drivers/base/memory.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Memory subsystem support
   4  *
   5  * Written by Matt Tolentino <matthew.e.tolentino@intel.com>
   6  *            Dave Hansen <haveblue@us.ibm.com>
   7  *
   8  * This file provides the necessary infrastructure to represent
   9  * a SPARSEMEM-memory-model system's physical memory in /sysfs.
  10  * All arch-independent code that assumes MEMORY_HOTPLUG requires
  11  * SPARSEMEM should be contained here, or in mm/memory_hotplug.c.
  12  */
  13
  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/memory_hotplug.h>
  21 #include <linux/mm.h>
  22 #include <linux/mutex.h>
  23 #include <linux/stat.h>
  24 #include <linux/slab.h>
  25
  26 #include <linux/atomic.h>
  27 #include <linux/uaccess.h>
  28
  29 static DEFINE_MUTEX(mem_sysfs_mutex);
  30
  31 #define MEMORY_CLASS_NAME       "memory"
  32
  33 #define to_memory_block(dev) container_of(dev, struct memory_block, dev)
  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 int memory_subsys_online(struct device *dev);
  43 static int memory_subsys_offline(struct device *dev);
  44
  45 static struct bus_type memory_subsys = {
  46         .name = MEMORY_CLASS_NAME,
  47         .dev_name = MEMORY_CLASS_NAME,
  48         .online = memory_subsys_online,
  49         .offline = memory_subsys_offline,
  50 };
  51
  52 static BLOCKING_NOTIFIER_HEAD(memory_chain);
  53
  54 int register_memory_notifier(struct notifier_block *nb)
  55 {
  56         return blocking_notifier_chain_register(&memory_chain, nb);
  57 }
  58 EXPORT_SYMBOL(register_memory_notifier);
  59
  60 void unregister_memory_notifier(struct notifier_block *nb)
  61 {
  62         blocking_notifier_chain_unregister(&memory_chain, nb);
  63 }
  64 EXPORT_SYMBOL(unregister_memory_notifier);
  65
  66 static ATOMIC_NOTIFIER_HEAD(memory_isolate_chain);
  67
  68 int register_memory_isolate_notifier(struct notifier_block *nb)
  69 {
  70         return atomic_notifier_chain_register(&memory_isolate_chain, nb);
  71 }
  72 EXPORT_SYMBOL(register_memory_isolate_notifier);
  73
  74 void unregister_memory_isolate_notifier(struct notifier_block *nb)
  75 {
  76         atomic_notifier_chain_unregister(&memory_isolate_chain, nb);
  77 }
  78 EXPORT_SYMBOL(unregister_memory_isolate_notifier);
  79
  80 static void memory_block_release(struct device *dev)
  81 {
  82         struct memory_block *mem = to_memory_block(dev);
  83
  84         kfree(mem);
  85 }
  86
  87 unsigned long __weak memory_block_size_bytes(void)
  88 {
  89         return MIN_MEMORY_BLOCK_SIZE;
  90 }
  91
  92 static unsigned long get_memory_block_size(void)
  93 {
  94         unsigned long block_sz;
  95
  96         block_sz = memory_block_size_bytes();
  97
  98         /* Validate blk_sz is a power of 2 and not less than section size */
  99         if ((block_sz & (block_sz - 1)) || (block_sz < MIN_MEMORY_BLOCK_SIZE)) {
 100                 WARN_ON(1);
 101                 block_sz = MIN_MEMORY_BLOCK_SIZE;
 102         }
 103
 104         return block_sz;
 105 }
 106
 107 /*
 108  * use this as the physical section index that this memsection
 109  * uses.
