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/delay.h>
  23 #include <linux/sched/isolation.h>
  24
  25 #include "base.h"
  26
  27 static DEFINE_PER_CPU(struct device *, cpu_sys_devices);
  28
  29 static int cpu_subsys_match(struct device *dev, struct device_driver *drv)
  30 {
  31         /* ACPI style match is the only one that may succeed. */
  32         if (acpi_driver_match_device(dev, drv))
  33                 return 1;
  34
  35         return 0;
  36 }
  37
  38 #ifdef CONFIG_HOTPLUG_CPU
  39 static void change_cpu_under_node(struct cpu *cpu,
  40                         unsigned int from_nid, unsigned int to_nid)
  41 {
  42         int cpuid = cpu->dev.id;
  43         unregister_cpu_under_node(cpuid, from_nid);
  44         register_cpu_under_node(cpuid, to_nid);
  45         cpu->node_id = to_nid;
  46 }
  47
  48 static int cpu_subsys_online(struct device *dev)
  49 {
  50         struct cpu *cpu = container_of(dev, struct cpu, dev);
  51         int cpuid = dev->id;
  52         int from_nid, to_nid;
  53         int ret;
  54         int retries = 0;
  55
  56         from_nid = cpu_to_node(cpuid);
  57         if (from_nid == NUMA_NO_NODE)
  58                 return -ENODEV;
  59
  60 retry:
  61         ret = cpu_device_up(dev);
  62
  63         /*
  64          * If -EBUSY is returned, it is likely that hotplug is temporarily
  65          * disabled when cpu_hotplug_disable() was called. This condition is
  66          * transient. So we retry after waiting for an exponentially
  67          * increasing delay up to a total of at least 620ms as some PCI
  68          * device initialization can take quite a while.
  69          */
  70         if (ret == -EBUSY) {
  71                 retries++;
  72                 if (retries > 5)
  73                         return ret;
  74                 msleep(10 * (1 << retries));
  75                 goto retry;
  76         }
  77
  78         /*
  79          * When hot adding memory to memoryless node and enabling a cpu
  80          * on the node, node number of the cpu may internally change.
  81          */
  82         to_nid = cpu_to_node(cpuid);
  83         if (from_nid != to_nid)
  84                 change_cpu_under_node(cpu, from_nid, to_nid);
  85
  86         return ret;
  87 }
  88
  89 static int cpu_subsys_offline(struct device *dev)
  90 {
  91         return cpu_device_down(dev);
  92 }
  93
  94 void unregister_cpu(struct cpu *cpu)
  95 {
  96         int logical_cpu = cpu->dev.id;
  97
  98         unregister_cpu_under_node(logical_cpu, cpu_to_node(logical_cpu));
  99
 100         device_unregister(&cpu->dev);
 101         per_cpu(cpu_sys_devices, logical_cpu) = NULL;
 102         return;
 103 }
 104
 105 #ifdef CONFIG_ARCH_CPU_PROBE_RELEASE
 106 static ssize_t cpu_probe_store(struct device *dev,
 107                                struct device_attribute *attr,
 108                                const char *buf,
 109                                size_t count)
 110 {
 111         ssize_t cnt;
 112         int ret;
 113
 114         ret = lock_device_hotplug_sysfs();
 115         if (ret)
 116                 return ret;
 117
 118         cnt = arch_cpu_probe(buf, count);
 119
 120         unlock_device_hotplug();
 121         return cnt;
 122 }
 123
 124 static ssize_t cpu_release_store(struct device *dev,
 125                                  struct device_attribute *attr,
 126                                  const char *buf,
 127                                  size_t count)
 128 {
 129         ssize_t cnt;
 130         int ret;
 131
 132         ret = lock_device_hotplug_sysfs();
 133         if (ret)
 134                 return ret;
 135
 136         cnt = arch_cpu_release(buf, count);
 137
 138         unlock_device_hotplug();
 139         return cnt;
 140 }
 141
 142 static DEVICE_ATTR(probe, S_IWUSR, NULL, cpu_probe_store);
 143 static DEVICE_ATTR(release, S_IWUSR, NULL, cpu_release_store);
 144 #endif /* CONFIG_ARCH_CPU_PROBE_RELEASE */
 145 #endif /* CONFIG_HOTPLUG_CPU */
 146
 147 #ifdef CONFIG_KEXEC
 148 #include <linux/kexec.h>
 149
 150 static ssize_t crash_notes_show(struct device *dev,
 151                                 struct device_attribute *attr,
 152                                 char *buf)
 153 {
 154         struct cpu *cpu = container_of(dev, struct cpu, dev);
 155         unsigned long long addr;
 156         int cpunum;
 157
 158         cpunum = cpu->dev.id;
 159
 160         /*
 161          * Might be reading other cpu's data based on which cpu read thread
 162          * has been scheduled. But cpu data (memory) is allocated once during
 163          * boot up and this data does not change there after. Hence this
 164          * operation should be safe. No locking required.
