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
  26 #include <asm/atomic.h>
  27 #include <asm/uaccess.h>
  28
  29 #define MEMORY_CLASS_NAME       "memory"
  30
  31 static struct sysdev_class memory_sysdev_class = {
  32         .name = MEMORY_CLASS_NAME,
  33 };
  34
  35 static const char *memory_uevent_name(struct kset *kset, struct kobject *kobj)
  36 {
  37         return MEMORY_CLASS_NAME;
  38 }
  39
  40 static int memory_uevent(struct kset *kset, struct kobject *obj, struct kobj_uevent_env *env)
  41 {
  42         int retval = 0;
  43
  44         return retval;
  45 }
  46
  47 static struct kset_uevent_ops memory_uevent_ops = {
  48         .name           = memory_uevent_name,
  49         .uevent         = memory_uevent,
  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 /*
  81  * register_memory - Setup a sysfs device for a memory block
  82  */
  83 static
  84 int register_memory(struct memory_block *memory, struct mem_section *section)
  85 {
  86         int error;
  87
  88         memory->sysdev.cls = &memory_sysdev_class;
  89         memory->sysdev.id = __section_nr(section);
  90
  91         error = sysdev_register(&memory->sysdev);
  92         return error;
  93 }
  94
  95 static void
  96 unregister_memory(struct memory_block *memory, struct mem_section *section)
  97 {
  98         BUG_ON(memory->sysdev.cls != &memory_sysdev_class);
  99         BUG_ON(memory->sysdev.id != __section_nr(section));
 100
 101         /* drop the ref. we got in remove_memory_block() */
 102         kobject_put(&memory->sysdev.kobj);
 103         sysdev_unregister(&memory->sysdev);
 104 }
 105
 106 /*
 107  * use this as the physical section index that this memsection
 108  * uses.
 109  */
 110
 111 static ssize_t show_mem_phys_index(struct sys_device *dev,
 112                         struct sysdev_attribute *attr, char *buf)
 113 {
 114         struct memory_block *mem =
 115                 container_of(dev, struct memory_block, sysdev);
 116         return sprintf(buf, "%08lx\n", mem->phys_index);
 117 }
 118
 119 /*
 120  * Show whether the section of memory is likely to be hot-removable
 121  */
 122 static ssize_t show_mem_removable(struct sys_device *dev,
 123                         struct sysdev_attribute *attr, char *buf)
 124 {
 125         unsigned long start_pfn;
 126         int ret;
 127         struct memory_block *mem =
 128                 container_of(dev, struct memory_block, sysdev);
 129
 130         start_pfn = section_nr_to_pfn(mem->phys_index);
 131         ret = is_mem_section_removable(start_pfn, PAGES_PER_SECTION);
 132         return sprintf(buf, "%d\n", ret);
 133 }
 134
 135 /*
 136  * online, offline, going offline, etc.
 137  */
 138 static ssize_t show_mem_state(struct sys_device *dev,
 139                         struct sysdev_attribute *attr, char *buf)
 140 {
 141         struct memory_block *mem =
 142                 container_of(dev, struct memory_block, sysdev);
 143         ssize_t len = 0;
 144
 145         /*
 146          * We can probably put these states in a nice little array
 147          * so that they're not open-coded
 148          */
 149         switch (mem->state) {
 150                 case MEM_ONLINE:
 151                         len = sprintf(buf, "online\n");
 152                         break;
 153                 case MEM_OFFLINE:
 154                         len = sprintf(buf, "offline\n");
 155                         break;
 156                 case MEM_GOING_OFFLINE:
 157                         len = sprintf(buf, "going-offline\n");
 158                         break;
 159                 default:
 160                         len = sprintf(buf, "ERROR-UNKNOWN-%ld\n",
 161                                         mem->state);
 162                         WARN_ON(1);
 163                         break;
 164         }
 165
 166         return len;
 167 }
 168
 169 int memory_notify(unsigned long val, void *v)
 170 {
 171         return blocking_notifier_call_chain(&memory_chain, val, v);
 172 }
 173
 174 int memory_isolate_notify(unsigned long val, void *v)
 175 {
 176         return atomic_notifier_call_chain(&memory_isolate_chain, val, v);
 177 }
 178
 179 /*
 180  * MEMORY_HOTPLUG depends on SPARSEMEM in mm/Kconfig, so it is
 181  * OK to have direct references to sparsemem variables in here.