 110  */
 111
 112 static ssize_t show_mem_start_phys_index(struct device *dev,
 113                         struct device_attribute *attr, char *buf)
 114 {
 115         struct memory_block *mem = to_memory_block(dev);
 116         unsigned long phys_index;
 117
 118         phys_index = mem->start_section_nr / sections_per_block;
 119         return sprintf(buf, "%08lx\n", phys_index);
 120 }
 121
 122 /*
 123  * Show whether the section of memory is likely to be hot-removable
 124  */
 125 static ssize_t show_mem_removable(struct device *dev,
 126                         struct device_attribute *attr, char *buf)
 127 {
 128         unsigned long i, pfn;
 129         int ret = 1;
 130         struct memory_block *mem = to_memory_block(dev);
 131
 132         if (mem->state != MEM_ONLINE)
 133                 goto out;
 134
 135         for (i = 0; i < sections_per_block; i++) {
 136                 if (!present_section_nr(mem->start_section_nr + i))
 137                         continue;
 138                 pfn = section_nr_to_pfn(mem->start_section_nr + i);
 139                 ret &= is_mem_section_removable(pfn, PAGES_PER_SECTION);
 140         }
 141
 142 out:
 143         return sprintf(buf, "%d\n", ret);
 144 }
 145
 146 /*
 147  * online, offline, going offline, etc.
 148  */
 149 static ssize_t show_mem_state(struct device *dev,
 150                         struct device_attribute *attr, char *buf)
 151 {
 152         struct memory_block *mem = to_memory_block(dev);
 153         ssize_t len = 0;
 154
 155         /*
 156          * We can probably put these states in a nice little array
 157          * so that they're not open-coded
 158          */
 159         switch (mem->state) {
 160         case MEM_ONLINE:
 161                 len = sprintf(buf, "online\n");
 162                 break;
 163         case MEM_OFFLINE:
 164                 len = sprintf(buf, "offline\n");
 165                 break;
 166         case MEM_GOING_OFFLINE:
 167                 len = sprintf(buf, "going-offline\n");
 168                 break;
 169         default:
 170                 len = sprintf(buf, "ERROR-UNKNOWN-%ld\n",
 171                                 mem->state);
 172                 WARN_ON(1);
 173                 break;
 174         }
 175
 176         return len;
 177 }
 178
 179 int memory_notify(unsigned long val, void *v)
 180 {
 181         return blocking_notifier_call_chain(&memory_chain, val, v);
 182 }
 183
 184 int memory_isolate_notify(unsigned long val, void *v)
 185 {
 186         return atomic_notifier_call_chain(&memory_isolate_chain, val, v);
 187 }
 188
 189 /*
 190  * The probe routines leave the pages reserved, just as the bootmem code does.
 191  * Make sure they're still that way.
 192  */
 193 static bool pages_correctly_reserved(unsigned long start_pfn)
 194 {
 195         int i, j;
 196         struct page *page;
 197         unsigned long pfn = start_pfn;
 198
 199         /*
 200          * memmap between sections is not contiguous except with
 201          * SPARSEMEM_VMEMMAP. We lookup the page once per section
 202          * and assume memmap is contiguous within each section
 203          */
 204         for (i = 0; i < sections_per_block; i++, pfn += PAGES_PER_SECTION) {
 205                 if (WARN_ON_ONCE(!pfn_valid(pfn)))
 206                         return false;
 207                 page = pfn_to_page(pfn);
 208
 209                 for (j = 0; j < PAGES_PER_SECTION; j++) {
 210                         if (PageReserved(page + j))
 211                                 continue;
 212
 213                         printk(KERN_WARNING "section number %ld page number %d "
 214                                 "not reserved, was it already online?\n",
 215                                 pfn_to_section_nr(pfn), j);
 216
 217                         return false;
 218                 }
 219         }
 220
 221         return true;
 222 }
 223
 224 /*
 225  * MEMORY_HOTPLUG depends on SPARSEMEM in mm/Kconfig, so it is
 226  * OK to have direct references to sparsemem variables in here.
 227  * Must already be protected by mem_hotplug_begin().
