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 int memhp_default_online_type = MMOP_OFFLINE; 71 #else 72 int memhp_default_online_type = MMOP_ONLINE; 73 #endif 74 75 static int __init setup_memhp_default_state(char *str) 76 { 77 const int online_type = memhp_online_type_from_str(str); 78 79 if (online_type >= 0) 80 memhp_default_online_type = online_type; 81 82 return 1; 83 } 84 __setup("memhp_default_state=", setup_memhp_default_state); 85 86 void mem_hotplug_begin(void) 87 { 88 cpus_read_lock(); 89 percpu_down_write(&mem_hotplug_lock); 90 } 91 92 void mem_hotplug_done(void) 93 { 94 percpu_up_write(&mem_hotplug_lock); 95 cpus_read_unlock(); 96 } 97 98 u64 max_mem_size = U64_MAX; 99 100 /* add this memory to iomem resource */ 101 static struct resource *register_memory_resource(u64 start, u64 size, 102 const char *resource_name) 103 { 104 struct resource *res; 105 unsigned long flags = IORESOURCE_SYSTEM_RAM | IORESOURCE_BUSY; 106 107 if (strcmp(resource_name, "System RAM")) 108 flags |= IORESOURCE_MEM_DRIVER_MANAGED; 109 110 /* 111 * Make sure value parsed from 'mem=' only restricts memory adding 112 * while booting, so that memory hotplug won't be impacted. Please 113 * refer to document of 'mem=' in kernel-parameters.txt for more 114 * details. 115 */ 116 if (start + size > max_mem_size && system_state < SYSTEM_RUNNING) 117 return ERR_PTR(-E2BIG); 118 119 /* 120 * Request ownership of the new memory range. This might be 121 * a child of an existing resource that was present but 122 * not marked as busy. 123 */ 124 res = __request_region(&iomem_resource, start, size, 125 resource_name, flags); 126 127 if (!res) { 128 pr_debug("Unable to reserve System RAM region: %016llx->%016llx\n", 129 start, start + size); 130 return ERR_PTR(-EEXIST); 131 } 132 return res; 133 } 134 135 static void release_memory_resource(struct resource *res) 136 { 137 if (!res) 138 return; 139 release_resource(res); 140 kfree(res); 141 } 142 143 #ifdef CONFIG_MEMORY_HOTPLUG_SPARSE 144 void get_page_bootmem(unsigned long info, struct page *page, 145 unsigned long type) 146 { 147 page->freelist = (void *)type; 148 SetPagePrivate(page); 149 set_page_private(page, info); 150 page_ref_inc(page); 151 } 152 153 void put_page_bootmem(struct page *page) 154 { 155 unsigned long type; 156 157 type = (unsigned long) page->freelist; 158 BUG_ON(type < MEMORY_HOTPLUG_MIN_BOOTMEM_TYPE || 159 type > MEMORY_HOTPLUG_MAX_BOOTMEM_TYPE); 160 161 if (page_ref_dec_return(page) == 1) { 162 page->freelist = NULL; 163 ClearPagePrivate(page); 164 set_page_private(page, 0); 165 INIT_LIST_HEAD(&page->lru); 166 free_reserved_page(page); 167 } 168 } 169 170 #ifdef CONFIG_HAVE_BOOTMEM_INFO_NODE 171 #ifndef CONFIG_SPARSEMEM_VMEMMAP 172 static void register_page_bootmem_info_section(unsigned long start_pfn) 173 { 174 unsigned long mapsize, section_nr, i; 175 struct mem_section *ms; 176 struct page *page, *memmap; 177 struct mem_section_usage *usage; 178 179 section_nr = pfn_to_section_nr(start_pfn); 180 ms = __nr_to_section(section_nr); 181 182 /* Get section's memmap address */ 183 memmap = sparse_decode_mem_map(ms->section_mem_map, section_nr); 184 185 /* 186 * Get page for the memmap's phys address 187 * XXX: need more consideration for sparse_vmemmap... 188 */ 189 page = virt_to_page(memmap); 190 mapsize = sizeof(struct page) * PAGES_PER_SECTION; 191 mapsize = PAGE_ALIGN(mapsize) >> PAGE_SHIFT; 192 193 /* remember memmap's page */ 194 for (i = 0; i < mapsize; i++, page++) 195 get_page_bootmem(section_nr, page, SECTION_INFO); 196 197 usage = ms->usage; 198 page = virt_to_page(usage); 199 200 mapsize = PAGE_ALIGN(mem_section_usage_size()) >> PAGE_SHIFT; 201 202 for (i = 0; i < mapsize; i++, page++) 203 get_page_bootmem(section_nr, page, MIX_SECTION_INFO); 204 205 } 206 #else /* CONFIG_SPARSEMEM_VMEMMAP */ 207 static void register_page_bootmem_info_section(unsigned long start_pfn) 208 { 209 unsigned long mapsize, section_nr, i; 210 struct mem_section *ms; 211 struct page *page, *memmap; 212 struct mem_section_usage *usage; 213 214 section_nr = pfn_to_section_nr(start_pfn); 215 ms = __nr_to_section(section_nr); 216 217 memmap = sparse_decode_mem_map(ms->section_mem_map, section_nr); 218 219 register_page_bootmem_memmap(section_nr, memmap, PAGES_PER_SECTION); 220 221 usage = ms->usage; 222 page = virt_to_page(usage); 223 224 mapsize = PAGE_ALIGN(mem_section_usage_size()) >> PAGE_SHIFT; 225 226 for (i = 0; i < mapsize; i++, page++) 227 get_page_bootmem(section_nr, page, MIX_SECTION_INFO); 228 } 229 #endif /* !CONFIG_SPARSEMEM_VMEMMAP */ 230 231 void __init register_page_bootmem_info_node(struct pglist_data *pgdat) 232 { 233 unsigned long i, pfn, end_pfn, nr_pages; 234 int node = pgdat->node_id; 235 struct page *page; 236 237 nr_pages = PAGE_ALIGN(sizeof(struct pglist_data)) >> PAGE_SHIFT; 238 page = virt_to_page(pgdat); 239 240 for (i = 0; i < nr_pages; i++, page++) 241 get_page_bootmem(node, page, NODE_INFO); 242 243 pfn = pgdat->node_start_pfn; 244 end_pfn = pgdat_end_pfn(pgdat); 245 246 /* register section info */ 247 for (; pfn < end_pfn; pfn += PAGES_PER_SECTION) { 248 /* 249 * Some platforms can assign the same pfn to multiple nodes - on 250 * node0 as well as nodeN. To avoid registering a pfn against 251 * multiple nodes we check that this pfn does not already 252 * reside in some other nodes. 253 */ 254 if (pfn_valid(pfn) && (early_pfn_to_nid(pfn) == node)) 255 register_page_bootmem_info_section(pfn); 256 } 257 } 258 #endif /* CONFIG_HAVE_BOOTMEM_INFO_NODE */ 259 260 static int check_pfn_span(unsigned long pfn, unsigned long nr_pages, 261 const char *reason) 262 { 263 /* 264 * Disallow all operations smaller than a sub-section and only 265 * allow operations smaller than a section for 266 * SPARSEMEM_VMEMMAP. Note that check_hotplug_memory_range() 267 * enforces a larger memory_block_size_bytes() granularity for 268 * memory that will be marked online, so this check should only 269 * fire for direct arch_{add,remove}_memory() users outside of 270 * add_memory_resource(). 271 */ 272 unsigned long min_align; 273 274 if (IS_ENABLED(CONFIG_SPARSEMEM_VMEMMAP)) 275 min_align = PAGES_PER_SUBSECTION; 276 else 277 min_align = PAGES_PER_SECTION; 278 if (!IS_ALIGNED(pfn, min_align) 279 || !IS_ALIGNED(nr_pages, min_align)) { 280 WARN(1, "Misaligned __%s_pages start: %#lx end: #%lx\n", 281 reason, pfn, pfn + nr_pages - 1); 282 return -EINVAL; 283 } 284 return 0; 285 } 286 287 static int check_hotplug_memory_addressable(unsigned long pfn, 288 unsigned long nr_pages) 289 { 290 const u64 max_addr = PFN_PHYS(pfn + nr_pages) - 1; 291 292 if (max_addr >> MAX_PHYSMEM_BITS) { 293 const u64 max_allowed = (1ull << (MAX_PHYSMEM_BITS + 1)) - 1; 294 WARN(1, 295 "Hotplugged memory exceeds maximum addressable address, range=%#llx-%#llx, maximum=%#llx\n", 296 (u64)PFN_PHYS(pfn), max_addr, max_allowed); 297 return -E2BIG; 298 } 299 300 return 0; 301 } 302 303 /* 304 * Reasonably generic function for adding memory. It is 305 * expected that archs that support memory hotplug will 306 * call this function after deciding the zone to which to 307 * add the new pages. 