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 EXPORT_SYMBOL_GPL(memory_block_size_bytes);
  92
  93 static unsigned long get_memory_block_size(void)
  94 {
  95         unsigned long block_sz;
  96
  97         block_sz = memory_block_size_bytes();
  98
  99         /* Validate blk_sz is a power of 2 and not less than section size */
 100         if ((block_sz & (block_sz - 1)) || (block_sz < MIN_MEMORY_BLOCK_SIZE)) {
 101                 WARN_ON(1);
 102                 block_sz = MIN_MEMORY_BLOCK_SIZE;
 103         }
 104
 105         return block_sz;
 106 }
 107
 108 /*
 109  * use this as the physical section index that this memsection
 110  * uses.
 111  */
 112
 113 static ssize_t phys_index_show(struct device *dev,
 114                                struct device_attribute *attr, char *buf)
 115 {
 116         struct memory_block *mem = to_memory_block(dev);
 117         unsigned long phys_index;
 118
 119         phys_index = mem->start_section_nr / sections_per_block;
 120         return sprintf(buf, "%08lx\n", phys_index);
 121 }
 122
 123 /*
 124  * Show whether the section of memory is likely to be hot-removable
 125  */
 126 static ssize_t removable_show(struct device *dev, struct device_attribute *attr,
 127                               char *buf)
 128 {
 129         unsigned long i, pfn;
 130         int ret = 1;
 131         struct memory_block *mem = to_memory_block(dev);
 132
 133         if (mem->state != MEM_ONLINE)
 134                 goto out;
 135
 136         for (i = 0; i < sections_per_block; i++) {
 137                 if (!present_section_nr(mem->start_section_nr + i))
 138                         continue;
 139                 pfn = section_nr_to_pfn(mem->start_section_nr + i);
 140                 ret &= is_mem_section_removable(pfn, PAGES_PER_SECTION);
 141         }
 142
 143 out:
 144         return sprintf(buf, "%d\n", ret);
 145 }
 146
 147 /*
 148  * online, offline, going offline, etc.
 149  */
 150 static ssize_t state_show(struct device *dev, struct device_attribute *attr,
 151                           char *buf)
 152 {
 153         struct memory_block *mem = to_memory_block(dev);
 154         ssize_t len = 0;
 155
 156         /*
 157          * We can probably put these states in a nice little array
 158          * so that they're not open-coded
 159          */
 160         switch (mem->state) {
 161         case MEM_ONLINE:
 162                 len = sprintf(buf, "online\n");
 163                 break;
 164         case MEM_OFFLINE:
 165                 len = sprintf(buf, "offline\n");
 166                 break;
 167         case MEM_GOING_OFFLINE:
 168                 len = sprintf(buf, "going-offline\n");
 169                 break;
 170         default:
 171                 len = sprintf(buf, "ERROR-UNKNOWN-%ld\n",
 172                                 mem->state);
 173                 WARN_ON(1);
 174                 break;
 175         }
 176
 177         return len;
 178 }
 179
 180 int memory_notify(unsigned long val, void *v)
 181 {
 182         return blocking_notifier_call_chain(&memory_chain, val, v);
 183 }
 184
 185 int memory_isolate_notify(unsigned long val, void *v)
 186 {
 187         return atomic_notifier_call_chain(&memory_isolate_chain, val, v);
 188 }
 189
 190 /*
 191  * The probe routines leave the pages uninitialized, just as the bootmem code
 192  * does. Make sure we do not access them, but instead use only information from
 193  * within sections.
