mm/memory_hotplug.c

   1 /*
   2  *  linux/mm/memory_hotplug.c
   3  *
   4  *  Copyright (C)
   5  */
   6
   7 #include <linux/stddef.h>
   8 #include <linux/mm.h>
   9 #include <linux/swap.h>
  10 #include <linux/interrupt.h>
  11 #include <linux/pagemap.h>
  12 #include <linux/bootmem.h>
  13 #include <linux/compiler.h>
  14 #include <linux/export.h>
  15 #include <linux/pagevec.h>
  16 #include <linux/writeback.h>
  17 #include <linux/slab.h>
  18 #include <linux/sysctl.h>
  19 #include <linux/cpu.h>
  20 #include <linux/memory.h>
  21 #include <linux/memory_hotplug.h>
  22 #include <linux/highmem.h>
  23 #include <linux/vmalloc.h>
  24 #include <linux/ioport.h>
  25 #include <linux/delay.h>
  26 #include <linux/migrate.h>
  27 #include <linux/page-isolation.h>
  28 #include <linux/pfn.h>
  29 #include <linux/suspend.h>
  30 #include <linux/mm_inline.h>
  31 #include <linux/firmware-map.h>
  32
  33 #include <asm/tlbflush.h>
  34
  35 #include "internal.h"
  36
  37 /*
  38  * online_page_callback contains pointer to current page onlining function.
  39  * Initially it is generic_online_page(). If it is required it could be
  40  * changed by calling set_online_page_callback() for callback registration
  41  * and restore_online_page_callback() for generic callback restore.
  42  */
  43
  44 static void generic_online_page(struct page *page);
  45
  46 static online_page_callback_t online_page_callback = generic_online_page;
  47
  48 DEFINE_MUTEX(mem_hotplug_mutex);
  49
  50 void lock_memory_hotplug(void)
  51 {
  52         mutex_lock(&mem_hotplug_mutex);
  53
  54         /* for exclusive hibernation if CONFIG_HIBERNATION=y */
  55         lock_system_sleep();
  56 }
  57
  58 void unlock_memory_hotplug(void)
  59 {
  60         unlock_system_sleep();
  61         mutex_unlock(&mem_hotplug_mutex);
  62 }
  63
  64
  65 /* add this memory to iomem resource */
  66 static struct resource *register_memory_resource(u64 start, u64 size)
  67 {
  68         struct resource *res;
  69         res = kzalloc(sizeof(struct resource), GFP_KERNEL);
  70         BUG_ON(!res);
  71
  72         res->name = "System RAM";
  73         res->start = start;
  74         res->end = start + size - 1;
  75         res->flags = IORESOURCE_MEM | IORESOURCE_BUSY;
  76         if (request_resource(&iomem_resource, res) < 0) {
  77                 printk("System RAM resource %pR cannot be added\n", res);
  78                 kfree(res);
  79                 res = NULL;
  80         }
  81         return res;
  82 }
  83
  84 static void release_memory_resource(struct resource *res)
  85 {
  86         if (!res)
  87                 return;
  88         release_resource(res);
  89         kfree(res);
  90         return;
  91 }
  92
  93 #ifdef CONFIG_MEMORY_HOTPLUG_SPARSE
  94 #ifndef CONFIG_SPARSEMEM_VMEMMAP
  95 static void get_page_bootmem(unsigned long info,  struct page *page,
  96                              unsigned long type)
  97 {
  98         page->lru.next = (struct list_head *) type;
  99         SetPagePrivate(page);
 100         set_page_private(page, info);
 101         atomic_inc(&page->_count);
 102 }
 103
 104 /* reference to __meminit __free_pages_bootmem is valid
 105  * so use __ref to tell modpost not to generate a warning */
 106 void __ref put_page_bootmem(struct page *page)
 107 {
 108         unsigned long type;
 109
 110         type = (unsigned long) page->lru.next;
 111         BUG_ON(type < MEMORY_HOTPLUG_MIN_BOOTMEM_TYPE ||
 112                type > MEMORY_HOTPLUG_MAX_BOOTMEM_TYPE);
 113
 114         if (atomic_dec_return(&page->_count) == 1) {
 115                 ClearPagePrivate(page);
 116                 set_page_private(page, 0);
 117                 INIT_LIST_HEAD(&page->lru);
 118                 __free_pages_bootmem(page, 0);
 119         }
 120
 121 }
 122
 123 static void register_page_bootmem_info_section(unsigned long start_pfn)
 124 {
 125         unsigned long *usemap, mapsize, section_nr, i;
 126         struct mem_section *ms;
 127         struct page *page, *memmap;
 128
 129         section_nr = pfn_to_section_nr(start_pfn);
 130         ms = __nr_to_section(section_nr);
 131
 132         /* Get section's memmap address */
 133         memmap = sparse_decode_mem_map(ms->section_mem_map, section_nr);
 134
 135         /*
 136          * Get page for the memmap's phys address
 137          * XXX: need more consideration for sparse_vmemmap...
