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 && 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 && 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 zone_start_pfn = zone->zone_start_pfn; 396 unsigned long z = zone_end_pfn(zone); /* zone_end_pfn namespace clash */ 397 unsigned long zone_end_pfn = z; 398 unsigned long pfn; 399 int nid = zone_to_nid(zone); 400 401 zone_span_writelock(zone); 402 if (zone_start_pfn == start_pfn) { 403 /* 404 * If the section is smallest section in the zone, it need 405 * shrink zone->zone_start_pfn and zone->zone_spanned_pages. 406 * In this case, we find second smallest valid mem_section 407 * for shrinking zone. 408 */ 409 pfn = find_smallest_section_pfn(nid, zone, end_pfn, 410 zone_end_pfn); 411 if (pfn) { 412 zone->zone_start_pfn = pfn; 413 zone->spanned_pages = zone_end_pfn - pfn; 414 } 415 } else if (zone_end_pfn == 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_start_pfn, 423 start_pfn); 424 if (pfn) 425 zone->spanned_pages = pfn - zone_start_pfn + 1; 426 } 427 428 /* 429 * The section is not biggest or smallest mem_section in the zone, it 430 * only creates a hole in the zone. So in this case, we need not 431 * change the zone. But perhaps, the zone has only hole data. Thus 432 * it check the zone has only hole or not. 433 */ 434 pfn = zone_start_pfn; 435 for (; pfn < zone_end_pfn; pfn += PAGES_PER_SUBSECTION) { 436 if (unlikely(!pfn_to_online_page(pfn))) 437 continue; 438 439 if (page_zone(pfn_to_page(pfn)) != zone) 440 continue; 441 442 /* Skip range to be removed */ 443 if (pfn >= start_pfn && pfn < end_pfn) 444 continue; 445 446 /* If we find valid section, we have nothing to do */ 447 zone_span_writeunlock(zone); 448 return; 449 } 450 451 /* The zone has no valid section */ 452 zone->zone_start_pfn = 0; 453 zone->spanned_pages = 0; 454 zone_span_writeunlock(zone); 455 } 456 457 static void update_pgdat_span(struct pglist_data *pgdat) 458 { 459 unsigned long node_start_pfn = 0, node_end_pfn = 0; 460 struct zone *zone; 461 462 for (zone = pgdat->node_zones; 463 zone < pgdat->node_zones + MAX_NR_ZONES; zone++) { 464 unsigned long zone_end_pfn = zone->zone_start_pfn + 465 zone->spanned_pages; 466 467 /* No need to lock the zones, they can't change. */ 468 if (!zone->spanned_pages) 469 continue; 470 if (!node_end_pfn) { 471 node_start_pfn = zone->zone_start_pfn; 472 node_end_pfn = zone_end_pfn; 473 continue; 474 } 475 476 if (zone_end_pfn > node_end_pfn) 477 node_end_pfn = zone_end_pfn; 478 if (zone->zone_start_pfn < node_start_pfn) 479 node_start_pfn = zone->zone_start_pfn; 480 } 481 482 pgdat->node_start_pfn = node_start_pfn; 483 pgdat->node_spanned_pages = node_end_pfn - node_start_pfn; 484 } 485 486 void __ref remove_pfn_range_from_zone(struct zone *zone, 487 unsigned long start_pfn, 488 unsigned long nr_pages) 489 { 490 struct pglist_data *pgdat = zone->zone_pgdat; 491 unsigned long flags; 492 493 #ifdef CONFIG_ZONE_DEVICE 494 /* 495 * Zone shrinking code cannot properly deal with ZONE_DEVICE. So 496 * we will not try to shrink the zones - which is okay as 497 * set_zone_contiguous() cannot deal with ZONE_DEVICE either way. 498 */ 499 if (zone_idx(zone) == ZONE_DEVICE) 500 return; 501 #endif 502 503 clear_zone_contiguous(zone); 504 505 pgdat_resize_lock(zone->zone_pgdat, &flags); 506 shrink_zone_span(zone, start_pfn, start_pfn + nr_pages); 507 update_pgdat_span(pgdat); 508 pgdat_resize_unlock(zone->zone_pgdat, &flags); 509 510 set_zone_contiguous(zone); 511 } 512 513 static void __remove_section(unsigned long pfn, unsigned long nr_pages, 514 unsigned long map_offset, 515 struct vmem_altmap *altmap) 516 { 517 struct mem_section *ms = __nr_to_section(pfn_to_section_nr(pfn)); 518 519 if (WARN_ON_ONCE(!valid_section(ms))) 520 return; 521 522 sparse_remove_section(ms, pfn, nr_pages, map_offset, altmap); 523 } 524 525 /** 526 * __remove_pages() - remove sections of pages 527 * @pfn: starting pageframe (must be aligned to start of a section) 528 * @nr_pages: number of pages to remove (must be multiple of section size) 529 * @altmap: alternative device page map or %NULL if default memmap is used 530 * 531 * Generic helper function to remove section mappings and sysfs entries 532 * for the section of the memory we are removing. Caller needs to make 533 * sure that pages are marked reserved and zones are adjust properly by 534 * calling offline_pages(). 535 */ 536 void __remove_pages(unsigned long pfn, unsigned long nr_pages, 537 struct vmem_altmap *altmap) 538 { 539 unsigned long map_offset = 0; 540 unsigned long nr, start_sec, end_sec; 541 542 map_offset = vmem_altmap_offset(altmap); 543 544 if (check_pfn_span(pfn, nr_pages, "remove")) 545 return; 546 547 start_sec = pfn_to_section_nr(pfn); 548 end_sec = pfn_to_section_nr(pfn + nr_pages - 1); 549 for (nr = start_sec; nr <= end_sec; nr++) { 550 unsigned long pfns; 551 552 cond_resched(); 553 pfns = min(nr_pages, PAGES_PER_SECTION 554 - (pfn & ~PAGE_SECTION_MASK)); 555 __remove_section(pfn, pfns, map_offset, altmap); 556 pfn += pfns; 557 nr_pages -= pfns; 558 map_offset = 0; 559 } 560 } 561 562 int set_online_page_callback(online_page_callback_t callback) 563 { 564 int rc = -EINVAL; 565 566 get_online_mems(); 567 mutex_lock(&online_page_callback_lock); 568 569 if (online_page_callback == generic_online_page) { 570 online_page_callback = callback; 571 rc = 0; 572 } 573 574 mutex_unlock(&online_page_callback_lock); 575 put_online_mems(); 576 577 return