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 <asm/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 struct sysdev_class memory_sysdev_class = {
  35         .name = MEMORY_CLASS_NAME,
  36 };
  37
  38 static const char *memory_uevent_name(struct kset *kset, struct kobject *kobj)
  39 {
  40         return MEMORY_CLASS_NAME;
  41 }
  42
  43 static int memory_uevent(struct kset *kset, struct kobject *obj, struct kobj_uevent_env *env)
  44 {
  45         int retval = 0;
  46
  47         return retval;
  48 }
  49
  50 static const struct kset_uevent_ops memory_uevent_ops = {
  51         .name           = memory_uevent_name,
  52         .uevent         = memory_uevent,
  53 };
  54
  55 static BLOCKING_NOTIFIER_HEAD(memory_chain);
  56
  57 int register_memory_notifier(struct notifier_block *nb)
  58 {
  59         return blocking_notifier_chain_register(&memory_chain, nb);
  60 }
  61 EXPORT_SYMBOL(register_memory_notifier);
  62
  63 void unregister_memory_notifier(struct notifier_block *nb)
  64 {
  65         blocking_notifier_chain_unregister(&memory_chain, nb);
  66 }
  67 EXPORT_SYMBOL(unregister_memory_notifier);
  68
  69 static ATOMIC_NOTIFIER_HEAD(memory_isolate_chain);
  70
  71 int register_memory_isolate_notifier(struct notifier_block *nb)
  72 {
  73         return atomic_notifier_chain_register(&memory_isolate_chain, nb);
  74 }
  75 EXPORT_SYMBOL(register_memory_isolate_notifier);
  76
  77 void unregister_memory_isolate_notifier(struct notifier_block *nb)
  78 {
  79         atomic_notifier_chain_unregister(&memory_isolate_chain, nb);
  80 }
  81 EXPORT_SYMBOL(unregister_memory_isolate_notifier);
  82
  83 /*
  84  * register_memory - Setup a sysfs device for a memory block
  85  */
  86 static
  87 int register_memory(struct memory_block *memory, struct mem_section *section)
  88 {
  89         int error;
  90
  91         memory->sysdev.cls = &memory_sysdev_class;
  92         memory->sysdev.id = __section_nr(section);
  93
  94         error = sysdev_register(&memory->sysdev);
  95         return error;
  96 }
  97
  98 static void
  99 unregister_memory(struct memory_block *memory, struct mem_section *section)
 100 {
 101         BUG_ON(memory->sysdev.cls != &memory_sysdev_class);
 102         BUG_ON(memory->sysdev.id != __section_nr(section));
 103
 104         /* drop the ref. we got in remove_memory_block() */
 105         kobject_put(&memory->sysdev.kobj);
 106         sysdev_unregister(&memory->sysdev);
 107 }
 108
 109 /*
 110  * use this as the physical section index that this memsection
 111  * uses.
 112  */
 113
 114 static ssize_t show_mem_phys_index(struct sys_device *dev,
 115                         struct sysdev_attribute *attr, char *buf)
 116 {
 117         struct memory_block *mem =
 118                 container_of(dev, struct memory_block, sysdev);
 119         return sprintf(buf, "%08lx\n", mem->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 sys_device *dev,
 126                         struct sysdev_attribute *attr, char *buf)
 127 {
 128         unsigned long start_pfn;
 129         int ret;
 130         struct memory_block *mem =
 131                 container_of(dev, struct memory_block, sysdev);
 132
 133         start_pfn = section_nr_to_pfn(mem->phys_index);
 134         ret = is_mem_section_removable(start_pfn, PAGES_PER_SECTION);
 135         return sprintf(buf, "%d\n", ret);
 136 }
 137
 138 /*
 139  * online, offline, going offline, etc.
