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 #define MIN_MEMORY_BLOCK_SIZE   (1 << SECTION_SIZE_BITS)
  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 struct sysdev_class memory_sysdev_class = {
  43         .name = MEMORY_CLASS_NAME,
  44 };
  45
  46 static const char *memory_uevent_name(struct kset *kset, struct kobject *kobj)
  47 {
  48         return MEMORY_CLASS_NAME;
  49 }
  50
  51 static int memory_uevent(struct kset *kset, struct kobject *obj,
  52                         struct kobj_uevent_env *env)
  53 {
  54         int retval = 0;
  55
  56         return retval;
  57 }
  58
  59 static const struct kset_uevent_ops memory_uevent_ops = {
  60         .name           = memory_uevent_name,
  61         .uevent         = memory_uevent,
  62 };
  63
  64 static BLOCKING_NOTIFIER_HEAD(memory_chain);
  65
  66 int register_memory_notifier(struct notifier_block *nb)
  67 {
  68         return blocking_notifier_chain_register(&memory_chain, nb);
  69 }
  70 EXPORT_SYMBOL(register_memory_notifier);
  71
  72 void unregister_memory_notifier(struct notifier_block *nb)
  73 {
  74         blocking_notifier_chain_unregister(&memory_chain, nb);
  75 }
  76 EXPORT_SYMBOL(unregister_memory_notifier);
  77
  78 static ATOMIC_NOTIFIER_HEAD(memory_isolate_chain);
  79
  80 int register_memory_isolate_notifier(struct notifier_block *nb)
  81 {
  82         return atomic_notifier_chain_register(&memory_isolate_chain, nb);
  83 }
  84 EXPORT_SYMBOL(register_memory_isolate_notifier);
  85
  86 void unregister_memory_isolate_notifier(struct notifier_block *nb)
  87 {
  88         atomic_notifier_chain_unregister(&memory_isolate_chain, nb);
  89 }
  90 EXPORT_SYMBOL(unregister_memory_isolate_notifier);
  91
  92 /*
  93  * register_memory - Setup a sysfs device for a memory block
  94  */
  95 static
  96 int register_memory(struct memory_block *memory)
  97 {
  98         int error;
  99
 100         memory->sysdev.cls = &memory_sysdev_class;
 101         memory->sysdev.id = memory->start_section_nr / sections_per_block;
 102
 103         error = sysdev_register(&memory->sysdev);
 104         return error;
 105 }
 106
 107 static void
 108 unregister_memory(struct memory_block *memory)
 109 {
 110         BUG_ON(memory->sysdev.cls != &memory_sysdev_class);
 111
 112         /* drop the ref. we got in remove_memory_block() */
 113         kobject_put(&memory->sysdev.kobj);
 114         sysdev_unregister(&memory->sysdev);
 115 }
 116
 117 unsigned long __weak memory_block_size_bytes(void)
 118 {
 119         return MIN_MEMORY_BLOCK_SIZE;
 120 }
 121
 122 static unsigned long get_memory_block_size(void)
 123 {
 124         unsigned long block_sz;
 125
 126         block_sz = memory_block_size_bytes();
 127
 128         /* Validate blk_sz is a power of 2 and not less than section size */
 129         if ((block_sz & (block_sz - 1)) || (block_sz < MIN_MEMORY_BLOCK_SIZE)) {
 130                 WARN_ON(1);
 131                 block_sz = MIN_MEMORY_BLOCK_SIZE;
 132         }
 133
 134         return block_sz;
 135 }
 136
 137 /*
 138  * use this as the physical section index that this memsection
 139  * uses.
 140  */
 141
 142 static ssize_t show_mem_start_phys_index(struct sys_device *dev,
 143                         struct sysdev_attribute *attr, char *buf)
 144 {
 145         struct memory_block *mem =
 146                 container_of(dev, struct memory_block, sysdev);
 147         unsigned long phys_index;
 148
 149         phys_index = mem->start_section_nr / sections_per_block;
 150         return sprintf(buf, "%08lx\n", phys_index);
 151 }
 152
 153 static ssize_t show_mem_end_phys_index(struct sys_device *dev,
 154                         struct sysdev_attribute *attr, char *buf)
 155 {
 156         struct memory_block *mem =
 157                 container_of(dev, struct memory_block, sysdev);
 158         unsigned long phys_index;
 159
 160         phys_index = mem->end_section_nr / sections_per_block;
 161         return sprintf(buf, "%08lx\n", phys_index);
 162 }
 163
 164 /*
 165  * Show whether the section of memory is likely to be hot-removable
 166  */
 167 static ssize_t show_mem_removable(struct sys_device *dev,
 168                         struct sysdev_attribute *attr, char *buf)
 169 {
 170         unsigned long i, pfn;
 171         int ret = 1;
 172         struct memory_block *mem =
 173                 container_of(dev, struct memory_block, sysdev);
 174
 175         for (i = 0; i < sections_per_block; i++) {
 176                 pfn = section_nr_to_pfn(mem->start_section_nr + i);
 177                 ret &= is_mem_section_removable(pfn, PAGES_PER_SECTION);
 178         }
 179
 180         return sprintf(buf, "%d\n", ret);
 181 }
 182
 183 /*
 184  * online, offline, going offline, etc.
