drivers/base/memory.c

   1 /*
   2  * Memory subsystem 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/module.h>
  14 #include <linux/init.h>
  15 #include <linux/topology.h>
  16 #include <linux/capability.h>
  17 #include <linux/device.h>
  18 #include <linux/memory.h>
  19 #include <linux/memory_hotplug.h>
  20 #include <linux/mm.h>
  21 #include <linux/mutex.h>
  22 #include <linux/stat.h>
  23 #include <linux/slab.h>
  24
  25 #include <linux/atomic.h>
  26 #include <asm/uaccess.h>
  27
  28 static DEFINE_MUTEX(mem_sysfs_mutex);
  29
  30 #define MEMORY_CLASS_NAME       "memory"
  31
  32 #define to_memory_block(dev) container_of(dev, struct memory_block, dev)
  33
  34 static int sections_per_block;
  35
  36 static inline int base_memory_block_id(int section_nr)
  37 {
  38         return section_nr / sections_per_block;
  39 }
  40
  41 static int memory_subsys_online(struct device *dev);
  42 static int memory_subsys_offline(struct device *dev);
  43
  44 static struct bus_type memory_subsys = {
  45         .name = MEMORY_CLASS_NAME,
  46         .dev_name = MEMORY_CLASS_NAME,
  47         .online = memory_subsys_online,
  48         .offline = memory_subsys_offline,
  49 };
  50
  51 static BLOCKING_NOTIFIER_HEAD(memory_chain);
  52
  53 int register_memory_notifier(struct notifier_block *nb)
  54 {
  55         return blocking_notifier_chain_register(&memory_chain, nb);
  56 }
  57 EXPORT_SYMBOL(register_memory_notifier);
  58
  59 void unregister_memory_notifier(struct notifier_block *nb)
  60 {
  61         blocking_notifier_chain_unregister(&memory_chain, nb);
  62 }
  63 EXPORT_SYMBOL(unregister_memory_notifier);
  64
  65 static ATOMIC_NOTIFIER_HEAD(memory_isolate_chain);
  66
  67 int register_memory_isolate_notifier(struct notifier_block *nb)
  68 {
  69         return atomic_notifier_chain_register(&memory_isolate_chain, nb);
  70 }
  71 EXPORT_SYMBOL(register_memory_isolate_notifier);
  72
  73 void unregister_memory_isolate_notifier(struct notifier_block *nb)
  74 {
  75         atomic_notifier_chain_unregister(&memory_isolate_chain, nb);
  76 }
  77 EXPORT_SYMBOL(unregister_memory_isolate_notifier);
  78
  79 static void memory_block_release(struct device *dev)
  80 {
  81         struct memory_block *mem = to_memory_block(dev);
  82
  83         kfree(mem);
  84 }
  85
  86 unsigned long __weak memory_block_size_bytes(void)
  87 {
  88         return MIN_MEMORY_BLOCK_SIZE;
  89 }
  90
  91 static unsigned long get_memory_block_size(void)
  92 {
  93         unsigned long block_sz;
  94
  95         block_sz = memory_block_size_bytes();
  96
  97         /* Validate blk_sz is a power of 2 and not less than section size */
  98         if ((block_sz & (block_sz - 1)) || (block_sz < MIN_MEMORY_BLOCK_SIZE)) {
  99                 WARN_ON(1);
 100                 block_sz = MIN_MEMORY_BLOCK_SIZE;
 101         }
 102
 103         return block_sz;
 104 }
 105
 106 /*
 107  * use this as the physical section index that this memsection
 108  * uses.
