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 /* find the smallest valid pfn in the range [start_pfn, end_pfn) */ 354 static unsigned long find_smallest_section_pfn(int nid, struct zone *zone, 355 unsigned long start_pfn, 356 unsigned long end_pfn) 357 { 358 for (; start_pfn < end_pfn; start_pfn += PAGES_PER_SUBSECTION) { 359 if (unlikely(!pfn_to_online_page(start_pfn))) 360 continue; 361 362 if (unlikely(pfn_to_nid(start_pfn) != nid)) 363 continue; 364 365 if (zone != page_zone(pfn_to_page(start_pfn))) 366 continue; 367 368 return start_pfn; 369 } 370 371 return 0; 372 } 373 374 /* find the biggest valid pfn in the range [start_pfn, end_pfn). */ 375 static unsigned long find_biggest_section_pfn(int nid, struct zone *zone, 376 unsigned long start_pfn, 377 unsigned long end_pfn) 378 { 379 unsigned long pfn; 380 381 /* pfn is the end pfn of a memory section. */ 382 pfn = end_pfn - 1; 383 for (; pfn >= start_pfn; pfn -= PAGES_PER_SUBSECTION) { 384 if (unlikely(!pfn_to_online_page(pfn))) 385 continue; 386 387 if (unlikely(pfn_to_nid(pfn) != nid)) 388 continue; 389 390 if (zone != page_zone(pfn_to_page(pfn))) 391 continue; 392 393 return pfn; 394 } 395 396 return 0; 397 } 398 399 static void shrink_zone_span(struct zone *zone, unsigned long start_pfn, 400 unsigned long end_pfn) 401 { 402 unsigned long pfn; 403 int nid = zone_to_nid(zone); 404 405 zone_span_writelock(zone); 406 if (zone->zone_start_pfn == start_pfn) { 407 /* 408 * If the section is smallest section in the zone, it need 409 * shrink zone->zone_start_pfn and zone->zone_spanned_pages. 410 * In this case, we find second smallest valid mem_section 411 * for shrinking zone. 412 */ 413 pfn = find_smallest_section_pfn(nid, zone, end_pfn, 414 zone_end_pfn(zone)); 415 if (pfn) { 416 zone->spanned_pages = zone_end_pfn(zone) - pfn; 417 zone->zone_start_pfn = pfn; 418 } else { 419 zone->zone_start_pfn = 0; 420 zone->spanned_pages = 0; 421 } 422 } else if (zone_end_pfn(zone) == end_pfn) { 423 /* 424 * If the section is biggest section in the zone, it need 425 * shrink zone->spanned_pages. 426 * In this case, we find second biggest valid mem_section for 427 * shrinking zone. 428 */ 429 pfn = find_biggest_section_pfn(nid, zone, zone->zone_start_pfn, 430 start_pfn); 431 if (pfn) 432 zone->spanned_pages = pfn - zone->zone_start_pfn + 1; 433 else { 434 zone->zone_start_pfn = 0; 435 zone->spanned_pages = 0; 436 } 437 } 438 zone_span_writeunlock(zone); 439 } 440 441 static void update_pgdat_span(struct pglist_data *pgdat) 442 { 443 unsigned long node_start_pfn = 0, node_end_pfn = 0; 444 struct zone *zone; 445 446 for (zone = pgdat->node_zones; 447 zone < pgdat->node_zones + MAX_NR_ZONES; zone++) { 448 unsigned long zone_end_pfn = zone->zone_start_pfn + 449 zone->spanned_pages; 450 451 /* No need to lock the zones, they can't change. */ 452 if (!zone->spanned_pages) 453 continue; 454 if (!node_end_pfn) { 455 node_start_pfn = zone->zone_start_pfn; 456 node_end_pfn = zone_end_pfn; 457 continue; 458 } 459 460 if (zone_end_pfn > node_end_pfn) 461 node_end_pfn = zone_end_pfn; 462 if (zone->zone_start_pfn < node_start_pfn) 463 node_start_pfn = zone->zone_start_pfn; 464 } 465 466 pgdat->node_start_pfn = node_start_pfn; 467 pgdat->node_spanned_pages = node_end_pfn - node_start_pfn; 468 } 469 470 void __ref remove_pfn_range_from_zone(struct zone *zone, 471 unsigned long start_pfn, 472 unsigned long nr_pages) 473 { 474 const unsigned long end_pfn = start_pfn + nr_pages; 475 struct pglist_data *pgdat = zone->zone_pgdat; 476 unsigned long pfn, cur_nr_pages, flags; 477 478 /* Poison struct pages because they are now uninitialized again. */ 479 for (pfn = start_pfn; pfn < end_pfn; pfn += cur_nr_pages) { 480 cond_resched(); 481 482 /* Select all remaining pages up to the next section boundary */ 483 cur_nr_pages = 484 min(end_pfn - pfn, SECTION_ALIGN_UP(pfn + 1) - pfn); 485 page_init_poison(pfn_to_page(pfn), 486 sizeof(struct page) * cur_nr_pages); 487 } 488 489 #ifdef CONFIG_ZONE_DEVICE 490 /* 491 * Zone shrinking code cannot properly deal with ZONE_DEVICE. So 492 * we will not try to shrink the zones - which is okay as 493 * set_zone_contiguous() cannot deal with ZONE_DEVICE either way. 494 */ 495 if (zone_idx(zone) == ZONE_DEVICE) 496 return; 497 #endif 498 499 clear_zone_contiguous(zone); 500 501 pgdat_resize_lock(zone->zone_pgdat, &flags); 502 shrink_zone_span(zone, start_pfn, start_pfn + nr_pages); 503 update_pgdat_span(pgdat); 504 pgdat_resize_unlock(zone->zone_pgdat, &flags); 505 506 set_zone_contiguous(zone); 507 } 508 509 static void __remove_section(unsigned long pfn, unsigned long nr_pages, 510 unsigned long map_offset, 511 struct vmem_altmap *altmap) 512 { 513 struct mem_section *ms = __pfn_to_section(pfn); 514 515 if (WARN_ON_ONCE(!valid_section(ms))) 516 return; 517 518 sparse_remove_section(ms, pfn, nr_pages, map_offset, altmap); 519 } 520 521 /** 522 * __remove_pages() - remove sections of pages 523 * @pfn: starting pageframe (must be aligned to start of a section) 524 * @nr_pages: number of pages to remove (must be multiple of section size) 525 * @altmap: alternative device page map or %NULL if default memmap is used 526 * 527 * Generic helper function to remove section mappings and sysfs entries 528 * for the section of the memory we are removing. Caller needs to make 529 * sure that pages are marked reserved and zones are adjust properly by 530 * calling offline_pages(). 