1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * linux/mm/memory_hotplug.c 4 * 5 * Copyright (C) 6 */ 7 8 #include <linux/stddef.h> 9 #include <linux/mm.h> 10 #include <linux/sched/signal.h> 11 #include <linux/swap.h> 12 #include <linux/interrupt.h> 13 #include <linux/pagemap.h> 14 #include <linux/compiler.h> 15 #include <linux/export.h> 16 #include <linux/pagevec.h> 17 #include <linux/writeback.h> 18 #include <linux/slab.h> 19 #include <linux/sysctl.h> 20 #include <linux/cpu.h> 21 #include <linux/memory.h> 22 #include <linux/memremap.h> 23 #include <linux/memory_hotplug.h> 24 #include <linux/highmem.h> 25 #include <linux/vmalloc.h> 26 #include <linux/ioport.h> 27 #include <linux/delay.h> 28 #include <linux/migrate.h> 29 #include <linux/page-isolation.h> 30 #include <linux/pfn.h> 31 #include <linux/suspend.h> 32 #include <linux/mm_inline.h> 33 #include <linux/firmware-map.h> 34 #include <linux/stop_machine.h> 35 #include <linux/hugetlb.h> 36 #include <linux/memblock.h> 37 #include <linux/compaction.h> 38 #include <linux/rmap.h> 39 40 #include <asm/tlbflush.h> 41 42 #include "internal.h" 43 #include "shuffle.h" 44 45 /* 46 * online_page_callback contains pointer to current page onlining function. 47 * Initially it is generic_online_page(). If it is required it could be 48 * changed by calling set_online_page_callback() for callback registration 49 * and restore_online_page_callback() for generic callback restore. 50 */ 51 52 static online_page_callback_t online_page_callback = generic_online_page; 53 static DEFINE_MUTEX(online_page_callback_lock); 54 55 DEFINE_STATIC_PERCPU_RWSEM(mem_hotplug_lock); 56 57 void get_online_mems(void) 58 { 59 percpu_down_read(&mem_hotplug_lock); 60 } 61 62 void put_online_mems(void) 63 { 64 percpu_up_read(&mem_hotplug_lock); 65 } 66 67 bool movable_node_enabled = false; 68 69 #ifndef CONFIG_MEMORY_HOTPLUG_DEFAULT_ONLINE 70 bool memhp_auto_online; 71 #else 72 bool memhp_auto_online = true; 73 #endif 74 EXPORT_SYMBOL_GPL(memhp_auto_online); 75 76 static int __init setup_memhp_default_state(char *str) 77 { 78 if (!strcmp(str, "online")) 79 memhp_auto_online = true; 80 else if (!strcmp(str, "offline")) 81 memhp_auto_online = false; 82 83 return 1; 84 } 85 __setup("memhp_default_state=", setup_memhp_default_state); 86 87 void mem_hotplug_begin(void) 88 { 89 cpus_read_lock(); 90 percpu_down_write(&mem_hotplug_lock); 91 } 92 93 void mem_hotplug_done(void) 94 { 95 percpu_up_write(&mem_hotplug_lock); 96 cpus_read_unlock(); 97 } 98 99 u64 max_mem_size = U64_MAX; 100 101 /* add this memory to iomem resource */ 102 static struct resource *register_memory_resource(u64 start, u64 size) 103 { 104 struct resource *res; 105 unsigned long flags = IORESOURCE_SYSTEM_RAM | IORESOURCE_BUSY; 106 char *resource_name = "System RAM"; 107 108 if (start + size > max_mem_size) 109 return ERR_PTR(-E2BIG); 110 111 /* 112 * Request ownership of the new memory range. This might be 113 * a child of an existing resource that was present but 114 * not marked as busy. 115 */ 116 res = __request_region(&iomem_resource, start, size, 117 resource_name, flags); 118 119 if (!res) { 120 pr_debug("Unable to reserve System RAM region: %016llx->%016llx\n", 121 start, start + size); 122 return ERR_PTR(-EEXIST); 123 } 124 return res; 125 } 126 127 static void release_memory_resource(struct resource *res) 128 { 129 if (!res) 130 return; 131 release_resource(res); 132 kfree(res); 133 } 134 135 #ifdef CONFIG_MEMORY_HOTPLUG_SPARSE 136 void get_page_bootmem(unsigned long info, struct page *page, 137 unsigned long type) 138 { 139 page->freelist = (void *)type; 140 SetPagePrivate(page); 141 set_page_private(page, info); 142 page_ref_inc(page); 143 } 144 145 void put_page_bootmem(struct page *page) 146 { 147 unsigned long type; 148 149 type = (unsigned long) page->freelist; 150 BUG_ON(type < MEMORY_HOTPLUG_MIN_BOOTMEM_TYPE || 151 type > MEMORY_HOTPLUG_MAX_BOOTMEM_TYPE); 152 153 if (page_ref_dec_return(page) == 1) { 154 page->freelist = NULL; 155 ClearPagePrivate(page); 156 set_page_private(page, 0); 157 INIT_LIST_HEAD(&page->lru); 158 free_reserved_page(page); 159 } 160 } 161 162 #ifdef CONFIG_HAVE_BOOTMEM_INFO_NODE 163 #ifndef CONFIG_SPARSEMEM_VMEMMAP 164 static void register_page_bootmem_info_section(unsigned long start_pfn) 165 { 166 unsigned long mapsize, section_nr, i; 167 struct mem_section *ms; 168 struct page *page, *memmap; 169 struct mem_section_usage *usage; 170 171 section_nr = pfn_to_section_nr(start_pfn); 172 ms = __nr_to_section(section_nr); 173 174 /* Get section's memmap address */ 175 memmap = sparse_decode_mem_map(ms->section_mem_map, section_nr); 176 177 /* 178 * Get page for the memmap's phys address 179 * XXX: need more consideration for sparse_vmemmap... 180 */ 181 page = virt_to_page(memmap); 182 mapsize = sizeof(struct page) * PAGES_PER_SECTION; 183 mapsize = PAGE_ALIGN(mapsize) >> PAGE_SHIFT; 184 185 /* remember memmap's page */ 186 for (i = 0; i < mapsize; i++, page++) 187 get_page_bootmem(section_nr, page, SECTION_INFO); 188 189 usage = ms->usage; 190 page = virt_to_page(usage); 191 192 mapsize = PAGE_ALIGN(mem_section_usage_size()) >> PAGE_SHIFT; 193 194 for (i = 0; i < mapsize; i++, page++) 195 get_page_bootmem(section_nr, page, MIX_SECTION_INFO); 196 197 } 198 #else /* CONFIG_SPARSEMEM_VMEMMAP */ 199 static void register_page_bootmem_info_section(unsigned long start_pfn) 200 { 201 unsigned long mapsize, section_nr, i; 202 struct mem_section *ms; 203 struct page *page, *memmap; 204 struct mem_section_usage *usage; 205 206 section_nr = pfn_to_section_nr(start_pfn); 207 ms = __nr_to_section(section_nr); 208 209 memmap = sparse_decode_mem_map(ms->section_mem_map, section_nr); 210 211 register_page_bootmem_memmap(section_nr, memmap, PAGES_PER_SECTION); 212 213 usage = ms->usage; 214 page = virt_to_page(usage); 215 216 mapsize = PAGE_ALIGN(mem_section_usage_size()) >> PAGE_SHIFT; 217 218 for (i = 0; i < mapsize; i++, page++) 219 get_page_bootmem(section_nr, page, MIX_SECTION_INFO); 220 } 221 #endif /* !CONFIG_SPARSEMEM_VMEMMAP */ 222 223 void __init register_page_bootmem_info_node(struct pglist_data *pgdat) 224 { 225 unsigned long i, pfn, end_pfn, nr_pages; 226 int node = pgdat->node_id; 227 struct page *page; 228 229 nr_pages = PAGE_ALIGN(sizeof(struct pglist_data)) >> PAGE_SHIFT; 230 page = virt_to_page(pgdat); 231 232 for (i = 0; i < nr_pages; i++, page++) 233 get_page_bootmem(node, page, NODE_INFO); 234 235 pfn = pgdat->node_start_pfn; 236 end_pfn = pgdat_end_pfn(pgdat); 237 238 /* register section info */ 239 for (; pfn < end_pfn; pfn += PAGES_PER_SECTION) { 240 /* 241 * Some platforms can assign the same pfn to multiple nodes - on 242 * node0 as well as nodeN. To avoid registering a pfn against 243 * multiple nodes we check that this pfn does not already 244 * reside in some other nodes. 