rc; 578 } 579 EXPORT_SYMBOL_GPL(set_online_page_callback); 580 581 int restore_online_page_callback(online_page_callback_t callback) 582 { 583 int rc = -EINVAL; 584 585 get_online_mems(); 586 mutex_lock(&online_page_callback_lock); 587 588 if (online_page_callback == callback) { 589 online_page_callback = generic_online_page; 590 rc = 0; 591 } 592 593 mutex_unlock(&online_page_callback_lock); 594 put_online_mems(); 595 596 return rc; 597 } 598 EXPORT_SYMBOL_GPL(restore_online_page_callback); 599 600 void generic_online_page(struct page *page, unsigned int order) 601 { 602 kernel_map_pages(page, 1 << order, 1); 603 __free_pages_core(page, order); 604 totalram_pages_add(1UL << order); 605 #ifdef CONFIG_HIGHMEM 606 if (PageHighMem(page)) 607 totalhigh_pages_add(1UL << order); 608 #endif 609 } 610 EXPORT_SYMBOL_GPL(generic_online_page); 611 612 static int online_pages_range(unsigned long start_pfn, unsigned long nr_pages, 613 void *arg) 614 { 615 const unsigned long end_pfn = start_pfn + nr_pages; 616 unsigned long pfn; 617 int order; 618 619 /* 620 * Online the pages. The callback might decide to keep some pages 621 * PG_reserved (to add them to the buddy later), but we still account 622 * them as being online/belonging to this zone ("present"). 623 */ 624 for (pfn = start_pfn; pfn < end_pfn; pfn += 1ul << order) { 625 order = min(MAX_ORDER - 1, get_order(PFN_PHYS(end_pfn - pfn))); 626 /* __free_pages_core() wants pfns to be aligned to the order */ 627 if (WARN_ON_ONCE(!IS_ALIGNED(pfn, 1ul << order))) 628 order = 0; 629 (*online_page_callback)(pfn_to_page(pfn), order); 630 } 631 632 /* mark all involved sections as online */ 633 online_mem_sections(start_pfn, end_pfn); 634 635 *(unsigned long *)arg += nr_pages; 636 return 0; 637 } 638 639 /* check which state of node_states will be changed when online memory */ 640 static void node_states_check_changes_online(unsigned long nr_pages, 641 struct zone *zone, struct memory_notify *arg) 642 { 643 int nid = zone_to_nid(zone); 644 645 arg->status_change_nid = NUMA_NO_NODE; 646 arg->status_change_nid_normal = NUMA_NO_NODE; 647 arg->status_change_nid_high = NUMA_NO_NODE; 648 649 if (!node_state(nid, N_MEMORY)) 650 arg->status_change_nid = nid; 651 if (zone_idx(zone) <= ZONE_NORMAL && !node_state(nid, N_NORMAL_MEMORY)) 652 arg->status_change_nid_normal = nid; 653 #ifdef CONFIG_HIGHMEM 654 if (zone_idx(zone) <= ZONE_HIGHMEM && !node_state(nid, N_HIGH_MEMORY)) 655 arg->status_change_nid_high = nid; 656 #endif 657 } 658 659 static void node_states_set_node(int node, struct memory_notify *arg) 660 { 661 if (arg->status_change_nid_normal >= 0) 662 node_set_state(node, N_NORMAL_MEMORY); 663 664 if (arg->status_change_nid_high >= 0) 665 node_set_state(node, N_HIGH_MEMORY); 666 667 if (arg->status_change_nid >= 0) 668 node_set_state(node, N_MEMORY); 669 } 670 671 static void __meminit resize_zone_range(struct zone *zone, unsigned long start_pfn, 672 unsigned long nr_pages) 673 { 674 unsigned long old_end_pfn = zone_end_pfn(zone); 675 676 if (zone_is_empty(zone) || start_pfn < zone->zone_start_pfn) 677 zone->zone_start_pfn = start_pfn; 678 679 zone->spanned_pages = max(start_pfn + nr_pages, old_end_pfn) - zone->zone_start_pfn; 680 } 681 682 static void __meminit resize_pgdat_range(struct pglist_data *pgdat, unsigned long start_pfn, 683 unsigned long nr_pages) 684 { 685 unsigned long old_end_pfn = pgdat_end_pfn(pgdat); 686 687 if (!pgdat->node_spanned_pages || start_pfn < pgdat->node_start_pfn) 688 pgdat->node_start_pfn = start_pfn; 689 690 pgdat->node_spanned_pages = max(start_pfn + nr_pages, old_end_pfn) - pgdat->node_start_pfn; 691 692 } 693 /* 694 * Associate the pfn range with the given zone, initializing the memmaps 695 * and resizing the pgdat/zone data to span the added pages. After this 696 * call, all affected pages are PG_reserved. 697 */ 698 void __ref move_pfn_range_to_zone(struct zone *zone, unsigned long start_pfn, 699 unsigned long nr_pages, struct vmem_altmap *altmap) 700 { 701 struct pglist_data *pgdat = zone->zone_pgdat; 702 int nid = pgdat->node_id; 703 unsigned long flags; 704 705 clear_zone_contiguous(zone); 706 707 /* TODO Huh pgdat is irqsave while zone is not. It used to be like that before */ 708 pgdat_resize_lock(pgdat, &flags); 709 zone_span_writelock(zone); 710 if (zone_is_empty(zone)) 711 init_currently_empty_zone(zone, start_pfn, nr_pages); 712 resize_zone_range(zone, start_pfn, nr_pages); 713 zone_span_writeunlock(zone); 714 resize_pgdat_range(pgdat, start_pfn, nr_pages); 715 pgdat_resize_unlock(pgdat, &flags); 716 717 /* 718 * TODO now we have a visible range of pages which are not associated 719 * with their zone properly. Not nice but set_pfnblock_flags_mask 720 * expects the zone spans the pfn range. All the pages in the range 721 * are reserved so nobody should be touching them so we should be safe 722 */ 723 memmap_init_zone(nr_pages, nid, zone_idx(zone), start_pfn, 724 MEMMAP_HOTPLUG, altmap); 725 726 set_zone_contiguous(zone); 727 } 728 729 /* 730 * Returns a default kernel memory zone for the given pfn range. 731 * If no kernel zone covers this pfn range it will automatically go 732 * to the ZONE_NORMAL. 