531 */ 532 void __remove_pages(unsigned long pfn, unsigned long nr_pages, 533 struct vmem_altmap *altmap) 534 { 535 const unsigned long end_pfn = pfn + nr_pages; 536 unsigned long cur_nr_pages; 537 unsigned long map_offset = 0; 538 539 map_offset = vmem_altmap_offset(altmap); 540 541 if (check_pfn_span(pfn, nr_pages, "remove")) 542 return; 543 544 for (; pfn < end_pfn; pfn += cur_nr_pages) { 545 cond_resched(); 546 /* Select all remaining pages up to the next section boundary */ 547 cur_nr_pages = min(end_pfn - pfn, 548 SECTION_ALIGN_UP(pfn + 1) - pfn); 549 __remove_section(pfn, cur_nr_pages, map_offset, altmap); 550 map_offset = 0; 551 } 552 } 553 554 int set_online_page_callback(online_page_callback_t callback) 555 { 556 int rc = -EINVAL; 557 558 get_online_mems(); 559 mutex_lock(&online_page_callback_lock); 560 561 if (online_page_callback == generic_online_page) { 562 online_page_callback = callback; 563 rc = 0; 564 } 565 566 mutex_unlock(&online_page_callback_lock); 567 put_online_mems(); 568 569 return rc; 570 } 571 EXPORT_SYMBOL_GPL(set_online_page_callback); 572 573 int restore_online_page_callback(online_page_callback_t callback) 574 { 575 int rc = -EINVAL; 576 577 get_online_mems(); 578 mutex_lock(&online_page_callback_lock); 579 580 if (online_page_callback == callback) { 581 online_page_callback = generic_online_page; 582 rc = 0; 583 } 584 585 mutex_unlock(&online_page_callback_lock); 586 put_online_mems(); 587 588 return rc; 589 } 590 EXPORT_SYMBOL_GPL(restore_online_page_callback); 591 592 void generic_online_page(struct page *page, unsigned int order) 593 { 594 /* 595 * Freeing the page with debug_pagealloc enabled will try to unmap it, 596 * so we should map it first. This is better than introducing a special 597 * case in page freeing fast path. 598 */ 599 if (debug_pagealloc_enabled_static()) 600 kernel_map_pages(page, 1 << order, 1); 601 __free_pages_core(page, order); 602 totalram_pages_add(1UL << order); 603 #ifdef CONFIG_HIGHMEM 604 if (PageHighMem(page)) 605 totalhigh_pages_add(1UL << order); 606 #endif 607 } 608 EXPORT_SYMBOL_GPL(generic_online_page); 609 610 static int online_pages_range(unsigned long start_pfn, unsigned long nr_pages, 611 void *arg) 612 { 613 const unsigned long end_pfn = start_pfn + nr_pages; 614 unsigned long pfn; 615 int order; 616 617 /* 618 * Online the pages. The callback might decide to keep some pages 619 * PG_reserved (to add them to the buddy later), but we still account 620 * them as being online/belonging to this zone ("present"). 621 */ 622 for (pfn = start_pfn; pfn < end_pfn; pfn += 1ul << order) { 623 order = min(MAX_ORDER - 1, get_order(PFN_PHYS(end_pfn - pfn))); 624 /* __free_pages_core() wants pfns to be aligned to the order */ 625 if (WARN_ON_ONCE(!IS_ALIGNED(pfn, 1ul << order))) 626 order = 0; 627 (*online_page_callback)(pfn_to_page(pfn), order); 628 } 629 630 /* mark all involved sections as online */ 631 online_mem_sections(start_pfn, end_pfn); 632 633 *(unsigned long *)arg += nr_pages; 634 return 0; 635 } 636 637 /* check which state of node_states will be changed when online memory */ 638 static void node_states_check_changes_online(unsigned long nr_pages, 639 struct zone *zone, struct memory_notify *arg) 640 { 641 int nid = zone_to_nid(zone); 642 643 arg->status_change_nid = NUMA_NO_NODE; 644 arg->status_change_nid_normal = NUMA_NO_NODE; 645 arg->status_change_nid_high = NUMA_NO_NODE; 646 647 if (!node_state(nid, N_MEMORY)) 648 arg->status_change_nid = nid; 649 if (zone_idx(zone) <= ZONE_NORMAL && !node_state(nid, N_NORMAL_MEMORY)) 650 arg->status_change_nid_normal = nid; 651 #ifdef CONFIG_HIGHMEM 652 if (zone_idx(zone) <= ZONE_HIGHMEM && !node_state(nid, N_HIGH_MEMORY)) 653 arg->status_change_nid_high = nid; 654 #endif 655 } 656 657 static void node_states_set_node(int node, struct memory_notify *arg) 658 { 659 if (arg->status_change_nid_normal >= 0) 660 node_set_state(node, N_NORMAL_MEMORY); 661 662 if (arg->status_change_nid_high >= 0) 663 node_set_state(node, N_HIGH_MEMORY); 664 665 if (arg->status_change_nid >= 0) 666 node_set_state(node, N_MEMORY); 667 } 668 669 static void __meminit resize_zone_range(struct zone *zone, unsigned long start_pfn, 670 unsigned long nr_pages) 671 { 672 unsigned long old_end_pfn = zone_end_pfn(zone); 673 674 if (zone_is_empty(zone) || start_pfn < zone->zone_start_pfn) 675 zone->zone_start_pfn = start_pfn; 676 677 zone->spanned_pages = max(start_pfn + nr_pages, old_end_pfn) - zone->zone_start_pfn; 678 } 679 680 static void __meminit resize_pgdat_range(struct pglist_data *pgdat, unsigned long start_pfn, 681 unsigned long nr_pages) 682 { 683 unsigned long old_end_pfn = pgdat_end_pfn(pgdat); 684 685 if (!pgdat->node_spanned_pages || start_pfn < pgdat->node_start_pfn) 686 pgdat->node_start_pfn = start_pfn; 687 688 pgdat->node_spanned_pages = max(start_pfn + nr_pages, old_end_pfn) - pgdat->node_start_pfn; 689 690 } 691 /* 692 * Associate the pfn range with the given zone, initializing the memmaps 693 * and resizing the pgdat/zone data to span the added pages. After this 694 * call, all affected pages are PG_reserved. 695 */ 696 void __ref move_pfn_range_to_zone(struct zone *zone, unsigned long start_pfn, 697 unsigned long nr_pages, struct vmem_altmap *altmap) 698 { 699 struct pglist_data *pgdat = zone->zone_pgdat; 700 int nid = pgdat->node_id; 701 unsigned long flags; 702 703 clear_zone_contiguous(zone); 704 705 /* TODO Huh pgdat is irqsave while zone is not. It used to be like that before */ 706 pgdat_resize_lock(pgdat, &flags); 707 zone_span_writelock(zone); 708 if (zone_is_empty(zone)) 709 init_currently_empty_zone(zone, start_pfn, nr_pages); 710 resize_zone_range(zone, start_pfn, nr_pages); 711 zone_span_writeunlock(zone); 712 resize_pgdat_range(pgdat, start_pfn, nr_pages); 713 pgdat_resize_unlock(pgdat, &flags); 714 715 /* 716 * TODO now we have a visible range of pages which are not associated 717 * with their zone properly. Not nice but set_pfnblock_flags_mask 718 * expects the zone spans the pfn range. All the pages in the range 719 * are reserved so nobody should be touching them so we should be safe 720 */ 721 memmap_init_zone(nr_pages, nid, zone_idx(zone), start_pfn, 722 MEMMAP_HOTPLUG, altmap); 723 724 set_zone_contiguous(zone); 725 } 726 727 /* 728 * Returns a default kernel memory zone for the given pfn range. 