 228  */
 229 static int
 230 memory_block_action(unsigned long phys_index, unsigned long action, int online_type)
 231 {
 232         unsigned long start_pfn;
 233         unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
 234         int ret;
 235
 236         start_pfn = section_nr_to_pfn(phys_index);
 237
 238         switch (action) {
 239         case MEM_ONLINE:
 240                 if (!pages_correctly_reserved(start_pfn))
 241                         return -EBUSY;
 242
 243                 ret = online_pages(start_pfn, nr_pages, online_type);
 244                 break;
 245         case MEM_OFFLINE:
 246                 ret = offline_pages(start_pfn, nr_pages);
 247                 break;
 248         default:
 249                 WARN(1, KERN_WARNING "%s(%ld, %ld) unknown action: "
 250                      "%ld\n", __func__, phys_index, action, action);
 251                 ret = -EINVAL;
 252         }
 253
 254         return ret;
 255 }
 256
 257 static int memory_block_change_state(struct memory_block *mem,
 258                 unsigned long to_state, unsigned long from_state_req)
 259 {
 260         int ret = 0;
 261
 262         if (mem->state != from_state_req)
 263                 return -EINVAL;
 264
 265         if (to_state == MEM_OFFLINE)
 266                 mem->state = MEM_GOING_OFFLINE;
 267
 268         ret = memory_block_action(mem->start_section_nr, to_state,
 269                                 mem->online_type);
 270
 271         mem->state = ret ? from_state_req : to_state;
 272
 273         return ret;
 274 }
 275
 276 /* The device lock serializes operations on memory_subsys_[online|offline] */
 277 static int memory_subsys_online(struct device *dev)
 278 {
 279         struct memory_block *mem = to_memory_block(dev);
 280         int ret;
 281
 282         if (mem->state == MEM_ONLINE)
 283                 return 0;
 284
 285         /*
 286          * If we are called from store_mem_state(), online_type will be
 287          * set >= 0 Otherwise we were called from the device online
 288          * attribute and need to set the online_type.
 289          */
 290         if (mem->online_type < 0)
 291                 mem->online_type = MMOP_ONLINE_KEEP;
 292
 293         /* Already under protection of mem_hotplug_begin() */
 294         ret = memory_block_change_state(mem, MEM_ONLINE, MEM_OFFLINE);
 295
 296         /* clear online_type */
 297         mem->online_type = -1;
 298
 299         return ret;
 300 }
 301
 302 static int memory_subsys_offline(struct device *dev)
 303 {
 304         struct memory_block *mem = to_memory_block(dev);
 305
 306         if (mem->state == MEM_OFFLINE)
 307                 return 0;
 308
 309         /* Can't offline block with non-present sections */
 310         if (mem->section_count != sections_per_block)
 311                 return -EINVAL;
 312
 313         return memory_block_change_state(mem, MEM_OFFLINE, MEM_ONLINE);
 314 }
 315
 316 static ssize_t
 317 store_mem_state(struct device *dev,
 318                 struct device_attribute *attr, const char *buf, size_t count)
 319 {
 320         struct memory_block *mem = to_memory_block(dev);
 321         int ret, online_type;
 322
 323         ret = lock_device_hotplug_sysfs();
 324         if (ret)
 325                 return ret;
 326
 327         if (sysfs_streq(buf, "online_kernel"))
 328                 online_type = MMOP_ONLINE_KERNEL;
 329         else if (sysfs_streq(buf, "online_movable"))
 330                 online_type = MMOP_ONLINE_MOVABLE;
 331         else if (sysfs_streq(buf, "online"))
 332                 online_type = MMOP_ONLINE_KEEP;
 333         else if (sysfs_streq(buf, "offline"))
 334                 online_type = MMOP_OFFLINE;
 335         else {
 336                 ret = -EINVAL;
 337                 goto err;
 338         }
 339
 340         /*
 341          * Memory hotplug needs to hold mem_hotplug_begin() for probe to find
 342          * the correct memory block to online before doing device_online(dev),
 343          * which will take dev->mutex.  Take the lock early to prevent an
 344          * inversion, memory_subsys_online() callbacks will be implemented by
 345          * assuming it's already protected.
 346          */
 347         mem_hotplug_begin();
 348
 349         switch (online_type) {
 350         case MMOP_ONLINE_KERNEL:
 351         case MMOP_ONLINE_MOVABLE:
 352         case MMOP_ONLINE_KEEP:
 353                 mem->online_type = online_type;
 354                 ret = device_online(&mem->dev);
 355                 break;
 356         case MMOP_OFFLINE:
 357                 ret = device_offline(&mem->dev);
 358                 break;
 359         default:
 360                 ret = -EINVAL; /* should never happen */
 361         }
 362
 363         mem_hotplug_done();
 364 err:
 365         unlock_device_hotplug();
 366
 367         if (ret < 0)
 368                 return ret;
 369         if (ret)
 370                 return -EINVAL;
 371
 372         return count;
 373 }
 374
 375 /*
 376  * phys_device is a bad name for this.  What I really want
 377  * is a way to differentiate between memory ranges that
 378  * are part of physical devices that constitute
 379  * a complete removable unit or fru.