733 */ 734 static struct zone *default_kernel_zone_for_pfn(int nid, unsigned long start_pfn, 735 unsigned long nr_pages) 736 { 737 struct pglist_data *pgdat = NODE_DATA(nid); 738 int zid; 739 740 for (zid = 0; zid <= ZONE_NORMAL; zid++) { 741 struct zone *zone = &pgdat->node_zones[zid]; 742 743 if (zone_intersects(zone, start_pfn, nr_pages)) 744 return zone; 745 } 746 747 return &pgdat->node_zones[ZONE_NORMAL]; 748 } 749 750 static inline struct zone *default_zone_for_pfn(int nid, unsigned long start_pfn, 751 unsigned long nr_pages) 752 { 753 struct zone *kernel_zone = default_kernel_zone_for_pfn(nid, start_pfn, 754 nr_pages); 755 struct zone *movable_zone = &NODE_DATA(nid)->node_zones[ZONE_MOVABLE]; 756 bool in_kernel = zone_intersects(kernel_zone, start_pfn, nr_pages); 757 bool in_movable = zone_intersects(movable_zone, start_pfn, nr_pages); 758 759 /* 760 * We inherit the existing zone in a simple case where zones do not 761 * overlap in the given range 762 */ 763 if (in_kernel ^ in_movable) 764 return (in_kernel) ? kernel_zone : movable_zone; 765 766 /* 767 * If the range doesn't belong to any zone or two zones overlap in the 768 * given range then we use movable zone only if movable_node is 769 * enabled because we always online to a kernel zone by default. 770 */ 771 return movable_node_enabled ? movable_zone : kernel_zone; 772 } 773 774 struct zone * zone_for_pfn_range(int online_type, int nid, unsigned start_pfn, 775 unsigned long nr_pages) 776 { 777 if (online_type == MMOP_ONLINE_KERNEL) 778 return default_kernel_zone_for_pfn(nid, start_pfn, nr_pages); 779 780 if (online_type == MMOP_ONLINE_MOVABLE) 781 return &NODE_DATA(nid)->node_zones[ZONE_MOVABLE]; 782 783 return default_zone_for_pfn(nid, start_pfn, nr_pages); 784 } 785 786 int __ref online_pages(unsigned long pfn, unsigned long nr_pages, int online_type) 787 { 788 unsigned long flags; 789 unsigned long onlined_pages = 0; 790 struct zone *zone; 791 int need_zonelists_rebuild = 0; 792 int nid; 793 int ret; 794 struct memory_notify arg; 795 struct memory_block *mem; 796 797 mem_hotplug_begin(); 798 799 /* 800 * We can't use pfn_to_nid() because nid might be stored in struct page 801 * which is not yet initialized. Instead, we find nid from memory block. 802 */ 803 mem = find_memory_block(__pfn_to_section(pfn)); 804 nid = mem->nid; 805 put_device(&mem->dev); 806 807 /* associate pfn range with the zone */ 808 zone = zone_for_pfn_range(online_type, nid, pfn, nr_pages); 809 move_pfn_range_to_zone(zone, pfn, nr_pages, NULL); 810 811 arg.start_pfn = pfn; 812 arg.nr_pages = nr_pages; 813 node_states_check_changes_online(nr_pages, zone, &arg); 814 815 ret = memory_notify(MEM_GOING_ONLINE, &arg); 816 ret = notifier_to_errno(ret); 817 if (ret) 818 goto failed_addition; 819 820 /* 821 * If this zone is not populated, then it is not in zonelist. 822 * This means the page allocator ignores this zone. 823 * So, zonelist must be updated after online. 824 */ 825 if (!populated_zone(zone)) { 826 need_zonelists_rebuild = 1; 827 setup_zone_pageset(zone); 828 } 829 830 ret = walk_system_ram_range(pfn, nr_pages, &onlined_pages, 831 online_pages_range); 832 if (ret) { 833 /* not a single memory resource was applicable */ 834 if (need_zonelists_rebuild) 835 zone_pcp_reset(zone); 836 goto failed_addition; 837 } 838 839 zone->present_pages += onlined_pages; 840 841 pgdat_resize_lock(zone->zone_pgdat, &flags); 842 zone->zone_pgdat->node_present_pages += onlined_pages; 843 pgdat_resize_unlock(zone->zone_pgdat, &flags); 844 845 shuffle_zone(zone); 846 847 node_states_set_node(nid, &arg); 848 if (need_zonelists_rebuild) 849 build_all_zonelists(NULL); 850 else 851 zone_pcp_update(zone); 852 853 init_per_zone_wmark_min(); 854 855 kswapd_run(nid); 856 kcompactd_run(nid); 857 858 vm_total_pages = nr_free_pagecache_pages(); 859 860 writeback_set_ratelimit(); 861 862 memory_notify(MEM_ONLINE, &arg); 863 mem_hotplug_done(); 864 return 0; 865 866 failed_addition: 867 pr_debug("online_pages [mem %#010llx-%#010llx] failed\n", 868 (unsigned long long) pfn << PAGE_SHIFT, 869 (((unsigned long long) pfn + nr_pages) << PAGE_SHIFT) - 1); 870 memory_notify(MEM_CANCEL_ONLINE, &arg); 871 remove_pfn_range_from_zone(zone, pfn, nr_pages); 872 mem_hotplug_done(); 873 return ret; 874 } 875 #endif /* CONFIG_MEMORY_HOTPLUG_SPARSE */ 876 877 static void reset_node_present_pages(pg_data_t *pgdat) 878 { 879 struct zone *z; 880 881 for (z = pgdat->node_zones; z < pgdat->node_zones + MAX_NR_ZONES; z++) 882 z->present_pages = 0; 883 884 pgdat->node_present_pages = 0; 885 } 886 887 /* we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG */ 888 static pg_data_t __ref *hotadd_new_pgdat(int nid, u64 start) 889 { 890 struct pglist_data *pgdat; 891 unsigned long start_pfn = PFN_DOWN(start); 892 893 pgdat = NODE_DATA(nid); 894 if (!pgdat) { 895 pgdat = arch_alloc_nodedata(nid); 896 if (!pgdat) 897 return NULL; 898 899 pgdat->per_cpu_nodestats = 900 alloc_percpu(struct per_cpu_nodestat); 901 arch_refresh_nodedata(nid, pgdat); 902 } else { 903 int cpu; 904 /* 905 * Reset the nr_zones, order and classzone_idx before reuse. 906 * Note that kswapd will init kswapd_classzone_idx properly 907 * when it starts in the near future. 908 */ 909 pgdat->nr_zones = 0; 910 pgdat->kswapd_order = 0; 911 pgdat->kswapd_classzone_idx = 0; 912 for_each_online_cpu(cpu) { 913 struct per_cpu_nodestat *p; 914 915 p = per_cpu_ptr(pgdat->per_cpu_nodestats, cpu); 916 memset(p, 0, sizeof(*p)); 917 } 918 } 919 920 /* we can use NODE_DATA(nid) from here */ 921 922 pgdat->node_id = nid; 923 pgdat->node_start_pfn = start_pfn; 924 925 /* init node's zones as empty zones, we don't have any present pages.