 185  */
 186 static ssize_t show_mem_state(struct sys_device *dev,
 187                         struct sysdev_attribute *attr, char *buf)
 188 {
 189         struct memory_block *mem =
 190                 container_of(dev, struct memory_block, sysdev);
 191         ssize_t len = 0;
 192
 193         /*
 194          * We can probably put these states in a nice little array
 195          * so that they're not open-coded
 196          */
 197         switch (mem->state) {
 198                 case MEM_ONLINE:
 199                         len = sprintf(buf, "online\n");
 200                         break;
 201                 case MEM_OFFLINE:
 202                         len = sprintf(buf, "offline\n");
 203                         break;
 204                 case MEM_GOING_OFFLINE:
 205                         len = sprintf(buf, "going-offline\n");
 206                         break;
 207                 default:
 208                         len = sprintf(buf, "ERROR-UNKNOWN-%ld\n",
 209                                         mem->state);
 210                         WARN_ON(1);
 211                         break;
 212         }
 213
 214         return len;
 215 }
 216
 217 int memory_notify(unsigned long val, void *v)
 218 {
 219         return blocking_notifier_call_chain(&memory_chain, val, v);
 220 }
 221
 222 int memory_isolate_notify(unsigned long val, void *v)
 223 {
 224         return atomic_notifier_call_chain(&memory_isolate_chain, val, v);
 225 }
 226
 227 /*
 228  * MEMORY_HOTPLUG depends on SPARSEMEM in mm/Kconfig, so it is
 229  * OK to have direct references to sparsemem variables in here.
 230  */
 231 static int
 232 memory_block_action(unsigned long phys_index, unsigned long action)
 233 {
 234         int i;
 235         unsigned long start_pfn, start_paddr;
 236         unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
 237         struct page *first_page;
 238         int ret;
 239
 240         first_page = pfn_to_page(phys_index << PFN_SECTION_SHIFT);
 241
 242         /*
 243          * The probe routines leave the pages reserved, just
 244          * as the bootmem code does.  Make sure they're still
 245          * that way.
 246          */
 247         if (action == MEM_ONLINE) {
 248                 for (i = 0; i < nr_pages; i++) {
 249                         if (PageReserved(first_page+i))
 250                                 continue;
 251
 252                         printk(KERN_WARNING "section number %ld page number %d "
 253                                 "not reserved, was it already online?\n",
 254                                 phys_index, i);
 255                         return -EBUSY;
 256                 }
 257         }
 258
 259         switch (action) {
 260                 case MEM_ONLINE:
 261                         start_pfn = page_to_pfn(first_page);
 262                         ret = online_pages(start_pfn, nr_pages);
 263                         break;
 264                 case MEM_OFFLINE:
 265                         start_paddr = page_to_pfn(first_page) << PAGE_SHIFT;
 266                         ret = remove_memory(start_paddr,
 267                                             nr_pages << PAGE_SHIFT);
 268                         break;
 269                 default:
 270                         WARN(1, KERN_WARNING "%s(%ld, %ld) unknown action: "
 271                              "%ld\n", __func__, phys_index, action, action);
 272                         ret = -EINVAL;
 273         }
 274
 275         return ret;
 276 }
 277
 278 static int memory_block_change_state(struct memory_block *mem,
 279                 unsigned long to_state, unsigned long from_state_req)
 280 {
 281         int ret = 0;
 282
 283         mutex_lock(&mem->state_mutex);
 284
 285         if (mem->state != from_state_req) {
 286                 ret = -EINVAL;
 287                 goto out;
 288         }
 289
 290         if (to_state == MEM_OFFLINE)
 291                 mem->state = MEM_GOING_OFFLINE;
 292
 293         ret = memory_block_action(mem->start_section_nr, to_state);
 294
 295         if (ret)
 296                 mem->state = from_state_req;
 297         else
 298                 mem->state = to_state;
 299
 300 out:
 301         mutex_unlock(&mem->state_mutex);
 302         return ret;
 303 }
 304
 305 static ssize_t
 306 store_mem_state(struct sys_device *dev,
 307                 struct sysdev_attribute *attr, const char *buf, size_t count)
 308 {
 309         struct memory_block *mem;
 310         int ret = -EINVAL;
 311
 312         mem = container_of(dev, struct memory_block, sysdev);
 313
 314         if (!strncmp(buf, "online", min((int)count, 6)))
 315                 ret = memory_block_change_state(mem, MEM_ONLINE, MEM_OFFLINE);
 316         else if(!strncmp(buf, "offline", min((int)count, 7)))
 317                 ret = memory_block_change_state(mem, MEM_OFFLINE, MEM_ONLINE);
 318
 319         if (ret)
 320                 return ret;
 321         return count;
 322 }
 323
 324 /*
 325  * phys_device is a bad name for this.  What I really want
 326  * is a way to differentiate between memory ranges that
 327  * are part of physical devices that constitute
 328  * a complete removable unit or fru.
