linux/drivers/base/memory.c
<<
>>
Prefs
   1/*
   2 * Memory subsystem 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/module.h>
  14#include <linux/init.h>
  15#include <linux/topology.h>
  16#include <linux/capability.h>
  17#include <linux/device.h>
  18#include <linux/memory.h>
  19#include <linux/kobject.h>
  20#include <linux/memory_hotplug.h>
  21#include <linux/mm.h>
  22#include <linux/mutex.h>
  23#include <linux/stat.h>
  24#include <linux/slab.h>
  25
  26#include <linux/atomic.h>
  27#include <asm/uaccess.h>
  28
  29static DEFINE_MUTEX(mem_sysfs_mutex);
  30
  31#define MEMORY_CLASS_NAME       "memory"
  32
  33static int sections_per_block;
  34
  35static inline int base_memory_block_id(int section_nr)
  36{
  37        return section_nr / sections_per_block;
  38}
  39
  40static struct bus_type memory_subsys = {
  41        .name = MEMORY_CLASS_NAME,
  42        .dev_name = MEMORY_CLASS_NAME,
  43};
  44
  45static BLOCKING_NOTIFIER_HEAD(memory_chain);
  46
  47int register_memory_notifier(struct notifier_block *nb)
  48{
  49        return blocking_notifier_chain_register(&memory_chain, nb);
  50}
  51EXPORT_SYMBOL(register_memory_notifier);
  52
  53void unregister_memory_notifier(struct notifier_block *nb)
  54{
  55        blocking_notifier_chain_unregister(&memory_chain, nb);
  56}
  57EXPORT_SYMBOL(unregister_memory_notifier);
  58
  59static ATOMIC_NOTIFIER_HEAD(memory_isolate_chain);
  60
  61int register_memory_isolate_notifier(struct notifier_block *nb)
  62{
  63        return atomic_notifier_chain_register(&memory_isolate_chain, nb);
  64}
  65EXPORT_SYMBOL(register_memory_isolate_notifier);
  66
  67void unregister_memory_isolate_notifier(struct notifier_block *nb)
  68{
  69        atomic_notifier_chain_unregister(&memory_isolate_chain, nb);
  70}
  71EXPORT_SYMBOL(unregister_memory_isolate_notifier);
  72
  73/*
  74 * register_memory - Setup a sysfs device for a memory block
  75 */
  76static
  77int register_memory(struct memory_block *memory)
  78{
  79        int error;
  80
  81        memory->dev.bus = &memory_subsys;
  82        memory->dev.id = memory->start_section_nr / sections_per_block;
  83
  84        error = device_register(&memory->dev);
  85        return error;
  86}
  87
  88static void
  89unregister_memory(struct memory_block *memory)
  90{
  91        BUG_ON(memory->dev.bus != &memory_subsys);
  92
  93        /* drop the ref. we got in remove_memory_block() */
  94        kobject_put(&memory->dev.kobj);
  95        device_unregister(&memory->dev);
  96}
  97
  98unsigned long __weak memory_block_size_bytes(void)
  99{
 100        return MIN_MEMORY_BLOCK_SIZE;
 101}
 102
 103static unsigned long get_memory_block_size(void)
 104{
 105        unsigned long block_sz;
 106
 107        block_sz = memory_block_size_bytes();
 108
 109        /* Validate blk_sz is a power of 2 and not less than section size */
 110        if ((block_sz & (block_sz - 1)) || (block_sz < MIN_MEMORY_BLOCK_SIZE)) {
 111                WARN_ON(1);
 112                block_sz = MIN_MEMORY_BLOCK_SIZE;
 113        }
 114
 115        return block_sz;
 116}
 117
 118/*
 119 * use this as the physical section index that this memsection
 120 * uses.