*/ 926 free_area_init_core_hotplug(nid); 927 928 /* 929 * The node we allocated has no zone fallback lists. For avoiding 930 * to access not-initialized zonelist, build here. 931 */ 932 build_all_zonelists(pgdat); 933 934 /* 935 * When memory is hot-added, all the memory is in offline state. So 936 * clear all zones' present_pages because they will be updated in 937 * online_pages() and offline_pages(). 938 */ 939 reset_node_managed_pages(pgdat); 940 reset_node_present_pages(pgdat); 941 942 return pgdat; 943 } 944 945 static void rollback_node_hotadd(int nid) 946 { 947 pg_data_t *pgdat = NODE_DATA(nid); 948 949 arch_refresh_nodedata(nid, NULL); 950 free_percpu(pgdat->per_cpu_nodestats); 951 arch_free_nodedata(pgdat); 952 } 953 954 955 /** 956 * try_online_node - online a node if offlined 957 * @nid: the node ID 958 * @start: start addr of the node 959 * @set_node_online: Whether we want to online the node 960 * called by cpu_up() to online a node without onlined memory. 961 * 962 * Returns: 963 * 1 -> a new node has been allocated 964 * 0 -> the node is already online 965 * -ENOMEM -> the node could not be allocated 966 */ 967 static int __try_online_node(int nid, u64 start, bool set_node_online) 968 { 969 pg_data_t *pgdat; 970 int ret = 1; 971 972 if (node_online(nid)) 973 return 0; 974 975 pgdat = hotadd_new_pgdat(nid, start); 976 if (!pgdat) { 977 pr_err("Cannot online node %d due to NULL pgdat\n", nid); 978 ret = -ENOMEM; 979 goto out; 980 } 981 982 if (set_node_online) { 983 node_set_online(nid); 984 ret = register_one_node(nid); 985 BUG_ON(ret); 986 } 987 out: 988 return ret; 989 } 990 991 /* 992 * Users of this function always want to online/register the node 993 */ 994 int try_online_node(int nid) 995 { 996 int ret; 997 998 mem_hotplug_begin(); 999 ret = __try_online_node(nid, 0, true); 1000 mem_hotplug_done(); 1001 return ret; 1002 } 1003 1004 static int check_hotplug_memory_range(u64 start, u64 size) 1005 { 1006 /* memory range must be block size aligned */ 1007 if (!size || !IS_ALIGNED(start, memory_block_size_bytes()) || 1008 !IS_ALIGNED(size, memory_block_size_bytes())) { 1009 pr_err("Block size [%#lx] unaligned hotplug range: start %#llx, size %#llx", 1010 memory_block_size_bytes(), start, size); 1011 return -EINVAL; 1012 } 1013 1014 return 0; 1015 } 1016 1017 static int online_memory_block(struct memory_block *mem, void *arg) 1018 { 1019 return device_online(&mem->dev); 1020 } 1021 1022 /* 1023 * NOTE: The caller must call lock_device_hotplug() to serialize hotplug 1024 * and online/offline operations (triggered e.g. by sysfs). 1025 * 1026 * we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG 1027 */ 1028 int __ref add_memory_resource(int nid, struct resource *res) 1029 { 1030 struct mhp_restrictions restrictions = {}; 1031 u64 start, size; 1032 bool new_node = false; 1033 int ret; 1034 1035 start = res->start; 1036 size = resource_size(res); 1037 1038 ret = check_hotplug_memory_range(start, size); 1039 if (ret) 1040 return ret; 1041 1042 mem_hotplug_begin(); 1043 1044 /* 1045 * Add new range to memblock so that when hotadd_new_pgdat() is called 1046 * to allocate new pgdat, get_pfn_range_for_nid() will be able to find 1047 * this new range and calculate total pages correctly. The range will 1048 * be removed at hot-remove time. 1049 */ 1050 memblock_add_node(start, size, nid); 1051 1052 ret = __try_online_node(nid, start, false); 1053 if (ret < 0) 1054 goto error; 1055 new_node = ret; 1056 1057 /* call arch's memory hotadd */ 1058 ret = arch_add_memory(nid, start, size, &restrictions); 1059 if (ret < 0) 1060 goto error; 1061 1062 /* create memory block devices after memory was added */ 1063 ret = create_memory_block_devices(start, size); 1064 if (ret) { 1065 arch_remove_memory(nid, start, size, NULL); 1066 goto error; 1067 } 1068 1069 if (new_node) { 1070 /* If sysfs file of new node can't be created, cpu on the node 1071 * can't be hot-added. There is no rollback way now. 1072 * So, check by BUG_ON() to catch it reluctantly.. 1073 * We online node here. We can't roll back from here. 1074 */ 1075 node_set_online(nid); 1076 ret = __register_one_node(nid); 1077 BUG_ON(ret); 1078 } 1079 1080 /* link memory sections under this node.*/ 1081 ret = link_mem_sections(nid, PFN_DOWN(start), PFN_UP(start + size - 1)); 1082 BUG_ON(ret); 1083 1084 /* create new memmap entry */ 1085 firmware_map_add_hotplug(start, start + size, "System RAM"); 1086 1087 /* device_online() will take the lock when calling online_pages() */ 1088 mem_hotplug_done(); 1089 1090 /* online pages if requested */ 1091 if (memhp_auto_online) 1092 walk_memory_blocks(start, size, NULL, online_memory_block); 1093 1094 return ret; 1095 error: 1096 /* rollback pgdat allocation and others */ 1097 if (new_node) 1098 rollback_node_hotadd(nid); 1099 memblock_remove(start, size); 1100 mem_hotplug_done(); 1101 return ret; 1102 } 1103 1104 /* requires device_hotplug_lock, see add_memory_resource() */ 1105 int __ref __add_memory(int nid, u64 start, u64 size) 1106 { 1107 struct resource *res; 1108 int ret; 1109 1110 res = register_memory_resource(start, size); 1111 if (IS_ERR(res)) 1112 return PTR_ERR(res); 1113 1114 ret = add_memory_resource(nid, res); 1115 if (ret < 0) 1116 release_memory_resource(res); 1117 return ret; 1118 } 1119 1120 int add_memory(int nid, u64 start, u64 size) 1121 { 1122 int rc; 1123 1124 lock_device_hotplug(); 1125 rc = __add_memory(nid, start, size); 1126 unlock_device_hotplug(); 1127 1128 return rc; 1129 } 1130 EXPORT_SYMBOL_GPL(add_memory); 1131 1132 #ifdef CONFIG_MEMORY_HOTREMOVE 1133 /* 1134 * A free page on the buddy free lists (not the per-cpu lists) has PageBuddy 1135 * set and the size of the free page is given by page_order(). Using this, 1136 * the function determines if the pageblock contains only free pages. 