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                 build_all_zonelists(NULL, zone);
 504         }
 505
 506         ret = walk_system_ram_range(pfn, nr_pages, &onlined_pages,
 507                 online_pages_range);
 508         if (ret) {
 509                 if (need_zonelists_rebuild)
 510                         zone_pcp_reset(zone);
 511                 mutex_unlock(&zonelists_mutex);
 512                 printk(KERN_DEBUG "online_pages [mem %#010llx-%#010llx] failed\n",
 513                        (unsigned long long) pfn << PAGE_SHIFT,
 514                        (((unsigned long long) pfn + nr_pages)
 515                             << PAGE_SHIFT) - 1);
 516                 memory_notify(MEM_CANCEL_ONLINE, &arg);
 517                 unlock_memory_hotplug();
 518                 return ret;
 519         }
 520
 521         zone->present_pages += onlined_pages;
 522         zone->zone_pgdat->node_present_pages += onlined_pages;
 523         if (onlined_pages) {
 524                 node_set_state(zone_to_nid(zone), N_HIGH_MEMORY);
 525                 if (need_zonelists_rebuild)
 526                         build_all_zonelists(NULL, NULL);
 527                 else
 528                         zone_pcp_update(zone);
 529         }
 530
 531         mutex_unlock(&zonelists_mutex);
 532
 533         init_per_zone_wmark_min();
 534
 535         if (onlined_pages)
 536                 kswapd_run(zone_to_nid(zone));
 537
 538         vm_total_pages = nr_free_pagecache_pages();
 539
 540         writeback_set_ratelimit();
 541
 542         if (onlined_pages)
 543                 memory_notify(MEM_ONLINE, &arg);
 544         unlock_memory_hotplug();
 545
 546         return 0;
 547 }
 548 #endif /* CONFIG_MEMORY_HOTPLUG_SPARSE */
 549
 550 /* we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG */
 551 static pg_data_t __ref *hotadd_new_pgdat(int nid, u64 start)
 552 {
 553         struct pglist_data *pgdat;
 554         unsigned long zones_size[MAX_NR_ZONES] = {0};
 555         unsigned long zholes_size[MAX_NR_ZONES] = {0};
 556         unsigned long start_pfn = start >> PAGE_SHIFT;
 557
 558         pgdat = arch_alloc_nodedata(nid);
 559         if (!pgdat)
 560                 return NULL;
 561
 562         arch_refresh_nodedata(nid, pgdat);
 563
 564         /* we can use NODE_DATA(nid) from here */
 565
 566         /* init node's zones as empty zones, we don't have any present pages.*/
 567         free_area_init_node(nid, zones_size, start_pfn, zholes_size);
 568
 569         /*
 570          * The node we allocated has no zone fallback lists. For avoiding
 571          * to access not-initialized zonelist, build here.
 572          */
 573         mutex_lock(&zonelists_mutex);
 574         build_all_zonelists(pgdat, NULL);
 575         mutex_unlock(&zonelists_mutex);
 576
 577         return pgdat;
 578 }
 579
 580 static void rollback_node_hotadd(int nid, pg_data_t *pgdat)
 581 {
 582         arch_refresh_nodedata(nid, NULL);
 583         arch_free_nodedata(pgdat);
 584         return;
 585 }
 586
 587
 588 /*
 589  * called by cpu_up() to online a node without onlined memory.