729 * If no kernel zone covers this pfn range it will automatically go 730 * to the ZONE_NORMAL. 731 */ 732 static struct zone *default_kernel_zone_for_pfn(int nid, unsigned long start_pfn, 733 unsigned long nr_pages) 734 { 735 struct pglist_data *pgdat = NODE_DATA(nid); 736 int zid; 737 738 for (zid = 0; zid <= ZONE_NORMAL; zid++) { 739 struct zone *zone = &pgdat->node_zones[zid]; 740 741 if (zone_intersects(zone, start_pfn, nr_pages)) 742 return zone; 743 } 744 745 return &pgdat->node_zones[ZONE_NORMAL]; 746 } 747 748 static inline struct zone *default_zone_for_pfn(int nid, unsigned long start_pfn, 749 unsigned long nr_pages) 750 { 751 struct zone *kernel_zone = default_kernel_zone_for_pfn(nid, start_pfn, 752 nr_pages); 753 struct zone *movable_zone = &NODE_DATA(nid)->node_zones[ZONE_MOVABLE]; 754 bool in_kernel = zone_intersects(kernel_zone, start_pfn, nr_pages); 755 bool in_movable = zone_intersects(movable_zone, start_pfn, nr_pages); 756 757 /* 758 * We inherit the existing zone in a simple case where zones do not 759 * overlap in the given range 760 */ 761 if (in_kernel ^ in_movable) 762 return (in_kernel) ? kernel_zone : movable_zone; 763 764 /* 765 * If the range doesn't belong to any zone or two zones overlap in the 766 * given range then we use movable zone only if movable_node is 767 * enabled because we always online to a kernel zone by default. 768 */ 769 return movable_node_enabled ? movable_zone : kernel_zone; 770 } 771 772 struct zone * zone_for_pfn_range(int online_type, int nid, unsigned start_pfn, 773 unsigned long nr_pages) 774 { 775 if (online_type == MMOP_ONLINE_KERNEL) 776 return default_kernel_zone_for_pfn(nid, start_pfn, nr_pages); 777 778 if (online_type == MMOP_ONLINE_MOVABLE) 779 return &NODE_DATA(nid)->node_zones[ZONE_MOVABLE]; 780 781 return default_zone_for_pfn(nid, start_pfn, nr_pages); 782 } 783 784 int __ref online_pages(unsigned long pfn, unsigned long nr_pages, 785 int online_type, int nid) 786 { 787 unsigned long flags; 788 unsigned long onlined_pages = 0; 789 struct zone *zone; 790 int need_zonelists_rebuild = 0; 791 int ret; 792 struct memory_notify arg; 793 794 mem_hotplug_begin(); 795 796 /* associate pfn range with the zone */ 797 zone = zone_for_pfn_range(online_type, nid, pfn, nr_pages); 798 move_pfn_range_to_zone(zone, pfn, nr_pages, NULL); 799 800 arg.start_pfn = pfn; 801 arg.nr_pages = nr_pages; 802 node_states_check_changes_online(nr_pages, zone, &arg); 803 804 ret = memory_notify(MEM_GOING_ONLINE, &arg); 805 ret = notifier_to_errno(ret); 806 if (ret) 807 goto failed_addition; 808 809 /* 810 * If this zone is not populated, then it is not in zonelist. 811 * This means the page allocator ignores this zone. 812 * So, zonelist must be updated after online. 813 */ 814 if (!populated_zone(zone)) { 815 need_zonelists_rebuild = 1; 816 setup_zone_pageset(zone); 817 } 818 819 ret = walk_system_ram_range(pfn, nr_pages, &onlined_pages, 820 online_pages_range); 821 if (ret) { 822 /* not a single memory resource was applicable */ 823 if (need_zonelists_rebuild) 824 zone_pcp_reset(zone); 825 goto failed_addition; 826 } 827 828 zone->present_pages += onlined_pages; 829 830 pgdat_resize_lock(zone->zone_pgdat, &flags); 831 zone->zone_pgdat->node_present_pages += onlined_pages; 832 pgdat_resize_unlock(zone->zone_pgdat, &flags); 833 834 shuffle_zone(zone); 835 836 node_states_set_node(nid, &arg); 837 if (need_zonelists_rebuild) 838 build_all_zonelists(NULL); 839 else 840 zone_pcp_update(zone); 841 842 init_per_zone_wmark_min(); 843 844 kswapd_run(nid); 845 kcompactd_run(nid); 846 847 vm_total_pages = nr_free_pagecache_pages(); 848 849 writeback_set_ratelimit(); 850 851 memory_notify(MEM_ONLINE, &arg); 852 mem_hotplug_done(); 853 return 0; 854 855 failed_addition: 856 pr_debug("online_pages [mem %#010llx-%#010llx] failed\n", 857 (unsigned long long) pfn << PAGE_SHIFT, 858 (((unsigned long long) pfn + nr_pages) << PAGE_SHIFT) - 1); 859 memory_notify(MEM_CANCEL_ONLINE, &arg); 860 remove_pfn_range_from_zone(zone, pfn, nr_pages); 861 mem_hotplug_done(); 862 return ret; 863 } 864 #endif /* CONFIG_MEMORY_HOTPLUG_SPARSE */ 865 866 static void reset_node_present_pages(pg_data_t *pgdat) 867 { 868 struct zone *z; 869 870 for (z = pgdat->node_zones; z < pgdat->node_zones + MAX_NR_ZONES; z++) 871 z->present_pages = 0; 872 873 pgdat->node_present_pages = 0; 874 } 875 876 /* we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG */ 877 static pg_data_t __ref *hotadd_new_pgdat(int nid) 878 { 879 struct pglist_data *pgdat; 880 881 pgdat = NODE_DATA(nid); 882 if (!pgdat) { 883 pgdat = arch_alloc_nodedata(nid); 884 if (!pgdat) 885 return NULL; 886 887 pgdat->per_cpu_nodestats = 888 alloc_percpu(struct per_cpu_nodestat); 889 arch_refresh_nodedata(nid, pgdat); 890 } else { 891 int cpu; 892 /* 893 * Reset the nr_zones, order and highest_zoneidx before reuse. 894 * Note that kswapd will init kswapd_highest_zoneidx properly 895 * when it starts in the near future. 896 */ 897 pgdat->nr_zones = 0; 898 pgdat->kswapd_order = 0; 899 pgdat->kswapd_highest_zoneidx = 0; 900 for_each_online_cpu(cpu) { 901 struct per_cpu_nodestat *p; 902 903 p = per_cpu_ptr(pgdat->per_cpu_nodestats, cpu); 904 memset(p, 0, sizeof(*p)); 905 } 906 } 907 908 /* we can use NODE_DATA(nid) from here */ 909 pgdat->node_id = nid; 910 pgdat->node_start_pfn = 0; 911 912 /* init node's zones as empty zones, we don't have any present pages.