308 */ 309 int __ref __add_pages(int nid, unsigned long pfn, unsigned long nr_pages, 310 struct mhp_params *params) 311 { 312 const unsigned long end_pfn = pfn + nr_pages; 313 unsigned long cur_nr_pages; 314 int err; 315 struct vmem_altmap *altmap = params->altmap; 316 317 if (WARN_ON_ONCE(!params->pgprot.pgprot)) 318 return -EINVAL; 319 320 err = check_hotplug_memory_addressable(pfn, nr_pages); 321 if (err) 322 return err; 323 324 if (altmap) { 325 /* 326 * Validate altmap is within bounds of the total request 327 */ 328 if (altmap->base_pfn != pfn 329 || vmem_altmap_offset(altmap) > nr_pages) { 330 pr_warn_once("memory add fail, invalid altmap\n"); 331 return -EINVAL; 332 } 333 altmap->alloc = 0; 334 } 335 336 err = check_pfn_span(pfn, nr_pages, "add"); 337 if (err) 338 return err; 339 340 for (; pfn < end_pfn; pfn += cur_nr_pages) { 341 /* Select all remaining pages up to the next section boundary */ 342 cur_nr_pages = min(end_pfn - pfn, 343 SECTION_ALIGN_UP(pfn + 1) - pfn); 344 err = sparse_add_section(nid, pfn, cur_nr_pages, altmap); 345 if (err) 346 break; 347 cond_resched(); 348 } 349 vmemmap_populate_print_last(); 350 return err; 351 } 352 353 #ifdef CONFIG_NUMA 354 int __weak memory_add_physaddr_to_nid(u64 start) 355 { 356 pr_info_once("Unknown target node for memory at 0x%llx, assuming node 0\n", 357 start); 358 return 0; 359 } 360 EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid); 361 #endif 362 363 /* find the smallest valid pfn in the range [start_pfn, end_pfn) */ 364 static unsigned long find_smallest_section_pfn(int nid, struct zone *zone, 365 unsigned long start_pfn, 366 unsigned long end_pfn) 367 { 368 for (; start_pfn < end_pfn; start_pfn += PAGES_PER_SUBSECTION) { 369 if (unlikely(!pfn_to_online_page(start_pfn))) 370 continue; 371 372 if (unlikely(pfn_to_nid(start_pfn) != nid)) 373 continue; 374 375 if (zone != page_zone(pfn_to_page(start_pfn))) 376 continue; 377 378 return start_pfn; 379 } 380 381 return 0; 382 } 383 384 /* find the biggest valid pfn in the range [start_pfn, end_pfn). */ 385 static unsigned long find_biggest_section_pfn(int nid, struct zone *zone, 386 unsigned long start_pfn, 387 unsigned long end_pfn) 388 { 389 unsigned long pfn; 390 391 /* pfn is the end pfn of a memory section. */ 392 pfn = end_pfn - 1; 393 for (; pfn >= start_pfn; pfn -= PAGES_PER_SUBSECTION) { 394 if (unlikely(!pfn_to_online_page(pfn))) 395 continue; 396 397 if (unlikely(pfn_to_nid(pfn) != nid)) 398 continue; 399 400 if (zone != page_zone(pfn_to_page(pfn))) 401 continue; 402 403 return pfn; 404 } 405 406 return 0; 407 } 408 409 static void shrink_zone_span(struct zone *zone, unsigned long start_pfn, 410 unsigned long end_pfn) 411 { 412 unsigned long pfn; 413 int nid = zone_to_nid(zone); 414 415 zone_span_writelock(zone); 416 if (zone->zone_start_pfn == start_pfn) { 417 /* 418 * If the section is smallest section in the zone, it need 419 * shrink zone->zone_start_pfn and zone->zone_spanned_pages. 420 * In this case, we find second smallest valid mem_section 421 * for shrinking zone. 422 */ 423 pfn = find_smallest_section_pfn(nid, zone, end_pfn, 424 zone_end_pfn(zone)); 425 if (pfn) { 426 zone->spanned_pages = zone_end_pfn(zone) - pfn; 427 zone->zone_start_pfn = pfn; 428 } else { 429 zone->zone_start_pfn = 0; 430 zone->spanned_pages = 0; 431 } 432 } else if (zone_end_pfn(zone) == end_pfn) { 433 /* 434 * If the section is biggest section in the zone, it need 435 * shrink zone->spanned_pages. 436 * In this case, we find second biggest valid mem_section for 437 * shrinking zone. 438 */ 439 pfn = find_biggest_section_pfn(nid, zone, zone->zone_start_pfn, 440 start_pfn); 441 if (pfn) 442 zone->spanned_pages = pfn - zone->zone_start_pfn + 1; 443 else { 444 zone->zone_start_pfn = 0; 445 zone->spanned_pages = 0; 446 } 447 } 448 zone_span_writeunlock(zone); 449 } 450 451 static void update_pgdat_span(struct pglist_data *pgdat) 452 { 453 unsigned long node_start_pfn = 0, node_end_pfn = 0; 454 struct zone *zone; 455 456 for (zone = pgdat->node_zones; 457 zone < pgdat->node_zones + MAX_NR_ZONES; zone++) { 458 unsigned long zone_end_pfn = zone->zone_start_pfn + 459 zone->spanned_pages; 460 461 /* No need to lock the zones, they can't change. */ 462 if (!zone->spanned_pages) 463 continue; 464 if (!node_end_pfn) { 465 node_start_pfn = zone->zone_start_pfn; 466 node_end_pfn = zone_end_pfn; 467 continue; 468 } 469 470 if (zone_end_pfn > node_end_pfn) 471 node_end_pfn = zone_end_pfn; 472 if (zone->zone_start_pfn < node_start_pfn) 473 node_start_pfn = zone->zone_start_pfn; 474 } 475 476 pgdat->node_start_pfn = node_start_pfn; 477 pgdat->node_spanned_pages = node_end_pfn - node_start_pfn; 478 } 479 480 void __ref remove_pfn_range_from_zone(struct zone *zone, 481 unsigned long start_pfn, 482 unsigned long nr_pages) 483 { 484 const unsigned long end_pfn = start_pfn + nr_pages; 485 struct pglist_data *pgdat = zone->zone_pgdat; 486 unsigned long pfn, cur_nr_pages, flags; 487 488 /* Poison struct pages because they are now uninitialized again. */ 489 for (pfn = start_pfn; pfn < end_pfn; pfn += cur_nr_pages) { 490 cond_resched(); 491 492 /* Select all remaining pages up to the next section boundary */ 493 cur_nr_pages = 494 min(end_pfn - pfn, SECTION_ALIGN_UP(pfn + 1) - pfn); 495 page_init_poison(pfn_to_page(pfn), 496 sizeof(struct page) * cur_nr_pages); 497 } 498 499 #ifdef CONFIG_ZONE_DEVICE 500 /* 501 * Zone shrinking code cannot properly deal with ZONE_DEVICE. So 502 * we will not try to shrink the zones - which is okay as 503 * set_zone_contiguous() cannot deal with ZONE_DEVICE either way. 504 */ 505 if (zone_idx(zone) == ZONE_DEVICE) 506 return; 507 #endif 508 509 clear_zone_contiguous(zone); 510 511 pgdat_resize_lock(zone->zone_pgdat, &flags); 512 shrink_zone_span(zone, start_pfn, start_pfn + nr_pages); 513 update_pgdat_span(pgdat); 514 pgdat_resize_unlock(zone->zone_pgdat, &flags); 515 516 set_zone_contiguous(zone); 517 } 518 519 static void __remove_section(unsigned long pfn, unsigned long nr_pages, 520 unsigned long map_offset, 521 struct vmem_altmap *altmap) 522 { 523 struct mem_section *ms = __pfn_to_section(pfn); 524 525 if (WARN_ON_ONCE(!valid_section(ms))) 526 return; 527 528 sparse_remove_section(ms, pfn, nr_pages, map_offset, altmap); 529 } 530 531 /** 532 * __remove_pages() - remove sections of pages 533 * @pfn: starting pageframe (must be aligned to start of a section) 534 * @nr_pages: number of pages to remove (must be multiple of section size) 535 * @altmap: alternative device page map or %NULL if default memmap is used 536 * 537 * Generic helper function to remove section mappings and sysfs entries 538 * for the section of the memory we are removing. Caller needs to make 539 * sure that pages are marked reserved and zones are adjust properly by 540 * calling offline_pages(). 