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/module.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/cpuset.h>
  26 #include <linux/delay.h>
  27 #include <linux/migrate.h>
  28 #include <linux/page-isolation.h>
  29
  30 #include <asm/tlbflush.h>
  31
  32 /* add this memory to iomem resource */
  33 static struct resource *register_memory_resource(u64 start, u64 size)
  34 {
  35         struct resource *res;
  36         res = kzalloc(sizeof(struct resource), GFP_KERNEL);
  37         BUG_ON(!res);
  38
  39         res->name = "System RAM";
  40         res->start = start;
  41         res->end = start + size - 1;
  42         res->flags = IORESOURCE_MEM | IORESOURCE_BUSY;
  43         if (request_resource(&iomem_resource, res) < 0) {
  44                 printk("System RAM resource %llx - %llx cannot be added\n",
  45                 (unsigned long long)res->start, (unsigned long long)res->end);
  46                 kfree(res);
  47                 res = NULL;
  48         }
  49         return res;
  50 }
  51
  52 static void release_memory_resource(struct resource *res)
  53 {
  54         if (!res)
  55                 return;
  56         release_resource(res);
  57         kfree(res);
  58         return;
  59 }
  60
  61
  62 #ifdef CONFIG_MEMORY_HOTPLUG_SPARSE
  63 static int __add_zone(struct zone *zone, unsigned long phys_start_pfn)
  64 {
  65         struct pglist_data *pgdat = zone->zone_pgdat;
  66         int nr_pages = PAGES_PER_SECTION;
  67         int nid = pgdat->node_id;
  68         int zone_type;
  69
  70         zone_type = zone - pgdat->node_zones;
  71         if (!zone->wait_table) {
  72                 int ret = 0;
  73                 ret = init_currently_empty_zone(zone, phys_start_pfn,
  74                                                 nr_pages, MEMMAP_HOTPLUG);
  75                 if (ret < 0)
  76                         return ret;
  77         }
  78         memmap_init_zone(nr_pages, nid, zone_type,
  79                          phys_start_pfn, MEMMAP_HOTPLUG);
  80         return 0;
  81 }
  82
  83 static int __add_section(struct zone *zone, unsigned long phys_start_pfn)
  84 {
  85         int nr_pages = PAGES_PER_SECTION;
  86         int ret;
  87
  88         if (pfn_valid(phys_start_pfn))
  89                 return -EEXIST;
  90
  91         ret = sparse_add_one_section(zone, phys_start_pfn, nr_pages);
  92
  93         if (ret < 0)
  94                 return ret;
  95
  96         ret = __add_zone(zone, phys_start_pfn);
  97
  98         if (ret < 0)
  99                 return ret;
 100
 101         return register_new_memory(__pfn_to_section(phys_start_pfn));
 102 }
 103
 104 static int __remove_section(struct zone *zone, struct mem_section *ms)
 105 {
 106         unsigned long flags;
 107         struct pglist_data *pgdat = zone->zone_pgdat;
 108         int ret = -EINVAL;
 109
 110         if (!valid_section(ms))
 111                 return ret;
 112
 113         ret = unregister_memory_section(ms);
 114         if (ret)
 115                 return ret;
 116
 117         pgdat_resize_lock(pgdat, &flags);
 118         sparse_remove_one_section(zone, ms);
 119         pgdat_resize_unlock(pgdat, &flags);
 120         return 0;
 121 }
 122
 123 /*
 124  * Reasonably generic function for adding memory.  It is
 125  * expected that archs that support memory hotplug will
 126  * call this function after deciding the zone to which to
 127  * add the new pages.
 128  */
 129 int __add_pages(struct zone *zone, unsigned long phys_start_pfn,
 130                  unsigned long nr_pages)
 131 {
 132         unsigned long i;
 133         int err = 0;
 134         int start_sec, end_sec;
 135         /* during initialize mem_map, align hot-added range to section */
 136         start_sec = pfn_to_section_nr(phys_start_pfn);
 137         end_sec = pfn_to_section_nr(phys_start_pfn + nr_pages - 1);
 138
 139         for (i = start_sec; i <= end_sec; i++) {
 140                 err = __add_section(zone, i << PFN_SECTION_SHIFT);
 141
 142                 /*
 143                  * EEXIST is finally dealt with by ioresource collision
 144                  * check. see add_memory() => register_memory_resource()
 145                  * Warning will be printed if there is collision.