*/ 913 free_area_init_core_hotplug(nid); 914 915 /* 916 * The node we allocated has no zone fallback lists. For avoiding 917 * to access not-initialized zonelist, build here. 918 */ 919 build_all_zonelists(pgdat); 920 921 /* 922 * When memory is hot-added, all the memory is in offline state. So 923 * clear all zones' present_pages because they will be updated in 924 * online_pages() and offline_pages(). 925 */ 926 reset_node_managed_pages(pgdat); 927 reset_node_present_pages(pgdat); 928 929 return pgdat; 930 } 931 932 static void rollback_node_hotadd(int nid) 933 { 934 pg_data_t *pgdat = NODE_DATA(nid); 935 936 arch_refresh_nodedata(nid, NULL); 937 free_percpu(pgdat->per_cpu_nodestats); 938 arch_free_nodedata(pgdat); 939 } 940 941 942 /** 943 * try_online_node - online a node if offlined 944 * @nid: the node ID 945 * @set_node_online: Whether we want to online the node 946 * called by cpu_up() to online a node without onlined memory. 947 * 948 * Returns: 949 * 1 -> a new node has been allocated 950 * 0 -> the node is already online 951 * -ENOMEM -> the node could not be allocated 952 */ 953 static int __try_online_node(int nid, bool set_node_online) 954 { 955 pg_data_t *pgdat; 956 int ret = 1; 957 958 if (node_online(nid)) 959 return 0; 960 961 pgdat = hotadd_new_pgdat(nid); 962 if (!pgdat) { 963 pr_err("Cannot online node %d due to NULL pgdat\n", nid); 964 ret = -ENOMEM; 965 goto out; 966 } 967 968 if (set_node_online) { 969 node_set_online(nid); 970 ret = register_one_node(nid); 971 BUG_ON(ret); 972 } 973 out: 974 return ret; 975 } 976 977 /* 978 * Users of this function always want to online/register the node 979 */ 980 int try_online_node(int nid) 981 { 982 int ret; 983 984 mem_hotplug_begin(); 985 ret = __try_online_node(nid, true); 986 mem_hotplug_done(); 987 return ret; 988 } 989 990 static int check_hotplug_memory_range(u64 start, u64 size) 991 { 992 /* memory range must be block size aligned */ 993 if (!size || !IS_ALIGNED(start, memory_block_size_bytes()) || 994 !IS_ALIGNED(size, memory_block_size_bytes())) { 995 pr_err("Block size [%#lx] unaligned hotplug range: start %#llx, size %#llx", 996 memory_block_size_bytes(), start, size); 997 return -EINVAL; 998 } 999 1000 return 0; 1001 } 1002 1003 static int online_memory_block(struct memory_block *mem, void *arg) 1004 { 1005 mem->online_type = memhp_default_online_type; 1006 return device_online(&mem->dev); 1007 } 1008 1009 /* 1010 * NOTE: The caller must call lock_device_hotplug() to serialize hotplug 1011 * and online/offline operations (triggered e.g. by sysfs). 1012 * 1013 * we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG 1014 */ 1015 int __ref add_memory_resource(int nid, struct resource *res) 1016 { 1017 struct mhp_params params = { .pgprot = PAGE_KERNEL }; 1018 u64 start, size; 1019 bool new_node = false; 1020 int ret; 1021 1022 start = res->start; 1023 size = resource_size(res); 1024 1025 ret = check_hotplug_memory_range(start, size); 1026 if (ret) 1027 return ret; 1028 1029 if (!node_possible(nid)) { 1030 WARN(1, "node %d was absent from the node_possible_map\n", nid); 1031 return -EINVAL; 1032 } 1033 1034 mem_hotplug_begin(); 1035 1036 if (IS_ENABLED(CONFIG_ARCH_KEEP_MEMBLOCK)) 1037 memblock_add_node(start, size, nid); 1038 1039 ret = __try_online_node(nid, false); 1040 if (ret < 0) 1041 goto error; 1042 new_node = ret; 1043 1044 /* call arch's memory hotadd */ 1045 ret = arch_add_memory(nid, start, size, ¶ms); 1046 if (ret < 0) 1047 goto error; 1048 1049 /* create memory block devices after memory was added */ 1050 ret = create_memory_block_devices(start, size); 1051 if (ret) { 1052 arch_remove_memory(nid, start, size, NULL); 1053 goto error; 1054 } 1055 1056 if (new_node) { 1057 /* If sysfs file of new node can't be created, cpu on the node 1058 * can't be hot-added. There is no rollback way now. 1059 * So, check by BUG_ON() to catch it reluctantly.. 1060 * We online node here. We can't roll back from here. 1061 */ 1062 node_set_online(nid); 1063 ret = __register_one_node(nid); 1064 BUG_ON(ret); 1065 } 1066 1067 /* link memory sections under this node.*/ 1068 ret = link_mem_sections(nid, PFN_DOWN(start), PFN_UP(start + size - 1)); 1069 BUG_ON(ret); 1070 1071 /* create new memmap entry */ 1072 if (!strcmp(res->name, "System RAM")) 1073 firmware_map_add_hotplug(start, start + size, "System RAM"); 1074 1075 /* device_online() will take the lock when calling online_pages() */ 1076 mem_hotplug_done(); 1077 1078 /* online pages if requested */ 1079 if (memhp_default_online_type != MMOP_OFFLINE) 1080 walk_memory_blocks(start, size, NULL, online_memory_block); 1081 1082 return ret; 1083 error: 1084 /* rollback pgdat allocation and others */ 1085 if (new_node) 1086 rollback_node_hotadd(nid); 1087 if (IS_ENABLED(CONFIG_ARCH_KEEP_MEMBLOCK)) 1088 memblock_remove(start, size); 1089 mem_hotplug_done(); 1090 return ret; 1091 } 1092 1093 /* requires device_hotplug_lock, see add_memory_resource() */ 1094 int __ref __add_memory(int nid, u64 start, u64 size) 1095 { 1096 struct resource *res; 1097 int ret; 1098 1099 res = register_memory_resource(start, size, "System RAM"); 1100 if (IS_ERR(res)) 1101 return PTR_ERR(res); 1102 1103 ret = add_memory_resource(nid, res); 1104 if (ret < 0) 1105 release_memory_resource(res); 1106 return ret; 1107 } 1108 1109 int add_memory(int nid, u64 start, u64 size) 1110 { 1111 int rc; 1112 1113 lock_device_hotplug(); 1114 rc = __add_memory(nid, start, size); 1115 unlock_device_hotplug(); 1116 1117 return rc; 1118 } 1119 EXPORT_SYMBOL_GPL(add_memory); 1120 1121 /* 1122 * Add special, driver-managed memory to the system as system RAM. Such 1123 * memory is not exposed via the raw firmware-provided memmap as system 1124 * RAM, instead, it is detected and added by a driver - during cold boot, 1125 * after a reboot, and after kexec. 