 109  */
 110
 111 static ssize_t show_mem_start_phys_index(struct device *dev,
 112                         struct device_attribute *attr, char *buf)
 113 {
 114         struct memory_block *mem = to_memory_block(dev);
 115         unsigned long phys_index;
 116
 117         phys_index = mem->start_section_nr / sections_per_block;
 118         return sprintf(buf, "%08lx\n", phys_index);
 119 }
 120
 121 /*
 122  * Show whether the section of memory is likely to be hot-removable
 123  */
 124 static ssize_t show_mem_removable(struct device *dev,
 125                         struct device_attribute *attr, char *buf)
 126 {
 127         unsigned long i, pfn;
 128         int ret = 1;
 129         struct memory_block *mem = to_memory_block(dev);
 130
 131         for (i = 0; i < sections_per_block; i++) {
 132                 if (!present_section_nr(mem->start_section_nr + i))
 133                         continue;
 134                 pfn = section_nr_to_pfn(mem->start_section_nr + i);
 135                 ret &= is_mem_section_removable(pfn, PAGES_PER_SECTION);
 136         }
 137
 138         return sprintf(buf, "%d\n", ret);
 139 }
 140
 141 /*
 142  * online, offline, going offline, etc.
 143  */
 144 static ssize_t show_mem_state(struct device *dev,
 145                         struct device_attribute *attr, char *buf)
 146 {
 147         struct memory_block *mem = to_memory_block(dev);
 148         ssize_t len = 0;
 149
 150         /*
 151          * We can probably put these states in a nice little array
 152          * so that they're not open-coded
 153          */
 154         switch (mem->state) {
 155         case MEM_ONLINE:
 156                 len = sprintf(buf, "online\n");
 157                 break;
 158         case MEM_OFFLINE:
 159                 len = sprintf(buf, "offline\n");
 160                 break;
 161         case MEM_GOING_OFFLINE:
 162                 len = sprintf(buf, "going-offline\n");
 163                 break;
 164         default:
 165                 len = sprintf(buf, "ERROR-UNKNOWN-%ld\n",
 166                                 mem->state);
 167                 WARN_ON(1);
 168                 break;
 169         }
 170
 171         return len;
 172 }
 173
 174 int memory_notify(unsigned long val, void *v)
 175 {
 176         return blocking_notifier_call_chain(&memory_chain, val, v);
 177 }
 178
 179 int memory_isolate_notify(unsigned long val, void *v)
 180 {
 181         return atomic_notifier_call_chain(&memory_isolate_chain, val, v);
 182 }
 183
 184 /*
 185  * The probe routines leave the pages reserved, just as the bootmem code does.
 186  * Make sure they're still that way.
 187  */
 188 static bool pages_correctly_reserved(unsigned long start_pfn)
 189 {
 190         int i, j;
 191         struct page *page;
 192         unsigned long pfn = start_pfn;
 193
 194         /*
 195          * memmap between sections is not contiguous except with
 196          * SPARSEMEM_VMEMMAP. We lookup the page once per section
 197          * and assume memmap is contiguous within each section
 198          */
 199         for (i = 0; i < sections_per_block; i++, pfn += PAGES_PER_SECTION) {
 200                 if (WARN_ON_ONCE(!pfn_valid(pfn)))
 201                         return false;
 202                 page = pfn_to_page(pfn);
 203
 204                 for (j = 0; j < PAGES_PER_SECTION; j++) {
 205                         if (PageReserved(page + j))
 206                                 continue;
 207
 208                         printk(KERN_WARNING "section number %ld page number %d "
 209                                 "not reserved, was it already online?\n",
 210                                 pfn_to_section_nr(pfn), j);
 211
 212                         return false;
 213                 }
 214         }
 215
 216         return true;
 217 }
 218
 219 /*
 220  * MEMORY_HOTPLUG depends on SPARSEMEM in mm/Kconfig, so it is
 221  * OK to have direct references to sparsemem variables in here.
 222  * Must already be protected by mem_hotplug_begin().
