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 /* 835 * When exposing larger, physically contiguous memory areas to the 836 * buddy, shuffling in the buddy (when freeing onlined pages, putting 837 * them either to the head or the tail of the freelist) is only helpful 838 * for maintaining the shuffle, but not for creating the initial 839 * shuffle. Shuffle the whole zone to make sure the just onlined pages 840 * are properly distributed across the whole freelist. 841 */ 842 shuffle_zone(zone); 843 844 node_states_set_node(nid, &arg); 845 if (need_zonelists_rebuild) 846 build_all_zonelists(NULL); 847 else 848 zone_pcp_update(zone); 849 850 init_per_zone_wmark_min(); 851 852 kswapd_run(nid); 853 kcompactd_run(nid); 854 855 writeback_set_ratelimit(); 856 857 memory_notify(MEM_ONLINE, &arg); 858 mem_hotplug_done(); 859 return 0; 860 861 failed_addition: 862 pr_debug("online_pages [mem %#010llx-%#010llx] failed\n", 863 (unsigned long long) pfn << PAGE_SHIFT, 864 (((unsigned long long) pfn + nr_pages) << PAGE_SHIFT) - 1); 865 memory_notify(MEM_CANCEL_ONLINE, &arg); 866 remove_pfn_range_from_zone(zone, pfn, nr_pages); 867 mem_hotplug_done(); 868 return ret; 869 } 870 #endif /* CONFIG_MEMORY_HOTPLUG_SPARSE */ 871 872 static void reset_node_present_pages(pg_data_t *pgdat) 873 { 874 struct zone *z; 875 876 for (z = pgdat->node_zones; z < pgdat->node_zones + MAX_NR_ZONES; z++) 877 z->present_pages = 0; 878 879 pgdat->node_present_pages = 0; 880 } 881 882 /* we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG */ 883 static pg_data_t __ref *hotadd_new_pgdat(int nid) 884 { 885 struct pglist_data *pgdat; 886 887 pgdat = NODE_DATA(nid); 888 if (!pgdat) { 889 pgdat = arch_alloc_nodedata(nid); 890 if (!pgdat) 891 return NULL; 892 893 pgdat->per_cpu_nodestats = 894 alloc_percpu(struct per_cpu_nodestat); 895 arch_refresh_nodedata(nid, pgdat); 896 } else { 897 int cpu; 898 /* 899 * Reset the nr_zones, order and highest_zoneidx before reuse. 900 * Note that kswapd will init kswapd_highest_zoneidx properly 901 * when it starts in the near future. 902 */ 903 pgdat->nr_zones = 0; 904 pgdat->kswapd_order = 0; 905 pgdat->kswapd_highest_zoneidx = 0; 906 for_each_online_cpu(cpu) { 907 struct per_cpu_nodestat *p; 908 909 p = per_cpu_ptr(pgdat->per_cpu_nodestats, cpu); 910 memset(p, 0, sizeof(*p)); 911 } 912 } 913 914 /* we can use NODE_DATA(nid) from here */ 915 pgdat->node_id = nid; 916 pgdat->node_start_pfn = 0; 917 918 /* init node's zones as empty zones, we don't have any present pages.*/ 919 free_area_init_core_hotplug(nid); 920 921 /* 922 * The node we allocated has no zone fallback lists. For avoiding 923 * to access not-initialized zonelist, build here. 924 */ 925 build_all_zonelists(pgdat); 926 927 /* 928 * When memory is hot-added, all the memory is in offline state. So 929 * clear all zones' present_pages because they will be updated in 930 * online_pages() and offline_pages(). 931 */ 932 reset_node_managed_pages(pgdat); 933 reset_node_present_pages(pgdat); 934 935 return pgdat; 936 } 937 938 static void rollback_node_hotadd(int nid) 939 { 940 pg_data_t *pgdat = NODE_DATA(nid); 941 942 arch_refresh_nodedata(nid, NULL); 943 free_percpu(pgdat->per_cpu_nodestats); 944 arch_free_nodedata(pgdat); 945 } 946 947 948 /** 949 * try_online_node - online a node if offlined 950 * @nid: the node ID 951 * @set_node_online: Whether we want to online the node 952 * called by cpu_up() to online a node without onlined memory. 