 165          */
 166         addr = per_cpu_ptr_to_phys(per_cpu_ptr(crash_notes, cpunum));
 167
 168         return sysfs_emit(buf, "%llx\n", addr);
 169 }
 170 static DEVICE_ATTR_ADMIN_RO(crash_notes);
 171
 172 static ssize_t crash_notes_size_show(struct device *dev,
 173                                      struct device_attribute *attr,
 174                                      char *buf)
 175 {
 176         return sysfs_emit(buf, "%zu\n", sizeof(note_buf_t));
 177 }
 178 static DEVICE_ATTR_ADMIN_RO(crash_notes_size);
 179
 180 static struct attribute *crash_note_cpu_attrs[] = {
 181         &dev_attr_crash_notes.attr,
 182         &dev_attr_crash_notes_size.attr,
 183         NULL
 184 };
 185
 186 static const struct attribute_group crash_note_cpu_attr_group = {
 187         .attrs = crash_note_cpu_attrs,
 188 };
 189 #endif
 190
 191 static const struct attribute_group *common_cpu_attr_groups[] = {
 192 #ifdef CONFIG_KEXEC
 193         &crash_note_cpu_attr_group,
 194 #endif
 195         NULL
 196 };
 197
 198 static const struct attribute_group *hotplugable_cpu_attr_groups[] = {
 199 #ifdef CONFIG_KEXEC
 200         &crash_note_cpu_attr_group,
 201 #endif
 202         NULL
 203 };
 204
 205 /*
 206  * Print cpu online, possible, present, and system maps
 207  */
 208
 209 struct cpu_attr {
 210         struct device_attribute attr;
 211         const struct cpumask *const map;
 212 };
 213
 214 static ssize_t show_cpus_attr(struct device *dev,
 215                               struct device_attribute *attr,
 216                               char *buf)
 217 {
 218         struct cpu_attr *ca = container_of(attr, struct cpu_attr, attr);
 219
 220         return cpumap_print_to_pagebuf(true, buf, ca->map);
 221 }
 222
 223 #define _CPU_ATTR(name, map) \
 224         { __ATTR(name, 0444, show_cpus_attr, NULL), map }
 225
 226 /* Keep in sync with cpu_subsys_attrs */
 227 static struct cpu_attr cpu_attrs[] = {
 228         _CPU_ATTR(online, &__cpu_online_mask),
 229         _CPU_ATTR(possible, &__cpu_possible_mask),
 230         _CPU_ATTR(present, &__cpu_present_mask),
 231 };
 232
 233 /*
 234  * Print values for NR_CPUS and offlined cpus
 235  */
 236 static ssize_t print_cpus_kernel_max(struct device *dev,
 237                                      struct device_attribute *attr, char *buf)
 238 {
 239         return sysfs_emit(buf, "%d\n", NR_CPUS - 1);
 240 }
 241 static DEVICE_ATTR(kernel_max, 0444, print_cpus_kernel_max, NULL);
 242
 243 /* arch-optional setting to enable display of offline cpus >= nr_cpu_ids */
 244 unsigned int total_cpus;
 245
 246 static ssize_t print_cpus_offline(struct device *dev,
 247                                   struct device_attribute *attr, char *buf)
 248 {
 249         int len = 0;
 250         cpumask_var_t offline;
 251
 252         /* display offline cpus < nr_cpu_ids */
 253         if (!alloc_cpumask_var(&offline, GFP_KERNEL))
