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/writeback.h> 17#include <linux/slab.h> 18#include <linux/sysctl.h> 19#include <linux/cpu.h> 20#include <linux/memory.h> 21#include <linux/memremap.h> 22#include <linux/memory_hotplug.h> 23#include <linux/vmalloc.h> 24#include <linux/ioport.h> 25#include <linux/delay.h> 26#include <linux/migrate.h> 27#include <linux/page-isolation.h> 28#include <linux/pfn.h> 29#include <linux/suspend.h> 30#include <linux/mm_inline.h> 31#include <linux/firmware-map.h> 32#include <linux/stop_machine.h> 33#include <linux/hugetlb.h> 34#include <linux/memblock.h> 35#include <linux/compaction.h> 36#include <linux/rmap.h> 37#include <linux/module.h> 38 39#include <asm/tlbflush.h> 40 41#include "internal.h" 42#include "shuffle.h" 43 44enum { 45 MEMMAP_ON_MEMORY_DISABLE = 0, 46 MEMMAP_ON_MEMORY_ENABLE, 47 MEMMAP_ON_MEMORY_FORCE, 48}; 49 50static int memmap_mode __read_mostly = MEMMAP_ON_MEMORY_DISABLE; 51 52static inline unsigned long memory_block_memmap_size(void) 53{ 54 return PHYS_PFN(memory_block_size_bytes()) * sizeof(struct page); 55} 56 57static inline unsigned long memory_block_memmap_on_memory_pages(void) 58{ 59 unsigned long nr_pages = PFN_UP(memory_block_memmap_size()); 60 61 /* 62 * In "forced" memmap_on_memory mode, we add extra pages to align the 63 * vmemmap size to cover full pageblocks. That way, we can add memory 64 * even if the vmemmap size is not properly aligned, however, we might waste 65 * memory. 66 */ 67 if (memmap_mode == MEMMAP_ON_MEMORY_FORCE) 68 return pageblock_align(nr_pages); 69 return nr_pages; 70} 71 72#ifdef CONFIG_MHP_MEMMAP_ON_MEMORY 73/* 74 * memory_hotplug.memmap_on_memory parameter 75 */ 76static int set_memmap_mode(const char *val, const struct kernel_param *kp) 77{ 78 int ret, mode; 79 bool enabled; 80 81 if (sysfs_streq(val, "force") || sysfs_streq(val, "FORCE")) { 82 mode = MEMMAP_ON_MEMORY_FORCE; 83 } else { 84 ret = kstrtobool(val, &enabled); 85 if (ret < 0) 86 return ret; 87 if (enabled) 88 mode = MEMMAP_ON_MEMORY_ENABLE; 89 else 90 mode = MEMMAP_ON_MEMORY_DISABLE; 91 } 92 *((int *)kp->arg) = mode; 93 if (mode == MEMMAP_ON_MEMORY_FORCE) { 94 unsigned long memmap_pages = memory_block_memmap_on_memory_pages(); 95 96 pr_info_once("Memory hotplug will waste %ld pages in each memory block\n", 97 memmap_pages - PFN_UP(memory_block_memmap_size())); 98 } 99 return 0; 100} 101 102static int get_memmap_mode(char *buffer, const struct kernel_param *kp) 103{ 104 int mode = *((int *)kp->arg); 105 106 if (mode == MEMMAP_ON_MEMORY_FORCE) 107 return sprintf(buffer, "force\n"); 108 return sprintf(buffer, "%c\n", mode ? 'Y' : 'N'); 109} 110 111static const struct kernel_param_ops memmap_mode_ops = { 112 .set = set_memmap_mode, 113 .get = get_memmap_mode, 114}; 115module_param_cb(memmap_on_memory, &memmap_mode_ops, &memmap_mode, 0444); 116MODULE_PARM_DESC(memmap_on_memory, "Enable memmap on memory for memory hotplug\n" 117 "With value \"force\" it could result in memory wastage due " 118 "to memmap size limitations (Y/N/force)"); 119 120static inline bool mhp_memmap_on_memory(void) 121{ 122 return memmap_mode != MEMMAP_ON_MEMORY_DISABLE; 123} 124#else 125static inline bool mhp_memmap_on_memory(void) 126{ 127 return false; 128} 129#endif 130 131enum { 132 ONLINE_POLICY_CONTIG_ZONES = 0, 133 ONLINE_POLICY_AUTO_MOVABLE, 134}; 135 136static const char * const online_policy_to_str[] = { 137 [ONLINE_POLICY_CONTIG_ZONES] = "contig-zones", 138 [ONLINE_POLICY_AUTO_MOVABLE] = "auto-movable", 139}; 140 141static int set_online_policy(const char *val, const struct kernel_param *kp) 142{ 143 int ret = sysfs_match_string(online_policy_to_str, val); 144 145 if (ret < 0) 146 return ret; 147 *((int *)kp->arg) = ret; 148 return 0; 149} 150 151static int get_online_policy(char *buffer, const struct kernel_param *kp) 152{ 153 return sprintf(buffer, "%s\n", online_policy_to_str[*((int *)kp->arg)]); 154} 155 156/* 157 * memory_hotplug.online_policy: configure online behavior when onlining without 158 * specifying a zone (MMOP_ONLINE) 159 * 160 * "contig-zones": keep zone contiguous 161 * "auto-movable": online memory to ZONE_MOVABLE if the configuration 162 * (auto_movable_ratio, auto_movable_numa_aware) allows for it 163 */ 164static int online_policy __read_mostly = ONLINE_POLICY_CONTIG_ZONES; 165static const struct kernel_param_ops online_policy_ops = { 166 .set = set_online_policy, 167 .get = get_online_policy, 168}; 169module_param_cb(online_policy, &online_policy_ops, &online_policy, 0644); 170MODULE_PARM_DESC(online_policy, 171 "Set the online policy (\"contig-zones\", \"auto-movable\") " 172 "Default: \"contig-zones\""); 173 174/* 175 * memory_hotplug.auto_movable_ratio: specify maximum MOVABLE:KERNEL ratio 176 * 177 * The ratio represent an upper limit and the kernel might decide to not 178 * online some memory to ZONE_MOVABLE -- e.g., because hotplugged KERNEL memory 179 * doesn't allow for more MOVABLE memory. 180 */ 181static unsigned int auto_movable_ratio __read_mostly = 301; 182module_param(auto_movable_ratio, uint, 0644); 183MODULE_PARM_DESC(auto_movable_ratio, 184 "Set the maximum ratio of MOVABLE:KERNEL memory in the system " 185 "in percent for \"auto-movable\" online policy. Default: 301"); 186 187/* 188 * memory_hotplug.auto_movable_numa_aware: consider numa node stats 189 */ 190#ifdef CONFIG_NUMA 191static bool auto_movable_numa_aware __read_mostly = true; 192module_param(auto_movable_numa_aware, bool, 0644); 193MODULE_PARM_DESC(auto_movable_numa_aware, 194 "Consider numa node stats in addition to global stats in " 195 "\"auto-movable\" online policy. Default: true"); 196#endif /* CONFIG_NUMA */ 197 198/* 199 * online_page_callback contains pointer to current page onlining function. 200 * Initially it is generic_online_page(). If it is required it could be 201 * changed by calling set_online_page_callback() for callback registration 202 * and restore_online_page_callback() for generic callback restore. 203 */ 204 205static online_page_callback_t online_page_callback = generic_online_page; 206static DEFINE_MUTEX(online_page_callback_lock); 207 208DEFINE_STATIC_PERCPU_RWSEM(mem_hotplug_lock); 209 210void get_online_mems(void) 211{ 212 percpu_down_read(&mem_hotplug_lock); 213} 214 215void put_online_mems(void) 216{ 217 percpu_up_read(&mem_hotplug_lock); 218} 219 220bool movable_node_enabled = false; 221 222#ifndef CONFIG_MEMORY_HOTPLUG_DEFAULT_ONLINE 223int mhp_default_online_type = MMOP_OFFLINE; 224#else 225int mhp_default_online_type = MMOP_ONLINE; 226#endif 227 228static int __init setup_memhp_default_state(char *str) 229{ 230 const int online_type = mhp_online_type_from_str(str); 231 232 if (online_type >= 0) 233 mhp_default_online_type = online_type; 234 235 return 1; 236} 237__setup("memhp_default_state=", setup_memhp_default_state); 238 239void mem_hotplug_begin(void) 240{ 241 cpus_read_lock(); 242 percpu_down_write(&mem_hotplug_lock); 243} 244 245void mem_hotplug_done(void) 246{ 247 percpu_up_write(&mem_hotplug_lock); 248 cpus_read_unlock(); 249} 250 251u64 max_mem_size = U64_MAX; 252 253/* add this memory to iomem resource */ 254static struct resource *register_memory_resource(u64 start, u64 size, 255 const char *resource_name) 256{ 257 struct resource *res; 258 unsigned long flags = IORESOURCE_SYSTEM_RAM | IORESOURCE_BUSY; 259 260 if (strcmp(resource_name, "System RAM")) 261 flags |= IORESOURCE_SYSRAM_DRIVER_MANAGED; 262 263 if (!mhp_range_allowed(start, size, true)) 264 return ERR_PTR(-E2BIG); 265 266 /* 267 * Make sure value parsed from 'mem=' only restricts memory adding 268 * while booting, so that memory hotplug won't be impacted. Please 269 * refer to document of 'mem=' in kernel-parameters.txt for more 270 * details. 271 */ 272 if (start + size > max_mem_size && system_state < SYSTEM_RUNNING) 273 return ERR_PTR(-E2BIG); 274 275 /* 276 * Request ownership of the new memory range. This might be 277 * a child of an existing resource that was present but 278 * not marked as busy. 279 */ 280 res = __request_region(&iomem_resource, start, size, 281 resource_name, flags); 282 283 if (!res) { 284 pr_debug("Unable to reserve System RAM region: %016llx->%016llx\n", 285 start, start + size); 286 return ERR_PTR(-EEXIST); 287 } 288 return res; 289} 290 291static void release_memory_resource(struct resource *res) 292{ 293 if (!res) 294 return; 295 release_resource(res); 296 kfree(res); 297} 298 299static int check_pfn_span(unsigned long pfn, unsigned long nr_pages) 300{ 301 /* 302 * Disallow all operations smaller than a sub-section and only 303 * allow operations smaller than a section for 304 * SPARSEMEM_VMEMMAP. Note that check_hotplug_memory_range() 305 * enforces a larger memory_block_size_bytes() granularity for 306 * memory that will be marked online, so this check should only 307 * fire for direct arch_{add,remove}_memory() users outside of 308 * add_memory_resource(). 