953 * 954 * Returns: 955 * 1 -> a new node has been allocated 956 * 0 -> the node is already online 957 * -ENOMEM -> the node could not be allocated 958 */ 959 static int __try_online_node(int nid, bool set_node_online) 960 { 961 pg_data_t *pgdat; 962 int ret = 1; 963 964 if (node_online(nid)) 965 return 0; 966 967 pgdat = hotadd_new_pgdat(nid); 968 if (!pgdat) { 969 pr_err("Cannot online node %d due to NULL pgdat\n", nid); 970 ret = -ENOMEM; 971 goto out; 972 } 973 974 if (set_node_online) { 975 node_set_online(nid); 976 ret = register_one_node(nid); 977 BUG_ON(ret); 978 } 979 out: 980 return ret; 981 } 982 983 /* 984 * Users of this function always want to online/register the node 985 */ 986 int try_online_node(int nid) 987 { 988 int ret; 989 990 mem_hotplug_begin(); 991 ret = __try_online_node(nid, true); 992 mem_hotplug_done(); 993 return ret; 994 } 995 996 static int check_hotplug_memory_range(u64 start, u64 size) 997 { 998 /* memory range must be block size aligned */ 999 if (!size || !IS_ALIGNED(start, memory_block_size_bytes()) || 1000 !IS_ALIGNED(size, memory_block_size_bytes())) { 1001 pr_err("Block size [%#lx] unaligned hotplug range: start %#llx, size %#llx", 1002 memory_block_size_bytes(), start, size); 1003 return -EINVAL; 1004 } 1005 1006 return 0; 1007 } 1008 1009 static int online_memory_block(struct memory_block *mem, void *arg) 1010 { 1011 mem->online_type = memhp_default_online_type; 1012 return device_online(&mem->dev); 1013 } 1014 1015 /* 1016 * NOTE: The caller must call lock_device_hotplug() to serialize hotplug 1017 * and online/offline operations (triggered e.g. by sysfs). 1018 * 1019 * we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG 1020 */ 1021 int __ref add_memory_resource(int nid, struct resource *res) 1022 { 1023 struct mhp_params params = { .pgprot = PAGE_KERNEL }; 1024 u64 start, size; 1025 bool new_node = false; 1026 int ret; 1027 1028 start = res->start; 1029 size = resource_size(res); 1030 1031 ret = check_hotplug_memory_range(start, size); 1032 if (ret) 1033 return ret; 1034 1035 if (!node_possible(nid)) { 1036 WARN(1, "node %d was absent from the node_possible_map\n", nid); 1037 return -EINVAL; 1038 } 1039 1040 mem_hotplug_begin(); 1041 1042 if (IS_ENABLED(CONFIG_ARCH_KEEP_MEMBLOCK)) 1043 memblock_add_node(start, size, nid); 1044 1045 ret = __try_online_node(nid, false); 1046 if (ret < 0) 1047 goto error; 1048 new_node = ret; 1049 1050 /* call arch's memory hotadd */ 1051 ret = arch_add_memory(nid, start, size, ¶ms); 1052 if (ret < 0) 1053 goto error; 1054 1055 /* create memory block devices after memory was added */ 1056 ret = create_memory_block_devices(start, size); 1057 if (ret) { 1058 arch_remove_memory(nid, start, size, NULL); 1059 goto error; 1060 } 1061 1062 if (new_node) { 1063 /* If sysfs file of new node can't be created, cpu on the node 1064 * can't be hot-added. There is no rollback way now. 1065 * So, check by BUG_ON() to catch it reluctantly.. 1066 * We online node here. We can't roll back from here. 1067 */ 1068 node_set_online(nid); 1069 ret = __register_one_node(nid); 1070 BUG_ON(ret); 1071 } 1072 1073 /* link memory sections under this node.*/ 1074 ret = link_mem_sections(nid, PFN_DOWN(start), PFN_UP(start + size - 1)); 1075 BUG_ON(ret); 1076 1077 /* create new memmap entry */ 1078 if (!strcmp(res->name, "System RAM")) 1079 firmware_map_add_hotplug(start, start + size, "System RAM"); 1080 1081 /* device_online() will take the lock when calling online_pages() */ 1082 mem_hotplug_done(); 1083 1084 /* online pages if requested */ 1085 if (memhp_default_online_type != MMOP_OFFLINE) 1086 walk_memory_blocks(start, size, NULL, online_memory_block); 1087 1088 return ret; 1089 error: 1090 /* rollback pgdat allocation and others */ 1091 if (new_node) 1092 rollback_node_hotadd(nid); 1093 if (IS_ENABLED(CONFIG_ARCH_KEEP_MEMBLOCK)) 1094 memblock_remove(start, size); 1095 mem_hotplug_done(); 1096 return ret; 1097 } 1098 1099 /* requires device_hotplug_lock, see add_memory_resource() */ 1100 int __ref __add_memory(int nid, u64 start, u64 size) 1101 { 1102 struct resource *res; 1103 int ret; 1104 1105 res = register_memory_resource(start, size, "System