 138          */
 139         page = virt_to_page(memmap);
 140         mapsize = sizeof(struct page) * PAGES_PER_SECTION;
 141         mapsize = PAGE_ALIGN(mapsize) >> PAGE_SHIFT;
 142
 143         /* remember memmap's page */
 144         for (i = 0; i < mapsize; i++, page++)
 145                 get_page_bootmem(section_nr, page, SECTION_INFO);
 146
 147         usemap = __nr_to_section(section_nr)->pageblock_flags;
 148         page = virt_to_page(usemap);
 149
 150         mapsize = PAGE_ALIGN(usemap_size()) >> PAGE_SHIFT;
 151
 152         for (i = 0; i < mapsize; i++, page++)
 153                 get_page_bootmem(section_nr, page, MIX_SECTION_INFO);
 154
 155 }
 156
 157 void register_page_bootmem_info_node(struct pglist_data *pgdat)
 158 {
 159         unsigned long i, pfn, end_pfn, nr_pages;
 160         int node = pgdat->node_id;
 161         struct page *page;
 162         struct zone *zone;
 163
 164         nr_pages = PAGE_ALIGN(sizeof(struct pglist_data)) >> PAGE_SHIFT;
 165         page = virt_to_page(pgdat);
 166
 167         for (i = 0; i < nr_pages; i++, page++)
 168                 get_page_bootmem(node, page, NODE_INFO);
 169
 170         zone = &pgdat->node_zones[0];
 171         for (; zone < pgdat->node_zones + MAX_NR_ZONES - 1; zone++) {
 172                 if (zone->wait_table) {
 173                         nr_pages = zone->wait_table_hash_nr_entries
 174                                 * sizeof(wait_queue_head_t);
 175                         nr_pages = PAGE_ALIGN(nr_pages) >> PAGE_SHIFT;
 176                         page = virt_to_page(zone->wait_table);
 177
 178                         for (i = 0; i < nr_pages; i++, page++)
 179                                 get_page_bootmem(node, page, NODE_INFO);
 180                 }
 181         }
 182
 183         pfn = pgdat->node_start_pfn;
 184         end_pfn = pfn + pgdat->node_spanned_pages;
 185
 186         /* register_section info */
 187         for (; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
 188                 /*
 189                  * Some platforms can assign the same pfn to multiple nodes - on
 190                  * node0 as well as nodeN.  To avoid registering a pfn against
 191                  * multiple nodes we check that this pfn does not already
 192                  * reside in some other node.
 193                  */
 194                 if (pfn_valid(pfn) && (pfn_to_nid(pfn) == node))
 195                         register_page_bootmem_info_section(pfn);
 196         }
 197 }
 198 #endif /* !CONFIG_SPARSEMEM_VMEMMAP */
 199
 200 static void grow_zone_span(struct zone *zone, unsigned long start_pfn,
 201                            unsigned long end_pfn)
 202 {
 203         unsigned long old_zone_end_pfn;
 204
 205         zone_span_writelock(zone);
 206
 207         old_zone_end_pfn = zone->zone_start_pfn + zone->spanned_pages;
 208         if (start_pfn < zone->zone_start_pfn)
 209                 zone->zone_start_pfn = start_pfn;
 210
 211         zone->spanned_pages = max(old_zone_end_pfn, end_pfn) -
 212                                 zone->zone_start_pfn;
 213
 214         zone_span_writeunlock(zone);
 215 }
 216
 217 static void grow_pgdat_span(struct pglist_data *pgdat, unsigned long start_pfn,
 218                             unsigned long end_pfn)
 219 {
 220         unsigned long old_pgdat_end_pfn =
 221                 pgdat->node_start_pfn + pgdat->node_spanned_pages;
 222
 223         if (start_pfn < pgdat->node_start_pfn)
 224                 pgdat->node_start_pfn = start_pfn;
 225
 226         pgdat->node_spanned_pages = max(old_pgdat_end_pfn, end_pfn) -
 227                                         pgdat->node_start_pfn;
 228 }
 229
 230 static int __meminit __add_zone(struct zone *zone, unsigned long phys_start_pfn)
 231 {
 232         struct pglist_data *pgdat = zone->zone_pgdat;
 233         int nr_pages = PAGES_PER_SECTION;
 234         int nid = pgdat->node_id;
 235         int zone_type;
 236         unsigned long flags;
 237
 238         zone_type = zone - pgdat->node_zones;
 239         if (!zone->wait_table) {
 240                 int ret;
 241
 242                 ret = init_currently_empty_zone(zone, phys_start_pfn,
 243                                                 nr_pages, MEMMAP_HOTPLUG);
 244                 if (ret)
 245                         return ret;
 246         }
 247         pgdat_resize_lock(zone->zone_pgdat, &flags);
 248         grow_zone_span(zone, phys_start_pfn, phys_start_pfn + nr_pages);
 249         grow_pgdat_span(zone->zone_pgdat, phys_start_pfn,
 250                         phys_start_pfn + nr_pages);
 251         pgdat_resize_unlock(zone->zone_pgdat, &flags);
 252         memmap_init_zone(nr_pages, nid, zone_type,
 253                          phys_start_pfn, MEMMAP_HOTPLUG);
 254         return 0;
 255 }
 256
 257 static int __meminit __add_section(int nid, struct zone *zone,
 258                                         unsigned long phys_start_pfn)
 259 {
 260         int nr_pages = PAGES_PER_SECTION;
 261         int ret;
 262
 263         if (pfn_valid(phys_start_pfn))
 264                 return -EEXIST;
 265
 266         ret = sparse_add_one_section(zone, phys_start_pfn, nr_pages);
 267
 268         if (ret < 0)
 269                 return ret;
 270
 271         ret = __add_zone(zone, phys_start_pfn);
 272
 273         if (ret < 0)
 274                 return ret;
 275
 276         return register_new_memory(nid, __pfn_to_section(phys_start_pfn));
 277 }
 278
 279 #ifdef CONFIG_SPARSEMEM_VMEMMAP
 280 static int __remove_section(struct zone *zone, struct mem_section *ms)
 281 {
 282         /*
 283          * XXX: Freeing memmap with vmemmap is not implement yet.
