drivers/base/memory.c

   1 /*
   2  * drivers/base/memory.c - basic Memory class 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/sysdev.h>
  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/kobject.h>
  21 #include <linux/memory_hotplug.h>
  22 #include <linux/mm.h>
  23 #include <linux/mutex.h>
  24 #include <linux/stat.h>
  25 #include <linux/slab.h>
  26
  27 #include <linux/atomic.h>
  28 #include <asm/uaccess.h>
  29
  30 static DEFINE_MUTEX(mem_sysfs_mutex);
  31
  32 #define MEMORY_CLASS_NAME       "memory"
  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 struct sysdev_class memory_sysdev_class = {
  42         .name = MEMORY_CLASS_NAME,
  43 };
  44
  45 static const char *memory_uevent_name(struct kset *kset, struct kobject *kobj)
  46 {
  47         return MEMORY_CLASS_NAME;
  48 }
  49
  50 static int memory_uevent(struct kset *kset, struct kobject *obj,
  51                         struct kobj_uevent_env *env)
  52 {
  53         int retval = 0;
  54
  55         return retval;
  56 }
  57
  58 static const struct kset_uevent_ops memory_uevent_ops = {
  59         .name           = memory_uevent_name,
  60         .uevent         = memory_uevent,
  61 };
  62
  63 static BLOCKING_NOTIFIER_HEAD(memory_chain);
  64
  65 int register_memory_notifier(struct notifier_block *nb)
  66 {
  67         return blocking_notifier_chain_register(&memory_chain, nb);
  68 }
  69 EXPORT_SYMBOL(register_memory_notifier);
  70
  71 void unregister_memory_notifier(struct notifier_block *nb)
  72 {
  73         blocking_notifier_chain_unregister(&memory_chain, nb);
  74 }
  75 EXPORT_SYMBOL(unregister_memory_notifier);
  76
  77 static ATOMIC_NOTIFIER_HEAD(memory_isolate_chain);
  78
  79 int register_memory_isolate_notifier(struct notifier_block *nb)
  80 {
  81         return atomic_notifier_chain_register(&memory_isolate_chain, nb);
  82 }
  83 EXPORT_SYMBOL(register_memory_isolate_notifier);
  84
  85 void unregister_memory_isolate_notifier(struct notifier_block *nb)
  86 {
  87         atomic_notifier_chain_unregister(&memory_isolate_chain, nb);
  88 }
  89 EXPORT_SYMBOL(unregister_memory_isolate_notifier);
  90
  91 /*
  92  * register_memory - Setup a sysfs device for a memory block
  93  */
  94 static
  95 int register_memory(struct memory_block *memory)
  96 {
  97         int error;
  98
  99         memory->sysdev.cls = &memory_sysdev_class;
 100         memory->sysdev.id = memory->start_section_nr / sections_per_block;
 101
 102         error = sysdev_register(&memory->sysdev);
 103         return error;
 104 }
 105
 106 static void
 107 unregister_memory(struct memory_block *memory)
 108 {
 109         BUG_ON(memory->sysdev.cls != &memory_sysdev_class);
 110
 111         /* drop the ref. we got in remove_memory_block() */
 112         kobject_put(&memory->sysdev.kobj);
 113         sysdev_unregister(&memory->sysdev);
 114 }
 115
 116 unsigned long __weak memory_block_size_bytes(void)
 117 {
 118         return MIN_MEMORY_BLOCK_SIZE;
 119 }
 120
 121 static unsigned long get_memory_block_size(void)
 122 {
 123         unsigned long block_sz;
 124
 125         block_sz = memory_block_size_bytes();
 126
 127         /* Validate blk_sz is a power of 2 and not less than section size */
 128         if ((block_sz & (block_sz - 1)) || (block_sz < MIN_MEMORY_BLOCK_SIZE)) {
 129                 WARN_ON(1);
 130                 block_sz = MIN_MEMORY_BLOCK_SIZE;
 131         }
 132
 133         return block_sz;
 134 }
 135
 136 /*
 137  * use this as the physical section index that this memsection
 138  * uses.