309 */ 310 unsigned long min_align; 311 312 if (IS_ENABLED(CONFIG_SPARSEMEM_VMEMMAP)) 313 min_align = PAGES_PER_SUBSECTION; 314 else 315 min_align = PAGES_PER_SECTION; 316 if (!IS_ALIGNED(pfn | nr_pages, min_align)) 317 return -EINVAL; 318 return 0; 319} 320 321/* 322 * Return page for the valid pfn only if the page is online. All pfn 323 * walkers which rely on the fully initialized page->flags and others 324 * should use this rather than pfn_valid && pfn_to_page 325 */ 326struct page *pfn_to_online_page(unsigned long pfn) 327{ 328 unsigned long nr = pfn_to_section_nr(pfn); 329 struct dev_pagemap *pgmap; 330 struct mem_section *ms; 331 332 if (nr >= NR_MEM_SECTIONS) 333 return NULL; 334 335 ms = __nr_to_section(nr); 336 if (!online_section(ms)) 337 return NULL; 338 339 /* 340 * Save some code text when online_section() + 341 * pfn_section_valid() are sufficient. 342 */ 343 if (IS_ENABLED(CONFIG_HAVE_ARCH_PFN_VALID) && !pfn_valid(pfn)) 344 return NULL; 345 346 if (!pfn_section_valid(ms, pfn)) 347 return NULL; 348 349 if (!online_device_section(ms)) 350 return pfn_to_page(pfn); 351 352 /* 353 * Slowpath: when ZONE_DEVICE collides with 354 * ZONE_{NORMAL,MOVABLE} within the same section some pfns in 355 * the section may be 'offline' but 'valid'. Only 356 * get_dev_pagemap() can determine sub-section online status. 357 */ 358 pgmap = get_dev_pagemap(pfn, NULL); 359 put_dev_pagemap(pgmap); 360 361 /* The presence of a pgmap indicates ZONE_DEVICE offline pfn */ 362 if (pgmap) 363 return NULL; 364 365 return pfn_to_page(pfn); 366} 367EXPORT_SYMBOL_GPL(pfn_to_online_page); 368 369int __ref __add_pages(int nid, unsigned long pfn, unsigned long nr_pages, 370 struct mhp_params *params) 371{ 372 const unsigned long end_pfn = pfn + nr_pages; 373 unsigned long cur_nr_pages; 374 int err; 375 struct vmem_altmap *altmap = params->altmap; 376 377 if (WARN_ON_ONCE(!pgprot_val(params->pgprot))) 378 return -EINVAL; 379 380 VM_BUG_ON(!mhp_range_allowed(PFN_PHYS(pfn), nr_pages * PAGE_SIZE, false)); 381 382 if (altmap) { 383 /* 384 * Validate altmap is within bounds of the total request 385 */ 386 if (altmap->base_pfn != pfn 387 || vmem_altmap_offset(altmap) > nr_pages) { 388 pr_warn_once("memory add fail, invalid altmap\n"); 389 return -EINVAL; 390 } 391 altmap->alloc = 0; 392 } 393 394 if (check_pfn_span(pfn, nr_pages)) { 395 WARN(1, "Misaligned %s start: %#lx end: %#lx\n", __func__, pfn, pfn + nr_pages - 1); 396 return -EINVAL; 397 } 398 399 for (; pfn < end_pfn; pfn += cur_nr_pages) { 400 /* Select all remaining pages up to the next section boundary */ 401 cur_nr_pages = min(end_pfn - pfn, 402 SECTION_ALIGN_UP(pfn + 1) - pfn); 403 err = sparse_add_section(nid, pfn, cur_nr_pages, altmap, 404 params->pgmap); 405 if (err) 406 break; 407 cond_resched(); 408 } 409 vmemmap_populate_print_last(); 410 return err; 411} 412 413/* find the smallest valid pfn in the range [start_pfn, end_pfn) */ 414static unsigned long find_smallest_section_pfn(int nid, struct zone *zone, 415 unsigned long start_pfn, 416 unsigned long end_pfn) 417{ 418 for (; start_pfn < end_pfn; start_pfn += PAGES_PER_SUBSECTION) { 419 if (unlikely(!pfn_to_online_page(start_pfn))) 420 continue; 421 422 if (unlikely(pfn_to_nid(start_pfn) != nid)) 423 continue; 424 425 if (zone != page_zone(pfn_to_page(start_pfn))) 426 continue; 427 428 return start_pfn; 429 } 430 431 return 0; 432} 433 434/* find the biggest valid pfn in the range [start_pfn, end_pfn). */ 435static unsigned long find_biggest_section_pfn(int nid, struct zone *zone, 436 unsigned long start_pfn, 437 unsigned long end_pfn) 438{ 439 unsigned long pfn; 440 441 /* pfn is the end pfn of a memory section. */ 442 pfn = end_pfn - 1; 443 for (; pfn >= start_pfn; pfn -= PAGES_PER_SUBSECTION) { 444 if (unlikely(!pfn_to_online_page(pfn))) 445 continue; 446 447 if (unlikely(pfn_to_nid(pfn) != nid)) 448 continue; 449 450 if (zone != page_zone(pfn_to_page(pfn))) 451 continue; 452 453 return pfn; 454 } 455 456 return 0; 457} 458 459static void shrink_zone_span(struct zone *zone, unsigned long start_pfn, 460 unsigned long end_pfn) 461{ 462 unsigned long pfn; 463 int nid = zone_to_nid(zone); 464 465 if (zone->zone_start_pfn == start_pfn) { 466 /* 467 * If the section is smallest section in the zone, it need 468 * shrink zone->zone_start_pfn and zone->zone_spanned_pages. 469 * In this case, we find second smallest valid mem_section 470 * for shrinking zone. 471 */ 472 pfn = find_smallest_section_pfn(nid, zone, end_pfn, 473 zone_end_pfn(zone)); 474 if (pfn) { 475 zone->spanned_pages = zone_end_pfn(zone) - pfn; 476 zone->zone_start_pfn = pfn; 477 } else { 478 zone->zone_start_pfn = 0; 479 zone->spanned_pages = 0; 480 } 481 } else if (zone_end_pfn(zone) == end_pfn) { 482 /* 483 * If the section is biggest section in the zone, it need 484 * shrink zone->spanned_pages. 485 * In this case, we find second biggest valid mem_section for 486 * shrinking zone. 487 */ 488 pfn = find_biggest_section_pfn(nid, zone, zone->zone_start_pfn, 489 start_pfn); 490 if (pfn) 491 zone->spanned_pages = pfn - zone->zone_start_pfn + 1; 492 else { 493 zone->zone_start_pfn = 0; 494 zone->spanned_pages = 0; 495 } 496 } 497} 498 499static void update_pgdat_span(struct pglist_data *pgdat) 500{ 501 unsigned long node_start_pfn = 0, node_end_pfn = 0; 502 struct zone *zone; 503 504 for (zone = pgdat->node_zones; 505 zone < pgdat->node_zones + MAX_NR_ZONES; zone++) { 506 unsigned long end_pfn = zone_end_pfn(zone); 507 508 /* No need to lock the zones, they can't change. */ 509 if (!zone->spanned_pages) 510 continue; 511 if (!node_end_pfn) { 512 node_start_pfn = zone->zone_start_pfn; 513 node_end_pfn = end_pfn; 514 continue; 515 } 516 517 if (end_pfn > node_end_pfn) 518 node_end_pfn = end_pfn; 519 if (zone->zone_start_pfn < node_start_pfn) 520 node_start_pfn = zone->zone_start_pfn; 521 } 522 523 pgdat->node_start_pfn = node_start_pfn; 524 pgdat->node_spanned_pages = node_end_pfn - node_start_pfn; 525} 526 527void __ref remove_pfn_range_from_zone(struct zone *zone, 528 unsigned long start_pfn, 529 unsigned long nr_pages) 530{ 531 const unsigned long end_pfn = start_pfn + nr_pages; 532 struct pglist_data *pgdat = zone->zone_pgdat; 533 unsigned long pfn, cur_nr_pages; 534 535 /* Poison struct pages because they are now uninitialized again. */ 536 for (pfn = start_pfn; pfn < end_pfn; pfn += cur_nr_pages) { 537 cond_resched(); 538 539 /* Select all remaining pages up to the next section boundary */ 540 cur_nr_pages = 541 min(end_pfn - pfn, SECTION_ALIGN_UP(pfn + 1) - pfn); 542 page_init_poison(pfn_to_page(pfn), 543 sizeof(struct page) * cur_nr_pages); 544 } 545 546 /* 547 * Zone shrinking code cannot properly deal with ZONE_DEVICE. So 548 * we will not try to shrink the zones - which is okay as 549 * set_zone_contiguous() cannot deal with ZONE_DEVICE either way. 550 */ 551 if (zone_is_zone_device(zone)) 552 return; 553 554 clear_zone_contiguous(zone); 555 556 shrink_zone_span(zone, start_pfn, start_pfn + nr_pages); 557 update_pgdat_span(pgdat); 558 559 set_zone_contiguous(zone); 560} 561 562/** 563 * __remove_pages() - remove sections of pages 564 * @pfn: starting pageframe (must be aligned to start of a section) 565 * @nr_pages: number of pages to remove (must be multiple of section size) 566 * @altmap: alternative device page map or %NULL if default memmap is used 567 * 568 * Generic helper function to remove section mappings and sysfs entries 569 * for the section of the memory we are removing. Caller needs to make 570 * sure that pages are marked reserved and zones are adjust properly by 571 * calling offline_pages(). 