 194  */
 195 static bool pages_correctly_probed(unsigned long start_pfn)
 196 {
 197         unsigned long section_nr = pfn_to_section_nr(start_pfn);
 198         unsigned long section_nr_end = section_nr + sections_per_block;
 199         unsigned long pfn = start_pfn;
 200
 201         /*
 202          * memmap between sections is not contiguous except with
 203          * SPARSEMEM_VMEMMAP. We lookup the page once per section
 204          * and assume memmap is contiguous within each section
 205          */
 206         for (; section_nr < section_nr_end; section_nr++) {
 207                 if (WARN_ON_ONCE(!pfn_valid(pfn)))
 208                         return false;
 209
 210                 if (!present_section_nr(section_nr)) {
 211                         pr_warn("section %ld pfn[%lx, %lx) not present\n",
 212                                 section_nr, pfn, pfn + PAGES_PER_SECTION);
 213                         return false;
 214                 } else if (!valid_section_nr(section_nr)) {
 215                         pr_warn("section %ld pfn[%lx, %lx) no valid memmap\n",
 216                                 section_nr, pfn, pfn + PAGES_PER_SECTION);
 217                         return false;
 218                 } else if (online_section_nr(section_nr)) {
 219                         pr_warn("section %ld pfn[%lx, %lx) is already online\n",
 220                                 section_nr, pfn, pfn + PAGES_PER_SECTION);
 221                         return false;
 222                 }
 223                 pfn += PAGES_PER_SECTION;
 224         }
 225
 226         return true;
 227 }
 228
 229 /*
 230  * MEMORY_HOTPLUG depends on SPARSEMEM in mm/Kconfig, so it is
 231  * OK to have direct references to sparsemem variables in here.
 232  */
 233 static int
 234 memory_block_action(unsigned long phys_index, unsigned long action, int online_type)
 235 {
 236         unsigned long start_pfn;
 237         unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
 238         int ret;
 239
 240         start_pfn = section_nr_to_pfn(phys_index);
 241
 242         switch (action) {
 243         case MEM_ONLINE:
 244                 if (!pages_correctly_probed(start_pfn))
 245                         return -EBUSY;
 246
 247                 ret = online_pages(start_pfn, nr_pages, online_type);
 248                 break;
 249         case MEM_OFFLINE:
 250                 ret = offline_pages(start_pfn, nr_pages);
 251                 break;
 252         default:
 253                 WARN(1, KERN_WARNING "%s(%ld, %ld) unknown action: "
 254                      "%ld\n", __func__, phys_index, action, action);
 255                 ret = -EINVAL;
 256         }
 257
 258         return ret;
 259 }
 260
 261 static int memory_block_change_state(struct memory_block *mem,
 262                 unsigned long to_state, unsigned long from_state_req)
 263 {
 264         int ret = 0;
 265
 266         if (mem->state != from_state_req)
 267                 return -EINVAL;
 268
 269         if (to_state == MEM_OFFLINE)
 270                 mem->state = MEM_GOING_OFFLINE;
 271
 272         ret = memory_block_action(mem->start_section_nr, to_state,
 273                                 mem->online_type);
 274
 275         mem->state = ret ? from_state_req : to_state;
 276
 277         return ret;
 278 }
 279
 280 /* The device lock serializes operations on memory_subsys_[online|offline] */
 281 static int memory_subsys_online(struct device *dev)
 282 {
 283         struct memory_block *mem = to_memory_block(dev);
 284         int ret;
 285
 286         if (mem->state == MEM_ONLINE)
 287                 return 0;
 288
 289         /*
 290          * If we are called from state_store(), online_type will be
 291          * set >= 0 Otherwise we were called from the device online
 292          * attribute and need to set the online_type.
 293          */
 294         if (mem->online_type < 0)
 295                 mem->online_type = MMOP_ONLINE_KEEP;
 296
 297         ret = memory_block_change_state(mem, MEM_ONLINE, MEM_OFFLINE);
 298
 299         /* clear online_type */
 300         mem->online_type = -1;
 301
 302         return ret;
 303 }
 304
 305 static int memory_subsys_offline(struct device *dev)
 306 {
 307         struct memory_block *mem = to_memory_block(dev);
 308
 309         if (mem->state == MEM_OFFLINE)
 310                 return 0;
 311
 312         /* Can't offline block with non-present sections */
 313         if (mem->section_count != sections_per_block)
 314                 return -EINVAL;
 315
 316         return memory_block_change_state(mem, MEM_OFFLINE, MEM_ONLINE);
 317 }
 318
 319 static ssize_t state_store(struct device *dev, struct device_attribute *attr,
 320                            const char *buf, size_t count)
 321 {
 322         struct memory_block *mem = to_memory_block(dev);
 323         int ret, online_type;
 324
 325         ret = lock_device_hotplug_sysfs();
