drivers/base/memory.c

   1 /*
   2  * drivers/base/memory.c - basic Memory class support
   3  *
   4  * Written by Matt Tolentino <matthew.e.tolentino@intel.com>
   5  *            Dave Hansen <haveblue@us.ibm.com>
   6  *
   7  * This file provides the necessary infrastructure to represent
   8  * a SPARSEMEM-memory-model system's physical memory in /sysfs.
   9  * All arch-independent code that assumes MEMORY_HOTPLUG requires
  10  * SPARSEMEM should be contained here, or in mm/memory_hotplug.c.
  11  */
  12
  13 #include <linux/sysdev.h>
  14 #include <linux/module.h>
  15 #include <linux/init.h>
  16 #include <linux/topology.h>
  17 #include <linux/capability.h>
  18 #include <linux/device.h>
  19 #include <linux/memory.h>
  20 #include <linux/kobject.h>
  21 #include <linux/memory_hotplug.h>
  22 #include <linux/mm.h>
  23 #include <linux/mutex.h>
  24 #include <linux/stat.h>
  25 #include <linux/slab.h>
  26
  27 #include <asm/atomic.h>
  28 #include <asm/uaccess.h>
  29
  30 static DEFINE_MUTEX(mem_sysfs_mutex);
  31
  32 #define MEMORY_CLASS_NAME       "memory"
  33 #define MIN_MEMORY_BLOCK_SIZE   (1 << SECTION_SIZE_BITS)
  34
  35 static int sections_per_block;
  36
  37 static inline int base_memory_block_id(int section_nr)
  38 {
  39         return section_nr / sections_per_block;
  40 }
  41
  42 static struct sysdev_class memory_sysdev_class = {
  43         .name = MEMORY_CLASS_NAME,
  44 };
  45
  46 static const char *memory_uevent_name(struct kset *kset, struct kobject *kobj)
  47 {
  48         return MEMORY_CLASS_NAME;
  49 }
  50
  51 static int memory_uevent(struct kset *kset, struct kobject *obj, struct kobj_uevent_env *env)
  52 {
  53         int retval = 0;
  54
  55         return retval;
  56 }
  57
  58 static const struct kset_uevent_ops memory_uevent_ops = {
  59         .name           = memory_uevent_name,
  60         .uevent         = memory_uevent,
  61 };
  62
  63 static BLOCKING_NOTIFIER_HEAD(memory_chain);
  64
  65 int register_memory_notifier(struct notifier_block *nb)
  66 {
  67         return blocking_notifier_chain_register(&memory_chain, nb);
  68 }
  69 EXPORT_SYMBOL(register_memory_notifier);
  70
  71 void unregister_memory_notifier(struct notifier_block *nb)
  72 {
  73         blocking_notifier_chain_unregister(&memory_chain, nb);
  74 }
  75 EXPORT_SYMBOL(unregister_memory_notifier);
  76
  77 static ATOMIC_NOTIFIER_HEAD(memory_isolate_chain);
  78
  79 int register_memory_isolate_notifier(struct notifier_block *nb)
  80 {
  81         return atomic_notifier_chain_register(&memory_isolate_chain, nb);
  82 }
  83 EXPORT_SYMBOL(register_memory_isolate_notifier);
  84
  85 void unregister_memory_isolate_notifier(struct notifier_block *nb)
  86 {
  87         atomic_notifier_chain_unregister(&memory_isolate_chain, nb);
  88 }
  89 EXPORT_SYMBOL(unregister_memory_isolate_notifier);
  90
  91 /*
  92  * register_memory - Setup a sysfs device for a memory block
  93  */
  94 static
  95 int register_memory(struct memory_block *memory)
  96 {
  97         int error;
  98
  99         memory->sysdev.cls = &memory_sysdev_class;
 100         memory->sysdev.id = memory->start_section_nr / sections_per_block;
 101
 102         error = sysdev_register(&memory->sysdev);
 103         return error;
 104 }
 105
 106 static void
 107 unregister_memory(struct memory_block *memory)
 108 {
 109         BUG_ON(memory->sysdev.cls != &memory_sysdev_class);
 110
 111         /* drop the ref. we got in remove_memory_block() */
 112         kobject_put(&memory->sysdev.kobj);
 113         sysdev_unregister(&memory->sysdev);
 114 }
 115
 116 unsigned long __weak memory_block_size_bytes(void)
 117 {
 118         return MIN_MEMORY_BLOCK_SIZE;
 119 }
 120
 121 static unsigned long get_memory_block_size(void)
 122 {
 123         unsigned long block_sz;
 124
 125         block_sz = memory_block_size_bytes();
 126
 127         /* Validate blk_sz is a power of 2 and not less than section size */
 128         if ((block_sz & (block_sz - 1)) || (block_sz < MIN_MEMORY_BLOCK_SIZE)) {
 129                 WARN_ON(1);
 130                 block_sz = MIN_MEMORY_BLOCK_SIZE;
 131         }
 132
 133         return block_sz;
 134 }
 135
 136 /*
 137  * use this as the physical section index that this memsection
 138  * uses.
