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 <linux/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         int ret;
 230
 231         start_pfn = section_nr_to_pfn(phys_index);
 232
 233         switch (action) {
 234         case MEM_ONLINE:
 235                 if (!pages_correctly_reserved(start_pfn))
 236                         return -EBUSY;
 237
 238                 ret = online_pages(start_pfn, nr_pages, online_type);
 239                 break;
 240         case MEM_OFFLINE:
 241                 ret = offline_pages(start_pfn, nr_pages);
 242                 break;
 243         default:
 244                 WARN(1, KERN_WARNING "%s(%ld, %ld) unknown action: "
 245                      "%ld\n", __func__, phys_index, action, action);
 246                 ret = -EINVAL;
 247         }
 248
 249         return ret;
 250 }
 251
 252 static int memory_block_change_state(struct memory_block *mem,
 253                 unsigned long to_state, unsigned long from_state_req)
 254 {
 255         int ret = 0;
 256
 257         if (mem->state != from_state_req)
 258                 return -EINVAL;
 259
 260         if (to_state == MEM_OFFLINE)
 261                 mem->state = MEM_GOING_OFFLINE;
 262
 263         ret = memory_block_action(mem->start_section_nr, to_state,
 264                                 mem->online_type);
 265
 266         mem->state = ret ? from_state_req : to_state;
 267
 268         return ret;
 269 }
 270
 271 /* The device lock serializes operations on memory_subsys_[online|offline] */
 272 static int memory_subsys_online(struct device *dev)
 273 {
 274         struct memory_block *mem = to_memory_block(dev);
 275         int ret;
 276
 277         if (mem->state == MEM_ONLINE)
 278                 return 0;
 279
 280         /*
 281          * If we are called from store_mem_state(), online_type will be
 282          * set >= 0 Otherwise we were called from the device online
 283          * attribute and need to set the online_type.
 284          */
 285         if (mem->online_type < 0)
 286                 mem->online_type = MMOP_ONLINE_KEEP;
 287
 288         /* Already under protection of mem_hotplug_begin() */
 289         ret = memory_block_change_state(mem, MEM_ONLINE, MEM_OFFLINE);
 290
 291         /* clear online_type */
 292         mem->online_type = -1;
 293
 294         return ret;
 295 }
 296
 297 static int memory_subsys_offline(struct device *dev)
 298 {
 299         struct memory_block *mem = to_memory_block(dev);
 300
 301         if (mem->state == MEM_OFFLINE)
 302                 return 0;
 303
 304         /* Can't offline block with non-present sections */
 305         if (mem->section_count != sections_per_block)
 306                 return -EINVAL;
 307
 308         return memory_block_change_state(mem, MEM_OFFLINE, MEM_ONLINE);
 309 }
 310
 311 static ssize_t
 312 store_mem_state(struct device *dev,
 313                 struct device_attribute *attr, const char *buf, size_t count)
 314 {
 315         struct memory_block *mem = to_memory_block(dev);
 316         int ret, online_type;
 317
 318         ret = lock_device_hotplug_sysfs();
 319         if (ret)
 320                 return ret;
 321
 322         if (sysfs_streq(buf, "online_kernel"))
 323                 online_type = MMOP_ONLINE_KERNEL;
 324         else if (sysfs_streq(buf, "online_movable"))
 325                 online_type = MMOP_ONLINE_MOVABLE;
 326         else if (sysfs_streq(buf, "online"))
 327                 online_type = MMOP_ONLINE_KEEP;
 328         else if (sysfs_streq(buf, "offline"))
 329                 online_type = MMOP_OFFLINE;
 330         else {
 331                 ret = -EINVAL;
 332                 goto err;
 333         }
 334
 335         /*
 336          * Memory hotplug needs to hold mem_hotplug_begin() for probe to find
 337          * the correct memory block to online before doing device_online(dev),
 338          * which will take dev->mutex.  Take the lock early to prevent an
 339          * inversion, memory_subsys_online() callbacks will be implemented by
 340          * assuming it's already protected.