 139  */
 140
 141 static ssize_t show_mem_start_phys_index(struct sys_device *dev,
 142                         struct sysdev_attribute *attr, char *buf)
 143 {
 144         struct memory_block *mem =
 145                 container_of(dev, struct memory_block, sysdev);
 146         unsigned long phys_index;
 147
 148         phys_index = mem->start_section_nr / sections_per_block;
 149         return sprintf(buf, "%08lx\n", phys_index);
 150 }
 151
 152 static ssize_t show_mem_end_phys_index(struct sys_device *dev,
 153                         struct sysdev_attribute *attr, char *buf)
 154 {
 155         struct memory_block *mem =
 156                 container_of(dev, struct memory_block, sysdev);
 157         unsigned long phys_index;
 158
 159         phys_index = mem->end_section_nr / sections_per_block;
 160         return sprintf(buf, "%08lx\n", phys_index);
 161 }
 162
 163 /*
 164  * Show whether the section of memory is likely to be hot-removable
 165  */
 166 static ssize_t show_mem_removable(struct sys_device *dev,
 167                         struct sysdev_attribute *attr, char *buf)
 168 {
 169         unsigned long i, pfn;
 170         int ret = 1;
 171         struct memory_block *mem =
 172                 container_of(dev, struct memory_block, sysdev);
 173
 174         for (i = 0; i < sections_per_block; i++) {
 175                 pfn = section_nr_to_pfn(mem->start_section_nr + i);
 176                 ret &= is_mem_section_removable(pfn, PAGES_PER_SECTION);
 177         }
 178
 179         return sprintf(buf, "%d\n", ret);
 180 }
 181
 182 /*
 183  * online, offline, going offline, etc.
 184  */
 185 static ssize_t show_mem_state(struct sys_device *dev,
 186                         struct sysdev_attribute *attr, char *buf)
 187 {
 188         struct memory_block *mem =
 189                 container_of(dev, struct memory_block, sysdev);
 190         ssize_t len = 0;
 191
 192         /*
 193          * We can probably put these states in a nice little array
 194          * so that they're not open-coded
 195          */
 196         switch (mem->state) {
 197                 case MEM_ONLINE:
 198                         len = sprintf(buf, "online\n");
 199                         break;
 200                 case MEM_OFFLINE:
 201                         len = sprintf(buf, "offline\n");
 202                         break;
 203                 case MEM_GOING_OFFLINE:
 204                         len = sprintf(buf, "going-offline\n");
 205                         break;
 206                 default:
 207                         len = sprintf(buf, "ERROR-UNKNOWN-%ld\n",
 208                                         mem->state);
 209                         WARN_ON(1);
 210                         break;
 211         }
 212
 213         return len;
 214 }
 215
 216 int memory_notify(unsigned long val, void *v)
 217 {
 218         return blocking_notifier_call_chain(&memory_chain, val, v);
 219 }
 220
 221 int memory_isolate_notify(unsigned long val, void *v)
 222 {
 223         return atomic_notifier_call_chain(&memory_isolate_chain, val, v);
 224 }
 225
 226 /*
 227  * The probe routines leave the pages reserved, just as the bootmem code does.
 228  * Make sure they're still that way.
 229  */
 230 static bool pages_correctly_reserved(unsigned long start_pfn,
 231                                         unsigned long nr_pages)
 232 {
 233         int i, j;
 234         struct page *page;
 235         unsigned long pfn = start_pfn;
 236
 237         /*
 238          * memmap between sections is not contiguous except with
 239          * SPARSEMEM_VMEMMAP. We lookup the page once per section
 240          * and assume memmap is contiguous within each section
 241          */
 242         for (i = 0; i < sections_per_block; i++, pfn += PAGES_PER_SECTION) {
 243                 if (WARN_ON_ONCE(!pfn_valid(pfn)))
 244                         return false;
 245                 page = pfn_to_page(pfn);
 246
 247                 for (j = 0; j < PAGES_PER_SECTION; j++) {
 248                         if (PageReserved(page + j))
 249                                 continue;
 250
 251                         printk(KERN_WARNING "section number %ld page number %d "
 252                                 "not reserved, was it already online?\n",
 253                                 pfn_to_section_nr(pfn), j);
 254
 255                         return false;
 256                 }
 257         }
 258
 259         return true;
 260 }
 261
 262 /*
 263  * MEMORY_HOTPLUG depends on SPARSEMEM in mm/Kconfig, so it is
 264  * OK to have direct references to sparsemem variables in here.