 254                 return -ENOMEM;
 255         cpumask_andnot(offline, cpu_possible_mask, cpu_online_mask);
 256         len += sysfs_emit_at(buf, len, "%*pbl", cpumask_pr_args(offline));
 257         free_cpumask_var(offline);
 258
 259         /* display offline cpus >= nr_cpu_ids */
 260         if (total_cpus && nr_cpu_ids < total_cpus) {
 261                 len += sysfs_emit_at(buf, len, ",");
 262
 263                 if (nr_cpu_ids == total_cpus-1)
 264                         len += sysfs_emit_at(buf, len, "%u", nr_cpu_ids);
 265                 else
 266                         len += sysfs_emit_at(buf, len, "%u-%d",
 267                                              nr_cpu_ids, total_cpus - 1);
 268         }
 269
 270         len += sysfs_emit_at(buf, len, "\n");
 271
 272         return len;
 273 }
 274 static DEVICE_ATTR(offline, 0444, print_cpus_offline, NULL);
 275
 276 static ssize_t print_cpus_isolated(struct device *dev,
 277                                   struct device_attribute *attr, char *buf)
 278 {
 279         int len;
 280         cpumask_var_t isolated;
 281
 282         if (!alloc_cpumask_var(&isolated, GFP_KERNEL))
 283                 return -ENOMEM;
 284
 285         cpumask_andnot(isolated, cpu_possible_mask,
 286                        housekeeping_cpumask(HK_TYPE_DOMAIN));
 287         len = sysfs_emit(buf, "%*pbl\n", cpumask_pr_args(isolated));
 288
 289         free_cpumask_var(isolated);
 290
 291         return len;
 292 }
 293 static DEVICE_ATTR(isolated, 0444, print_cpus_isolated, NULL);
 294
 295 #ifdef CONFIG_NO_HZ_FULL
 296 static ssize_t print_cpus_nohz_full(struct device *dev,
 297                                     struct device_attribute *attr, char *buf)
 298 {
 299         return sysfs_emit(buf, "%*pbl\n", cpumask_pr_args(tick_nohz_full_mask));
 300 }
 301 static DEVICE_ATTR(nohz_full, 0444, print_cpus_nohz_full, NULL);
 302 #endif
 303
 304 static void cpu_device_release(struct device *dev)
 305 {
 306         /*
 307          * This is an empty function to prevent the driver core from spitting a
 308          * warning at us.  Yes, I know this is directly opposite of what the
 309          * documentation for the driver core and kobjects say, and the author
 310          * of this code has already been publically ridiculed for doing
 311          * something as foolish as this.  However, at this point in time, it is
 312          * the only way to handle the issue of statically allocated cpu
 313          * devices.  The different architectures will have their cpu device
 314          * code reworked to properly handle this in the near future, so this
 315          * function will then be changed to correctly free up the memory held
 316          * by the cpu device.
 317          *
 318          * Never copy this way of doing things, or you too will be made fun of
 319          * on the linux-kernel list, you have been warned.