RAM"); 1106 if (IS_ERR(res)) 1107 return PTR_ERR(res); 1108 1109 ret = add_memory_resource(nid, res); 1110 if (ret < 0) 1111 release_memory_resource(res); 1112 return ret; 1113 } 1114 1115 int add_memory(int nid, u64 start, u64 size) 1116 { 1117 int rc; 1118 1119 lock_device_hotplug(); 1120 rc = __add_memory(nid, start, size); 1121 unlock_device_hotplug(); 1122 1123 return rc; 1124 } 1125 EXPORT_SYMBOL_GPL(add_memory); 1126 1127 /* 1128 * Add special, driver-managed memory to the system as system RAM. Such 1129 * memory is not exposed via the raw firmware-provided memmap as system 1130 * RAM, instead, it is detected and added by a driver - during cold boot, 1131 * after a reboot, and after kexec. 1132 * 1133 * Reasons why this memory should not be used for the initial memmap of a 1134 * kexec kernel or for placing kexec images: 1135 * - The booting kernel is in charge of determining how this memory will be 1136 * used (e.g., use persistent memory as system RAM) 1137 * - Coordination with a hypervisor is required before this memory 1138 * can be used (e.g., inaccessible parts). 1139 * 1140 * For this memory, no entries in /sys/firmware/memmap ("raw firmware-provided 1141 * memory map") are created. Also, the created memory resource is flagged 1142 * with IORESOURCE_MEM_DRIVER_MANAGED, so in-kernel users can special-case 1143 * this memory as well (esp., not place kexec images onto it). 1144 * 1145 * The resource_name (visible via /proc/iomem) has to have the format 1146 * "System RAM ($DRIVER)". 1147 */ 1148 int add_memory_driver_managed(int nid, u64 start, u64 size, 1149 const char *resource_name) 1150 { 1151 struct resource *res; 1152 int rc; 1153 1154 if (!resource_name || 1155 strstr(resource_name, "System RAM (") != resource_name || 1156 resource_name[strlen(resource_name) - 1] != ')') 1157 return -EINVAL; 1158 1159 lock_device_hotplug(); 1160 1161 res = register_memory_resource(start, size, resource_name); 1162 if (IS_ERR(res)) { 1163 rc = PTR_ERR(res); 1164 goto out_unlock; 1165 } 1166 1167 rc = add_memory_resource(nid, res); 1168 if (rc < 0) 1169 release_memory_resource(res); 1170 1171 out_unlock: 1172 unlock_device_hotplug(); 1173 return rc; 1174 } 1175 EXPORT_SYMBOL_GPL(add_memory_driver_managed); 1176 1177 #ifdef CONFIG_MEMORY_HOTREMOVE 1178 /* 1179 * Confirm all pages in a range [start, end) belong to the same zone (skipping 1180 * memory holes). When true, return the zone. 1181 */ 1182 struct zone *test_pages_in_a_zone(unsigned long start_pfn, 1183 unsigned long end_pfn) 1184 { 1185 unsigned long pfn, sec_end_pfn; 1186 struct zone *zone = NULL; 1187 struct page *page; 1188 int i; 1189 for (pfn = start_pfn, sec_end_pfn = SECTION_ALIGN_UP(start_pfn + 1); 1190 pfn < end_pfn; 1191 pfn = sec_end_pfn, sec_end_pfn += PAGES_PER_SECTION) { 1192 /* Make sure the memory section is present first */ 1193 if (!present_section_nr(pfn_to_section_nr(pfn))) 1194 continue; 1195 for (; pfn < sec_end_pfn && pfn < end_pfn; 1196 pfn += MAX_ORDER_NR_PAGES) { 1197 i = 0; 1198 /* This is just a CONFIG_HOLES_IN_ZONE check.*/ 1199 while ((i < MAX_ORDER_NR_PAGES) && 1200 !pfn_valid_within(pfn + i)) 1201 i++; 1202 if (i == MAX_ORDER_NR_PAGES || pfn + i >= end_pfn) 1203 continue; 1204 /* Check if we got outside of the zone */ 1205 if (zone && !zone_spans_pfn(zone, pfn + i)) 1206 return NULL; 1207 page = pfn_to_page(pfn + i); 1208 if (zone && page_zone(page) != zone) 1209 return NULL; 1210 zone = page_zone(page); 1211 } 1212 } 1213 1214 return zone; 1215 } 1216 1217 /* 1218 * Scan pfn range [start,end) to find movable/migratable pages (LRU pages, 1219 * non-lru movable pages and hugepages). Will skip over most unmovable 1220 * pages (esp., pages that can be skipped when offlining), but bail out on 1221 * definitely unmovable pages. 