541 */ 542 void __remove_pages(unsigned long pfn, unsigned long nr_pages, 543 struct vmem_altmap *altmap) 544 { 545 const unsigned long end_pfn = pfn + nr_pages; 546 unsigned long cur_nr_pages; 547 unsigned long map_offset = 0; 548 549 map_offset = vmem_altmap_offset(altmap); 550 551 if (check_pfn_span(pfn, nr_pages, "remove")) 552 return; 553 554 for (; pfn < end_pfn; pfn += cur_nr_pages) { 555 cond_resched(); 556 /* Select all remaining pages up to the next section boundary */ 557 cur_nr_pages = min(end_pfn - pfn, 558 SECTION_ALIGN_UP(pfn + 1) - pfn); 559 __remove_section(pfn, cur_nr_pages, map_offset, altmap); 560 map_offset = 0; 561 } 562 } 563 564 int set_online_page_callback(online_page_callback_t callback) 565 { 566 int rc = -EINVAL; 567 568 get_online_mems(); 569 mutex_lock(&online_page_callback_lock); 570 571 if (online_page_callback == generic_online_page) { 572 online_page_callback = callback; 573 rc = 0; 574 } 575 576 mutex_unlock(&online_page_callback_lock); 577 put_online_mems(); 578 579 return rc; 580 } 581 EXPORT_SYMBOL_GPL(set_online_page_callback); 582 583 int restore_online_page_callback(online_page_callback_t callback) 584 { 585 int rc = -EINVAL; 586 587 get_online_mems(); 588 mutex_lock(&online_page_callback_lock); 589 590 if (online_page_callback == callback) { 591 online_page_callback = generic_online_page; 592 rc = 0; 593 } 594 595 mutex_unlock(&online_page_callback_lock); 596 put_online_mems(); 597 598 return rc; 599 } 600 EXPORT_SYMBOL_GPL(restore_online_page_callback); 601 602 void generic_online_page(struct page *page, unsigned int order) 603 { 604 /* 605 * Freeing the page with debug_pagealloc enabled will try to unmap it, 606 * so we should map it first. This is better than introducing a special 607 * case in page freeing fast path. 608 */ 609 if (debug_pagealloc_enabled_static()) 610 kernel_map_pages(page, 1 << order, 1); 611 __free_pages_core(page, order); 612 totalram_pages_add(1UL << order); 613 #ifdef CONFIG_HIGHMEM 614 if (PageHighMem(page)) 615 totalhigh_pages_add(1UL << order); 616 #endif 617 } 618 EXPORT_SYMBOL_GPL(generic_online_page); 619 620 static int online_pages_range(unsigned long start_pfn, unsigned long nr_pages, 621 void *arg) 622 { 623 const unsigned long end_pfn = start_pfn + nr_pages; 624 unsigned long pfn; 625 int order; 626 627 /* 628 * Online the pages. The callback might decide to keep some pages 629 * PG_reserved (to add them to the buddy later), but we still account 630 * them as being online/belonging to this zone ("present"). 631 */ 632 for (pfn = start_pfn; pfn < end_pfn; pfn += 1ul << order) { 633 order = min(MAX_ORDER - 1, get_order(PFN_PHYS(end_pfn - pfn))); 634 /* __free_pages_core() wants pfns to be aligned to the order */ 635 if (WARN_ON_ONCE(!IS_ALIGNED(pfn, 1ul << order))) 636 order = 0; 637 (*online_page_callback)(pfn_to_page(pfn), order); 638 } 639 640 /* mark all involved sections as online */ 641 online_mem_sections(start_pfn, end_pfn); 642 643 *(unsigned long *)arg += nr_pages; 644 return 0; 645 } 646 647 /* check which state of node_states will be changed when online memory */ 648 static void node_states_check_changes_online(unsigned long nr_pages, 649 struct zone *zone, struct memory_notify *arg) 650 { 651 int nid = zone_to_nid(zone); 652 653 arg->status_change_nid = NUMA_NO_NODE; 654 arg->status_change_nid_normal = NUMA_NO_NODE; 655 arg->status_change_nid_high = NUMA_NO_NODE; 656 657 if (!node_state(nid, N_MEMORY)) 658 arg->status_change_nid = nid; 659 if (zone_idx(zone) <= ZONE_NORMAL && !node_state(nid, N_NORMAL_MEMORY)) 660 arg->status_change_nid_normal = nid; 661 #ifdef CONFIG_HIGHMEM 662 if (zone_idx(zone) <= ZONE_HIGHMEM && !node_state(nid, N_HIGH_MEMORY)) 663 arg->status_change_nid_high = nid; 664 #endif 665 } 666 667 static void node_states_set_node(int node, struct memory_notify *arg) 668 { 669 if (arg->status_change_nid_normal >= 0) 670 node_set_state(node, N_NORMAL_MEMORY); 671 672 if (arg->status_change_nid_high >= 0) 673 node_set_state(node, N_HIGH_MEMORY); 674 675 if (arg->status_change_nid >= 0) 676 node_set_state(node, N_MEMORY); 677 } 678 679 static void __meminit resize_zone_range(struct zone *zone, unsigned long start_pfn, 680 unsigned long nr_pages) 681 { 682 unsigned long old_end_pfn = zone_end_pfn(zone); 683 684 if (zone_is_empty(zone) || start_pfn < zone->zone_start_pfn) 685 zone->zone_start_pfn = start_pfn; 686 687 zone->spanned_pages = max(start_pfn + nr_pages, old_end_pfn) - zone->zone_start_pfn; 688 } 689 690 static void __meminit resize_pgdat_range(struct pglist_data *pgdat, unsigned long start_pfn, 691 unsigned long nr_pages) 692 { 693 unsigned long old_end_pfn = pgdat_end_pfn(pgdat); 694 695 if (!pgdat->node_spanned_pages || start_pfn < pgdat->node_start_pfn) 696 pgdat->node_start_pfn = start_pfn; 697 698 pgdat->node_spanned_pages = max(start_pfn + nr_pages, old_end_pfn) - pgdat->node_start_pfn; 699 700 } 701 /* 702 * Associate the pfn range with the given zone, initializing the memmaps 703 * and resizing the pgdat/zone data to span the added pages. After this 704 * call, all affected pages are PG_reserved. 705 */ 706 void __ref move_pfn_range_to_zone(struct zone *zone, unsigned long start_pfn, 707 unsigned long nr_pages, struct vmem_altmap *altmap) 708 { 709 struct pglist_data *pgdat = zone->zone_pgdat; 710 int nid = pgdat->node_id; 711 unsigned long flags; 712 713 clear_zone_contiguous(zone); 714 715 /* TODO Huh pgdat is irqsave while zone is not. It used to be like that before */ 716 pgdat_resize_lock(pgdat, &flags); 717 zone_span_writelock(zone); 718 if (zone_is_empty(zone)) 719 init_currently_empty_zone(zone, start_pfn, nr_pages); 720 resize_zone_range(zone, start_pfn, nr_pages); 721 zone_span_writeunlock(zone); 722 resize_pgdat_range(pgdat, start_pfn, nr_pages); 723 pgdat_resize_unlock(pgdat, &flags); 724 725 /* 726 * TODO now we have a visible range of pages which are not associated 727 * with their zone properly. Not nice but set_pfnblock_flags_mask 728 * expects the zone spans the pfn range. All the pages in the range 729 * are reserved so nobody should be touching them so we should be safe 730 */ 731 memmap_init_zone(nr_pages, nid, zone_idx(zone), start_pfn, 732 MEMMAP_HOTPLUG, altmap); 733 734 set_zone_contiguous(zone); 735 } 736 737 /* 738 * Returns a default kernel memory zone for the given pfn range. 739 * If no kernel zone covers this pfn range it will automatically go 740 * to the ZONE_NORMAL. 