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 show_crash_notes(struct device *dev, struct device_attribute *attr,
 143                                 char *buf)
 144 {
 145         struct cpu *cpu = container_of(dev, struct cpu, dev);
 146         ssize_t rc;
 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         rc = sprintf(buf, "%Lx\n", addr);
 160         return rc;
 161 }
 162 static DEVICE_ATTR(crash_notes, 0400, show_crash_notes, NULL);
 163
 164 static ssize_t show_crash_notes_size(struct device *dev,
 165                                      struct device_attribute *attr,
 166                                      char *buf)
 167 {
 168         ssize_t rc;
 169
 170         rc = sprintf(buf, "%zu\n", sizeof(note_buf_t));
 171         return rc;
 172 }
 173 static DEVICE_ATTR(crash_notes_size, 0400, show_crash_notes_size, NULL);
 174
 175 static struct attribute *crash_note_cpu_attrs[] = {
 176         &dev_attr_crash_notes.attr,
 177         &dev_attr_crash_notes_size.attr,
 178         NULL
 179 };
 180
 181 static struct attribute_group crash_note_cpu_attr_group = {
 182         .attrs = crash_note_cpu_attrs,
 183 };
 184 #endif
 185
 186 static const struct attribute_group *common_cpu_attr_groups[] = {
 187 #ifdef CONFIG_KEXEC
 188         &crash_note_cpu_attr_group,
 189 #endif
 190         NULL
 191 };
 192
 193 static const struct attribute_group *hotplugable_cpu_attr_groups[] = {
 194 #ifdef CONFIG_KEXEC
 195         &crash_note_cpu_attr_group,
 196 #endif
 197         NULL
 198 };
 199
 200 /*
 201  * Print cpu online, possible, present, and system maps
 202  */
 203
 204 struct cpu_attr {
 205         struct device_attribute attr;
 206         const struct cpumask *const map;
 207 };
 208
 209 static ssize_t show_cpus_attr(struct device *dev,
 210                               struct device_attribute *attr,
 211                               char *buf)
 212 {
 213         struct cpu_attr *ca = container_of(attr, struct cpu_attr, attr);
 214
 215         return cpumap_print_to_pagebuf(true, buf, ca->map);
 216 }
 217
 218 #define _CPU_ATTR(name, map) \
 219         { __ATTR(name, 0444, show_cpus_attr, NULL), map }
 220
 221 /* Keep in sync with cpu_subsys_attrs */
 222 static struct cpu_attr cpu_attrs[] = {
 223         _CPU_ATTR(online, &__cpu_online_mask),
 224         _CPU_ATTR(possible, &__cpu_possible_mask),
 225         _CPU_ATTR(present, &__cpu_present_mask),
 226 };
 227
 228 /*
 229  * Print values for NR_CPUS and offlined cpus
 230  */
 231 static ssize_t print_cpus_kernel_max(struct device *dev,
 232                                      struct device_attribute *attr, char *buf)
 233 {
 234         return sprintf(buf, "%d\n", NR_CPUS - 1);
 235 }
 236 static DEVICE_ATTR(kernel_max, 0444, print_cpus_kernel_max, NULL);
 237
 238 /* arch-optional setting to enable display of offline cpus >= nr_cpu_ids */
 239 unsigned int total_cpus;
 240
 241 static ssize_t print_cpus_offline(struct device *dev,
 242                                   struct device_attribute *attr, char *buf)
 243 {
 244         int n = 0, len = PAGE_SIZE-2;
 245         cpumask_var_t offline;
 246
 247         /* display offline cpus < nr_cpu_ids */
 248         if (!alloc_cpumask_var(&offline, GFP_KERNEL))
 249                 return -ENOMEM;
 250         cpumask_andnot(offline, cpu_possible_mask, cpu_online_mask);
 251         n = scnprintf(buf, len, "%*pbl", cpumask_pr_args(offline));
 252         free_cpumask_var(offline);
 253
 254         /* display offline cpus >= nr_cpu_ids */
 255         if (total_cpus && nr_cpu_ids < total_cpus) {
 256                 if (n && n < len)
 257                         buf[n++] = ',';
 258
 259                 if (nr_cpu_ids == total_cpus-1)
 260                         n += scnprintf(&buf[n], len - n, "%u", nr_cpu_ids);
 261                 else
 262                         n += scnprintf(&buf[n], len - n, "%u-%d",