 139  */
 140
 141 static ssize_t show_mem_start_phys_index(struct sys_device *dev,
 142                         struct sysdev_attribute *attr, char *buf)
 143 {
 144         struct memory_block *mem =
 145                 container_of(dev, struct memory_block, sysdev);
 146         unsigned long phys_index;
 147
 148         phys_index = mem->start_section_nr / sections_per_block;
 149         return sprintf(buf, "%08lx\n", phys_index);
 150 }
 151
 152 static ssize_t show_mem_end_phys_index(struct sys_device *dev,
 153                         struct sysdev_attribute *attr, char *buf)
 154 {
 155         struct memory_block *mem =
 156                 container_of(dev, struct memory_block, sysdev);
 157         unsigned long phys_index;
 158
 159         phys_index = mem->end_section_nr / sections_per_block;
 160         return sprintf(buf, "%08lx\n", phys_index);
 161 }
 162
 163 /*
 164  * Show whether the section of memory is likely to be hot-removable
 165  */
 166 static ssize_t show_mem_removable(struct sys_device *dev,
 167                         struct sysdev_attribute *attr, char *buf)
 168 {
 169         unsigned long i, pfn;
 170         int ret = 1;
 171         struct memory_block *mem =
 172                 container_of(dev, struct memory_block, sysdev);
 173
 174         for (i = 0; i < sections_per_block; i++) {
 175                 pfn = section_nr_to_pfn(mem->start_section_nr + i);
 176                 ret &= is_mem_section_removable(pfn, PAGES_PER_SECTION);
 177         }
 178
 179         return sprintf(buf, "%d\n", ret);
 180 }
 181
 182 /*
 183  * online, offline, going offline, etc.
 184  */
 185 static ssize_t show_mem_state(struct sys_device *dev,
 186                         struct sysdev_attribute *attr, char *buf)
 187 {
 188         struct memory_block *mem =
 189                 container_of(dev, struct memory_block, sysdev);
 190         ssize_t len = 0;
 191
 192         /*
 193          * We can probably put these states in a nice little array
 194          * so that they're not open-coded
 195          */
 196         switch (mem->state) {
 197                 case MEM_ONLINE:
 198                         len = sprintf(buf, "online\n");
 199                         break;
 200                 case MEM_OFFLINE:
 201                         len = sprintf(buf, "offline\n");
 202                         break;
 203                 case MEM_GOING_OFFLINE:
 204                         len = sprintf(buf, "going-offline\n");
 205                         break;
 206                 default:
 207                         len = sprintf(buf, "ERROR-UNKNOWN-%ld\n",
 208                                         mem->state);
 209                         WARN_ON(1);
 210                         break;
 211         }
 212
 213         return len;
 214 }
 215
 216 int memory_notify(unsigned long val, void *v)
 217 {
 218         return blocking_notifier_call_chain(&memory_chain, val, v);
 219 }
 220
 221 int memory_isolate_notify(unsigned long val, void *v)
 222 {
 223         return atomic_notifier_call_chain(&memory_isolate_chain, val, v);
 224 }
 225
 226 /*
 227  * MEMORY_HOTPLUG depends on SPARSEMEM in mm/Kconfig, so it is
 228  * OK to have direct references to sparsemem variables in here.