 326         if (ret)
 327                 return ret;
 328
 329         if (sysfs_streq(buf, "online_kernel"))
 330                 online_type = MMOP_ONLINE_KERNEL;
 331         else if (sysfs_streq(buf, "online_movable"))
 332                 online_type = MMOP_ONLINE_MOVABLE;
 333         else if (sysfs_streq(buf, "online"))
 334                 online_type = MMOP_ONLINE_KEEP;
 335         else if (sysfs_streq(buf, "offline"))
 336                 online_type = MMOP_OFFLINE;
 337         else {
 338                 ret = -EINVAL;
 339                 goto err;
 340         }
 341
 342         switch (online_type) {
 343         case MMOP_ONLINE_KERNEL:
 344         case MMOP_ONLINE_MOVABLE:
 345         case MMOP_ONLINE_KEEP:
 346                 /* mem->online_type is protected by device_hotplug_lock */
 347                 mem->online_type = online_type;
 348                 ret = device_online(&mem->dev);
 349                 break;
 350         case MMOP_OFFLINE:
 351                 ret = device_offline(&mem->dev);
 352                 break;
 353         default:
 354                 ret = -EINVAL; /* should never happen */
 355         }
 356
 357 err:
 358         unlock_device_hotplug();
 359
 360         if (ret < 0)
 361                 return ret;
 362         if (ret)
 363                 return -EINVAL;
 364
 365         return count;
 366 }
 367
 368 /*
 369  * phys_device is a bad name for this.  What I really want
 370  * is a way to differentiate between memory ranges that
 371  * are part of physical devices that constitute
 372  * a complete removable unit or fru.
 373  * i.e. do these ranges belong to the same physical device,
 374  * s.t. if I offline all of these sections I can then
 375  * remove the physical device?
 376  */
 377 static ssize_t phys_device_show(struct device *dev,
 378                                 struct device_attribute *attr, char *buf)
 379 {
 380         struct memory_block *mem = to_memory_block(dev);
 381         return sprintf(buf, "%d\n", mem->phys_device);
 382 }
 383
 384 #ifdef CONFIG_MEMORY_HOTREMOVE
 385 static void print_allowed_zone(char *buf, int nid, unsigned long start_pfn,
 386                 unsigned long nr_pages, int online_type,
 387                 struct zone *default_zone)
 388 {
 389         struct zone *zone;
 390
 391         zone = zone_for_pfn_range(online_type, nid, start_pfn, nr_pages);
 392         if (zone != default_zone) {
 393                 strcat(buf, " ");
 394                 strcat(buf, zone->name);
 395         }
 396 }
 397
 398 static ssize_t valid_zones_show(struct device *dev,
 399                                 struct device_attribute *attr, char *buf)
 400 {
 401         struct memory_block *mem = to_memory_block(dev);
 402         unsigned long start_pfn = section_nr_to_pfn(mem->start_section_nr);
 403         unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
 404         unsigned long valid_start_pfn, valid_end_pfn;
 405         struct zone *default_zone;
 406         int nid;
 407
 408         /*
 409          * Check the existing zone. Make sure that we do that only on the
 410          * online nodes otherwise the page_zone is not reliable
 411          */
 412         if (mem->state == MEM_ONLINE) {
 413                 /*
 414                  * The block contains more than one zone can not be offlined.
 415                  * This can happen e.g. for ZONE_DMA and ZONE_DMA32
 416                  */
 417                 if (!test_pages_in_a_zone(start_pfn, start_pfn + nr_pages,
 418                                           &valid_start_pfn, &valid_end_pfn))
 419                         return sprintf(buf, "none\n");
 420                 start_pfn = valid_start_pfn;
 421                 strcat(buf, page_zone(pfn_to_page(start_pfn))->name);
 422                 goto out;
 423         }
 424
 425         nid = mem->nid;
 426         default_zone = zone_for_pfn_range(MMOP_ONLINE_KEEP, nid, start_pfn, nr_pages);
 427         strcat(buf, default_zone->name);
 428
 429         print_allowed_zone(buf, nid, start_pfn, nr_pages, MMOP_ONLINE_KERNEL,
 430                         default_zone);
 431         print_allowed_zone(buf, nid, start_pfn, nr_pages, MMOP_ONLINE_MOVABLE,
 432                         default_zone);
 433 out:
 434         strcat(buf, "\n");
 435
 436         return strlen(buf);
 437 }
 438 static DEVICE_ATTR_RO(valid_zones);
 439 #endif
 440
 441 static DEVICE_ATTR_RO(phys_index);
 442 static DEVICE_ATTR_RW(state);
 443 static DEVICE_ATTR_RO(phys_device);
 444 static DEVICE_ATTR_RO(removable);
 445
 446 /*
 447  * Block size attribute stuff
 448  */
 449 static ssize_t block_size_bytes_show(struct device *dev,
 450                                      struct device_attribute *attr, char *buf)
 451 {
 452         return sprintf(buf, "%lx\n", get_memory_block_size());
 453 }
 454
 455 static DEVICE_ATTR_RO(block_size_bytes);
 456
 457 /*
 458  * Memory auto online policy.