 182  */
 183 static int
 184 memory_block_action(struct memory_block *mem, unsigned long action)
 185 {
 186         int i;
 187         unsigned long psection;
 188         unsigned long start_pfn, start_paddr;
 189         struct page *first_page;
 190         int ret;
 191         int old_state = mem->state;
 192
 193         psection = mem->phys_index;
 194         first_page = pfn_to_page(psection << PFN_SECTION_SHIFT);
 195
 196         /*
 197          * The probe routines leave the pages reserved, just
 198          * as the bootmem code does.  Make sure they're still
 199          * that way.
 200          */
 201         if (action == MEM_ONLINE) {
 202                 for (i = 0; i < PAGES_PER_SECTION; i++) {
 203                         if (PageReserved(first_page+i))
 204                                 continue;
 205
 206                         printk(KERN_WARNING "section number %ld page number %d "
 207                                 "not reserved, was it already online? \n",
 208                                 psection, i);
 209                         return -EBUSY;
 210                 }
 211         }
 212
 213         switch (action) {
 214                 case MEM_ONLINE:
 215                         start_pfn = page_to_pfn(first_page);
 216                         ret = online_pages(start_pfn, PAGES_PER_SECTION);
 217                         break;
 218                 case MEM_OFFLINE:
 219                         mem->state = MEM_GOING_OFFLINE;
 220                         start_paddr = page_to_pfn(first_page) << PAGE_SHIFT;
 221                         ret = remove_memory(start_paddr,
 222                                             PAGES_PER_SECTION << PAGE_SHIFT);
 223                         if (ret) {
 224                                 mem->state = old_state;
 225                                 break;
 226                         }
 227                         break;
 228                 default:
 229                         WARN(1, KERN_WARNING "%s(%p, %ld) unknown action: %ld\n",
 230                                         __func__, mem, action, action);
 231                         ret = -EINVAL;
 232         }
 233
 234         return ret;
 235 }
 236
 237 static int memory_block_change_state(struct memory_block *mem,
 238                 unsigned long to_state, unsigned long from_state_req)
 239 {
 240         int ret = 0;
 241         mutex_lock(&mem->state_mutex);
 242
 243         if (mem->state != from_state_req) {
 244                 ret = -EINVAL;
 245                 goto out;
 246         }
 247
 248         ret = memory_block_action(mem, to_state);
 249         if (!ret)
 250                 mem->state = to_state;
 251
 252 out:
 253         mutex_unlock(&mem->state_mutex);
 254         return ret;
 255 }
 256
 257 static ssize_t
 258 store_mem_state(struct sys_device *dev,
 259                 struct sysdev_attribute *attr, const char *buf, size_t count)
 260 {
 261         struct memory_block *mem;
 262         unsigned int phys_section_nr;
 263         int ret = -EINVAL;
 264
 265         mem = container_of(dev, struct memory_block, sysdev);
 266         phys_section_nr = mem->phys_index;
 267
 268         if (!present_section_nr(phys_section_nr))
 269                 goto out;
 270
 271         if (!strncmp(buf, "online", min((int)count, 6)))
 272                 ret = memory_block_change_state(mem, MEM_ONLINE, MEM_OFFLINE);
 273         else if(!strncmp(buf, "offline", min((int)count, 7)))
 274                 ret = memory_block_change_state(mem, MEM_OFFLINE, MEM_ONLINE);
 275 out:
 276         if (ret)
 277                 return ret;
 278         return count;
 279 }
 280
 281 /*
 282  * phys_device is a bad name for this.  What I really want
 283  * is a way to differentiate between memory ranges that
 284  * are part of physical devices that constitute
 285  * a complete removable unit or fru.
 286  * i.e. do these ranges belong to the same physical device,
 287  * s.t. if I offline all of these sections I can then
 288  * remove the physical device?