 341          */
 342         mem_hotplug_begin();
 343
 344         switch (online_type) {
 345         case MMOP_ONLINE_KERNEL:
 346         case MMOP_ONLINE_MOVABLE:
 347         case MMOP_ONLINE_KEEP:
 348                 mem->online_type = online_type;
 349                 ret = device_online(&mem->dev);
 350                 break;
 351         case MMOP_OFFLINE:
 352                 ret = device_offline(&mem->dev);
 353                 break;
 354         default:
 355                 ret = -EINVAL; /* should never happen */
 356         }
 357
 358         mem_hotplug_done();
 359 err:
 360         unlock_device_hotplug();
 361
 362         if (ret < 0)
 363                 return ret;
 364         if (ret)
 365                 return -EINVAL;
 366
 367         return count;
 368 }
 369
 370 /*
 371  * phys_device is a bad name for this.  What I really want
 372  * is a way to differentiate between memory ranges that
 373  * are part of physical devices that constitute
 374  * a complete removable unit or fru.
 375  * i.e. do these ranges belong to the same physical device,
 376  * s.t. if I offline all of these sections I can then
 377  * remove the physical device?
 378  */
 379 static ssize_t show_phys_device(struct device *dev,
 380                                 struct device_attribute *attr, char *buf)
 381 {
 382         struct memory_block *mem = to_memory_block(dev);
 383         return sprintf(buf, "%d\n", mem->phys_device);
 384 }
 385
 386 #ifdef CONFIG_MEMORY_HOTREMOVE
 387 static ssize_t show_valid_zones(struct device *dev,
 388                                 struct device_attribute *attr, char *buf)
 389 {
 390         struct memory_block *mem = to_memory_block(dev);
 391         unsigned long start_pfn, end_pfn;
 392         unsigned long valid_start, valid_end, valid_pages;
 393         unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
 394         struct zone *zone;
 395         int zone_shift = 0;
 396
 397         start_pfn = section_nr_to_pfn(mem->start_section_nr);
 398         end_pfn = start_pfn + nr_pages;
 399
 400         /* The block contains more than one zone can not be offlined. */
 401         if (!test_pages_in_a_zone(start_pfn, end_pfn, &valid_start, &valid_end))
 402                 return sprintf(buf, "none\n");
 403
 404         zone = page_zone(pfn_to_page(valid_start));
 405         valid_pages = valid_end - valid_start;
 406
 407         /* MMOP_ONLINE_KEEP */
 408         sprintf(buf, "%s", zone->name);
 409
 410         /* MMOP_ONLINE_KERNEL */
 411         zone_can_shift(valid_start, valid_pages, ZONE_NORMAL, &zone_shift);
 412         if (zone_shift) {
 413                 strcat(buf, " ");
 414                 strcat(buf, (zone + zone_shift)->name);
 415         }
 416
 417         /* MMOP_ONLINE_MOVABLE */
 418         zone_can_shift(valid_start, valid_pages, ZONE_MOVABLE, &zone_shift);
 419         if (zone_shift) {
 420                 strcat(buf, " ");
 421                 strcat(buf, (zone + zone_shift)->name);
 422         }
 423
 424         strcat(buf, "\n");
 425
 426         return strlen(buf);
 427 }
 428 static DEVICE_ATTR(valid_zones, 0444, show_valid_zones, NULL);
 429 #endif
 430
 431 static DEVICE_ATTR(phys_index, 0444, show_mem_start_phys_index, NULL);
 432 static DEVICE_ATTR(state, 0644, show_mem_state, store_mem_state);
 433 static DEVICE_ATTR(phys_device, 0444, show_phys_device, NULL);
 434 static DEVICE_ATTR(removable, 0444, show_mem_removable, NULL);
 435
 436 /*
 437  * Block size attribute stuff
 438  */
 439 static ssize_t
 440 print_block_size(struct device *dev, struct device_attribute *attr,
 441                  char *buf)
 442 {
 443         return sprintf(buf, "%lx\n", get_memory_block_size());
 444 }
 445
 446 static DEVICE_ATTR(block_size_bytes, 0444, print_block_size, NULL);
 447
 448 /*
 449  * Memory auto online policy.