 229  */
 230 static int
 231 memory_section_action(unsigned long phys_index, unsigned long action)
 232 {
 233         int i;
 234         unsigned long start_pfn, start_paddr;
 235         struct page *first_page;
 236         int ret;
 237
 238         first_page = pfn_to_page(phys_index << PFN_SECTION_SHIFT);
 239
 240         /*
 241          * The probe routines leave the pages reserved, just
 242          * as the bootmem code does.  Make sure they're still
 243          * that way.
 244          */
 245         if (action == MEM_ONLINE) {
 246                 for (i = 0; i < PAGES_PER_SECTION; i++) {
 247                         if (PageReserved(first_page+i))
 248                                 continue;
 249
 250                         printk(KERN_WARNING "section number %ld page number %d "
 251                                 "not reserved, was it already online?\n",
 252                                 phys_index, i);
 253                         return -EBUSY;
 254                 }
 255         }
 256
 257         switch (action) {
 258                 case MEM_ONLINE:
 259                         start_pfn = page_to_pfn(first_page);
 260                         ret = online_pages(start_pfn, PAGES_PER_SECTION);
 261                         break;
 262                 case MEM_OFFLINE:
 263                         start_paddr = page_to_pfn(first_page) << PAGE_SHIFT;
 264                         ret = remove_memory(start_paddr,
 265                                             PAGES_PER_SECTION << PAGE_SHIFT);
 266                         break;
 267                 default:
 268                         WARN(1, KERN_WARNING "%s(%ld, %ld) unknown action: "
 269                              "%ld\n", __func__, phys_index, action, action);
 270                         ret = -EINVAL;
 271         }
 272
 273         return ret;
 274 }
 275
 276 static int memory_block_change_state(struct memory_block *mem,
 277                 unsigned long to_state, unsigned long from_state_req)
 278 {
 279         int i, ret = 0;
 280
 281         mutex_lock(&mem->state_mutex);
 282
 283         if (mem->state != from_state_req) {
 284                 ret = -EINVAL;
 285                 goto out;
 286         }
 287
 288         if (to_state == MEM_OFFLINE)
 289                 mem->state = MEM_GOING_OFFLINE;
 290
 291         for (i = 0; i < sections_per_block; i++) {
 292                 ret = memory_section_action(mem->start_section_nr + i,
 293                                             to_state);
 294                 if (ret)
 295                         break;
 296         }
 297
 298         if (ret) {
 299                 for (i = 0; i < sections_per_block; i++)
 300                         memory_section_action(mem->start_section_nr + i,
 301                                               from_state_req);
 302
 303                 mem->state = from_state_req;
 304         } else
 305                 mem->state = to_state;
 306
 307 out:
 308         mutex_unlock(&mem->state_mutex);
 309         return ret;
 310 }
 311
 312 static ssize_t
 313 store_mem_state(struct sys_device *dev,
 314                 struct sysdev_attribute *attr, const char *buf, size_t count)
 315 {
 316         struct memory_block *mem;
 317         int ret = -EINVAL;
 318
 319         mem = container_of(dev, struct memory_block, sysdev);
 320
 321         if (!strncmp(buf, "online", min((int)count, 6)))
 322                 ret = memory_block_change_state(mem, MEM_ONLINE, MEM_OFFLINE);
 323         else if(!strncmp(buf, "offline", min((int)count, 7)))
 324                 ret = memory_block_change_state(mem, MEM_OFFLINE, MEM_ONLINE);
 325
 326         if (ret)
 327                 return ret;
 328         return count;
 329 }
 330
 331 /*
 332  * phys_device is a bad name for this.  What I really want
 333  * is a way to differentiate between memory ranges that
 334  * are part of physical devices that constitute
 335  * a complete removable unit or fru.
 336  * i.e. do these ranges belong to the same physical device,
 337  * s.t. if I offline all of these sections I can then
 338  * remove the physical device?