 320          */
 321 }
 322
 323 #ifdef CONFIG_GENERIC_CPU_AUTOPROBE
 324 static ssize_t print_cpu_modalias(struct device *dev,
 325                                   struct device_attribute *attr,
 326                                   char *buf)
 327 {
 328         int len = 0;
 329         u32 i;
 330
 331         len += sysfs_emit_at(buf, len,
 332                              "cpu:type:" CPU_FEATURE_TYPEFMT ":feature:",
 333                              CPU_FEATURE_TYPEVAL);
 334
 335         for (i = 0; i < MAX_CPU_FEATURES; i++)
 336                 if (cpu_have_feature(i)) {
 337                         if (len + sizeof(",XXXX\n") >= PAGE_SIZE) {
 338                                 WARN(1, "CPU features overflow page\n");
 339                                 break;
 340                         }
 341                         len += sysfs_emit_at(buf, len, ",%04X", i);
 342                 }
 343         len += sysfs_emit_at(buf, len, "\n");
 344         return len;
 345 }
 346
 347 static int cpu_uevent(const struct device *dev, struct kobj_uevent_env *env)
 348 {
 349         char *buf = kzalloc(PAGE_SIZE, GFP_KERNEL);
 350         if (buf) {
 351                 print_cpu_modalias(NULL, NULL, buf);
 352                 add_uevent_var(env, "MODALIAS=%s", buf);
 353                 kfree(buf);
 354         }
 355         return 0;
 356 }
 357 #endif
 358
 359 struct bus_type cpu_subsys = {
 360         .name = "cpu",
 361         .dev_name = "cpu",
 362         .match = cpu_subsys_match,
 363 #ifdef CONFIG_HOTPLUG_CPU
 364         .online = cpu_subsys_online,
 365         .offline = cpu_subsys_offline,
 366 #endif
 367 #ifdef CONFIG_GENERIC_CPU_AUTOPROBE
 368         .uevent = cpu_uevent,
 369 #endif
 370 };
 371 EXPORT_SYMBOL_GPL(cpu_subsys);
 372
 373 /*
 374  * register_cpu - Setup a sysfs device for a CPU.
 375  * @cpu - cpu->hotpluggable field set to 1 will generate a control file in
 376  *        sysfs for this CPU.
 377  * @num - CPU number to use when creating the device.
 378  *
 379  * Initialize and register the CPU device.
 380  */
 381 int register_cpu(struct cpu *cpu, int num)
 382 {
 383         int error;
 384
 385         cpu->node_id = cpu_to_node(num);
 386         memset(&cpu->dev, 0x00, sizeof(struct device));
 387         cpu->dev.id = num;
 388         cpu->dev.bus = &cpu_subsys;
 389         cpu->dev.release = cpu_device_release;
 390         cpu->dev.offline_disabled = !cpu->hotpluggable;
 391         cpu->dev.offline = !cpu_online(num);
 392         cpu->dev.of_node = of_get_cpu_node(num, NULL);
 393         cpu->dev.groups = common_cpu_attr_groups;
 394         if (cpu->hotpluggable)
 395                 cpu->dev.groups = hotplugable_cpu_attr_groups;
 396         error = device_register(&cpu->dev);
 397         if (error) {
 398                 put_device(&cpu->dev);
 399                 return error;
 400         }
 401
 402         per_cpu(cpu_sys_devices, num) = &cpu->dev;
 403         register_cpu_under_node(num, cpu_to_node(num));
 404         dev_pm_qos_expose_latency_limit(&cpu->dev,
 405                                         PM_QOS_RESUME_LATENCY_NO_CONSTRAINT);
 406
 407         return 0;
 408 }
 409
 410 struct device *get_cpu_device(unsigned int cpu)
 411 {
 412         if (cpu < nr_cpu_ids && cpu_possible(cpu))
 413                 return per_cpu(cpu_sys_devices, cpu);
 414         else
 415                 return NULL;
 416 }
 417 EXPORT_SYMBOL_GPL(get_cpu_device);
 418
 419 static void device_create_release(struct device *dev)
 420 {
 421         kfree(dev);
 422 }
 423
 424 __printf(4, 0)
 425 static struct device *
 426 __cpu_device_create(struct device *parent, void *drvdata,
 427                     const struct attribute_group **groups,
 428                     const char *fmt, va_list args)
 429 {
 430         struct device *dev = NULL;
 431         int retval = -ENOMEM;
 432
 433         dev = kzalloc(sizeof(*dev), GFP_KERNEL);
 434         if (!dev)
 435                 goto error;
 436
 437         device_initialize(dev);
 438         dev->parent = parent;
 439         dev->groups = groups;
 440         dev->release = device_create_release;
 441         device_set_pm_not_required(dev);
 442         dev_set_drvdata(dev, drvdata);
 443
 444         retval = kobject_set_name_vargs(&dev->kobj, fmt, args);
 445         if (retval)
 446                 goto error;
 447
 448         retval = device_add(dev);
 449         if (retval)
 450                 goto error;
 451
 452         return dev;
 453
 454 error:
 455         put_device(dev);
 456         return ERR_PTR(retval);
 457 }
 458
 459 struct device *cpu_device_create(struct device *parent, void *drvdata,
 460                                  const struct attribute_group **groups,
 461                                  const char *fmt, ...)
