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  * MEMORY_HOTPLUG depends on SPARSEMEM in mm/Kconfig, so it is
 228  * OK to have direct references to sparsemem variables in here.
 229  */
 230 static int
 231 memory_block_action(unsigned long phys_index, unsigned long action)
 232 {
 233         int i;
 234         unsigned long start_pfn, start_paddr;
 235         unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
 236         struct page *first_page;
 237         int ret;
 238
 239         first_page = pfn_to_page(phys_index << PFN_SECTION_SHIFT);
 240
 241         /*
 242          * The probe routines leave the pages reserved, just
 243          * as the bootmem code does.  Make sure they're still
 244          * that way.
 245          */
 246         if (action == MEM_ONLINE) {
 247                 for (i = 0; i < nr_pages; i++) {
 248                         if (PageReserved(first_page+i))
 249                                 continue;
 250
 251                         printk(KERN_WARNING "section number %ld page number %d "
 252                                 "not reserved, was it already online?\n",
 253                                 phys_index, i);
 254                         return -EBUSY;
 255                 }
 256         }
 257
 258         switch (action) {
 259                 case MEM_ONLINE:
 260                         start_pfn = page_to_pfn(first_page);
 261                         ret = online_pages(start_pfn, nr_pages);
 262                         break;
 263                 case MEM_OFFLINE:
 264                         start_paddr = page_to_pfn(first_page) << PAGE_SHIFT;
 265                         ret = remove_memory(start_paddr,
 266                                             nr_pages << PAGE_SHIFT);
 267                         break;
 268                 default:
 269                         WARN(1, KERN_WARNING "%s(%ld, %ld) unknown action: "
 270                              "%ld\n", __func__, phys_index, action, action);
 271                         ret = -EINVAL;
 272         }
 273
 274         return ret;
 275 }
 276
 277 static int memory_block_change_state(struct memory_block *mem,
 278                 unsigned long to_state, unsigned long from_state_req)
 279 {
 280         int ret = 0;
 281
 282         mutex_lock(&mem->state_mutex);
 283
 284         if (mem->state != from_state_req) {
 285                 ret = -EINVAL;
 286                 goto out;
 287         }
 288
 289         if (to_state == MEM_OFFLINE)
 290                 mem->state = MEM_GOING_OFFLINE;
 291
 292         ret = memory_block_action(mem->start_section_nr, to_state);
 293
 294         if (ret)
 295                 mem->state = from_state_req;
 296         else
 297                 mem->state = to_state;
 298
 299 out:
 300         mutex_unlock(&mem->state_mutex);
 301         return ret;
 302 }
 303
 304 static ssize_t
 305 store_mem_state(struct sys_device *dev,
 306                 struct sysdev_attribute *attr, const char *buf, size_t count)
 307 {
 308         struct memory_block *mem;
 309         int ret = -EINVAL;
 310
 311         mem = container_of(dev, struct memory_block, sysdev);
 312
 313         if (!strncmp(buf, "online", min((int)count, 6)))
 314                 ret = memory_block_change_state(mem, MEM_ONLINE, MEM_OFFLINE);
 315         else if(!strncmp(buf, "offline", min((int)count, 7)))
 316                 ret = memory_block_change_state(mem, MEM_OFFLINE, MEM_ONLINE);
 317
 318         if (ret)
 319                 return ret;
 320         return count;
 321 }
 322
 323 /*
 324  * phys_device is a bad name for this.  What I really want
 325  * is a way to differentiate between memory ranges that
 326  * are part of physical devices that constitute
 327  * a complete removable unit or fru.
 328  * i.e. do these ranges belong to the same physical device,
 329  * s.t. if I offline all of these sections I can then
 330  * remove the physical device?