1137 * Due to buddy contraints, a free page at least the size of a pageblock will 1138 * be located at the start of the pageblock 1139 */ 1140 static inline int pageblock_free(struct page *page) 1141 { 1142 return PageBuddy(page) && page_order(page) >= pageblock_order; 1143 } 1144 1145 /* Return the pfn of the start of the next active pageblock after a given pfn */ 1146 static unsigned long next_active_pageblock(unsigned long pfn) 1147 { 1148 struct page *page = pfn_to_page(pfn); 1149 1150 /* Ensure the starting page is pageblock-aligned */ 1151 BUG_ON(pfn & (pageblock_nr_pages - 1)); 1152 1153 /* If the entire pageblock is free, move to the end of free page */ 1154 if (pageblock_free(page)) { 1155 int order; 1156 /* be careful. we don't have locks, page_order can be changed.*/ 1157 order = page_order(page); 1158 if ((order < MAX_ORDER) && (order >= pageblock_order)) 1159 return pfn + (1 << order); 1160 } 1161 1162 return pfn + pageblock_nr_pages; 1163 } 1164 1165 static bool is_pageblock_removable_nolock(unsigned long pfn) 1166 { 1167 struct page *page = pfn_to_page(pfn); 1168 struct zone *zone; 1169 1170 /* 1171 * We have to be careful here because we are iterating over memory 1172 * sections which are not zone aware so we might end up outside of 1173 * the zone but still within the section. 1174 * We have to take care about the node as well. If the node is offline 1175 * its NODE_DATA will be NULL - see page_zone. 1176 */ 1177 if (!node_online(page_to_nid(page))) 1178 return false; 1179 1180 zone = page_zone(page); 1181 pfn = page_to_pfn(page); 1182 if (!zone_spans_pfn(zone, pfn)) 1183 return false; 1184 1185 return !has_unmovable_pages(zone, page, 0, MIGRATE_MOVABLE, 1186 MEMORY_OFFLINE); 1187 } 1188 1189 /* Checks if this range of memory is likely to be hot-removable. */ 1190 bool is_mem_section_removable(unsigned long start_pfn, unsigned long nr_pages) 1191 { 1192 unsigned long end_pfn, pfn; 1193 1194 end_pfn = min(start_pfn + nr_pages, 1195 zone_end_pfn(page_zone(pfn_to_page(start_pfn)))); 1196 1197 /* Check the starting page of each pageblock within the range */ 1198 for (pfn = start_pfn; pfn < end_pfn; pfn = next_active_pageblock(pfn)) { 1199 if (!is_pageblock_removable_nolock(pfn)) 1200 return false; 1201 cond_resched(); 1202 } 1203 1204 /* All pageblocks in the memory block are likely to be hot-removable */ 1205 return true; 1206 } 1207 1208 /* 1209 * Confirm all pages in a range [start, end) belong to the same zone. 1210 * When true, return its valid [start, end). 1211 */ 1212 int test_pages_in_a_zone(unsigned long start_pfn, unsigned long end_pfn, 1213 unsigned long *valid_start, unsigned long *valid_end) 1214 { 1215 unsigned long pfn, sec_end_pfn; 1216 unsigned long start, end; 1217 struct zone *zone = NULL; 1218 struct page *page; 1219 int i; 1220 for (pfn = start_pfn, sec_end_pfn = SECTION_ALIGN_UP(start_pfn + 1); 1221 pfn < end_pfn; 1222 pfn = sec_end_pfn, sec_end_pfn += PAGES_PER_SECTION) { 1223 /* Make sure the memory section is present first */ 1224 if (!present_section_nr(pfn_to_section_nr(pfn))) 1225 continue; 1226 for (; pfn < sec_end_pfn && pfn < end_pfn; 1227 pfn += MAX_ORDER_NR_PAGES) { 1228 i = 0; 1229 /* This is just a CONFIG_HOLES_IN_ZONE check.*/ 1230 while ((i < MAX_ORDER_NR_PAGES) && 1231 !pfn_valid_within(pfn + i)) 1232 i++; 1233 if (i == MAX_ORDER_NR_PAGES || pfn + i >= end_pfn) 1234 continue; 1235 /* Check if we got outside of the zone */ 1236 if (zone && !zone_spans_pfn(zone, pfn + i)) 1237 return 0; 1238 page = pfn_to_page(pfn + i); 1239 if (zone && page_zone(page) != zone) 1240 return 0; 1241 if (!zone) 1242 start = pfn + i; 1243 zone = page_zone(page); 1244 end = pfn + MAX_ORDER_NR_PAGES; 1245 } 1246 } 1247 1248 if (zone) { 1249 *valid_start = start; 1250 *valid_end = min(end, end_pfn); 1251 return 1; 1252 } else { 1253 return 0; 1254 } 1255 } 1256 1257 /* 1258 * Scan pfn range [start,end) to find movable/migratable pages (LRU pages, 1259 * non-lru movable pages and hugepages). We scan pfn because it's much 1260 * easier than scanning over linked list. This function returns the pfn 1261 * of the first found movable page if it's found, otherwise 0. 1262 */ 1263 static unsigned long scan_movable_pages(unsigned long start, unsigned long end) 1264 { 1265 unsigned long pfn; 1266 1267 for (pfn = start; pfn < end; pfn++) { 1268 struct page *page, *head; 1269 unsigned long skip; 1270 1271 if (!pfn_valid(pfn)) 1272 continue; 1273 page = pfn_to_page(pfn); 1274 if (PageLRU(page)) 1275 return pfn; 1276 if (__PageMovable(page)) 1277 return pfn; 1278 1279 if (!PageHuge(page)) 1280 continue; 1281 head = compound_head(page); 1282 if (page_huge_active(head)) 1283 return pfn; 1284 skip = compound_nr(head) - (page - head); 1285 pfn += skip - 1; 1286 } 1287 return 0; 1288 } 1289 1290 static struct page *new_node_page(struct page *page, unsigned long private) 1291 { 1292 int nid = page_to_nid(page); 1293 nodemask_t nmask = node_states[N_MEMORY]; 1294 1295 /* 1296 * try to allocate from a different node but reuse this node if there 1297 * are no other online nodes to be used (e.g. we are offlining a part 1298 * of the only existing node) 1299 */ 1300 node_clear(nid, nmask); 