245 */ 246 if (pfn_valid(pfn) && (early_pfn_to_nid(pfn) == node)) 247 register_page_bootmem_info_section(pfn); 248 } 249 } 250 #endif /* CONFIG_HAVE_BOOTMEM_INFO_NODE */ 251 252 static int check_pfn_span(unsigned long pfn, unsigned long nr_pages, 253 const char *reason) 254 { 255 /* 256 * Disallow all operations smaller than a sub-section and only 257 * allow operations smaller than a section for 258 * SPARSEMEM_VMEMMAP. Note that check_hotplug_memory_range() 259 * enforces a larger memory_block_size_bytes() granularity for 260 * memory that will be marked online, so this check should only 261 * fire for direct arch_{add,remove}_memory() users outside of 262 * add_memory_resource(). 263 */ 264 unsigned long min_align; 265 266 if (IS_ENABLED(CONFIG_SPARSEMEM_VMEMMAP)) 267 min_align = PAGES_PER_SUBSECTION; 268 else 269 min_align = PAGES_PER_SECTION; 270 if (!IS_ALIGNED(pfn, min_align) 271 || !IS_ALIGNED(nr_pages, min_align)) { 272 WARN(1, "Misaligned __%s_pages start: %#lx end: #%lx\n", 273 reason, pfn, pfn + nr_pages - 1); 274 return -EINVAL; 275 } 276 return 0; 277 } 278 279 static int check_hotplug_memory_addressable(unsigned long pfn, 280 unsigned long nr_pages) 281 { 282 const u64 max_addr = PFN_PHYS(pfn + nr_pages) - 1; 283 284 if (max_addr >> MAX_PHYSMEM_BITS) { 285 const u64 max_allowed = (1ull << (MAX_PHYSMEM_BITS + 1)) - 1; 286 WARN(1, 287 "Hotplugged memory exceeds maximum addressable address, range=%#llx-%#llx, maximum=%#llx\n", 288 (u64)PFN_PHYS(pfn), max_addr, max_allowed); 289 return -E2BIG; 290 } 291 292 return 0; 293 } 294 295 /* 296 * Reasonably generic function for adding memory. It is 297 * expected that archs that support memory hotplug will 298 * call this function after deciding the zone to which to 299 * add the new pages. 300 */ 301 int __ref __add_pages(int nid, unsigned long pfn, unsigned long nr_pages, 302 struct mhp_restrictions *restrictions) 303 { 304 int err; 305 unsigned long nr, start_sec, end_sec; 306 struct vmem_altmap *altmap = restrictions->altmap; 307 308 err = check_hotplug_memory_addressable(pfn, nr_pages); 309 if (err) 310 return err; 311 312 if (altmap) { 313 /* 314 * Validate altmap is within bounds of the total request 315 */ 316 if (altmap->base_pfn != pfn 317 || vmem_altmap_offset(altmap) > nr_pages) { 318 pr_warn_once("memory add fail, invalid altmap\n"); 319 return -EINVAL; 320 } 321 altmap->alloc = 0; 322 } 323 324 err = check_pfn_span(pfn, nr_pages, "add"); 325 if (err) 326 return err; 327 328 start_sec = pfn_to_section_nr(pfn); 329 end_sec = pfn_to_section_nr(pfn + nr_pages - 1); 330 for (nr = start_sec; nr <= end_sec; nr++) { 331 unsigned long pfns; 332 333 pfns = min(nr_pages, PAGES_PER_SECTION 334 - (pfn & ~PAGE_SECTION_MASK)); 335 err = sparse_add_section(nid, pfn, pfns, altmap); 336 if (err) 337 break; 338 pfn += pfns; 339 nr_pages -= pfns; 340 cond_resched(); 341 } 342 vmemmap_populate_print_last(); 343 return err; 344 } 345 346 /* find the smallest valid pfn in the range [start_pfn, end_pfn) */ 347 static unsigned long find_smallest_section_pfn(int nid, struct zone *zone, 348 unsigned long start_pfn, 349 unsigned long end_pfn) 350 { 351 for (; start_pfn < end_pfn; start_pfn += PAGES_PER_SUBSECTION) { 352 if (unlikely(!pfn_to_online_page(start_pfn))) 353 continue; 354 355 if (unlikely(pfn_to_nid(start_pfn) != nid)) 356 continue; 357 358 if (zone != page_zone(pfn_to_page(start_pfn))) 359 continue; 360 361 return start_pfn; 362 } 363 364 return 0; 365 } 366 367 /* find the biggest valid pfn in the range [start_pfn, end_pfn). */ 368 static unsigned long find_biggest_section_pfn(int nid, struct zone *zone, 369 unsigned long start_pfn, 370 unsigned long end_pfn) 371 { 372 unsigned long pfn; 373 374 /* pfn is the end pfn of a memory section. */ 375 pfn = end_pfn - 1; 376 for (; pfn >= start_pfn; pfn -= PAGES_PER_SUBSECTION) { 377 if (unlikely(!pfn_to_online_page(pfn))) 378 continue; 379 380 if (unlikely(pfn_to_nid(pfn) != nid)) 381 continue; 382 383 if (zone != page_zone(pfn_to_page(pfn))) 384 continue; 385 386 return pfn; 387 } 388 389 return 0; 390 } 391 392 static void shrink_zone_span(struct zone *zone, unsigned long start_pfn, 393 unsigned long end_pfn) 394 { 395 unsigned long pfn; 396 int nid = zone_to_nid(zone); 397 398 zone_span_writelock(zone); 399 if (zone->zone_start_pfn == start_pfn) { 400 /* 401 * If the section is smallest section in the zone, it need 402 * shrink zone->zone_start_pfn and zone->zone_spanned_pages. 403 * In this case, we find second smallest valid mem_section 404 * for shrinking zone. 405 */ 406 pfn = find_smallest_section_pfn(nid, zone, end_pfn, 407 zone_end_pfn(zone)); 408 if (pfn) { 409 zone->spanned_pages = zone_end_pfn(zone) - pfn; 410 zone->zone_start_pfn = pfn; 411 } else { 412 zone->zone_start_pfn = 0; 413 zone->spanned_pages = 0; 414 } 415 } else if (zone_end_pfn(zone) == end_pfn) { 416 /* 417 * If the section is biggest section in the zone, it need 418 * shrink zone->spanned_pages. 419 * In this case, we find second biggest valid mem_section for 420 * shrinking zone. 421 */ 422 pfn = find_biggest_section_pfn(nid, zone, zone->zone_start_pfn, 423 start_pfn); 424 if (pfn) 425 zone->spanned_pages = pfn - zone->zone_start_pfn + 1; 426 else { 427 zone->zone_start_pfn = 0; 428 zone->spanned_pages = 0; 429 } 430 } 431 zone_span_writeunlock(zone); 432 } 433 434 static void update_pgdat_span(struct pglist_data *pgdat) 435 { 436 unsigned long node_start_pfn = 0, node_end_pfn = 0; 437 struct zone *zone; 438 439 for (zone = pgdat->node_zones; 440 zone < pgdat->node_zones + MAX_NR_ZONES; zone++) { 441 unsigned long zone_end_pfn = zone->zone_start_pfn + 442 zone->spanned_pages; 443 444 /* No need to lock the zones, they can't change. */ 445 if (!zone->spanned_pages) 446 continue; 447 if (!node_end_pfn) { 448 node_start_pfn = zone->zone_start_pfn; 449 node_end_pfn = zone_end_pfn; 450 continue; 451 } 452 453 if (zone_end_pfn > node_end_pfn) 454 node_end_pfn = zone_end_pfn; 455 if (zone->zone_start_pfn < node_start_pfn) 456 node_start_pfn = zone->zone_start_pfn; 457 } 458 459 pgdat->node_start_pfn = node_start_pfn; 460 pgdat->node_spanned_pages = node_end_pfn - node_start_pfn; 461 } 462 463 void __ref remove_pfn_range_from_zone(struct zone *zone, 464 unsigned long start_pfn, 465 unsigned long nr_pages) 466 { 467 struct pglist_data *pgdat = zone->zone_pgdat; 468 unsigned long flags; 469 470 /* Poison struct pages because they are now uninitialized again. */ 471 page_init_poison(pfn_to_page(start_pfn), sizeof(struct page) * nr_pages); 472 473 #ifdef CONFIG_ZONE_DEVICE 474 /* 475 * Zone shrinking code cannot properly deal with ZONE_DEVICE. So 476 * we will not try to shrink the zones - which is okay as 477 * set_zone_contiguous() cannot deal with ZONE_DEVICE either way. 478 */ 479 if (zone_idx(zone) == ZONE_DEVICE) 480 return; 481 #endif 482 483 clear_zone_contiguous(zone); 484 485 pgdat_resize_lock(zone->zone_pgdat, &flags); 486 shrink_zone_span(zone, start_pfn, start_pfn + nr_pages); 487 update_pgdat_span(pgdat); 488 pgdat_resize_unlock(zone->zone_pgdat, &flags); 489 490 set_zone_contiguous(zone); 491 } 492 493 static void __remove_section(unsigned long pfn, unsigned long nr_pages, 494 unsigned long map_offset, 495 struct vmem_altmap *altmap) 496 { 497 struct mem_section *ms = __nr_to_section(pfn_to_section_nr(pfn)); 498 499 if (WARN_ON_ONCE(!valid_section(ms))) 500 return; 501 502 sparse_remove_section(ms, pfn, nr_pages, map_offset, altmap); 503 } 504 505 /** 506 * __remove_pages() - remove sections of pages 507 * @pfn: starting pageframe (must be aligned to start of a section) 508 * @nr_pages: number of pages to remove (must be multiple of section size) 509 * @altmap: alternative device page map or %NULL if default memmap is used 510 * 511 * Generic helper function to remove section mappings and sysfs entries 512 * for the section of the memory we are removing. Caller needs to make 513 * sure that pages are marked reserved and zones are adjust properly by 514 * calling offline_pages(). 515 */ 516 void __remove_pages(unsigned long pfn, unsigned long nr_pages, 517 struct vmem_altmap *altmap) 518 { 519 const unsigned long end_pfn = pfn + nr_pages; 520 unsigned long cur_nr_pages; 521 unsigned long map_offset = 0; 522 523 map_offset = vmem_altmap_offset(altmap); 524 525 if (check_pfn_span(pfn, nr_pages, "remove")) 526 return; 527 528 for (; pfn < end_pfn; pfn += cur_nr_pages) { 529 cond_resched(); 530 /* Select all remaining pages up to the next section boundary */ 531 cur_nr_pages = min(end_pfn - pfn, -(pfn | PAGE_SECTION_MASK)); 532 __remove_section(pfn, cur_nr_pages, map_offset, altmap); 533 map_offset = 0; 534 } 535 } 536 537 int set_online_page_callback(online_page_callback_t callback) 538 { 539 int rc = -EINVAL; 540 541 get_online_mems(); 542 mutex_lock(&online_page_callback_lock); 543 544 if (online_page_callback == generic_online_page) { 545 online_page_callback = callback; 546 rc = 0; 547 } 548 549 mutex_unlock(&online_page_callback_lock); 550 put_online_mems(); 551 552 return rc; 553 } 554 EXPORT_SYMBOL_GPL(set_online_page_callback); 555 556 int restore_online_page_callback(online_page_callback_t callback) 557 { 558 int rc = -EINVAL; 559 560 get_online_mems(); 561 mutex_lock(&online_page_callback_lock); 562 563 if (online_page_callback == callback) { 564 online_page_callback = generic_online_page; 565 rc = 0; 566 } 567 568 mutex_unlock(&online_page_callback_lock); 569 put_online_mems(); 570 571 return rc; 572 } 573 EXPORT_SYMBOL_GPL(restore_online_page_callback); 574 575 void generic_online_page(struct page *page, unsigned int order) 576 { 577 kernel_map_pages(page, 1 << order, 1); 578 __free_pages_core(page, order); 579 totalram_pages_add(1UL << order); 580 #ifdef CONFIG_HIGHMEM 581 if (PageHighMem(page)) 582 totalhigh_pages_add(1UL << order); 583 #endif 584 } 585 EXPORT_SYMBOL_GPL(generic_online_page); 586 587 static int online_pages_range(unsigned long start_pfn, unsigned long nr_pages, 588 void *arg) 589 { 590 const unsigned long end_pfn = start_pfn + nr_pages; 591 unsigned long pfn; 592 int order; 593 594 /* 595 * Online the pages. The callback might decide to keep some pages 596 * PG_reserved (to add them to the buddy later), but we still account 597 * them as being online/belonging to this zone ("present"). 598 */ 599 for (pfn = start_pfn; pfn < end_pfn; pfn += 1ul << order) { 600 order = min(MAX_ORDER - 1, get_order(PFN_PHYS(end_pfn - pfn))); 601 /* __free_pages_core() wants pfns to be aligned to the order */ 602 if (WARN_ON_ONCE(!IS_ALIGNED(pfn, 1ul << order))) 603 order = 0; 604 (*online_page_callback)(pfn_to_page(pfn), order); 605 } 606 607 /* mark all involved sections as online */ 608 online_mem_sections(start_pfn, end_pfn); 609 610 *(unsigned long *)arg += nr_pages; 611 return 0; 612 } 613 614 /* check which state of node_states will be changed when online memory */ 615 static void node_states_check_changes_online(unsigned long nr_pages, 616 struct zone *zone, struct memory_notify *arg) 617 { 618 int nid = zone_to_nid(zone); 619 620 arg->status_change_nid = NUMA_NO_NODE; 621 arg->status_change_nid_normal = NUMA_NO_NODE; 622 arg->status_change_nid_high = NUMA_NO_NODE; 623 624 if (!node_state(nid, N_MEMORY)) 625 arg->status_change_nid = nid; 626 if (zone_idx(zone) <= ZONE_NORMAL && !node_state(nid, N_NORMAL_MEMORY)) 627 arg->status_change_nid_normal = nid; 628 #ifdef CONFIG_HIGHMEM 629 if (zone_idx(zone) <= ZONE_HIGHMEM && !node_state(nid, N_HIGH_MEMORY)) 630 arg->status_change_nid_high = nid; 631 #endif 632 } 633 634 static void node_states_set_node(int node, struct memory_notify *arg) 635 { 636 if (arg->status_change_nid_normal >= 0) 637 node_set_state(node, N_NORMAL_MEMORY); 638 639 if (arg->status_change_nid_high >= 0) 640 node_set_state(node, N_HIGH_MEMORY); 641 642 if (arg->status_change_nid >= 0) 643 node_set_state(node, N_MEMORY); 644 } 645 646 static void __meminit resize_zone_range(struct zone *zone, unsigned long start_pfn, 647 unsigned long nr_pages) 648 { 649 unsigned long old_end_pfn = zone_end_pfn(zone); 650 651 if (zone_is_empty(zone) || start_pfn < zone->zone_start_pfn) 652 zone->zone_start_pfn = start_pfn; 653 654 zone->spanned_pages = max(start_pfn + nr_pages, old_end_pfn) - zone->zone_start_pfn; 655 } 656 657 static void __meminit resize_pgdat_range(struct pglist_data *pgdat, unsigned long start_pfn, 658 unsigned long nr_pages) 659 { 660 unsigned long old_end_pfn = pgdat_end_pfn(pgdat); 661 662 if (!pgdat->node_spanned_pages || start_pfn < pgdat->node_start_pfn) 663 pgdat->node_start_pfn = start_pfn; 664 665 pgdat->node_spanned_pages = max(start_pfn + nr_pages, old_end_pfn) - pgdat->node_start_pfn; 666 667 } 668 /* 669 * Associate the pfn range with the given zone, initializing the memmaps 670 * and resizing the pgdat/zone data to span the added pages. After this 671 * call, all affected pages are PG_reserved. 