 450  */
 451
 452 static ssize_t
 453 show_auto_online_blocks(struct device *dev, struct device_attribute *attr,
 454                         char *buf)
 455 {
 456         if (memhp_auto_online)
 457                 return sprintf(buf, "online\n");
 458         else
 459                 return sprintf(buf, "offline\n");
 460 }
 461
 462 static ssize_t
 463 store_auto_online_blocks(struct device *dev, struct device_attribute *attr,
 464                          const char *buf, size_t count)
 465 {
 466         if (sysfs_streq(buf, "online"))
 467                 memhp_auto_online = true;
 468         else if (sysfs_streq(buf, "offline"))
 469                 memhp_auto_online = false;
 470         else
 471                 return -EINVAL;
 472
 473         return count;
 474 }
 475
 476 static DEVICE_ATTR(auto_online_blocks, 0644, show_auto_online_blocks,
 477                    store_auto_online_blocks);
 478
 479 /*
 480  * Some architectures will have custom drivers to do this, and
 481  * will not need to do it from userspace.  The fake hot-add code
 482  * as well as ppc64 will do all of their discovery in userspace
 483  * and will require this interface.
 484  */
 485 #ifdef CONFIG_ARCH_MEMORY_PROBE
 486 static ssize_t
 487 memory_probe_store(struct device *dev, struct device_attribute *attr,
 488                    const char *buf, size_t count)
 489 {
 490         u64 phys_addr;
 491         int nid, ret;
 492         unsigned long pages_per_block = PAGES_PER_SECTION * sections_per_block;
 493
 494         ret = kstrtoull(buf, 0, &phys_addr);
 495         if (ret)
 496                 return ret;
 497
 498         if (phys_addr & ((pages_per_block << PAGE_SHIFT) - 1))
 499                 return -EINVAL;
 500
 501         nid = memory_add_physaddr_to_nid(phys_addr);
 502         ret = add_memory(nid, phys_addr,
 503                          MIN_MEMORY_BLOCK_SIZE * sections_per_block);
 504
 505         if (ret)
 506                 goto out;
 507
 508         ret = count;
 509 out:
 510         return ret;
 511 }
 512
 513 static DEVICE_ATTR(probe, S_IWUSR, NULL, memory_probe_store);
 514 #endif
 515
 516 #ifdef CONFIG_MEMORY_FAILURE
 517 /*
 518  * Support for offlining pages of memory
 519  */
 520
 521 /* Soft offline a page */
 522 static ssize_t
 523 store_soft_offline_page(struct device *dev,
 524                         struct device_attribute *attr,
 525                         const char *buf, size_t count)
 526 {
 527         int ret;
 528         u64 pfn;
 529         if (!capable(CAP_SYS_ADMIN))
 530                 return -EPERM;
 531         if (kstrtoull(buf, 0, &pfn) < 0)
 532                 return -EINVAL;
 533         pfn >>= PAGE_SHIFT;
 534         if (!pfn_valid(pfn))
 535                 return -ENXIO;
 536         ret = soft_offline_page(pfn_to_page(pfn), 0);
 537         return ret == 0 ? count : ret;
 538 }
 539
 540 /* Forcibly offline a page, including killing processes. */
 541 static ssize_t
 542 store_hard_offline_page(struct device *dev,
 543                         struct device_attribute *attr,
 544                         const char *buf, size_t count)
 545 {
 546         int ret;
 547         u64 pfn;
 548         if (!capable(CAP_SYS_ADMIN))
 549                 return -EPERM;
 550         if (kstrtoull(buf, 0, &pfn) < 0)
 551                 return -EINVAL;
 552         pfn >>= PAGE_SHIFT;
 553         ret = memory_failure(pfn, 0, 0);
 554         return ret ? ret : count;
 555 }
 556
 557 static DEVICE_ATTR(soft_offline_page, S_IWUSR, NULL, store_soft_offline_page);
 558 static DEVICE_ATTR(hard_offline_page, S_IWUSR, NULL, store_hard_offline_page);
 559 #endif
 560
 561 /*
 562  * Note that phys_device is optional.  It is here to allow for
 563  * differentiation between which *physical* devices each
 564  * section belongs to...
 565  */
 566 int __weak arch_get_memory_phys_device(unsigned long start_pfn)
 567 {
 568         return 0;
 569 }
 570
 571 /*
 572  * A reference for the returned object is held and the reference for the
 573  * hinted object is released.