572 */ 573void __remove_pages(unsigned long pfn, unsigned long nr_pages, 574 struct vmem_altmap *altmap) 575{ 576 const unsigned long end_pfn = pfn + nr_pages; 577 unsigned long cur_nr_pages; 578 579 if (check_pfn_span(pfn, nr_pages)) { 580 WARN(1, "Misaligned %s start: %#lx end: %#lx\n", __func__, pfn, pfn + nr_pages - 1); 581 return; 582 } 583 584 for (; pfn < end_pfn; pfn += cur_nr_pages) { 585 cond_resched(); 586 /* Select all remaining pages up to the next section boundary */ 587 cur_nr_pages = min(end_pfn - pfn, 588 SECTION_ALIGN_UP(pfn + 1) - pfn); 589 sparse_remove_section(pfn, cur_nr_pages, altmap); 590 } 591} 592 593int set_online_page_callback(online_page_callback_t callback) 594{ 595 int rc = -EINVAL; 596 597 get_online_mems(); 598 mutex_lock(&online_page_callback_lock); 599 600 if (online_page_callback == generic_online_page) { 601 online_page_callback = callback; 602 rc = 0; 603 } 604 605 mutex_unlock(&online_page_callback_lock); 606 put_online_mems(); 607 608 return rc; 609} 610EXPORT_SYMBOL_GPL(set_online_page_callback); 611 612int restore_online_page_callback(online_page_callback_t callback) 613{ 614 int rc = -EINVAL; 615 616 get_online_mems(); 617 mutex_lock(&online_page_callback_lock); 618 619 if (online_page_callback == callback) { 620 online_page_callback = generic_online_page; 621 rc = 0; 622 } 623 624 mutex_unlock(&online_page_callback_lock); 625 put_online_mems(); 626 627 return rc; 628} 629EXPORT_SYMBOL_GPL(restore_online_page_callback); 630 631void generic_online_page(struct page *page, unsigned int order) 632{ 633 /* 634 * Freeing the page with debug_pagealloc enabled will try to unmap it, 635 * so we should map it first. This is better than introducing a special 636 * case in page freeing fast path. 637 */ 638 debug_pagealloc_map_pages(page, 1 << order); 639 __free_pages_core(page, order); 640 totalram_pages_add(1UL << order); 641} 642EXPORT_SYMBOL_GPL(generic_online_page); 643 644static void online_pages_range(unsigned long start_pfn, unsigned long nr_pages) 645{ 646 const unsigned long end_pfn = start_pfn + nr_pages; 647 unsigned long pfn; 648 649 /* 650 * Online the pages in MAX_PAGE_ORDER aligned chunks. The callback might 651 * decide to not expose all pages to the buddy (e.g., expose them 652 * later). We account all pages as being online and belonging to this 653 * zone ("present"). 654 * When using memmap_on_memory, the range might not be aligned to 655 * MAX_ORDER_NR_PAGES - 1, but pageblock aligned. __ffs() will detect 656 * this and the first chunk to online will be pageblock_nr_pages. 657 */ 658 for (pfn = start_pfn; pfn < end_pfn;) { 659 int order; 660 661 /* 662 * Free to online pages in the largest chunks alignment allows. 663 * 664 * __ffs() behaviour is undefined for 0. start == 0 is 665 * MAX_PAGE_ORDER-aligned, Set order to MAX_PAGE_ORDER for 666 * the case. 667 */ 668 if (pfn) 669 order = min_t(int, MAX_PAGE_ORDER, __ffs(pfn)); 670 else 671 order = MAX_PAGE_ORDER; 672 673 (*online_page_callback)(pfn_to_page(pfn), order); 674 pfn += (1UL << order); 675 } 676 677 /* mark all involved sections as online */ 678 online_mem_sections(start_pfn, end_pfn); 679} 680 681/* check which state of node_states will be changed when online memory */ 682static void node_states_check_changes_online(unsigned long nr_pages, 683 struct zone *zone, struct memory_notify *arg) 684{ 685 int nid = zone_to_nid(zone); 686 687 arg->status_change_nid = NUMA_NO_NODE; 688 arg->status_change_nid_normal = NUMA_NO_NODE; 689 690 if (!node_state(nid, N_MEMORY)) 691 arg->status_change_nid = nid; 692 if (zone_idx(zone) <= ZONE_NORMAL && !node_state(nid, N_NORMAL_MEMORY)) 693 arg->status_change_nid_normal = nid; 694} 695 696static void node_states_set_node(int node, struct memory_notify *arg) 697{ 698 if (arg->status_change_nid_normal >= 0) 699 node_set_state(node, N_NORMAL_MEMORY); 700 701 if (arg->status_change_nid >= 0) 702 node_set_state(node, N_MEMORY); 703} 704 705static void __meminit resize_zone_range(struct zone *zone, unsigned long start_pfn, 706 unsigned long nr_pages) 707{ 708 unsigned long old_end_pfn = zone_end_pfn(zone); 709 710 if (zone_is_empty(zone) || start_pfn < zone->zone_start_pfn) 711 zone->zone_start_pfn = start_pfn; 712 713 zone->spanned_pages = max(start_pfn + nr_pages, old_end_pfn) - zone->zone_start_pfn; 714} 715 716static void __meminit resize_pgdat_range(struct pglist_data *pgdat, unsigned long start_pfn, 717 unsigned long nr_pages) 718{ 719 unsigned long old_end_pfn = pgdat_end_pfn(pgdat); 720 721 if (!pgdat->node_spanned_pages || start_pfn < pgdat->node_start_pfn) 722 pgdat->node_start_pfn = start_pfn; 723 724 pgdat->node_spanned_pages = max(start_pfn + nr_pages, old_end_pfn) - pgdat->node_start_pfn; 725 726} 727 728#ifdef CONFIG_ZONE_DEVICE 729static void section_taint_zone_device(unsigned long pfn) 730{ 731 struct mem_section *ms = __pfn_to_section(pfn); 732 733 ms->section_mem_map |= SECTION_TAINT_ZONE_DEVICE; 734} 735#else 736static inline void section_taint_zone_device(unsigned long pfn) 737{ 738} 739#endif 740 741/* 742 * Associate the pfn range with the given zone, initializing the memmaps 743 * and resizing the pgdat/zone data to span the added pages. After this 744 * call, all affected pages are PG_reserved. 745 * 746 * All aligned pageblocks are initialized to the specified migratetype 747 * (usually MIGRATE_MOVABLE). Besides setting the migratetype, no related 748 * zone stats (e.g., nr_isolate_pageblock) are touched. 749 */ 750void __ref move_pfn_range_to_zone(struct zone *zone, unsigned long start_pfn, 751 unsigned long nr_pages, 752 struct vmem_altmap *altmap, int migratetype) 753{ 754 struct pglist_data *pgdat = zone->zone_pgdat; 755 int nid = pgdat->node_id; 756 757 clear_zone_contiguous(zone); 758 759 if (zone_is_empty(zone)) 760 init_currently_empty_zone(zone, start_pfn, nr_pages); 761 resize_zone_range(zone, start_pfn, nr_pages); 762 resize_pgdat_range(pgdat, start_pfn, nr_pages); 763 764 /* 765 * Subsection population requires care in pfn_to_online_page(). 766 * Set the taint to enable the slow path detection of 767 * ZONE_DEVICE pages in an otherwise ZONE_{NORMAL,MOVABLE} 768 * section. 769 */ 770 if (zone_is_zone_device(zone)) { 771 if (!IS_ALIGNED(start_pfn, PAGES_PER_SECTION)) 772 section_taint_zone_device(start_pfn); 773 if (!IS_ALIGNED(start_pfn + nr_pages, PAGES_PER_SECTION)) 774 section_taint_zone_device(start_pfn + nr_pages); 775 } 776 777 /* 778 * TODO now we have a visible range of pages which are not associated 779 * with their zone properly. Not nice but set_pfnblock_flags_mask 780 * expects the zone spans the pfn range. All the pages in the range 781 * are reserved so nobody should be touching them so we should be safe 782 */ 783 memmap_init_range(nr_pages, nid, zone_idx(zone), start_pfn, 0, 784 MEMINIT_HOTPLUG, altmap, migratetype); 785 786 set_zone_contiguous(zone); 787} 788 789struct auto_movable_stats { 790 unsigned long kernel_early_pages; 791 unsigned long movable_pages; 792}; 793 794static void auto_movable_stats_account_zone(struct auto_movable_stats *stats, 795 struct zone *zone) 796{ 797 if (zone_idx(zone) == ZONE_MOVABLE) { 798 stats->movable_pages += zone->present_pages; 799 } else { 800 stats->kernel_early_pages += zone->present_early_pages; 801#ifdef CONFIG_CMA 802 /* 803 * CMA pages (never on hotplugged memory) behave like 804 * ZONE_MOVABLE. 805 */ 806 stats->movable_pages += zone->cma_pages; 807 stats->kernel_early_pages -= zone->cma_pages; 808#endif /* CONFIG_CMA */ 809 } 810} 811struct auto_movable_group_stats { 812 unsigned long movable_pages; 813 unsigned long req_kernel_early_pages; 814}; 815 816static int auto_movable_stats_account_group(struct memory_group *group, 817 void *arg) 818{ 819 const int ratio = READ_ONCE(auto_movable_ratio); 820 struct auto_movable_group_stats *stats = arg; 821 long pages; 822 823 /* 824 * We don't support modifying the config while the auto-movable online 825 * policy is already enabled. Just avoid the division by zero below. 826 */ 827 if (!ratio) 828 return 0; 829 830 /* 831 * Calculate how many early kernel pages this group requires to 832 * satisfy the configured zone ratio. 833 */ 834 pages = group->present_movable_pages * 100 / ratio; 835 pages -= group->present_kernel_pages; 836 837 if (pages > 0) 838 stats->req_kernel_early_pages += pages; 839 stats->movable_pages += group->present_movable_pages; 840 return 0; 841} 842 843static bool auto_movable_can_online_movable(int nid, struct memory_group *group, 844 unsigned long nr_pages) 845{ 846 unsigned long kernel_early_pages, movable_pages; 847 struct auto_movable_group_stats group_stats = {}; 848 struct auto_movable_stats stats = {}; 849 pg_data_t *pgdat = NODE_DATA(nid); 850 struct zone *zone; 851 int i; 852 853 /* Walk all relevant zones and collect MOVABLE vs. KERNEL stats. */ 854 if (nid == NUMA_NO_NODE) { 855 /* TODO: cache values */ 856 for_each_populated_zone(zone) 857 auto_movable_stats_account_zone(&stats, zone); 858 } else { 859 for (i = 0; i < MAX_NR_ZONES; i++) { 860 zone = pgdat->node_zones + i; 861 if (populated_zone(zone)) 862 auto_movable_stats_account_zone(&stats, zone); 863 } 864 } 865 866 kernel_early_pages = stats.kernel_early_pages; 867 movable_pages = stats.movable_pages; 868 869 /* 870 * Kernel memory inside dynamic memory group allows for more MOVABLE 871 * memory within the same group. Remove the effect of all but the 872 * current group from the stats. 