1126 * 1127 * Reasons why this memory should not be used for the initial memmap of a 1128 * kexec kernel or for placing kexec images: 1129 * - The booting kernel is in charge of determining how this memory will be 1130 * used (e.g., use persistent memory as system RAM) 1131 * - Coordination with a hypervisor is required before this memory 1132 * can be used (e.g., inaccessible parts). 1133 * 1134 * For this memory, no entries in /sys/firmware/memmap ("raw firmware-provided 1135 * memory map") are created. Also, the created memory resource is flagged 1136 * with IORESOURCE_MEM_DRIVER_MANAGED, so in-kernel users can special-case 1137 * this memory as well (esp., not place kexec images onto it). 1138 * 1139 * The resource_name (visible via /proc/iomem) has to have the format 1140 * "System RAM ($DRIVER)". 1141 */ 1142 int add_memory_driver_managed(int nid, u64 start, u64 size, 1143 const char *resource_name) 1144 { 1145 struct resource *res; 1146 int rc; 1147 1148 if (!resource_name || 1149 strstr(resource_name, "System RAM (") != resource_name || 1150 resource_name[strlen(resource_name) - 1] != ')') 1151 return -EINVAL; 1152 1153 lock_device_hotplug(); 1154 1155 res = register_memory_resource(start, size, resource_name); 1156 if (IS_ERR(res)) { 1157 rc = PTR_ERR(res); 1158 goto out_unlock; 1159 } 1160 1161 rc = add_memory_resource(nid, res); 1162 if (rc < 0) 1163 release_memory_resource(res); 1164 1165 out_unlock: 1166 unlock_device_hotplug(); 1167 return rc; 1168 } 1169 EXPORT_SYMBOL_GPL(add_memory_driver_managed); 1170 1171 #ifdef CONFIG_MEMORY_HOTREMOVE 1172 /* 1173 * Confirm all pages in a range [start, end) belong to the same zone (skipping 1174 * memory holes). When true, return the zone. 1175 */ 1176 struct zone *test_pages_in_a_zone(unsigned long start_pfn, 1177 unsigned long end_pfn) 1178 { 1179 unsigned long pfn, sec_end_pfn; 1180 struct zone *zone = NULL; 1181 struct page *page; 1182 int i; 1183 for (pfn = start_pfn, sec_end_pfn = SECTION_ALIGN_UP(start_pfn + 1); 1184 pfn < end_pfn; 1185 pfn = sec_end_pfn, sec_end_pfn += PAGES_PER_SECTION) { 1186 /* Make sure the memory section is present first */ 1187 if (!present_section_nr(pfn_to_section_nr(pfn))) 1188 continue; 1189 for (; pfn < sec_end_pfn && pfn < end_pfn; 1190 pfn += MAX_ORDER_NR_PAGES) { 1191 i = 0; 1192 /* This is just a CONFIG_HOLES_IN_ZONE check.*/ 1193 while ((i < MAX_ORDER_NR_PAGES) && 1194 !pfn_valid_within(pfn + i)) 1195 i++; 1196 if (i == MAX_ORDER_NR_PAGES || pfn + i >= end_pfn) 1197 continue; 1198 /* Check if we got outside of the zone */ 1199 if (zone && !zone_spans_pfn(zone, pfn + i)) 1200 return NULL; 1201 page = pfn_to_page(pfn + i); 1202 if (zone && page_zone(page) != zone) 1203 return NULL; 1204 zone = page_zone(page); 1205 } 1206 } 1207 1208 return zone; 1209 } 1210 1211 /* 1212 * Scan pfn range [start,end) to find movable/migratable pages (LRU pages, 1213 * non-lru movable pages and hugepages). Will skip over most unmovable 1214 * pages (esp., pages that can be skipped when offlining), but bail out on 1215 * definitely unmovable pages. 1216 * 1217 * Returns: 1218 * 0 in case a movable page is found and movable_pfn was updated. 1219 * -ENOENT in case no movable page was found. 1220 * -EBUSY in case a definitely unmovable page was found. 1221 */ 1222 static int scan_movable_pages(unsigned long start, unsigned long end, 1223 unsigned long *movable_pfn) 1224 { 1225 unsigned long pfn; 1226 1227 for (pfn = start; pfn < end; pfn++) { 1228 struct page *page, *head; 1229 unsigned long skip; 1230 1231 if (!pfn_valid(pfn)) 1232 continue; 1233 page = pfn_to_page(pfn); 1234 if (PageLRU(page)) 1235 goto found; 1236 if (__PageMovable(page)) 1237 goto found; 1238 1239 /* 1240 * PageOffline() pages that are not marked __PageMovable() and 1241 * have a reference count > 0 (after MEM_GOING_OFFLINE) are 1242 * definitely unmovable. If their reference count would be 0, 1243 * they could at least be skipped when offlining memory. 1244 */ 1245 if (PageOffline(page) && page_count(page)) 1246 return -EBUSY; 1247 1248 if (!PageHuge(page)) 1249 continue; 1250 head = compound_head(page); 1251 if (page_huge_active(head)) 1252 goto found; 1253 skip = compound_nr(head) - (page - head); 1254 pfn += skip - 1; 1255 } 1256 return -ENOENT; 1257 found: 1258 *movable_pfn = pfn; 1259 return 0; 1260 } 1261 1262 static struct page *new_node_page(struct page *page, unsigned long private) 1263 { 1264 int nid = page_to_nid(page); 1265 nodemask_t nmask = node_states[N_MEMORY]; 1266 1267 /* 1268 * try to allocate from a different node but reuse this node if there 1269 * are no other online nodes to be used (e.g. we are offlining a part 1270 * of the only existing node) 1271 */ 1272 node_clear(nid, nmask); 1273 if (nodes_empty(nmask)) 1274 node_set(nid, nmask); 1275 1276 return new_page_nodemask(page, nid, &nmask); 1277 } 1278 1279 static int 1280 do_migrate_range(unsigned long start_pfn, unsigned long end_pfn) 1281 { 1282 unsigned long pfn; 1283 struct page *page; 1284 int ret = 0; 1285 LIST_HEAD(source); 1286 1287 for (pfn = start_pfn; pfn < end_pfn; pfn++) { 1288 if (!pfn_valid(pfn)) 1289 continue; 1290 page = pfn_to_page(pfn); 1291 1292 if (PageHuge(page)) { 1293 struct page *head = compound_head(page); 1294 pfn = page_to_pfn(head) + compound_nr(head) - 1; 1295 isolate_huge_page(head, &source); 1296 continue; 1297 } else if (PageTransHuge(page)) 1298 pfn = page_to_pfn(compound_head(page)) 1299 + hpage_nr_pages(page) - 1; 1300 1301 /* 1302 * HWPoison pages have elevated reference counts so the migration would 1303 * fail on them. It also doesn't make any sense to migrate them in the 1304 * first place. Still try to unmap such a page in case it is still mapped 1305 * (e.g. current hwpoison implementation doesn't unmap KSM pages but keep 1306 * the unmap as the catch all safety net). 