 339  */
 340 static ssize_t show_phys_device(struct sys_device *dev,
 341                                 struct sysdev_attribute *attr, char *buf)
 342 {
 343         struct memory_block *mem =
 344                 container_of(dev, struct memory_block, sysdev);
 345         return sprintf(buf, "%d\n", mem->phys_device);
 346 }
 347
 348 static SYSDEV_ATTR(phys_index, 0444, show_mem_start_phys_index, NULL);
 349 static SYSDEV_ATTR(end_phys_index, 0444, show_mem_end_phys_index, NULL);
 350 static SYSDEV_ATTR(state, 0644, show_mem_state, store_mem_state);
 351 static SYSDEV_ATTR(phys_device, 0444, show_phys_device, NULL);
 352 static SYSDEV_ATTR(removable, 0444, show_mem_removable, NULL);
 353
 354 #define mem_create_simple_file(mem, attr_name)  \
 355         sysdev_create_file(&mem->sysdev, &attr_##attr_name)
 356 #define mem_remove_simple_file(mem, attr_name)  \
 357         sysdev_remove_file(&mem->sysdev, &attr_##attr_name)
 358
 359 /*
 360  * Block size attribute stuff
 361  */
 362 static ssize_t
 363 print_block_size(struct sysdev_class *class, struct sysdev_class_attribute *attr,
 364                  char *buf)
 365 {
 366         return sprintf(buf, "%lx\n", get_memory_block_size());
 367 }
 368
 369 static SYSDEV_CLASS_ATTR(block_size_bytes, 0444, print_block_size, NULL);
 370
 371 static int block_size_init(void)
 372 {
 373         return sysfs_create_file(&memory_sysdev_class.kset.kobj,
 374                                 &attr_block_size_bytes.attr);
 375 }
 376
 377 /*
 378  * Some architectures will have custom drivers to do this, and
 379  * will not need to do it from userspace.  The fake hot-add code
 380  * as well as ppc64 will do all of their discovery in userspace
 381  * and will require this interface.
 382  */
 383 #ifdef CONFIG_ARCH_MEMORY_PROBE
 384 static ssize_t
 385 memory_probe_store(struct class *class, struct class_attribute *attr,
 386                    const char *buf, size_t count)
 387 {
 388         u64 phys_addr;
 389         int nid;
 390         int i, ret;
 391
 392         phys_addr = simple_strtoull(buf, NULL, 0);
 393
 394         for (i = 0; i < sections_per_block; i++) {
 395                 nid = memory_add_physaddr_to_nid(phys_addr);
 396                 ret = add_memory(nid, phys_addr,
 397                                  PAGES_PER_SECTION << PAGE_SHIFT);
 398                 if (ret)
 399                         break;
 400
 401                 phys_addr += MIN_MEMORY_BLOCK_SIZE;
 402         }
 403
 404         if (ret)
 405                 count = ret;
 406
 407         return count;
 408 }
 409 static CLASS_ATTR(probe, S_IWUSR, NULL, memory_probe_store);
 410
 411 static int memory_probe_init(void)
 412 {
 413         return sysfs_create_file(&memory_sysdev_class.kset.kobj,
 414                                 &class_attr_probe.attr);
 415 }
 416 #else
 417 static inline int memory_probe_init(void)
 418 {
 419         return 0;
 420 }
 421 #endif
 422
 423 #ifdef CONFIG_MEMORY_FAILURE
 424 /*
 425  * Support for offlining pages of memory
 426  */
 427
 428 /* Soft offline a page */
 429 static ssize_t
 430 store_soft_offline_page(struct class *class,
 431                         struct class_attribute *attr,
 432                         const char *buf, size_t count)
 433 {
 434         int ret;
 435         u64 pfn;
 436         if (!capable(CAP_SYS_ADMIN))
 437                 return -EPERM;
 438         if (strict_strtoull(buf, 0, &pfn) < 0)
 439                 return -EINVAL;
 440         pfn >>= PAGE_SHIFT;
 441         if (!pfn_valid(pfn))
 442                 return -ENXIO;
 443         ret = soft_offline_page(pfn_to_page(pfn), 0);
 444         return ret == 0 ? count : ret;
 445 }
 446
 447 /* Forcibly offline a page, including killing processes. */
 448 static ssize_t
 449 store_hard_offline_page(struct class *class,
 450                         struct class_attribute *attr,
 451                         const char *buf, size_t count)
 452 {
 453         int ret;
 454         u64 pfn;