873 */ 874 walk_dynamic_memory_groups(nid, auto_movable_stats_account_group, 875 group, &group_stats); 876 if (kernel_early_pages <= group_stats.req_kernel_early_pages) 877 return false; 878 kernel_early_pages -= group_stats.req_kernel_early_pages; 879 movable_pages -= group_stats.movable_pages; 880 881 if (group && group->is_dynamic) 882 kernel_early_pages += group->present_kernel_pages; 883 884 /* 885 * Test if we could online the given number of pages to ZONE_MOVABLE 886 * and still stay in the configured ratio. 887 */ 888 movable_pages += nr_pages; 889 return movable_pages <= (auto_movable_ratio * kernel_early_pages) / 100; 890} 891 892/* 893 * Returns a default kernel memory zone for the given pfn range. 894 * If no kernel zone covers this pfn range it will automatically go 895 * to the ZONE_NORMAL. 896 */ 897static struct zone *default_kernel_zone_for_pfn(int nid, unsigned long start_pfn, 898 unsigned long nr_pages) 899{ 900 struct pglist_data *pgdat = NODE_DATA(nid); 901 int zid; 902 903 for (zid = 0; zid < ZONE_NORMAL; zid++) { 904 struct zone *zone = &pgdat->node_zones[zid]; 905 906 if (zone_intersects(zone, start_pfn, nr_pages)) 907 return zone; 908 } 909 910 return &pgdat->node_zones[ZONE_NORMAL]; 911} 912 913/* 914 * Determine to which zone to online memory dynamically based on user 915 * configuration and system stats. We care about the following ratio: 916 * 917 * MOVABLE : KERNEL 918 * 919 * Whereby MOVABLE is memory in ZONE_MOVABLE and KERNEL is memory in 920 * one of the kernel zones. CMA pages inside one of the kernel zones really 921 * behaves like ZONE_MOVABLE, so we treat them accordingly. 922 * 923 * We don't allow for hotplugged memory in a KERNEL zone to increase the 924 * amount of MOVABLE memory we can have, so we end up with: 925 * 926 * MOVABLE : KERNEL_EARLY 927 * 928 * Whereby KERNEL_EARLY is memory in one of the kernel zones, available sinze 929 * boot. We base our calculation on KERNEL_EARLY internally, because: 930 * 931 * a) Hotplugged memory in one of the kernel zones can sometimes still get 932 * hotunplugged, especially when hot(un)plugging individual memory blocks. 933 * There is no coordination across memory devices, therefore "automatic" 934 * hotunplugging, as implemented in hypervisors, could result in zone 935 * imbalances. 936 * b) Early/boot memory in one of the kernel zones can usually not get 937 * hotunplugged again (e.g., no firmware interface to unplug, fragmented 938 * with unmovable allocations). While there are corner cases where it might 939 * still work, it is barely relevant in practice. 940 * 941 * Exceptions are dynamic memory groups, which allow for more MOVABLE 942 * memory within the same memory group -- because in that case, there is 943 * coordination within the single memory device managed by a single driver. 944 * 945 * We rely on "present pages" instead of "managed pages", as the latter is 946 * highly unreliable and dynamic in virtualized environments, and does not 947 * consider boot time allocations. For example, memory ballooning adjusts the 948 * managed pages when inflating/deflating the balloon, and balloon compaction 949 * can even migrate inflated pages between zones. 950 * 951 * Using "present pages" is better but some things to keep in mind are: 952 * 953 * a) Some memblock allocations, such as for the crashkernel area, are 954 * effectively unused by the kernel, yet they account to "present pages". 955 * Fortunately, these allocations are comparatively small in relevant setups 956 * (e.g., fraction of system memory). 957 * b) Some hotplugged memory blocks in virtualized environments, esecially 958 * hotplugged by virtio-mem, look like they are completely present, however, 959 * only parts of the memory block are actually currently usable. 960 * "present pages" is an upper limit that can get reached at runtime. As 961 * we base our calculations on KERNEL_EARLY, this is not an issue. 962 */ 963static struct zone *auto_movable_zone_for_pfn(int nid, 964 struct memory_group *group, 965 unsigned long pfn, 966 unsigned long nr_pages) 967{ 968 unsigned long online_pages = 0, max_pages, end_pfn; 969 struct page *page; 970 971 if (!auto_movable_ratio) 972 goto kernel_zone; 973 974 if (group && !group->is_dynamic) { 975 max_pages = group->s.max_pages; 976 online_pages = group->present_movable_pages; 977 978 /* If anything is !MOVABLE online the rest !MOVABLE. */ 979 if (group->present_kernel_pages) 980 goto kernel_zone; 981 } else if (!group || group->d.unit_pages == nr_pages) { 982 max_pages = nr_pages; 983 } else { 984 max_pages = group->d.unit_pages; 985 /* 986 * Take a look at all online sections in the current unit. 987 * We can safely assume that all pages within a section belong 988 * to the same zone, because dynamic memory groups only deal 989 * with hotplugged memory. 990 */ 991 pfn = ALIGN_DOWN(pfn, group->d.unit_pages); 992 end_pfn = pfn + group->d.unit_pages; 993 for (; pfn < end_pfn; pfn += PAGES_PER_SECTION) { 994 page = pfn_to_online_page(pfn); 995 if (!page) 996 continue; 997 /* If anything is !MOVABLE online the rest !MOVABLE. */ 998 if (!is_zone_movable_page(page)) 999 goto kernel_zone; 1000 online_pages += PAGES_PER_SECTION; 1001 } 1002 } 1003 1004 /* 1005 * Online MOVABLE if we could *currently* online all remaining parts 1006 * MOVABLE. We expect to (add+) online them immediately next, so if 1007 * nobody interferes, all will be MOVABLE if possible. 1008 */ 1009 nr_pages = max_pages - online_pages; 1010 if (!auto_movable_can_online_movable(NUMA_NO_NODE, group, nr_pages)) 1011 goto kernel_zone; 1012 1013#ifdef CONFIG_NUMA 1014 if (auto_movable_numa_aware && 1015 !auto_movable_can_online_movable(nid, group, nr_pages)) 1016 goto kernel_zone; 1017#endif /* CONFIG_NUMA */ 1018 1019 return &NODE_DATA(nid)->node_zones[ZONE_MOVABLE]; 1020kernel_zone: 1021 return default_kernel_zone_for_pfn(nid, pfn, nr_pages); 1022} 1023 1024static inline struct zone *default_zone_for_pfn(int nid, unsigned long start_pfn, 1025 unsigned long nr_pages) 1026{ 1027 struct zone *kernel_zone = default_kernel_zone_for_pfn(nid, start_pfn, 1028 nr_pages); 1029 struct zone *movable_zone = &NODE_DATA(nid)->node_zones[ZONE_MOVABLE]; 1030 bool in_kernel = zone_intersects(kernel_zone, start_pfn, nr_pages); 1031 bool in_movable = zone_intersects(movable_zone, start_pfn, nr_pages); 1032 1033 /* 1034 * We inherit the existing zone in a simple case where zones do not 1035 * overlap in the given range 1036 */ 1037 if (in_kernel ^ in_movable) 1038 return (in_kernel) ? kernel_zone : movable_zone; 1039 1040 /* 1041 * If the range doesn't belong to any zone or two zones overlap in the 1042 * given range then we use movable zone only if movable_node is 1043 * enabled because we always online to a kernel zone by default. 1044 */ 1045 return movable_node_enabled ? movable_zone : kernel_zone; 1046} 1047 1048struct zone *zone_for_pfn_range(int online_type, int nid, 1049 struct memory_group *group, unsigned long start_pfn, 1050 unsigned long nr_pages) 1051{ 1052 if (online_type == MMOP_ONLINE_KERNEL) 1053 return default_kernel_zone_for_pfn(nid, start_pfn, nr_pages); 1054 1055 if (online_type == MMOP_ONLINE_MOVABLE) 1056 return &NODE_DATA(nid)->node_zones[ZONE_MOVABLE]; 1057 1058 if (online_policy == ONLINE_POLICY_AUTO_MOVABLE) 1059 return auto_movable_zone_for_pfn(nid, group, start_pfn, nr_pages); 1060 1061 return default_zone_for_pfn(nid, start_pfn, nr_pages); 1062} 1063 1064/* 1065 * This function should only be called by memory_block_{online,offline}, 1066 * and {online,offline}_pages. 1067 */ 1068void adjust_present_page_count(struct page *page, struct memory_group *group, 1069 long nr_pages) 1070{ 1071 struct zone *zone = page_zone(page); 1072 const bool movable = zone_idx(zone) == ZONE_MOVABLE; 1073 1074 /* 1075 * We only support onlining/offlining/adding/removing of complete 1076 * memory blocks; therefore, either all is either early or hotplugged. 