672 */ 673 void __ref move_pfn_range_to_zone(struct zone *zone, unsigned long start_pfn, 674 unsigned long nr_pages, struct vmem_altmap *altmap) 675 { 676 struct pglist_data *pgdat = zone->zone_pgdat; 677 int nid = pgdat->node_id; 678 unsigned long flags; 679 680 clear_zone_contiguous(zone); 681 682 /* TODO Huh pgdat is irqsave while zone is not. It used to be like that before */ 683 pgdat_resize_lock(pgdat, &flags); 684 zone_span_writelock(zone); 685 if (zone_is_empty(zone)) 686 init_currently_empty_zone(zone, start_pfn, nr_pages); 687 resize_zone_range(zone, start_pfn, nr_pages); 688 zone_span_writeunlock(zone); 689 resize_pgdat_range(pgdat, start_pfn, nr_pages); 690 pgdat_resize_unlock(pgdat, &flags); 691 692 /* 693 * TODO now we have a visible range of pages which are not associated 694 * with their zone properly. Not nice but set_pfnblock_flags_mask 695 * expects the zone spans the pfn range. All the pages in the range 696 * are reserved so nobody should be touching them so we should be safe 697 */ 698 memmap_init_zone(nr_pages, nid, zone_idx(zone), start_pfn, 699 MEMMAP_HOTPLUG, altmap); 700 701 set_zone_contiguous(zone); 702 } 703 704 /* 705 * Returns a default kernel memory zone for the given pfn range. 706 * If no kernel zone covers this pfn range it will automatically go 707 * to the ZONE_NORMAL. 708 */ 709 static struct zone *default_kernel_zone_for_pfn(int nid, unsigned long start_pfn, 710 unsigned long nr_pages) 711 { 712 struct pglist_data *pgdat = NODE_DATA(nid); 713 int zid; 714 715 for (zid = 0; zid <= ZONE_NORMAL; zid++) { 716 struct zone *zone = &pgdat->node_zones[zid]; 717 718 if (zone_intersects(zone, start_pfn, nr_pages)) 719 return zone; 720 } 721 722 return &pgdat->node_zones[ZONE_NORMAL]; 723 } 724 725 static inline struct zone *default_zone_for_pfn(int nid, unsigned long start_pfn, 726 unsigned long nr_pages) 727 { 728 struct zone *kernel_zone = default_kernel_zone_for_pfn(nid, start_pfn, 729 nr_pages); 730 struct zone *movable_zone = &NODE_DATA(nid)->node_zones[ZONE_MOVABLE]; 731 bool in_kernel = zone_intersects(kernel_zone, start_pfn, nr_pages); 732 bool in_movable = zone_intersects(movable_zone, start_pfn, nr_pages); 733 734 /* 735 * We inherit the existing zone in a simple case where zones do not 736 * overlap in the given range 737 */ 738 if (in_kernel ^ in_movable) 739 return (in_kernel) ? kernel_zone : movable_zone; 740 741 /* 742 * If the range doesn't belong to any zone or two zones overlap in the 743 * given range then we use movable zone only if movable_node is 744 * enabled because we always online to a kernel zone by default. 745 */ 746 return movable_node_enabled ? movable_zone : kernel_zone; 747 } 748 749 struct zone * zone_for_pfn_range(int online_type, int nid, unsigned start_pfn, 750 unsigned long nr_pages) 751 { 752 if (online_type == MMOP_ONLINE_KERNEL) 753 return default_kernel_zone_for_pfn(nid, start_pfn, nr_pages); 754 755 if (online_type == MMOP_ONLINE_MOVABLE) 756 return &NODE_DATA(nid)->node_zones[ZONE_MOVABLE]; 757 758 return default_zone_for_pfn(nid, start_pfn, nr_pages); 759 } 760 761 int __ref online_pages(unsigned long pfn, unsigned long nr_pages, 762 int online_type, int nid) 763 { 764 unsigned long flags; 765 unsigned long onlined_pages = 0; 766 struct zone *zone; 767 int need_zonelists_rebuild = 0; 768 int ret; 769 struct memory_notify arg; 770 771 mem_hotplug_begin(); 772 773 /* associate pfn range with the zone */ 774 zone = zone_for_pfn_range(online_type, nid, pfn, nr_pages); 775 move_pfn_range_to_zone(zone, pfn, nr_pages, NULL); 776 777 arg.start_pfn = pfn; 778 arg.nr_pages = nr_pages; 779 node_states_check_changes_online(nr_pages, zone, &arg); 780 781 ret = memory_notify(MEM_GOING_ONLINE, &arg); 782 ret = notifier_to_errno(ret); 783 if (ret) 784 goto failed_addition; 785 786 /* 787 * If this zone is not populated, then it is not in zonelist. 788 * This means the page allocator ignores this zone. 789 * So, zonelist must be updated after online. 790 */ 791 if (!populated_zone(zone)) { 792 need_zonelists_rebuild = 1; 793 setup_zone_pageset(zone); 794 } 795 796 ret = walk_system_ram_range(pfn, nr_pages, &onlined_pages, 797 online_pages_range); 798 if (ret) { 799 /* not a single memory resource was applicable */ 800 if (need_zonelists_rebuild) 801 zone_pcp_reset(zone); 802 goto failed_addition; 803 } 804 805 zone->present_pages += onlined_pages; 806 807 pgdat_resize_lock(zone->zone_pgdat, &flags); 808 zone->zone_pgdat->node_present_pages += onlined_pages; 809 pgdat_resize_unlock(zone->zone_pgdat, &flags); 810 811 shuffle_zone(zone); 812 813 node_states_set_node(nid, &arg); 814 if (need_zonelists_rebuild) 815 build_all_zonelists(NULL); 816 else 817 zone_pcp_update(zone); 818 819 init_per_zone_wmark_min(); 820 821 kswapd_run(nid); 822 kcompactd_run(nid); 823 824 vm_total_pages = nr_free_pagecache_pages(); 825 826 writeback_set_ratelimit(); 827 828 memory_notify(MEM_ONLINE, &arg); 829 mem_hotplug_done(); 830 return 0; 831 832 failed_addition: 833 pr_debug("online_pages [mem %#010llx-%#010llx] failed\n", 834 (unsigned long long) pfn << PAGE_SHIFT, 835 (((unsigned long long) pfn + nr_pages) << PAGE_SHIFT) - 1); 836 memory_notify(MEM_CANCEL_ONLINE, &arg); 837 remove_pfn_range_from_zone(zone, pfn, nr_pages); 838 mem_hotplug_done(); 839 return ret; 840 } 841 #endif /* CONFIG_MEMORY_HOTPLUG_SPARSE */ 842 843 static void reset_node_present_pages(pg_data_t *pgdat) 844 { 845 struct zone *z; 846 847 for (z = pgdat->node_zones; z < pgdat->node_zones + MAX_NR_ZONES; z++) 848 z->present_pages = 0; 849 850 pgdat->node_present_pages = 0; 851 } 852 853 /* we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG */ 854 static pg_data_t __ref *hotadd_new_pgdat(int nid, u64 start) 855 { 856 struct pglist_data *pgdat; 857 unsigned long start_pfn = PFN_DOWN(start); 858 859 pgdat = NODE_DATA(nid); 860 if (!pgdat) { 861 pgdat = arch_alloc_nodedata(nid); 862 if (!pgdat) 863 return NULL; 864 865 pgdat->per_cpu_nodestats = 866 alloc_percpu(struct per_cpu_nodestat); 867 arch_refresh_nodedata(nid, pgdat); 868 } else { 869 int cpu; 870 /* 871 * Reset the nr_zones, order and classzone_idx before reuse. 872 * Note that kswapd will init kswapd_classzone_idx properly 873 * when it starts in the near future. 874 */ 875 pgdat->nr_zones = 0; 876 pgdat->kswapd_order = 0; 877 pgdat->kswapd_classzone_idx = 0; 878 for_each_online_cpu(cpu) { 879 struct per_cpu_nodestat *p; 880 881 p = per_cpu_ptr(pgdat->per_cpu_nodestats, cpu); 882 memset(p, 0, sizeof(*p)); 883 } 884 } 885 886 /* we can use NODE_DATA(nid) from here */ 887 888 pgdat->node_id = nid; 889 pgdat->node_start_pfn = start_pfn; 890 891 /* init node's zones as empty zones, we don't have any present pages.*/ 892 free_area_init_core_hotplug(nid); 893 894 /* 895 * The node we allocated has no zone fallback lists. For avoiding 896 * to access not-initialized zonelist, build here. 897 */ 898 build_all_zonelists(pgdat); 899 900 /* 901 * When memory is hot-added, all the memory is in offline state. So 902 * clear all zones' present_pages because they will be updated in 903 * online_pages() and offline_pages(). 