1307 */ 1308 if (PageHWPoison(page)) { 1309 if (WARN_ON(PageLRU(page))) 1310 isolate_lru_page(page); 1311 if (page_mapped(page)) 1312 try_to_unmap(page, TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS); 1313 continue; 1314 } 1315 1316 if (!get_page_unless_zero(page)) 1317 continue; 1318 /* 1319 * We can skip free pages. And we can deal with pages on 1320 * LRU and non-lru movable pages. 1321 */ 1322 if (PageLRU(page)) 1323 ret = isolate_lru_page(page); 1324 else 1325 ret = isolate_movable_page(page, ISOLATE_UNEVICTABLE); 1326 if (!ret) { /* Success */ 1327 list_add_tail(&page->lru, &source); 1328 if (!__PageMovable(page)) 1329 inc_node_page_state(page, NR_ISOLATED_ANON + 1330 page_is_file_lru(page)); 1331 1332 } else { 1333 pr_warn("failed to isolate pfn %lx\n", pfn); 1334 dump_page(page, "isolation failed"); 1335 } 1336 put_page(page); 1337 } 1338 if (!list_empty(&source)) { 1339 /* Allocate a new page from the nearest neighbor node */ 1340 ret = migrate_pages(&source, new_node_page, NULL, 0, 1341 MIGRATE_SYNC, MR_MEMORY_HOTPLUG); 1342 if (ret) { 1343 list_for_each_entry(page, &source, lru) { 1344 pr_warn("migrating pfn %lx failed ret:%d ", 1345 page_to_pfn(page), ret); 1346 dump_page(page, "migration failure"); 1347 } 1348 putback_movable_pages(&source); 1349 } 1350 } 1351 1352 return ret; 1353 } 1354 1355 /* Mark all sections offline and remove all free pages from the buddy. */ 1356 static int 1357 offline_isolated_pages_cb(unsigned long start, unsigned long nr_pages, 1358 void *data) 1359 { 1360 unsigned long *offlined_pages = (unsigned long *)data; 1361 1362 *offlined_pages += __offline_isolated_pages(start, start + nr_pages); 1363 return 0; 1364 } 1365 1366 /* 1367 * Check all pages in range, recorded as memory resource, are isolated. 1368 */ 1369 static int 1370 check_pages_isolated_cb(unsigned long start_pfn, unsigned long nr_pages, 1371 void *data) 1372 { 1373 return test_pages_isolated(start_pfn, start_pfn + nr_pages, 1374 MEMORY_OFFLINE); 1375 } 1376 1377 static int __init cmdline_parse_movable_node(char *p) 1378 { 1379 movable_node_enabled = true; 1380 return 0; 1381 } 1382 early_param("movable_node", cmdline_parse_movable_node); 1383 1384 /* check which state of node_states will be changed when offline memory */ 1385 static void node_states_check_changes_offline(unsigned long nr_pages, 1386 struct zone *zone, struct memory_notify *arg) 1387 { 1388 struct pglist_data *pgdat = zone->zone_pgdat; 1389 unsigned long present_pages = 0; 1390 enum zone_type zt; 1391 1392 arg->status_change_nid = NUMA_NO_NODE; 1393 arg->status_change_nid_normal = NUMA_NO_NODE; 1394 arg->status_change_nid_high = NUMA_NO_NODE; 1395 1396 /* 1397 * Check whether node_states[N_NORMAL_MEMORY] will be changed. 1398 * If the memory to be offline is within the range 1399 * [0..ZONE_NORMAL], and it is the last present memory there, 1400 * the zones in that range will become empty after the offlining, 1401 * thus we can determine that we need to clear the node from 1402 * node_states[N_NORMAL_MEMORY]. 1403 */ 1404 for (zt = 0; zt <= ZONE_NORMAL; zt++) 1405 present_pages += pgdat->node_zones[zt].present_pages; 1406 if (zone_idx(zone) <= ZONE_NORMAL && nr_pages >= present_pages) 1407 arg->status_change_nid_normal = zone_to_nid(zone); 1408 1409 #ifdef CONFIG_HIGHMEM 1410 /* 1411 * node_states[N_HIGH_MEMORY] contains nodes which 1412 * have normal memory or high memory. 1413 * Here we add the present_pages belonging to ZONE_HIGHMEM. 1414 * If the zone is within the range of [0..ZONE_HIGHMEM), and 1415 * we determine that the zones in that range become empty, 1416 * we need to clear the node for N_HIGH_MEMORY. 1417 */ 1418 present_pages += pgdat->node_zones[ZONE_HIGHMEM].present_pages; 1419 if (zone_idx(zone) <= ZONE_HIGHMEM && nr_pages >= present_pages) 1420 arg->status_change_nid_high = zone_to_nid(zone); 1421 #endif 1422 1423 /* 1424 * We have accounted the pages from [0..ZONE_NORMAL), and 1425 * in case of CONFIG_HIGHMEM the pages from ZONE_HIGHMEM 1426 * as well. 1427 * Here we count the possible pages from ZONE_MOVABLE. 1428 * If after having accounted all the pages, we see that the nr_pages 1429 * to be offlined is over or equal to the accounted pages, 1430 * we know that the node will become empty, and so, we can clear 1431 * it for N_MEMORY as well. 1432 */ 1433 present_pages += pgdat->node_zones[ZONE_MOVABLE].present_pages; 1434 1435 if (nr_pages >= present_pages) 1436 arg->status_change_nid = zone_to_nid(zone); 1437 } 1438 1439 static void node_states_clear_node(int node, struct memory_notify *arg) 1440 { 1441 if (arg->status_change_nid_normal >= 0) 1442 node_clear_state(node, N_NORMAL_MEMORY); 1443 1444 if (arg->status_change_nid_high >= 0) 1445 node_clear_state(node, N_HIGH_MEMORY); 1446 1447 if (arg->status_change_nid >= 0) 1448 node_clear_state(node, N_MEMORY); 1449 } 1450 1451 static int count_system_ram_pages_cb(unsigned long start_pfn, 1452 unsigned long nr_pages, void *data) 1453 { 1454 unsigned long *nr_system_ram_pages = data; 1455 1456 *nr_system_ram_pages += nr_pages; 1457 return 0; 1458 } 1459 1460 static int __ref __offline_pages(unsigned long start_pfn, 1461 unsigned long end_pfn) 1462 { 1463 unsigned long pfn, nr_pages = 0; 1464 unsigned long offlined_pages = 0; 1465 int ret, node, nr_isolate_pageblock; 1466 unsigned long flags; 1467 struct zone *zone; 1468 struct memory_notify arg; 1469 char *reason; 1470 1471 mem_hotplug_begin(); 1472 1473 /* 1474 * Don't allow to offline memory blocks that contain holes. 