 121 */
 122
 123static ssize_t show_mem_start_phys_index(struct device *dev,
 124                        struct device_attribute *attr, char *buf)
 125{
 126        struct memory_block *mem =
 127                container_of(dev, struct memory_block, dev);
 128        unsigned long phys_index;
 129
 130        phys_index = mem->start_section_nr / sections_per_block;
 131        return sprintf(buf, "%08lx\n", phys_index);
 132}
 133
 134static ssize_t show_mem_end_phys_index(struct device *dev,
 135                        struct device_attribute *attr, char *buf)
 136{
 137        struct memory_block *mem =
 138                container_of(dev, struct memory_block, dev);
 139        unsigned long phys_index;
 140
 141        phys_index = mem->end_section_nr / sections_per_block;
 142        return sprintf(buf, "%08lx\n", phys_index);
 143}
 144
 145/*
 146 * Show whether the section of memory is likely to be hot-removable
 147 */
 148static ssize_t show_mem_removable(struct device *dev,
 149                        struct device_attribute *attr, char *buf)
 150{
 151        unsigned long i, pfn;
 152        int ret = 1;
 153        struct memory_block *mem =
 154                container_of(dev, struct memory_block, dev);
 155
 156        for (i = 0; i < sections_per_block; i++) {
 157                pfn = section_nr_to_pfn(mem->start_section_nr + i);
 158                ret &= is_mem_section_removable(pfn, PAGES_PER_SECTION);
 159        }
 160
 161        return sprintf(buf, "%d\n", ret);
 162}
 163
 164/*
 165 * online, offline, going offline, etc.
 166 */
 167static ssize_t show_mem_state(struct device *dev,
 168                        struct device_attribute *attr, char *buf)
 169{
 170        struct memory_block *mem =
 171                container_of(dev, struct memory_block, dev);
 172        ssize_t len = 0;
 173
 174        /*
 175         * We can probably put these states in a nice little array
 176         * so that they're not open-coded
 177         */
 178        switch (mem->state) {
 179                case MEM_ONLINE:
 180                        len = sprintf(buf, "online\n");
 181                        break;
 182                case MEM_OFFLINE:
 183                        len = sprintf(buf, "offline\n");
 184                        break;
 185                case MEM_GOING_OFFLINE:
 186                        len = sprintf(buf, "going-offline\n");
 187                        break;
 188                default:
 189                        len = sprintf(buf, "ERROR-UNKNOWN-%ld\n",
 190                                        mem->state);
 191                        WARN_ON(1);
 192                        break;
 193        }
 194
 195        return len;
 196}
 197
 198int memory_notify(unsigned long val, void *v)
 199{
 200        return blocking_notifier_call_chain(&memory_chain, val, v);
 201}
 202
 203int memory_isolate_notify(unsigned long val, void *v)
 204{
 205        return atomic_notifier_call_chain(&memory_isolate_chain, val, v);
 206}
 207
 208/*
 209 * The probe routines leave the pages reserved, just as the bootmem code does.
 210 * Make sure they're still that way.
 211 */
 212static bool pages_correctly_reserved(unsigned long start_pfn,
 213                                        unsigned long nr_pages)
 214{
 215        int i, j;
 216        struct page *page;
 217        unsigned long pfn = start_pfn;
 218
 219        /*
 220         * memmap between sections is not contiguous except with
 221         * SPARSEMEM_VMEMMAP. We lookup the page once per section
 222         * and assume memmap is contiguous within each section
 223         */
 224        for (i = 0; i < sections_per_block; i++, pfn += PAGES_PER_SECTION) {
 225                if (WARN_ON_ONCE(!pfn_valid(pfn)))
 226                        return false;
 227                page = pfn_to_page(pfn);
 228
 229                for (j = 0; j < PAGES_PER_SECTION; j++) {
 230                        if (PageReserved(page + j))
 231                                continue;
 232
 233                        printk(KERN_WARNING "section number %ld page number %d "
 234                                "not reserved, was it already online?\n",
 235                                pfn_to_section_nr(pfn), j);
 236
 237                        return false;
 238                }
 239        }
 240
 241        return true;
 242}
 243
 244/*
 245 * MEMORY_HOTPLUG depends on SPARSEMEM in mm/Kconfig, so it is
 246 * OK to have direct references to sparsemem variables in here.