 284          *      This should be removed later.
 285          */
 286         return -EBUSY;
 287 }
 288 #else
 289 static int __remove_section(struct zone *zone, struct mem_section *ms)
 290 {
 291         unsigned long flags;
 292         struct pglist_data *pgdat = zone->zone_pgdat;
 293         int ret = -EINVAL;
 294
 295         if (!valid_section(ms))
 296                 return ret;
 297
 298         ret = unregister_memory_section(ms);
 299         if (ret)
 300                 return ret;
 301
 302         pgdat_resize_lock(pgdat, &flags);
 303         sparse_remove_one_section(zone, ms);
 304         pgdat_resize_unlock(pgdat, &flags);
 305         return 0;
 306 }
 307 #endif
 308
 309 /*
 310  * Reasonably generic function for adding memory.  It is
 311  * expected that archs that support memory hotplug will
 312  * call this function after deciding the zone to which to
 313  * add the new pages.
 314  */
 315 int __ref __add_pages(int nid, struct zone *zone, unsigned long phys_start_pfn,
 316                         unsigned long nr_pages)
 317 {
 318         unsigned long i;
 319         int err = 0;
 320         int start_sec, end_sec;
 321         /* during initialize mem_map, align hot-added range to section */
 322         start_sec = pfn_to_section_nr(phys_start_pfn);
 323         end_sec = pfn_to_section_nr(phys_start_pfn + nr_pages - 1);
 324
 325         for (i = start_sec; i <= end_sec; i++) {
 326                 err = __add_section(nid, zone, i << PFN_SECTION_SHIFT);
 327
 328                 /*
 329                  * EEXIST is finally dealt with by ioresource collision
 330                  * check. see add_memory() => register_memory_resource()
 331                  * Warning will be printed if there is collision.
 332                  */
 333                 if (err && (err != -EEXIST))
 334                         break;
 335                 err = 0;
 336         }
 337
 338         return err;
 339 }
 340 EXPORT_SYMBOL_GPL(__add_pages);
 341
 342 /**
 343  * __remove_pages() - remove sections of pages from a zone
 344  * @zone: zone from which pages need to be removed
 345  * @phys_start_pfn: starting pageframe (must be aligned to start of a section)
 346  * @nr_pages: number of pages to remove (must be multiple of section size)
 347  *
 348  * Generic helper function to remove section mappings and sysfs entries
 349  * for the section of the memory we are removing. Caller needs to make
 350  * sure that pages are marked reserved and zones are adjust properly by
 351  * calling offline_pages().