 331  */
 332 static ssize_t show_phys_device(struct sys_device *dev,
 333                                 struct sysdev_attribute *attr, char *buf)
 334 {
 335         struct memory_block *mem =
 336                 container_of(dev, struct memory_block, sysdev);
 337         return sprintf(buf, "%d\n", mem->phys_device);
 338 }
 339
 340 static SYSDEV_ATTR(phys_index, 0444, show_mem_start_phys_index, NULL);
 341 static SYSDEV_ATTR(end_phys_index, 0444, show_mem_end_phys_index, NULL);
 342 static SYSDEV_ATTR(state, 0644, show_mem_state, store_mem_state);
 343 static SYSDEV_ATTR(phys_device, 0444, show_phys_device, NULL);
 344 static SYSDEV_ATTR(removable, 0444, show_mem_removable, NULL);
 345
 346 #define mem_create_simple_file(mem, attr_name)  \
 347         sysdev_create_file(&mem->sysdev, &attr_##attr_name)
 348 #define mem_remove_simple_file(mem, attr_name)  \
 349         sysdev_remove_file(&mem->sysdev, &attr_##attr_name)
 350
 351 /*
 352  * Block size attribute stuff
 353  */
 354 static ssize_t
 355 print_block_size(struct sysdev_class *class, struct sysdev_class_attribute *attr,
 356                  char *buf)
 357 {
 358         return sprintf(buf, "%lx\n", get_memory_block_size());
 359 }
 360
 361 static SYSDEV_CLASS_ATTR(block_size_bytes, 0444, print_block_size, NULL);
 362
 363 static int block_size_init(void)
 364 {
 365         return sysfs_create_file(&memory_sysdev_class.kset.kobj,
 366                                 &attr_block_size_bytes.attr);
 367 }
 368
 369 /*
 370  * Some architectures will have custom drivers to do this, and
 371  * will not need to do it from userspace.  The fake hot-add code
 372  * as well as ppc64 will do all of their discovery in userspace
 373  * and will require this interface.
 374  */
 375 #ifdef CONFIG_ARCH_MEMORY_PROBE
 376 static ssize_t
 377 memory_probe_store(struct class *class, struct class_attribute *attr,
 378                    const char *buf, size_t count)
 379 {
 380         u64 phys_addr;
 381         int nid;
 382         int i, ret;
 383
 384         phys_addr = simple_strtoull(buf, NULL, 0);
 385
 386         for (i = 0; i < sections_per_block; i++) {
 387                 nid = memory_add_physaddr_to_nid(phys_addr);
 388                 ret = add_memory(nid, phys_addr,
 389                                  PAGES_PER_SECTION << PAGE_SHIFT);
 390                 if (ret)
 391                         goto out;
 392
 393                 phys_addr += MIN_MEMORY_BLOCK_SIZE;
 394         }
 395
 396         ret = count;
 397 out:
 398         return ret;
 399 }
 400 static CLASS_ATTR(probe, S_IWUSR, NULL, memory_probe_store);
 401
 402 static int memory_probe_init(void)
 403 {
 404         return sysfs_create_file(&memory_sysdev_class.kset.kobj,
 405                                 &class_attr_probe.attr);
 406 }
 407 #else
 408 static inline int memory_probe_init(void)
 409 {
 410         return 0;
 411 }
 412 #endif
 413
 414 #ifdef CONFIG_MEMORY_FAILURE
 415 /*
 416  * Support for offlining pages of memory
 417  */
 418
 419 /* Soft offline a page */
 420 static ssize_t
 421 store_soft_offline_page(struct class *class,
 422                         struct class_attribute *attr,
 423                         const char *buf, size_t count)
 424 {
 425         int ret;
 426         u64 pfn;
 427         if (!capable(CAP_SYS_ADMIN))
 428                 return -EPERM;
 429         if (strict_strtoull(buf, 0, &pfn) < 0)
 430                 return -EINVAL;
 431         pfn >>= PAGE_SHIFT;
 432         if (!pfn_valid(pfn))
 433                 return -ENXIO;
 434         ret = soft_offline_page(pfn_to_page(pfn), 0);
 435         return ret == 0 ? count : ret;
 436 }
 437
 438 /* Forcibly offline a page, including killing processes. */
 439 static ssize_t
 440 store_hard_offline_page(struct class *class,
 441                         struct class_attribute *attr,
 442                         const char *buf, size_t count)
 443 {
 444         int ret;
 445         u64 pfn;
 446         if (!capable(CAP_SYS_ADMIN))