 590  */
 591 int mem_online_node(int nid)
 592 {
 593         pg_data_t       *pgdat;
 594         int     ret;
 595
 596         lock_memory_hotplug();
 597         pgdat = hotadd_new_pgdat(nid, 0);
 598         if (!pgdat) {
 599                 ret = -ENOMEM;
 600                 goto out;
 601         }
 602         node_set_online(nid);
 603         ret = register_one_node(nid);
 604         BUG_ON(ret);
 605
 606 out:
 607         unlock_memory_hotplug();
 608         return ret;
 609 }
 610
 611 /* we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG */
 612 int __ref add_memory(int nid, u64 start, u64 size)
 613 {
 614         pg_data_t *pgdat = NULL;
 615         int new_pgdat = 0;
 616         struct resource *res;
 617         int ret;
 618
 619         lock_memory_hotplug();
 620
 621         res = register_memory_resource(start, size);
 622         ret = -EEXIST;
 623         if (!res)
 624                 goto out;
 625
 626         if (!node_online(nid)) {
 627                 pgdat = hotadd_new_pgdat(nid, start);
 628                 ret = -ENOMEM;
 629                 if (!pgdat)
 630                         goto error;
 631                 new_pgdat = 1;
 632         }
 633
 634         /* call arch's memory hotadd */
 635         ret = arch_add_memory(nid, start, size);
 636
 637         if (ret < 0)
 638                 goto error;
 639
 640         /* we online node here. we can't roll back from here. */
 641         node_set_online(nid);
 642
 643         if (new_pgdat) {
 644                 ret = register_one_node(nid);
 645                 /*
 646                  * If sysfs file of new node can't create, cpu on the node
 647                  * can't be hot-added. There is no rollback way now.
 648                  * So, check by BUG_ON() to catch it reluctantly..
 649                  */
 650                 BUG_ON(ret);
 651         }
 652
 653         /* create new memmap entry */
 654         firmware_map_add_hotplug(start, start + size, "System RAM");
 655
 656         goto out;
 657
 658 error:
 659         /* rollback pgdat allocation and others */
 660         if (new_pgdat)
 661                 rollback_node_hotadd(nid, pgdat);
 662         if (res)
 663                 release_memory_resource(res);
 664
 665 out:
 666         unlock_memory_hotplug();
 667         return ret;
 668 }
 669 EXPORT_SYMBOL_GPL(add_memory);
 670
 671 #ifdef CONFIG_MEMORY_HOTREMOVE
 672 /*
 673  * A free page on the buddy free lists (not the per-cpu lists) has PageBuddy
 674  * set and the size of the free page is given by page_order(). Using this,
 675  * the function determines if the pageblock contains only free pages.
 676  * Due to buddy contraints, a free page at least the size of a pageblock will
 677  * be located at the start of the pageblock
 678  */
 679 static inline int pageblock_free(struct page *page)
 680 {
 681         return PageBuddy(page) && page_order(page) >= pageblock_order;
 682 }
 683
 684 /* Return the start of the next active pageblock after a given page */
 685 static struct page *next_active_pageblock(struct page *page)
 686 {
 687         /* Ensure the starting page is pageblock-aligned */
 688         BUG_ON(page_to_pfn(page) & (pageblock_nr_pages - 1));
 689
 690         /* If the entire pageblock is free, move to the end of free page */
 691         if (pageblock_free(page)) {
 692                 int order;
 693                 /* be careful. we don't have locks, page_order can be changed.*/
 694                 order = page_order(page);
 695                 if ((order < MAX_ORDER) && (order >= pageblock_order))
 696                         return page + (1 << order);
 697         }
 698
 699         return page + pageblock_nr_pages;
 700 }
 701
 702 /* Checks if this range of memory is likely to be hot-removable. */
 703 int is_mem_section_removable(unsigned long start_pfn, unsigned long nr_pages)
 704 {
 705         struct page *page = pfn_to_page(start_pfn);
 706         struct page *end_page = page + nr_pages;
 707
 708         /* Check the starting page of each pageblock within the range */
 709         for (; page < end_page; page = next_active_pageblock(page)) {
 710                 if (!is_pageblock_removable_nolock(page))
 711                         return 0;
 712                 cond_resched();
 713         }
 714
 715         /* All pageblocks in the memory block are likely to be hot-removable */
 716         return 1;
 717 }
 718
 719 /*
 720  * Confirm all pages in a range [start, end) is belongs to the same zone.