 455         if (!capable(CAP_SYS_ADMIN))
 456                 return -EPERM;
 457         if (strict_strtoull(buf, 0, &pfn) < 0)
 458                 return -EINVAL;
 459         pfn >>= PAGE_SHIFT;
 460         ret = __memory_failure(pfn, 0, 0);
 461         return ret ? ret : count;
 462 }
 463
 464 static CLASS_ATTR(soft_offline_page, 0644, NULL, store_soft_offline_page);
 465 static CLASS_ATTR(hard_offline_page, 0644, NULL, store_hard_offline_page);
 466
 467 static __init int memory_fail_init(void)
 468 {
 469         int err;
 470
 471         err = sysfs_create_file(&memory_sysdev_class.kset.kobj,
 472                                 &class_attr_soft_offline_page.attr);
 473         if (!err)
 474                 err = sysfs_create_file(&memory_sysdev_class.kset.kobj,
 475                                 &class_attr_hard_offline_page.attr);
 476         return err;
 477 }
 478 #else
 479 static inline int memory_fail_init(void)
 480 {
 481         return 0;
 482 }
 483 #endif
 484
 485 /*
 486  * Note that phys_device is optional.  It is here to allow for
 487  * differentiation between which *physical* devices each
 488  * section belongs to...
 489  */
 490 int __weak arch_get_memory_phys_device(unsigned long start_pfn)
 491 {
 492         return 0;
 493 }
 494
 495 struct memory_block *find_memory_block_hinted(struct mem_section *section,
 496                                               struct memory_block *hint)
 497 {
 498         struct kobject *kobj;
 499         struct sys_device *sysdev;
 500         struct memory_block *mem;
 501         char name[sizeof(MEMORY_CLASS_NAME) + 9 + 1];
 502         int block_id = base_memory_block_id(__section_nr(section));
 503
 504         kobj = hint ? &hint->sysdev.kobj : NULL;
 505
 506         /*
 507          * This only works because we know that section == sysdev->id
 508          * slightly redundant with sysdev_register()
 509          */
 510         sprintf(&name[0], "%s%d", MEMORY_CLASS_NAME, block_id);
 511
 512         kobj = kset_find_obj_hinted(&memory_sysdev_class.kset, name, kobj);
 513         if (!kobj)
 514                 return NULL;
 515
 516         sysdev = container_of(kobj, struct sys_device, kobj);
 517         mem = container_of(sysdev, struct memory_block, sysdev);
 518
 519         return mem;
 520 }
 521
 522 /*
 523  * For now, we have a linear search to go find the appropriate
 524  * memory_block corresponding to a particular phys_index. If
 525  * this gets to be a real problem, we can always use a radix
 526  * tree or something here.
 527  *
 528  * This could be made generic for all sysdev classes.
 529  */
 530 struct memory_block *find_memory_block(struct mem_section *section)
 531 {
 532         return find_memory_block_hinted(section, NULL);
 533 }
 534
 535 static int init_memory_block(struct memory_block **memory,
 536                              struct mem_section *section, unsigned long state)
 537 {
 538         struct memory_block *mem;
 539         unsigned long start_pfn;
 540         int scn_nr;
 541         int ret = 0;
 542
 543         mem = kzalloc(sizeof(*mem), GFP_KERNEL);
 544         if (!mem)
 545                 return -ENOMEM;
 546
 547         scn_nr = __section_nr(section);
 548         mem->start_section_nr =
 549                         base_memory_block_id(scn_nr) * sections_per_block;
 550         mem->end_section_nr = mem->start_section_nr + sections_per_block - 1;
 551         mem->state = state;
 552         mem->section_count++;
 553         mutex_init(&mem->state_mutex);
 554         start_pfn = section_nr_to_pfn(mem->start_section_nr);
 555         mem->phys_device = arch_get_memory_phys_device(start_pfn);
 556
 557         ret = register_memory(mem);
 558         if (!ret)
 559                 ret = mem_create_simple_file(mem, phys_index);
 560         if (!ret)
 561                 ret = mem_create_simple_file(mem, end_phys_index);
 562         if (!ret)
 563                 ret = mem_create_simple_file(mem, state);
 564         if (!ret)