 263                                                       nr_cpu_ids, total_cpus-1);
 264         }
 265
 266         n += scnprintf(&buf[n], len - n, "\n");
 267         return n;
 268 }
 269 static DEVICE_ATTR(offline, 0444, print_cpus_offline, NULL);
 270
 271 static ssize_t print_cpus_isolated(struct device *dev,
 272                                   struct device_attribute *attr, char *buf)
 273 {
 274         int n;
 275         cpumask_var_t isolated;
 276
 277         if (!alloc_cpumask_var(&isolated, GFP_KERNEL))
 278                 return -ENOMEM;
 279
 280         cpumask_andnot(isolated, cpu_possible_mask,
 281                        housekeeping_cpumask(HK_FLAG_DOMAIN));
 282         n = sprintf(buf, "%*pbl\n", cpumask_pr_args(isolated));
 283
 284         free_cpumask_var(isolated);
 285
 286         return n;
 287 }
 288 static DEVICE_ATTR(isolated, 0444, print_cpus_isolated, NULL);
 289
 290 #ifdef CONFIG_NO_HZ_FULL
 291 static ssize_t print_cpus_nohz_full(struct device *dev,
 292                                   struct device_attribute *attr, char *buf)
 293 {
 294         return sprintf(buf, "%*pbl\n", cpumask_pr_args(tick_nohz_full_mask));
 295 }
 296 static DEVICE_ATTR(nohz_full, 0444, print_cpus_nohz_full, NULL);
 297 #endif
 298
 299 static void cpu_device_release(struct device *dev)
 300 {
 301         /*
 302          * This is an empty function to prevent the driver core from spitting a
 303          * warning at us.  Yes, I know this is directly opposite of what the
 304          * documentation for the driver core and kobjects say, and the author
 305          * of this code has already been publically ridiculed for doing
 306          * something as foolish as this.  However, at this point in time, it is
 307          * the only way to handle the issue of statically allocated cpu
 308          * devices.  The different architectures will have their cpu device
 309          * code reworked to properly handle this in the near future, so this
 310          * function will then be changed to correctly free up the memory held
 311          * by the cpu device.
 312          *
 313          * Never copy this way of doing things, or you too will be made fun of
 314          * on the linux-kernel list, you have been warned.
 315          */
 316 }
 317
 318 #ifdef CONFIG_GENERIC_CPU_AUTOPROBE
 319 static ssize_t print_cpu_modalias(struct device *dev,
 320                                   struct device_attribute *attr,
 321                                   char *buf)
 322 {
 323         ssize_t n;
 324         u32 i;
 325
 326         n = sprintf(buf, "cpu:type:" CPU_FEATURE_TYPEFMT ":feature:",
 327                     CPU_FEATURE_TYPEVAL);
 328
 329         for (i = 0; i < MAX_CPU_FEATURES; i++)
 330                 if (cpu_have_feature(i)) {
 331                         if (PAGE_SIZE < n + sizeof(",XXXX\n")) {
 332                                 WARN(1, "CPU features overflow page\n");
 333                                 break;
 334                         }
 335                         n += sprintf(&buf[n], ",%04X", i);
 336                 }
 337         buf[n++] = '\n';
 338         return n;
 339 }
 340
 341 static int cpu_uevent(struct device *dev, struct kobj_uevent_env *env)
 342 {
 343         char *buf = kzalloc(PAGE_SIZE, GFP_KERNEL);
 344         if (buf) {
 345                 print_cpu_modalias(NULL, NULL, buf);
 346                 add_uevent_var(env, "MODALIAS=%s", buf);
 347                 kfree(buf);
 348         }
 349         return 0;
 350 }
 351 #endif
 352
 353 /*
 354  * register_cpu - Setup a sysfs device for a CPU.
 355  * @cpu - cpu->hotpluggable field set to 1 will generate a control file in
 356  *        sysfs for this CPU.
 357  * @num - CPU number to use when creating the device.
 358  *
 359  * Initialize and register the CPU device.