 721  */
 722 static int test_pages_in_a_zone(unsigned long start_pfn, unsigned long end_pfn)
 723 {
 724         unsigned long pfn;
 725         struct zone *zone = NULL;
 726         struct page *page;
 727         int i;
 728         for (pfn = start_pfn;
 729              pfn < end_pfn;
 730              pfn += MAX_ORDER_NR_PAGES) {
 731                 i = 0;
 732                 /* This is just a CONFIG_HOLES_IN_ZONE check.*/
 733                 while ((i < MAX_ORDER_NR_PAGES) && !pfn_valid_within(pfn + i))
 734                         i++;
 735                 if (i == MAX_ORDER_NR_PAGES)
 736                         continue;
 737                 page = pfn_to_page(pfn + i);
 738                 if (zone && page_zone(page) != zone)
 739                         return 0;
 740                 zone = page_zone(page);
 741         }
 742         return 1;
 743 }
 744
 745 /*
 746  * Scanning pfn is much easier than scanning lru list.
 747  * Scan pfn from start to end and Find LRU page.
 748  */
 749 static unsigned long scan_lru_pages(unsigned long start, unsigned long end)
 750 {
 751         unsigned long pfn;
 752         struct page *page;
 753         for (pfn = start; pfn < end; pfn++) {
 754                 if (pfn_valid(pfn)) {
 755                         page = pfn_to_page(pfn);
 756                         if (PageLRU(page))
 757                                 return pfn;
 758                 }
 759         }
 760         return 0;
 761 }
 762
 763 #define NR_OFFLINE_AT_ONCE_PAGES        (256)
 764 static int
 765 do_migrate_range(unsigned long start_pfn, unsigned long end_pfn)
 766 {
 767         unsigned long pfn;
 768         struct page *page;
 769         int move_pages = NR_OFFLINE_AT_ONCE_PAGES;
 770         int not_managed = 0;
 771         int ret = 0;
 772         LIST_HEAD(source);
 773
 774         for (pfn = start_pfn; pfn < end_pfn && move_pages > 0; pfn++) {
 775                 if (!pfn_valid(pfn))
 776                         continue;
 777                 page = pfn_to_page(pfn);
 778                 if (!get_page_unless_zero(page))
 779                         continue;
 780                 /*
 781                  * We can skip free pages. And we can only deal with pages on
 782                  * LRU.
 783                  */
 784                 ret = isolate_lru_page(page);
 785                 if (!ret) { /* Success */
 786                         put_page(page);
 787                         list_add_tail(&page->lru, &source);
 788                         move_pages--;
 789                         inc_zone_page_state(page, NR_ISOLATED_ANON +
 790                                             page_is_file_cache(page));
 791
 792                 } else {
 793 #ifdef CONFIG_DEBUG_VM
 794                         printk(KERN_ALERT "removing pfn %lx from LRU failed\n",
 795                                pfn);
 796                         dump_page(page);
 797 #endif
 798                         put_page(page);
 799                         /* Because we don't have big zone->lock. we should
 800                            check this again here. */
 801                         if (page_count(page)) {
 802                                 not_managed++;
 803                                 ret = -EBUSY;
 804                                 break;
 805                         }
 806                 }
 807         }
 808         if (!list_empty(&source)) {
 809                 if (not_managed) {
 810                         putback_lru_pages(&source);
 811                         goto out;
 812                 }
 813
 814                 /*
 815                  * alloc_migrate_target should be improooooved!!
 816                  * migrate_pages returns # of failed pages.
 817                  */
 818                 ret = migrate_pages(&source, alloc_migrate_target, 0,
 819                                                         true, MIGRATE_SYNC);
 820                 if (ret)
 821                         putback_lru_pages(&source);
 822         }
 823 out:
 824         return ret;
 825 }
 826
 827 /*
 828  * remove from free_area[] and mark all as Reserved.
 829  */
 830 static int
 831 offline_isolated_pages_cb(unsigned long start, unsigned long nr_pages,
 832                         void *data)
 833 {
 834         __offline_isolated_pages(start, start + nr_pages);
 835         return 0;
 836 }
 837
 838 static void
 839 offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
 840 {
 841         walk_system_ram_range(start_pfn, end_pfn - start_pfn, NULL,
 842                                 offline_isolated_pages_cb);
 843 }
 844
 845 /*
 846  * Check all pages in range, recoreded as memory resource, are isolated.