 146                  */
 147                 if (err && (err != -EEXIST))
 148                         break;
 149                 err = 0;
 150         }
 151
 152         return err;
 153 }
 154 EXPORT_SYMBOL_GPL(__add_pages);
 155
 156 /**
 157  * __remove_pages() - remove sections of pages from a zone
 158  * @zone: zone from which pages need to be removed
 159  * @phys_start_pfn: starting pageframe (must be aligned to start of a section)
 160  * @nr_pages: number of pages to remove (must be multiple of section size)
 161  *
 162  * Generic helper function to remove section mappings and sysfs entries
 163  * for the section of the memory we are removing. Caller needs to make
 164  * sure that pages are marked reserved and zones are adjust properly by
 165  * calling offline_pages().
 166  */
 167 int __remove_pages(struct zone *zone, unsigned long phys_start_pfn,
 168                  unsigned long nr_pages)
 169 {
 170         unsigned long i, ret = 0;
 171         int sections_to_remove;
 172
 173         /*
 174          * We can only remove entire sections
 175          */
 176         BUG_ON(phys_start_pfn & ~PAGE_SECTION_MASK);
 177         BUG_ON(nr_pages % PAGES_PER_SECTION);
 178
 179         release_mem_region(phys_start_pfn << PAGE_SHIFT, nr_pages * PAGE_SIZE);
 180
 181         sections_to_remove = nr_pages / PAGES_PER_SECTION;
 182         for (i = 0; i < sections_to_remove; i++) {
 183                 unsigned long pfn = phys_start_pfn + i*PAGES_PER_SECTION;
 184                 ret = __remove_section(zone, __pfn_to_section(pfn));
 185                 if (ret)
 186                         break;
 187         }
 188         return ret;
 189 }
 190 EXPORT_SYMBOL_GPL(__remove_pages);
 191
 192 static void grow_zone_span(struct zone *zone,
 193                 unsigned long start_pfn, unsigned long end_pfn)
 194 {
 195         unsigned long old_zone_end_pfn;
 196
 197         zone_span_writelock(zone);
 198
 199         old_zone_end_pfn = zone->zone_start_pfn + zone->spanned_pages;
 200         if (start_pfn < zone->zone_start_pfn)
 201                 zone->zone_start_pfn = start_pfn;
 202
 203         zone->spanned_pages = max(old_zone_end_pfn, end_pfn) -
 204                                 zone->zone_start_pfn;
 205
 206         zone_span_writeunlock(zone);
 207 }
 208
 209 static void grow_pgdat_span(struct pglist_data *pgdat,
 210                 unsigned long start_pfn, unsigned long end_pfn)
 211 {
 212         unsigned long old_pgdat_end_pfn =
 213                 pgdat->node_start_pfn + pgdat->node_spanned_pages;
 214
 215         if (start_pfn < pgdat->node_start_pfn)
 216                 pgdat->node_start_pfn = start_pfn;
 217
 218         pgdat->node_spanned_pages = max(old_pgdat_end_pfn, end_pfn) -
 219                                         pgdat->node_start_pfn;
 220 }
 221
 222 void online_page(struct page *page)
 223 {
 224         totalram_pages++;
 225         num_physpages++;
 226
 227 #ifdef CONFIG_HIGHMEM
 228         if (PageHighMem(page))
 229                 totalhigh_pages++;
 230 #endif
 231
 232 #ifdef CONFIG_FLATMEM
 233         max_mapnr = max(page_to_pfn(page), max_mapnr);
 234 #endif
 235
 236         ClearPageReserved(page);
 237         init_page_count(page);
 238         __free_page(page);
 239 }
 240
 241 static int online_pages_range(unsigned long start_pfn, unsigned long nr_pages,
 242                         void *arg)
 243 {
 244         unsigned long i;
 245         unsigned long onlined_pages = *(unsigned long *)arg;
 246         struct page *page;
 247         if (PageReserved(pfn_to_page(start_pfn)))
 248                 for (i = 0; i < nr_pages; i++) {
 249                         page = pfn_to_page(start_pfn + i);
 250                         online_page(page);
 251                         onlined_pages++;
 252                 }
