *cpu = container_of(dev, struct cpu, dev);
 147         unsigned long long addr;
 148         int cpunum;
 149
 150         cpunum = cpu->dev.id;
 151
 152         /*
 153          * Might be reading other cpu's data based on which cpu read thread
 154          * has been scheduled. But cpu data (memory) is allocated once during
 155          * boot up and this data does not change there after. Hence this
 156          * operation should be safe. No locking required.
 157          */
 158         addr = per_cpu_ptr_to_phys(per_cpu_ptr(crash_notes, cpunum));
 159
 160         return sysfs_emit(buf, "%llx\n", addr);
 161 }
 162 static DEVICE_ATTR_ADMIN_RO(crash_notes);
 163
 164 static ssize_t crash_notes_size_show(struct device *dev,
 165                                      struct device_attribute *attr,
 166                                      char *buf)
 167 {
 168         return sysfs_emit(buf, "%zu\n", sizeof(note_buf_t));
 169 }
 170 static DEVICE_ATTR_ADMIN_RO(crash_notes_size);
 171
 172 static struct attribute *crash_note_cpu_attrs[] = {
 173         &dev_attr_crash_notes.attr,
 174         &dev_attr_crash_notes_size.attr,
 175         NULL
 176 };
 177
 178 static const struct attribute_group crash_note_cpu_attr_group = {
 179         .attrs = crash_note_cpu_attrs,
 180 };
 181 #endif
 182
 183 static const struct attribute_group *common_cpu_attr_groups[] = {
 184 #ifdef CONFIG_KEXEC
 185         &crash_note_cpu_attr_group,
 186 #endif
 187         NULL
 188 };
 189
 190 static const struct attribute_group *hotplugable_cpu_attr_groups[] = {
 191 #ifdef CONFIG_KEXEC
 192         &crash_note_cpu_attr_group,
 193 #endif
 194         NULL
 195 };
 196
 197 /*
 198  * Print cpu online, possible, present, and system maps
 199  */
 200
 201 struct cpu_attr {
 202         struct device_attribute attr;
 203         const struct cpumask *const map;
 204 };
 205
 206 static ssize_t show_cpus_attr(struct device *dev,
 207                               struct device_attribute *attr,
 208                               char *buf)
 209 {
 210         struct cpu_attr *ca = container_of(attr, struct cpu_attr, attr);
 211
 212         return cpumap_print_to_pagebuf(true, buf, ca->map);
 213 }
 214
 215 #define _CPU_ATTR(name, map) \
 216         { __ATTR(name, 0444, show_cpus_attr, NULL), map }
 217
 218 /* Keep in sync with cpu_subsys_attrs */
 219 static struct cpu_attr cpu_attrs[] = {
 220         _CPU_ATTR(online, &__cpu_online_mask),
 221         _CPU_ATTR(possible, &__cpu_possible_mask),
 222         _CPU_ATTR(present, &__cpu_present_mask),
 223 };
 224
 225 /*
 226  * Print values for NR_CPUS and offlined cpus
 227  */
 228 static ssize_t print_cpus_kernel_max(struct device *dev,
 229                                      struct device_attribute *attr, char *buf)
 230 {
 231         return sysfs_emit(buf, "%d\n", NR_CPUS - 1);
 232 }
 233 static DEVICE_ATTR(kernel_max, 0444, print_cpus_kernel_max, NULL);
 234
 235 /* arch-optional setting to enable display of offline cpus >= nr_cpu_ids */
 236 unsigned int total_cpus;
 237
 238 static ssize_t print_cpus_offline(struct device *dev,
 239                                   struct device_attribute *attr, char *buf)
 240 {
 241         int len = 0;
 242         cpumask_var_t offline;
 243
 244         /* display offline cpus < nr_cpu_ids */
 245         if (!alloc_cpumask_var(&offline, GFP_KERNEL))
 246                 return -ENOMEM;
 247         cpumask_andnot(offline, cpu_possible_mask, cpu_online_mask);
 248         len += sysfs_emit_at(buf, len, "%*pbl", cpumask_pr_args(offline));
 249         free_cpumask_var(offline);
 250
 251         /* display offline cpus >= nr_cpu_ids */
 252         if (total_cpus && nr_cpu_ids < total_cpus) {
 253                 len += sysfs_emit_at(buf, len, ",");
 254
 255                 if (nr_cpu_ids == total_cpus-1)
 256                         len += sysfs_emit_at(buf, len, "%u", nr_cpu_ids);
 257                 else
 258                         len += sysfs_emit_at(buf, len, "%u-%d",