741 */ 742 static struct zone *default_kernel_zone_for_pfn(int nid, unsigned long start_pfn, 743 unsigned long nr_pages) 744 { 745 struct pglist_data *pgdat = NODE_DATA(nid); 746 int zid; 747 748 for (zid = 0; zid <= ZONE_NORMAL; zid++) { 749 struct zone *zone = &pgdat->node_zones[zid]; 750 751 if (zone_intersects(zone, start_pfn, nr_pages)) 752 return zone; 753 } 754 755 return &pgdat->node_zones[ZONE_NORMAL]; 756 } 757 758 static inline struct zone *default_zone_for_pfn(int nid, unsigned long start_pfn, 759 unsigned long nr_pages) 760 { 761 struct zone *kernel_zone = default_kernel_zone_for_pfn(nid, start_pfn, 762 nr_pages); 763 struct zone *movable_zone = &NODE_DATA(nid)->node_zones[ZONE_MOVABLE]; 764 bool in_kernel = zone_intersects(kernel_zone, start_pfn, nr_pages); 765 bool in_movable = zone_intersects(movable_zone, start_pfn, nr_pages); 766 767 /* 768 * We inherit the existing zone in a simple case where zones do not 769 * overlap in the given range 770 */ 771 if (in_kernel ^ in_movable) 772 return (in_kernel) ? kernel_zone : movable_zone; 773 774 /* 775 * If the range doesn't belong to any zone or two zones overlap in the 776 * given range then we use movable zone only if movable_node is 777 * enabled because we always online to a kernel zone by default. 778 */ 779 return movable_node_enabled ? movable_zone : kernel_zone; 780 } 781 782 struct zone * zone_for_pfn_range(int online_type, int nid, unsigned start_pfn, 783 unsigned long nr_pages) 784 { 785 if (online_type == MMOP_ONLINE_KERNEL) 786 return default_kernel_zone_for_pfn(nid, start_pfn, nr_pages); 787 788 if (online_type == MMOP_ONLINE_MOVABLE) 789 return &NODE_DATA(nid)->node_zones[ZONE_MOVABLE]; 790 791 return default_zone_for_pfn(nid, start_pfn, nr_pages); 792 } 793 794 int __ref online_pages(unsigned long pfn, unsigned long nr_pages, 795 int online_type, int nid) 796 { 797 unsigned long flags; 798 unsigned long onlined_pages = 0; 799 struct zone *zone; 800 int need_zonelists_rebuild = 0; 801 int ret; 802 struct memory_notify arg; 803 804 mem_hotplug_begin(); 805 806 /* associate pfn range with the zone */ 807 zone = zone_for_pfn_range(online_type, nid, pfn, nr_pages); 808 move_pfn_range_to_zone(zone, pfn, nr_pages, NULL); 809 810 arg.start_pfn = pfn; 811 arg.nr_pages = nr_pages; 812 node_states_check_changes_online(nr_pages, zone, &arg); 813 814 ret = memory_notify(MEM_GOING_ONLINE, &arg); 815 ret = notifier_to_errno(ret); 816 if (ret) 817 goto failed_addition; 818 819 /* 820 * If this zone is not populated, then it is not in zonelist. 821 * This means the page allocator ignores this zone. 822 * So, zonelist must be updated after online. 823 */ 824 if (!populated_zone(zone)) { 825 need_zonelists_rebuild = 1; 826 setup_zone_pageset(zone); 827 } 828 829 ret = walk_system_ram_range(pfn, nr_pages, &onlined_pages, 830 online_pages_range); 831 if (ret) { 832 /* not a single memory resource was applicable */ 833 if (need_zonelists_rebuild) 834 zone_pcp_reset(zone); 835 goto failed_addition; 836 } 837 838 zone->present_pages += onlined_pages; 839 840 pgdat_resize_lock(zone->zone_pgdat, &flags); 841 zone->zone_pgdat->node_present_pages += onlined_pages; 842 pgdat_resize_unlock(zone->zone_pgdat, &flags); 843 844 /* 845 * When exposing larger, physically contiguous memory areas to the 846 * buddy, shuffling in the buddy (when freeing onlined pages, putting 847 * them either to the head or the tail of the freelist) is only helpful 848 * for maintaining the shuffle, but not for creating the initial 849 * shuffle. Shuffle the whole zone to make sure the just onlined pages 850 * are properly distributed across the whole freelist. 851 */ 852 shuffle_zone(zone); 853 854 node_states_set_node(nid, &arg); 855 if (need_zonelists_rebuild) 856 build_all_zonelists(NULL); 857 zone_pcp_update(zone); 858 859 init_per_zone_wmark_min(); 860 861 kswapd_run(nid); 862 kcompactd_run(nid); 863 864 writeback_set_ratelimit(); 865 866 memory_notify(MEM_ONLINE, &arg); 867 mem_hotplug_done(); 868 return 0; 869 870 failed_addition: 871 pr_debug("online_pages [mem %#010llx-%#010llx] failed\n", 872 (unsigned long long) pfn << PAGE_SHIFT, 873 (((unsigned long long) pfn + nr_pages) << PAGE_SHIFT) - 1); 874 memory_notify(MEM_CANCEL_ONLINE, &arg); 875 remove_pfn_range_from_zone(zone, pfn, nr_pages); 876 mem_hotplug_done(); 877 return ret; 878 } 879 #endif /* CONFIG_MEMORY_HOTPLUG_SPARSE */ 880 881 static void reset_node_present_pages(pg_data_t *pgdat) 882 { 883 struct zone *z; 884 885 for (z = pgdat->node_zones; z < pgdat->node_zones + MAX_NR_ZONES; z++) 886 z->present_pages = 0; 887 888 pgdat->node_present_pages = 0; 889 } 890 891 /* we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG */ 892 static pg_data_t __ref *hotadd_new_pgdat(int nid) 893 { 894 struct pglist_data *pgdat; 895 896 pgdat = NODE_DATA(nid); 897 if (!pgdat) { 898 pgdat = arch_alloc_nodedata(nid); 899 if (!pgdat) 900 return NULL; 901 902 pgdat->per_cpu_nodestats = 903 alloc_percpu(struct per_cpu_nodestat); 904 arch_refresh_nodedata(nid, pgdat); 905 } else { 906 int cpu; 907 /* 908 * Reset the nr_zones, order and highest_zoneidx before reuse. 909 * Note that kswapd will init kswapd_highest_zoneidx properly 910 * when it starts in the near future. 911 */ 912 pgdat->nr_zones = 0; 913 pgdat->kswapd_order = 0; 914 pgdat->kswapd_highest_zoneidx = 0; 915 for_each_online_cpu(cpu) { 916 struct per_cpu_nodestat *p; 917 918 p = per_cpu_ptr(pgdat->per_cpu_nodestats, cpu); 919 memset(p, 0, sizeof(*p)); 920 } 921 } 922 923 /* we can use NODE_DATA(nid) from here */ 924 pgdat->node_id = nid; 925 pgdat->node_start_pfn = 0; 926 927 /* init node's zones as empty zones, we don't have any present pages.*/ 928 free_area_init_core_hotplug(nid); 929 930 /* 931 * The node we allocated has no zone fallback lists. For avoiding 932 * to access not-initialized zonelist, build here. 933 */ 934 build_all_zonelists(pgdat); 935 936 /* 937 * When memory is hot-added, all the memory is in offline state. So 938 * clear all zones' present_pages because they will be updated in 939 * online_pages() and offline_pages(). 940 */ 941 reset_node_managed_pages(pgdat); 942 reset_node_present_pages(pgdat); 943 944 return pgdat; 945 } 946 947 static void rollback_node_hotadd(int nid) 948 { 949 pg_data_t *pgdat = NODE_DATA(nid); 950 951 arch_refresh_nodedata(nid, NULL); 952 free_percpu(pgdat->per_cpu_nodestats); 953 arch_free_nodedata(pgdat); 954 } 955 956 957 /** 958 * try_online_node - online a node if offlined 959 * @nid: the node ID 960 * @set_node_online: Whether we want to online the node 961 * called by cpu_up() to online a node without onlined memory. 