 253         *(unsigned long *)arg = onlined_pages;
 254         return 0;
 255 }
 256
 257
 258 int online_pages(unsigned long pfn, unsigned long nr_pages)
 259 {
 260         unsigned long flags;
 261         unsigned long onlined_pages = 0;
 262         struct zone *zone;
 263         int need_zonelists_rebuild = 0;
 264         int nid;
 265         int ret;
 266         struct memory_notify arg;
 267
 268         arg.start_pfn = pfn;
 269         arg.nr_pages = nr_pages;
 270         arg.status_change_nid = -1;
 271
 272         nid = page_to_nid(pfn_to_page(pfn));
 273         if (node_present_pages(nid) == 0)
 274                 arg.status_change_nid = nid;
 275
 276         ret = memory_notify(MEM_GOING_ONLINE, &arg);
 277         ret = notifier_to_errno(ret);
 278         if (ret) {
 279                 memory_notify(MEM_CANCEL_ONLINE, &arg);
 280                 return ret;
 281         }
 282         /*
 283          * This doesn't need a lock to do pfn_to_page().
 284          * The section can't be removed here because of the
 285          * memory_block->state_mutex.
 286          */
 287         zone = page_zone(pfn_to_page(pfn));
 288         pgdat_resize_lock(zone->zone_pgdat, &flags);
 289         grow_zone_span(zone, pfn, pfn + nr_pages);
 290         grow_pgdat_span(zone->zone_pgdat, pfn, pfn + nr_pages);
 291         pgdat_resize_unlock(zone->zone_pgdat, &flags);
 292
 293         /*
 294          * If this zone is not populated, then it is not in zonelist.
 295          * This means the page allocator ignores this zone.
 296          * So, zonelist must be updated after online.
 297          */
 298         if (!populated_zone(zone))
 299                 need_zonelists_rebuild = 1;
 300
 301         walk_memory_resource(pfn, nr_pages, &onlined_pages,
 302                 online_pages_range);
 303         zone->present_pages += onlined_pages;
 304         zone->zone_pgdat->node_present_pages += onlined_pages;
 305
 306         setup_per_zone_pages_min();
 307         if (onlined_pages) {
 308                 kswapd_run(zone_to_nid(zone));
 309                 node_set_state(zone_to_nid(zone), N_HIGH_MEMORY);
 310         }
 311
 312         if (need_zonelists_rebuild)
 313                 build_all_zonelists();
 314         vm_total_pages = nr_free_pagecache_pages();
 315         writeback_set_ratelimit();
 316
 317         if (onlined_pages)
 318                 memory_notify(MEM_ONLINE, &arg);
 319
 320         return 0;
 321 }
 322 #endif /* CONFIG_MEMORY_HOTPLUG_SPARSE */
 323
 324 static pg_data_t *hotadd_new_pgdat(int nid, u64 start)
 325 {
 326         struct pglist_data *pgdat;
 327         unsigned long zones_size[MAX_NR_ZONES] = {0};
 328         unsigned long zholes_size[MAX_NR_ZONES] = {0};
 329         unsigned long start_pfn = start >> PAGE_SHIFT;
 330
 331         pgdat = arch_alloc_nodedata(nid);
 332         if (!pgdat)
 333                 return NULL;
 334
 335         arch_refresh_nodedata(nid, pgdat);
 336
 337         /* we can use NODE_DATA(nid) from here */
 338
 339         /* init node's zones as empty zones, we don't have any present pages.*/
 340         free_area_init_node(nid, pgdat, zones_size, start_pfn, zholes_size);
 341
 342         return pgdat;
 343 }
 344
 345 static void rollback_node_hotadd(int nid, pg_data_t *pgdat)
 346 {
 347         arch_refresh_nodedata(nid, NULL);
 348         arch_free_nodedata(pgdat);
 349         return;
 350 }
 351
 352
 353 int add_memory(int nid, u64 start, u64 size)
 354 {
 355         pg_data_t *pgdat = NULL;
 356         int new_pgdat = 0;
 357         struct resource *res;
 358         int ret;
 359
 360         res = register_memory_resource(start, size);
 361         if (!res)
 362                 return -EEXIST;
 363
 364         if (!node_online(nid)) {
 365                 pgdat = hotadd_new_pgdat(nid, start);