1077 */ 1078 if (early_section(__pfn_to_section(page_to_pfn(page)))) 1079 zone->present_early_pages += nr_pages; 1080 zone->present_pages += nr_pages; 1081 zone->zone_pgdat->node_present_pages += nr_pages; 1082 1083 if (group && movable) 1084 group->present_movable_pages += nr_pages; 1085 else if (group && !movable) 1086 group->present_kernel_pages += nr_pages; 1087} 1088 1089int mhp_init_memmap_on_memory(unsigned long pfn, unsigned long nr_pages, 1090 struct zone *zone, bool mhp_off_inaccessible) 1091{ 1092 unsigned long end_pfn = pfn + nr_pages; 1093 int ret, i; 1094 1095 ret = kasan_add_zero_shadow(__va(PFN_PHYS(pfn)), PFN_PHYS(nr_pages)); 1096 if (ret) 1097 return ret; 1098 1099 /* 1100 * Memory block is accessible at this stage and hence poison the struct 1101 * pages now. If the memory block is accessible during memory hotplug 1102 * addition phase, then page poisining is already performed in 1103 * sparse_add_section(). 1104 */ 1105 if (mhp_off_inaccessible) 1106 page_init_poison(pfn_to_page(pfn), sizeof(struct page) * nr_pages); 1107 1108 move_pfn_range_to_zone(zone, pfn, nr_pages, NULL, MIGRATE_UNMOVABLE); 1109 1110 for (i = 0; i < nr_pages; i++) 1111 SetPageVmemmapSelfHosted(pfn_to_page(pfn + i)); 1112 1113 /* 1114 * It might be that the vmemmap_pages fully span sections. If that is 1115 * the case, mark those sections online here as otherwise they will be 1116 * left offline. 1117 */ 1118 if (nr_pages >= PAGES_PER_SECTION) 1119 online_mem_sections(pfn, ALIGN_DOWN(end_pfn, PAGES_PER_SECTION)); 1120 1121 return ret; 1122} 1123 1124void mhp_deinit_memmap_on_memory(unsigned long pfn, unsigned long nr_pages) 1125{ 1126 unsigned long end_pfn = pfn + nr_pages; 1127 1128 /* 1129 * It might be that the vmemmap_pages fully span sections. If that is 1130 * the case, mark those sections offline here as otherwise they will be 1131 * left online. 1132 */ 1133 if (nr_pages >= PAGES_PER_SECTION) 1134 offline_mem_sections(pfn, ALIGN_DOWN(end_pfn, PAGES_PER_SECTION)); 1135 1136 /* 1137 * The pages associated with this vmemmap have been offlined, so 1138 * we can reset its state here. 1139 */ 1140 remove_pfn_range_from_zone(page_zone(pfn_to_page(pfn)), pfn, nr_pages); 1141 kasan_remove_zero_shadow(__va(PFN_PHYS(pfn)), PFN_PHYS(nr_pages)); 1142} 1143 1144/* 1145 * Must be called with mem_hotplug_lock in write mode. 1146 */ 1147int __ref online_pages(unsigned long pfn, unsigned long nr_pages, 1148 struct zone *zone, struct memory_group *group) 1149{ 1150 unsigned long flags; 1151 int need_zonelists_rebuild = 0; 1152 const int nid = zone_to_nid(zone); 1153 int ret; 1154 struct memory_notify arg; 1155 1156 /* 1157 * {on,off}lining is constrained to full memory sections (or more 1158 * precisely to memory blocks from the user space POV). 1159 * memmap_on_memory is an exception because it reserves initial part 1160 * of the physical memory space for vmemmaps. That space is pageblock 1161 * aligned. 1162 */ 1163 if (WARN_ON_ONCE(!nr_pages || !pageblock_aligned(pfn) || 1164 !IS_ALIGNED(pfn + nr_pages, PAGES_PER_SECTION))) 1165 return -EINVAL; 1166 1167 1168 /* associate pfn range with the zone */ 1169 move_pfn_range_to_zone(zone, pfn, nr_pages, NULL, MIGRATE_ISOLATE); 1170 1171 arg.start_pfn = pfn; 1172 arg.nr_pages = nr_pages; 1173 node_states_check_changes_online(nr_pages, zone, &arg); 1174 1175 ret = memory_notify(MEM_GOING_ONLINE, &arg); 1176 ret = notifier_to_errno(ret); 1177 if (ret) 1178 goto failed_addition; 1179 1180 /* 1181 * Fixup the number of isolated pageblocks before marking the sections 1182 * onlining, such that undo_isolate_page_range() works correctly. 1183 */ 1184 spin_lock_irqsave(&zone->lock, flags); 1185 zone->nr_isolate_pageblock += nr_pages / pageblock_nr_pages; 1186 spin_unlock_irqrestore(&zone->lock, flags); 1187 1188 /* 1189 * If this zone is not populated, then it is not in zonelist. 1190 * This means the page allocator ignores this zone. 1191 * So, zonelist must be updated after online. 1192 */ 1193 if (!populated_zone(zone)) { 1194 need_zonelists_rebuild = 1; 1195 setup_zone_pageset(zone); 1196 } 1197 1198 online_pages_range(pfn, nr_pages); 1199 adjust_present_page_count(pfn_to_page(pfn), group, nr_pages); 1200 1201 node_states_set_node(nid, &arg); 1202 if (need_zonelists_rebuild) 1203 build_all_zonelists(NULL); 1204 1205 /* Basic onlining is complete, allow allocation of onlined pages. */ 1206 undo_isolate_page_range(pfn, pfn + nr_pages, MIGRATE_MOVABLE); 1207 1208 /* 1209 * Freshly onlined pages aren't shuffled (e.g., all pages are placed to 1210 * the tail of the freelist when undoing isolation). Shuffle the whole 1211 * zone to make sure the just onlined pages are properly distributed 1212 * across the whole freelist - to create an initial shuffle. 1213 */ 1214 shuffle_zone(zone); 1215 1216 /* reinitialise watermarks and update pcp limits */ 1217 init_per_zone_wmark_min(); 1218 1219 kswapd_run(nid); 1220 kcompactd_run(nid); 1221 1222 writeback_set_ratelimit(); 1223 1224 memory_notify(MEM_ONLINE, &arg); 1225 return 0; 1226 1227failed_addition: 1228 pr_debug("online_pages [mem %#010llx-%#010llx] failed\n", 1229 (unsigned long long) pfn << PAGE_SHIFT, 1230 (((unsigned long long) pfn + nr_pages) << PAGE_SHIFT) - 1); 1231 memory_notify(MEM_CANCEL_ONLINE, &arg); 1232 remove_pfn_range_from_zone(zone, pfn, nr_pages); 1233 return ret; 1234} 1235 1236/* we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG */ 1237static pg_data_t __ref *hotadd_init_pgdat(int nid) 1238{ 1239 struct pglist_data *pgdat; 1240 1241 /* 1242 * NODE_DATA is preallocated (free_area_init) but its internal 1243 * state is not allocated completely. Add missing pieces. 1244 * Completely offline nodes stay around and they just need 1245 * reintialization. 1246 */ 1247 pgdat = NODE_DATA(nid); 1248 1249 /* init node's zones as empty zones, we don't have any present pages.*/ 1250 free_area_init_core_hotplug(pgdat); 1251 1252 /* 1253 * The node we allocated has no zone fallback lists. For avoiding 1254 * to access not-initialized zonelist, build here. 1255 */ 1256 build_all_zonelists(pgdat); 1257 1258 return pgdat; 1259} 1260 1261/* 1262 * __try_online_node - online a node if offlined 1263 * @nid: the node ID 1264 * @set_node_online: Whether we want to online the node 1265 * called by cpu_up() to online a node without onlined memory. 1266 * 1267 * Returns: 1268 * 1 -> a new node has been allocated 1269 * 0 -> the node is already online 1270 * -ENOMEM -> the node could not be allocated 1271 */ 1272static int __try_online_node(int nid, bool set_node_online) 1273{ 1274 pg_data_t *pgdat; 1275 int ret = 1; 1276 1277 if (node_online(nid)) 1278 return 0; 1279 1280 pgdat = hotadd_init_pgdat(nid); 1281 if (!pgdat) { 1282 pr_err("Cannot online node %d due to NULL pgdat\n", nid); 1283 ret = -ENOMEM; 1284 goto out; 1285 } 1286 1287 if (set_node_online) { 1288 node_set_online(nid); 1289 ret = register_one_node(nid); 1290 BUG_ON(ret); 1291 } 1292out: 1293 return ret; 1294} 1295 1296/* 1297 * Users of this function always want to online/register the node 1298 */ 1299int try_online_node(int nid) 1300{ 1301 int ret; 1302 1303 mem_hotplug_begin(); 1304 ret = __try_online_node(nid, true); 1305 mem_hotplug_done(); 1306 return ret; 1307} 1308 1309static int check_hotplug_memory_range(u64 start, u64 size) 1310{ 1311 /* memory range must be block size aligned */ 1312 if (!size || !IS_ALIGNED(start, memory_block_size_bytes()) || 1313 !IS_ALIGNED(size, memory_block_size_bytes())) { 1314 pr_err("Block size [%#lx] unaligned hotplug range: start %#llx, size %#llx", 1315 memory_block_size_bytes(), start, size); 1316 return -EINVAL; 1317 } 1318 1319 return 0; 1320} 1321 1322static int online_memory_block(struct memory_block *mem, void *arg) 1323{ 1324 mem->online_type = mhp_default_online_type; 1325 return device_online(&mem->dev); 1326} 1327 1328#ifndef arch_supports_memmap_on_memory 1329static inline bool arch_supports_memmap_on_memory(unsigned long vmemmap_size) 1330{ 1331 /* 1332 * As default, we want the vmemmap to span a complete PMD such that we 1333 * can map the vmemmap using a single PMD if supported by the 1334 * architecture. 