1475 * Consequently, memory blocks with holes can never get onlined 1476 * via the hotplug path - online_pages() - as hotplugged memory has 1477 * no holes. This way, we e.g., don't have to worry about marking 1478 * memory holes PG_reserved, don't need pfn_valid() checks, and can 1479 * avoid using walk_system_ram_range() later. 1480 */ 1481 walk_system_ram_range(start_pfn, end_pfn - start_pfn, &nr_pages, 1482 count_system_ram_pages_cb); 1483 if (nr_pages != end_pfn - start_pfn) { 1484 ret = -EINVAL; 1485 reason = "memory holes"; 1486 goto failed_removal; 1487 } 1488 1489 /* This makes hotplug much easier...and readable. 1490 we assume this for now. .*/ 1491 zone = test_pages_in_a_zone(start_pfn, end_pfn); 1492 if (!zone) { 1493 ret = -EINVAL; 1494 reason = "multizone range"; 1495 goto failed_removal; 1496 } 1497 node = zone_to_nid(zone); 1498 1499 /* set above range as isolated */ 1500 ret = start_isolate_page_range(start_pfn, end_pfn, 1501 MIGRATE_MOVABLE, 1502 MEMORY_OFFLINE | REPORT_FAILURE); 1503 if (ret < 0) { 1504 reason = "failure to isolate range"; 1505 goto failed_removal; 1506 } 1507 nr_isolate_pageblock = ret; 1508 1509 arg.start_pfn = start_pfn; 1510 arg.nr_pages = nr_pages; 1511 node_states_check_changes_offline(nr_pages, zone, &arg); 1512 1513 ret = memory_notify(MEM_GOING_OFFLINE, &arg); 1514 ret = notifier_to_errno(ret); 1515 if (ret) { 1516 reason = "notifier failure"; 1517 goto failed_removal_isolated; 1518 } 1519 1520 do { 1521 pfn = start_pfn; 1522 do { 1523 if (signal_pending(current)) { 1524 ret = -EINTR; 1525 reason = "signal backoff"; 1526 goto failed_removal_isolated; 1527 } 1528 1529 cond_resched(); 1530 lru_add_drain_all(); 1531 1532 ret = scan_movable_pages(pfn, end_pfn, &pfn); 1533 if (!ret) { 1534 /* 1535 * TODO: fatal migration failures should bail 1536 * out 1537 */ 1538 do_migrate_range(pfn, end_pfn); 1539 } 1540 } while (!ret); 1541 1542 if (ret != -ENOENT) { 1543 reason = "unmovable page"; 1544 goto failed_removal_isolated; 1545 } 1546 1547 /* 1548 * Dissolve free hugepages in the memory block before doing 1549 * offlining actually in order to make hugetlbfs's object 1550 * counting consistent. 1551 */ 1552 ret = dissolve_free_huge_pages(start_pfn, end_pfn); 1553 if (ret) { 1554 reason = "failure to dissolve huge pages"; 1555 goto failed_removal_isolated; 1556 } 1557 /* check again */ 1558 ret = walk_system_ram_range(start_pfn, end_pfn - start_pfn, 1559 NULL, check_pages_isolated_cb); 1560 } while (ret); 1561 1562 /* Ok, all of our target is isolated. 1563 We cannot do rollback at this point. */ 1564 walk_system_ram_range(start_pfn, end_pfn - start_pfn, 1565 &offlined_pages, offline_isolated_pages_cb); 1566 pr_info("Offlined Pages %ld\n", offlined_pages); 1567 /* 1568 * Onlining will reset pagetype flags and makes migrate type 1569 * MOVABLE, so just need to decrease the number of isolated 1570 * pageblocks zone counter here. 1571 */ 1572 spin_lock_irqsave(&zone->lock, flags); 1573 zone->nr_isolate_pageblock -= nr_isolate_pageblock; 1574 spin_unlock_irqrestore(&zone->lock, flags); 1575 1576 /* removal success */ 1577 adjust_managed_page_count(pfn_to_page(start_pfn), -offlined_pages); 1578 zone->present_pages -= offlined_pages; 1579 1580 pgdat_resize_lock(zone->zone_pgdat, &flags); 1581 zone->zone_pgdat->node_present_pages -= offlined_pages; 1582 pgdat_resize_unlock(zone->zone_pgdat, &flags); 1583 1584 init_per_zone_wmark_min(); 1585 1586 if (!populated_zone(zone)) { 1587 zone_pcp_reset(zone); 1588 build_all_zonelists(NULL); 1589 } else 1590 zone_pcp_update(zone); 1591 1592 node_states_clear_node(node, &arg); 1593 if (arg.status_change_nid >= 0) { 1594 kswapd_stop(node); 1595 kcompactd_stop(node); 1596 } 1597 1598 vm_total_pages = nr_free_pagecache_pages(); 1599 writeback_set_ratelimit(); 1600 1601 memory_notify(MEM_OFFLINE, &arg); 1602 remove_pfn_range_from_zone(zone, start_pfn, nr_pages); 1603 mem_hotplug_done(); 1604 return 0; 1605 1606 failed_removal_isolated: 1607 undo_isolate_page_range(start_pfn, end_pfn, MIGRATE_MOVABLE); 1608 memory_notify(MEM_CANCEL_OFFLINE, &arg); 1609 failed_removal: 1610 pr_debug("memory offlining [mem %#010llx-%#010llx] failed due to %s\n", 1611 (unsigned long long) start_pfn << PAGE_SHIFT, 1612 ((unsigned long long) end_pfn << PAGE_SHIFT) - 1, 1613 reason); 1614 /* pushback to free area */ 1615 mem_hotplug_done(); 1616 return ret; 1617 } 1618 1619 int offline_pages(unsigned long start_pfn, unsigned long nr_pages) 1620 { 1621 return __offline_pages(start_pfn, start_pfn + nr_pages); 1622 } 1623 1624 static int check_memblock_offlined_cb(struct memory_block *mem, void *arg) 1625 { 1626 int ret = !is_memblock_offlined(mem); 1627 1628 if (unlikely(ret)) { 1629 phys_addr_t beginpa, endpa; 1630 1631 beginpa = PFN_PHYS(section_nr_to_pfn(mem->start_section_nr)); 1632 endpa = beginpa + memory_block_size_bytes() - 1; 1633 pr_warn("removing memory fails, because memory [%pa-%pa] is onlined\n", 1634 &beginpa, &endpa); 1635 1636 return -EBUSY; 1637 } 1638 return 0; 1639 } 1640 1641 static int check_cpu_on_node(pg_data_t *pgdat) 1642 { 1643 int cpu; 1644 1645 for_each_present_cpu(cpu) { 1646 if (cpu_to_node(cpu) == pgdat->node_id) 1647 /* 1648 * the cpu on this node isn't removed, and we can't 1649 * offline this node. 1650 */ 1651 return -EBUSY; 1652 } 1653 1654 return 0; 1655 } 1656 1657 static int check_no_memblock_for_node_cb(struct memory_block *mem, void *arg) 1658 { 1659 int nid = *(int *)arg; 1660 1661 /* 1662 * If a memory block belongs to multiple nodes, the stored nid is not 1663 * reliable. However, such blocks are always online (e.g., cannot get 1664 * offlined) and, therefore, are still spanned by the node. 1665 */ 1666 return mem->nid == nid ? -EEXIST : 0; 1667 } 1668 1669 /** 1670 * try_offline_node 1671 * @nid: the node ID 1672 * 1673 * Offline a node if all memory sections and cpus of the node are removed. 