 329  * i.e. do these ranges belong to the same physical device,
 330  * s.t. if I offline all of these sections I can then
 331  * remove the physical device?
 332  */
 333 static ssize_t show_phys_device(struct sys_device *dev,
 334                                 struct sysdev_attribute *attr, char *buf)
 335 {
 336         struct memory_block *mem =
 337                 container_of(dev, struct memory_block, sysdev);
 338         return sprintf(buf, "%d\n", mem->phys_device);
 339 }
 340
 341 static SYSDEV_ATTR(phys_index, 0444, show_mem_start_phys_index, NULL);
 342 static SYSDEV_ATTR(end_phys_index, 0444, show_mem_end_phys_index, NULL);
 343 static SYSDEV_ATTR(state, 0644, show_mem_state, store_mem_state);
 344 static SYSDEV_ATTR(phys_device, 0444, show_phys_device, NULL);
 345 static SYSDEV_ATTR(removable, 0444, show_mem_removable, NULL);
 346
 347 #define mem_create_simple_file(mem, attr_name)  \
 348         sysdev_create_file(&mem->sysdev, &attr_##attr_name)
 349 #define mem_remove_simple_file(mem, attr_name)  \
 350         sysdev_remove_file(&mem->sysdev, &attr_##attr_name)
 351
 352 /*
 353  * Block size attribute stuff
 354  */
 355 static ssize_t
 356 print_block_size(struct sysdev_class *class, struct sysdev_class_attribute *attr,
 357                  char *buf)
 358 {
 359         return sprintf(buf, "%lx\n", get_memory_block_size());
 360 }
 361
 362 static SYSDEV_CLASS_ATTR(block_size_bytes, 0444, print_block_size, NULL);
 363
 364 static int block_size_init(void)
 365 {
 366         return sysfs_create_file(&memory_sysdev_class.kset.kobj,
 367                                 &attr_block_size_bytes.attr);
 368 }
 369
 370 /*
 371  * Some architectures will have custom drivers to do this, and
 372  * will not need to do it from userspace.  The fake hot-add code
 373  * as well as ppc64 will do all of their discovery in userspace
 374  * and will require this interface.
 375  */
 376 #ifdef CONFIG_ARCH_MEMORY_PROBE
 377 static ssize_t
 378 memory_probe_store(struct class *class, struct class_attribute *attr,
 379                    const char *buf, size_t count)
 380 {
 381         u64 phys_addr;
 382         int nid;
 383         int i, ret;
 384
 385         phys_addr = simple_strtoull(buf, NULL, 0);
 386
 387         for (i = 0; i < sections_per_block; i++) {
 388                 nid = memory_add_physaddr_to_nid(phys_addr);
 389                 ret = add_memory(nid, phys_addr,
 390                                  PAGES_PER_SECTION << PAGE_SHIFT);
 391                 if (ret)
 392                         goto out;
 393
 394                 phys_addr += MIN_MEMORY_BLOCK_SIZE;
 395         }
 396
 397         ret = count;
 398 out:
 399         return ret;
 400 }
 401 static CLASS_ATTR(probe, S_IWUSR, NULL, memory_probe_store);
 402
 403 static int memory_probe_init(void)
 404 {
 405         return sysfs_create_file(&memory_sysdev_class.kset.kobj,
 406                                 &class_attr_probe.attr);
 407 }
 408 #else
 409 static inline int memory_probe_init(void)
 410 {
 411         return 0;
 412 }
 413 #endif
 414
 415 #ifdef CONFIG_MEMORY_FAILURE
 416 /*
 417  * Support for offlining pages of memory
 418  */
 419
 420 /* Soft offline a page */
 421 static ssize_t
 422 store_soft_offline_page(struct class *class,
 423                         struct class_attribute *attr,
 424                         const char *buf, size_t count)