1301 if (nodes_empty(nmask)) 1302 node_set(nid, nmask); 1303 1304 return new_page_nodemask(page, nid, &nmask); 1305 } 1306 1307 static int 1308 do_migrate_range(unsigned long start_pfn, unsigned long end_pfn) 1309 { 1310 unsigned long pfn; 1311 struct page *page; 1312 int ret = 0; 1313 LIST_HEAD(source); 1314 1315 for (pfn = start_pfn; pfn < end_pfn; pfn++) { 1316 if (!pfn_valid(pfn)) 1317 continue; 1318 page = pfn_to_page(pfn); 1319 1320 if (PageHuge(page)) { 1321 struct page *head = compound_head(page); 1322 pfn = page_to_pfn(head) + compound_nr(head) - 1; 1323 isolate_huge_page(head, &source); 1324 continue; 1325 } else if (PageTransHuge(page)) 1326 pfn = page_to_pfn(compound_head(page)) 1327 + hpage_nr_pages(page) - 1; 1328 1329 /* 1330 * HWPoison pages have elevated reference counts so the migration would 1331 * fail on them. It also doesn't make any sense to migrate them in the 1332 * first place. Still try to unmap such a page in case it is still mapped 1333 * (e.g. current hwpoison implementation doesn't unmap KSM pages but keep 1334 * the unmap as the catch all safety net). 1335 */ 1336 if (PageHWPoison(page)) { 1337 if (WARN_ON(PageLRU(page))) 1338 isolate_lru_page(page); 1339 if (page_mapped(page)) 1340 try_to_unmap(page, TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS); 1341 continue; 1342 } 1343 1344 if (!get_page_unless_zero(page)) 1345 continue; 1346 /* 1347 * We can skip free pages. And we can deal with pages on 1348 * LRU and non-lru movable pages. 1349 */ 1350 if (PageLRU(page)) 1351 ret = isolate_lru_page(page); 1352 else 1353 ret = isolate_movable_page(page, ISOLATE_UNEVICTABLE); 1354 if (!ret) { /* Success */ 1355 list_add_tail(&page->lru, &source); 1356 if (!__PageMovable(page)) 1357 inc_node_page_state(page, NR_ISOLATED_ANON + 1358 page_is_file_cache(page)); 1359 1360 } else { 1361 pr_warn("failed to isolate pfn %lx\n", pfn); 1362 dump_page(page, "isolation failed"); 1363 } 1364 put_page(page); 1365 } 1366 if (!list_empty(&source)) { 1367 /* Allocate a new page from the nearest neighbor node */ 1368 ret = migrate_pages(&source, new_node_page, NULL, 0, 1369 MIGRATE_SYNC, MR_MEMORY_HOTPLUG); 1370 if (ret) { 1371 list_for_each_entry(page, &source, lru) { 1372 pr_warn("migrating pfn %lx failed ret:%d ", 1373 page_to_pfn(page), ret); 1374 dump_page(page, "migration failure"); 1375 } 1376 putback_movable_pages(&source); 1377 } 1378 } 1379 1380 return ret; 1381 } 1382 1383 /* Mark all sections offline and remove all free pages from the buddy. */ 1384 static int 1385 offline_isolated_pages_cb(unsigned long start, unsigned long nr_pages, 1386 void *data) 1387 { 1388 unsigned long *offlined_pages = (unsigned long *)data; 1389 1390 *offlined_pages += __offline_isolated_pages(start, start + nr_pages); 1391 return 0; 1392 } 1393 1394 /* 1395 * Check all pages in range, recoreded as memory resource, are isolated. 1396 */ 1397 static int 1398 check_pages_isolated_cb(unsigned long start_pfn, unsigned long nr_pages, 1399 void *data) 1400 { 1401 return test_pages_isolated(start_pfn, start_pfn + nr_pages, 1402 MEMORY_OFFLINE); 1403 } 1404 1405 static int __init cmdline_parse_movable_node(char *p) 1406 { 1407 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP 1408 movable_node_enabled = true; 1409 #else 1410 pr_warn("movable_node parameter depends on CONFIG_HAVE_MEMBLOCK_NODE_MAP to work properly\n"); 1411 #endif 1412 return 0; 1413 } 1414 early_param("movable_node", cmdline_parse_movable_node); 1415 1416 /* check which state of node_states will be changed when offline memory */ 1417 static void node_states_check_changes_offline(unsigned long nr_pages, 1418 struct zone *zone, struct memory_notify *arg) 1419 { 1420 struct pglist_data *pgdat = zone->zone_pgdat; 1421 unsigned long present_pages = 0; 1422 enum zone_type zt; 1423 1424 arg->status_change_nid = NUMA_NO_NODE; 1425 arg->status_change_nid_normal = NUMA_NO_NODE; 1426 arg->status_change_nid_high = NUMA_NO_NODE; 1427 1428 /* 1429 * Check whether node_states[N_NORMAL_MEMORY] will be changed. 1430 * If the memory to be offline is within the range 1431 * [0..ZONE_NORMAL], and it is the last present memory there, 1432 * the zones in that range will become empty after the offlining, 1433 * thus we can determine that we need to clear the node from 1434 * node_states[N_NORMAL_MEMORY]. 1435 */ 1436 for (zt = 0; zt <= ZONE_NORMAL; zt++) 1437 present_pages += pgdat->node_zones[zt].present_pages; 1438 if (zone_idx(zone) <= ZONE_NORMAL && nr_pages >= present_pages) 1439 arg->status_change_nid_normal = zone_to_nid(zone); 1440 1441 #ifdef CONFIG_HIGHMEM 1442 /* 1443 * node_states[N_HIGH_MEMORY] contains nodes which 1444 * have normal memory or high memory. 1445 * Here we add the present_pages belonging to ZONE_HIGHMEM. 1446 * If the zone is within the range of [0..ZONE_HIGHMEM), and 1447 * we determine that the zones in that range become empty, 1448 * we need to clear the node for N_HIGH_MEMORY. 1449 */ 1450 present_pages += pgdat->node_zones[ZONE_HIGHMEM].present_pages; 1451 if (zone_idx(zone) <= ZONE_HIGHMEM && nr_pages >= present_pages) 1452 arg->status_change_nid_high = zone_to_nid(zone); 1453 #endif 1454 1455 /* 1456 * We have accounted the pages from [0..ZONE_NORMAL), and 1457 * in case of CONFIG_HIGHMEM the pages from ZONE_HIGHMEM 1458 * as well. 1459 * Here we count the possible pages from ZONE_MOVABLE. 