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