962 * 963 * Returns: 964 * 1 -> a new node has been allocated 965 * 0 -> the node is already online 966 * -ENOMEM -> the node could not be allocated 967 */ 968 static int __try_online_node(int nid, bool set_node_online) 969 { 970 pg_data_t *pgdat; 971 int ret = 1; 972 973 if (node_online(nid)) 974 return 0; 975 976 pgdat = hotadd_new_pgdat(nid); 977 if (!pgdat) { 978 pr_err("Cannot online node %d due to NULL pgdat\n", nid); 979 ret = -ENOMEM; 980 goto out; 981 } 982 983 if (set_node_online) { 984 node_set_online(nid); 985 ret = register_one_node(nid); 986 BUG_ON(ret); 987 } 988 out: 989 return ret; 990 } 991 992 /* 993 * Users of this function always want to online/register the node 994 */ 995 int try_online_node(int nid) 996 { 997 int ret; 998 999 mem_hotplug_begin(); 1000 ret = __try_online_node(nid, true); 1001 mem_hotplug_done(); 1002 return ret; 1003 } 1004 1005 static int check_hotplug_memory_range(u64 start, u64 size) 1006 { 1007 /* memory range must be block size aligned */ 1008 if (!size || !IS_ALIGNED(start, memory_block_size_bytes()) || 1009 !IS_ALIGNED(size, memory_block_size_bytes())) { 1010 pr_err("Block size [%#lx] unaligned hotplug range: start %#llx, size %#llx", 1011 memory_block_size_bytes(), start, size); 1012 return -EINVAL; 1013 } 1014 1015 return 0; 1016 } 1017 1018 static int online_memory_block(struct memory_block *mem, void *arg) 1019 { 1020 mem->online_type = memhp_default_online_type; 1021 return device_online(&mem->dev); 1022 } 1023 1024 /* 1025 * NOTE: The caller must call lock_device_hotplug() to serialize hotplug 1026 * and online/offline operations (triggered e.g. by sysfs). 1027 * 1028 * we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG 1029 */ 1030 int __ref add_memory_resource(int nid, struct resource *res) 1031 { 1032 struct mhp_params params = { .pgprot = PAGE_KERNEL }; 1033 u64 start, size; 1034 bool new_node = false; 1035 int ret; 1036 1037 start = res->start; 1038 size = resource_size(res); 1039 1040 ret = check_hotplug_memory_range(start, size); 1041 if (ret) 1042 return ret; 1043 1044 if (!node_possible(nid)) { 1045 WARN(1, "node %d was absent from the node_possible_map\n", nid); 1046 return -EINVAL; 1047 } 1048 1049 mem_hotplug_begin(); 1050 1051 if (IS_ENABLED(CONFIG_ARCH_KEEP_MEMBLOCK)) 1052 memblock_add_node(start, size, nid); 1053 1054 ret = __try_online_node(nid, false); 1055 if (ret < 0) 1056 goto error; 1057 new_node = ret; 1058 1059 /* call arch's memory hotadd */ 1060 ret = arch_add_memory(nid, start, size, ¶ms); 1061 if (ret < 0) 1062 goto error; 1063 1064 /* create memory block devices after memory was added */ 1065 ret = create_memory_block_devices(start, size); 1066 if (ret) { 1067 arch_remove_memory(nid, start, size, NULL); 1068 goto error; 1069 } 1070 1071 if (new_node) { 1072 /* If sysfs file of new node can't be created, cpu on the node 1073 * can't be hot-added. There is no rollback way now. 1074 * So, check by BUG_ON() to catch it reluctantly.. 1075 * We online node here. We can't roll back from here. 1076 */ 1077 node_set_online(nid); 1078 ret = __register_one_node(nid); 1079 BUG_ON(ret); 1080 } 1081 1082 /* link memory sections under this node.*/ 1083 ret = link_mem_sections(nid, PFN_DOWN(start), PFN_UP(start + size - 1)); 1084 BUG_ON(ret); 1085 1086 /* create new memmap entry */ 1087 if (!strcmp(res->name, "System RAM")) 1088 firmware_map_add_hotplug(start, start + size, "System RAM"); 1089 1090 /* device_online() will take the lock when calling online_pages() */ 1091 mem_hotplug_done(); 1092 1093 /* online pages if requested */ 1094 if (memhp_default_online_type != MMOP_OFFLINE) 1095 walk_memory_blocks(start, size, NULL, online_memory_block); 1096 1097 return ret; 1098 error: 1099 /* rollback pgdat allocation and others */ 1100 if (new_node) 1101 rollback_node_hotadd(nid); 1102 if (IS_ENABLED(CONFIG_ARCH_KEEP_MEMBLOCK)) 1103 memblock_remove(start, size); 1104 mem_hotplug_done(); 1105 return ret; 1106 } 1107 1108 /* requires device_hotplug_lock, see add_memory_resource() */ 1109 int __ref __add_memory(int nid, u64 start, u64 size) 1110 { 1111 struct resource *res; 1112 int ret; 1113 1114 res = register_memory_resource(start, size, "System RAM"); 1115 if (IS_ERR(res)) 1116 return PTR_ERR(res); 1117 1118 ret = add_memory_resource(nid, res); 1119 if (ret < 0) 1120 release_memory_resource(res); 1121 return ret; 1122 } 1123 1124 int add_memory(int nid, u64 start, u64 size) 1125 { 1126 int rc; 1127 1128 lock_device_hotplug(); 1129 rc = __add_memory(nid, start, size); 1130 unlock_device_hotplug(); 1131 1132 return rc; 1133 } 1134 EXPORT_SYMBOL_GPL(add_memory); 1135 1136 /* 1137 * Add special, driver-managed memory to the system as system RAM. Such 1138 * memory is not exposed via the raw firmware-provided memmap as system 1139 * RAM, instead, it is detected and added by a driver - during cold boot, 1140 * after a reboot, and after kexec. 1141 * 1142 * Reasons why this memory should not be used for the initial memmap of a 1143 * kexec kernel or for placing kexec images: 1144 * - The booting kernel is in charge of determining how this memory will be 1145 * used (e.g., use persistent memory as system RAM) 1146 * - Coordination with a hypervisor is required before this memory 1147 * can be used (e.g., inaccessible parts). 1148 * 1149 * For this memory, no entries in /sys/firmware/memmap ("raw firmware-provided 1150 * memory map") are created. Also, the created memory resource is flagged 1151 * with IORESOURCE_MEM_DRIVER_MANAGED, so in-kernel users can special-case 1152 * this memory as well (esp., not place kexec images onto it). 1153 * 1154 * The resource_name (visible via /proc/iomem) has to have the format 1155 * "System RAM ($DRIVER)". 1156 */ 1157 int add_memory_driver_managed(int nid, u64 start, u64 size, 1158 const char *resource_name) 1159 { 1160 struct resource *res; 1161 int rc; 1162 1163 if (!resource_name || 1164 strstr(resource_name, "System RAM (") != resource_name || 1165 resource_name[strlen(resource_name) - 1] != ')') 1166 return -EINVAL; 1167 1168 lock_device_hotplug(); 1169 1170 res = register_memory_resource(start, size, resource_name); 1171 if (IS_ERR(res)) { 1172 rc = PTR_ERR(res); 1173 goto out_unlock; 1174 } 1175 1176 rc = add_memory_resource(nid, res); 1177 if (rc < 0) 1178 release_memory_resource(res); 1179 1180 out_unlock: 1181 unlock_device_hotplug(); 1182 return rc; 1183 } 1184 EXPORT_SYMBOL_GPL(add_memory_driver_managed); 1185 1186 #ifdef CONFIG_MEMORY_HOTREMOVE 1187 /* 1188 * Confirm all pages in a range [start, end) belong to the same zone (skipping 1189 * memory holes). When true, return the zone. 1190 */ 1191 struct zone *test_pages_in_a_zone(unsigned long start_pfn, 1192 unsigned long end_pfn) 1193 { 1194 unsigned long pfn, sec_end_pfn; 1195 struct zone *zone = NULL; 1196 struct page *page; 1197 int i; 1198 for (pfn = start_pfn, sec_end_pfn = SECTION_ALIGN_UP(start_pfn + 1); 1199 pfn < end_pfn; 1200 pfn = sec_end_pfn, sec_end_pfn += PAGES_PER_SECTION) { 1201 /* Make sure the memory section is present first */ 1202 if (!present_section_nr(pfn_to_section_nr(pfn))) 1203 continue; 1204 for (; pfn < sec_end_pfn && pfn < end_pfn; 1205 pfn += MAX_ORDER_NR_PAGES) { 1206 i = 0; 1207 /* This is just a CONFIG_HOLES_IN_ZONE check.