1460 * If after having accounted all the pages, we see that the nr_pages 1461 * to be offlined is over or equal to the accounted pages, 1462 * we know that the node will become empty, and so, we can clear 1463 * it for N_MEMORY as well. 1464 */ 1465 present_pages += pgdat->node_zones[ZONE_MOVABLE].present_pages; 1466 1467 if (nr_pages >= present_pages) 1468 arg->status_change_nid = zone_to_nid(zone); 1469 } 1470 1471 static void node_states_clear_node(int node, struct memory_notify *arg) 1472 { 1473 if (arg->status_change_nid_normal >= 0) 1474 node_clear_state(node, N_NORMAL_MEMORY); 1475 1476 if (arg->status_change_nid_high >= 0) 1477 node_clear_state(node, N_HIGH_MEMORY); 1478 1479 if (arg->status_change_nid >= 0) 1480 node_clear_state(node, N_MEMORY); 1481 } 1482 1483 static int count_system_ram_pages_cb(unsigned long start_pfn, 1484 unsigned long nr_pages, void *data) 1485 { 1486 unsigned long *nr_system_ram_pages = data; 1487 1488 *nr_system_ram_pages += nr_pages; 1489 return 0; 1490 } 1491 1492 static int __ref __offline_pages(unsigned long start_pfn, 1493 unsigned long end_pfn) 1494 { 1495 unsigned long pfn, nr_pages = 0; 1496 unsigned long offlined_pages = 0; 1497 int ret, node, nr_isolate_pageblock; 1498 unsigned long flags; 1499 unsigned long valid_start, valid_end; 1500 struct zone *zone; 1501 struct memory_notify arg; 1502 char *reason; 1503 1504 mem_hotplug_begin(); 1505 1506 /* 1507 * Don't allow to offline memory blocks that contain holes. 1508 * Consequently, memory blocks with holes can never get onlined 1509 * via the hotplug path - online_pages() - as hotplugged memory has 1510 * no holes. This way, we e.g., don't have to worry about marking 1511 * memory holes PG_reserved, don't need pfn_valid() checks, and can 1512 * avoid using walk_system_ram_range() later. 1513 */ 1514 walk_system_ram_range(start_pfn, end_pfn - start_pfn, &nr_pages, 1515 count_system_ram_pages_cb); 1516 if (nr_pages != end_pfn - start_pfn) { 1517 ret = -EINVAL; 1518 reason = "memory holes"; 1519 goto failed_removal; 1520 } 1521 1522 /* This makes hotplug much easier...and readable. 1523 we assume this for now. .*/ 1524 if (!test_pages_in_a_zone(start_pfn, end_pfn, &valid_start, 1525 &valid_end)) { 1526 ret = -EINVAL; 1527 reason = "multizone range"; 1528 goto failed_removal; 1529 } 1530 1531 zone = page_zone(pfn_to_page(valid_start)); 1532 node = zone_to_nid(zone); 1533 1534 /* set above range as isolated */ 1535 ret = start_isolate_page_range(start_pfn, end_pfn, 1536 MIGRATE_MOVABLE, 1537 MEMORY_OFFLINE | REPORT_FAILURE); 1538 if (ret < 0) { 1539 reason = "failure to isolate range"; 1540 goto failed_removal; 1541 } 1542 nr_isolate_pageblock = ret; 1543 1544 arg.start_pfn = start_pfn; 1545 arg.nr_pages = nr_pages; 1546 node_states_check_changes_offline(nr_pages, zone, &arg); 1547 1548 ret = memory_notify(MEM_GOING_OFFLINE, &arg); 1549 ret = notifier_to_errno(ret); 1550 if (ret) { 1551 reason = "notifier failure"; 1552 goto failed_removal_isolated; 1553 } 1554 1555 do { 1556 for (pfn = start_pfn; pfn;) { 1557 if (signal_pending(current)) { 1558 ret = -EINTR; 1559 reason = "signal backoff"; 1560 goto failed_removal_isolated; 1561 } 1562 1563 cond_resched(); 1564 lru_add_drain_all(); 1565 1566 pfn = scan_movable_pages(pfn, end_pfn); 1567 if (pfn) { 1568 /* 1569 * TODO: fatal migration failures should bail 1570 * out 1571 */ 1572 do_migrate_range(pfn, end_pfn); 1573 } 1574 } 1575 1576 /* 1577 * Dissolve free hugepages in the memory block before doing 1578 * offlining actually in order to make hugetlbfs's object 1579 * counting consistent. 1580 */ 1581 ret = dissolve_free_huge_pages(start_pfn, end_pfn); 1582 if (ret) { 1583 reason = "failure to dissolve huge pages"; 1584 goto failed_removal_isolated; 1585 } 1586 /* check again */ 1587 ret = walk_system_ram_range(start_pfn, end_pfn - start_pfn, 1588 NULL, check_pages_isolated_cb); 1589 } while (ret); 1590 1591 /* Ok, all of our target is isolated. 1592 We cannot do rollback at this point. */ 1593 walk_system_ram_range(start_pfn, end_pfn - start_pfn, 1594 &offlined_pages, offline_isolated_pages_cb); 1595 pr_info("Offlined Pages %ld\n", offlined_pages); 1596 /* 1597 * Onlining will reset pagetype flags and makes migrate type 1598 * MOVABLE, so just need to decrease the number of isolated 1599 * pageblocks zone counter here. 1600 */ 1601 spin_lock_irqsave(&zone->lock, flags); 1602 zone->nr_isolate_pageblock -= nr_isolate_pageblock; 1603 spin_unlock_irqrestore(&zone->lock, flags); 1604 1605 /* removal success */ 1606 adjust_managed_page_count(pfn_to_page(start_pfn), -offlined_pages); 1607 zone->present_pages -= offlined_pages; 1608 1609 pgdat_resize_lock(zone->zone_pgdat, &flags); 1610 zone->zone_pgdat->node_present_pages -= offlined_pages; 1611 pgdat_resize_unlock(zone->zone_pgdat, &flags); 1612 1613 init_per_zone_wmark_min(); 1614 1615 if (!populated_zone(zone)) { 1616 zone_pcp_reset(zone); 1617 build_all_zonelists(NULL); 1618 } else 1619 zone_pcp_update(zone); 1620 1621 node_states_clear_node(node, &arg); 1622 if (arg.status_change_nid >= 0) { 1623 kswapd_stop(node); 1624 kcompactd_stop(node); 1625 } 1626 1627 vm_total_pages = nr_free_pagecache_pages(); 1628 writeback_set_ratelimit(); 1629 1630 memory_notify(MEM_OFFLINE, &arg); 1631 remove_pfn_range_from_zone(zone, start_pfn, nr_pages); 1632 mem_hotplug_done(); 1633 return 0; 1634 1635 failed_removal_isolated: 1636 undo_isolate_page_range(start_pfn, end_pfn, MIGRATE_MOVABLE); 1637 memory_notify(MEM_CANCEL_OFFLINE, &arg); 1638 failed_removal: 1639 pr_debug("memory