 247 */
 248static int
 249memory_block_action(unsigned long phys_index, unsigned long action)
 250{
 251        unsigned long start_pfn;
 252        unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
 253        struct page *first_page;
 254        int ret;
 255
 256        first_page = pfn_to_page(phys_index << PFN_SECTION_SHIFT);
 257        start_pfn = page_to_pfn(first_page);
 258
 259        switch (action) {
 260                case MEM_ONLINE:
 261                        if (!pages_correctly_reserved(start_pfn, nr_pages))
 262                                return -EBUSY;
 263
 264                        ret = online_pages(start_pfn, nr_pages);
 265                        break;
 266                case MEM_OFFLINE:
 267                        ret = offline_pages(start_pfn, nr_pages);
 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
 278static 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        if (mem->state != from_state_req) {
 284                ret = -EINVAL;
 285                goto out;
 286        }
 287
 288        if (to_state == MEM_OFFLINE)
 289                mem->state = MEM_GOING_OFFLINE;
 290
 291        ret = memory_block_action(mem->start_section_nr, to_state);
 292
 293        if (ret) {
 294                mem->state = from_state_req;
 295                goto out;
 296        }
 297
 298        mem->state = to_state;
 299        switch (mem->state) {
 300        case MEM_OFFLINE:
 301                kobject_uevent(&mem->dev.kobj, KOBJ_OFFLINE);
 302                break;
 303        case MEM_ONLINE:
 304                kobject_uevent(&mem->dev.kobj, KOBJ_ONLINE);
 305                break;
 306        default:
 307                break;
 308        }
 309out:
 310        return ret;
 311}
 312
 313static int memory_block_change_state(struct memory_block *mem,
 314                unsigned long to_state, unsigned long from_state_req)
 315{
 316        int ret;
 317
 318        mutex_lock(&mem->state_mutex);
 319        ret = __memory_block_change_state(mem, to_state, from_state_req);
 320        mutex_unlock(&mem->state_mutex);
 321
 322        return ret;
 323}
 324static ssize_t
 325store_mem_state(struct device *dev,
 326                struct device_attribute *attr, const char *buf, size_t count)
 327{
 328        struct memory_block *mem;
 329        int ret = -EINVAL;
 330
 331        mem = container_of(dev, struct memory_block, dev);
 332
 333        if (!strncmp(buf, "online", min((int)count, 6)))
 334                ret = memory_block_change_state(mem, MEM_ONLINE, MEM_OFFLINE);
 335        else if(!strncmp(buf, "offline", min((int)count, 7)))
 336                ret = memory_block_change_state(mem, MEM_OFFLINE, MEM_ONLINE);
 337
 338        if (ret)
 339                return ret;
 340        return count;
 341}
 342
 343/*
 344 * phys_device is a bad name for this.  What I really want
 345 * is a way to differentiate between memory ranges that
 346 * are part of physical devices that constitute
 347 * a complete removable unit or fru.
 348 * i.e. do these ranges belong to the same physical device,
 349 * s.t. if I offline all of these sections I can then
 350 * remove the physical device?
 351 */
 352static ssize_t show_phys_device(struct device *dev,
 353                                struct device_attribute *attr, char *buf)
 354{
 355        struct memory_block *mem =
 356                container_of(dev, struct memory_block, dev);
 357        return sprintf(buf, "%d\n", mem->phys_device);
 358}
 359
 360static DEVICE_ATTR(phys_index, 0444, show_mem_start_phys_index, NULL);
 361static DEVICE_ATTR(end_phys_index, 0444, show_mem_end_phys_index, NULL);
 362static DEVICE_ATTR(state, 0644, show_mem_state, store_mem_state);
 363static DEVICE_ATTR(phys_device, 0444, show_phys_device, NULL);
 364static DEVICE_ATTR(removable, 0444, show_mem_removable, NULL);
 365
 366#define mem_create_simple_file(mem, attr_name)  \
 367        device_create_file(&mem->dev, &dev_attr_##attr_name)
 368#define mem_remove_simple_file(mem, attr_name)  \
 369        device_remove_file(&mem->dev, &dev_attr_##attr_name)
 370
 371/*
 372 * Block size attribute stuff
 373 */
 374static ssize_t
 375print_block_size(struct device *dev, struct device_attribute *attr,
 376                 char *buf)
 377{
 378        return sprintf(buf, "%lx\n", get_memory_block_size());
 379}
 380
 381static DEVICE_ATTR(block_size_bytes, 0444, print_block_size, NULL);
 382
 383static int block_size_init(void)
 384{
 385        return device_create_file(memory_subsys.dev_root,
 386                                  &dev_attr_block_size_bytes);
 387}
 388
 389/*
 390 * Some architectures will have custom drivers to do this, and
 391 * will not need to do it from userspace.  The fake hot-add code
 392 * as well as ppc64 will do all of their discovery in userspace
 393 * and will require this interface.