 140  */
 141 static ssize_t show_mem_state(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         ssize_t len = 0;
 147
 148         /*
 149          * We can probably put these states in a nice little array
 150          * so that they're not open-coded
 151          */
 152         switch (mem->state) {
 153                 case MEM_ONLINE:
 154                         len = sprintf(buf, "online\n");
 155                         break;
 156                 case MEM_OFFLINE:
 157                         len = sprintf(buf, "offline\n");
 158                         break;
 159                 case MEM_GOING_OFFLINE:
 160                         len = sprintf(buf, "going-offline\n");
 161                         break;
 162                 default:
 163                         len = sprintf(buf, "ERROR-UNKNOWN-%ld\n",
 164                                         mem->state);
 165                         WARN_ON(1);
 166                         break;
 167         }
 168
 169         return len;
 170 }
 171
 172 int memory_notify(unsigned long val, void *v)
 173 {
 174         return blocking_notifier_call_chain(&memory_chain, val, v);
 175 }
 176
 177 int memory_isolate_notify(unsigned long val, void *v)
 178 {
 179         return atomic_notifier_call_chain(&memory_isolate_chain, val, v);
 180 }
 181
 182 /*
 183  * MEMORY_HOTPLUG depends on SPARSEMEM in mm/Kconfig, so it is
 184  * OK to have direct references to sparsemem variables in here.
 185  */
 186 static int
 187 memory_block_action(struct memory_block *mem, unsigned long action)
 188 {
 189         int i;
 190         unsigned long psection;
 191         unsigned long start_pfn, start_paddr;
 192         struct page *first_page;
 193         int ret;
 194         int old_state = mem->state;
 195
 196         psection = mem->phys_index;
 197         first_page = pfn_to_page(psection << PFN_SECTION_SHIFT);
 198
 199         /*
 200          * The probe routines leave the pages reserved, just
 201          * as the bootmem code does.  Make sure they're still
 202          * that way.
 203          */
 204         if (action == MEM_ONLINE) {
 205                 for (i = 0; i < PAGES_PER_SECTION; i++) {
 206                         if (PageReserved(first_page+i))
 207                                 continue;
 208
 209                         printk(KERN_WARNING "section number %ld page number %d "
 210                                 "not reserved, was it already online? \n",
 211                                 psection, i);
 212                         return -EBUSY;
 213                 }
 214         }
 215
 216         switch (action) {
 217                 case MEM_ONLINE:
 218                         start_pfn = page_to_pfn(first_page);
 219                         ret = online_pages(start_pfn, PAGES_PER_SECTION);
 220                         break;
 221                 case MEM_OFFLINE:
 222                         mem->state = MEM_GOING_OFFLINE;
 223                         start_paddr = page_to_pfn(first_page) << PAGE_SHIFT;
 224                         ret = remove_memory(start_paddr,
 225                                             PAGES_PER_SECTION << PAGE_SHIFT);
 226                         if (ret) {
 227                                 mem->state = old_state;
 228                                 break;
 229                         }
 230                         break;
 231                 default:
 232                         WARN(1, KERN_WARNING "%s(%p, %ld) unknown action: %ld\n",
 233                                         __func__, mem, action, action);
 234                         ret = -EINVAL;
 235         }
 236
 237         return ret;
 238 }
 239
 240 static int memory_block_change_state(struct memory_block *mem,
 241                 unsigned long to_state, unsigned long from_state_req)
 242 {
 243         int ret = 0;
 244         mutex_lock(&mem->state_mutex);
 245
 246         if (mem->state != from_state_req) {
 247                 ret = -EINVAL;
 248                 goto out;
 249         }
 250
 251         ret = memory_block_action(mem, to_state);
 252         if (!ret)
 253                 mem->state = to_state;
 254
 255 out:
 256         mutex_unlock(&mem->state_mutex);
 257         return ret;
 258 }
 259
 260 static ssize_t
 261 store_mem_state(struct sys_device *dev,
 262                 struct sysdev_attribute *attr, const char *buf, size_t count)
 263 {
 264         struct memory_block *mem;
 265         unsigned int phys_section_nr;
 266         int ret = -EINVAL;
 267
 268         mem = container_of(dev, struct memory_block, sysdev);
 269         phys_section_nr = mem->phys_index;
 270
 271         if (!present_section_nr(phys_section_nr))
 272                 goto out;
 273
 274         if (!strncmp(buf, "online", min((int)count, 6)))
 275                 ret = memory_block_change_state(mem, MEM_ONLINE, MEM_OFFLINE);
 276         else if(!strncmp(buf, "offline", min((int)count, 7)))
 277                 ret = memory_block_change_state(mem, MEM_OFFLINE, MEM_ONLINE);
 278 out:
 279         if (ret)
 280                 return ret;
 281         return count;
 282 }
 283
 284 /*
 285  * phys_device is a bad name for this.  What I really want
 286  * is a way to differentiate between memory ranges that
 287  * are part of physical devices that constitute
 288  * a complete removable unit or fru.