904 */ 905 reset_node_managed_pages(pgdat); 906 reset_node_present_pages(pgdat); 907 908 return pgdat; 909 } 910 911 static void rollback_node_hotadd(int nid) 912 { 913 pg_data_t *pgdat = NODE_DATA(nid); 914 915 arch_refresh_nodedata(nid, NULL); 916 free_percpu(pgdat->per_cpu_nodestats); 917 arch_free_nodedata(pgdat); 918 } 919 920 921 /** 922 * try_online_node - online a node if offlined 923 * @nid: the node ID 924 * @start: start addr of the node 925 * @set_node_online: Whether we want to online the node 926 * called by cpu_up() to online a node without onlined memory. 927 * 928 * Returns: 929 * 1 -> a new node has been allocated 930 * 0 -> the node is already online 931 * -ENOMEM -> the node could not be allocated 932 */ 933 static int __try_online_node(int nid, u64 start, bool set_node_online) 934 { 935 pg_data_t *pgdat; 936 int ret = 1; 937 938 if (node_online(nid)) 939 return 0; 940 941 pgdat = hotadd_new_pgdat(nid, start); 942 if (!pgdat) { 943 pr_err("Cannot online node %d due to NULL pgdat\n", nid); 944 ret = -ENOMEM; 945 goto out; 946 } 947 948 if (set_node_online) { 949 node_set_online(nid); 950 ret = register_one_node(nid); 951 BUG_ON(ret); 952 } 953 out: 954 return ret; 955 } 956 957 /* 958 * Users of this function always want to online/register the node 959 */ 960 int try_online_node(int nid) 961 { 962 int ret; 963 964 mem_hotplug_begin(); 965 ret = __try_online_node(nid, 0, true); 966 mem_hotplug_done(); 967 return ret; 968 } 969 970 static int check_hotplug_memory_range(u64 start, u64 size) 971 { 972 /* memory range must be block size aligned */ 973 if (!size || !IS_ALIGNED(start, memory_block_size_bytes()) || 974 !IS_ALIGNED(size, memory_block_size_bytes())) { 975 pr_err("Block size [%#lx] unaligned hotplug range: start %#llx, size %#llx", 976 memory_block_size_bytes(), start, size); 977 return -EINVAL; 978 } 979 980 return 0; 981 } 982 983 static int online_memory_block(struct memory_block *mem, void *arg) 984 { 985 return device_online(&mem->dev); 986 } 987 988 /* 989 * NOTE: The caller must call lock_device_hotplug() to serialize hotplug 990 * and online/offline operations (triggered e.g. by sysfs). 991 * 992 * we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG 993 */ 994 int __ref add_memory_resource(int nid, struct resource *res) 995 { 996 struct mhp_restrictions restrictions = {}; 997 u64 start, size; 998 bool new_node = false; 999 int ret; 1000 1001 start = res->start; 1002 size = resource_size(res); 1003 1004 ret = check_hotplug_memory_range(start, size); 1005 if (ret) 1006 return ret; 1007 1008 mem_hotplug_begin(); 1009 1010 /* 1011 * Add new range to memblock so that when hotadd_new_pgdat() is called 1012 * to allocate new pgdat, get_pfn_range_for_nid() will be able to find 1013 * this new range and calculate total pages correctly. The range will 1014 * be removed at hot-remove time. 1015 */ 1016 memblock_add_node(start, size, nid); 1017 1018 ret = __try_online_node(nid, start, false); 1019 if (ret < 0) 1020 goto error; 1021 new_node = ret; 1022 1023 /* call arch's memory hotadd */ 1024 ret = arch_add_memory(nid, start, size, &restrictions); 1025 if (ret < 0) 1026 goto error; 1027 1028 /* create memory block devices after memory was added */ 1029 ret = create_memory_block_devices(start, size); 1030 if (ret) { 1031 arch_remove_memory(nid, start, size, NULL); 1032 goto error; 1033 } 1034 1035 if (new_node) { 1036 /* If sysfs file of new node can't be created, cpu on the node 1037 * can't be hot-added. There is no rollback way now. 1038 * So, check by BUG_ON() to catch it reluctantly.. 1039 * We online node here. We can't roll back from here. 1040 */ 1041 node_set_online(nid); 1042 ret = __register_one_node(nid); 1043 BUG_ON(ret); 1044 } 1045 1046 /* link memory sections under this node.*/ 1047 ret = link_mem_sections(nid, PFN_DOWN(start), PFN_UP(start + size - 1)); 1048 BUG_ON(ret); 1049 1050 /* create new memmap entry */ 1051 firmware_map_add_hotplug(start, start + size, "System RAM"); 1052 1053 /* device_online() will take the lock when calling online_pages() */ 1054 mem_hotplug_done(); 1055 1056 /* online pages if requested */ 1057 if (memhp_auto_online) 1058 walk_memory_blocks(start, size, NULL, online_memory_block); 1059 1060 return ret; 1061 error: 1062 /* rollback pgdat allocation and others */ 1063 if (new_node) 1064 rollback_node_hotadd(nid); 1065 memblock_remove(start, size); 1066 mem_hotplug_done(); 1067 return ret; 1068 } 1069 1070 /* requires device_hotplug_lock, see add_memory_resource() */ 1071 int __ref __add_memory(int nid, u64 start, u64 size) 1072 { 1073 struct resource *res; 1074 int ret; 1075 1076 res = register_memory_resource(start, size); 1077 if (IS_ERR(res)) 1078 return PTR_ERR(res); 1079 1080 ret = add_memory_resource(nid, res); 1081 if (ret < 0) 1082 release_memory_resource(res); 1083 return ret; 1084 } 1085 1086 int add_memory(int nid, u64 start, u64 size) 1087 { 1088 int rc; 1089 1090 lock_device_hotplug(); 1091 rc = __add_memory(nid, start, size); 1092 unlock_device_hotplug(); 1093 1094 return rc; 1095 } 1096 EXPORT_SYMBOL_GPL(add_memory); 1097 1098 #ifdef CONFIG_MEMORY_HOTREMOVE 1099 /* 1100 * A free page on the buddy free lists (not the per-cpu lists) has PageBuddy 1101 * set and the size of the free page is given by page_order(). Using this, 1102 * the function determines if the pageblock contains only free pages. 1103 * Due to buddy contraints, a free page at least the size of a pageblock will 1104 * be located at the start of the pageblock 1105 */ 1106 static inline int pageblock_free(struct page *page) 1107 { 1108 return PageBuddy(page) && page_order(page) >= pageblock_order; 1109 } 1110 1111 /* Return the pfn of the start of the next active pageblock after a given pfn */ 1112 static unsigned long next_active_pageblock(unsigned long pfn) 1113 { 1114 struct page *page = pfn_to_page(pfn); 1115 1116 /* Ensure the starting page is pageblock-aligned */ 1117 BUG_ON(pfn & (pageblock_nr_pages - 1)); 1118 1119 /* If the entire pageblock is free, move to the end of free page */ 1120 if (pageblock_free(page)) { 1121 int order; 1122 /* be careful. we don't have locks, page_order can be changed.*/ 1123 order = page_order(page); 1124 if ((order < MAX_ORDER) && (order >= pageblock_order)) 1125 return pfn + (1 << order); 1126 } 1127 1128 return pfn + pageblock_nr_pages; 1129 } 1130 1131 static bool is_pageblock_removable_nolock(unsigned long pfn) 1132 { 1133 struct page *page = pfn_to_page(pfn); 1134 struct zone *zone; 1135 1136 /* 1137 * We have to be careful here because we are iterating over memory 1138 * sections which are not zone aware so we might end up outside of 1139 * the zone but still within the section. 