1222 * 1223 * Returns: 1224 * 0 in case a movable page is found and movable_pfn was updated. 1225 * -ENOENT in case no movable page was found. 1226 * -EBUSY in case a definitely unmovable page was found. 1227 */ 1228 static int scan_movable_pages(unsigned long start, unsigned long end, 1229 unsigned long *movable_pfn) 1230 { 1231 unsigned long pfn; 1232 1233 for (pfn = start; pfn < end; pfn++) { 1234 struct page *page, *head; 1235 unsigned long skip; 1236 1237 if (!pfn_valid(pfn)) 1238 continue; 1239 page = pfn_to_page(pfn); 1240 if (PageLRU(page)) 1241 goto found; 1242 if (__PageMovable(page)) 1243 goto found; 1244 1245 /* 1246 * PageOffline() pages that are not marked __PageMovable() and 1247 * have a reference count > 0 (after MEM_GOING_OFFLINE) are 1248 * definitely unmovable. If their reference count would be 0, 1249 * they could at least be skipped when offlining memory. 1250 */ 1251 if (PageOffline(page) && page_count(page)) 1252 return -EBUSY; 1253 1254 if (!PageHuge(page)) 1255 continue; 1256 head = compound_head(page); 1257 if (page_huge_active(head)) 1258 goto found; 1259 skip = compound_nr(head) - (page - head); 1260 pfn += skip - 1; 1261 } 1262 return -ENOENT; 1263 found: 1264 *movable_pfn = pfn; 1265 return 0; 1266 } 1267 1268 static struct page *new_node_page(struct page *page, unsigned long private) 1269 { 1270 int nid = page_to_nid(page); 1271 nodemask_t nmask = node_states[N_MEMORY]; 1272 1273 /* 1274 * try to allocate from a different node but reuse this node if there 1275 * are no other online nodes to be used (e.g. we are offlining a part 1276 * of the only existing node) 1277 */ 1278 node_clear(nid, nmask); 1279 if (nodes_empty(nmask)) 1280 node_set(nid, nmask); 1281 1282 return new_page_nodemask(page, nid, &nmask); 1283 } 1284 1285 static int 1286 do_migrate_range(unsigned long start_pfn, unsigned long end_pfn) 1287 { 1288 unsigned long pfn; 1289 struct page *page; 1290 int ret = 0; 1291 LIST_HEAD(source); 1292 1293 for (pfn = start_pfn; pfn < end_pfn; pfn++) { 1294 if (!pfn_valid(pfn)) 1295 continue; 1296 page = pfn_to_page(pfn); 1297 1298 if (PageHuge(page)) { 1299 struct page *head = compound_head(page); 1300 pfn = page_to_pfn(head) + compound_nr(head) - 1; 1301 isolate_huge_page(head, &source); 1302 continue; 1303 } else if (PageTransHuge(page)) 1304 pfn = page_to_pfn(compound_head(page)) 1305 + hpage_nr_pages(page) - 1; 1306 1307 /* 1308 * HWPoison pages have elevated reference counts so the migration would 1309 * fail on them. It also doesn't make any sense to migrate them in the 1310 * first place. Still try to unmap such a page in case it is still mapped 1311 * (e.g. current hwpoison implementation doesn't unmap KSM pages but keep 1312 * the unmap as the catch all safety net). 1313 */ 1314 if (PageHWPoison(page)) { 1315 if (WARN_ON(PageLRU(page))) 1316 isolate_lru_page(page); 1317 if (page_mapped(page)) 1318 try_to_unmap(page, TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS); 1319 continue; 1320 } 1321 1322 if (!get_page_unless_zero(page)) 1323 continue; 1324 /* 1325 * We can skip free pages. And we can deal with pages on 1326 * LRU and non-lru movable pages. 1327 */ 1328 if (PageLRU(page)) 1329 ret = isolate_lru_page(page); 1330 else 1331 ret = isolate_movable_page(page, ISOLATE_UNEVICTABLE); 1332 if (!ret) { /* Success */ 1333 list_add_tail(&page->lru, &source); 1334 if (!