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