1140 * We have to take care about the node as well. If the node is offline 1141 * its NODE_DATA will be NULL - see page_zone. 1142 */ 1143 if (!node_online(page_to_nid(page))) 1144 return false; 1145 1146 zone = page_zone(page); 1147 pfn = page_to_pfn(page); 1148 if (!zone_spans_pfn(zone, pfn)) 1149 return false; 1150 1151 return !has_unmovable_pages(zone, page, MIGRATE_MOVABLE, 1152 MEMORY_OFFLINE); 1153 } 1154 1155 /* Checks if this range of memory is likely to be hot-removable. */ 1156 bool is_mem_section_removable(unsigned long start_pfn, unsigned long nr_pages) 1157 { 1158 unsigned long end_pfn, pfn; 1159 1160 end_pfn = min(start_pfn + nr_pages, 1161 zone_end_pfn(page_zone(pfn_to_page(start_pfn)))); 1162 1163 /* Check the starting page of each pageblock within the range */ 1164 for (pfn = start_pfn; pfn < end_pfn; pfn = next_active_pageblock(pfn)) { 1165 if (!is_pageblock_removable_nolock(pfn)) 1166 return false; 1167 cond_resched(); 1168 } 1169 1170 /* All pageblocks in the memory block are likely to be hot-removable */ 1171 return true; 1172 } 1173 1174 /* 1175 * Confirm all pages in a range [start, end) belong to the same zone (skipping 1176 * memory holes). When true, return the zone. 1177 */ 1178 struct zone *test_pages_in_a_zone(unsigned long start_pfn, 1179 unsigned long end_pfn) 1180 { 1181 unsigned long pfn, sec_end_pfn; 1182 struct zone *zone = NULL; 1183 struct page *page; 1184 int i; 1185 for (pfn = start_pfn, sec_end_pfn = SECTION_ALIGN_UP(start_pfn + 1); 1186 pfn < end_pfn; 1187 pfn = sec_end_pfn, sec_end_pfn += PAGES_PER_SECTION) { 1188 /* Make sure the memory section is present first */ 1189 if (!present_section_nr(pfn_to_section_nr(pfn))) 1190 continue; 1191 for (; pfn < sec_end_pfn && pfn < end_pfn; 1192 pfn += MAX_ORDER_NR_PAGES) { 1193 i = 0; 1194 /* This is just a CONFIG_HOLES_IN_ZONE check.*/ 1195 while ((i < MAX_ORDER_NR_PAGES) && 1196 !pfn_valid_within(pfn + i)) 1197 i++; 1198 if (i == MAX_ORDER_NR_PAGES || pfn + i >= end_pfn) 1199 continue; 1200 /* Check if we got outside of the zone */ 1201 if (zone && !zone_spans_pfn(zone, pfn + i)) 1202 return NULL; 1203 page = pfn_to_page(pfn + i); 1204 if (zone && page_zone(page) != zone) 1205 return NULL; 1206 zone = page_zone(page); 1207 } 1208 } 1209 1210 return zone; 1211 } 1212 1213 /* 1214 * Scan pfn range [start,end) to find movable/migratable pages (LRU pages, 1215 * non-lru movable pages and hugepages). We scan pfn because it's much 1216 * easier than scanning over linked list. This function returns the pfn 1217 * of the first found movable page if it's found, otherwise 0. 1218 */ 1219 static unsigned long scan_movable_pages(unsigned long start, unsigned long end) 1220 { 1221 unsigned long pfn; 1222 1223 for (pfn = start; pfn < end; pfn++) { 1224 struct page *page, *head; 1225 unsigned long skip; 1226 1227 if (!pfn_valid(pfn)) 1228 continue; 1229 page = pfn_to_page(pfn); 1230 if (PageLRU(page)) 1231 return pfn; 1232 if (__PageMovable(page)) 1233 return pfn; 1234 1235 if (!PageHuge(page)) 1236 continue; 1237 head = compound_head(page); 1238 if (page_huge_active(head)) 1239 return pfn; 1240 skip = compound_nr(head) - (page - head); 1241 pfn += skip - 1; 1242 } 1243 return 0; 1244 } 1245 1246 static struct page *new_node_page(struct page *page, unsigned long private) 1247 { 1248 int nid = page_to_nid(page); 1249 nodemask_t nmask = node_states[N_MEMORY]; 1250 1251 /* 1252 * try to allocate from a different node but reuse this node if there 1253 * are no other online nodes to be used (e.g. we are offlining a part 1254 * of the only existing node) 1255 */ 1256 node_clear(nid, nmask); 1257 if (nodes_empty(nmask)) 1258 node_set(nid, nmask); 1259 1260 return new_page_nodemask(page, nid, &nmask); 1261 } 1262 1263 static int 1264 do_migrate_range(unsigned long start_pfn, unsigned long end_pfn) 1265 { 1266 unsigned long pfn; 1267 struct page *page; 1268 int ret = 0; 1269 LIST_HEAD(source); 1270 1271 for (pfn = start_pfn; pfn < end_pfn; pfn++) { 1272 if (!pfn_valid(pfn)) 1273 continue; 1274 page = pfn_to_page(pfn); 1275 1276 if (PageHuge(page)) { 1277 struct page *head = compound_head(page); 1278 pfn = page_to_pfn(head) + compound_nr(head) - 1; 1279 isolate_huge_page(head, &source); 1280 continue; 1281 } else if (PageTransHuge(page)) 1282 pfn = page_to_pfn(compound_head(page)) 1283 + hpage_nr_pages(page) - 1; 1284 1285 /* 1286 * HWPoison pages have elevated reference counts so the migration would 1287 * fail on them. It also doesn't make any sense to migrate them in the 1288 * first place. Still try to unmap such a page in case it is still mapped 1289 * (e.g. current hwpoison implementation doesn't unmap KSM pages but keep 1290 * the unmap as the catch all safety net). 1291 */ 1292 if (PageHWPoison(page)) { 1293 if (WARN_ON(PageLRU(page))) 1294 isolate_lru_page(page); 1295 if (page_mapped(page)) 1296 try_to_unmap(page, TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS); 1297 continue; 1298 } 1299 1300 if (!get_page_unless_zero(page)) 1301 continue; 1302 /* 1303 * We can skip free pages. And we can deal with pages on 1304 * LRU and non-lru movable pages. 1305 */ 1306 if (PageLRU(page)) 1307 ret = isolate_lru_page(page); 1308 else 1309 ret = isolate_movable_page(page, ISOLATE_UNEVICTABLE); 1310 if (!ret) { /* Success */ 1311 list_add_tail(&page->lru, &source); 1312 if (!__PageMovable(page)) 1313 inc_node_page_state(page, NR_ISOLATED_ANON + 1314 page_is_file_cache(page)); 1315 1316 } else { 1317 pr_warn("failed to isolate pfn %lx\n", pfn); 1318 dump_page(page, "isolation failed"); 1319 } 1320 put_page(page); 1321 } 1322 if (!list_empty(&source)) { 1323 /* Allocate a new page from the nearest neighbor node */ 1324 ret = migrate_pages(&source, new_node_page, NULL, 0, 1325 MIGRATE_SYNC, MR_MEMORY_HOTPLUG); 1326 if (ret) { 1327 list_for_each_entry(page, &source, lru) { 1328 pr_warn("migrating pfn %lx failed ret:%d ", 1329 page_to_pfn(page), ret); 1330 dump_page(page, "migration failure"); 1331 } 1332 putback_movable_pages(&source); 1333 } 1334 } 1335 1336 return ret; 1337 } 1338 1339 /* Mark all sections offline and remove all free pages from the buddy. */ 1340 static int 1341 offline_isolated_pages_cb(unsigned long start, unsigned long nr_pages, 1342 void *data) 1343 { 1344 unsigned long *offlined_pages = (unsigned long *)data; 1345 1346 *offlined_pages += __offline_isolated_pages(start, start + nr_pages); 1347 return 0; 1348 } 1349 1350 /* 1351 * Check all pages in range, recoreded as memory resource, are isolated. 