 265  */
 266 static int
 267 memory_block_action(unsigned long phys_index, unsigned long action)
 268 {
 269         unsigned long start_pfn, start_paddr;
 270         unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
 271         struct page *first_page;
 272         int ret;
 273
 274         first_page = pfn_to_page(phys_index << PFN_SECTION_SHIFT);
 275
 276         switch (action) {
 277                 case MEM_ONLINE:
 278                         start_pfn = page_to_pfn(first_page);
 279
 280                         if (!pages_correctly_reserved(start_pfn, nr_pages))
 281                                 return -EBUSY;
 282
 283                         ret = online_pages(start_pfn, nr_pages);
 284                         break;
 285                 case MEM_OFFLINE:
 286                         start_paddr = page_to_pfn(first_page) << PAGE_SHIFT;
 287                         ret = remove_memory(start_paddr,
 288                                             nr_pages << PAGE_SHIFT);
 289                         break;
 290                 default:
 291                         WARN(1, KERN_WARNING "%s(%ld, %ld) unknown action: "
 292                              "%ld\n", __func__, phys_index, action, action);
 293                         ret = -EINVAL;
 294         }
 295
 296         return ret;
 297 }
 298
 299 static int memory_block_change_state(struct memory_block *mem,
 300                 unsigned long to_state, unsigned long from_state_req)
 301 {
 302         int ret = 0;
 303
 304         mutex_lock(&mem->state_mutex);
 305
 306         if (mem->state != from_state_req) {
 307                 ret = -EINVAL;
 308                 goto out;
 309         }
 310
 311         if (to_state == MEM_OFFLINE)
 312                 mem->state = MEM_GOING_OFFLINE;
 313
 314         ret = memory_block_action(mem->start_section_nr, to_state);
 315
 316         if (ret)
 317                 mem->state = from_state_req;
 318         else
 319                 mem->state = to_state;
 320
 321 out:
 322         mutex_unlock(&mem->state_mutex);
 323         return ret;
 324 }
 325
 326 static ssize_t
 327 store_mem_state(struct sys_device *dev,
 328                 struct sysdev_attribute *attr, const char *buf, size_t count)
 329 {
 330         struct memory_block *mem;
 331         int ret = -EINVAL;
 332
 333         mem = container_of(dev, struct memory_block, sysdev);
 334
 335         if (!strncmp(buf, "online", min((int)count, 6)))
 336                 ret = memory_block_change_state(mem, MEM_ONLINE, MEM_OFFLINE);
 337         else if(!strncmp(buf, "offline", min((int)count, 7)))
 338                 ret = memory_block_change_state(mem, MEM_OFFLINE, MEM_ONLINE);
 339
 340         if (ret)
 341                 return ret;
 342         return count;
 343 }
 344
 345 /*
 346  * phys_device is a bad name for this.  What I really want
 347  * is a way to differentiate between memory ranges that
 348  * are part of physical devices that constitute
 349  * a complete removable unit or fru.
 350  * i.e. do these ranges belong to the same physical device,
 351  * s.t. if I offline all of these sections I can then
 352  * remove the physical device?