*/ 1208 while ((i < MAX_ORDER_NR_PAGES) && 1209 !pfn_valid_within(pfn + i)) 1210 i++; 1211 if (i == MAX_ORDER_NR_PAGES || pfn + i >= end_pfn) 1212 continue; 1213 /* Check if we got outside of the zone */ 1214 if (zone && !zone_spans_pfn(zone, pfn + i)) 1215 return NULL; 1216 page = pfn_to_page(pfn + i); 1217 if (zone && page_zone(page) != zone) 1218 return NULL; 1219 zone = page_zone(page); 1220 } 1221 } 1222 1223 return zone; 1224 } 1225 1226 /* 1227 * Scan pfn range [start,end) to find movable/migratable pages (LRU pages, 1228 * non-lru movable pages and hugepages). Will skip over most unmovable 1229 * pages (esp., pages that can be skipped when offlining), but bail out on 1230 * definitely unmovable pages. 1231 * 1232 * Returns: 1233 * 0 in case a movable page is found and movable_pfn was updated. 1234 * -ENOENT in case no movable page was found. 1235 * -EBUSY in case a definitely unmovable page was found. 1236 */ 1237 static int scan_movable_pages(unsigned long start, unsigned long end, 1238 unsigned long *movable_pfn) 1239 { 1240 unsigned long pfn; 1241 1242 for (pfn = start; pfn < end; pfn++) { 1243 struct page *page, *head; 1244 unsigned long skip; 1245 1246 if (!pfn_valid(pfn)) 1247 continue; 1248 page = pfn_to_page(pfn); 1249 if (PageLRU(page)) 1250 goto found; 1251 if (__PageMovable(page)) 1252 goto found; 1253 1254 /* 1255 * PageOffline() pages that are not marked __PageMovable() and 1256 * have a reference count > 0 (after MEM_GOING_OFFLINE) are 1257 * definitely unmovable. If their reference count would be 0, 1258 * they could at least be skipped when offlining memory. 1259 */ 1260 if (PageOffline(page) && page_count(page)) 1261 return -EBUSY; 1262 1263 if (!PageHuge(page)) 1264 continue; 1265 head = compound_head(page); 1266 if (page_huge_active(head)) 1267 goto found; 1268 skip = compound_nr(head) - (page - head); 1269 pfn += skip - 1; 1270 } 1271 return -ENOENT; 1272 found: 1273 *movable_pfn = pfn; 1274 return 0; 1275 } 1276 1277 static struct page *new_node_page(struct page *page, unsigned long private) 1278 { 1279 nodemask_t nmask = node_states[N_MEMORY]; 1280 struct migration_target_control mtc = { 1281 .nid = page_to_nid(page), 1282 .nmask = &nmask, 1283 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL, 1284 }; 1285 1286 /* 1287 * try to allocate from a different node but reuse this node if there 1288 * are no other online nodes to be used (e.g. we are offlining a part 1289 * of the only existing node) 1290 */ 1291 node_clear(mtc.nid, nmask); 1292 if (nodes_empty(nmask)) 1293 node_set(mtc.nid, nmask); 1294 1295 return alloc_migration_target(page, (unsigned long)&mtc); 1296 } 1297 1298 static int 1299 do_migrate_range(unsigned long start_pfn, unsigned long end_pfn) 1300 { 1301 unsigned long pfn; 1302 struct page *page, *head; 1303 int ret = 0; 1304 LIST_HEAD(source); 1305 1306 for (pfn = start_pfn; pfn < end_pfn; pfn++) { 1307 if (!pfn_valid(pfn)) 1308 continue; 1309 page = pfn_to_page(pfn); 1310 head = compound_head(page); 1311 1312 if (PageHuge(page)) { 1313 pfn = page_to_pfn(head) + compound_nr(head) - 1; 1314 isolate_huge_page(head, &source); 1315 continue; 1316 } else if (PageTransHuge(page)) 1317 pfn = page_to_pfn(head) + thp_nr_pages(page) - 1; 1318 1319 /* 1320 * HWPoison pages have elevated reference counts so the migration would 1321 * fail on them. It also doesn't make any sense to migrate them in the 1322 * first place. Still try to unmap such a page in case it is still mapped 1323 * (e.g. current hwpoison implementation doesn't unmap KSM pages but keep 1324 * the unmap as the catch all safety net). 1325 */ 1326 if (PageHWPoison(page)) { 1327 if (WARN_ON(PageLRU(page))) 1328 isolate_lru_page(page); 1329 if (page_mapped(page)) 1330 try_to_unmap(page, TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS); 1331 continue; 1332 } 1333 1334 if (!get_page_unless_zero(page)) 1335 continue; 1336 /* 1337 * We can skip free pages. And we can deal with pages on 1338 * LRU and non-lru movable pages. 1339 */ 1340 if (PageLRU(page)) 1341 ret = isolate_lru_page(page); 1342 else 1343 ret = isolate_movable_page(page, ISOLATE_UNEVICTABLE); 1344 if (!ret) { /* Success */ 1345 list_add_tail(&page->lru, &source); 1346 if (!__PageMovable(page)) 1347 inc_node_page_state(page, NR_ISOLATED_ANON + 1348 page_is_file_lru(page)); 1349 1350 } else { 1351 pr_warn("failed to isolate pfn %lx\n", pfn); 1352 dump_page(page, "isolation failed"); 1353 } 1354 put_page(page); 1355 } 1356 if (!list_empty(&source)) { 1357 /* Allocate a new page from the nearest neighbor node */ 1358 ret = migrate_pages(&source, new_node_page, NULL, 0, 1359 MIGRATE_SYNC, MR_MEMORY_HOTPLUG); 1360 if (ret) { 1361 list_for_each_entry(page, &source, lru) { 1362 pr_warn("migrating pfn %lx failed ret:%d ", 1363 page_to_pfn(page), ret); 1364 dump_page(page, "migration failure"); 1365 } 1366 putback_movable_pages(&source); 1367 } 1368 } 1369 1370 return ret; 1371 } 1372 1373 /* Mark all sections offline and remove all free pages from the buddy. */ 1374 static int 1375 offline_isolated_pages_cb(unsigned long start, unsigned long nr_pages, 1376 void *data) 1377 { 1378 unsigned long *offlined_pages = (unsigned long *)data; 1379 1380 *offlined_pages += __offline_isolated_pages(start, start + nr_pages); 1381 return 0; 1382 } 1383 1384 /* 1385 * Check all pages in range, recorded as memory resource, are isolated. 1386 */ 1387 static int 1388 check_pages_isolated_cb(unsigned long start_pfn, unsigned long nr_pages, 1389 void *data) 1390 { 1391 return test_pages_isolated(start_pfn, start_pfn + nr_pages, 1392 MEMORY_OFFLINE); 1393 } 1394 1395 static int __init cmdline_parse_movable_node(char *p) 1396 { 1397 movable_node_enabled = true; 1398 return 0; 1399 } 1400 early_param("movable_node", cmdline_parse_movable_node); 1401 1402 /* check which state of node_states will be changed when offline memory */ 1403 static void node_states_check_changes_offline(unsigned long nr_pages, 1404 struct zone *zone, struct memory_notify *arg) 1405 { 1406 struct pglist_data *pgdat = zone->zone_pgdat; 1407 unsigned long present_pages = 0; 1408 enum zone_type zt; 1409 1410 arg->status_change_nid = NUMA_NO_NODE; 1411 arg->status_change_nid_normal = NUMA_NO_NODE; 1412 arg->status_change_nid_high = NUMA_NO_NODE; 1413 1414 /* 1415 * Check whether node_states[N_NORMAL_MEMORY] will be changed. 1416 * If the memory to be offline is within the range 1417 * [0..ZONE_NORMAL], and it is the last present memory there, 1418 * the zones in that range will become empty after the offlining, 1419 * thus we can determine that we need to clear the node from 1420 * node_states[N_NORMAL_MEMORY]. 