__PageMovable(page)) 1335 inc_node_page_state(page, NR_ISOLATED_ANON + 1336 page_is_file_lru(page)); 1337 1338 } else { 1339 pr_warn("failed to isolate pfn %lx\n", pfn); 1340 dump_page(page, "isolation failed"); 1341 } 1342 put_page(page); 1343 } 1344 if (!list_empty(&source)) { 1345 /* Allocate a new page from the nearest neighbor node */ 1346 ret = migrate_pages(&source, new_node_page, NULL, 0, 1347 MIGRATE_SYNC, MR_MEMORY_HOTPLUG); 1348 if (ret) { 1349 list_for_each_entry(page, &source, lru) { 1350 pr_warn("migrating pfn %lx failed ret:%d ", 1351 page_to_pfn(page), ret); 1352 dump_page(page, "migration failure"); 1353 } 1354 putback_movable_pages(&source); 1355 } 1356 } 1357 1358 return ret; 1359 } 1360 1361 /* Mark all sections offline and remove all free pages from the buddy. */ 1362 static int 1363 offline_isolated_pages_cb(unsigned long start, unsigned long nr_pages, 1364 void *data) 1365 { 1366 unsigned long *offlined_pages = (unsigned long *)data; 1367 1368 *offlined_pages += __offline_isolated_pages(start, start + nr_pages); 1369 return 0; 1370 } 1371 1372 /* 1373 * Check all pages in range, recorded as memory resource, are isolated. 1374 */ 1375 static int 1376 check_pages_isolated_cb(unsigned long start_pfn, unsigned long nr_pages, 1377 void *data) 1378 { 1379 return test_pages_isolated(start_pfn, start_pfn + nr_pages, 1380 MEMORY_OFFLINE); 1381 } 1382 1383 static int __init cmdline_parse_movable_node(char *p) 1384 { 1385 movable_node_enabled = true; 1386 return 0; 1387 } 1388 early_param("movable_node", cmdline_parse_movable_node); 1389 1390 /* check which state of node_states will be changed when offline memory */ 1391 static void node_states_check_changes_offline(unsigned long nr_pages, 1392 struct zone *zone, struct memory_notify *arg) 1393 { 1394 struct pglist_data *pgdat = zone->zone_pgdat; 1395 unsigned long present_pages = 0; 1396 enum zone_type zt; 1397 1398 arg->status_change_nid = NUMA_NO_NODE; 1399 arg->status_change_nid_normal = NUMA_NO_NODE; 1400 arg->status_change_nid_high = NUMA_NO_NODE; 1401 1402 /* 1403 * Check whether node_states[N_NORMAL_MEMORY] will be changed. 1404 * If the memory to be offline is within the range 1405 * [0..ZONE_NORMAL], and it is the last present memory there, 1406 * the zones in that range will become empty after the offlining, 1407 * thus we can determine that we need to clear the node from 1408 * node_states[N_NORMAL_MEMORY]. 1409 */ 1410 for (zt = 0; zt <= ZONE_NORMAL; zt++) 1411 present_pages += pgdat->node_zones[zt].present_pages; 1412 if (zone_idx(zone) <= ZONE_NORMAL && nr_pages >= present_pages) 1413 arg->status_change_nid_normal = zone_to_nid(zone); 1414 1415 #ifdef CONFIG_HIGHMEM 1416 /* 1417 * node_states[N_HIGH_MEMORY] contains nodes which 1418 * have normal memory or high memory. 1419 * Here we add the present_pages belonging to ZONE_HIGHMEM. 1420 * If the zone is within the range of [0..ZONE_HIGHMEM), and 1421 * we determine that the zones in that range become empty, 1422 * we need to clear the node for N_HIGH_MEMORY. 1423 */ 1424 present_pages += pgdat->node_zones[ZONE_HIGHMEM].present_pages; 1425 if (zone_idx(zone) <= ZONE_HIGHMEM && nr_pages >= present_pages) 1426 arg->status_change_nid_high = zone_to_nid(zone); 1427 #endif 1428 1429 /* 1430 * We have accounted the pages from [0..ZONE_NORMAL), and 1431 * in case of CONFIG_HIGHMEM the pages from ZONE_HIGHMEM 1432 * as well. 1433 * Here we count the possible pages from ZONE_MOVABLE. 1434 * If after having accounted all the pages, we see that the nr_pages 1435 * to be offlined is over or equal to the accounted pages, 1436 * we know that the node will become empty, and so, we can clear 1437 * it for N_MEMORY as well. 1438 */ 1439 present_pages += pgdat->node_zones[ZONE_MOVABLE].present_pages; 1440 1441 if (nr_pages >= present_pages) 1442 arg->status_change_nid = zone_to_nid(zone); 1443 } 1444 1445 static void node_states_clear_node(int node, struct memory_notify *arg) 1446 { 1447 if (arg->status_change_nid_normal >= 0) 1448 node_clear_state(node, N_NORMAL_MEMORY); 1449 1450 if (arg->status_change_nid_high >= 0) 1451 node_clear_state(node, N_HIGH_MEMORY); 1452 1453 if (arg->status_change_nid >= 0) 1454 node_clear_state(node, N_MEMORY); 1455 } 1456 1457 static int