 353  */
 354 static ssize_t show_phys_device(struct sys_device *dev,
 355                                 struct sysdev_attribute *attr, char *buf)
 356 {
 357         struct memory_block *mem =
 358                 container_of(dev, struct memory_block, sysdev);
 359         return sprintf(buf, "%d\n", mem->phys_device);
 360 }
 361
 362 static SYSDEV_ATTR(phys_index, 0444, show_mem_start_phys_index, NULL);
 363 static SYSDEV_ATTR(end_phys_index, 0444, show_mem_end_phys_index, NULL);
 364 static SYSDEV_ATTR(state, 0644, show_mem_state, store_mem_state);
 365 static SYSDEV_ATTR(phys_device, 0444, show_phys_device, NULL);
 366 static SYSDEV_ATTR(removable, 0444, show_mem_removable, NULL);
 367
 368 #define mem_create_simple_file(mem, attr_name)  \
 369         sysdev_create_file(&mem->sysdev, &attr_##attr_name)
 370 #define mem_remove_simple_file(mem, attr_name)  \
 371         sysdev_remove_file(&mem->sysdev, &attr_##attr_name)
 372
 373 /*
 374  * Block size attribute stuff
 375  */
 376 static ssize_t
 377 print_block_size(struct sysdev_class *class, struct sysdev_class_attribute *attr,
 378                  char *buf)
 379 {
 380         return sprintf(buf, "%lx\n", get_memory_block_size());
 381 }
 382
 383 static SYSDEV_CLASS_ATTR(block_size_bytes, 0444, print_block_size, NULL);
 384
 385 static int block_size_init(void)
 386 {
 387         return sysfs_create_file(&memory_sysdev_class.kset.kobj,
 388                                 &attr_block_size_bytes.attr);
 389 }
 390
 391 /*
 392  * Some architectures will have custom drivers to do this, and
 393  * will not need to do it from userspace.  The fake hot-add code
 394  * as well as ppc64 will do all of their discovery in userspace
 395  * and will require this interface.
 396  */
 397 #ifdef CONFIG_ARCH_MEMORY_PROBE
 398 static ssize_t
 399 memory_probe_store(struct class *class, struct class_attribute *attr,
 400                    const char *buf, size_t count)
 401 {
 402         u64 phys_addr;
 403         int nid;
 404         int i, ret;
 405         unsigned long pages_per_block = PAGES_PER_SECTION * sections_per_block;
 406
 407         phys_addr = simple_strtoull(buf, NULL, 0);
 408
 409         if (phys_addr & ((pages_per_block << PAGE_SHIFT) - 1))
 410                 return -EINVAL;
 411
 412         for (i = 0; i < sections_per_block; i++) {
 413                 nid = memory_add_physaddr_to_nid(phys_addr);
 414                 ret = add_memory(nid, phys_addr,
 415                                  PAGES_PER_SECTION << PAGE_SHIFT);
 416                 if (ret)
 417                         goto out;
 418
 419                 phys_addr += MIN_MEMORY_BLOCK_SIZE;
 420         }
 421
 422         ret = count;
 423 out:
 424         return ret;
 425 }
 426 static CLASS_ATTR(probe, S_IWUSR, NULL, memory_probe_store);
 427
 428 static int memory_probe_init(void)
 429 {
 430         return sysfs_create_file(&memory_sysdev_class.kset.kobj,
 431                                 &class_attr_probe.attr);
 432 }
 433 #else
 434 static inline int memory_probe_init(void)
 435 {
 436         return 0;
 437 }
 438 #endif
 439
 440 #ifdef CONFIG_MEMORY_FAILURE
 441 /*
 442  * Support for offlining pages of memory
 443  */
 444
 445 /* Soft offline a page */
 446 static ssize_t
 447 store_soft_offline_page(struct class *class,
 448                         struct class_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         if (!pfn_valid(pfn))
 459                 return -ENXIO;
 460         ret = soft_offline_page(pfn_to_page(pfn), 0);
 461         return ret == 0 ? count : ret;
 462 }
 463
 464 /* Forcibly offline a page, including killing processes. */
 465 static ssize_t
 466 store_hard_offline_page(struct class *class,
 467                         struct class_attribute *attr,
 468                         const char *buf, size_t count)
 469 {
 470         int ret;
 471         u64 pfn;
 472         if (!capable(CAP_SYS_ADMIN))
 473                 return -EPERM;
 474         if (strict_strtoull(buf, 0, &pfn) < 0)
 475                 return -EINVAL;
 476         pfn >>= PAGE_SHIFT;
 477         ret = __memory_failure(pfn, 0, 0);
 478         return ret ? ret : count;
 479 }
 480
 481 static CLASS_ATTR(soft_offline_page, 0644, NULL, store_soft_offline_page);
 482 static CLASS_ATTR(hard_offline_page, 0644, NULL, store_hard_offline_page);
 483
 484 static __init int memory_fail_init(void)
 485 {
 486         int err;
 487
 488         err = sysfs_create_file(&memory_sysdev_class.kset.kobj,
 489                                 &class_attr_soft_offline_page.attr);
 490         if (!err)
 491                 err = sysfs_create_file(&memory_sysdev_class.kset.kobj,
 492                                 &class_attr_hard_offline_page.attr);
 493         return err;
 494 }
 495 #else
 496 static inline int memory_fail_init(void)
 497 {
 498         return 0;
 499 }
 500 #endif
 501
 502 /*
 503  * Note that phys_device is optional.  It is here to allow for
 504  * differentiation between which *physical* devices each
 505  * section belongs to...