 574  */
 575 struct memory_block *find_memory_block_hinted(struct mem_section *section,
 576                                               struct memory_block *hint)
 577 {
 578         int block_id = base_memory_block_id(__section_nr(section));
 579         struct device *hintdev = hint ? &hint->dev : NULL;
 580         struct device *dev;
 581
 582         dev = subsys_find_device_by_id(&memory_subsys, block_id, hintdev);
 583         if (hint)
 584                 put_device(&hint->dev);
 585         if (!dev)
 586                 return NULL;
 587         return to_memory_block(dev);
 588 }
 589
 590 /*
 591  * For now, we have a linear search to go find the appropriate
 592  * memory_block corresponding to a particular phys_index. If
 593  * this gets to be a real problem, we can always use a radix
 594  * tree or something here.
 595  *
 596  * This could be made generic for all device subsystems.
 597  */
 598 struct memory_block *find_memory_block(struct mem_section *section)
 599 {
 600         return find_memory_block_hinted(section, NULL);
 601 }
 602
 603 static struct attribute *memory_memblk_attrs[] = {
 604         &dev_attr_phys_index.attr,
 605         &dev_attr_state.attr,
 606         &dev_attr_phys_device.attr,
 607         &dev_attr_removable.attr,
 608 #ifdef CONFIG_MEMORY_HOTREMOVE
 609         &dev_attr_valid_zones.attr,
 610 #endif
 611         NULL
 612 };
 613
 614 static struct attribute_group memory_memblk_attr_group = {
 615         .attrs = memory_memblk_attrs,
 616 };
 617
 618 static const struct attribute_group *memory_memblk_attr_groups[] = {
 619         &memory_memblk_attr_group,
 620         NULL,
 621 };
 622
 623 /*
 624  * register_memory - Setup a sysfs device for a memory block
 625  */
 626 static
 627 int register_memory(struct memory_block *memory)
 628 {
 629         memory->dev.bus = &memory_subsys;
 630         memory->dev.id = memory->start_section_nr / sections_per_block;
 631         memory->dev.release = memory_block_release;
 632         memory->dev.groups = memory_memblk_attr_groups;
 633         memory->dev.offline = memory->state == MEM_OFFLINE;
 634
 635         return device_register(&memory->dev);
 636 }
 637
 638 static int init_memory_block(struct memory_block **memory,
 639                              struct mem_section *section, unsigned long state)
 640 {
 641         struct memory_block *mem;
 642         unsigned long start_pfn;
 643         int scn_nr;
 644         int ret = 0;
 645
 646         mem = kzalloc(sizeof(*mem), GFP_KERNEL);
 647         if (!mem)
 648                 return -ENOMEM;
 649
 650         scn_nr = __section_nr(section);
 651         mem->start_section_nr =
 652                         base_memory_block_id(scn_nr) * sections_per_block;
 653         mem->end_section_nr = mem->start_section_nr + sections_per_block - 1;
 654         mem->state = state;
 655         start_pfn = section_nr_to_pfn(mem->start_section_nr);
 656         mem->phys_device = arch_get_memory_phys_device(start_pfn);
 657
 658         ret = register_memory(mem);
 659
 660         *memory = mem;
 661         return ret;
 662 }
 663
 664 static int add_memory_block(int base_section_nr)
 665 {
 666         struct memory_block *mem;
 667         int i, ret, section_count = 0, section_nr;
 668
 669         for (i = base_section_nr;
 670              (i < base_section_nr + sections_per_block) && i < NR_MEM_SECTIONS;
 671              i++) {
 672                 if (!present_section_nr(i))
 673                         continue;
 674                 if (section_count == 0)
 675                         section_nr = i;
 676                 section_count++;
 677         }
 678
 679         if (section_count == 0)
 680                 return 0;
 681         ret = init_memory_block(&mem, __nr_to_section(section_nr), MEM_ONLINE);
 682         if (ret)
 683                 return ret;
 684         mem->section_count = section_count;
 685         return 0;
 686 }
 687
 688 static bool is_zone_device_section(struct mem_section *ms)
 689 {
 690         struct page *page;
 691
 692         page = sparse_decode_mem_map(ms->section_mem_map, __section_nr(ms));
 693         return is_zone_device_page(page);
 694 }
 695
 696 /*
 697  * need an interface for the VM to add new memory regions,
 698  * but without onlining it.