 223  */
 224 static int
 225 memory_block_action(unsigned long phys_index, unsigned long action, int online_type)
 226 {
 227         unsigned long start_pfn;
 228         unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
 229         struct page *first_page;
 230         int ret;
 231
 232         start_pfn = section_nr_to_pfn(phys_index);
 233         first_page = pfn_to_page(start_pfn);
 234
 235         switch (action) {
 236         case MEM_ONLINE:
 237                 if (!pages_correctly_reserved(start_pfn))
 238                         return -EBUSY;
 239
 240                 ret = online_pages(start_pfn, nr_pages, online_type);
 241                 break;
 242         case MEM_OFFLINE:
 243                 ret = offline_pages(start_pfn, nr_pages);
 244                 break;
 245         default:
 246                 WARN(1, KERN_WARNING "%s(%ld, %ld) unknown action: "
 247                      "%ld\n", __func__, phys_index, action, action);
 248                 ret = -EINVAL;
 249         }
 250
 251         return ret;
 252 }
 253
 254 int memory_block_change_state(struct memory_block *mem,
 255                 unsigned long to_state, unsigned long from_state_req)
 256 {
 257         int ret = 0;
 258
 259         if (mem->state != from_state_req)
 260                 return -EINVAL;
 261
 262         if (to_state == MEM_OFFLINE)
 263                 mem->state = MEM_GOING_OFFLINE;
 264
 265         ret = memory_block_action(mem->start_section_nr, to_state,
 266                                 mem->online_type);
 267
 268         mem->state = ret ? from_state_req : to_state;
 269
 270         return ret;
 271 }
 272
 273 /* The device lock serializes operations on memory_subsys_[online|offline] */
 274 static int memory_subsys_online(struct device *dev)
 275 {
 276         struct memory_block *mem = to_memory_block(dev);
 277         int ret;
 278
 279         if (mem->state == MEM_ONLINE)
 280                 return 0;
 281
 282         /*
 283          * If we are called from store_mem_state(), online_type will be
 284          * set >= 0 Otherwise we were called from the device online
 285          * attribute and need to set the online_type.
 286          */
 287         if (mem->online_type < 0)
 288                 mem->online_type = MMOP_ONLINE_KEEP;
 289
 290         /* Already under protection of mem_hotplug_begin() */
 291         ret = memory_block_change_state(mem, MEM_ONLINE, MEM_OFFLINE);
 292
 293         /* clear online_type */
 294         mem->online_type = -1;
 295
 296         return ret;
 297 }
 298
 299 static int memory_subsys_offline(struct device *dev)
 300 {
 301         struct memory_block *mem = to_memory_block(dev);
 302
 303         if (mem->state == MEM_OFFLINE)
 304                 return 0;
 305
 306         /* Can't offline block with non-present sections */
 307         if (mem->section_count != sections_per_block)
 308                 return -EINVAL;
 309
 310         return memory_block_change_state(mem, MEM_OFFLINE, MEM_ONLINE);
 311 }
 312
 313 static ssize_t
 314 store_mem_state(struct device *dev,
 315                 struct device_attribute *attr, const char *buf, size_t count)
 316 {
 317         struct memory_block *mem = to_memory_block(dev);
 318         int ret, online_type;
 319
 320         ret = lock_device_hotplug_sysfs();
 321         if (ret)
 322                 return ret;
 323
 324         if (sysfs_streq(buf, "online_kernel"))
 325                 online_type = MMOP_ONLINE_KERNEL;
 326         else if (sysfs_streq(buf, "online_movable"))
 327                 online_type = MMOP_ONLINE_MOVABLE;
 328         else if (sysfs_streq(buf, "online"))
 329                 online_type = MMOP_ONLINE_KEEP;
 330         else if (sysfs_streq(buf, "offline"))
 331                 online_type = MMOP_OFFLINE;
 332         else {
 333                 ret = -EINVAL;
 334                 goto err;
 335         }
 336
 337         /*
 338          * Memory hotplug needs to hold mem_hotplug_begin() for probe to find
 339          * the correct memory block to online before doing device_online(dev),
 340          * which will take dev->mutex.  Take the lock early to prevent an
 341          * inversion, memory_subsys_online() callbacks will be implemented by
 342          * assuming it's already protected.
 343          */
 344         mem_hotplug_begin();
 345
 346         switch (online_type) {
 347         case MMOP_ONLINE_KERNEL:
 348         case MMOP_ONLINE_MOVABLE:
 349         case MMOP_ONLINE_KEEP:
 350                 mem->online_type = online_type;
 351                 ret = device_online(&mem->dev);
 352                 break;
 353         case MMOP_OFFLINE:
 354                 ret = device_offline(&mem->dev);
 355                 break;
 356         default:
 357                 ret = -EINVAL; /* should never happen */
 358         }
 359
 360         mem_hotplug_done();
 361 err:
 362         unlock_device_hotplug();
 363
 364         if (ret < 0)
 365                 return ret;
 366         if (ret)
 367                 return -EINVAL;
 368
 369         return count;
 370 }
 371
 372 /*
 373  * phys_device is a bad name for this.  What I really want
 374  * is a way to differentiate between memory ranges that
 375  * are part of physical devices that constitute
 376  * a complete removable unit or fru.