 565                 ret = mem_create_simple_file(mem, phys_device);
 566         if (!ret)
 567                 ret = mem_create_simple_file(mem, removable);
 568
 569         *memory = mem;
 570         return ret;
 571 }
 572
 573 static int add_memory_section(int nid, struct mem_section *section,
 574                         unsigned long state, enum mem_add_context context)
 575 {
 576         struct memory_block *mem;
 577         int ret = 0;
 578
 579         mutex_lock(&mem_sysfs_mutex);
 580
 581         mem = find_memory_block(section);
 582         if (mem) {
 583                 mem->section_count++;
 584                 kobject_put(&mem->sysdev.kobj);
 585         } else
 586                 ret = init_memory_block(&mem, section, state);
 587
 588         if (!ret) {
 589                 if (context == HOTPLUG &&
 590                     mem->section_count == sections_per_block)
 591                         ret = register_mem_sect_under_node(mem, nid);
 592         }
 593
 594         mutex_unlock(&mem_sysfs_mutex);
 595         return ret;
 596 }
 597
 598 int remove_memory_block(unsigned long node_id, struct mem_section *section,
 599                 int phys_device)
 600 {
 601         struct memory_block *mem;
 602
 603         mutex_lock(&mem_sysfs_mutex);
 604         mem = find_memory_block(section);
 605         unregister_mem_sect_under_nodes(mem, __section_nr(section));
 606
 607         mem->section_count--;
 608         if (mem->section_count == 0) {
 609                 mem_remove_simple_file(mem, phys_index);
 610                 mem_remove_simple_file(mem, end_phys_index);
 611                 mem_remove_simple_file(mem, state);
 612                 mem_remove_simple_file(mem, phys_device);
 613                 mem_remove_simple_file(mem, removable);
 614                 unregister_memory(mem);
 615                 kfree(mem);
 616         } else
 617                 kobject_put(&mem->sysdev.kobj);
 618
 619         mutex_unlock(&mem_sysfs_mutex);
 620         return 0;
 621 }
 622
 623 /*
 624  * need an interface for the VM to add new memory regions,
 625  * but without onlining it.
 626  */
 627 int register_new_memory(int nid, struct mem_section *section)
 628 {
 629         return add_memory_section(nid, section, MEM_OFFLINE, HOTPLUG);
 630 }
 631
 632 int unregister_memory_section(struct mem_section *section)
 633 {
 634         if (!present_section(section))
 635                 return -EINVAL;
 636
 637         return remove_memory_block(0, section, 0);
 638 }
 639
 640 /*
 641  * Initialize the sysfs support for memory devices...
 642  */
 643 int __init memory_dev_init(void)
 644 {
 645         unsigned int i;
 646         int ret;
 647         int err;
 648         unsigned long block_sz;
 649
 650         memory_sysdev_class.kset.uevent_ops = &memory_uevent_ops;
 651         ret = sysdev_class_register(&memory_sysdev_class);
 652         if (ret)
 653                 goto out;
 654
 655         block_sz = get_memory_block_size();
 656         sections_per_block = block_sz / MIN_MEMORY_BLOCK_SIZE;
 657
 658         /*
 659          * Create entries for memory sections that were found
 660          * during boot and have been initialized
 661          */
 662         for (i = 0; i < NR_MEM_SECTIONS; i++) {
 663                 if (!present_section_nr(i))
 664                         continue;
 665                 err = add_memory_section(0, __nr_to_section(i), MEM_ONLINE,
 666                                          BOOT);
 667                 if (!ret)
 668                         ret = err;
 669         }
 670
 671         err = memory_probe_init();
 672         if (!ret)
 673                 ret = err;
 674         err = memory_fail_init();
 675         if (!ret)
 676                 ret = err;
 677         err = block_size_init();
 678         if (!ret)
 679                 ret = err;
 680 out:
 681         if (ret)
 682                 printk(KERN_ERR "%s() failed: %d\n", __func__, ret);
 683         return ret;
 684 }