 847  */
 848 static int
 849 check_pages_isolated_cb(unsigned long start_pfn, unsigned long nr_pages,
 850                         void *data)
 851 {
 852         int ret;
 853         long offlined = *(long *)data;
 854         ret = test_pages_isolated(start_pfn, start_pfn + nr_pages, true);
 855         offlined = nr_pages;
 856         if (!ret)
 857                 *(long *)data += offlined;
 858         return ret;
 859 }
 860
 861 static long
 862 check_pages_isolated(unsigned long start_pfn, unsigned long end_pfn)
 863 {
 864         long offlined = 0;
 865         int ret;
 866
 867         ret = walk_system_ram_range(start_pfn, end_pfn - start_pfn, &offlined,
 868                         check_pages_isolated_cb);
 869         if (ret < 0)
 870                 offlined = (long)ret;
 871         return offlined;
 872 }
 873
 874 static int __ref __offline_pages(unsigned long start_pfn,
 875                   unsigned long end_pfn, unsigned long timeout)
 876 {
 877         unsigned long pfn, nr_pages, expire;
 878         long offlined_pages;
 879         int ret, drain, retry_max, node;
 880         struct zone *zone;
 881         struct memory_notify arg;
 882
 883         BUG_ON(start_pfn >= end_pfn);
 884         /* at least, alignment against pageblock is necessary */
 885         if (!IS_ALIGNED(start_pfn, pageblock_nr_pages))
 886                 return -EINVAL;
 887         if (!IS_ALIGNED(end_pfn, pageblock_nr_pages))
 888                 return -EINVAL;
 889         /* This makes hotplug much easier...and readable.
 890            we assume this for now. .*/
 891         if (!test_pages_in_a_zone(start_pfn, end_pfn))
 892                 return -EINVAL;
 893
 894         lock_memory_hotplug();
 895
 896         zone = page_zone(pfn_to_page(start_pfn));
 897         node = zone_to_nid(zone);
 898         nr_pages = end_pfn - start_pfn;
 899
 900         /* set above range as isolated */
 901         ret = start_isolate_page_range(start_pfn, end_pfn,
 902                                        MIGRATE_MOVABLE, true);
 903         if (ret)
 904                 goto out;
 905
 906         arg.start_pfn = start_pfn;
 907         arg.nr_pages = nr_pages;
 908         arg.status_change_nid = -1;
 909         if (nr_pages >= node_present_pages(node))
 910                 arg.status_change_nid = node;
 911
 912         ret = memory_notify(MEM_GOING_OFFLINE, &arg);
 913         ret = notifier_to_errno(ret);
 914         if (ret)
 915                 goto failed_removal;
 916
 917         pfn = start_pfn;
 918         expire = jiffies + timeout;
 919         drain = 0;
 920         retry_max = 5;
 921 repeat:
 922         /* start memory hot removal */
 923         ret = -EAGAIN;
 924         if (time_after(jiffies, expire))
 925                 goto failed_removal;
 926         ret = -EINTR;
 927         if (signal_pending(current))
 928                 goto failed_removal;
 929         ret = 0;
 930         if (drain) {
 931                 lru_add_drain_all();
 932                 cond_resched();
 933                 drain_all_pages();
 934         }
 935
 936         pfn = scan_lru_pages(start_pfn, end_pfn);
 937         if (pfn) { /* We have page on LRU */
 938                 ret = do_migrate_range(pfn, end_pfn);
 939                 if (!ret) {
 940                         drain = 1;
 941                         goto repeat;
 942                 } else {
 943                         if (ret < 0)
 944                                 if (--retry_max == 0)
 945                                         goto failed_removal;
 946                         yield();
 947                         drain = 1;
 948                         goto repeat;
 949                 }
 950         }
 951         /* drain all zone's lru pagevec, this is asyncronous... */
 952         lru_add_drain_all();
 953         yield();
 954         /* drain pcp pages , this is synchrouns. */
 955         drain_all_pages();
 956         /* check again */
 957         offlined_pages = check_pages_isolated(start_pfn, end_pfn);
 958         if (offlined_pages < 0) {
 959                 ret = -EBUSY;
 960                 goto failed_removal;
 961         }
 962         printk(KERN_INFO "Offlined Pages %ld\n", offlined_pages);
 963         /* Ok, all of our target is islaoted.