 699  */
 700 int register_new_memory(int nid, struct mem_section *section)
 701 {
 702         int ret = 0;
 703         struct memory_block *mem;
 704
 705         if (is_zone_device_section(section))
 706                 return 0;
 707
 708         mutex_lock(&mem_sysfs_mutex);
 709
 710         mem = find_memory_block(section);
 711         if (mem) {
 712                 mem->section_count++;
 713                 put_device(&mem->dev);
 714         } else {
 715                 ret = init_memory_block(&mem, section, MEM_OFFLINE);
 716                 if (ret)
 717                         goto out;
 718                 mem->section_count++;
 719         }
 720
 721         if (mem->section_count == sections_per_block)
 722                 ret = register_mem_sect_under_node(mem, nid);
 723 out:
 724         mutex_unlock(&mem_sysfs_mutex);
 725         return ret;
 726 }
 727
 728 #ifdef CONFIG_MEMORY_HOTREMOVE
 729 static void
 730 unregister_memory(struct memory_block *memory)
 731 {
 732         BUG_ON(memory->dev.bus != &memory_subsys);
 733
 734         /* drop the ref. we got in remove_memory_block() */
 735         put_device(&memory->dev);
 736         device_unregister(&memory->dev);
 737 }
 738
 739 static int remove_memory_section(unsigned long node_id,
 740                                struct mem_section *section, int phys_device)
 741 {
 742         struct memory_block *mem;
 743
 744         if (is_zone_device_section(section))
 745                 return 0;
 746
 747         mutex_lock(&mem_sysfs_mutex);
 748         mem = find_memory_block(section);
 749         unregister_mem_sect_under_nodes(mem, __section_nr(section));
 750
 751         mem->section_count--;
 752         if (mem->section_count == 0)
 753                 unregister_memory(mem);
 754         else
 755                 put_device(&mem->dev);
 756
 757         mutex_unlock(&mem_sysfs_mutex);
 758         return 0;
 759 }
 760
 761 int unregister_memory_section(struct mem_section *section)
 762 {
 763         if (!present_section(section))
 764                 return -EINVAL;
 765
 766         return remove_memory_section(0, section, 0);
 767 }
 768 #endif /* CONFIG_MEMORY_HOTREMOVE */
 769
 770 /* return true if the memory block is offlined, otherwise, return false */
 771 bool is_memblock_offlined(struct memory_block *mem)
 772 {
 773         return mem->state == MEM_OFFLINE;
 774 }
 775
 776 static struct attribute *memory_root_attrs[] = {
 777 #ifdef CONFIG_ARCH_MEMORY_PROBE
 778         &dev_attr_probe.attr,
 779 #endif
 780
 781 #ifdef CONFIG_MEMORY_FAILURE
 782         &dev_attr_soft_offline_page.attr,
 783         &dev_attr_hard_offline_page.attr,
 784 #endif
 785
 786         &dev_attr_block_size_bytes.attr,
 787         &dev_attr_auto_online_blocks.attr,
 788         NULL
 789 };
 790
 791 static struct attribute_group memory_root_attr_group = {
 792         .attrs = memory_root_attrs,
 793 };
 794
 795 static const struct attribute_group *memory_root_attr_groups[] = {
 796         &memory_root_attr_group,
 797         NULL,
 798 };
 799
 800 /*
 801  * Initialize the sysfs support for memory devices...
 802  */
 803 int __init memory_dev_init(void)
 804 {
 805         unsigned int i;
 806         int ret;
 807         int err;
 808         unsigned long block_sz;
 809
 810         ret = subsys_system_register(&memory_subsys, memory_root_attr_groups);
 811         if (ret)
 812                 goto out;
 813
 814         block_sz = get_memory_block_size();
 815         sections_per_block = block_sz / MIN_MEMORY_BLOCK_SIZE;
 816
 817         /*
 818          * Create entries for memory sections that were found
 819          * during boot and have been initialized
 820          */
 821         mutex_lock(&mem_sysfs_mutex);
 822         for (i = 0; i < NR_MEM_SECTIONS; i += sections_per_block) {
 823                 err = add_memory_block(i);
 824                 if (!ret)
 825                         ret = err;
 826         }
 827         mutex_unlock(&mem_sysfs_mutex);
 828
 829 out:
 830         if (ret)
 831                 printk(KERN_ERR "%s() failed: %d\n", __func__, ret);
 832         return ret;
 833 }