1335 */ 1336 return IS_ALIGNED(vmemmap_size, PMD_SIZE); 1337} 1338#endif 1339 1340bool mhp_supports_memmap_on_memory(void) 1341{ 1342 unsigned long vmemmap_size = memory_block_memmap_size(); 1343 unsigned long memmap_pages = memory_block_memmap_on_memory_pages(); 1344 1345 /* 1346 * Besides having arch support and the feature enabled at runtime, we 1347 * need a few more assumptions to hold true: 1348 * 1349 * a) The vmemmap pages span complete PMDs: We don't want vmemmap code 1350 * to populate memory from the altmap for unrelated parts (i.e., 1351 * other memory blocks) 1352 * 1353 * b) The vmemmap pages (and thereby the pages that will be exposed to 1354 * the buddy) have to cover full pageblocks: memory onlining/offlining 1355 * code requires applicable ranges to be page-aligned, for example, to 1356 * set the migratetypes properly. 1357 * 1358 * TODO: Although we have a check here to make sure that vmemmap pages 1359 * fully populate a PMD, it is not the right place to check for 1360 * this. A much better solution involves improving vmemmap code 1361 * to fallback to base pages when trying to populate vmemmap using 1362 * altmap as an alternative source of memory, and we do not exactly 1363 * populate a single PMD. 1364 */ 1365 if (!mhp_memmap_on_memory()) 1366 return false; 1367 1368 /* 1369 * Make sure the vmemmap allocation is fully contained 1370 * so that we always allocate vmemmap memory from altmap area. 1371 */ 1372 if (!IS_ALIGNED(vmemmap_size, PAGE_SIZE)) 1373 return false; 1374 1375 /* 1376 * start pfn should be pageblock_nr_pages aligned for correctly 1377 * setting migrate types 1378 */ 1379 if (!pageblock_aligned(memmap_pages)) 1380 return false; 1381 1382 if (memmap_pages == PHYS_PFN(memory_block_size_bytes())) 1383 /* No effective hotplugged memory doesn't make sense. */ 1384 return false; 1385 1386 return arch_supports_memmap_on_memory(vmemmap_size); 1387} 1388EXPORT_SYMBOL_GPL(mhp_supports_memmap_on_memory); 1389 1390static void __ref remove_memory_blocks_and_altmaps(u64 start, u64 size) 1391{ 1392 unsigned long memblock_size = memory_block_size_bytes(); 1393 u64 cur_start; 1394 1395 /* 1396 * For memmap_on_memory, the altmaps were added on a per-memblock 1397 * basis; we have to process each individual memory block. 1398 */ 1399 for (cur_start = start; cur_start < start + size; 1400 cur_start += memblock_size) { 1401 struct vmem_altmap *altmap = NULL; 1402 struct memory_block *mem; 1403 1404 mem = find_memory_block(pfn_to_section_nr(PFN_DOWN(cur_start))); 1405 if (WARN_ON_ONCE(!mem)) 1406 continue; 1407 1408 altmap = mem->altmap; 1409 mem->altmap = NULL; 1410 1411 remove_memory_block_devices(cur_start, memblock_size); 1412 1413 arch_remove_memory(cur_start, memblock_size, altmap); 1414 1415 /* Verify that all vmemmap pages have actually been freed. */ 1416 WARN(altmap->alloc, "Altmap not fully unmapped"); 1417 kfree(altmap); 1418 } 1419} 1420 1421static int create_altmaps_and_memory_blocks(int nid, struct memory_group *group, 1422 u64 start, u64 size, mhp_t mhp_flags) 1423{ 1424 unsigned long memblock_size = memory_block_size_bytes(); 1425 u64 cur_start; 1426 int ret; 1427 1428 for (cur_start = start; cur_start < start + size; 1429 cur_start += memblock_size) { 1430 struct mhp_params params = { .pgprot = 1431 pgprot_mhp(PAGE_KERNEL) }; 1432 struct vmem_altmap mhp_altmap = { 1433 .base_pfn = PHYS_PFN(cur_start), 1434 .end_pfn = PHYS_PFN(cur_start + memblock_size - 1), 1435 }; 1436 1437 mhp_altmap.free = memory_block_memmap_on_memory_pages(); 1438 if (mhp_flags & MHP_OFFLINE_INACCESSIBLE) 1439 mhp_altmap.inaccessible = true; 1440 params.altmap = kmemdup(&mhp_altmap, sizeof(struct vmem_altmap), 1441 GFP_KERNEL); 1442 if (!params.altmap) { 1443 ret = -ENOMEM; 1444 goto out; 1445 } 1446 1447 /* call arch's memory hotadd */ 1448 ret = arch_add_memory(nid, cur_start, memblock_size, ¶ms); 1449 if (ret < 0) { 1450 kfree(params.altmap); 1451 goto out; 1452 } 1453 1454 /* create memory block devices after memory was added */ 1455 ret = create_memory_block_devices(cur_start, memblock_size, 1456 params.altmap, group); 1457 if (ret) { 1458 arch_remove_memory(cur_start, memblock_size, NULL); 1459 kfree(params.altmap); 1460 goto out; 1461 } 1462 } 1463 1464 return 0; 1465out: 1466 if (ret && cur_start != start) 1467 remove_memory_blocks_and_altmaps(start, cur_start - start); 1468 return ret; 1469} 1470 1471/* 1472 * NOTE: The caller must call lock_device_hotplug() to serialize hotplug 1473 * and online/offline operations (triggered e.g. by sysfs). 1474 * 1475 * we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG 1476 */ 1477int __ref add_memory_resource(int nid, struct resource *res, mhp_t mhp_flags) 1478{ 1479 struct mhp_params params = { .pgprot = pgprot_mhp(PAGE_KERNEL) }; 1480 enum memblock_flags memblock_flags = MEMBLOCK_NONE; 1481 struct memory_group *group = NULL; 1482 u64 start, size; 1483 bool new_node = false; 1484 int ret; 1485 1486 start = res->start; 1487 size = resource_size(res); 1488 1489 ret = check_hotplug_memory_range(start, size); 1490 if (ret) 1491 return ret; 1492 1493 if (mhp_flags & MHP_NID_IS_MGID) { 1494 group = memory_group_find_by_id(nid); 1495 if (!group) 1496 return -EINVAL; 1497 nid = group->nid; 1498 } 1499 1500 if (!node_possible(nid)) { 1501 WARN(1, "node %d was absent from the node_possible_map\n", nid); 1502 return -EINVAL; 1503 } 1504 1505 mem_hotplug_begin(); 1506 1507 if (IS_ENABLED(CONFIG_ARCH_KEEP_MEMBLOCK)) { 1508 if (res->flags & IORESOURCE_SYSRAM_DRIVER_MANAGED) 1509 memblock_flags = MEMBLOCK_DRIVER_MANAGED; 1510 ret = memblock_add_node(start, size, nid, memblock_flags); 1511 if (ret) 1512 goto error_mem_hotplug_end; 1513 } 1514 1515 ret = __try_online_node(nid, false); 1516 if (ret < 0) 1517 goto error; 1518 new_node = ret; 1519 1520 /* 1521 * Self hosted memmap array 1522 */ 1523 if ((mhp_flags & MHP_MEMMAP_ON_MEMORY) && 1524 mhp_supports_memmap_on_memory()) { 1525 ret = create_altmaps_and_memory_blocks(nid, group, start, size, mhp_flags); 1526 if (ret) 1527 goto error; 1528 } else { 1529 ret = arch_add_memory(nid, start, size, ¶ms); 1530 if (ret < 0) 1531 goto error; 1532 1533 /* create memory block devices after memory was added */ 1534 ret = create_memory_block_devices(start, size, NULL, group); 1535 if (ret) { 1536 arch_remove_memory(start, size, params.altmap); 1537 goto error; 1538 } 1539 } 1540 1541 if (new_node) { 1542 /* If sysfs file of new node can't be created, cpu on the node 1543 * can't be hot-added. There is no rollback way now. 1544 * So, check by BUG_ON() to catch it reluctantly.. 1545 * We online node here. We can't roll back from here. 1546 */ 1547 node_set_online(nid); 1548 ret = __register_one_node(nid); 1549 BUG_ON(ret); 1550 } 1551 1552 register_memory_blocks_under_node(nid, PFN_DOWN(start), 1553 PFN_UP(start + size - 1), 1554 MEMINIT_HOTPLUG); 1555 1556 /* create new memmap entry */ 1557 if (!strcmp(res->name, "System RAM")) 1558 firmware_map_add_hotplug(start, start + size, "System RAM"); 1559 1560 /* device_online() will take the lock when calling online_pages() */ 1561 mem_hotplug_done(); 1562 1563 /* 1564 * In case we're allowed to merge the resource, flag it and trigger 1565 * merging now that adding succeeded. 1566 */ 1567 if (mhp_flags & MHP_MERGE_RESOURCE) 1568 merge_system_ram_resource(res); 1569 1570 /* online pages if requested */ 1571 if (mhp_default_online_type != MMOP_OFFLINE) 1572 walk_memory_blocks(start, size, NULL, online_memory_block); 1573 1574 return ret; 1575error: 1576 if (IS_ENABLED(CONFIG_ARCH_KEEP_MEMBLOCK)) 1577 memblock_remove(start, size); 1578error_mem_hotplug_end: 1579 mem_hotplug_done(); 1580 return ret; 1581} 1582 1583/* requires device_hotplug_lock, see add_memory_resource() */ 1584int __ref __add_memory(int nid, u64 start, u64 size, mhp_t mhp_flags) 1585{ 1586 struct resource *res; 1587 int ret; 1588 1589 res = register_memory_resource(start, size, "System RAM"); 1590 if (IS_ERR(res)) 1591 return PTR_ERR(res); 1592 1593 ret = add_memory_resource(nid, res, mhp_flags); 1594 if (ret < 0) 1595 release_memory_resource(res); 1596 return ret; 1597} 1598 1599int add_memory(int nid, u64 start, u64 size, mhp_t mhp_flags) 1600{ 1601 int rc; 1602 1603 lock_device_hotplug(); 1604 rc = __add_memory(nid, start, size, mhp_flags); 1605 unlock_device_hotplug(); 1606 1607 return rc; 1608} 1609EXPORT_SYMBOL_GPL(add_memory); 1610 1611/* 1612 * Add special, driver-managed memory to the system as system RAM. Such 1613 * memory is not exposed via the raw firmware-provided memmap as system 1614 * RAM, instead, it is detected and added by a driver - during cold boot, 1615 * after a reboot, and after kexec. 