 459  */
 460
 461 static ssize_t auto_online_blocks_show(struct device *dev,
 462                                        struct device_attribute *attr, char *buf)
 463 {
 464         if (memhp_auto_online)
 465                 return sprintf(buf, "online\n");
 466         else
 467                 return sprintf(buf, "offline\n");
 468 }
 469
 470 static ssize_t auto_online_blocks_store(struct device *dev,
 471                                         struct device_attribute *attr,
 472                                         const char *buf, size_t count)
 473 {
 474         if (sysfs_streq(buf, "online"))
 475                 memhp_auto_online = true;
 476         else if (sysfs_streq(buf, "offline"))
 477                 memhp_auto_online = false;
 478         else
 479                 return -EINVAL;
 480
 481         return count;
 482 }
 483
 484 static DEVICE_ATTR_RW(auto_online_blocks);
 485
 486 /*
 487  * Some architectures will have custom drivers to do this, and
 488  * will not need to do it from userspace.  The fake hot-add code
 489  * as well as ppc64 will do all of their discovery in userspace
 490  * and will require this interface.
 491  */
 492 #ifdef CONFIG_ARCH_MEMORY_PROBE
 493 static ssize_t probe_store(struct device *dev, struct device_attribute *attr,
 494                            const char *buf, size_t count)
 495 {
 496         u64 phys_addr;
 497         int nid, ret;
 498         unsigned long pages_per_block = PAGES_PER_SECTION * sections_per_block;
 499
 500         ret = kstrtoull(buf, 0, &phys_addr);
 501         if (ret)
 502                 return ret;
 503
 504         if (phys_addr & ((pages_per_block << PAGE_SHIFT) - 1))
 505                 return -EINVAL;
 506
 507         ret = lock_device_hotplug_sysfs();
 508         if (ret)
 509                 goto out;
 510
 511         nid = memory_add_physaddr_to_nid(phys_addr);
 512         ret = __add_memory(nid, phys_addr,
 513                            MIN_MEMORY_BLOCK_SIZE * sections_per_block);
 514
 515         if (ret)
 516                 goto out;
 517
 518         ret = count;
 519 out:
 520         unlock_device_hotplug();
 521         return ret;
 522 }
 523
 524 static DEVICE_ATTR_WO(probe);
 525 #endif
 526
 527 #ifdef CONFIG_MEMORY_FAILURE
 528 /*
 529  * Support for offlining pages of memory
 530  */
 531
 532 /* Soft offline a page */
 533 static ssize_t soft_offline_page_store(struct device *dev,
 534                                        struct device_attribute *attr,
 535                                        const char *buf, size_t count)
 536 {
 537         int ret;
 538         u64 pfn;
 539         if (!capable(CAP_SYS_ADMIN))
 540                 return -EPERM;
 541         if (kstrtoull(buf, 0, &pfn) < 0)
 542                 return -EINVAL;
 543         pfn >>= PAGE_SHIFT;
 544         if (!pfn_valid(pfn))
 545                 return -ENXIO;
 546         ret = soft_offline_page(pfn_to_page(pfn), 0);
 547         return ret == 0 ? count : ret;
 548 }
 549
 550 /* Forcibly offline a page, including killing processes. */
 551 static ssize_t hard_offline_page_store(struct device *dev,
 552                                        struct device_attribute *attr,
 553                                        const char *buf, size_t count)
 554 {
 555         int ret;
 556         u64 pfn;
 557         if (!capable(CAP_SYS_ADMIN))
 558                 return -EPERM;
 559         if (kstrtoull(buf, 0, &pfn) < 0)
 560                 return -EINVAL;
 561         pfn >>= PAGE_SHIFT;
 562         ret = memory_failure(pfn, 0);
 563         return ret ? ret : count;
 564 }
 565
 566 static DEVICE_ATTR_WO(soft_offline_page);
 567 static DEVICE_ATTR_WO(hard_offline_page);
 568 #endif
 569
 570 /*
 571  * Note that phys_device is optional.  It is here to allow for
 572  * differentiation between which *physical* devices each
 573  * section belongs to...