 352  */
 353 int __remove_pages(struct zone *zone, unsigned long phys_start_pfn,
 354                  unsigned long nr_pages)
 355 {
 356         unsigned long i, ret = 0;
 357         int sections_to_remove;
 358
 359         /*
 360          * We can only remove entire sections
 361          */
 362         BUG_ON(phys_start_pfn & ~PAGE_SECTION_MASK);
 363         BUG_ON(nr_pages % PAGES_PER_SECTION);
 364
 365         release_mem_region(phys_start_pfn << PAGE_SHIFT, nr_pages * PAGE_SIZE);
 366
 367         sections_to_remove = nr_pages / PAGES_PER_SECTION;
 368         for (i = 0; i < sections_to_remove; i++) {
 369                 unsigned long pfn = phys_start_pfn + i*PAGES_PER_SECTION;
 370                 ret = __remove_section(zone, __pfn_to_section(pfn));
 371                 if (ret)
 372                         break;
 373         }
 374         return ret;
 375 }
 376 EXPORT_SYMBOL_GPL(__remove_pages);
 377
 378 int set_online_page_callback(online_page_callback_t callback)
 379 {
 380         int rc = -EINVAL;
 381
 382         lock_memory_hotplug();
 383
 384         if (online_page_callback == generic_online_page) {
 385                 online_page_callback = callback;
 386                 rc = 0;
 387         }
 388
 389         unlock_memory_hotplug();
 390
 391         return rc;
 392 }
 393 EXPORT_SYMBOL_GPL(set_online_page_callback);
 394
 395 int restore_online_page_callback(online_page_callback_t callback)
 396 {
 397         int rc = -EINVAL;
 398
 399         lock_memory_hotplug();
 400
 401         if (online_page_callback == callback) {
 402                 online_page_callback = generic_online_page;
 403                 rc = 0;
 404         }
 405
 406         unlock_memory_hotplug();
 407
 408         return rc;
 409 }
 410 EXPORT_SYMBOL_GPL(restore_online_page_callback);
 411
 412 void __online_page_set_limits(struct page *page)
 413 {
 414         unsigned long pfn = page_to_pfn(page);
 415
 416         if (pfn >= num_physpages)
 417                 num_physpages = pfn + 1;
 418 }
 419 EXPORT_SYMBOL_GPL(__online_page_set_limits);
 420
 421 void __online_page_increment_counters(struct page *page)
 422 {
 423         totalram_pages++;
 424
 425 #ifdef CONFIG_HIGHMEM
 426         if (PageHighMem(page))
 427                 totalhigh_pages++;
 428 #endif
 429 }
 430 EXPORT_SYMBOL_GPL(__online_page_increment_counters);
 431
 432 void __online_page_free(struct page *page)
 433 {
 434         ClearPageReserved(page);
 435         init_page_count(page);
 436         __free_page(page);
 437 }
 438 EXPORT_SYMBOL_GPL(__online_page_free);
 439
 440 static void generic_online_page(struct page *page)
 441 {
 442         __online_page_set_limits(page);
 443         __online_page_increment_counters(page);
 444         __online_page_free(page);
 445 }
 446
 447 static int online_pages_range(unsigned long start_pfn, unsigned long nr_pages,
 448                         void *arg)
 449 {
 450         unsigned long i;
 451         unsigned long onlined_pages = *(unsigned long *)arg;
 452         struct page *page;
 453         if (PageReserved(pfn_to_page(start_pfn)))
 454                 for (i = 0; i < nr_pages; i++) {
 455                         page = pfn_to_page(start_pfn + i);
 456                         (*online_page_callback)(page);
 457                         onlined_pages++;
 458                 }
 459         *(unsigned long *)arg = onlined_pages;
 460         return 0;
 461 }
 462
 463
 464 int __ref online_pages(unsigned long pfn, unsigned long nr_pages)
 465 {
 466         unsigned long onlined_pages = 0;
 467         struct zone *zone;
 468         int need_zonelists_rebuild = 0;
 469         int nid;
 470         int ret;
 471         struct memory_notify arg;
 472
 473         lock_memory_hotplug();
 474         arg.start_pfn = pfn;
 475         arg.nr_pages = nr_pages;
 476         arg.status_change_nid = -1;
 477
 478         nid = page_to_nid(pfn_to_page(pfn));
 479         if (node_present_pages(nid) == 0)
 480                 arg.status_change_nid = nid;
 481
 482         ret = memory_notify(MEM_GOING_ONLINE, &arg);
 483         ret = notifier_to_errno(ret);
 484         if (ret) {
 485                 memory_notify(MEM_CANCEL_ONLINE, &arg);
 486                 unlock_memory_hotplug();
 487                 return ret;
 488         }
 489         /*
 490          * This doesn't need a lock to do pfn_to_page().
 491          * The section can't be removed here because of the
 492          * memory_block->state_mutex.
 493          */
 494         zone = page_zone(pfn_to_page(pfn));
 495         /*
 496          * If this zone is not populated, then it is not in zonelist.
 497          * This means the page allocator ignores this zone.
 498          * So, zonelist must be updated after online.