1616 * 1617 * Reasons why this memory should not be used for the initial memmap of a 1618 * kexec kernel or for placing kexec images: 1619 * - The booting kernel is in charge of determining how this memory will be 1620 * used (e.g., use persistent memory as system RAM) 1621 * - Coordination with a hypervisor is required before this memory 1622 * can be used (e.g., inaccessible parts). 1623 * 1624 * For this memory, no entries in /sys/firmware/memmap ("raw firmware-provided 1625 * memory map") are created. Also, the created memory resource is flagged 1626 * with IORESOURCE_SYSRAM_DRIVER_MANAGED, so in-kernel users can special-case 1627 * this memory as well (esp., not place kexec images onto it). 1628 * 1629 * The resource_name (visible via /proc/iomem) has to have the format 1630 * "System RAM ($DRIVER)". 1631 */ 1632int add_memory_driver_managed(int nid, u64 start, u64 size, 1633 const char *resource_name, mhp_t mhp_flags) 1634{ 1635 struct resource *res; 1636 int rc; 1637 1638 if (!resource_name || 1639 strstr(resource_name, "System RAM (") != resource_name || 1640 resource_name[strlen(resource_name) - 1] != ')') 1641 return -EINVAL; 1642 1643 lock_device_hotplug(); 1644 1645 res = register_memory_resource(start, size, resource_name); 1646 if (IS_ERR(res)) { 1647 rc = PTR_ERR(res); 1648 goto out_unlock; 1649 } 1650 1651 rc = add_memory_resource(nid, res, mhp_flags); 1652 if (rc < 0) 1653 release_memory_resource(res); 1654 1655out_unlock: 1656 unlock_device_hotplug(); 1657 return rc; 1658} 1659EXPORT_SYMBOL_GPL(add_memory_driver_managed); 1660 1661/* 1662 * Platforms should define arch_get_mappable_range() that provides 1663 * maximum possible addressable physical memory range for which the 1664 * linear mapping could be created. The platform returned address 1665 * range must adhere to these following semantics. 1666 * 1667 * - range.start <= range.end 1668 * - Range includes both end points [range.start..range.end] 1669 * 1670 * There is also a fallback definition provided here, allowing the 1671 * entire possible physical address range in case any platform does 1672 * not define arch_get_mappable_range(). 1673 */ 1674struct range __weak arch_get_mappable_range(void) 1675{ 1676 struct range mhp_range = { 1677 .start = 0UL, 1678 .end = -1ULL, 1679 }; 1680 return mhp_range; 1681} 1682 1683struct range mhp_get_pluggable_range(bool need_mapping) 1684{ 1685 const u64 max_phys = (1ULL << MAX_PHYSMEM_BITS) - 1; 1686 struct range mhp_range; 1687 1688 if (need_mapping) { 1689 mhp_range = arch_get_mappable_range(); 1690 if (mhp_range.start > max_phys) { 1691 mhp_range.start = 0; 1692 mhp_range.end = 0; 1693 } 1694 mhp_range.end = min_t(u64, mhp_range.end, max_phys); 1695 } else { 1696 mhp_range.start = 0; 1697 mhp_range.end = max_phys; 1698 } 1699 return mhp_range; 1700} 1701EXPORT_SYMBOL_GPL(mhp_get_pluggable_range); 1702 1703bool mhp_range_allowed(u64 start, u64 size, bool need_mapping) 1704{ 1705 struct range mhp_range = mhp_get_pluggable_range(need_mapping); 1706 u64 end = start + size; 1707 1708 if (start < end && start >= mhp_range.start && (end - 1) <= mhp_range.end) 1709 return true; 1710 1711 pr_warn("Hotplug memory [%#llx-%#llx] exceeds maximum addressable range [%#llx-%#llx]\n", 1712 start, end, mhp_range.start, mhp_range.end); 1713 return false; 1714} 1715 1716#ifdef CONFIG_MEMORY_HOTREMOVE 1717/* 1718 * Scan pfn range [start,end) to find movable/migratable pages (LRU pages, 1719 * non-lru movable pages and hugepages). Will skip over most unmovable 1720 * pages (esp., pages that can be skipped when offlining), but bail out on 1721 * definitely unmovable pages. 1722 * 1723 * Returns: 1724 * 0 in case a movable page is found and movable_pfn was updated. 1725 * -ENOENT in case no movable page was found. 1726 * -EBUSY in case a definitely unmovable page was found. 1727 */ 1728static int scan_movable_pages(unsigned long start, unsigned long end, 1729 unsigned long *movable_pfn) 1730{ 1731 unsigned long pfn; 1732 1733 for (pfn = start; pfn < end; pfn++) { 1734 struct page *page, *head; 1735 unsigned long skip; 1736 1737 if (!pfn_valid(pfn)) 1738 continue; 1739 page = pfn_to_page(pfn); 1740 if (PageLRU(page)) 1741 goto found; 1742 if (__PageMovable(page)) 1743 goto found; 1744 1745 /* 1746 * PageOffline() pages that are not marked __PageMovable() and 1747 * have a reference count > 0 (after MEM_GOING_OFFLINE) are 1748 * definitely unmovable. If their reference count would be 0, 1749 * they could at least be skipped when offlining memory. 1750 */ 1751 if (PageOffline(page) && page_count(page)) 1752 return -EBUSY; 1753 1754 if (!PageHuge(page)) 1755 continue; 1756 head = compound_head(page); 1757 /* 1758 * This test is racy as we hold no reference or lock. The 1759 * hugetlb page could have been free'ed and head is no longer 1760 * a hugetlb page before the following check. In such unlikely 1761 * cases false positives and negatives are possible. Calling 1762 * code must deal with these scenarios. 1763 */ 1764 if (HPageMigratable(head)) 1765 goto found; 1766 skip = compound_nr(head) - (pfn - page_to_pfn(head)); 1767 pfn += skip - 1; 1768 } 1769 return -ENOENT; 1770found: 1771 *movable_pfn = pfn; 1772 return 0; 1773} 1774 1775static void do_migrate_range(unsigned long start_pfn, unsigned long end_pfn) 1776{ 1777 unsigned long pfn; 1778 struct page *page, *head; 1779 LIST_HEAD(source); 1780 static DEFINE_RATELIMIT_STATE(migrate_rs, DEFAULT_RATELIMIT_INTERVAL, 1781 DEFAULT_RATELIMIT_BURST); 1782 1783 for (pfn = start_pfn; pfn < end_pfn; pfn++) { 1784 struct folio *folio; 1785 bool isolated; 1786 1787 if (!pfn_valid(pfn)) 1788 continue; 1789 page = pfn_to_page(pfn); 1790 folio = page_folio(page); 1791 head = &folio->page; 1792 1793 if (PageHuge(page)) { 1794 pfn = page_to_pfn(head) + compound_nr(head) - 1; 1795 isolate_hugetlb(folio, &source); 1796 continue; 1797 } else if (PageTransHuge(page)) 1798 pfn = page_to_pfn(head) + thp_nr_pages(page) - 1; 1799 1800 /* 1801 * HWPoison pages have elevated reference counts so the migration would 1802 * fail on them. It also doesn't make any sense to migrate them in the 1803 * first place. Still try to unmap such a page in case it is still mapped 1804 * (e.g. current hwpoison implementation doesn't unmap KSM pages but keep 1805 * the unmap as the catch all safety net). 1806 */ 1807 if (PageHWPoison(page)) { 1808 if (WARN_ON(folio_test_lru(folio))) 1809 folio_isolate_lru(folio); 1810 if (folio_mapped(folio)) 1811 try_to_unmap(folio, TTU_IGNORE_MLOCK); 1812 continue; 1813 } 1814 1815 if (!get_page_unless_zero(page)) 1816 continue; 1817 /* 1818 * We can skip free pages. And we can deal with pages on 1819 * LRU and non-lru movable pages. 1820 */ 1821 if (PageLRU(page)) 1822 isolated = isolate_lru_page(page); 1823 else 1824 isolated = isolate_movable_page(page, ISOLATE_UNEVICTABLE); 1825 if (isolated) { 1826 list_add_tail(&page->lru, &source); 1827 if (!__PageMovable(page)) 1828 inc_node_page_state(page, NR_ISOLATED_ANON + 1829 page_is_file_lru(page)); 1830 1831 } else { 1832 if (__ratelimit(&migrate_rs)) { 1833 pr_warn("failed to isolate pfn %lx\n", pfn); 1834 dump_page(page, "isolation failed"); 1835 } 1836 } 1837 put_page(page); 1838 } 1839 if (!list_empty(&source)) { 1840 nodemask_t nmask = node_states[N_MEMORY]; 1841 struct migration_target_control mtc = { 1842 .nmask = &nmask, 1843 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL, 1844 }; 1845 int ret; 1846 1847 /* 1848 * We have checked that migration range is on a single zone so 1849 * we can use the nid of the first page to all the others. 1850 */ 1851 mtc.nid = page_to_nid(list_first_entry(&source, struct page, lru)); 1852 1853 /* 1854 * try to allocate from a different node but reuse this node 1855 * if there are no other online nodes to be used (e.g. we are 1856 * offlining a part of the only existing node) 1857 */ 1858 node_clear(mtc.nid, nmask); 1859 if (nodes_empty(nmask)) 1860 node_set(mtc.nid, nmask); 1861 ret = migrate_pages(&source, alloc_migration_target, NULL, 1862 (unsigned long)&mtc, MIGRATE_SYNC, MR_MEMORY_HOTPLUG, NULL); 1863 if (ret) { 1864 list_for_each_entry(page, &source, lru) { 1865 if (__ratelimit(&migrate_rs)) { 1866 pr_warn("migrating pfn %lx failed ret:%d\n", 1867 page_to_pfn(page), ret); 1868 dump_page(page, "migration failure"); 1869 } 1870 } 1871 putback_movable_pages(&source); 1872 } 1873 } 1874} 1875 1876static int __init cmdline_parse_movable_node(char *p) 1877{ 1878 movable_node_enabled = true; 1879 return 0; 1880} 1881early_param("movable_node", cmdline_parse_movable_node); 1882 1883/* check which state of node_states will be changed when offline memory */ 1884static void node_states_check_changes_offline(unsigned long nr_pages, 1885 struct zone *zone, struct memory_notify *arg) 1886{ 1887 struct pglist_data *pgdat = zone->zone_pgdat; 1888 unsigned long present_pages = 0; 1889 enum zone_type zt; 1890 1891 arg->status_change_nid = NUMA_NO_NODE; 1892 arg->status_change_nid_normal = NUMA_NO_NODE; 1893 1894 /* 1895 * Check whether node_states[N_NORMAL_MEMORY] will be changed. 