 289  */
 290 static ssize_t show_phys_device(struct sys_device *dev,
 291                                 struct sysdev_attribute *attr, char *buf)
 292 {
 293         struct memory_block *mem =
 294                 container_of(dev, struct memory_block, sysdev);
 295         return sprintf(buf, "%d\n", mem->phys_device);
 296 }
 297
 298 static SYSDEV_ATTR(phys_index, 0444, show_mem_phys_index, NULL);
 299 static SYSDEV_ATTR(state, 0644, show_mem_state, store_mem_state);
 300 static SYSDEV_ATTR(phys_device, 0444, show_phys_device, NULL);
 301 static SYSDEV_ATTR(removable, 0444, show_mem_removable, NULL);
 302
 303 #define mem_create_simple_file(mem, attr_name)  \
 304         sysdev_create_file(&mem->sysdev, &attr_##attr_name)
 305 #define mem_remove_simple_file(mem, attr_name)  \
 306         sysdev_remove_file(&mem->sysdev, &attr_##attr_name)
 307
 308 /*
 309  * Block size attribute stuff
 310  */
 311 static ssize_t
 312 print_block_size(struct class *class, char *buf)
 313 {
 314         return sprintf(buf, "%#lx\n", (unsigned long)PAGES_PER_SECTION * PAGE_SIZE);
 315 }
 316
 317 static CLASS_ATTR(block_size_bytes, 0444, print_block_size, NULL);
 318
 319 static int block_size_init(void)
 320 {
 321         return sysfs_create_file(&memory_sysdev_class.kset.kobj,
 322                                 &class_attr_block_size_bytes.attr);
 323 }
 324
 325 /*
 326  * Some architectures will have custom drivers to do this, and
 327  * will not need to do it from userspace.  The fake hot-add code
 328  * as well as ppc64 will do all of their discovery in userspace
 329  * and will require this interface.
 330  */
 331 #ifdef CONFIG_ARCH_MEMORY_PROBE
 332 static ssize_t
 333 memory_probe_store(struct class *class, const char *buf, size_t count)
 334 {
 335         u64 phys_addr;
 336         int nid;
 337         int ret;
 338
 339         phys_addr = simple_strtoull(buf, NULL, 0);
 340
 341         nid = memory_add_physaddr_to_nid(phys_addr);
 342         ret = add_memory(nid, phys_addr, PAGES_PER_SECTION << PAGE_SHIFT);
 343
 344         if (ret)
 345                 count = ret;
 346
 347         return count;
 348 }
 349 static CLASS_ATTR(probe, S_IWUSR, NULL, memory_probe_store);
 350
 351 static int memory_probe_init(void)
 352 {
 353         return sysfs_create_file(&memory_sysdev_class.kset.kobj,
 354                                 &class_attr_probe.attr);
 355 }
 356 #else
 357 static inline int memory_probe_init(void)
 358 {
 359         return 0;
 360 }
 361 #endif
 362
 363 #ifdef CONFIG_MEMORY_FAILURE
 364 /*
 365  * Support for offlining pages of memory
 366  */
 367
 368 /* Soft offline a page */
 369 static ssize_t
 370 store_soft_offline_page(struct class *class, const char *buf, size_t count)
 371 {
 372         int ret;
 373         u64 pfn;
 374         if (!capable(CAP_SYS_ADMIN))
 375                 return -EPERM;
 376         if (strict_strtoull(buf, 0, &pfn) < 0)
 377                 return -EINVAL;
 378         pfn >>= PAGE_SHIFT;
 379         if (!pfn_valid(pfn))
 380                 return -ENXIO;
 381         ret = soft_offline_page(pfn_to_page(pfn), 0);
 382         return ret == 0 ? count : ret;
 383 }
 384
 385 /* Forcibly offline a page, including killing processes. */
 386 static ssize_t
 387 store_hard_offline_page(struct class *class, const char *buf, size_t count)
 388 {
 389         int ret;
 390         u64 pfn;
 391         if (!capable(CAP_SYS_ADMIN))
 392                 return -EPERM;
 393         if (strict_strtoull(buf, 0, &pfn) < 0)
 394                 return -EINVAL;
 395         pfn >>= PAGE_SHIFT;
 396         ret = __memory_failure(pfn, 0, 0);
 397         return ret ? ret : count;
 398 }
 399
 400 static CLASS_ATTR(soft_offline_page, 0644, NULL, store_soft_offline_page);
 401 static CLASS_ATTR(hard_offline_page, 0644, NULL, store_hard_offline_page);
 402
 403 static __init int memory_fail_init(void)
 404 {
 405         int err;
 406
 407         err = sysfs_create_file(&memory_sysdev_class.kset.kobj,
 408                                 &class_attr_soft_offline_page.attr);
 409         if (!err)
 410                 err = sysfs_create_file(&memory_sysdev_class.kset.kobj,
 411                                 &class_attr_hard_offline_page.attr);
 412         return err;
 413 }
 414 #else
 415 static inline int memory_fail_init(void)
 416 {
 417         return 0;
 418 }
 419 #endif
 420
 421 /*
 422  * Note that phys_device is optional.  It is here to allow for
 423  * differentiation between which *physical* devices each
 424  * section belongs to...