 499          */
 500         mutex_lock(&zonelists_mutex);
 501         if (!populated_zone(zone))
 502                 need_zonelists_rebuild = 1;
 503
 504         ret = walk_system_ram_range(pfn, nr_pages, &onlined_pages,
 505                 online_pages_range);
 506         if (ret) {
 507                 mutex_unlock(&zonelists_mutex);
 508                 printk(KERN_DEBUG "online_pages [mem %#010llx-%#010llx] failed\n",
 509                        (unsigned long long) pfn << PAGE_SHIFT,
 510                        (((unsigned long long) pfn + nr_pages)
 511                             << PAGE_SHIFT) - 1);
 512                 memory_notify(MEM_CANCEL_ONLINE, &arg);
 513                 unlock_memory_hotplug();
 514                 return ret;
 515         }
 516
 517         zone->present_pages += onlined_pages;
 518         zone->zone_pgdat->node_present_pages += onlined_pages;
 519         if (onlined_pages) {
 520                 node_set_state(zone_to_nid(zone), N_HIGH_MEMORY);
 521                 if (need_zonelists_rebuild)
 522                         build_all_zonelists(NULL, zone);
 523                 else
 524                         zone_pcp_update(zone);
 525         }
 526
 527         mutex_unlock(&zonelists_mutex);
 528
 529         init_per_zone_wmark_min();
 530
 531         if (onlined_pages)
 532                 kswapd_run(zone_to_nid(zone));
 533
 534         vm_total_pages = nr_free_pagecache_pages();
 535
 536         writeback_set_ratelimit();
 537
 538         if (onlined_pages)
 539                 memory_notify(MEM_ONLINE, &arg);
 540         unlock_memory_hotplug();
 541
 542         return 0;
 543 }
 544 #endif /* CONFIG_MEMORY_HOTPLUG_SPARSE */
 545
 546 /* we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG */
 547 static pg_data_t __ref *hotadd_new_pgdat(int nid, u64 start)
 548 {
 549         struct pglist_data *pgdat;
 550         unsigned long zones_size[MAX_NR_ZONES] = {0};
 551         unsigned long zholes_size[MAX_NR_ZONES] = {0};
 552         unsigned long start_pfn = start >> PAGE_SHIFT;
 553
 554         pgdat = arch_alloc_nodedata(nid);
 555         if (!pgdat)
 556                 return NULL;
 557
 558         arch_refresh_nodedata(nid, pgdat);
 559
 560         /* we can use NODE_DATA(nid) from here */
 561
 562         /* init node's zones as empty zones, we don't have any present pages.*/
 563         free_area_init_node(nid, zones_size, start_pfn, zholes_size);
 564
 565         /*
 566          * The node we allocated has no zone fallback lists. For avoiding
 567          * to access not-initialized zonelist, build here.
 568          */
 569         mutex_lock(&zonelists_mutex);
 570         build_all_zonelists(pgdat, NULL);
 571         mutex_unlock(&zonelists_mutex);
 572
 573         return pgdat;
 574 }
 575
 576 static void rollback_node_hotadd(int nid, pg_data_t *pgdat)
 577 {
 578         arch_refresh_nodedata(nid, NULL);
 579         arch_free_nodedata(pgdat);
 580         return;
 581 }
 582
 583
 584 /*
 585  * called by cpu_up() to online a node without onlined memory.
 586  */
 587 int mem_online_node(int nid)
 588 {
 589         pg_data_t       *pgdat;
 590         int     ret;
 591
 592         lock_memory_hotplug();
 593         pgdat = hotadd_new_pgdat(nid, 0);
 594         if (!pgdat) {
 595                 ret = -ENOMEM;
 596                 goto out;
 597         }
 598         node_set_online(nid);
 599         ret = register_one_node(nid);
 600         BUG_ON(ret);
 601
 602 out:
 603         unlock_memory_hotplug();
 604         return ret;
 605 }
 606
 607 /* we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG */
 608 int __ref add_memory(int nid, u64 start, u64 size)
 609 {
 610         pg_data_t *pgdat = NULL;
 611         int new_pgdat = 0;
 612         struct resource *res;
 613         int ret;
 614
 615         lock_memory_hotplug();
 616
 617         res = register_memory_resource(start, size);
 618         ret = -EEXIST;
 619         if (!res)
 620                 goto out;
 621
 622         if (!node_online(nid)) {
 623                 pgdat = hotadd_new_pgdat(nid, start);
 624                 ret = -ENOMEM;
 625                 if (!pgdat)
 626                         goto error;
 627                 new_pgdat = 1;
 628         }
 629
 630         /* call arch's memory hotadd */
 631         ret = arch_add_memory(nid, start, size);
 632
 633         if (ret < 0)
 634                 goto error;
 635
 636         /* we online node here. we can't roll back from here. */
 637         node_set_online(nid);
 638
 639         if (new_pgdat) {
 640                 ret = register_one_node(nid);
 641                 /*
 642                  * If sysfs file of new node can't create, cpu on the node
 643                  * can't be hot-added. There is no rollback way now.
 644                  * So, check by BUG_ON() to catch it reluctantly..