 394 */
 395#ifdef CONFIG_ARCH_MEMORY_PROBE
 396static ssize_t
 397memory_probe_store(struct device *dev, struct device_attribute *attr,
 398                   const char *buf, size_t count)
 399{
 400        u64 phys_addr;
 401        int nid;
 402        int i, ret;
 403        unsigned long pages_per_block = PAGES_PER_SECTION * sections_per_block;
 404
 405        phys_addr = simple_strtoull(buf, NULL, 0);
 406
 407        if (phys_addr & ((pages_per_block << PAGE_SHIFT) - 1))
 408                return -EINVAL;
 409
 410        for (i = 0; i < sections_per_block; i++) {
 411                nid = memory_add_physaddr_to_nid(phys_addr);
 412                ret = add_memory(nid, phys_addr,
 413                                 PAGES_PER_SECTION << PAGE_SHIFT);
 414                if (ret)
 415                        goto out;
 416
 417                phys_addr += MIN_MEMORY_BLOCK_SIZE;
 418        }
 419
 420        ret = count;
 421out:
 422        return ret;
 423}
 424static DEVICE_ATTR(probe, S_IWUSR, NULL, memory_probe_store);
 425
 426static int memory_probe_init(void)
 427{
 428        return device_create_file(memory_subsys.dev_root, &dev_attr_probe);
 429}
 430#else
 431static inline int memory_probe_init(void)
 432{
 433        return 0;
 434}
 435#endif
 436
 437#ifdef CONFIG_MEMORY_FAILURE
 438/*
 439 * Support for offlining pages of memory
 440 */
 441
 442/* Soft offline a page */
 443static ssize_t
 444store_soft_offline_page(struct device *dev,
 445                        struct device_attribute *attr,
 446                        const char *buf, size_t count)
 447{
 448        int ret;
 449        u64 pfn;
 450        if (!capable(CAP_SYS_ADMIN))
 451                return -EPERM;
 452        if (strict_strtoull(buf, 0, &pfn) < 0)
 453                return -EINVAL;
 454        pfn >>= PAGE_SHIFT;
 455        if (!pfn_valid(pfn))
 456                return -ENXIO;
 457        ret = soft_offline_page(pfn_to_page(pfn), 0);
 458        return ret == 0 ? count : ret;
 459}
 460
 461/* Forcibly offline a page, including killing processes. */
 462static ssize_t
 463store_hard_offline_page(struct device *dev,
 464                        struct device_attribute *attr,
 465                        const char *buf, size_t count)
 466{
 467        int ret;
 468        u64 pfn;
 469        if (!capable(CAP_SYS_ADMIN))
 470                return -EPERM;
 471        if (strict_strtoull(buf, 0, &pfn) < 0)
 472                return -EINVAL;
 473        pfn >>= PAGE_SHIFT;
 474        ret = memory_failure(pfn, 0, 0);
 475        return ret ? ret : count;
 476}
 477
 478static DEVICE_ATTR(soft_offline_page, 0644, NULL, store_soft_offline_page);
 479static DEVICE_ATTR(hard_offline_page, 0644, NULL, store_hard_offline_page);
 480
 481static __init int memory_fail_init(void)
 482{
 483        int err;
 484
 485        err = device_create_file(memory_subsys.dev_root,
 486                                &dev_attr_soft_offline_page);
 487        if (!err)
 488                err = device_create_file(memory_subsys.dev_root,
 489                                &dev_attr_hard_offline_page);
 490        return err;
 491}
 492#else
 493static inline int memory_fail_init(void)
 494{
 495        return 0;
 496}
 497#endif
 498
 499/*
 500 * Note that phys_device is optional.  It is here to allow for
 501 * differentiation between which *physical* devices each
 502 * section belongs to...