 425  */
 426
 427 static int add_memory_block(int nid, struct mem_section *section,
 428                         unsigned long state, int phys_device,
 429                         enum mem_add_context context)
 430 {
 431         struct memory_block *mem = kzalloc(sizeof(*mem), GFP_KERNEL);
 432         int ret = 0;
 433
 434         if (!mem)
 435                 return -ENOMEM;
 436
 437         mem->phys_index = __section_nr(section);
 438         mem->state = state;
 439         mutex_init(&mem->state_mutex);
 440         mem->phys_device = phys_device;
 441
 442         ret = register_memory(mem, section);
 443         if (!ret)
 444                 ret = mem_create_simple_file(mem, phys_index);
 445         if (!ret)
 446                 ret = mem_create_simple_file(mem, state);
 447         if (!ret)
 448                 ret = mem_create_simple_file(mem, phys_device);
 449         if (!ret)
 450                 ret = mem_create_simple_file(mem, removable);
 451         if (!ret) {
 452                 if (context == HOTPLUG)
 453                         ret = register_mem_sect_under_node(mem, nid);
 454         }
 455
 456         return ret;
 457 }
 458
 459 /*
 460  * For now, we have a linear search to go find the appropriate
 461  * memory_block corresponding to a particular phys_index. If
 462  * this gets to be a real problem, we can always use a radix
 463  * tree or something here.
 464  *
 465  * This could be made generic for all sysdev classes.
 466  */
 467 struct memory_block *find_memory_block(struct mem_section *section)
 468 {
 469         struct kobject *kobj;
 470         struct sys_device *sysdev;
 471         struct memory_block *mem;
 472         char name[sizeof(MEMORY_CLASS_NAME) + 9 + 1];
 473
 474         /*
 475          * This only works because we know that section == sysdev->id
 476          * slightly redundant with sysdev_register()
 477          */
 478         sprintf(&name[0], "%s%d", MEMORY_CLASS_NAME, __section_nr(section));
 479
 480         kobj = kset_find_obj(&memory_sysdev_class.kset, name);
 481         if (!kobj)
 482                 return NULL;
 483
 484         sysdev = container_of(kobj, struct sys_device, kobj);
 485         mem = container_of(sysdev, struct memory_block, sysdev);
 486
 487         return mem;
 488 }
 489
 490 int remove_memory_block(unsigned long node_id, struct mem_section *section,
 491                 int phys_device)
 492 {
 493         struct memory_block *mem;
 494
 495         mem = find_memory_block(section);
 496         unregister_mem_sect_under_nodes(mem);
 497         mem_remove_simple_file(mem, phys_index);
 498         mem_remove_simple_file(mem, state);
 499         mem_remove_simple_file(mem, phys_device);
 500         mem_remove_simple_file(mem, removable);
 501         unregister_memory(mem, section);
 502
 503         return 0;
 504 }
 505
 506 /*
 507  * need an interface for the VM to add new memory regions,
 508  * but without onlining it.
 509  */
 510 int register_new_memory(int nid, struct mem_section *section)
 511 {
 512         return add_memory_block(nid, section, MEM_OFFLINE, 0, HOTPLUG);
 513 }
 514
 515 int unregister_memory_section(struct mem_section *section)
 516 {
 517         if (!present_section(section))
 518                 return -EINVAL;
 519
 520         return remove_memory_block(0, section, 0);
 521 }
 522
 523 /*
 524  * Initialize the sysfs support for memory devices...
 525  */
 526 int __init memory_dev_init(void)
 527 {
 528         unsigned int i;
 529         int ret;
 530         int err;
 531
 532         memory_sysdev_class.kset.uevent_ops = &memory_uevent_ops;
 533         ret = sysdev_class_register(&memory_sysdev_class);
 534         if (ret)
 535                 goto out;
 536
 537         /*
 538          * Create entries for memory sections that were found
 539          * during boot and have been initialized
 540          */
 541         for (i = 0; i < NR_MEM_SECTIONS; i++) {
 542                 if (!present_section_nr(i))
 543                         continue;
 544                 err = add_memory_block(0, __nr_to_section(i), MEM_ONLINE,
 545                                         0, BOOT);
 546                 if (!ret)
 547                         ret = err;
 548         }
 549
 550         err = memory_probe_init();
 551         if (!ret)
 552                 ret = err;
 553         err = memory_fail_init();
 554         if (!ret)
 555                 ret = err;
 556         err = block_size_init();
 557         if (!ret)
 558                 ret = err;
 559 out:
 560         if (ret)
 561                 printk(KERN_ERR "%s() failed: %d\n", __func__, ret);
 562         return ret;
 563 }