 645                  */
 646                 BUG_ON(ret);
 647         }
 648
 649         /* create new memmap entry */
 650         firmware_map_add_hotplug(start, start + size, "System RAM");
 651
 652         goto out;
 653
 654 error:
 655         /* rollback pgdat allocation and others */
 656         if (new_pgdat)
 657                 rollback_node_hotadd(nid, pgdat);
 658         if (res)
 659                 release_memory_resource(res);
 660
 661 out:
 662         unlock_memory_hotplug();
 663         return ret;
 664 }
 665 EXPORT_SYMBOL_GPL(add_memory);
 666
 667 #ifdef CONFIG_MEMORY_HOTREMOVE
 668 /*
 669  * A free page on the buddy free lists (not the per-cpu lists) has PageBuddy
 670  * set and the size of the free page is given by page_order(). Using this,
 671  * the function determines if the pageblock contains only free pages.
 672  * Due to buddy contraints, a free page at least the size of a pageblock will
 673  * be located at the start of the pageblock
 674  */
 675 static inline int pageblock_free(struct page *page)
 676 {
 677         return PageBuddy(page) && page_order(page) >= pageblock_order;
 678 }
 679
 680 /* Return the start of the next active pageblock after a given page */
 681 static struct page *next_active_pageblock(struct page *page)
 682 {
 683         /* Ensure the starting page is pageblock-aligned */
 684         BUG_ON(page_to_pfn(page) & (pageblock_nr_pages - 1));
 685
 686         /* If the entire pageblock is free, move to the end of free page */
 687         if (pageblock_free(page)) {
 688                 int order;
 689                 /* be careful. we don't have locks, page_order can be changed.*/
 690                 order = page_order(page);
 691                 if ((order < MAX_ORDER) && (order >= pageblock_order))
 692                         return page + (1 << order);
 693         }
 694
 695         return page + pageblock_nr_pages;
 696 }
 697
 698 /* Checks if this range of memory is likely to be hot-removable. */
 699 int is_mem_section_removable(unsigned long start_pfn, unsigned long nr_pages)
 700 {
 701         struct page *page = pfn_to_page(start_pfn);
 702         struct page *end_page = page + nr_pages;
 703
 704         /* Check the starting page of each pageblock within the range */
 705         for (; page < end_page; page = next_active_pageblock(page)) {
 706                 if (!is_pageblock_removable_nolock(page))
 707                         return 0;
 708                 cond_resched();
 709         }
 710
 711         /* All pageblocks in the memory block are likely to be hot-removable */
 712         return 1;
 713 }
 714
 715 /*
 716  * Confirm all pages in a range [start, end) is belongs to the same zone.
 717  */
 718 static int test_pages_in_a_zone(unsigned long start_pfn, unsigned long end_pfn)
 719 {
 720         unsigned long pfn;
 721         struct zone *zone = NULL;
 722         struct page *page;
 723         int i;
 724         for (pfn = start_pfn;
 725              pfn < end_pfn;
 726              pfn += MAX_ORDER_NR_PAGES) {
 727                 i = 0;
 728                 /* This is just a CONFIG_HOLES_IN_ZONE check.*/
 729                 while ((i < MAX_ORDER_NR_PAGES) && !pfn_valid_within(pfn + i))
 730                         i++;
 731                 if (i == MAX_ORDER_NR_PAGES)
 732                         continue;
 733                 page = pfn_to_page(pfn + i);
 734                 if (zone && page_zone(page) != zone)
 735                         return 0;
 736                 zone = page_zone(page);
 737         }
 738         return 1;
 739 }
 740
 741 /*
 742  * Scanning pfn is much easier than scanning lru list.
 743  * Scan pfn from start to end and Find LRU page.
 744  */
 745 static unsigned long scan_lru_pages(unsigned long start, unsigned long end)
 746 {
 747         unsigned long pfn;
 748         struct page *page;
 749         for (pfn = start; pfn < end; pfn++) {
 750                 if (pfn_valid(pfn)) {
 751                         page = pfn_to_page(pfn);
 752                         if (PageLRU(page))
 753                                 return pfn;
 754                 }
 755         }
 756         return 0;
 757 }
 758
 759 #define NR_OFFLINE_AT_ONCE_PAGES        (256)
 760 static int
 761 do_migrate_range(unsigned long start_pfn, unsigned long end_pfn)
 762 {
 763         unsigned long pfn;
 764         struct page *page;
 765         int move_pages = NR_OFFLINE_AT_ONCE_PAGES;
 766         int not_managed = 0;
 767         int ret = 0;
 768         LIST_HEAD(source);
 769
 770         for (pfn = start_pfn; pfn < end_pfn && move_pages > 0; pfn++) {
 771                 if (!pfn_valid(pfn))
 772                         continue;
 773                 page = pfn_to_page(pfn);
 774                 if (!get_page_unless_zero(page))
 775                         continue;
 776                 /*
 777                  * We can skip free pages. And we can only deal with pages on
 778                  * LRU.