count_system_ram_pages_cb(unsigned long start_pfn, 1458 unsigned long nr_pages, void *data) 1459 { 1460 unsigned long *nr_system_ram_pages = data; 1461 1462 *nr_system_ram_pages += nr_pages; 1463 return 0; 1464 } 1465 1466 static int __ref __offline_pages(unsigned long start_pfn, 1467 unsigned long end_pfn) 1468 { 1469 unsigned long pfn, nr_pages = 0; 1470 unsigned long offlined_pages = 0; 1471 int ret, node, nr_isolate_pageblock; 1472 unsigned long flags; 1473 struct zone *zone; 1474 struct memory_notify arg; 1475 char *reason; 1476 1477 mem_hotplug_begin(); 1478 1479 /* 1480 * Don't allow to offline memory blocks that contain holes. 1481 * Consequently, memory blocks with holes can never get onlined 1482 * via the hotplug path - online_pages() - as hotplugged memory has 1483 * no holes. This way, we e.g., don't have to worry about marking 1484 * memory holes PG_reserved, don't need pfn_valid() checks, and can 1485 * avoid using walk_system_ram_range() later. 1486 */ 1487 walk_system_ram_range(start_pfn, end_pfn - start_pfn, &nr_pages, 1488 count_system_ram_pages_cb); 1489 if (nr_pages != end_pfn - start_pfn) { 1490 ret = -EINVAL; 1491 reason = "memory holes"; 1492 goto failed_removal; 1493 } 1494 1495 /* This makes hotplug much easier...and readable. 1496 we assume this for now. .*/ 1497 zone = test_pages_in_a_zone(start_pfn, end_pfn); 1498 if (!zone) { 1499 ret = -EINVAL; 1500 reason = "multizone range"; 1501 goto failed_removal; 1502 } 1503 node = zone_to_nid(zone); 1504 1505 /* set above range as isolated */ 1506 ret = start_isolate_page_range(start_pfn, end_pfn, 1507 MIGRATE_MOVABLE, 1508 MEMORY_OFFLINE | REPORT_FAILURE); 1509 if (ret < 0) { 1510 reason = "failure to isolate range"; 1511 goto failed_removal; 1512 } 1513 nr_isolate_pageblock = ret; 1514 1515 arg.start_pfn = start_pfn; 1516 arg.nr_pages = nr_pages; 1517 node_states_check_changes_offline(nr_pages, zone, &arg); 1518 1519 ret = memory_notify(MEM_GOING_OFFLINE, &arg); 1520 ret = notifier_to_errno(ret); 1521 if (ret) { 1522 reason = "notifier failure"; 1523 goto failed_removal_isolated; 1524 } 1525 1526 do { 1527 pfn = start_pfn; 1528 do { 1529 if (signal_pending(current)) { 1530 ret = -EINTR; 1531 reason = "signal backoff"; 1532 goto failed_removal_isolated; 1533 } 1534 1535 cond_resched(); 1536 lru_add_drain_all(); 1537 1538 ret = scan_movable_pages(pfn, end_pfn, &pfn); 1539 if (!ret) { 1540 /* 1541 * TODO: fatal migration failures should bail 1542 * out 1543 */ 1544 do_migrate_range(pfn, end_pfn); 1545 } 1546 } while (!ret); 1547 1548 if (ret != -ENOENT) { 1549 reason = "unmovable page"; 1550 goto failed_removal_isolated; 1551 } 1552 1553 /* 1554 * Dissolve free hugepages in the memory block before doing 1555 * offlining actually in order to make hugetlbfs's object 1556 * counting consistent. 1557 */ 1558 ret = dissolve_free_huge_pages(start_pfn, end_pfn); 1559 if (ret) { 1560 reason = "failure to dissolve huge pages"; 1561 goto failed_removal_isolated; 1562 } 1563 /* check again */ 1564 ret = walk_system_ram_range(start_pfn, end_pfn - start_pfn, 1565 NULL, check_pages_isolated_cb); 1566 } while (ret); 1567 1568 /* Ok, all of our target is isolated. 1569 We cannot do rollback at this point. */ 1570 walk_system_ram_range(start_pfn, end_pfn - start_pfn, 1571 &offlined_pages, offline_isolated_pages_cb); 1572 pr_info("Offlined Pages %ld\n", offlined_pages); 1573 /* 1574 * Onlining will reset pagetype flags and makes migrate type 1575 * MOVABLE, so just need to decrease the number of isolated 1576 * pageblocks zone counter here. 1577 */ 1578 spin_lock_irqsave(&zone->lock, flags); 1579 zone->nr_isolate_pageblock -= nr_isolate_pageblock; 1580 spin_unlock_irqrestore(&zone->lock, flags); 1581 1582 /* removal success */ 1583 adjust_managed_page_count(pfn_to_page(start_pfn), -offlined_pages); 1584 