1674 * 1675 * NOTE: The caller must call lock_device_hotplug() to serialize hotplug 1676 * and online/offline operations before this call. 1677 */ 1678 void try_offline_node(int nid) 1679 { 1680 pg_data_t *pgdat = NODE_DATA(nid); 1681 int rc; 1682 1683 /* 1684 * If the node still spans pages (especially ZONE_DEVICE), don't 1685 * offline it. A node spans memory after move_pfn_range_to_zone(), 1686 * e.g., after the memory block was onlined. 1687 */ 1688 if (pgdat->node_spanned_pages) 1689 return; 1690 1691 /* 1692 * Especially offline memory blocks might not be spanned by the 1693 * node. They will get spanned by the node once they get onlined. 1694 * However, they link to the node in sysfs and can get onlined later. 1695 */ 1696 rc = for_each_memory_block(&nid, check_no_memblock_for_node_cb); 1697 if (rc) 1698 return; 1699 1700 if (check_cpu_on_node(pgdat)) 1701 return; 1702 1703 /* 1704 * all memory/cpu of this node are removed, we can offline this 1705 * node now. 1706 */ 1707 node_set_offline(nid); 1708 unregister_one_node(nid); 1709 } 1710 EXPORT_SYMBOL(try_offline_node); 1711 1712 static void __release_memory_resource(resource_size_t start, 1713 resource_size_t size) 1714 { 1715 int ret; 1716 1717 /* 1718 * When removing memory in the same granularity as it was added, 1719 * this function never fails. It might only fail if resources 1720 * have to be adjusted or split. We'll ignore the error, as 1721 * removing of memory cannot fail. 1722 */ 1723 ret = release_mem_region_adjustable(&iomem_resource, start, size); 1724 if (ret) { 1725 resource_size_t endres = start + size - 1; 1726 1727 pr_warn("Unable to release resource <%pa-%pa> (%d)\n", 1728 &start, &endres, ret); 1729 } 1730 } 1731 1732 static int __ref try_remove_memory(int nid, u64 start, u64 size) 1733 { 1734 int rc = 0; 1735 1736 BUG_ON(check_hotplug_memory_range(start, size)); 1737 1738 /* 1739 * All memory blocks must be offlined before removing memory. Check 1740 * whether all memory blocks in question are offline and return error 1741 * if this is not the case. 1742 */ 1743 rc = walk_memory_blocks(start, size, NULL, check_memblock_offlined_cb); 1744 if (rc) 1745 goto done; 1746 1747 /* remove memmap entry */ 1748 firmware_map_remove(start, start + size, "System RAM"); 1749 1750 /* 1751 * Memory block device removal under the device_hotplug_lock is 1752 * a barrier against racing online attempts. 1753 */ 1754 remove_memory_block_devices(start, size); 1755 1756 mem_hotplug_begin(); 1757 1758 arch_remove_memory(nid, start, size, NULL); 1759 1760 if (IS_ENABLED(CONFIG_ARCH_KEEP_MEMBLOCK)) { 1761 memblock_free(start, size); 1762 memblock_remove(start, size); 1763 } 1764 1765 __release_memory_resource(start, size); 1766 1767 try_offline_node(nid); 1768 1769 done: 1770 mem_hotplug_done(); 1771 return rc; 1772 } 1773 1774 /** 1775 * remove_memory 1776 * @nid: the node ID 1777 * @start: physical address of the region to remove 1778 * @size: size of the region to remove 1779 * 1780 * NOTE: The caller must call lock_device_hotplug() to serialize hotplug 1781 * and online/offline operations before this call, as required by 1782 * try_offline_node(). 1783 */ 1784 void __remove_memory(int nid, u64 start, u64 size) 1785 { 1786 1787 /* 1788 * trigger BUG() if some memory is not offlined prior to calling this 1789 * function 1790 */ 1791 if (try_remove_memory(nid, start, size)) 1792 BUG(); 1793 } 1794 1795 /* 1796 * Remove memory if every memory block is offline, otherwise return -EBUSY is 1797 * some memory is not offline 1798 */ 1799 int remove_memory(int nid, u64 start, u64 size) 1800 { 1801 int rc; 1802 1803 lock_device_hotplug(); 1804 rc = try_remove_memory(nid, start, size); 1805 unlock_device_hotplug(); 1806 1807 return rc; 1808 } 1809 EXPORT_SYMBOL_GPL(remove_memory); 1810 1811 /* 1812 * Try to offline and remove a memory block. Might take a long time to 1813 * finish in case memory is still in use. Primarily useful for memory devices 1814 * that logically unplugged all memory (so it's no longer in use) and want to 1815 * offline + remove the memory block. 1816 */ 1817 int offline_and_remove_memory(int nid, u64 start, u64 size) 1818 { 1819 struct memory_block *mem; 1820 int rc = -EINVAL; 1821 1822 if (!IS_ALIGNED(start, memory_block_size_bytes()) || 1823 size != memory_block_size_bytes()) 1824 return rc; 1825 1826 lock_device_hotplug(); 1827 mem = find_memory_block(__pfn_to_section(PFN_DOWN(start))); 1828 if (mem) 1829 rc = device_offline(&mem->dev); 1830 /* Ignore if the device is already offline. */ 1831 if (rc > 0) 1832 rc = 0; 1833 1834 /* 1835 * In case we succeeded to offline the memory block, remove it. 1836 * This cannot fail as it cannot get onlined in the meantime. 1837 */ 1838 if (!rc) { 1839 rc = try_remove_memory(nid, start, size); 1840 WARN_ON_ONCE(rc); 1841 } 1842 unlock_device_hotplug(); 1843 1844 return rc; 1845 } 1846 EXPORT_SYMBOL_GPL(offline_and_remove_memory); 1847 #endif /* CONFIG_MEMORY_HOTREMOVE */ 1848