 779                  */
 780                 ret = isolate_lru_page(page);
 781                 if (!ret) { /* Success */
 782                         put_page(page);
 783                         list_add_tail(&page->lru, &source);
 784                         move_pages--;
 785                         inc_zone_page_state(page, NR_ISOLATED_ANON +
 786                                             page_is_file_cache(page));
 787
 788                 } else {
 789 #ifdef CONFIG_DEBUG_VM
 790                         printk(KERN_ALERT "removing pfn %lx from LRU failed\n",
 791                                pfn);
 792                         dump_page(page);
 793 #endif
 794                         put_page(page);
 795                         /* Because we don't have big zone->lock. we should
 796                            check this again here. */
 797                         if (page_count(page)) {
 798                                 not_managed++;
 799                                 ret = -EBUSY;
 800                                 break;
 801                         }
 802                 }
 803         }
 804         if (!list_empty(&source)) {
 805                 if (not_managed) {
 806                         putback_lru_pages(&source);
 807                         goto out;
 808                 }
 809
 810                 /*
 811                  * alloc_migrate_target should be improooooved!!
 812                  * migrate_pages returns # of failed pages.
 813                  */
 814                 ret = migrate_pages(&source, alloc_migrate_target, 0,
 815                                                         true, MIGRATE_SYNC);
 816                 if (ret)
 817                         putback_lru_pages(&source);
 818         }
 819 out:
 820         return ret;
 821 }
 822
 823 /*
 824  * remove from free_area[] and mark all as Reserved.
 825  */
 826 static int
 827 offline_isolated_pages_cb(unsigned long start, unsigned long nr_pages,
 828                         void *data)
 829 {
 830         __offline_isolated_pages(start, start + nr_pages);
 831         return 0;
 832 }
 833
 834 static void
 835 offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
 836 {
 837         walk_system_ram_range(start_pfn, end_pfn - start_pfn, NULL,
 838                                 offline_isolated_pages_cb);
 839 }
 840
 841 /*
 842  * Check all pages in range, recoreded as memory resource, are isolated.
 843  */
 844 static int
 845 check_pages_isolated_cb(unsigned long start_pfn, unsigned long nr_pages,
 846                         void *data)
 847 {
 848         int ret;
 849         long offlined = *(long *)data;
 850         ret = test_pages_isolated(start_pfn, start_pfn + nr_pages);
 851         offlined = nr_pages;
 852         if (!ret)
 853                 *(long *)data += offlined;
 854         return ret;
 855 }
 856
 857 static long
 858 check_pages_isolated(unsigned long start_pfn, unsigned long end_pfn)
 859 {
 860         long offlined = 0;
 861         int ret;
 862
 863         ret = walk_system_ram_range(start_pfn, end_pfn - start_pfn, &offlined,
 864                         check_pages_isolated_cb);
 865         if (ret < 0)
 866                 offlined = (long)ret;
 867         return offlined;
 868 }
 869
 870 static int __ref __offline_pages(unsigned long start_pfn,
 871                   unsigned long end_pfn, unsigned long timeout)
 872 {
 873         unsigned long pfn, nr_pages, expire;
 874         long offlined_pages;
 875         int ret, drain, retry_max, node;
 876         struct zone *zone;
 877         struct memory_notify arg;
 878
 879         BUG_ON(start_pfn >= end_pfn);
 880         /* at least, alignment against pageblock is necessary */
 881         if (!IS_ALIGNED(start_pfn, pageblock_nr_pages))
 882                 return -EINVAL;
 883         if (!IS_ALIGNED(end_pfn, pageblock_nr_pages))
 884                 return -EINVAL;
 885         /* This makes hotplug much easier...and readable.
 886            we assume this for now. .*/
 887         if (!test_pages_in_a_zone(start_pfn, end_pfn))
 888                 return -EINVAL;
 889
 890         lock_memory_hotplug();
 891
 892         zone = page_zone(pfn_to_page(start_pfn));
 893         node = zone_to_nid(zone);
 894         nr_pages = end_pfn - start_pfn;
 895
 896         /* set above range as isolated */
 897         ret = start_isolate_page_range(start_pfn, end_pfn, MIGRATE_MOVABLE);
 898         if (ret)
 899                 goto out;
 900
 901         arg.start_pfn = start_pfn;
 902         arg.nr_pages = nr_pages;
 903         arg.status_change_nid = -1;
 904         if (nr_pages >= node_present_pages(node))
 905                 arg.status_change_nid = node;
 906
 907         ret = memory_notify(MEM_GOING_OFFLINE, &arg);
 908         ret = notifier_to_errno(ret);
 909         if (ret)
 910                 goto failed_removal;
 911
 912         pfn = start_pfn;
 913         expire = jiffies + timeout;
 914         drain = 0;
 915         retry_max = 5;
 916 repeat:
 917         /* start memory hot removal */
 918         ret = -EAGAIN;
 919         if (time_after(jiffies, expire))
 920                 goto failed_removal;
 921         ret = -EINTR;
 922         if (signal_pending(current))
 923                 goto failed_removal;
 924         ret = 0;
 925         if (drain) {
 926                 lru_add_drain_all();
 927                 cond_resched();
 928                 drain_all_pages();
 929         }
 930
 931         pfn = scan_lru_pages(start_pfn, end_pfn);
 932         if (pfn) { /* We have page on LRU */
 933                 ret = do_migrate_range(pfn, end_pfn);
 934                 if (!ret) {
 935                         drain = 1;
 936                         goto repeat;
 937                 } else {
 938                         if (ret < 0)
 939                                 if (--retry_max == 0)
 940                                         goto failed_removal;
 941                         yield();
 942                         drain = 1;
 943                         goto repeat;
 944                 }
 945         }
 946         /* drain all zone's lru pagevec, this is asyncronous... */
 947         lru_add_drain_all();
 948         yield();
 949         /* drain pcp pages , this is synchrouns. */
 950         drain_all_pages();
 951         /* check again */
 952         offlined_pages = check_pages_isolated(start_pfn, end_pfn);
 953         if (offlined_pages < 0) {
 954                 ret = -EBUSY;
 955                 goto failed_removal;
 956         }
 957         printk(KERN_INFO "Offlined Pages %ld\n", offlined_pages);
 958         /* Ok, all of our target is islaoted.