 380  * i.e. do these ranges belong to the same physical device,
 381  * s.t. if I offline all of these sections I can then
 382  * remove the physical device?
 383  */
 384 static ssize_t show_phys_device(struct device *dev,
 385                                 struct device_attribute *attr, char *buf)
 386 {
 387         struct memory_block *mem = to_memory_block(dev);
 388         return sprintf(buf, "%d\n", mem->phys_device);
 389 }
 390
 391 #ifdef CONFIG_MEMORY_HOTREMOVE
 392 static void print_allowed_zone(char *buf, int nid, unsigned long start_pfn,
 393                 unsigned long nr_pages, int online_type,
 394                 struct zone *default_zone)
 395 {
 396         struct zone *zone;
 397
 398         zone = zone_for_pfn_range(online_type, nid, start_pfn, nr_pages);
 399         if (zone != default_zone) {
 400                 strcat(buf, " ");
 401                 strcat(buf, zone->name);
 402         }
 403 }
 404
 405 static ssize_t show_valid_zones(struct device *dev,
 406                                 struct device_attribute *attr, char *buf)
 407 {
 408         struct memory_block *mem = to_memory_block(dev);
 409         unsigned long start_pfn = section_nr_to_pfn(mem->start_section_nr);
 410         unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
 411         unsigned long valid_start_pfn, valid_end_pfn;
 412         struct zone *default_zone;
 413         int nid;
 414
 415         /*
 416          * The block contains more than one zone can not be offlined.
 417          * This can happen e.g. for ZONE_DMA and ZONE_DMA32
 418          */
 419         if (!test_pages_in_a_zone(start_pfn, start_pfn + nr_pages, &valid_start_pfn, &valid_end_pfn))
 420                 return sprintf(buf, "none\n");
 421
 422         start_pfn = valid_start_pfn;
 423         nr_pages = valid_end_pfn - start_pfn;
 424
 425         /*
 426          * Check the existing zone. Make sure that we do that only on the
 427          * online nodes otherwise the page_zone is not reliable
 428          */
 429         if (mem->state == MEM_ONLINE) {
 430                 strcat(buf, page_zone(pfn_to_page(start_pfn))->name);
 431                 goto out;
 432         }
 433
 434         nid = pfn_to_nid(start_pfn);
 435         default_zone = zone_for_pfn_range(MMOP_ONLINE_KEEP, nid, start_pfn, nr_pages);
 436         strcat(buf, default_zone->name);
 437
 438         print_allowed_zone(buf, nid, start_pfn, nr_pages, MMOP_ONLINE_KERNEL,
 439                         default_zone);
 440         print_allowed_zone(buf, nid, start_pfn, nr_pages, MMOP_ONLINE_MOVABLE,
 441                         default_zone);
 442 out:
 443         strcat(buf, "\n");
 444
 445         return strlen(buf);
 446 }
 447 static DEVICE_ATTR(valid_zones, 0444, show_valid_zones, NULL);
 448 #endif
 449
 450 static DEVICE_ATTR(phys_index, 0444, show_mem_start_phys_index, NULL);
 451 static DEVICE_ATTR(state, 0644, show_mem_state, store_mem_state);
 452 static DEVICE_ATTR(phys_device, 0444, show_phys_device, NULL);
 453 static DEVICE_ATTR(removable, 0444, show_mem_removable, NULL);
 454
 455 /*
 456  * Block size attribute stuff
 457  */
 458 static ssize_t
 459 print_block_size(struct device *dev, struct device_attribute *attr,
 460                  char *buf)
 461 {
 462         return sprintf(buf, "%lx\n", get_memory_block_size());
 463 }
 464
 465 static DEVICE_ATTR(block_size_bytes, 0444, print_block_size, NULL);
 466
 467 /*
 468  * Memory auto online policy.