 289  * i.e. do these ranges belong to the same physical device,
 290  * s.t. if I offline all of these sections I can then
 291  * remove the physical device?
 292  */
 293 static ssize_t show_phys_device(struct sys_device *dev,
 294                                 struct sysdev_attribute *attr, char *buf)
 295 {
 296         struct memory_block *mem =
 297                 container_of(dev, struct memory_block, sysdev);
 298         return sprintf(buf, "%d\n", mem->phys_device);
 299 }
 300
 301 static SYSDEV_ATTR(phys_index, 0444, show_mem_phys_index, NULL);
 302 static SYSDEV_ATTR(state, 0644, show_mem_state, store_mem_state);
 303 static SYSDEV_ATTR(phys_device, 0444, show_phys_device, NULL);
 304 static SYSDEV_ATTR(removable, 0444, show_mem_removable, NULL);
 305
 306 #define mem_create_simple_file(mem, attr_name)  \
 307         sysdev_create_file(&mem->sysdev, &attr_##attr_name)
 308 #define mem_remove_simple_file(mem, attr_name)  \
 309         sysdev_remove_file(&mem->sysdev, &attr_##attr_name)
 310
 311 /*
 312  * Block size attribute stuff
 313  */
 314 static ssize_t
 315 print_block_size(struct sysdev_class *class, struct sysdev_class_attribute *attr,
 316                  char *buf)
 317 {
 318         return sprintf(buf, "%lx\n", (unsigned long)PAGES_PER_SECTION * PAGE_SIZE);
 319 }
 320
 321 static SYSDEV_CLASS_ATTR(block_size_bytes, 0444, print_block_size, NULL);
 322
 323 static int block_size_init(void)
 324 {
 325         return sysfs_create_file(&memory_sysdev_class.kset.kobj,
 326                                 &attr_block_size_bytes.attr);
 327 }
 328
 329 /*
 330  * Some architectures will have custom drivers to do this, and
 331  * will not need to do it from userspace.  The fake hot-add code
 332  * as well as ppc64 will do all of their discovery in userspace
 333  * and will require this interface.