 959            We cannot do rollback at this point. */
 960         offline_isolated_pages(start_pfn, end_pfn);
 961         /* reset pagetype flags and makes migrate type to be MOVABLE */
 962         undo_isolate_page_range(start_pfn, end_pfn, MIGRATE_MOVABLE);
 963         /* removal success */
 964         zone->present_pages -= offlined_pages;
 965         zone->zone_pgdat->node_present_pages -= offlined_pages;
 966         totalram_pages -= offlined_pages;
 967
 968         init_per_zone_wmark_min();
 969
 970         if (!populated_zone(zone)) {
 971                 zone_pcp_reset(zone);
 972                 mutex_lock(&zonelists_mutex);
 973                 build_all_zonelists(NULL, NULL);
 974                 mutex_unlock(&zonelists_mutex);
 975         } else
 976                 zone_pcp_update(zone);
 977
 978         if (!node_present_pages(node)) {
 979                 node_clear_state(node, N_HIGH_MEMORY);
 980                 kswapd_stop(node);
 981         }
 982
 983         vm_total_pages = nr_free_pagecache_pages();
 984         writeback_set_ratelimit();
 985
 986         memory_notify(MEM_OFFLINE, &arg);
 987         unlock_memory_hotplug();
 988         return 0;
 989
 990 failed_removal:
 991         printk(KERN_INFO "memory offlining [mem %#010llx-%#010llx] failed\n",
 992                (unsigned long long) start_pfn << PAGE_SHIFT,
 993                ((unsigned long long) end_pfn << PAGE_SHIFT) - 1);
 994         memory_notify(MEM_CANCEL_OFFLINE, &arg);
 995         /* pushback to free area */
 996         undo_isolate_page_range(start_pfn, end_pfn, MIGRATE_MOVABLE);
 997
 998 out:
 999         unlock_memory_hotplug();
1000         return ret;
1001 }
1002
1003 int offline_pages(unsigned long start_pfn, unsigned long nr_pages)
1004 {
1005         return __offline_pages(start_pfn, start_pfn + nr_pages, 120 * HZ);
1006 }
1007
1008 int remove_memory(u64 start, u64 size)
1009 {
1010         struct memory_block *mem = NULL;
1011         struct mem_section *section;
1012         unsigned long start_pfn, end_pfn;
1013         unsigned long pfn, section_nr;
1014         int ret;
1015
1016         start_pfn = PFN_DOWN(start);
1017         end_pfn = start_pfn + PFN_DOWN(size);
1018
1019         for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
1020                 section_nr = pfn_to_section_nr(pfn);
1021                 if (!present_section_nr(section_nr))
1022                         continue;
1023
1024                 section = __nr_to_section(section_nr);
1025                 /* same memblock? */
1026                 if (mem)
1027                         if ((section_nr >= mem->start_section_nr) &&
1028                             (section_nr <= mem->end_section_nr))
1029                                 continue;
1030
1031                 mem = find_memory_block_hinted(section, mem);
1032                 if (!mem)
1033                         continue;
1034
1035                 ret = offline_memory_block(mem);
1036                 if (ret) {
1037                         kobject_put(&mem->dev.kobj);
1038                         return ret;
1039                 }
1040         }
1041
1042         if (mem)
1043                 kobject_put(&mem->dev.kobj);
1044
1045         return 0;
1046 }
1047 #else
1048 int offline_pages(unsigned long start_pfn, unsigned long nr_pages)
1049 {
1050         return -EINVAL;
1051 }
1052 int remove_memory(u64 start, u64 size)
1053 {
1054         return -EINVAL;
1055 }
1056 #endif /* CONFIG_MEMORY_HOTREMOVE */
1057 EXPORT_SYMBOL_GPL(remove_memory);