 259                                              nr_cpu_ids, total_cpus - 1);
 260         }
 261
 262         len += sysfs_emit_at(buf, len, "\n");
 263
 264         return len;
 265 }
 266 static DEVICE_ATTR(offline, 0444, print_cpus_offline, NULL);
 267
 268 static ssize_t print_cpus_isolated(struct device *dev,
 269                                   struct device_attribute *attr, char *buf)
 270 {
 271         int len;
 272         cpumask_var_t isolated;
 273
 274         if (!alloc_cpumask_var(&isolated, GFP_KERNEL))
 275                 return -ENOMEM;
 276
 277         cpumask_andnot(isolated, cpu_possible_mask,
 278                        housekeeping_cpumask(HK_FLAG_DOMAIN));
 279         len = sysfs_emit(buf, "%*pbl\n", cpumask_pr_args(isolated));
 280
 281         free_cpumask_var(isolated);
 282
 283         return len;
 284 }
 285 static DEVICE_ATTR(isolated, 0444, print_cpus_isolated, NULL);
 286
 287 #ifdef CONFIG_NO_HZ_FULL
 288 static ssize_t print_cpus_nohz_full(struct device *dev,
 289                                     struct device_attribute *attr, char *buf)
 290 {
 291         return sysfs_emit(buf, "%*pbl\n", cpumask_pr_args(tick_nohz_full_mask));
 292 }
 293 static DEVICE_ATTR(nohz_full, 0444, print_cpus_nohz_full, NULL);
 294 #endif
 295
 296 static void cpu_device_release(struct device *dev)
 297 {
 298         /*
 299          * This is an empty function to prevent the driver core from spitting a
 300          * warning at us.  Yes, I know this is directly opposite of what the
 301          * documentation for the driver core and kobjects say, and the author
 302          * of this code has already been publically ridiculed for doing
 303          * something as foolish as this.  However, at this point in time, it is
 304          * the only way to handle the issue of statically allocated cpu
 305          * devices.  The different architectures will have their cpu device
 306          * code reworked to properly handle this in the near future, so this
 307          * function will then be changed to correctly free up the memory held
 308          * by the cpu device.
 309          *
 310          * Never copy this way of doing things, or you too will be made fun of
 311          * on the linux-kernel list, you have been warned.
 312          */
 313 }
 314
 315 #ifdef CONFIG_GENERIC_CPU_AUTOPROBE
 316 static ssize_t print_cpu_modalias(struct device *dev,
 317                                   struct device_attribute *attr,
 318                                   char *buf)
 319 {
 320         int len = 0;
 321         u32 i;
 322
 323         len += sysfs_emit_at(buf, len,
 324                              "cpu:type:" CPU_FEATURE_TYPEFMT ":feature:",
 325                              CPU_FEATURE_TYPEVAL);
 326
 327         for (i = 0; i < MAX_CPU_FEATURES; i++)
 328                 if (cpu_have_feature(i)) {
 329                         if (len + sizeof(",XXXX\n") >= PAGE_SIZE) {
 330                                 WARN(1, "CPU features overflow page\n");
 331                                 break;
 332                         }
 333                         len += sysfs_emit_at(buf, len, ",%04X", i);
 334                 }
 335         len += sysfs_emit_at(buf, len, "\n");
 336         return len;
 337 }
 338
 339 static int cpu_uevent(struct device *dev, struct kobj_uevent_env *env)
 340 {
 341         char *buf = kzalloc(PAGE_SIZE, GFP_KERNEL);
 342         if (buf) {
 343                 print_cpu_modalias(NULL, NULL, buf);
 344                 add_uevent_var(env, "MODALIAS=%s", buf);
 345                 kfree(buf);
 346         }
 347         return 0;
 348 }
 349 #endif
 350
 351 /*
 352  * register_cpu - Setup a sysfs device for a CPU.
 353  * @cpu - cpu->hotpluggable field set to 1 will generate a control file in
 354  *        sysfs for this CPU.
 355  * @num - CPU number to use when creating the device.
 356  *
 357  * Initialize and register the CPU device.