 503 */
 504int __weak arch_get_memory_phys_device(unsigned long start_pfn)
 505{
 506        return 0;
 507}
 508
 509/*
 510 * A reference for the returned object is held and the reference for the
 511 * hinted object is released.
 512 */
 513struct memory_block *find_memory_block_hinted(struct mem_section *section,
 514                                              struct memory_block *hint)
 515{
 516        int block_id = base_memory_block_id(__section_nr(section));
 517        struct device *hintdev = hint ? &hint->dev : NULL;
 518        struct device *dev;
 519
 520        dev = subsys_find_device_by_id(&memory_subsys, block_id, hintdev);
 521        if (hint)
 522                put_device(&hint->dev);
 523        if (!dev)
 524                return NULL;
 525        return container_of(dev, struct memory_block, dev);
 526}
 527
 528/*
 529 * For now, we have a linear search to go find the appropriate
 530 * memory_block corresponding to a particular phys_index. If
 531 * this gets to be a real problem, we can always use a radix
 532 * tree or something here.
 533 *
 534 * This could be made generic for all device subsystems.
 535 */
 536struct memory_block *find_memory_block(struct mem_section *section)
 537{
 538        return find_memory_block_hinted(section, NULL);
 539}
 540
 541static int init_memory_block(struct memory_block **memory,
 542                             struct mem_section *section, unsigned long state)
 543{
 544        struct memory_block *mem;
 545        unsigned long start_pfn;
 546        int scn_nr;
 547        int ret = 0;
 548
 549        mem = kzalloc(sizeof(*mem), GFP_KERNEL);
 550        if (!mem)
 551                return -ENOMEM;
 552
 553        scn_nr = __section_nr(section);
 554        mem->start_section_nr =
 555                        base_memory_block_id(scn_nr) * sections_per_block;
 556        mem->end_section_nr = mem->start_section_nr + sections_per_block - 1;
 557        mem->state = state;
 558        mem->section_count++;
 559        mutex_init(&mem->state_mutex);
 560        start_pfn = section_nr_to_pfn(mem->start_section_nr);
 561        mem->phys_device = arch_get_memory_phys_device(start_pfn);
 562
 563        ret = register_memory(mem);
 564        if (!ret)
 565                ret = mem_create_simple_file(mem, phys_index);
 566        if (!ret)
 567                ret = mem_create_simple_file(mem, end_phys_index);
 568        if (!ret)
 569                ret = mem_create_simple_file(mem, state);
 570        if (!ret)
 571                ret = mem_create_simple_file(mem, phys_device);
 572        if (!ret)
 573                ret = mem_create_simple_file(mem, removable);
 574
 575        *memory = mem;
 576        return ret;
 577}
 578
 579static int add_memory_section(int nid, struct mem_section *section,
 580                        struct memory_block **mem_p,
 581                        unsigned long state, enum mem_add_context context)
 582{
 583        struct memory_block *mem = NULL;
 584        int scn_nr = __section_nr(section);
 585        int ret = 0;
 586
 587        mutex_lock(&mem_sysfs_mutex);
 588
 589        if (context == BOOT) {
 590                /* same memory block ? */
 591                if (mem_p && *mem_p)
 592                        if (scn_nr >= (*mem_p)->start_section_nr &&
 593                            scn_nr <= (*mem_p)->end_section_nr) {
 594                                mem = *mem_p;
 595                                kobject_get(&mem->dev.kobj);
 596                        }
 597        } else
 598                mem = find_memory_block(section);
 599
 600        if (mem) {
 601                mem->section_count++;
 602                kobject_put(&mem->dev.kobj);
 603        } else {
 604                ret = init_memory_block(&mem, section, state);
 605                /* store memory_block pointer for next loop */
 606                if (!ret && context == BOOT)
 607                        if (mem_p)