 360  */
 361 int register_cpu(struct cpu *cpu, int num)
 362 {
 363         int error;
 364
 365         cpu->node_id = cpu_to_node(num);
 366         memset(&cpu->dev, 0x00, sizeof(struct device));
 367         cpu->dev.id = num;
 368         cpu->dev.bus = &cpu_subsys;
 369         cpu->dev.release = cpu_device_release;
 370         cpu->dev.offline_disabled = !cpu->hotpluggable;
 371         cpu->dev.offline = !cpu_online(num);
 372         cpu->dev.of_node = of_get_cpu_node(num, NULL);
 373 #ifdef CONFIG_GENERIC_CPU_AUTOPROBE
 374         cpu->dev.bus->uevent = cpu_uevent;
 375 #endif
 376         cpu->dev.groups = common_cpu_attr_groups;
 377         if (cpu->hotpluggable)
 378                 cpu->dev.groups = hotplugable_cpu_attr_groups;
 379         error = device_register(&cpu->dev);
 380         if (error) {
 381                 put_device(&cpu->dev);
 382                 return error;
 383         }
 384
 385         per_cpu(cpu_sys_devices, num) = &cpu->dev;
 386         register_cpu_under_node(num, cpu_to_node(num));
 387         dev_pm_qos_expose_latency_limit(&cpu->dev,
 388                                         PM_QOS_RESUME_LATENCY_NO_CONSTRAINT);
 389
 390         return 0;
 391 }
 392
 393 struct device *get_cpu_device(unsigned cpu)
 394 {
 395         if (cpu < nr_cpu_ids && cpu_possible(cpu))
 396                 return per_cpu(cpu_sys_devices, cpu);
 397         else
 398                 return NULL;
 399 }
 400 EXPORT_SYMBOL_GPL(get_cpu_device);
 401
 402 static void device_create_release(struct device *dev)
 403 {
 404         kfree(dev);
 405 }
 406
 407 __printf(4, 0)
 408 static struct device *
 409 __cpu_device_create(struct device *parent, void *drvdata,
 410                     const struct attribute_group **groups,
 411                     const char *fmt, va_list args)
 412 {
 413         struct device *dev = NULL;
 414         int retval = -ENODEV;
 415
 416         dev = kzalloc(sizeof(*dev), GFP_KERNEL);
 417         if (!dev) {
 418                 retval = -ENOMEM;
 419                 goto error;
 420         }
 421
 422         device_initialize(dev);
 423         dev->parent = parent;
 424         dev->groups = groups;
 425         dev->release = device_create_release;
 426         device_set_pm_not_required(dev);
 427         dev_set_drvdata(dev, drvdata);
 428
 429         retval = kobject_set_name_vargs(&dev->kobj, fmt, args);
 430         if (retval)
 431                 goto error;
 432
 433         retval = device_add(dev);
 434         if (retval)
 435                 goto error;
 436
 437         return dev;
 438
 439 error:
 440         put_device(dev);
 441         return ERR_PTR(retval);
 442 }
 443
 444 struct device *cpu_device_create(struct device *parent, void *drvdata,
 445                                  const struct attribute_group **groups,
 446                                  const char *fmt, ...)
 447 {
 448         va_list vargs;
 449         struct device *dev;
 450
 451         va_start(vargs, fmt);
 452         dev = __cpu_device_create(parent, drvdata, groups, fmt, vargs);
 453         va_end(vargs);
 454         return dev;
 455 }
 456 EXPORT_SYMBOL_GPL(cpu_device_create);
 457
 458 #ifdef CONFIG_GENERIC_CPU_AUTOPROBE
 459 static DEVICE_ATTR(modalias, 0444, print_cpu_modalias, NULL);
 460 #endif
 461
 462 static struct attribute *cpu_root_attrs[] = {
 463 #ifdef CONFIG_ARCH_CPU_PROBE_RELEASE
 464         &dev_attr_probe.attr,
 465         &dev_attr_release.attr,
 466 #endif
 467         &cpu_attrs[0].attr.attr,
 468         &cpu_attrs[1].attr.attr,
 469         &cpu_attrs[2].attr.attr,
 470         &dev_attr_kernel_max.attr,
 471         &dev_attr_offline.attr,
 472         &dev_attr_isolated.attr,
 473 #ifdef CONFIG_NO_HZ_FULL
 474         &dev_attr_nohz_full.attr,
 475 #endif
 476 #ifdef CONFIG_GENERIC_CPU_AUTOPROBE
 477         &dev_attr_modalias.attr,
 478 #endif
 479         NULL
 480 };
 481
 482 static struct attribute_group cpu_root_attr_group = {
 483         .attrs = cpu_root_attrs,
 484 };
 485
 486 static const struct attribute_group *cpu_root_attr_groups[] = {
 487         &cpu_root_attr_group,
 488         NULL,
 489 };
 490
 491 bool cpu_is_hotpluggable(unsigned cpu)
 492 {