 447                 return -EPERM;
 448         if (strict_strtoull(buf, 0, &pfn) < 0)
 449                 return -EINVAL;
 450         pfn >>= PAGE_SHIFT;
 451         ret = __memory_failure(pfn, 0, 0);
 452         return ret ? ret : count;
 453 }
 454
 455 static CLASS_ATTR(soft_offline_page, 0644, NULL, store_soft_offline_page);
 456 static CLASS_ATTR(hard_offline_page, 0644, NULL, store_hard_offline_page);
 457
 458 static __init int memory_fail_init(void)
 459 {
 460         int err;
 461
 462         err = sysfs_create_file(&memory_sysdev_class.kset.kobj,
 463                                 &class_attr_soft_offline_page.attr);
 464         if (!err)
 465                 err = sysfs_create_file(&memory_sysdev_class.kset.kobj,
 466                                 &class_attr_hard_offline_page.attr);
 467         return err;
 468 }
 469 #else
 470 static inline int memory_fail_init(void)
 471 {
 472         return 0;
 473 }
 474 #endif
 475
 476 /*
 477  * Note that phys_device is optional.  It is here to allow for
 478  * differentiation between which *physical* devices each
 479  * section belongs to...
 480  */
 481 int __weak arch_get_memory_phys_device(unsigned long start_pfn)
 482 {
 483         return 0;
 484 }
 485
 486 struct memory_block *find_memory_block_hinted(struct mem_section *section,
 487                                               struct memory_block *hint)
 488 {
 489         struct kobject *kobj;
 490         struct sys_device *sysdev;
 491         struct memory_block *mem;
 492         char name[sizeof(MEMORY_CLASS_NAME) + 9 + 1];
 493         int block_id = base_memory_block_id(__section_nr(section));
 494
 495         kobj = hint ? &hint->sysdev.kobj : NULL;
 496
 497         /*
 498          * This only works because we know that section == sysdev->id
 499          * slightly redundant with sysdev_register()
 500          */
 501         sprintf(&name[0], "%s%d", MEMORY_CLASS_NAME, block_id);
 502
 503         kobj = kset_find_obj_hinted(&memory_sysdev_class.kset, name, kobj);
 504         if (!kobj)
 505                 return NULL;
 506
 507         sysdev = container_of(kobj, struct sys_device, kobj);
 508         mem = container_of(sysdev, struct memory_block, sysdev);
 509
 510         return mem;
 511 }
 512
 513 /*
 514  * For now, we have a linear search to go find the appropriate
 515  * memory_block corresponding to a particular phys_index. If
 516  * this gets to be a real problem, we can always use a radix
 517  * tree or something here.
 518  *
 519  * This could be made generic for all sysdev classes.
 520  */
 521 struct memory_block *find_memory_block(struct mem_section *section)
 522 {
 523         return find_memory_block_hinted(section, NULL);
 524 }
 525
 526 static int init_memory_block(struct memory_block **memory,
 527                              struct mem_section *section, unsigned long state)
 528 {
 529         struct memory_block *mem;
 530         unsigned long start_pfn;
 531         int scn_nr;
 532         int ret = 0;
 533
 534         mem = kzalloc(sizeof(*mem), GFP_KERNEL);
 535         if (!mem)
 536                 return -ENOMEM;
 537
 538         scn_nr = __section_nr(section);
 539         mem->start_section_nr =
 540                         base_memory_block_id(scn_nr) * sections_per_block;
 541         mem->end_section_nr = mem->start_section_nr + sections_per_block - 1;
 542         mem->state = state;
 543         mem->section_count++;
 544         mutex_init(&mem->state_mutex);
 545         start_pfn = section_nr_to_pfn(mem->start_section_nr);
 546         mem->phys_device = arch_get_memory_phys_device(start_pfn);
 547
 548         ret = register_memory(mem);
 549         if (!ret)
 550                 ret = mem_create_simple_file(mem, phys_index);
 551         if (!ret)
 552                 ret = mem_create_simple_file(mem, end_phys_index);
 553         if (!ret)
 554                 ret = mem_create_simple_file(mem, state);
 555         if (!ret)