 462 {
 463         va_list vargs;
 464         struct device *dev;
 465
 466         va_start(vargs, fmt);
 467         dev = __cpu_device_create(parent, drvdata, groups, fmt, vargs);
 468         va_end(vargs);
 469         return dev;
 470 }
 471 EXPORT_SYMBOL_GPL(cpu_device_create);
 472
 473 #ifdef CONFIG_GENERIC_CPU_AUTOPROBE
 474 static DEVICE_ATTR(modalias, 0444, print_cpu_modalias, NULL);
 475 #endif
 476
 477 static struct attribute *cpu_root_attrs[] = {
 478 #ifdef CONFIG_ARCH_CPU_PROBE_RELEASE
 479         &dev_attr_probe.attr,
 480         &dev_attr_release.attr,
 481 #endif
 482         &cpu_attrs[0].attr.attr,
 483         &cpu_attrs[1].attr.attr,
 484         &cpu_attrs[2].attr.attr,
 485         &dev_attr_kernel_max.attr,
 486         &dev_attr_offline.attr,
 487         &dev_attr_isolated.attr,
 488 #ifdef CONFIG_NO_HZ_FULL
 489         &dev_attr_nohz_full.attr,
 490 #endif
 491 #ifdef CONFIG_GENERIC_CPU_AUTOPROBE
 492         &dev_attr_modalias.attr,
 493 #endif
 494         NULL
 495 };
 496
 497 static const struct attribute_group cpu_root_attr_group = {
 498         .attrs = cpu_root_attrs,
 499 };
 500
 501 static const struct attribute_group *cpu_root_attr_groups[] = {
 502         &cpu_root_attr_group,
 503         NULL,
 504 };
 505
 506 bool cpu_is_hotpluggable(unsigned int cpu)
 507 {
 508         struct device *dev = get_cpu_device(cpu);
 509         return dev && container_of(dev, struct cpu, dev)->hotpluggable
 510                 && tick_nohz_cpu_hotpluggable(cpu);
 511 }
 512 EXPORT_SYMBOL_GPL(cpu_is_hotpluggable);
 513
 514 #ifdef CONFIG_GENERIC_CPU_DEVICES
 515 static DEFINE_PER_CPU(struct cpu, cpu_devices);
 516 #endif
 517
 518 static void __init cpu_dev_register_generic(void)
 519 {
 520 #ifdef CONFIG_GENERIC_CPU_DEVICES
 521         int i;
 522
 523         for_each_possible_cpu(i) {
 524                 if (register_cpu(&per_cpu(cpu_devices, i), i))
 525                         panic("Failed to register CPU device");
 526         }
 527 #endif
 528 }
 529
 530 #ifdef CONFIG_GENERIC_CPU_VULNERABILITIES
 531
 532 ssize_t __weak cpu_show_meltdown(struct device *dev,
 533                                  struct device_attribute *attr, char *buf)
 534 {
 535         return sysfs_emit(buf, "Not affected\n");
 536 }
 537
 538 ssize_t __weak cpu_show_spectre_v1(struct device *dev,
 539                                    struct device_attribute *attr, char *buf)
 540 {
 541         return sysfs_emit(buf, "Not affected\n");
 542 }
 543
 544 ssize_t __weak cpu_show_spectre_v2(struct device *dev,
 545                                    struct device_attribute *attr, char *buf)
 546 {
 547         return sysfs_emit(buf, "Not affected\n");
 548 }
 549
 550 ssize_t __weak cpu_show_spec_store_bypass(struct device *dev,
 551                                           struct device_attribute *attr, char *buf)
 552 {
 553         return sysfs_emit(buf, "Not affected\n");
 554 }
 555
 556 ssize_t __weak cpu_show_l1tf(struct device *dev,
 557                              struct device_attribute *attr, char *buf)
 558 {
 559         return sysfs_emit(buf, "Not affected\n");
 560 }
 561
 562 ssize_t __weak cpu_show_mds(struct device *dev,
 563                             struct device_attribute *attr, char *buf)
 564 {
 565         return sysfs_emit(buf, "Not affected\n");
 566 }
 567