 366                 if (!pgdat)
 367                         return -ENOMEM;
 368                 new_pgdat = 1;
 369         }
 370
 371         /* call arch's memory hotadd */
 372         ret = arch_add_memory(nid, start, size);
 373
 374         if (ret < 0)
 375                 goto error;
 376
 377         /* we online node here. we can't roll back from here. */
 378         node_set_online(nid);
 379
 380         cpuset_track_online_nodes();
 381
 382         if (new_pgdat) {
 383                 ret = register_one_node(nid);
 384                 /*
 385                  * If sysfs file of new node can't create, cpu on the node
 386                  * can't be hot-added. There is no rollback way now.
 387                  * So, check by BUG_ON() to catch it reluctantly..
 388                  */
 389                 BUG_ON(ret);
 390         }
 391
 392         return ret;
 393 error:
 394         /* rollback pgdat allocation and others */
 395         if (new_pgdat)
 396                 rollback_node_hotadd(nid, pgdat);
 397         if (res)
 398                 release_memory_resource(res);
 399
 400         return ret;
 401 }
 402 EXPORT_SYMBOL_GPL(add_memory);
 403
 404 #ifdef CONFIG_MEMORY_HOTREMOVE
 405 /*
 406  * Confirm all pages in a range [start, end) is belongs to the same zone.
 407  */
 408 static int test_pages_in_a_zone(unsigned long start_pfn, unsigned long end_pfn)
 409 {
 410         unsigned long pfn;
 411         struct zone *zone = NULL;
 412         struct page *page;
 413         int i;
 414         for (pfn = start_pfn;
 415              pfn < end_pfn;
 416              pfn += MAX_ORDER_NR_PAGES) {
 417                 i = 0;
 418                 /* This is just a CONFIG_HOLES_IN_ZONE check.*/
 419                 while ((i < MAX_ORDER_NR_PAGES) && !pfn_valid_within(pfn + i))
 420                         i++;
 421                 if (i == MAX_ORDER_NR_PAGES)
 422                         continue;
 423                 page = pfn_to_page(pfn + i);
 424                 if (zone && page_zone(page) != zone)
 425                         return 0;
 426                 zone = page_zone(page);
 427         }
 428         return 1;
 429 }
 430
 431 /*
 432  * Scanning pfn is much easier than scanning lru list.
 433  * Scan pfn from start to end and Find LRU page.
 434  */
 435 int scan_lru_pages(unsigned long start, unsigned long end)
 436 {
 437         unsigned long pfn;
 438         struct page *page;
 439         for (pfn = start; pfn < end; pfn++) {
 440                 if (pfn_valid(pfn)) {
 441                         page = pfn_to_page(pfn);
 442                         if (PageLRU(page))
 443                                 return pfn;
 444                 }
 445         }
 446         return 0;
 447 }
 448
 449 static struct page *
 450 hotremove_migrate_alloc(struct page *page,
 451                         unsigned long private,
 452                         int **x)
 453 {
 454         /* This should be improoooooved!! */
 455         return alloc_page(GFP_HIGHUSER_PAGECACHE);
 456 }
 457
 458
 459 #define NR_OFFLINE_AT_ONCE_PAGES        (256)
 460 static int
 461 do_migrate_range(unsigned long start_pfn, unsigned long end_pfn)
 462 {
 463         unsigned long pfn;
 464         struct page *page;
 465         int move_pages = NR_OFFLINE_AT_ONCE_PAGES;
 466         int not_managed = 0;
 467         int ret = 0;
 468         LIST_HEAD(source);
 469
 470         for (pfn = start_pfn; pfn < end_pfn && move_pages > 0; pfn++) {
 471                 if (!pfn_valid(pfn))
 472                         continue;
 473                 page = pfn_to_page(pfn);
 474                 if (!page_count(page))
 475                         continue;
 476                 /*
 477                  * We can skip free pages. And we can only deal with pages on
 478                  * LRU.