1421 */ 1422 for (zt = 0; zt <= ZONE_NORMAL; zt++) 1423 present_pages += pgdat->node_zones[zt].present_pages; 1424 if (zone_idx(zone) <= ZONE_NORMAL && nr_pages >= present_pages) 1425 arg->status_change_nid_normal = zone_to_nid(zone); 1426 1427 #ifdef CONFIG_HIGHMEM 1428 /* 1429 * node_states[N_HIGH_MEMORY] contains nodes which 1430 * have normal memory or high memory. 1431 * Here we add the present_pages belonging to ZONE_HIGHMEM. 1432 * If the zone is within the range of [0..ZONE_HIGHMEM), and 1433 * we determine that the zones in that range become empty, 1434 * we need to clear the node for N_HIGH_MEMORY. 1435 */ 1436 present_pages += pgdat->node_zones[ZONE_HIGHMEM].present_pages; 1437 if (zone_idx(zone) <= ZONE_HIGHMEM && nr_pages >= present_pages) 1438 arg->status_change_nid_high = zone_to_nid(zone); 1439 #endif 1440 1441 /* 1442 * We have accounted the pages from [0..ZONE_NORMAL), and 1443 * in case of CONFIG_HIGHMEM the pages from ZONE_HIGHMEM 1444 * as well. 1445 * Here we count the possible pages from ZONE_MOVABLE. 1446 * If after having accounted all the pages, we see that the nr_pages 1447 * to be offlined is over or equal to the accounted pages, 1448 * we know that the node will become empty, and so, we can clear 1449 * it for N_MEMORY as well. 1450 */ 1451 present_pages += pgdat->node_zones[ZONE_MOVABLE].present_pages; 1452 1453 if (nr_pages >= present_pages) 1454 arg->status_change_nid = zone_to_nid(zone); 1455 } 1456 1457 static void node_states_clear_node(int node, struct memory_notify *arg) 1458 { 1459 if (arg->status_change_nid_normal >= 0) 1460 node_clear_state(node, N_NORMAL_MEMORY); 1461 1462 if (arg->status_change_nid_high >= 0) 1463 node_clear_state(node, N_HIGH_MEMORY); 1464 1465 if (arg->status_change_nid >= 0) 1466 node_clear_state(node, N_MEMORY); 1467 } 1468 1469 static int count_system_ram_pages_cb(unsigned long start_pfn, 1470 unsigned long nr_pages, void *data) 1471 { 1472 unsigned long *nr_system_ram_pages = data; 1473 1474 *nr_system_ram_pages += nr_pages; 1475 return 0; 1476 } 1477 1478 static int __ref __offline_pages(unsigned long start_pfn, 1479 unsigned long end_pfn) 1480 { 1481 unsigned long pfn, nr_pages = 0; 1482 unsigned long offlined_pages = 0; 1483 int ret, node, nr_isolate_pageblock; 1484 unsigned long flags; 1485 struct zone *zone; 1486 struct memory_notify arg; 1487 char *reason; 1488 1489 mem_hotplug_begin(); 1490 1491 /* 1492 * Don't allow to offline memory blocks that contain holes. 1493 * Consequently, memory blocks with holes can never get onlined 1494 * via the hotplug path - online_pages() - as hotplugged memory has 1495 * no holes. This way, we e.g., don't have to worry about marking 1496 * memory holes PG_reserved, don't need pfn_valid() checks, and can 1497 * avoid using walk_system_ram_range() later. 1498 */ 1499 walk_system_ram_range(start_pfn, end_pfn - start_pfn, &nr_pages, 1500 count_system_ram_pages_cb); 1501 if (nr_pages != end_pfn - start_pfn) { 1502 ret = -EINVAL; 1503 reason = "memory holes"; 1504 goto failed_removal; 1505 } 1506 1507 /* This makes hotplug much easier...and readable. 1508 we assume this for now. .*/ 1509 zone = test_pages_in_a_zone(start_pfn, end_pfn); 1510 if (!zone) { 1511 ret = -EINVAL; 1512 reason = "multizone range"; 1513 goto failed_removal; 1514 } 1515 node = zone_to_nid(zone); 1516 1517 /* set above range as isolated */ 1518 ret = start_isolate_page_range(start_pfn, end_pfn, 1519 MIGRATE_MOVABLE, 1520 MEMORY_OFFLINE | REPORT_FAILURE); 1521 if (ret < 0) { 1522 reason = "failure to isolate range"; 1523 goto failed_removal; 1524 } 1525 nr_isolate_pageblock = ret; 1526 1527 arg.start_pfn = start_pfn; 1528 arg.nr_pages = nr_pages; 1529 node_states_check_changes_offline(nr_pages, zone, &arg); 1530 1531 ret = memory_notify(MEM_GOING_OFFLINE, &arg); 1532 ret = notifier_to_errno(ret); 1533 if (ret) { 1534 reason = "notifier failure"; 1535 goto failed_removal_isolated; 1536 } 1537 1538 do { 1539 pfn = start_pfn; 1540 do { 1541 if (signal_pending(current)) { 1542 ret = -EINTR; 1543 reason = "signal backoff"; 1544 goto failed_removal_isolated; 1545 } 1546 1547 cond_resched(); 1548 lru_add_drain_all(); 1549 1550 ret = scan_movable_pages(pfn, end_pfn, &pfn); 1551 if (!ret) { 1552 /* 1553 * TODO: fatal migration failures should bail 1554 * out 1555 */ 1556 do_migrate_range(pfn, end_pfn); 1557 } 1558 } while (!ret); 1559 1560 if (ret != -ENOENT) { 1561 reason = "unmovable page"; 1562 goto failed_removal_isolated; 1563 } 1564 1565 /* 1566 * Dissolve free hugepages in the memory block before doing 1567 * offlining actually in order to make hugetlbfs's object 1568 * counting consistent. 1569 */ 1570 ret = dissolve_free_huge_pages(start_pfn, end_pfn); 1571 if (ret) { 1572 reason = "failure to dissolve huge pages"; 1573 goto failed_removal_isolated; 1574 } 1575 /* check again */ 1576 ret = walk_system_ram_range(start_pfn, end_pfn - start_pfn, 1577 NULL, check_pages_isolated_cb); 1578 /* 1579 * per-cpu pages are drained in start_isolate_page_range, but if 1580 * there are still pages that are not free, make sure that we 1581 * drain again, because when we isolated range we might 1582 * have raced with another thread that was adding pages to pcp 1583 * list. 1584 * 1585 * Forward progress should be still guaranteed because 1586 * pages on the pcp list can only belong to MOVABLE_ZONE 1587 * because has_unmovable_pages explicitly checks for 1588 * PageBuddy on freed pages on other zones. 1589 */ 1590 if (ret) 1591 drain_all_pages(zone); 1592 } while (ret); 1593 1594 /* Ok, all of our target is isolated. 1595 We cannot do rollback at this point. */ 1596 walk_system_ram_range(start_pfn, end_pfn - start_pfn, 1597 &offlined_pages, offline_isolated_pages_cb); 1598 pr_info("Offlined Pages %ld\n", offlined_pages); 1599 /* 1600 * Onlining will reset pagetype flags and makes migrate type 1601 * MOVABLE, so just need to decrease the number of isolated 1602 * pageblocks zone counter here. 1603 */ 1604 spin_lock_irqsave(&zone->lock, flags); 1605 zone->nr_isolate_pageblock -= nr_isolate_pageblock; 1606 spin_unlock_irqrestore(&zone->lock, flags); 1607 1608 /* removal success */ 1609 adjust_managed_page_count(pfn_to_page(start_pfn), -offlined_pages); 1610 zone->present_pages -= offlined_pages; 1611 1612 pgdat_resize_lock(zone->zone_pgdat, &flags); 1613 zone->zone_pgdat->node_present_pages -= offlined_pages; 1614 pgdat_resize_unlock(zone->zone_pgdat, &flags); 1615 1616 init_per_zone_wmark_min(); 1617 1618 if (!populated_zone(zone)) { 1619 zone_pcp_reset(zone); 1620 build_all_zonelists(NULL); 1621 } else 1622 zone_pcp_update(zone); 1623 1624 node_states_clear_node(node, &arg); 1625 if (arg.status_change_nid >= 0) { 1626 kswapd_stop(node); 1627 kcompactd_stop(node); 1628 } 1629 1630 writeback_set_ratelimit(); 1631 