 425 {
 426         int ret;
 427         u64 pfn;
 428         if (!capable(CAP_SYS_ADMIN))
 429                 return -EPERM;
 430         if (strict_strtoull(buf, 0, &pfn) < 0)
 431                 return -EINVAL;
 432         pfn >>= PAGE_SHIFT;
 433         if (!pfn_valid(pfn))
 434                 return -ENXIO;
 435         ret = soft_offline_page(pfn_to_page(pfn), 0);
 436         return ret == 0 ? count : ret;
 437 }
 438
 439 /* Forcibly offline a page, including killing processes. */
 440 static ssize_t
 441 store_hard_offline_page(struct class *class,
 442                         struct class_attribute *attr,
 443                         const char *buf, size_t count)
 444 {
 445         int ret;
 446         u64 pfn;
 447         if (!capable(CAP_SYS_ADMIN))
 448                 return -EPERM;
 449         if (strict_strtoull(buf, 0, &pfn) < 0)
 450                 return -EINVAL;
 451         pfn >>= PAGE_SHIFT;
 452         ret = __memory_failure(pfn, 0, 0);
 453         return ret ? ret : count;
 454 }
 455
 456 static CLASS_ATTR(soft_offline_page, 0644, NULL, store_soft_offline_page);
 457 static CLASS_ATTR(hard_offline_page, 0644, NULL, store_hard_offline_page);
 458
 459 static __init int memory_fail_init(void)
 460 {
 461         int err;
 462
 463         err = sysfs_create_file(&memory_sysdev_class.kset.kobj,
 464                                 &class_attr_soft_offline_page.attr);
 465         if (!err)
 466                 err = sysfs_create_file(&memory_sysdev_class.kset.kobj,
 467                                 &class_attr_hard_offline_page.attr);
 468         return err;
 469 }
 470 #else
 471 static inline int memory_fail_init(void)
 472 {
 473         return 0;
 474 }
 475 #endif
 476
 477 /*
 478  * Note that phys_device is optional.  It is here to allow for
 479  * differentiation between which *physical* devices each
 480  * section belongs to...
 481  */
 482 int __weak arch_get_memory_phys_device(unsigned long start_pfn)
 483 {
 484         return 0;
 485 }
 486
 487 struct memory_block *find_memory_block_hinted(struct mem_section *section,
 488                                               struct memory_block *hint)
 489 {
 490         struct kobject *kobj;
 491         struct sys_device *sysdev;
 492         struct memory_block *mem;
 493         char name[sizeof(MEMORY_CLASS_NAME) + 9 + 1];
 494         int block_id = base_memory_block_id(__section_nr(section));
 495
 496         kobj = hint ? &hint->sysdev.kobj : NULL;
 497
 498         /*
 499          * This only works because we know that section == sysdev->id
 500          * slightly redundant with sysdev_register()
 501          */
 502         sprintf(&name[0], "%s%d", MEMORY_CLASS_NAME, block_id);
 503
 504         kobj = kset_find_obj_hinted(&memory_sysdev_class.kset, name, kobj);
 505         if (!kobj)
 506                 return NULL;
 507
 508         sysdev = container_of(kobj, struct sys_device, kobj);
 509         mem = container_of(sysdev, struct memory_block, sysdev);
 510
 511         return mem;
 512 }
 513
 514 /*
 515  * For now, we have a linear search to go find the appropriate
 516  * memory_block corresponding to a particular phys_index. If
 517  * this gets to be a real problem, we can always use a radix
 518  * tree or something here.
 519  *
 520  * This could be made generic for all sysdev classes.