 377  * i.e. do these ranges belong to the same physical device,
 378  * s.t. if I offline all of these sections I can then
 379  * remove the physical device?
 380  */
 381 static ssize_t show_phys_device(struct device *dev,
 382                                 struct device_attribute *attr, char *buf)
 383 {
 384         struct memory_block *mem = to_memory_block(dev);
 385         return sprintf(buf, "%d\n", mem->phys_device);
 386 }
 387
 388 #ifdef CONFIG_MEMORY_HOTREMOVE
 389 static ssize_t show_valid_zones(struct device *dev,
 390                                 struct device_attribute *attr, char *buf)
 391 {
 392         struct memory_block *mem = to_memory_block(dev);
 393         unsigned long start_pfn, end_pfn;
 394         unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
 395         struct page *first_page;
 396         struct zone *zone;
 397         int zone_shift = 0;
 398
 399         start_pfn = section_nr_to_pfn(mem->start_section_nr);
 400         end_pfn = start_pfn + nr_pages;
 401         first_page = pfn_to_page(start_pfn);
 402
 403         /* The block contains more than one zone can not be offlined. */
 404         if (!test_pages_in_a_zone(start_pfn, end_pfn))
 405                 return sprintf(buf, "none\n");
 406
 407         zone = page_zone(first_page);
 408
 409         /* MMOP_ONLINE_KEEP */
 410         sprintf(buf, "%s", zone->name);
 411
 412         /* MMOP_ONLINE_KERNEL */
 413         zone_shift = zone_can_shift(start_pfn, nr_pages, ZONE_NORMAL);
 414         if (zone_shift) {
 415                 strcat(buf, " ");
 416                 strcat(buf, (zone + zone_shift)->name);
 417         }
 418
 419         /* MMOP_ONLINE_MOVABLE */
 420         zone_shift = zone_can_shift(start_pfn, nr_pages, ZONE_MOVABLE);
 421         if (zone_shift) {
 422                 strcat(buf, " ");
 423                 strcat(buf, (zone + zone_shift)->name);
 424         }
 425
 426         strcat(buf, "\n");
 427
 428         return strlen(buf);
 429 }
 430 static DEVICE_ATTR(valid_zones, 0444, show_valid_zones, NULL);
 431 #endif
 432
 433 static DEVICE_ATTR(phys_index, 0444, show_mem_start_phys_index, NULL);
 434 static DEVICE_ATTR(state, 0644, show_mem_state, store_mem_state);
 435 static DEVICE_ATTR(phys_device, 0444, show_phys_device, NULL);
 436 static DEVICE_ATTR(removable, 0444, show_mem_removable, NULL);
 437
 438 /*
 439  * Block size attribute stuff
 440  */
 441 static ssize_t
 442 print_block_size(struct device *dev, struct device_attribute *attr,
 443                  char *buf)
 444 {
 445         return sprintf(buf, "%lx\n", get_memory_block_size());
 446 }
 447
 448 static DEVICE_ATTR(block_size_bytes, 0444, print_block_size, NULL);
 449
 450 /*
 451  * Memory auto online policy.
 452  */
 453
 454 static ssize_t
 455 show_auto_online_blocks(struct device *dev, struct device_attribute *attr,
 456                         char *buf)
 457 {
 458         if (memhp_auto_online)
 459                 return sprintf(buf, "online\n");
 460         else
 461                 return sprintf(buf, "offline\n");
 462 }
 463
 464 static ssize_t
 465 store_auto_online_blocks(struct device *dev, struct device_attribute *attr,
 466                          const char *buf, size_t count)
 467 {
 468         if (sysfs_streq(buf, "online"))
 469                 memhp_auto_online = true;
 470         else if (sysfs_streq(buf, "offline"))
 471                 memhp_auto_online = false;
 472         else
 473                 return -EINVAL;
 474
 475         return count;
 476 }
 477
 478 static DEVICE_ATTR(auto_online_blocks, 0644, show_auto_online_blocks,
 479                    store_auto_online_blocks);
 480
 481 /*
 482  * Some architectures will have custom drivers to do this, and
 483  * will not need to do it from userspace.  The fake hot-add code
 484  * as well as ppc64 will do all of their discovery in userspace
 485  * and will require this interface.