 964            We cannot do rollback at this point. */
 965         offline_isolated_pages(start_pfn, end_pfn);
 966         /* reset pagetype flags and makes migrate type to be MOVABLE */
 967         undo_isolate_page_range(start_pfn, end_pfn, MIGRATE_MOVABLE);
 968         /* removal success */
 969         zone->present_pages -= offlined_pages;
 970         zone->zone_pgdat->node_present_pages -= offlined_pages;
 971         totalram_pages -= offlined_pages;
 972
 973         init_per_zone_wmark_min();
 974
 975         if (!populated_zone(zone)) {
 976                 zone_pcp_reset(zone);
 977                 mutex_lock(&zonelists_mutex);
 978                 build_all_zonelists(NULL, NULL);
 979                 mutex_unlock(&zonelists_mutex);
 980         } else
 981                 zone_pcp_update(zone);
 982
 983         if (!node_present_pages(node)) {
 984                 node_clear_state(node, N_HIGH_MEMORY);
 985                 kswapd_stop(node);
 986         }
 987
 988         vm_total_pages = nr_free_pagecache_pages();
 989         writeback_set_ratelimit();
 990
 991         memory_notify(MEM_OFFLINE, &arg);
 992         unlock_memory_hotplug();
 993         return 0;
 994
 995 failed_removal:
 996         printk(KERN_INFO "memory offlining [mem %#010llx-%#010llx] failed\n",
 997                (unsigned long long) start_pfn << PAGE_SHIFT,
 998                ((unsigned long long) end_pfn << PAGE_SHIFT) - 1);
 999         memory_notify(MEM_CANCEL_OFFLINE, &arg);
1000         /* pushback to free area */
1001         undo_isolate_page_range(start_pfn, end_pfn, MIGRATE_MOVABLE);
1002
1003 out:
1004         unlock_memory_hotplug();
1005         return ret;
1006 }
1007
1008 int offline_pages(unsigned long start_pfn, unsigned long nr_pages)
1009 {
1010         return __offline_pages(start_pfn, start_pfn + nr_pages, 120 * HZ);
1011 }
1012
1013 int remove_memory(u64 start, u64 size)
1014 {
1015         struct memory_block *mem = NULL;
1016         struct mem_section *section;
1017         unsigned long start_pfn, end_pfn;
1018         unsigned long pfn, section_nr;
1019         int ret;
1020
1021         start_pfn = PFN_DOWN(start);
1022         end_pfn = start_pfn + PFN_DOWN(size);
1023
1024         for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
1025                 section_nr = pfn_to_section_nr(pfn);
1026                 if (!present_section_nr(section_nr))
1027                         continue;
1028
1029                 section = __nr_to_section(section_nr);
1030                 /* same memblock? */
1031                 if (mem)
1032                         if ((section_nr >= mem->start_section_nr) &&
1033                             (section_nr <= mem->end_section_nr))
1034                                 continue;
1035
1036                 mem = find_memory_block_hinted(section, mem);
1037                 if (!mem)
1038                         continue;
1039
1040                 ret = offline_memory_block(mem);
1041                 if (ret) {
1042                         kobject_put(&mem->dev.kobj);
1043                         return ret;
1044                 }
1045         }
1046
1047         if (mem)
1048                 kobject_put(&mem->dev.kobj);
1049
1050         return 0;
1051 }
1052 #else
1053 int offline_pages(unsigned long start_pfn, unsigned long nr_pages)
1054 {
1055         return -EINVAL;
1056 }
1057 int remove_memory(u64 start, u64 size)
1058 {
1059         return -EINVAL;
1060 }
1061 #endif /* CONFIG_MEMORY_HOTREMOVE */
1062 EXPORT_SYMBOL_GPL(remove_memory);