 521  */
 522 struct memory_block *find_memory_block(struct mem_section *section)
 523 {
 524         return find_memory_block_hinted(section, NULL);
 525 }
 526
 527 static int init_memory_block(struct memory_block **memory,
 528                              struct mem_section *section, unsigned long state)
 529 {
 530         struct memory_block *mem;
 531         unsigned long start_pfn;
 532         int scn_nr;
 533         int ret = 0;
 534
 535         mem = kzalloc(sizeof(*mem), GFP_KERNEL);
 536         if (!mem)
 537                 return -ENOMEM;
 538
 539         scn_nr = __section_nr(section);
 540         mem->start_section_nr =
 541                         base_memory_block_id(scn_nr) * sections_per_block;
 542         mem->end_section_nr = mem->start_section_nr + sections_per_block - 1;
 543         mem->state = state;
 544         mem->section_count++;
 545         mutex_init(&mem->state_mutex);
 546         start_pfn = section_nr_to_pfn(mem->start_section_nr);
 547         mem->phys_device = arch_get_memory_phys_device(start_pfn);
 548
 549         ret = register_memory(mem);
 550         if (!ret)
 551                 ret = mem_create_simple_file(mem, phys_index);
 552         if (!ret)
 553                 ret = mem_create_simple_file(mem, end_phys_index);
 554         if (!ret)
 555                 ret = mem_create_simple_file(mem, state);
 556         if (!ret)
 557                 ret = mem_create_simple_file(mem, phys_device);
 558         if (!ret)
 559                 ret = mem_create_simple_file(mem, removable);
 560
 561         *memory = mem;
 562         return ret;
 563 }
 564
 565 static int add_memory_section(int nid, struct mem_section *section,
 566                         unsigned long state, enum mem_add_context context)
 567 {
 568         struct memory_block *mem;
 569         int ret = 0;
 570
 571         mutex_lock(&mem_sysfs_mutex);
 572
 573         mem = find_memory_block(section);
 574         if (mem) {
 575                 mem->section_count++;
 576                 kobject_put(&mem->sysdev.kobj);
 577         } else
 578                 ret = init_memory_block(&mem, section, state);
 579
 580         if (!ret) {
 581                 if (context == HOTPLUG &&
 582                     mem->section_count == sections_per_block)
 583                         ret = register_mem_sect_under_node(mem, nid);
 584         }
 585
 586         mutex_unlock(&mem_sysfs_mutex);
 587         return ret;
 588 }
 589
 590 int remove_memory_block(unsigned long node_id, struct mem_section *section,
 591                 int phys_device)
 592 {
 593         struct memory_block *mem;
 594
 595         mutex_lock(&mem_sysfs_mutex);
 596         mem = find_memory_block(section);
 597         unregister_mem_sect_under_nodes(mem, __section_nr(section));
 598
 599         mem->section_count--;
 600         if (mem->section_count == 0) {
 601                 mem_remove_simple_file(mem, phys_index);
 602                 mem_remove_simple_file(mem, end_phys_index);
 603                 mem_remove_simple_file(mem, state);
 604                 mem_remove_simple_file(mem, phys_device);
 605                 mem_remove_simple_file(mem, removable);
 606                 unregister_memory(mem);
 607                 kfree(mem);
 608         } else
 609                 kobject_put(&mem->sysdev.kobj);
 610
 611         mutex_unlock(&mem_sysfs_mutex);
 612         return 0;
 613 }
 614
 615 /*
 616  * need an interface for the VM to add new memory regions,
 617  * but without onlining it.
 618  */
 619 int register_new_memory(int nid, struct mem_section *section)
 620 {
 621         return add_memory_section(nid, section, MEM_OFFLINE, HOTPLUG);
 622 }
 623
 624 int unregister_memory_section(struct mem_section *section)
 625 {
 626         if (!present_section(section))
 627                 return -EINVAL;
 628
 629         return remove_memory_block(0, section, 0);
 630 }
 631
 632 /*
 633  * Initialize the sysfs support for memory devices...
 634  */
 635 int __init memory_dev_init(void)
 636 {
 637         unsigned int i;
 638         int ret;
 639         int err;
 640         unsigned long block_sz;
 641
 642         memory_sysdev_class.kset.uevent_ops = &memory_uevent_ops;
 643         ret = sysdev_class_register(&memory_sysdev_class);
 644         if (ret)
 645                 goto out;
 646
 647         block_sz = get_memory_block_size();
 648         sections_per_block = block_sz / MIN_MEMORY_BLOCK_SIZE;
 649
 650         /*
 651          * Create entries for memory sections that were found
 652          * during boot and have been initialized
 653          */
 654         for (i = 0; i < NR_MEM_SECTIONS; i++) {
 655                 if (!present_section_nr(i))
 656                         continue;
 657                 err = add_memory_section(0, __nr_to_section(i), MEM_ONLINE,
 658                                          BOOT);
 659                 if (!ret)
 660                         ret = err;
 661         }
 662
 663         err = memory_probe_init();
 664         if (!ret)
 665                 ret = err;
 666         err = memory_fail_init();
 667         if (!ret)
 668                 ret = err;
 669         err = block_size_init();
 670         if (!ret)
 671                 ret = err;
 672 out:
 673         if (ret)
 674                 printk(KERN_ERR "%s() failed: %d\n", __func__, ret);
 675         return ret;
 676 }