 486  */
 487 #ifdef CONFIG_ARCH_MEMORY_PROBE
 488 static ssize_t
 489 memory_probe_store(struct device *dev, struct device_attribute *attr,
 490                    const char *buf, size_t count)
 491 {
 492         u64 phys_addr;
 493         int nid, ret;
 494         unsigned long pages_per_block = PAGES_PER_SECTION * sections_per_block;
 495
 496         ret = kstrtoull(buf, 0, &phys_addr);
 497         if (ret)
 498                 return ret;
 499
 500         if (phys_addr & ((pages_per_block << PAGE_SHIFT) - 1))
 501                 return -EINVAL;
 502
 503         nid = memory_add_physaddr_to_nid(phys_addr);
 504         ret = add_memory(nid, phys_addr,
 505                          MIN_MEMORY_BLOCK_SIZE * sections_per_block);
 506
 507         if (ret)
 508                 goto out;
 509
 510         ret = count;
 511 out:
 512         return ret;
 513 }
 514
 515 static DEVICE_ATTR(probe, S_IWUSR, NULL, memory_probe_store);
 516 #endif
 517
 518 #ifdef CONFIG_MEMORY_FAILURE
 519 /*
 520  * Support for offlining pages of memory
 521  */
 522
 523 /* Soft offline a page */
 524 static ssize_t
 525 store_soft_offline_page(struct device *dev,
 526                         struct device_attribute *attr,
 527                         const char *buf, size_t count)
 528 {
 529         int ret;
 530         u64 pfn;
 531         if (!capable(CAP_SYS_ADMIN))
 532                 return -EPERM;
 533         if (kstrtoull(buf, 0, &pfn) < 0)
 534                 return -EINVAL;
 535         pfn >>= PAGE_SHIFT;
 536         if (!pfn_valid(pfn))
 537                 return -ENXIO;
 538         ret = soft_offline_page(pfn_to_page(pfn), 0);
 539         return ret == 0 ? count : ret;
 540 }
 541
 542 /* Forcibly offline a page, including killing processes. */
 543 static ssize_t
 544 store_hard_offline_page(struct device *dev,
 545                         struct device_attribute *attr,
 546                         const char *buf, size_t count)
 547 {
 548         int ret;
 549         u64 pfn;
 550         if (!capable(CAP_SYS_ADMIN))
 551                 return -EPERM;
 552         if (kstrtoull(buf, 0, &pfn) < 0)
 553                 return -EINVAL;
 554         pfn >>= PAGE_SHIFT;
 555         ret = memory_failure(pfn, 0, 0);
 556         return ret ? ret : count;
 557 }
 558
 559 static DEVICE_ATTR(soft_offline_page, S_IWUSR, NULL, store_soft_offline_page);
 560 static DEVICE_ATTR(hard_offline_page, S_IWUSR, NULL, store_hard_offline_page);
 561 #endif
 562
 563 /*
 564  * Note that phys_device is optional.  It is here to allow for
 565  * differentiation between which *physical* devices each
 566  * section belongs to...
 567  */
 568 int __weak arch_get_memory_phys_device(unsigned long start_pfn)
 569 {
 570         return 0;
 571 }
 572
 573 /*
 574  * A reference for the returned object is held and the reference for the
 575  * hinted object is released.
 576  */
 577 struct memory_block *find_memory_block_hinted(struct mem_section *section,
 578                                               struct memory_block *hint)
 579 {
 580         int block_id = base_memory_block_id(__section_nr(section));
 581         struct device *hintdev = hint ? &hint->dev : NULL;
 582         struct device *dev;
 583
 584         dev = subsys_find_device_by_id(&memory_subsys, block_id, hintdev);
 585         if (hint)
 586                 put_device(&hint->dev);
 587         if (!dev)
 588                 return NULL;
 589         return to_memory_block(dev);
 590 }
 591
 592 /*
 593  * For now, we have a linear search to go find the appropriate
 594  * memory_block corresponding to a particular phys_index. If
 595  * this gets to be a real problem, we can always use a radix
 596  * tree or something here.