1352 */ 1353 static int 1354 check_pages_isolated_cb(unsigned long start_pfn, unsigned long nr_pages, 1355 void *data) 1356 { 1357 return test_pages_isolated(start_pfn, start_pfn + nr_pages, 1358 MEMORY_OFFLINE); 1359 } 1360 1361 static int __init cmdline_parse_movable_node(char *p) 1362 { 1363 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP 1364 movable_node_enabled = true; 1365 #else 1366 pr_warn("movable_node parameter depends on CONFIG_HAVE_MEMBLOCK_NODE_MAP to work properly\n"); 1367 #endif 1368 return 0; 1369 } 1370 early_param("movable_node", cmdline_parse_movable_node); 1371 1372 /* check which state of node_states will be changed when offline memory */ 1373 static void node_states_check_changes_offline(unsigned long nr_pages, 1374 struct zone *zone, struct memory_notify *arg) 1375 { 1376 struct pglist_data *pgdat = zone->zone_pgdat; 1377 unsigned long present_pages = 0; 1378 enum zone_type zt; 1379 1380 arg->status_change_nid = NUMA_NO_NODE; 1381 arg->status_change_nid_normal = NUMA_NO_NODE; 1382 arg->status_change_nid_high = NUMA_NO_NODE; 1383 1384 /* 1385 * Check whether node_states[N_NORMAL_MEMORY] will be changed. 1386 * If the memory to be offline is within the range 1387 * [0..ZONE_NORMAL], and it is the last present memory there, 1388 * the zones in that range will become empty after the offlining, 1389 * thus we can determine that we need to clear the node from 1390 * node_states[N_NORMAL_MEMORY]. 1391 */ 1392 for (zt = 0; zt <= ZONE_NORMAL; zt++) 1393 present_pages += pgdat->node_zones[zt].present_pages; 1394 if (zone_idx(zone) <= ZONE_NORMAL && nr_pages >= present_pages) 1395 arg->status_change_nid_normal = zone_to_nid(zone); 1396 1397 #ifdef CONFIG_HIGHMEM 1398 /* 1399 * node_states[N_HIGH_MEMORY] contains nodes which 1400 * have normal memory or high memory. 1401 * Here we add the present_pages belonging to ZONE_HIGHMEM. 1402 * If the zone is within the range of [0..ZONE_HIGHMEM), and 1403 * we determine that the zones in that range become empty, 1404 * we need to clear the node for N_HIGH_MEMORY. 1405 */ 1406 present_pages += pgdat->node_zones[ZONE_HIGHMEM].present_pages; 1407 if (zone_idx(zone) <= ZONE_HIGHMEM && nr_pages >= present_pages) 1408 arg->status_change_nid_high = zone_to_nid(zone); 1409 #endif 1410 1411 /* 1412 * We have accounted the pages from [0..ZONE_NORMAL), and 1413 * in case of CONFIG_HIGHMEM the pages from ZONE_HIGHMEM 1414 * as well. 1415 * Here we count the possible pages from ZONE_MOVABLE. 1416 * If after having accounted all the pages, we see that the nr_pages 1417 * to be offlined is over or equal to the accounted pages, 1418 * we know that the node will become empty, and so, we can clear 1419 * it for N_MEMORY as well. 1420 */ 1421 present_pages += pgdat->node_zones[ZONE_MOVABLE].present_pages; 1422 1423 if (nr_pages >= present_pages) 1424 arg->status_change_nid = zone_to_nid(zone); 1425 } 1426 1427 static void node_states_clear_node(int node, struct memory_notify *arg) 1428 { 1429 if (arg->status_change_nid_normal >= 0) 1430 node_clear_state(node, N_NORMAL_MEMORY); 1431 1432 if (arg->status_change_nid_high >= 0) 1433 node_clear_state(node, N_HIGH_MEMORY); 1434 1435 if (arg->status_change_nid >= 0) 1436 node_clear_state(node, N_MEMORY); 1437 } 1438 1439 static int count_system_ram_pages_cb(unsigned long start_pfn, 1440 unsigned long nr_pages, void *data) 1441 { 1442 unsigned long *nr_system_ram_pages = data; 1443 1444 *nr_system_ram_pages += nr_pages; 1445 return 0; 1446 } 1447 1448 static int __ref __offline_pages(unsigned long start_pfn, 1449 unsigned long end_pfn) 1450 { 1451 unsigned long pfn, nr_pages = 0; 1452 unsigned long offlined_pages = 0; 1453 int ret, node, nr_isolate_pageblock; 1454 unsigned long flags; 1455 struct zone *zone; 1456 struct memory_notify arg; 1457 char *reason; 1458 1459 mem_hotplug_begin(); 1460 1461 /* 1462 * Don't allow to offline memory blocks that contain holes. 1463 * Consequently, memory blocks with holes can never get onlined 1464 * via the hotplug path - online_pages() - as hotplugged memory has 1465 * no holes. This way, we e.g., don't have to worry about marking 1466 * memory holes PG_reserved, don't need pfn_valid() checks, and can 1467 * avoid using walk_system_ram_range() later. 1468 */ 1469 walk_system_ram_range(start_pfn, end_pfn - start_pfn, &nr_pages, 1470 count_system_ram_pages_cb); 1471 if (nr_pages != end_pfn - start_pfn) { 1472 ret = -EINVAL; 1473 reason = "memory holes"; 1474 goto failed_removal; 1475 } 1476 1477 /* This makes hotplug much easier...and readable. 1478 we assume this for now. .*/ 1479 zone = test_pages_in_a_zone(start_pfn, end_pfn); 1480 if (!zone) { 1481 ret = -EINVAL; 1482 reason = "multizone range"; 1483 goto failed_removal; 1484 } 1485 node = zone_to_nid(zone); 1486 1487 /* set above range as isolated */ 1488 ret = start_isolate_page_range(start_pfn, end_pfn, 1489 MIGRATE_MOVABLE, 1490 MEMORY_OFFLINE | REPORT_FAILURE); 1491 if (ret < 0) { 1492 reason = "failure to isolate range"; 1493 goto failed_removal; 1494 } 1495 nr_isolate_pageblock = ret; 1496 1497 arg.start_pfn = start_pfn; 1498 arg.nr_pages = nr_pages; 1499 node_states_check_changes_offline(nr_pages, zone, &arg); 1500 1501 ret = memory_notify(MEM_GOING_OFFLINE, &arg); 1502 ret = notifier_to_errno(ret); 1503 if (ret) { 1504 reason = "notifier failure"; 1505 goto failed_removal_isolated; 1506 } 1507 1508 do { 1509 for (pfn = start_pfn; pfn;) { 1510 if (signal_pending(current)) { 1511 ret = -EINTR; 1512 reason = "signal backoff"; 1513 goto failed_removal_isolated; 1514 } 1515 1516 cond_resched(); 1517 lru_add_drain_all(); 1518 1519 pfn = scan_movable_pages(pfn, end_pfn); 1520 if (pfn) { 1521 /* 1522 * TODO: fatal migration failures should bail 1523 * out 1524 */ 1525 do_migrate_range(pfn, end_pfn); 1526 } 1527 } 1528 1529 /* 1530 * Dissolve free hugepages in the memory block before doing 1531 * offlining actually in order to make hugetlbfs's object 1532 * counting consistent. 1533 */ 1534 ret = dissolve_free_huge_pages(start_pfn, end_pfn); 1535 if (ret) { 1536 reason = "failure to dissolve huge pages"; 1537 goto failed_removal_isolated; 1538 } 1539 /* check again */ 1540 ret = walk_system_ram_range(start_pfn, end_pfn - start_pfn, 1541 NULL, check_pages_isolated_cb); 1542 } while (ret); 1543 1544 /* Ok, all of our target is isolated. 