 568 ssize_t __weak cpu_show_tsx_async_abort(struct device *dev,
 569                                         struct device_attribute *attr,
 570                                         char *buf)
 571 {
 572         return sysfs_emit(buf, "Not affected\n");
 573 }
 574
 575 ssize_t __weak cpu_show_itlb_multihit(struct device *dev,
 576                                       struct device_attribute *attr, char *buf)
 577 {
 578         return sysfs_emit(buf, "Not affected\n");
 579 }
 580
 581 ssize_t __weak cpu_show_srbds(struct device *dev,
 582                               struct device_attribute *attr, char *buf)
 583 {
 584         return sysfs_emit(buf, "Not affected\n");
 585 }
 586
 587 ssize_t __weak cpu_show_mmio_stale_data(struct device *dev,
 588                                         struct device_attribute *attr, char *buf)
 589 {
 590         return sysfs_emit(buf, "Not affected\n");
 591 }
 592
 593 ssize_t __weak cpu_show_retbleed(struct device *dev,
 594                                  struct device_attribute *attr, char *buf)
 595 {
 596         return sysfs_emit(buf, "Not affected\n");
 597 }
 598
 599 static DEVICE_ATTR(meltdown, 0444, cpu_show_meltdown, NULL);
 600 static DEVICE_ATTR(spectre_v1, 0444, cpu_show_spectre_v1, NULL);
 601 static DEVICE_ATTR(spectre_v2, 0444, cpu_show_spectre_v2, NULL);
 602 static DEVICE_ATTR(spec_store_bypass, 0444, cpu_show_spec_store_bypass, NULL);
 603 static DEVICE_ATTR(l1tf, 0444, cpu_show_l1tf, NULL);
 604 static DEVICE_ATTR(mds, 0444, cpu_show_mds, NULL);
 605 static DEVICE_ATTR(tsx_async_abort, 0444, cpu_show_tsx_async_abort, NULL);
 606 static DEVICE_ATTR(itlb_multihit, 0444, cpu_show_itlb_multihit, NULL);
 607 static DEVICE_ATTR(srbds, 0444, cpu_show_srbds, NULL);
 608 static DEVICE_ATTR(mmio_stale_data, 0444, cpu_show_mmio_stale_data, NULL);
 609 static DEVICE_ATTR(retbleed, 0444, cpu_show_retbleed, NULL);
 610
 611 static struct attribute *cpu_root_vulnerabilities_attrs[] = {
 612         &dev_attr_meltdown.attr,
 613         &dev_attr_spectre_v1.attr,
 614         &dev_attr_spectre_v2.attr,
 615         &dev_attr_spec_store_bypass.attr,
 616         &dev_attr_l1tf.attr,
 617         &dev_attr_mds.attr,
 618         &dev_attr_tsx_async_abort.attr,
 619         &dev_attr_itlb_multihit.attr,
 620         &dev_attr_srbds.attr,
 621         &dev_attr_mmio_stale_data.attr,
 622         &dev_attr_retbleed.attr,
 623         NULL
 624 };
 625
 626 static const struct attribute_group cpu_root_vulnerabilities_group = {
 627         .name  = "vulnerabilities",
 628         .attrs = cpu_root_vulnerabilities_attrs,
 629 };
 630
 631 static void __init cpu_register_vulnerabilities(void)
 632 {
 633         struct device *dev = bus_get_dev_root(&cpu_subsys);
 634
 635         if (dev) {
 636                 if (sysfs_create_group(&dev->kobj, &cpu_root_vulnerabilities_group))
 637                         pr_err("Unable to register CPU vulnerabilities\n");
 638                 put_device(dev);
 639         }
 640 }
 641
 642 #else
 643 static inline void cpu_register_vulnerabilities(void) { }
 644 #endif
 645
 646 void __init cpu_dev_init(void)
 647 {
 648         if (subsys_system_register(&cpu_subsys, cpu_root_attr_groups))
 649                 panic("Failed to register CPU subsystem");
 650
 651         cpu_dev_register_generic();
 652         cpu_register_vulnerabilities();
 653 }