 597  *
 598  * This could be made generic for all device subsystems.
 599  */
 600 struct memory_block *find_memory_block(struct mem_section *section)
 601 {
 602         return find_memory_block_hinted(section, NULL);
 603 }
 604
 605 static struct attribute *memory_memblk_attrs[] = {
 606         &dev_attr_phys_index.attr,
 607         &dev_attr_state.attr,
 608         &dev_attr_phys_device.attr,
 609         &dev_attr_removable.attr,
 610 #ifdef CONFIG_MEMORY_HOTREMOVE
 611         &dev_attr_valid_zones.attr,
 612 #endif
 613         NULL
 614 };
 615
 616 static struct attribute_group memory_memblk_attr_group = {
 617         .attrs = memory_memblk_attrs,
 618 };
 619
 620 static const struct attribute_group *memory_memblk_attr_groups[] = {
 621         &memory_memblk_attr_group,
 622         NULL,
 623 };
 624
 625 /*
 626  * register_memory - Setup a sysfs device for a memory block
 627  */
 628 static
 629 int register_memory(struct memory_block *memory)
 630 {
 631         memory->dev.bus = &memory_subsys;
 632         memory->dev.id = memory->start_section_nr / sections_per_block;
 633         memory->dev.release = memory_block_release;
 634         memory->dev.groups = memory_memblk_attr_groups;
 635         memory->dev.offline = memory->state == MEM_OFFLINE;
 636
 637         return device_register(&memory->dev);
 638 }
 639
 640 static int init_memory_block(struct memory_block **memory,
 641                              struct mem_section *section, unsigned long state)
 642 {
 643         struct memory_block *mem;
 644         unsigned long start_pfn;
 645         int scn_nr;
 646         int ret = 0;
 647
 648         mem = kzalloc(sizeof(*mem), GFP_KERNEL);
 649         if (!mem)
 650                 return -ENOMEM;
 651
 652         scn_nr = __section_nr(section);
 653         mem->start_section_nr =
 654                         base_memory_block_id(scn_nr) * sections_per_block;
 655         mem->end_section_nr = mem->start_section_nr + sections_per_block - 1;
 656         mem->state = state;
 657         start_pfn = section_nr_to_pfn(mem->start_section_nr);
 658         mem->phys_device = arch_get_memory_phys_device(start_pfn);
 659
 660         ret = register_memory(mem);
 661
 662         *memory = mem;
 663         return ret;
 664 }
 665
 666 static int add_memory_block(int base_section_nr)
 667 {
 668         struct memory_block *mem;
 669         int i, ret, section_count = 0, section_nr;
 670
 671         for (i = base_section_nr;
 672              (i < base_section_nr + sections_per_block) && i < NR_MEM_SECTIONS;
 673              i++) {
 674                 if (!present_section_nr(i))
 675                         continue;
 676                 if (section_count == 0)
 677                         section_nr = i;
 678                 section_count++;
 679         }
 680
 681         if (section_count == 0)
 682                 return 0;
 683         ret = init_memory_block(&mem, __nr_to_section(section_nr), MEM_ONLINE);
 684         if (ret)
 685                 return ret;
 686         mem->section_count = section_count;
 687         return 0;
 688 }
 689
 690 static bool is_zone_device_section(struct mem_section *ms)
 691 {
 692         struct page *page;
 693
 694         page = sparse_decode_mem_map(ms->section_mem_map, __section_nr(ms));
 695         return is_zone_device_page(page);
 696 }
 697
 698 /*
 699  * need an interface for the VM to add new memory regions,
 700  * but without onlining it.