 556                 ret = mem_create_simple_file(mem, phys_device);
 557         if (!ret)
 558                 ret = mem_create_simple_file(mem, removable);
 559
 560         *memory = mem;
 561         return ret;
 562 }
 563
 564 static int add_memory_section(int nid, struct mem_section *section,
 565                         unsigned long state, enum mem_add_context context)
 566 {
 567         struct memory_block *mem;
 568         int ret = 0;
 569
 570         mutex_lock(&mem_sysfs_mutex);
 571
 572         mem = find_memory_block(section);
 573         if (mem) {
 574                 mem->section_count++;
 575                 kobject_put(&mem->sysdev.kobj);
 576         } else
 577                 ret = init_memory_block(&mem, section, state);
 578
 579         if (!ret) {
 580                 if (context == HOTPLUG &&
 581                     mem->section_count == sections_per_block)
 582                         ret = register_mem_sect_under_node(mem, nid);
 583         }
 584
 585         mutex_unlock(&mem_sysfs_mutex);
 586         return ret;
 587 }
 588
 589 int remove_memory_block(unsigned long node_id, struct mem_section *section,
 590                 int phys_device)
 591 {
 592         struct memory_block *mem;
 593
 594         mutex_lock(&mem_sysfs_mutex);
 595         mem = find_memory_block(section);
 596         unregister_mem_sect_under_nodes(mem, __section_nr(section));
 597
 598         mem->section_count--;
 599         if (mem->section_count == 0) {
 600                 mem_remove_simple_file(mem, phys_index);
 601                 mem_remove_simple_file(mem, end_phys_index);
 602                 mem_remove_simple_file(mem, state);
 603                 mem_remove_simple_file(mem, phys_device);
 604                 mem_remove_simple_file(mem, removable);
 605                 unregister_memory(mem);
 606                 kfree(mem);
 607         } else
 608                 kobject_put(&mem->sysdev.kobj);
 609
 610         mutex_unlock(&mem_sysfs_mutex);
 611         return 0;
 612 }
 613
 614 /*
 615  * need an interface for the VM to add new memory regions,
 616  * but without onlining it.
 617  */
 618 int register_new_memory(int nid, struct mem_section *section)
 619 {
 620         return add_memory_section(nid, section, MEM_OFFLINE, HOTPLUG);
 621 }
 622
 623 int unregister_memory_section(struct mem_section *section)
 624 {
 625         if (!present_section(section))
 626                 return -EINVAL;
 627
 628         return remove_memory_block(0, section, 0);
 629 }
 630
 631 /*
 632  * Initialize the sysfs support for memory devices...
 633  */
 634 int __init memory_dev_init(void)
 635 {
 636         unsigned int i;
 637         int ret;
 638         int err;
 639         unsigned long block_sz;
 640
 641         memory_sysdev_class.kset.uevent_ops = &memory_uevent_ops;
 642         ret = sysdev_class_register(&memory_sysdev_class);
 643         if (ret)
 644                 goto out;
 645
 646         block_sz = get_memory_block_size();
 647         sections_per_block = block_sz / MIN_MEMORY_BLOCK_SIZE;
 648
 649         /*
 650          * Create entries for memory sections that were found
 651          * during boot and have been initialized
 652          */
 653         for (i = 0; i < NR_MEM_SECTIONS; i++) {
 654                 if (!present_section_nr(i))
 655                         continue;
 656                 err = add_memory_section(0, __nr_to_section(i), MEM_ONLINE,
 657                                          BOOT);
 658                 if (!ret)
 659                         ret = err;
 660         }
 661
 662         err = memory_probe_init();
 663         if (!ret)
 664                 ret = err;
 665         err = memory_fail_init();
 666         if (!ret)
 667                 ret = err;
 668         err = block_size_init();
 669         if (!ret)
 670                 ret = err;
 671 out:
 672         if (ret)
 673                 printk(KERN_ERR "%s() failed: %d\n", __func__, ret);
 674         return ret;
 675 }