 608                                *mem_p = mem;
 609        }
 610
 611        if (!ret) {
 612                if (context == HOTPLUG &&
 613                    mem->section_count == sections_per_block)
 614                        ret = register_mem_sect_under_node(mem, nid);
 615        }
 616
 617        mutex_unlock(&mem_sysfs_mutex);
 618        return ret;
 619}
 620
 621int remove_memory_block(unsigned long node_id, struct mem_section *section,
 622                int phys_device)
 623{
 624        struct memory_block *mem;
 625
 626        mutex_lock(&mem_sysfs_mutex);
 627        mem = find_memory_block(section);
 628        unregister_mem_sect_under_nodes(mem, __section_nr(section));
 629
 630        mem->section_count--;
 631        if (mem->section_count == 0) {
 632                mem_remove_simple_file(mem, phys_index);
 633                mem_remove_simple_file(mem, end_phys_index);
 634                mem_remove_simple_file(mem, state);
 635                mem_remove_simple_file(mem, phys_device);
 636                mem_remove_simple_file(mem, removable);
 637                unregister_memory(mem);
 638                kfree(mem);
 639        } else
 640                kobject_put(&mem->dev.kobj);
 641
 642        mutex_unlock(&mem_sysfs_mutex);
 643        return 0;
 644}
 645
 646/*
 647 * need an interface for the VM to add new memory regions,
 648 * but without onlining it.
 649 */
 650int register_new_memory(int nid, struct mem_section *section)
 651{
 652        return add_memory_section(nid, section, NULL, MEM_OFFLINE, HOTPLUG);
 653}
 654
 655int unregister_memory_section(struct mem_section *section)
 656{
 657        if (!present_section(section))
 658                return -EINVAL;
 659
 660        return remove_memory_block(0, section, 0);
 661}
 662
 663/*
 664 * offline one memory block. If the memory block has been offlined, do nothing.
 665 */
 666int offline_memory_block(struct memory_block *mem)
 667{
 668        int ret = 0;
 669
 670        mutex_lock(&mem->state_mutex);
 671        if (mem->state != MEM_OFFLINE)
 672                ret = __memory_block_change_state(mem, MEM_OFFLINE, MEM_ONLINE);
 673        mutex_unlock(&mem->state_mutex);
 674
 675        return ret;
 676}
 677
 678/*
 679 * Initialize the sysfs support for memory devices...
 680 */
 681int __init memory_dev_init(void)
 682{
 683        unsigned int i;
 684        int ret;
 685        int err;
 686        unsigned long block_sz;
 687        struct memory_block *mem = NULL;
 688
 689        ret = subsys_system_register(&memory_subsys, NULL);
 690        if (ret)
 691                goto out;
 692
 693        block_sz = get_memory_block_size();
 694        sections_per_block = block_sz / MIN_MEMORY_BLOCK_SIZE;
 695
 696        /*
 697         * Create entries for memory sections that were found
 698         * during boot and have been initialized
 699         */
 700        for (i = 0; i < NR_MEM_SECTIONS; i++) {
 701                if (!present_section_nr(i))
 702                        continue;
 703                /* don't need to reuse memory_block if only one per block */
 704                err = add_memory_section(0, __nr_to_section(i),
 705                                 (sections_per_block == 1) ? NULL : &mem,
 706                                         MEM_ONLINE,
 707                                         BOOT);
 708                if (!ret)
 709                        ret = err;
 710        }
 711
 712        err = memory_probe_init();
 713        if (!ret)
 714                ret = err;
 715        err = memory_fail_init();
 716        if (!ret)
 717                ret = err;
 718        err = block_size_init();
 719        if (!ret)
 720                ret = err;
 721out:
 722        if (ret)
 723                printk(KERN_ERR "%s() failed: %d\n", __func__, ret);
 724        return ret;
 725}
 726
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.