 334  */
 335 #ifdef CONFIG_ARCH_MEMORY_PROBE
 336 static ssize_t
 337 memory_probe_store(struct class *class, struct class_attribute *attr,
 338                    const char *buf, size_t count)
 339 {
 340         u64 phys_addr;
 341         int nid;
 342         int ret;
 343
 344         phys_addr = simple_strtoull(buf, NULL, 0);
 345
 346         nid = memory_add_physaddr_to_nid(phys_addr);
 347         ret = add_memory(nid, phys_addr, PAGES_PER_SECTION << PAGE_SHIFT);
 348
 349         if (ret)
 350                 count = ret;
 351
 352         return count;
 353 }
 354 static CLASS_ATTR(probe, S_IWUSR, NULL, memory_probe_store);
 355
 356 static int memory_probe_init(void)
 357 {
 358         return sysfs_create_file(&memory_sysdev_class.kset.kobj,
 359                                 &class_attr_probe.attr);
 360 }
 361 #else
 362 static inline int memory_probe_init(void)
 363 {
 364         return 0;
 365 }
 366 #endif
 367
 368 #ifdef CONFIG_MEMORY_FAILURE
 369 /*
 370  * Support for offlining pages of memory
 371  */
 372
 373 /* Soft offline a page */
 374 static ssize_t
 375 store_soft_offline_page(struct class *class,
 376                         struct class_attribute *attr,
 377                         const char *buf, size_t count)
 378 {
 379         int ret;
 380         u64 pfn;
 381         if (!capable(CAP_SYS_ADMIN))
 382                 return -EPERM;
 383         if (strict_strtoull(buf, 0, &pfn) < 0)
 384                 return -EINVAL;
 385         pfn >>= PAGE_SHIFT;
 386         if (!pfn_valid(pfn))
 387                 return -ENXIO;
 388         ret = soft_offline_page(pfn_to_page(pfn), 0);
 389         return ret == 0 ? count : ret;
 390 }
 391
 392 /* Forcibly offline a page, including killing processes. */
 393 static ssize_t
 394 store_hard_offline_page(struct class *class,
 395                         struct class_attribute *attr,
 396                         const char *buf, size_t count)
 397 {
 398         int ret;
 399         u64 pfn;
 400         if (!capable(CAP_SYS_ADMIN))
 401                 return -EPERM;
 402         if (strict_strtoull(buf, 0, &pfn) < 0)
 403                 return -EINVAL;
 404         pfn >>= PAGE_SHIFT;
 405         ret = __memory_failure(pfn, 0, 0);
 406         return ret ? ret : count;
 407 }
 408
 409 static CLASS_ATTR(soft_offline_page, 0644, NULL, store_soft_offline_page);
 410 static CLASS_ATTR(hard_offline_page, 0644, NULL, store_hard_offline_page);
 411
 412 static __init int memory_fail_init(void)
 413 {
 414         int err;
 415
 416         err = sysfs_create_file(&memory_sysdev_class.kset.kobj,
 417                                 &class_attr_soft_offline_page.attr);
 418         if (!err)
 419                 err = sysfs_create_file(&memory_sysdev_class.kset.kobj,
 420                                 &class_attr_hard_offline_page.attr);
 421         return err;
 422 }
 423 #else
 424 static inline int memory_fail_init(void)
 425 {
 426         return 0;
 427 }
 428 #endif
 429
 430 /*
 431  * Note that phys_device is optional.  It is here to allow for
 432  * differentiation between which *physical* devices each
 433  * section belongs to...
 434  */
 435 int __weak arch_get_memory_phys_device(unsigned long start_pfn)
 436 {
 437         return 0;
 438 }
 439
 440 struct memory_block *find_memory_block_hinted(struct mem_section *section,
 441                                               struct memory_block *hint)
 442 {
 443         struct kobject *kobj;
 444         struct sys_device *sysdev;
 445         struct memory_block *mem;
 446         char name[sizeof(MEMORY_CLASS_NAME) + 9 + 1];
 447
 448         kobj = hint ? &hint->sysdev.kobj : NULL;
 449
 450         /*
 451          * This only works because we know that section == sysdev->id
 452          * slightly redundant with sysdev_register()
 453          */
 454         sprintf(&name[0], "%s%d", MEMORY_CLASS_NAME, __section_nr(section));
 455
 456         kobj = kset_find_obj_hinted(&memory_sysdev_class.kset, name, kobj);
 457         if (!kobj)
 458                 return NULL;
 459
 460         sysdev = container_of(kobj, struct sys_device, kobj);
 461         mem = container_of(sysdev, struct memory_block, sysdev);
 462
 463         return mem;
 464 }
 465
 466 /*
 467  * For now, we have a linear search to go find the appropriate
 468  * memory_block corresponding to a particular phys_index. If
 469  * this gets to be a real problem, we can always use a radix
 470  * tree or something here.
 471  *
 472  * This could be made generic for all sysdev classes.