 479                  */
 480                 ret = isolate_lru_page(page, &source);
 481                 if (!ret) { /* Success */
 482                         move_pages--;
 483                 } else {
 484                         /* Becasue we don't have big zone->lock. we should
 485                            check this again here. */
 486                         if (page_count(page))
 487                                 not_managed++;
 488 #ifdef CONFIG_DEBUG_VM
 489                         printk(KERN_INFO "removing from LRU failed"
 490                                          " %lx/%d/%lx\n",
 491                                 pfn, page_count(page), page->flags);
 492 #endif
 493                 }
 494         }
 495         ret = -EBUSY;
 496         if (not_managed) {
 497                 if (!list_empty(&source))
 498                         putback_lru_pages(&source);
 499                 goto out;
 500         }
 501         ret = 0;
 502         if (list_empty(&source))
 503                 goto out;
 504         /* this function returns # of failed pages */
 505         ret = migrate_pages(&source, hotremove_migrate_alloc, 0);
 506
 507 out:
 508         return ret;
 509 }
 510
 511 /*
 512  * remove from free_area[] and mark all as Reserved.
 513  */
 514 static int
 515 offline_isolated_pages_cb(unsigned long start, unsigned long nr_pages,
 516                         void *data)
 517 {
 518         __offline_isolated_pages(start, start + nr_pages);
 519         return 0;
 520 }
 521
 522 static void
 523 offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
 524 {
 525         walk_memory_resource(start_pfn, end_pfn - start_pfn, NULL,
 526                                 offline_isolated_pages_cb);
 527 }
 528
 529 /*
 530  * Check all pages in range, recoreded as memory resource, are isolated.
 531  */
 532 static int
 533 check_pages_isolated_cb(unsigned long start_pfn, unsigned long nr_pages,
 534                         void *data)
 535 {
 536         int ret;
 537         long offlined = *(long *)data;
 538         ret = test_pages_isolated(start_pfn, start_pfn + nr_pages);
 539         offlined = nr_pages;
 540         if (!ret)
 541                 *(long *)data += offlined;
 542         return ret;
 543 }
 544
 545 static long
 546 check_pages_isolated(unsigned long start_pfn, unsigned long end_pfn)
 547 {
 548         long offlined = 0;
 549         int ret;
 550
 551         ret = walk_memory_resource(start_pfn, end_pfn - start_pfn, &offlined,
 552                         check_pages_isolated_cb);
 553         if (ret < 0)
 554                 offlined = (long)ret;
 555         return offlined;
 556 }
 557
 558 int offline_pages(unsigned long start_pfn,
 559                   unsigned long end_pfn, unsigned long timeout)
 560 {
 561         unsigned long pfn, nr_pages, expire;
 562         long offlined_pages;
 563         int ret, drain, retry_max, node;
 564         struct zone *zone;
 565         struct memory_notify arg;
 566
 567         BUG_ON(start_pfn >= end_pfn);
 568         /* at least, alignment against pageblock is necessary */
 569         if (!IS_ALIGNED(start_pfn, pageblock_nr_pages))
 570                 return -EINVAL;
 571         if (!IS_ALIGNED(end_pfn, pageblock_nr_pages))
 572                 return -EINVAL;
 573         /* This makes hotplug much easier...and readable.