 493         struct device *dev = get_cpu_device(cpu);
 494         return dev && container_of(dev, struct cpu, dev)->hotpluggable;
 495 }
 496 EXPORT_SYMBOL_GPL(cpu_is_hotpluggable);
 497
 498 #ifdef CONFIG_GENERIC_CPU_DEVICES
 499 static DEFINE_PER_CPU(struct cpu, cpu_devices);
 500 #endif
 501
 502 static void __init cpu_dev_register_generic(void)
 503 {
 504 #ifdef CONFIG_GENERIC_CPU_DEVICES
 505         int i;
 506
 507         for_each_possible_cpu(i) {
 508                 if (register_cpu(&per_cpu(cpu_devices, i), i))
 509                         panic("Failed to register CPU device");
 510         }
 511 #endif
 512 }
 513
 514 #ifdef CONFIG_GENERIC_CPU_VULNERABILITIES
 515
 516 ssize_t __weak cpu_show_meltdown(struct device *dev,
 517                                  struct device_attribute *attr, char *buf)
 518 {
 519         return sprintf(buf, "Not affected\n");
 520 }
 521
 522 ssize_t __weak cpu_show_spectre_v1(struct device *dev,
 523                                    struct device_attribute *attr, char *buf)
 524 {
 525         return sprintf(buf, "Not affected\n");
 526 }
 527
 528 ssize_t __weak cpu_show_spectre_v2(struct device *dev,
 529                                    struct device_attribute *attr, char *buf)
 530 {
 531         return sprintf(buf, "Not affected\n");
 532 }
 533
 534 ssize_t __weak cpu_show_spec_store_bypass(struct device *dev,
 535                                           struct device_attribute *attr, char *buf)
 536 {
 537         return sprintf(buf, "Not affected\n");
 538 }
 539
 540 ssize_t __weak cpu_show_l1tf(struct device *dev,
 541                              struct device_attribute *attr, char *buf)
 542 {
 543         return sprintf(buf, "Not affected\n");
 544 }
 545
 546 ssize_t __weak cpu_show_mds(struct device *dev,
 547                             struct device_attribute *attr, char *buf)
 548 {
 549         return sprintf(buf, "Not affected\n");
 550 }
 551
 552 ssize_t __weak cpu_show_tsx_async_abort(struct device *dev,
 553                                         struct device_attribute *attr,
 554                                         char *buf)
 555 {
 556         return sprintf(buf, "Not affected\n");
 557 }
 558
 559 ssize_t __weak cpu_show_itlb_multihit(struct device *dev,
 560                             struct device_attribute *attr, char *buf)
 561 {
 562         return sprintf(buf, "Not affected\n");
 563 }
 564
 565 static DEVICE_ATTR(meltdown, 0444, cpu_show_meltdown, NULL);
 566 static DEVICE_ATTR(spectre_v1, 0444, cpu_show_spectre_v1, NULL);
 567 static DEVICE_ATTR(spectre_v2, 0444, cpu_show_spectre_v2, NULL);
 568 static DEVICE_ATTR(spec_store_bypass, 0444, cpu_show_spec_store_bypass, NULL);
 569 static DEVICE_ATTR(l1tf, 0444, cpu_show_l1tf, NULL);
 570 static DEVICE_ATTR(mds, 0444, cpu_show_mds, NULL);
 571 static DEVICE_ATTR(tsx_async_abort, 0444, cpu_show_tsx_async_abort, NULL);
 572 static DEVICE_ATTR(itlb_multihit, 0444, cpu_show_itlb_multihit, NULL);
 573
 574 static struct attribute *cpu_root_vulnerabilities_attrs[] = {
 575         &dev_attr_meltdown.attr,
 576         &dev_attr_spectre_v1.attr,
 577         &dev_attr_spectre_v2.attr,
 578         &dev_attr_spec_store_bypass.attr,
 579         &dev_attr_l1tf.attr,
 580         &dev_attr_mds.attr,
 581         &dev_attr_tsx_async_abort.attr,
 582         &dev_attr_itlb_multihit.attr,
 583         NULL
 584 };
 585
 586 static const struct attribute_group cpu_root_vulnerabilities_group = {
 587         .name  = "vulnerabilities",
 588         .attrs = cpu_root_vulnerabilities_attrs,
 589 };
 590
 591 static void __init cpu_register_vulnerabilities(void)
 592 {
 593         if (sysfs_create_group(&cpu_subsys.dev_root->kobj,
 594                                &cpu_root_vulnerabilities_group))
 595                 pr_err("Unable to register CPU vulnerabilities\n");
 596 }
 597
 598 #else
 599 static inline void cpu_register_vulnerabilities(void) { }
 600 #endif
 601
 602 void __init cpu_dev_init(void)
 603 {
 604         if (subsys_system_register(&cpu_subsys, cpu_root_attr_groups))
 605                 panic("Failed to register CPU subsystem");
 606
 607         cpu_dev_register_generic();
 608         cpu_register_vulnerabilities();
 609 }