 506  */
 507 int __weak arch_get_memory_phys_device(unsigned long start_pfn)
 508 {
 509         return 0;
 510 }
 511
 512 struct memory_block *find_memory_block_hinted(struct mem_section *section,
 513                                               struct memory_block *hint)
 514 {
 515         struct kobject *kobj;
 516         struct sys_device *sysdev;
 517         struct memory_block *mem;
 518         char name[sizeof(MEMORY_CLASS_NAME) + 9 + 1];
 519         int block_id = base_memory_block_id(__section_nr(section));
 520
 521         kobj = hint ? &hint->sysdev.kobj : NULL;
 522
 523         /*
 524          * This only works because we know that section == sysdev->id
 525          * slightly redundant with sysdev_register()
 526          */
 527         sprintf(&name[0], "%s%d", MEMORY_CLASS_NAME, block_id);
 528
 529         kobj = kset_find_obj_hinted(&memory_sysdev_class.kset, name, kobj);
 530         if (!kobj)
 531                 return NULL;
 532
 533         sysdev = container_of(kobj, struct sys_device, kobj);
 534         mem = container_of(sysdev, struct memory_block, sysdev);
 535
 536         return mem;
 537 }
 538
 539 /*
 540  * For now, we have a linear search to go find the appropriate
 541  * memory_block corresponding to a particular phys_index. If
 542  * this gets to be a real problem, we can always use a radix
 543  * tree or something here.
 544  *
 545  * This could be made generic for all sysdev classes.
 546  */
 547 struct memory_block *find_memory_block(struct mem_section *section)
 548 {
 549         return find_memory_block_hinted(section, NULL);
 550 }
 551
 552 static int init_memory_block(struct memory_block **memory,
 553                              struct mem_section *section, unsigned long state)
 554 {
 555         struct memory_block *mem;
 556         unsigned long start_pfn;
 557         int scn_nr;
 558         int ret = 0;
 559
 560         mem = kzalloc(sizeof(*mem), GFP_KERNEL);
 561         if (!mem)
 562                 return -ENOMEM;
 563
 564         scn_nr = __section_nr(section);
 565         mem->start_section_nr =
 566                         base_memory_block_id(scn_nr) * sections_per_block;
 567         mem->end_section_nr = mem->start_section_nr + sections_per_block - 1;
 568         mem->state = state;
 569         mem->section_count++;
 570         mutex_init(&mem->state_mutex);
 571         start_pfn = section_nr_to_pfn(mem->start_section_nr);
 572         mem->phys_device = arch_get_memory_phys_device(start_pfn);
 573
 574         ret = register_memory(mem);
 575         if (!ret)
 576                 ret = mem_create_simple_file(mem, phys_index);
 577         if (!ret)
 578                 ret = mem_create_simple_file(mem, end_phys_index);
 579         if (!ret)
 580                 ret = mem_create_simple_file(mem, state);
 581         if (!ret)
 582                 ret = mem_create_simple_file(mem, phys_device);
 583         if (!ret)
 584                 ret = mem_create_simple_file(mem, removable);
 585
 586         *memory = mem;
 587         return ret;
 588 }
 589
 590 static int add_memory_section(int nid, struct mem_section *section,
 591                         unsigned long state, enum mem_add_context context)
 592 {
 593         struct memory_block *mem;
 594         int ret = 0;
 595
 596         mutex_lock(&mem_sysfs_mutex);
 597
 598         mem = find_memory_block(section);
 599         if (mem) {