 358  */
 359 int register_cpu(struct cpu *cpu, int num)
 360 {
 361         int error;
 362
 363         cpu->node_id = cpu_to_node(num);
 364         memset(&cpu->dev, 0x00, sizeof(struct device));
 365         cpu->dev.id = num;
 366         cpu->dev.bus = &cpu_subsys;
 367         cpu->dev.release = cpu_device_release;
 368         cpu->dev.offline_disabled = !cpu->hotpluggable;
 369         cpu->dev.offline = !cpu_online(num);
 370         cpu->dev.of_node = of_get_cpu_node(num, NULL);
 371 #ifdef CONFIG_GENERIC_CPU_AUTOPROBE
 372         cpu->dev.bus->uevent = cpu_uevent;
 373 #endif
 374         cpu->dev.groups = common_cpu_attr_groups;
 375         if (cpu->hotpluggable)
 376                 cpu->dev.groups = hotplugable_cpu_attr_groups;
 377         error = device_register(&cpu->dev);
 378         if (error) {
 379                 put_device(&cpu->dev);
 380                 return error;
 381         }
 382
 383         per_cpu(cpu_sys_devices, num) = &cpu->dev;
 384         register_cpu_under_node(num, cpu_to_node(num));
 385         dev_pm_qos_expose_latency_limit(&cpu->dev,
 386                                         PM_QOS_RESUME_LATENCY_NO_CONSTRAINT);
 387
 388         return 0;
 389 }
 390
 391 struct device *get_cpu_device(unsigned int cpu)
 392 {
 393         if (cpu < nr_cpu_ids && cpu_possible(cpu))
 394                 return per_cpu(cpu_sys_devices, cpu);
 395         else
 396                 return NULL;
 397 }
 398 EXPORT_SYMBOL_GPL(get_cpu_device);
 399
 400 static void device_create_release(struct device *dev)
 401 {
 402         kfree(dev);
 403 }
 404
 405 __printf(4, 0)
 406 static struct device *
 407 __cpu_device_create(struct device *parent, void *drvdata,
 408                     const struct attribute_group **groups,
 409                     const char *fmt, va_list args)
 410 {
 411         struct device *dev = NULL;
 412         int retval = -ENOMEM;
 413
 414         dev = kzalloc(sizeof(*dev), GFP_KERNEL);
 415         if (!dev)
 416                 goto error;
 417
 418         device_initialize(dev);
 419         dev->parent = parent;
 420         dev->groups = groups;
 421         dev->release = device_create_release;
 422         device_set_pm_not_required(dev);
 423         dev_set_drvdata(dev, drvdata);
 424
 425         retval = kobject_set_name_vargs(&dev->kobj, fmt, args);
 426         if (retval)
 427                 goto error;
 428
 429         retval = device_add(dev);
 430         if (retval)
 431                 goto error;
 432
 433         return dev;
 434
 435 error:
 436         put_device(dev);
 437         return ERR_PTR(retval);
 438 }
 439
 440 struct device *cpu_device_create(struct device *parent, void *drvdata,
 441                                  const struct attribute_group **groups,
 442                                  const char *fmt, ...)
 443 {
 444         va_list vargs;
 445         struct device *dev;
 446
 447         va_start(vargs, fmt);
 448         dev = __cpu_device_create(parent, drvdata, groups, fmt, vargs);
 449         va_end(vargs);
 450         return dev;
 451 }
 452 EXPORT_SYMBOL_GPL(cpu_device_create);
 453
 454 #ifdef CONFIG_GENERIC_CPU_AUTOPROBE
 455 static DEVICE_ATTR(modalias, 0444, print_cpu_modalias, NULL);
 456 #endif
 457
 458 static struct attribute *cpu_root_attrs[] = {
 459 #ifdef CONFIG_ARCH_CPU_PROBE_RELEASE
 460         &dev_attr_probe.attr,
 461         &dev_attr_release.attr,
 462 #endif
 463         &cpu_attrs[0].attr.attr,
 464         &cpu_attrs[1].attr.attr,
 465         &cpu_attrs[2].attr.attr,
 466         &dev_attr_kernel_max.attr,
 467         &dev_attr_offline.attr,
 468         &dev_attr_isolated.attr,
 469 #ifdef CONFIG_NO_HZ_FULL
 470         &dev_attr_nohz_full.attr,
 471 #endif
 472 #ifdef CONFIG_GENERIC_CPU_AUTOPROBE
 473         &dev_attr_modalias.attr,
 474 #endif
 475         NULL
 476 };
 477
 478 static const struct attribute_group cpu_root_attr_group = {
 479         .attrs = cpu_root_attrs,
 480 };
 481
 482 static const struct attribute_group *cpu_root_attr_groups[] = {
 483         &cpu_root_attr_group,
 484         NULL,
 485 };
 486
 487 bool cpu_is_hotpluggable(unsigned int cpu)
 488 {
 489         struct device *dev = get_cpu_device(cpu);
 490         return dev && container_of(dev, struct cpu, dev)->hotpluggable;
 491 }
 492 EXPORT_SYMBOL_GPL(cpu_is_hotpluggable);
 493