offlining [mem %#010llx-%#010llx] failed due to %s\n", 1640 (unsigned long long) start_pfn << PAGE_SHIFT, 1641 ((unsigned long long) end_pfn << PAGE_SHIFT) - 1, 1642 reason); 1643 /* pushback to free area */ 1644 mem_hotplug_done(); 1645 return ret; 1646 } 1647 1648 int offline_pages(unsigned long start_pfn, unsigned long nr_pages) 1649 { 1650 return __offline_pages(start_pfn, start_pfn + nr_pages); 1651 } 1652 1653 static int check_memblock_offlined_cb(struct memory_block *mem, void *arg) 1654 { 1655 int ret = !is_memblock_offlined(mem); 1656 1657 if (unlikely(ret)) { 1658 phys_addr_t beginpa, endpa; 1659 1660 beginpa = PFN_PHYS(section_nr_to_pfn(mem->start_section_nr)); 1661 endpa = beginpa + memory_block_size_bytes() - 1; 1662 pr_warn("removing memory fails, because memory [%pa-%pa] is onlined\n", 1663 &beginpa, &endpa); 1664 1665 return -EBUSY; 1666 } 1667 return 0; 1668 } 1669 1670 static int check_cpu_on_node(pg_data_t *pgdat) 1671 { 1672 int cpu; 1673 1674 for_each_present_cpu(cpu) { 1675 if (cpu_to_node(cpu) == pgdat->node_id) 1676 /* 1677 * the cpu on this node isn't removed, and we can't 1678 * offline this node. 1679 */ 1680 return -EBUSY; 1681 } 1682 1683 return 0; 1684 } 1685 1686 static int check_no_memblock_for_node_cb(struct memory_block *mem, void *arg) 1687 { 1688 int nid = *(int *)arg; 1689 1690 /* 1691 * If a memory block belongs to multiple nodes, the stored nid is not 1692 * reliable. However, such blocks are always online (e.g., cannot get 1693 * offlined) and, therefore, are still spanned by the node. 1694 */ 1695 return mem->nid == nid ? -EEXIST : 0; 1696 } 1697 1698 /** 1699 * try_offline_node 1700 * @nid: the node ID 1701 * 1702 * Offline a node if all memory sections and cpus of the node are removed. 1703 * 1704 * NOTE: The caller must call lock_device_hotplug() to serialize hotplug 1705 * and online/offline operations before this call. 1706 */ 1707 void try_offline_node(int nid) 1708 { 1709 pg_data_t *pgdat = NODE_DATA(nid); 1710 int rc; 1711 1712 /* 1713 * If the node still spans pages (especially ZONE_DEVICE), don't 1714 * offline it. A node spans memory after move_pfn_range_to_zone(), 1715 * e.g., after the memory block was onlined. 1716 */ 1717 if (pgdat->node_spanned_pages) 1718 return; 1719 1720 /* 1721 * Especially offline memory blocks might not be spanned by the 1722 * node. They will get spanned by the node once they get onlined. 1723 * However, they link to the node in sysfs and can get onlined later. 1724 */ 1725 rc = for_each_memory_block(&nid, check_no_memblock_for_node_cb); 1726 if (rc) 1727 return; 1728 1729 if (check_cpu_on_node(pgdat)) 1730 return; 1731 1732 /* 1733 * all memory/cpu of this node are removed, we can offline this 1734 * node now. 1735 */ 1736 node_set_offline(nid); 1737 unregister_one_node(nid); 1738 } 1739 EXPORT_SYMBOL(try_offline_node); 1740 1741 static void __release_memory_resource(resource_size_t start, 1742 resource_size_t size) 1743 { 1744 int ret; 1745 1746 /* 1747 * When removing memory in the same granularity as it was added, 1748 * this function never fails. It might only fail if resources 1749 * have to be adjusted or split. We'll ignore the error, as 1750 * removing of memory cannot fail. 1751 */ 1752 ret = release_mem_region_adjustable(&iomem_resource, start, size); 1753 if (ret) { 1754 resource_size_t endres = start + size - 1; 1755 1756 pr_warn("Unable to release resource <%pa-%pa> (%d)\n", 1757 &start, &endres, ret); 1758 } 1759 } 1760 1761 static int __ref try_remove_memory(int nid, u64 start, u64 size) 1762 { 1763 int rc = 0; 1764 1765 BUG_ON(check_hotplug_memory_range(start, size)); 1766 1767 mem_hotplug_begin(); 1768 1769 /* 1770 * All memory blocks must be offlined before removing memory. Check 1771 * whether all memory blocks in question are offline and return error 1772 * if this is not the case. 1773 */ 1774 rc = walk_memory_blocks(start, size, NULL, check_memblock_offlined_cb); 1775 if (rc) 1776 goto done; 1777 1778 /* remove memmap entry */ 1779 firmware_map_remove(start, start + size, "System RAM"); 1780 1781 /* remove memory block devices before removing memory */ 1782 remove_memory_block_devices(start, size); 1783 1784 arch_remove_memory(nid, start, size, NULL); 1785 memblock_free(start, size); 1786 memblock_remove(start, size); 1787 __release_memory_resource(start, size); 1788 1789 try_offline_node(nid); 1790 1791 done: 1792 mem_hotplug_done(); 1793 return rc; 1794 } 1795 1796 /** 1797 * remove_memory 1798 * @nid: the node ID 1799 * @start: physical address of the region to remove 1800 * @size: size of the region to remove 1801 * 1802 * NOTE: The caller must call lock_device_hotplug() to serialize hotplug 1803 * and online/offline operations before this call, as required by 1804 * try_offline_node(). 1805 */ 1806 void __remove_memory(int nid, u64 start, u64 size) 1807 { 1808 1809 /* 1810 * trigger BUG() if some memory is not offlined prior to calling this 1811 * function 1812 */ 1813 if (try_remove_memory(nid, start, size)) 1814 BUG(); 1815 } 1816 1817 /* 1818 * Remove memory if every memory block is offline, otherwise return -EBUSY is 1819 * some memory is not offline 1820 */ 1821 int remove_memory(int nid, u64 start, u64 size) 1822 { 1823 int rc; 1824 1825 lock_device_hotplug(); 1826 rc = try_remove_memory(nid, start, size); 1827 unlock_device_hotplug(); 1828 1829 return rc; 1830 } 1831 EXPORT_SYMBOL_GPL(remove_memory); 1832 #endif /* CONFIG_MEMORY_HOTREMOVE */ 1833