 473  */
 474 struct memory_block *find_memory_block(struct mem_section *section)
 475 {
 476         return find_memory_block_hinted(section, NULL);
 477 }
 478
 479 static int add_memory_block(int nid, struct mem_section *section,
 480                         unsigned long state, enum mem_add_context context)
 481 {
 482         struct memory_block *mem = kzalloc(sizeof(*mem), GFP_KERNEL);
 483         unsigned long start_pfn;
 484         int ret = 0;
 485
 486         if (!mem)
 487                 return -ENOMEM;
 488
 489         mutex_lock(&mem_sysfs_mutex);
 490
 491         mem->phys_index = __section_nr(section);
 492         mem->state = state;
 493         mem->section_count++;
 494         mutex_init(&mem->state_mutex);
 495         start_pfn = section_nr_to_pfn(mem->phys_index);
 496         mem->phys_device = arch_get_memory_phys_device(start_pfn);
 497
 498         ret = register_memory(mem, section);
 499         if (!ret)
 500                 ret = mem_create_simple_file(mem, phys_index);
 501         if (!ret)
 502                 ret = mem_create_simple_file(mem, state);
 503         if (!ret)
 504                 ret = mem_create_simple_file(mem, phys_device);
 505         if (!ret)
 506                 ret = mem_create_simple_file(mem, removable);
 507         if (!ret) {
 508                 if (context == HOTPLUG)
 509                         ret = register_mem_sect_under_node(mem, nid);
 510         }
 511
 512         mutex_unlock(&mem_sysfs_mutex);
 513         return ret;
 514 }
 515
 516 int remove_memory_block(unsigned long node_id, struct mem_section *section,
 517                 int phys_device)
 518 {
 519         struct memory_block *mem;
 520
 521         mutex_lock(&mem_sysfs_mutex);
 522         mem = find_memory_block(section);
 523
 524         mem->section_count--;
 525         if (mem->section_count == 0) {
 526                 unregister_mem_sect_under_nodes(mem);
 527                 mem_remove_simple_file(mem, phys_index);
 528                 mem_remove_simple_file(mem, state);
 529                 mem_remove_simple_file(mem, phys_device);
 530                 mem_remove_simple_file(mem, removable);
 531                 unregister_memory(mem, section);
 532         }
 533
 534         mutex_unlock(&mem_sysfs_mutex);
 535         return 0;
 536 }
 537
 538 /*
 539  * need an interface for the VM to add new memory regions,
 540  * but without onlining it.
 541  */
 542 int register_new_memory(int nid, struct mem_section *section)
 543 {
 544         return add_memory_block(nid, section, MEM_OFFLINE, HOTPLUG);
 545 }
 546
 547 int unregister_memory_section(struct mem_section *section)
 548 {
 549         if (!present_section(section))
 550                 return -EINVAL;
 551
 552         return remove_memory_block(0, section, 0);
 553 }
 554
 555 /*
 556  * Initialize the sysfs support for memory devices...
 557  */
 558 int __init memory_dev_init(void)
 559 {
 560         unsigned int i;
 561         int ret;
 562         int err;
 563
 564         memory_sysdev_class.kset.uevent_ops = &memory_uevent_ops;
 565         ret = sysdev_class_register(&memory_sysdev_class);
 566         if (ret)
 567                 goto out;
 568
 569         /*
 570          * Create entries for memory sections that were found
 571          * during boot and have been initialized
 572          */
 573         for (i = 0; i < NR_MEM_SECTIONS; i++) {
 574                 if (!present_section_nr(i))
 575                         continue;
 576                 err = add_memory_block(0, __nr_to_section(i), MEM_ONLINE,
 577                                        BOOT);
 578                 if (!ret)
 579                         ret = err;
 580         }
 581
 582         err = memory_probe_init();
 583         if (!ret)
 584                 ret = err;
 585         err = memory_fail_init();
 586         if (!ret)
 587                 ret = err;
 588         err = block_size_init();
 589         if (!ret)
 590                 ret = err;
 591 out:
 592         if (ret)
 593                 printk(KERN_ERR "%s() failed: %d\n", __func__, ret);
 594         return ret;
 595 }