 600                 mem->section_count++;
 601                 kobject_put(&mem->sysdev.kobj);
 602         } else
 603                 ret = init_memory_block(&mem, section, state);
 604
 605         if (!ret) {
 606                 if (context == HOTPLUG &&
 607                     mem->section_count == sections_per_block)
 608                         ret = register_mem_sect_under_node(mem, nid);
 609         }
 610
 611         mutex_unlock(&mem_sysfs_mutex);
 612         return ret;
 613 }
 614
 615 int remove_memory_block(unsigned long node_id, struct mem_section *section,
 616                 int phys_device)
 617 {
 618         struct memory_block *mem;
 619
 620         mutex_lock(&mem_sysfs_mutex);
 621         mem = find_memory_block(section);
 622         unregister_mem_sect_under_nodes(mem, __section_nr(section));
 623
 624         mem->section_count--;
 625         if (mem->section_count == 0) {
 626                 mem_remove_simple_file(mem, phys_index);
 627                 mem_remove_simple_file(mem, end_phys_index);
 628                 mem_remove_simple_file(mem, state);
 629                 mem_remove_simple_file(mem, phys_device);
 630                 mem_remove_simple_file(mem, removable);
 631                 unregister_memory(mem);
 632                 kfree(mem);
 633         } else
 634                 kobject_put(&mem->sysdev.kobj);
 635
 636         mutex_unlock(&mem_sysfs_mutex);
 637         return 0;
 638 }
 639
 640 /*
 641  * need an interface for the VM to add new memory regions,
 642  * but without onlining it.
 643  */
 644 int register_new_memory(int nid, struct mem_section *section)
 645 {
 646         return add_memory_section(nid, section, MEM_OFFLINE, HOTPLUG);
 647 }
 648
 649 int unregister_memory_section(struct mem_section *section)
 650 {
 651         if (!present_section(section))
 652                 return -EINVAL;
 653
 654         return remove_memory_block(0, section, 0);
 655 }
 656
 657 /*
 658  * Initialize the sysfs support for memory devices...
 659  */
 660 int __init memory_dev_init(void)
 661 {
 662         unsigned int i;
 663         int ret;
 664         int err;
 665         unsigned long block_sz;
 666
 667         memory_sysdev_class.kset.uevent_ops = &memory_uevent_ops;
 668         ret = sysdev_class_register(&memory_sysdev_class);
 669         if (ret)
 670                 goto out;
 671
 672         block_sz = get_memory_block_size();
 673         sections_per_block = block_sz / MIN_MEMORY_BLOCK_SIZE;
 674
 675         /*
 676          * Create entries for memory sections that were found
 677          * during boot and have been initialized
 678          */
 679         for (i = 0; i < NR_MEM_SECTIONS; i++) {
 680                 if (!present_section_nr(i))
 681                         continue;
 682                 err = add_memory_section(0, __nr_to_section(i), MEM_ONLINE,
 683                                          BOOT);
 684                 if (!ret)
 685                         ret = err;
 686         }
 687
 688         err = memory_probe_init();
 689         if (!ret)
 690                 ret = err;
 691         err = memory_fail_init();
 692         if (!ret)
 693                 ret = err;
 694         err = block_size_init();
 695         if (!ret)
 696                 ret = err;
 697 out:
 698         if (ret)
 699                 printk(KERN_ERR "%s() failed: %d\n", __func__, ret);
 700         return ret;
 701 }