 574  */
 575 int __weak arch_get_memory_phys_device(unsigned long start_pfn)
 576 {
 577         return 0;
 578 }
 579
 580 /*
 581  * A reference for the returned object is held and the reference for the
 582  * hinted object is released.
 583  */
 584 struct memory_block *find_memory_block_hinted(struct mem_section *section,
 585                                               struct memory_block *hint)
 586 {
 587         int block_id = base_memory_block_id(__section_nr(section));
 588         struct device *hintdev = hint ? &hint->dev : NULL;
 589         struct device *dev;
 590
 591         dev = subsys_find_device_by_id(&memory_subsys, block_id, hintdev);
 592         if (hint)
 593                 put_device(&hint->dev);
 594         if (!dev)
 595                 return NULL;
 596         return to_memory_block(dev);
 597 }
 598
 599 /*
 600  * For now, we have a linear search to go find the appropriate
 601  * memory_block corresponding to a particular phys_index. If
 602  * this gets to be a real problem, we can always use a radix
 603  * tree or something here.
 604  *
 605  * This could be made generic for all device subsystems.
 606  */
 607 struct memory_block *find_memory_block(struct mem_section *section)
 608 {
 609         return find_memory_block_hinted(section, NULL);
 610 }
 611
 612 static struct attribute *memory_memblk_attrs[] = {
 613         &dev_attr_phys_index.attr,
 614         &dev_attr_state.attr,
 615         &dev_attr_phys_device.attr,
 616         &dev_attr_removable.attr,
 617 #ifdef CONFIG_MEMORY_HOTREMOVE
 618         &dev_attr_valid_zones.attr,
 619 #endif
 620         NULL
 621 };
 622
 623 static struct attribute_group memory_memblk_attr_group = {
 624         .attrs = memory_memblk_attrs,
 625 };
 626
 627 static const struct attribute_group *memory_memblk_attr_groups[] = {
 628         &memory_memblk_attr_group,
 629         NULL,
 630 };
 631
 632 /*
 633  * register_memory - Setup a sysfs device for a memory block
 634  */
 635 static
 636 int register_memory(struct memory_block *memory)
 637 {
 638         int ret;
 639
 640         memory->dev.bus = &memory_subsys;
 641         memory->dev.id = memory->start_section_nr / sections_per_block;
 642         memory->dev.release = memory_block_release;
 643         memory->dev.groups = memory_memblk_attr_groups;
 644         memory->dev.offline = memory->state == MEM_OFFLINE;
 645
 646         ret = device_register(&memory->dev);
 647         if (ret)
 648                 put_device(&memory->dev);
 649
 650         return ret;
 651 }
 652
 653 static int init_memory_block(struct memory_block **memory,
 654                              struct mem_section *section, unsigned long state)
 655 {
 656         struct memory_block *mem;
 657         unsigned long start_pfn;
 658         int scn_nr;
 659         int ret = 0;
 660
 661         mem = kzalloc(sizeof(*mem), GFP_KERNEL);
 662         if (!mem)
 663                 return -ENOMEM;
 664
 665         scn_nr = __section_nr(section);
 666         mem->start_section_nr =
 667                         base_memory_block_id(scn_nr) * sections_per_block;
 668         mem->end_section_nr = mem->start_section_nr + sections_per_block - 1;
 669         mem->state = state;
 670         start_pfn = section_nr_to_pfn(mem->start_section_nr);
 671         mem->phys_device = arch_get_memory_phys_device(start_pfn);
 672
 673         ret = register_memory(mem);
 674
 675         *memory = mem;
 676         return ret;
 677 }
 678
 679 static int add_memory_block(int base_section_nr)
 680 {
 681         struct memory_block *mem;
 682         int i, ret, section_count = 0, section_nr;
 683
 684         for (i = base_section_nr;
 685              i < base_section_nr + sections_per_block;
 686              i++) {
 687                 if (!present_section_nr(i))
 688                         continue;
 689                 if (section_count == 0)
 690                         section_nr = i;
 691                 section_count++;
 692         }
 693
 694         if (section_count == 0)
 695                 return 0;
 696         ret = init_memory_block(&mem, __nr_to_section(section_nr), MEM_ONLINE);
 697         if (ret)
 698                 return ret;
 699         mem->section_count = section_count;
 700         return 0;
 701 }
 702
 703 /*
 704  * need an interface for the VM to add new memory regions,
 705  * but without onlining it.