1545 We cannot do rollback at this point. */ 1546 walk_system_ram_range(start_pfn, end_pfn - start_pfn, 1547 &offlined_pages, offline_isolated_pages_cb); 1548 pr_info("Offlined Pages %ld\n", offlined_pages); 1549 /* 1550 * Onlining will reset pagetype flags and makes migrate type 1551 * MOVABLE, so just need to decrease the number of isolated 1552 * pageblocks zone counter here. 1553 */ 1554 spin_lock_irqsave(&zone->lock, flags); 1555 zone->nr_isolate_pageblock -= nr_isolate_pageblock; 1556 spin_unlock_irqrestore(&zone->lock, flags); 1557 1558 /* removal success */ 1559 adjust_managed_page_count(pfn_to_page(start_pfn), -offlined_pages); 1560 zone->present_pages -= offlined_pages; 1561 1562 pgdat_resize_lock(zone->zone_pgdat, &flags); 1563 zone->zone_pgdat->node_present_pages -= offlined_pages; 1564 pgdat_resize_unlock(zone->zone_pgdat, &flags); 1565 1566 init_per_zone_wmark_min(); 1567 1568 if (!populated_zone(zone)) { 1569 zone_pcp_reset(zone); 1570 build_all_zonelists(NULL); 1571 } else 1572 zone_pcp_update(zone); 1573 1574 node_states_clear_node(node, &arg); 1575 if (arg.status_change_nid >= 0) { 1576 kswapd_stop(node); 1577 kcompactd_stop(node); 1578 } 1579 1580 vm_total_pages = nr_free_pagecache_pages(); 1581 writeback_set_ratelimit(); 1582 1583 memory_notify(MEM_OFFLINE, &arg); 1584 remove_pfn_range_from_zone(zone, start_pfn, nr_pages); 1585 mem_hotplug_done(); 1586 return 0; 1587 1588 failed_removal_isolated: 1589 undo_isolate_page_range(start_pfn, end_pfn, MIGRATE_MOVABLE); 1590 memory_notify(MEM_CANCEL_OFFLINE, &arg); 1591 failed_removal: 1592 pr_debug("memory offlining [mem %#010llx-%#010llx] failed due to %s\n", 1593 (unsigned long long) start_pfn << PAGE_SHIFT, 1594 ((unsigned long long) end_pfn << PAGE_SHIFT) - 1, 1595 reason); 1596 /* pushback to free area */ 1597 mem_hotplug_done(); 1598 return ret; 1599 } 1600 1601 int offline_pages(unsigned long start_pfn, unsigned long nr_pages) 1602 { 1603 return __offline_pages(start_pfn, start_pfn + nr_pages); 1604 } 1605 1606 static int check_memblock_offlined_cb(struct memory_block *mem, void *arg) 1607 { 1608 int ret = !is_memblock_offlined(mem); 1609 1610 if (unlikely(ret)) { 1611 phys_addr_t beginpa, endpa; 1612 1613 beginpa = PFN_PHYS(section_nr_to_pfn(mem->start_section_nr)); 1614 endpa = beginpa + memory_block_size_bytes() - 1; 1615 pr_warn("removing memory fails, because memory [%pa-%pa] is onlined\n", 1616 &beginpa, &endpa); 1617 1618 return -EBUSY; 1619 } 1620 return 0; 1621 } 1622 1623 static int check_cpu_on_node(pg_data_t *pgdat) 1624 { 1625 int cpu; 1626 1627 for_each_present_cpu(cpu) { 1628 if (cpu_to_node(cpu) == pgdat->node_id) 1629 /* 1630 * the cpu on this node isn't removed, and we can't 1631 * offline this node. 1632 */ 1633 return -EBUSY; 1634 } 1635 1636 return 0; 1637 } 1638 1639 static int check_no_memblock_for_node_cb(struct memory_block *mem, void *arg) 1640 { 1641 int nid = *(int *)arg; 1642 1643 /* 1644 * If a memory block belongs to multiple nodes, the stored nid is not 1645 * reliable. However, such blocks are always online (e.g., cannot get 1646 * offlined) and, therefore, are still spanned by the node. 1647 */ 1648 return mem->nid == nid ? -EEXIST : 0; 1649 } 1650 1651 /** 1652 * try_offline_node 1653 * @nid: the node ID 1654 * 1655 * Offline a node if all memory sections and cpus of the node are removed. 1656 * 1657 * NOTE: The caller must call lock_device_hotplug() to serialize hotplug 1658 * and online/offline operations before this call. 1659 */ 1660 void try_offline_node(int nid) 1661 { 1662 pg_data_t *pgdat = NODE_DATA(nid); 1663 int rc; 1664 1665 /* 1666 * If the node still spans pages (especially ZONE_DEVICE), don't 1667 * offline it. A node spans memory after move_pfn_range_to_zone(), 1668 * e.g., after the memory block was onlined. 1669 */ 1670 if (pgdat->node_spanned_pages) 1671 return; 1672 1673 /* 1674 * Especially offline memory blocks might not be spanned by the 1675 * node. They will get spanned by the node once they get onlined. 1676 * However, they link to the node in sysfs and can get onlined later. 1677 */ 1678 rc = for_each_memory_block(&nid, check_no_memblock_for_node_cb); 1679 if (rc) 1680 return; 1681 1682 if (check_cpu_on_node(pgdat)) 1683 return; 1684 1685 /* 1686 * all memory/cpu of this node are removed, we can offline this 1687 * node now. 1688 */ 1689 node_set_offline(nid); 1690 unregister_one_node(nid); 1691 } 1692 EXPORT_SYMBOL(try_offline_node); 1693 1694 static void __release_memory_resource(resource_size_t start, 1695 resource_size_t size) 1696 { 1697 int ret; 1698 1699 /* 1700 * When removing memory in the same granularity as it was added, 1701 * this function never fails. It might only fail if resources 1702 * have to be adjusted or split. We'll ignore the error, as 1703 * removing of memory cannot fail. 1704 */ 1705 ret = release_mem_region_adjustable(&iomem_resource, start, size); 1706 if (ret) { 1707 resource_size_t endres = start + size - 1; 1708 1709 pr_warn("Unable to release resource <%pa-%pa> (%d)\n", 1710 &start, &endres, ret); 1711 } 1712 } 1713 1714 static int __ref try_remove_memory(int nid, u64 start, u64 size) 1715 { 1716 int rc = 0; 1717 1718 BUG_ON(check_hotplug_memory_range(start, size)); 1719 1720 /* 1721 * All memory blocks must be offlined before removing memory. Check 1722 * whether all memory blocks in question are offline and return error 1723 * if this is not the case. 1724 */ 1725 rc = walk_memory_blocks(start, size, NULL, check_memblock_offlined_cb); 1726 if (rc) 1727 goto done; 1728 1729 /* remove memmap entry */ 1730 firmware_map_remove(start, start + size, "System RAM"); 1731 1732 /* 1733 * Memory block device removal under the device_hotplug_lock is 1734 * a barrier against racing online attempts. 1735 */ 1736 remove_memory_block_devices(start, size); 1737 1738 mem_hotplug_begin(); 1739 1740 arch_remove_memory(nid, start, size, NULL); 1741 memblock_free(start, size); 1742 memblock_remove(start, size); 1743 __release_memory_resource(start, size); 1744 1745 try_offline_node(nid); 1746 1747 done: 1748 mem_hotplug_done(); 1749 return rc; 1750 } 1751 1752 /** 1753 * remove_memory 1754 * @nid: the node ID 1755 * @start: physical address of the region to remove 1756 * @size: size of the region to remove 1757 * 1758 * NOTE: The caller must call lock_device_hotplug() to serialize hotplug 1759 * and online/offline operations before this call, as required by 1760 * try_offline_node(). 1761 */ 1762 void __remove_memory(int nid, u64 start, u64 size) 1763 { 1764 1765 /* 1766 * trigger BUG() if some memory is not offlined prior to calling this 1767 * function 1768 */ 1769 if (try_remove_memory(nid, start, size)) 1770 BUG(); 1771 } 1772 1773 /* 1774 * Remove memory if every memory block is offline, otherwise return -EBUSY is 1775 * some memory is not offline 1776 */ 1777 int remove_memory(int nid, u64 start, u64 size) 1778 { 1779 int rc; 1780 1781 lock_device_hotplug(); 1782 rc = try_remove_memory(nid, start, size); 1783 unlock_device_hotplug(); 1784 1785 return rc; 1786 } 1787 EXPORT_SYMBOL_GPL(remove_memory); 1788 #endif /* CONFIG_MEMORY_HOTREMOVE */ 1789