 469  */
 470
 471 static ssize_t
 472 show_auto_online_blocks(struct device *dev, struct device_attribute *attr,
 473                         char *buf)
 474 {
 475         if (memhp_auto_online)
 476                 return sprintf(buf, "online\n");
 477         else
 478                 return sprintf(buf, "offline\n");
 479 }
 480
 481 static ssize_t
 482 store_auto_online_blocks(struct device *dev, struct device_attribute *attr,
 483                          const char *buf, size_t count)
 484 {
 485         if (sysfs_streq(buf, "online"))
 486                 memhp_auto_online = true;
 487         else if (sysfs_streq(buf, "offline"))
 488                 memhp_auto_online = false;
 489         else
 490                 return -EINVAL;
 491
 492         return count;
 493 }
 494
 495 static DEVICE_ATTR(auto_online_blocks, 0644, show_auto_online_blocks,
 496                    store_auto_online_blocks);
 497
 498 /*
 499  * Some architectures will have custom drivers to do this, and
 500  * will not need to do it from userspace.  The fake hot-add code
 501  * as well as ppc64 will do all of their discovery in userspace
 502  * and will require this interface.
 503  */
 504 #ifdef CONFIG_ARCH_MEMORY_PROBE
 505 static ssize_t
 506 memory_probe_store(struct device *dev, struct device_attribute *attr,
 507                    const char *buf, size_t count)
 508 {
 509         u64 phys_addr;
 510         int nid, ret;
 511         unsigned long pages_per_block = PAGES_PER_SECTION * sections_per_block;
 512
 513         ret = kstrtoull(buf, 0, &phys_addr);
 514         if (ret)
 515                 return ret;
 516
 517         if (phys_addr & ((pages_per_block << PAGE_SHIFT) - 1))
 518                 return -EINVAL;
 519
 520         nid = memory_add_physaddr_to_nid(phys_addr);
 521         ret = add_memory(nid, phys_addr,
 522                          MIN_MEMORY_BLOCK_SIZE * sections_per_block);
 523
 524         if (ret)
 525                 goto out;
 526
 527         ret = count;
 528 out:
 529         return ret;
 530 }
 531
 532 static DEVICE_ATTR(probe, S_IWUSR, NULL, memory_probe_store);
 533 #endif
 534
 535 #ifdef CONFIG_MEMORY_FAILURE
 536 /*
 537  * Support for offlining pages of memory
 538  */
 539
 540 /* Soft offline a page */
 541 static ssize_t
 542 store_soft_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         if (!pfn_valid(pfn))
 554                 return -ENXIO;
 555         ret = soft_offline_page(pfn_to_page(pfn), 0);
 556         return ret == 0 ? count : ret;
 557 }
 558
 559 /* Forcibly offline a page, including killing processes. */
 560 static ssize_t
 561 store_hard_offline_page(struct device *dev,
 562                         struct device_attribute *attr,
 563                         const char *buf, size_t count)
 564 {
 565         int ret;
 566         u64 pfn;
 567         if (!capable(CAP_SYS_ADMIN))
 568                 return -EPERM;
 569         if (kstrtoull(buf, 0, &pfn) < 0)
 570                 return -EINVAL;
 571         pfn >>= PAGE_SHIFT;
 572         ret = memory_failure(pfn, 0, 0);
 573         return ret ? ret : count;
 574 }
 575
 576 static DEVICE_ATTR(soft_offline_page, S_IWUSR, NULL, store_soft_offline_page);
 577 static DEVICE_ATTR(hard_offline_page, S_IWUSR, NULL, store_hard_offline_page);
 578 #endif
 579
 580 /*
 581  * Note that phys_device is optional.  It is here to allow for
 582  * differentiation between which *physical* devices each
 583  * section belongs to...
 584  */
 585 int __weak arch_get_memory_phys_device(unsigned long start_pfn)
 586 {
 587         return 0;
 588 }
 589
 590 /*
 591  * A reference for the returned object is held and the reference for the
 592  * hinted object is released.