 574            we assume this for now. .*/
 575         if (!test_pages_in_a_zone(start_pfn, end_pfn))
 576                 return -EINVAL;
 577
 578         zone = page_zone(pfn_to_page(start_pfn));
 579         node = zone_to_nid(zone);
 580         nr_pages = end_pfn - start_pfn;
 581
 582         /* set above range as isolated */
 583         ret = start_isolate_page_range(start_pfn, end_pfn);
 584         if (ret)
 585                 return ret;
 586
 587         arg.start_pfn = start_pfn;
 588         arg.nr_pages = nr_pages;
 589         arg.status_change_nid = -1;
 590         if (nr_pages >= node_present_pages(node))
 591                 arg.status_change_nid = node;
 592
 593         ret = memory_notify(MEM_GOING_OFFLINE, &arg);
 594         ret = notifier_to_errno(ret);
 595         if (ret)
 596                 goto failed_removal;
 597
 598         pfn = start_pfn;
 599         expire = jiffies + timeout;
 600         drain = 0;
 601         retry_max = 5;
 602 repeat:
 603         /* start memory hot removal */
 604         ret = -EAGAIN;
 605         if (time_after(jiffies, expire))
 606                 goto failed_removal;
 607         ret = -EINTR;
 608         if (signal_pending(current))
 609                 goto failed_removal;
 610         ret = 0;
 611         if (drain) {
 612                 lru_add_drain_all();
 613                 flush_scheduled_work();
 614                 cond_resched();
 615                 drain_all_pages();
 616         }
 617
 618         pfn = scan_lru_pages(start_pfn, end_pfn);
 619         if (pfn) { /* We have page on LRU */
 620                 ret = do_migrate_range(pfn, end_pfn);
 621                 if (!ret) {
 622                         drain = 1;
 623                         goto repeat;
 624                 } else {
 625                         if (ret < 0)
 626                                 if (--retry_max == 0)
 627                                         goto failed_removal;
 628                         yield();
 629                         drain = 1;
 630                         goto repeat;
 631                 }
 632         }
 633         /* drain all zone's lru pagevec, this is asyncronous... */
 634         lru_add_drain_all();
 635         flush_scheduled_work();
 636         yield();
 637         /* drain pcp pages , this is synchrouns. */
 638         drain_all_pages();
 639         /* check again */
 640         offlined_pages = check_pages_isolated(start_pfn, end_pfn);
 641         if (offlined_pages < 0) {
 642                 ret = -EBUSY;
 643                 goto failed_removal;
 644         }
 645         printk(KERN_INFO "Offlined Pages %ld\n", offlined_pages);
 646         /* Ok, all of our target is islaoted.
 647            We cannot do rollback at this point. */
 648         offline_isolated_pages(start_pfn, end_pfn);
 649         /* reset pagetype flags and makes migrate type to be MOVABLE */
 650         undo_isolate_page_range(start_pfn, end_pfn);
 651         /* removal success */
 652         zone->present_pages -= offlined_pages;
 653         zone->zone_pgdat->node_present_pages -= offlined_pages;
 654         totalram_pages -= offlined_pages;
 655         num_physpages -= offlined_pages;
 656
 657         vm_total_pages = nr_free_pagecache_pages();
 658         writeback_set_ratelimit();
 659
 660         memory_notify(MEM_OFFLINE, &arg);
 661         return 0;
 662
 663 failed_removal:
 664         printk(KERN_INFO "memory offlining %lx to %lx failed\n",
 665                 start_pfn, end_pfn);
 666         memory_notify(MEM_CANCEL_OFFLINE, &arg);
 667         /* pushback to free area */
 668         undo_isolate_page_range(start_pfn, end_pfn);
 669
 670         return ret;
 671 }
 672 #else
 673 int remove_memory(u64 start, u64 size)
 674 {
 675         return -EINVAL;
 676 }
 677 EXPORT_SYMBOL_GPL(remove_memory);
 678 #endif /* CONFIG_MEMORY_HOTREMOVE */