1632 memory_notify(MEM_OFFLINE, &arg); 1633 remove_pfn_range_from_zone(zone, start_pfn, nr_pages); 1634 mem_hotplug_done(); 1635 return 0; 1636 1637 failed_removal_isolated: 1638 undo_isolate_page_range(start_pfn, end_pfn, MIGRATE_MOVABLE); 1639 memory_notify(MEM_CANCEL_OFFLINE, &arg); 1640 failed_removal: 1641 pr_debug("memory offlining [mem %#010llx-%#010llx] failed due to %s\n", 1642 (unsigned long long) start_pfn << PAGE_SHIFT, 1643 ((unsigned long long) end_pfn << PAGE_SHIFT) - 1, 1644 reason); 1645 /* pushback to free area */ 1646 mem_hotplug_done(); 1647 return ret; 1648 } 1649 1650 int offline_pages(unsigned long start_pfn, unsigned long nr_pages) 1651 { 1652 return __offline_pages(start_pfn, start_pfn + nr_pages); 1653 } 1654 1655 static int check_memblock_offlined_cb(struct memory_block *mem, void *arg) 1656 { 1657 int ret = !is_memblock_offlined(mem); 1658 1659 if (unlikely(ret)) { 1660 phys_addr_t beginpa, endpa; 1661 1662 beginpa = PFN_PHYS(section_nr_to_pfn(mem->start_section_nr)); 1663 endpa = beginpa + memory_block_size_bytes() - 1; 1664 pr_warn("removing memory fails, because memory [%pa-%pa] is onlined\n", 1665 &beginpa, &endpa); 1666 1667 return -EBUSY; 1668 } 1669 return 0; 1670 } 1671 1672 static int check_cpu_on_node(pg_data_t *pgdat) 1673 { 1674 int cpu; 1675 1676 for_each_present_cpu(cpu) { 1677 if (cpu_to_node(cpu) == pgdat->node_id) 1678 /* 1679 * the cpu on this node isn't removed, and we can't 1680 * offline this node. 1681 */ 1682 return -EBUSY; 1683 } 1684 1685 return 0; 1686 } 1687 1688 static int check_no_memblock_for_node_cb(struct memory_block *mem, void *arg) 1689 { 1690 int nid = *(int *)arg; 1691 1692 /* 1693 * If a memory block belongs to multiple nodes, the stored nid is not 1694 * reliable. However, such blocks are always online (e.g., cannot get 1695 * offlined) and, therefore, are still spanned by the node. 1696 */ 1697 return mem->nid == nid ? -EEXIST : 0; 1698 } 1699 1700 /** 1701 * try_offline_node 1702 * @nid: the node ID 1703 * 1704 * Offline a node if all memory sections and cpus of the node are removed. 1705 * 1706 * NOTE: The caller must call lock_device_hotplug() to serialize hotplug 1707 * and online/offline operations before this call. 1708 */ 1709 void try_offline_node(int nid) 1710 { 1711 pg_data_t *pgdat = NODE_DATA(nid); 1712 int rc; 1713 1714 /* 1715 * If the node still spans pages (especially ZONE_DEVICE), don't 1716 * offline it. A node spans memory after move_pfn_range_to_zone(), 1717 * e.g., after the memory block was onlined. 1718 */ 1719 if (pgdat->node_spanned_pages) 1720 return; 1721 1722 /* 1723 * Especially offline memory blocks might not be spanned by the 1724 * node. They will get spanned by the node once they get onlined. 1725 * However, they link to the node in sysfs and can get onlined later. 1726 */ 1727 rc = for_each_memory_block(&nid, check_no_memblock_for_node_cb); 1728 if (rc) 1729 return; 1730 1731 if (check_cpu_on_node(pgdat)) 1732 return; 1733 1734 /* 1735 * all memory/cpu of this node are removed, we can offline this 1736 * node now. 1737 */ 1738 node_set_offline(nid); 1739 unregister_one_node(nid); 1740 } 1741 EXPORT_SYMBOL(try_offline_node); 1742 1743 static void __release_memory_resource(resource_size_t start, 1744 resource_size_t size) 1745 { 1746 int ret; 1747 1748 /* 1749 * When removing memory in the same granularity as it was added, 1750 * this function never fails. It might only fail if resources 1751 * have to be adjusted or split. We'll ignore the error, as 1752 * removing of memory cannot fail. 1753 */ 1754 ret = release_mem_region_adjustable(&iomem_resource, start, size); 1755 if (ret) { 1756 resource_size_t endres = start + size - 1; 1757 1758 pr_warn("Unable to release resource <%pa-%pa> (%d)\n", 1759 &start, &endres, ret); 1760 } 1761 } 1762 1763 static int __ref try_remove_memory(int nid, u64 start, u64 size) 1764 { 1765 int rc = 0; 1766 1767 BUG_ON(check_hotplug_memory_range(start, size)); 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 return rc; 1777 1778 /* remove memmap entry */ 1779 firmware_map_remove(start, start + size, "System RAM"); 1780 1781 /* 1782 * Memory block device removal under the device_hotplug_lock is 1783 * a barrier against racing online attempts. 1784 */ 1785 remove_memory_block_devices(start, size); 1786 1787 mem_hotplug_begin(); 1788 1789 arch_remove_memory(nid, start, size, NULL); 1790 1791 if (IS_ENABLED(CONFIG_ARCH_KEEP_MEMBLOCK)) { 1792 memblock_free(start, size); 1793 memblock_remove(start, size); 1794 } 1795 1796 __release_memory_resource(start, size); 1797 1798 try_offline_node(nid); 1799 1800 mem_hotplug_done(); 1801 return 0; 1802 } 1803 1804 /** 1805 * remove_memory 1806 * @nid: the node ID 1807 * @start: physical address of the region to remove 1808 * @size: size of the region to remove 1809 * 1810 * NOTE: The caller must call lock_device_hotplug() to serialize hotplug 1811 * and online/offline operations before this call, as required by 1812 * try_offline_node(). 1813 */ 1814 void __remove_memory(int nid, u64 start, u64 size) 1815 { 1816 1817 /* 1818 * trigger BUG() if some memory is not offlined prior to calling this 1819 * function 1820 */ 1821 if (try_remove_memory(nid, start, size)) 1822 BUG(); 1823 } 1824 1825 /* 1826 * Remove memory if every memory block is offline, otherwise return -EBUSY is 1827 * some memory is not offline 1828 */ 1829 int remove_memory(int nid, u64 start, u64 size) 1830 { 1831 int rc; 1832 1833 lock_device_hotplug(); 1834 rc = try_remove_memory(nid, start, size); 1835 unlock_device_hotplug(); 1836 1837 return rc; 1838 } 1839 EXPORT_SYMBOL_GPL(remove_memory); 1840 1841 /* 1842 * Try to offline and remove a memory block. Might take a long time to 1843 * finish in case memory is still in use. Primarily useful for memory devices 1844 * that logically unplugged all memory (so it's no longer in use) and want to 1845 * offline + remove the memory block. 1846 */ 1847 int offline_and_remove_memory(int nid, u64 start, u64 size) 1848 { 1849 struct memory_block *mem; 1850 int rc = -EINVAL; 1851 1852 if (!IS_ALIGNED(start, memory_block_size_bytes()) || 1853 size != memory_block_size_bytes()) 1854 return rc; 1855 1856 lock_device_hotplug(); 1857 mem = find_memory_block(__pfn_to_section(PFN_DOWN(start))); 1858 if (mem) 1859 rc = device_offline(&mem->dev); 1860 /* Ignore if the device is already offline. */ 1861 if (rc > 0) 1862 rc = 0; 1863 1864 /* 1865 * In case we succeeded to offline the memory block, remove it. 1866 * This cannot fail as it cannot get onlined in the meantime. 1867 */ 1868 if (!rc) { 1869 rc = try_remove_memory(nid, start, size); 1870 WARN_ON_ONCE(rc); 1871 } 1872 unlock_device_hotplug(); 1873 1874 return rc; 1875 } 1876 EXPORT_SYMBOL_GPL(offline_and_remove_memory); 1877 #endif /* CONFIG_MEMORY_HOTREMOVE */ 1878