 593  */
 594 struct memory_block *find_memory_block_hinted(struct mem_section *section,
 595                                               struct memory_block *hint)
 596 {
 597         int block_id = base_memory_block_id(__section_nr(section));
 598         struct device *hintdev = hint ? &hint->dev : NULL;
 599         struct device *dev;
 600
 601         dev = subsys_find_device_by_id(&memory_subsys, block_id, hintdev);
 602         if (hint)
 603                 put_device(&hint->dev);
 604         if (!dev)
 605                 return NULL;
 606         return to_memory_block(dev);
 607 }
 608
 609 /*
 610  * For now, we have a linear search to go find the appropriate
 611  * memory_block corresponding to a particular phys_index. If
 612  * this gets to be a real problem, we can always use a radix
 613  * tree or something here.
 614  *
 615  * This could be made generic for all device subsystems.
 616  */
 617 struct memory_block *find_memory_block(struct mem_section *section)
 618 {
 619         return find_memory_block_hinted(section, NULL);
 620 }
 621
 622 static struct attribute *memory_memblk_attrs[] = {
 623         &dev_attr_phys_index.attr,
 624         &dev_attr_state.attr,
 625         &dev_attr_phys_device.attr,
 626         &dev_attr_removable.attr,
 627 #ifdef CONFIG_MEMORY_HOTREMOVE
 628         &dev_attr_valid_zones.attr,
 629 #endif
 630         NULL
 631 };
 632
 633 static struct attribute_group memory_memblk_attr_group = {
 634         .attrs = memory_memblk_attrs,
 635 };
 636
 637 static const struct attribute_group *memory_memblk_attr_groups[] = {
 638         &memory_memblk_attr_group,
 639         NULL,
 640 };
 641
 642 /*
 643  * register_memory - Setup a sysfs device for a memory block
 644  */
 645 static
 646 int register_memory(struct memory_block *memory)
 647 {
 648         memory->dev.bus = &memory_subsys;
 649         memory->dev.id = memory->start_section_nr / sections_per_block;
 650         memory->dev.release = memory_block_release;
 651         memory->dev.groups = memory_memblk_attr_groups;
 652         memory->dev.offline = memory->state == MEM_OFFLINE;
 653
 654         return device_register(&memory->dev);
 655 }
 656
 657 static int init_memory_block(struct memory_block **memory,
 658                              struct mem_section *section, unsigned long state)
 659 {
 660         struct memory_block *mem;
 661         unsigned long start_pfn;
 662         int scn_nr;
 663         int ret = 0;
 664
 665         mem = kzalloc(sizeof(*mem), GFP_KERNEL);
 666         if (!mem)
 667                 return -ENOMEM;
 668
 669         scn_nr = __section_nr(section);
 670         mem->start_section_nr =
 671                         base_memory_block_id(scn_nr) * sections_per_block;
 672         mem->end_section_nr = mem->start_section_nr + sections_per_block - 1;
 673         mem->state = state;
 674         start_pfn = section_nr_to_pfn(mem->start_section_nr);
 675         mem->phys_device = arch_get_memory_phys_device(start_pfn);
 676
 677         ret = register_memory(mem);
 678
 679         *memory = mem;
 680         return ret;
 681 }
 682
 683 static int add_memory_block(int base_section_nr)
 684 {
 685         struct memory_block *mem;
 686         int i, ret, section_count = 0, section_nr;
 687
 688         for (i = base_section_nr;
 689              (i < base_section_nr + sections_per_block) && i < NR_MEM_SECTIONS;
 690              i++) {
 691                 if (!present_section_nr(i))
 692                         continue;
 693                 if (section_count == 0)
 694                         section_nr = i;
 695                 section_count++;
 696         }
 697
 698         if (section_count == 0)
 699                 return 0;
 700         ret = init_memory_block(&mem, __nr_to_section(section_nr), MEM_ONLINE);
 701         if (ret)
 702                 return ret;
 703         mem->section_count = section_count;
 704         return 0;
 705 }
 706
 707 /*
 708  * need an interface for the VM to add new memory regions,
 709  * but without onlining it.