zone->present_pages -= offlined_pages; 1585 1586 pgdat_resize_lock(zone->zone_pgdat, &flags); 1587 zone->zone_pgdat->node_present_pages -= offlined_pages; 1588 pgdat_resize_unlock(zone->zone_pgdat, &flags); 1589 1590 init_per_zone_wmark_min(); 1591 1592 if (!populated_zone(zone)) { 1593 zone_pcp_reset(zone); 1594 build_all_zonelists(NULL); 1595 } else 1596 zone_pcp_update(zone); 1597 1598 node_states_clear_node(node, &arg); 1599 if (arg.status_change_nid >= 0) { 1600 kswapd_stop(node); 1601 kcompactd_stop(node); 1602 } 1603 1604 writeback_set_ratelimit(); 1605 1606 memory_notify(MEM_OFFLINE, &arg); 1607 remove_pfn_range_from_zone(zone, start_pfn, nr_pages); 1608 mem_hotplug_done(); 1609 return 0; 1610 1611 failed_removal_isolated: 1612 undo_isolate_page_range(start_pfn, end_pfn, MIGRATE_MOVABLE); 1613 memory_notify(MEM_CANCEL_OFFLINE, &arg); 1614 failed_removal: 1615 pr_debug("memory offlining [mem %#010llx-%#010llx] failed due to %s\n", 1616 (unsigned long long) start_pfn << PAGE_SHIFT, 1617 ((unsigned long long) end_pfn << PAGE_SHIFT) - 1, 1618 reason); 1619 /* pushback to free area */ 1620 mem_hotplug_done(); 1621 return ret; 1622 } 1623 1624 int offline_pages(unsigned long start_pfn, unsigned long nr_pages) 1625 { 1626 return __offline_pages(start_pfn, start_pfn + nr_pages); 1627 } 1628 1629 static int check_memblock_offlined_cb(struct memory_block *mem, void *arg) 1630 { 1631 int ret = !is_memblock_offlined(mem); 1632 1633 if (unlikely(ret)) { 1634 phys_addr_t beginpa, endpa; 1635 1636 beginpa = PFN_PHYS(section_nr_to_pfn(mem->start_section_nr)); 1637 endpa = beginpa + memory_block_size_bytes() - 1; 1638 pr_warn("removing memory fails, because memory [%pa-%pa] is onlined\n", 1639 &beginpa, &endpa); 1640 1641 return -EBUSY; 1642 } 1643 return 0; 1644 } 1645 1646 static int check_cpu_on_node(pg_data_t *pgdat) 1647 { 1648 int cpu; 1649 1650 for_each_present_cpu(cpu) { 1651 if (cpu_to_node(cpu) == pgdat->node_id) 1652 /* 1653 * the cpu on this node isn't removed, and we can't 1654 * offline this node. 1655 */ 1656 return -EBUSY; 1657 } 1658 1659 return 0; 1660 } 1661 1662 static int check_no_memblock_for_node_cb(struct memory_block *mem, void *arg) 1663 { 1664 int nid = *(int *)arg; 1665 1666 /* 1667 * If a memory block belongs to multiple nodes, the stored nid is not 1668 * reliable. However, such blocks are always online (e.g., cannot get 1669 * offlined) and, therefore, are still spanned by the node. 1670 */ 1671 return mem->nid == nid ? -EEXIST : 0; 1672 } 1673 1674 /** 1675 * try_offline_node 1676 * @nid: the node ID 1677 * 1678 * Offline a node if all memory sections and cpus of the node are removed. 1679 * 1680 * NOTE: The caller must call lock_device_hotplug() to serialize hotplug 1681 * and online/offline operations before this call. 1682 */ 1683 void try_offline_node(int nid) 1684 { 1685 pg_data_t *pgdat = NODE_DATA(nid); 1686 int rc; 1687 1688 /* 1689 * If the node still spans pages (especially ZONE_DEVICE), don't 1690 * offline it. A node spans memory after move_pfn_range_to_zone(), 1691 * e.g., after the memory block was onlined. 1692 */ 1693 if (pgdat->node_spanned_pages) 1694 return; 1695 1696 /* 1697 * Especially offline memory blocks might not be spanned by the 1698 * node. They will get spanned by the node once they get onlined. 1699 * However, they link to the node in sysfs and can get onlined later. 