 706  */
 707 int hotplug_memory_register(int nid, struct mem_section *section)
 708 {
 709         int ret = 0;
 710         struct memory_block *mem;
 711
 712         mutex_lock(&mem_sysfs_mutex);
 713
 714         mem = find_memory_block(section);
 715         if (mem) {
 716                 mem->section_count++;
 717                 put_device(&mem->dev);
 718         } else {
 719                 ret = init_memory_block(&mem, section, MEM_OFFLINE);
 720                 if (ret)
 721                         goto out;
 722                 mem->section_count++;
 723         }
 724
 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_section() */
 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         mutex_lock(&mem_sysfs_mutex);
 747
 748         /*
 749          * Some users of the memory hotplug do not want/need memblock to
 750          * track all sections. Skip over those.
 751          */
 752         mem = find_memory_block(section);
 753         if (!mem)
 754                 goto out_unlock;
 755
 756         unregister_mem_sect_under_nodes(mem, __section_nr(section));
 757
 758         mem->section_count--;
 759         if (mem->section_count == 0)
 760                 unregister_memory(mem);
 761         else
 762                 put_device(&mem->dev);
 763
 764 out_unlock:
 765         mutex_unlock(&mem_sysfs_mutex);
 766         return 0;
 767 }
 768
 769 int unregister_memory_section(struct mem_section *section)
 770 {
 771         if (!present_section(section))
 772                 return -EINVAL;
 773
 774         return remove_memory_section(0, section, 0);
 775 }
 776 #endif /* CONFIG_MEMORY_HOTREMOVE */
 777
 778 /* return true if the memory block is offlined, otherwise, return false */
 779 bool is_memblock_offlined(struct memory_block *mem)
 780 {
 781         return mem->state == MEM_OFFLINE;
 782 }
 783
 784 static struct attribute *memory_root_attrs[] = {
 785 #ifdef CONFIG_ARCH_MEMORY_PROBE
 786         &dev_attr_probe.attr,
 787 #endif
 788
 789 #ifdef CONFIG_MEMORY_FAILURE
 790         &dev_attr_soft_offline_page.attr,
 791         &dev_attr_hard_offline_page.attr,
 792 #endif
 793
 794         &dev_attr_block_size_bytes.attr,
 795         &dev_attr_auto_online_blocks.attr,
 796         NULL
 797 };
 798
 799 static struct attribute_group memory_root_attr_group = {
 800         .attrs = memory_root_attrs,
 801 };
 802
 803 static const struct attribute_group *memory_root_attr_groups[] = {
 804         &memory_root_attr_group,
 805         NULL,
 806 };
 807
 808 /*
 809  * Initialize the sysfs support for memory devices...
 810  */
 811 int __init memory_dev_init(void)
 812 {
 813         unsigned int i;
 814         int ret;
 815         int err;
 816         unsigned long block_sz;
 817
 818         ret = subsys_system_register(&memory_subsys, memory_root_attr_groups);
 819         if (ret)
 820                 goto out;
 821
 822         block_sz = get_memory_block_size();
 823         sections_per_block = block_sz / MIN_MEMORY_BLOCK_SIZE;
 824
 825         /*
 826          * Create entries for memory sections that were found
 827          * during boot and have been initialized
 828          */
 829         mutex_lock(&mem_sysfs_mutex);
 830         for (i = 0; i <= __highest_present_section_nr;
 831                 i += sections_per_block) {
 832                 err = add_memory_block(i);
 833                 if (!ret)
 834                         ret = err;
 835         }
 836         mutex_unlock(&mem_sysfs_mutex);
 837
 838 out:
 839         if (ret)
 840                 printk(KERN_ERR "%s() failed: %d\n", __func__, ret);
 841         return ret;
 842 }