 701  */
 702 int register_new_memory(int nid, struct mem_section *section)
 703 {
 704         int ret = 0;
 705         struct memory_block *mem;
 706
 707         if (is_zone_device_section(section))
 708                 return 0;
 709
 710         mutex_lock(&mem_sysfs_mutex);
 711
 712         mem = find_memory_block(section);
 713         if (mem) {
 714                 mem->section_count++;
 715                 put_device(&mem->dev);
 716         } else {
 717                 ret = init_memory_block(&mem, section, MEM_OFFLINE);
 718                 if (ret)
 719                         goto out;
 720                 mem->section_count++;
 721         }
 722
 723         if (mem->section_count == sections_per_block)
 724                 ret = register_mem_sect_under_node(mem, nid);
 725 out:
 726         mutex_unlock(&mem_sysfs_mutex);
 727         return ret;
 728 }
 729
 730 #ifdef CONFIG_MEMORY_HOTREMOVE
 731 static void
 732 unregister_memory(struct memory_block *memory)
 733 {
 734         BUG_ON(memory->dev.bus != &memory_subsys);
 735
 736         /* drop the ref. we got in remove_memory_block() */
 737         put_device(&memory->dev);
 738         device_unregister(&memory->dev);
 739 }
 740
 741 static int remove_memory_section(unsigned long node_id,
 742                                struct mem_section *section, int phys_device)
 743 {
 744         struct memory_block *mem;
 745
 746         if (is_zone_device_section(section))
 747                 return 0;
 748
 749         mutex_lock(&mem_sysfs_mutex);
 750         mem = find_memory_block(section);
 751         unregister_mem_sect_under_nodes(mem, __section_nr(section));
 752
 753         mem->section_count--;
 754         if (mem->section_count == 0)
 755                 unregister_memory(mem);
 756         else
 757                 put_device(&mem->dev);
 758
 759         mutex_unlock(&mem_sysfs_mutex);
 760         return 0;
 761 }
 762
 763 int unregister_memory_section(struct mem_section *section)
 764 {
 765         if (!present_section(section))
 766                 return -EINVAL;
 767
 768         return remove_memory_section(0, section, 0);
 769 }
 770 #endif /* CONFIG_MEMORY_HOTREMOVE */
 771
 772 /* return true if the memory block is offlined, otherwise, return false */
 773 bool is_memblock_offlined(struct memory_block *mem)
 774 {
 775         return mem->state == MEM_OFFLINE;
 776 }
 777
 778 static struct attribute *memory_root_attrs[] = {
 779 #ifdef CONFIG_ARCH_MEMORY_PROBE
 780         &dev_attr_probe.attr,
 781 #endif
 782
 783 #ifdef CONFIG_MEMORY_FAILURE
 784         &dev_attr_soft_offline_page.attr,
 785         &dev_attr_hard_offline_page.attr,
 786 #endif
 787
 788         &dev_attr_block_size_bytes.attr,
 789         &dev_attr_auto_online_blocks.attr,
 790         NULL
 791 };
 792
 793 static struct attribute_group memory_root_attr_group = {
 794         .attrs = memory_root_attrs,
 795 };
 796
 797 static const struct attribute_group *memory_root_attr_groups[] = {
 798         &memory_root_attr_group,
 799         NULL,
 800 };
 801
 802 /*
 803  * Initialize the sysfs support for memory devices...
 804  */
 805 int __init memory_dev_init(void)
 806 {
 807         unsigned int i;
 808         int ret;
 809         int err;
 810         unsigned long block_sz;
 811
 812         ret = subsys_system_register(&memory_subsys, memory_root_attr_groups);
 813         if (ret)
 814                 goto out;
 815
 816         block_sz = get_memory_block_size();
 817         sections_per_block = block_sz / MIN_MEMORY_BLOCK_SIZE;
 818
 819         /*
 820          * Create entries for memory sections that were found
 821          * during boot and have been initialized
 822          */
 823         mutex_lock(&mem_sysfs_mutex);
 824         for (i = 0; i < NR_MEM_SECTIONS; i += sections_per_block) {
 825                 err = add_memory_block(i);
 826                 if (!ret)
 827                         ret = err;
 828         }
 829         mutex_unlock(&mem_sysfs_mutex);
 830
 831 out:
 832         if (ret)
 833                 printk(KERN_ERR "%s() failed: %d\n", __func__, ret);
 834         return ret;
 835 }