 494 #ifdef CONFIG_GENERIC_CPU_DEVICES
 495 static DEFINE_PER_CPU(struct cpu, cpu_devices);
 496 #endif
 497
 498 static void __init cpu_dev_register_generic(void)
 499 {
 500 #ifdef CONFIG_GENERIC_CPU_DEVICES
 501         int i;
 502
 503         for_each_possible_cpu(i) {
 504                 if (register_cpu(&per_cpu(cpu_devices, i), i))
 505                         panic("Failed to register CPU device");
 506         }
 507 #endif
 508 }
 509
 510 #ifdef CONFIG_GENERIC_CPU_VULNERABILITIES
 511
 512 ssize_t __weak cpu_show_meltdown(struct device *dev,
 513                                  struct device_attribute *attr, char *buf)
 514 {
 515         return sysfs_emit(buf, "Not affected\n");
 516 }
 517
 518 ssize_t __weak cpu_show_spectre_v1(struct device *dev,
 519                                    struct device_attribute *attr, char *buf)
 520 {
 521         return sysfs_emit(buf, "Not affected\n");
 522 }
 523
 524 ssize_t __weak cpu_show_spectre_v2(struct device *dev,
 525                                    struct device_attribute *attr, char *buf)
 526 {
 527         return sysfs_emit(buf, "Not affected\n");
 528 }
 529
 530 ssize_t __weak cpu_show_spec_store_bypass(struct device *dev,
 531                                           struct device_attribute *attr, char *buf)
 532 {
 533         return sysfs_emit(buf, "Not affected\n");
 534 }
 535
 536 ssize_t __weak cpu_show_l1tf(struct device *dev,
 537                              struct device_attribute *attr, char *buf)
 538 {
 539         return sysfs_emit(buf, "Not affected\n");
 540 }
 541
 542 ssize_t __weak cpu_show_mds(struct device *dev,
 543                             struct device_attribute *attr, char *buf)
 544 {
 545         return sysfs_emit(buf, "Not affected\n");
 546 }
 547
 548 ssize_t __weak cpu_show_tsx_async_abort(struct device *dev,
 549                                         struct device_attribute *attr,
 550                                         char *buf)
 551 {
 552         return sysfs_emit(buf, "Not affected\n");
 553 }
 554
 555 ssize_t __weak cpu_show_itlb_multihit(struct device *dev,
 556                                       struct device_attribute *attr, char *buf)
 557 {
 558         return sysfs_emit(buf, "Not affected\n");
 559 }
 560
 561 ssize_t __weak cpu_show_srbds(struct device *dev,
 562                               struct device_attribute *attr, char *buf)
 563 {
 564         return sysfs_emit(buf, "Not affected\n");
 565 }
 566
 567 static DEVICE_ATTR(meltdown, 0444, cpu_show_meltdown, NULL);
 568 static DEVICE_ATTR(spectre_v1, 0444, cpu_show_spectre_v1, NULL);
 569 static DEVICE_ATTR(spectre_v2, 0444, cpu_show_spectre_v2, NULL);
 570 static DEVICE_ATTR(spec_store_bypass, 0444, cpu_show_spec_store_bypass, NULL);
 571 static DEVICE_ATTR(l1tf, 0444, cpu_show_l1tf, NULL);
 572 static DEVICE_ATTR(mds, 0444, cpu_show_mds, NULL);
 573 static DEVICE_ATTR(tsx_async_abort, 0444, cpu_show_tsx_async_abort, NULL);
 574 static DEVICE_ATTR(itlb_multihit, 0444, cpu_show_itlb_multihit, NULL);
 575 static DEVICE_ATTR(srbds, 0444, cpu_show_srbds, NULL);
 576
 577 static struct attribute *cpu_root_vulnerabilities_attrs[] = {
 578         &dev_attr_meltdown.attr,
 579         &dev_attr_spectre_v1.attr,
 580         &dev_attr_spectre_v2.attr,
 581         &dev_attr_spec_store_bypass.attr,
 582         &dev_attr_l1tf.attr,
 583         &dev_attr_mds.attr,
 584         &dev_attr_tsx_async_abort.attr,
 585         &dev_attr_itlb_multihit.attr,
 586         &dev_attr_srbds.attr,
 587         NULL
 588 };
 589
 590 static const struct attribute_group cpu_root_vulnerabilities_group = {
 591         .name  = "vulnerabilities",
 592         .attrs = cpu_root_vulnerabilities_attrs,
 593 };
 594
 595 static void __init cpu_register_vulnerabilities(void)
 596 {
 597         if (sysfs_create_group(&cpu_subsys.dev_root->kobj,
 598                                &cpu_root_vulnerabilities_group))
 599                 pr_err("Unable to register CPU vulnerabilities\n");
 600 }
 601
 602 #else
 603 static inline void cpu_register_vulnerabilities(void) { }
 604 #endif
 605
 606 void __init cpu_dev_init(void)
 607 {
 608         if (subsys_system_register(&cpu_subsys, cpu_root_attr_groups))
 609                 panic("Failed to register CPU subsystem");
 610
 611         cpu_dev_register_generic();
 612         cpu_register_vulnerabilities();
 613 }