1700 */ 1701 rc = for_each_memory_block(&nid, check_no_memblock_for_node_cb); 1702 if (rc) 1703 return; 1704 1705 if (check_cpu_on_node(pgdat)) 1706 return; 1707 1708 /* 1709 * all memory/cpu of this node are removed, we can offline this 1710 * node now. 1711 */ 1712 node_set_offline(nid); 1713 unregister_one_node(nid); 1714 } 1715 EXPORT_SYMBOL(try_offline_node); 1716 1717 static void __release_memory_resource(resource_size_t start, 1718 resource_size_t size) 1719 { 1720 int ret; 1721 1722 /* 1723 * When removing memory in the same granularity as it was added, 1724 * this function never fails. It might only fail if resources 1725 * have to be adjusted or split. We'll ignore the error, as 1726 * removing of memory cannot fail. 1727 */ 1728 ret = release_mem_region_adjustable(&iomem_resource, start, size); 1729 if (ret) { 1730 resource_size_t endres = start + size - 1; 1731 1732 pr_warn("Unable to release resource <%pa-%pa> (%d)\n", 1733 &start, &endres, ret); 1734 } 1735 } 1736 1737 static int __ref try_remove_memory(int nid, u64 start, u64 size) 1738 { 1739 int rc = 0; 1740 1741 BUG_ON(check_hotplug_memory_range(start, size)); 1742 1743 /* 1744 * All memory blocks must be offlined before removing memory. Check 1745 * whether all memory blocks in question are offline and return error 1746 * if this is not the case. 1747 */ 1748 rc = walk_memory_blocks(start, size, NULL, check_memblock_offlined_cb); 1749 if (rc) 1750 goto done; 1751 1752 /* remove memmap entry */ 1753 firmware_map_remove(start, start + size, "System RAM"); 1754 1755 /* 1756 * Memory block device removal under the device_hotplug_lock is 1757 * a barrier against racing online attempts. 1758 */ 1759 remove_memory_block_devices(start, size); 1760 1761 mem_hotplug_begin(); 1762 1763 arch_remove_memory(nid, start, size, NULL); 1764 1765 if (IS_ENABLED(CONFIG_ARCH_KEEP_MEMBLOCK)) { 1766 memblock_free(start, size); 1767 memblock_remove(start, size); 1768 } 1769 1770 __release_memory_resource(start, size); 1771 1772 try_offline_node(nid); 1773 1774 done: 1775 mem_hotplug_done(); 1776 return rc; 1777 } 1778 1779 /** 1780 * remove_memory 1781 * @nid: the node ID 1782 * @start: physical address of the region to remove 1783 * @size: size of the region to remove 1784 * 1785 * NOTE: The caller must call lock_device_hotplug() to serialize hotplug 1786 * and online/offline operations before this call, as required by 1787 * try_offline_node(). 1788 */ 1789 void __remove_memory(int nid, u64 start, u64 size) 1790 { 1791 1792 /* 1793 * trigger BUG() if some memory is not offlined prior to calling this 1794 * function 1795 */ 1796 if (try_remove_memory(nid, start, size)) 1797 BUG(); 1798 } 1799 1800 /* 1801 * Remove memory if every memory block is offline, otherwise return -EBUSY is 1802 * some memory is not offline 1803 */ 1804 int remove_memory(int nid, u64 start, u64 size) 1805 { 1806 int rc; 1807 1808 lock_device_hotplug(); 1809 rc = try_remove_memory(nid, start, size); 1810 unlock_device_hotplug(); 1811 1812 return rc; 1813 } 1814 EXPORT_SYMBOL_GPL(remove_memory); 1815 1816 /* 1817 * Try to offline and remove a memory block. Might take a long time to 1818 * finish in case memory is still in use. Primarily useful for memory devices 1819 * that logically unplugged all memory (so it's no longer in use) and want to 1820 * offline + remove the memory block. 1821 */ 1822 int offline_and_remove_memory(int nid, u64 start, u64 size) 1823 { 1824 struct memory_block *mem; 1825 int rc = -EINVAL; 1826 1827 if (!IS_ALIGNED(start, memory_block_size_bytes()) || 1828 size != memory_block_size_bytes()) 1829 return rc; 1830 1831 lock_device_hotplug(); 1832 mem = find_memory_block(__pfn_to_section(PFN_DOWN(start))); 1833 if (mem) 1834 rc = device_offline(&mem->dev); 1835 /* Ignore if the device is already offline. */ 1836 if (rc > 0) 1837 rc = 0; 1838 1839 /* 1840 * In case we succeeded to offline the memory block, remove it. 1841 * This cannot fail as it cannot get onlined in the meantime. 1842 */ 1843 if (!rc) { 1844 rc = try_remove_memory(nid, start, size); 1845 WARN_ON_ONCE(rc); 1846 } 1847 unlock_device_hotplug(); 1848 1849 return rc; 1850 } 1851 EXPORT_SYMBOL_GPL(offline_and_remove_memory); 1852 #endif /* CONFIG_MEMORY_HOTREMOVE */ 1853