 710  */
 711 int register_new_memory(int nid, struct mem_section *section)
 712 {
 713         int ret = 0;
 714         struct memory_block *mem;
 715
 716         mutex_lock(&mem_sysfs_mutex);
 717
 718         mem = find_memory_block(section);
 719         if (mem) {
 720                 mem->section_count++;
 721                 put_device(&mem->dev);
 722         } else {
 723                 ret = init_memory_block(&mem, section, MEM_OFFLINE);
 724                 if (ret)
 725                         goto out;
 726                 mem->section_count++;
 727         }
 728
 729         if (mem->section_count == sections_per_block)
 730                 ret = register_mem_sect_under_node(mem, nid);
 731 out:
 732         mutex_unlock(&mem_sysfs_mutex);
 733         return ret;
 734 }
 735
 736 #ifdef CONFIG_MEMORY_HOTREMOVE
 737 static void
 738 unregister_memory(struct memory_block *memory)
 739 {
 740         BUG_ON(memory->dev.bus != &memory_subsys);
 741
 742         /* drop the ref. we got in remove_memory_block() */
 743         put_device(&memory->dev);
 744         device_unregister(&memory->dev);
 745 }
 746
 747 static int remove_memory_section(unsigned long node_id,
 748                                struct mem_section *section, int phys_device)
 749 {
 750         struct memory_block *mem;
 751
 752         mutex_lock(&mem_sysfs_mutex);
 753
 754         /*
 755          * Some users of the memory hotplug do not want/need memblock to
 756          * track all sections. Skip over those.
 757          */
 758         mem = find_memory_block(section);
 759         if (!mem)
 760                 goto out_unlock;
 761
 762         unregister_mem_sect_under_nodes(mem, __section_nr(section));
 763
 764         mem->section_count--;
 765         if (mem->section_count == 0)
 766                 unregister_memory(mem);
 767         else
 768                 put_device(&mem->dev);
 769
 770 out_unlock:
 771         mutex_unlock(&mem_sysfs_mutex);
 772         return 0;
 773 }
 774
 775 int unregister_memory_section(struct mem_section *section)
 776 {
 777         if (!present_section(section))
 778                 return -EINVAL;
 779
 780         return remove_memory_section(0, section, 0);
 781 }
 782 #endif /* CONFIG_MEMORY_HOTREMOVE */
 783
 784 /* return true if the memory block is offlined, otherwise, return false */
 785 bool is_memblock_offlined(struct memory_block *mem)
 786 {
 787         return mem->state == MEM_OFFLINE;
 788 }
 789
 790 static struct attribute *memory_root_attrs[] = {
 791 #ifdef CONFIG_ARCH_MEMORY_PROBE
 792         &dev_attr_probe.attr,
 793 #endif
 794
 795 #ifdef CONFIG_MEMORY_FAILURE
 796         &dev_attr_soft_offline_page.attr,
 797         &dev_attr_hard_offline_page.attr,
 798 #endif
 799
 800         &dev_attr_block_size_bytes.attr,
 801         &dev_attr_auto_online_blocks.attr,
 802         NULL
 803 };
 804
 805 static struct attribute_group memory_root_attr_group = {
 806         .attrs = memory_root_attrs,
 807 };
 808
 809 static const struct attribute_group *memory_root_attr_groups[] = {
 810         &memory_root_attr_group,
 811         NULL,
 812 };
 813
 814 /*
 815  * Initialize the sysfs support for memory devices...
 816  */
 817 int __init memory_dev_init(void)
 818 {
 819         unsigned int i;
 820         int ret;
 821         int err;
 822         unsigned long block_sz;
 823
 824         ret = subsys_system_register(&memory_subsys, memory_root_attr_groups);
 825         if (ret)
 826                 goto out;
 827
 828         block_sz = get_memory_block_size();
 829         sections_per_block = block_sz / MIN_MEMORY_BLOCK_SIZE;
 830
 831         /*
 832          * Create entries for memory sections that were found
 833          * during boot and have been initialized
 834          */
 835         mutex_lock(&mem_sysfs_mutex);
 836         for (i = 0; i < NR_MEM_SECTIONS; i += sections_per_block) {
 837                 /* Don't iterate over sections we know are !present: */
 838                 if (i > __highest_present_section_nr)
 839                         break;
 840
 841                 err = add_memory_block(i);
 842                 if (!ret)
 843                         ret = err;
 844         }
 845         mutex_unlock(&mem_sysfs_mutex);
 846
 847 out:
 848         if (ret)
 849                 printk(KERN_ERR "%s() failed: %d\n", __func__, ret);
 850         return ret;
 851 }