1896 * If the memory to be offline is within the range 1897 * [0..ZONE_NORMAL], and it is the last present memory there, 1898 * the zones in that range will become empty after the offlining, 1899 * thus we can determine that we need to clear the node from 1900 * node_states[N_NORMAL_MEMORY]. 1901 */ 1902 for (zt = 0; zt <= ZONE_NORMAL; zt++) 1903 present_pages += pgdat->node_zones[zt].present_pages; 1904 if (zone_idx(zone) <= ZONE_NORMAL && nr_pages >= present_pages) 1905 arg->status_change_nid_normal = zone_to_nid(zone); 1906 1907 /* 1908 * We have accounted the pages from [0..ZONE_NORMAL); ZONE_HIGHMEM 1909 * does not apply as we don't support 32bit. 1910 * Here we count the possible pages from ZONE_MOVABLE. 1911 * If after having accounted all the pages, we see that the nr_pages 1912 * to be offlined is over or equal to the accounted pages, 1913 * we know that the node will become empty, and so, we can clear 1914 * it for N_MEMORY as well. 1915 */ 1916 present_pages += pgdat->node_zones[ZONE_MOVABLE].present_pages; 1917 1918 if (nr_pages >= present_pages) 1919 arg->status_change_nid = zone_to_nid(zone); 1920} 1921 1922static void node_states_clear_node(int node, struct memory_notify *arg) 1923{ 1924 if (arg->status_change_nid_normal >= 0) 1925 node_clear_state(node, N_NORMAL_MEMORY); 1926 1927 if (arg->status_change_nid >= 0) 1928 node_clear_state(node, N_MEMORY); 1929} 1930 1931static int count_system_ram_pages_cb(unsigned long start_pfn, 1932 unsigned long nr_pages, void *data) 1933{ 1934 unsigned long *nr_system_ram_pages = data; 1935 1936 *nr_system_ram_pages += nr_pages; 1937 return 0; 1938} 1939 1940/* 1941 * Must be called with mem_hotplug_lock in write mode. 1942 */ 1943int __ref offline_pages(unsigned long start_pfn, unsigned long nr_pages, 1944 struct zone *zone, struct memory_group *group) 1945{ 1946 const unsigned long end_pfn = start_pfn + nr_pages; 1947 unsigned long pfn, system_ram_pages = 0; 1948 const int node = zone_to_nid(zone); 1949 unsigned long flags; 1950 struct memory_notify arg; 1951 char *reason; 1952 int ret; 1953 1954 /* 1955 * {on,off}lining is constrained to full memory sections (or more 1956 * precisely to memory blocks from the user space POV). 1957 * memmap_on_memory is an exception because it reserves initial part 1958 * of the physical memory space for vmemmaps. That space is pageblock 1959 * aligned. 1960 */ 1961 if (WARN_ON_ONCE(!nr_pages || !pageblock_aligned(start_pfn) || 1962 !IS_ALIGNED(start_pfn + nr_pages, PAGES_PER_SECTION))) 1963 return -EINVAL; 1964 1965 /* 1966 * Don't allow to offline memory blocks that contain holes. 1967 * Consequently, memory blocks with holes can never get onlined 1968 * via the hotplug path - online_pages() - as hotplugged memory has 1969 * no holes. This way, we e.g., don't have to worry about marking 1970 * memory holes PG_reserved, don't need pfn_valid() checks, and can 1971 * avoid using walk_system_ram_range() later. 1972 */ 1973 walk_system_ram_range(start_pfn, nr_pages, &system_ram_pages, 1974 count_system_ram_pages_cb); 1975 if (system_ram_pages != nr_pages) { 1976 ret = -EINVAL; 1977 reason = "memory holes"; 1978 goto failed_removal; 1979 } 1980 1981 /* 1982 * We only support offlining of memory blocks managed by a single zone, 1983 * checked by calling code. This is just a sanity check that we might 1984 * want to remove in the future. 1985 */ 1986 if (WARN_ON_ONCE(page_zone(pfn_to_page(start_pfn)) != zone || 1987 page_zone(pfn_to_page(end_pfn - 1)) != zone)) { 1988 ret = -EINVAL; 1989 reason = "multizone range"; 1990 goto failed_removal; 1991 } 1992 1993 /* 1994 * Disable pcplists so that page isolation cannot race with freeing 1995 * in a way that pages from isolated pageblock are left on pcplists. 1996 */ 1997 zone_pcp_disable(zone); 1998 lru_cache_disable(); 1999 2000 /* set above range as isolated */ 2001 ret = start_isolate_page_range(start_pfn, end_pfn, 2002 MIGRATE_MOVABLE, 2003 MEMORY_OFFLINE | REPORT_FAILURE, 2004 GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL); 2005 if (ret) { 2006 reason = "failure to isolate range"; 2007 goto failed_removal_pcplists_disabled; 2008 } 2009 2010 arg.start_pfn = start_pfn; 2011 arg.nr_pages = nr_pages; 2012 node_states_check_changes_offline(nr_pages, zone, &arg); 2013 2014 ret = memory_notify(MEM_GOING_OFFLINE, &arg); 2015 ret = notifier_to_errno(ret); 2016 if (ret) { 2017 reason = "notifier failure"; 2018 goto failed_removal_isolated; 2019 } 2020 2021 do { 2022 pfn = start_pfn; 2023 do { 2024 /* 2025 * Historically we always checked for any signal and 2026 * can't limit it to fatal signals without eventually 2027 * breaking user space. 2028 */ 2029 if (signal_pending(current)) { 2030 ret = -EINTR; 2031 reason = "signal backoff"; 2032 goto failed_removal_isolated; 2033 } 2034 2035 cond_resched(); 2036 2037 ret = scan_movable_pages(pfn, end_pfn, &pfn); 2038 if (!ret) { 2039 /* 2040 * TODO: fatal migration failures should bail 2041 * out 2042 */ 2043 do_migrate_range(pfn, end_pfn); 2044 } 2045 } while (!ret); 2046 2047 if (ret != -ENOENT) { 2048 reason = "unmovable page"; 2049 goto failed_removal_isolated; 2050 } 2051 2052 /* 2053 * Dissolve free hugepages in the memory block before doing 2054 * offlining actually in order to make hugetlbfs's object 2055 * counting consistent. 2056 */ 2057 ret = dissolve_free_huge_pages(start_pfn, end_pfn); 2058 if (ret) { 2059 reason = "failure to dissolve huge pages"; 2060 goto failed_removal_isolated; 2061 } 2062 2063 ret = test_pages_isolated(start_pfn, end_pfn, MEMORY_OFFLINE); 2064 2065 } while (ret); 2066 2067 /* Mark all sections offline and remove free pages from the buddy. */ 2068 __offline_isolated_pages(start_pfn, end_pfn); 2069 pr_debug("Offlined Pages %ld\n", nr_pages); 2070 2071 /* 2072 * The memory sections are marked offline, and the pageblock flags 2073 * effectively stale; nobody should be touching them. Fixup the number 2074 * of isolated pageblocks, memory onlining will properly revert this. 2075 */ 2076 spin_lock_irqsave(&zone->lock, flags); 2077 zone->nr_isolate_pageblock -= nr_pages / pageblock_nr_pages; 2078 spin_unlock_irqrestore(&zone->lock, flags); 2079 2080 lru_cache_enable(); 2081 zone_pcp_enable(zone); 2082 2083 /* removal success */ 2084 adjust_managed_page_count(pfn_to_page(start_pfn), -nr_pages); 2085 adjust_present_page_count(pfn_to_page(start_pfn), group, -nr_pages); 2086 2087 /* reinitialise watermarks and update pcp limits */ 2088 init_per_zone_wmark_min(); 2089 2090 /* 2091 * Make sure to mark the node as memory-less before rebuilding the zone 2092 * list. Otherwise this node would still appear in the fallback lists. 2093 */ 2094 node_states_clear_node(node, &arg); 2095 if (!populated_zone(zone)) { 2096 zone_pcp_reset(zone); 2097 build_all_zonelists(NULL); 2098 } 2099 2100 if (arg.status_change_nid >= 0) { 2101 kcompactd_stop(node); 2102 kswapd_stop(node); 2103 } 2104 2105 writeback_set_ratelimit(); 2106 2107 memory_notify(MEM_OFFLINE, &arg); 2108 remove_pfn_range_from_zone(zone, start_pfn, nr_pages); 2109 return 0; 2110 2111failed_removal_isolated: 2112 /* pushback to free area */ 2113 undo_isolate_page_range(start_pfn, end_pfn, MIGRATE_MOVABLE); 2114 memory_notify(MEM_CANCEL_OFFLINE, &arg); 2115failed_removal_pcplists_disabled: 2116 lru_cache_enable(); 2117 zone_pcp_enable(zone); 2118failed_removal: 2119 pr_debug("memory offlining [mem %#010llx-%#010llx] failed due to %s\n", 2120 (unsigned long long) start_pfn << PAGE_SHIFT, 2121 ((unsigned long long) end_pfn << PAGE_SHIFT) - 1, 2122 reason); 2123 return ret; 2124} 2125 2126static int check_memblock_offlined_cb(struct memory_block *mem, void *arg) 2127{ 2128 int *nid = arg; 2129 2130 *nid = mem->nid; 2131 if (unlikely(mem->state != MEM_OFFLINE)) { 2132 phys_addr_t beginpa, endpa; 2133 2134 beginpa = PFN_PHYS(section_nr_to_pfn(mem->start_section_nr)); 2135 endpa = beginpa + memory_block_size_bytes() - 1; 2136 pr_warn("removing memory fails, because memory [%pa-%pa] is onlined\n", 2137 &beginpa, &endpa); 2138 2139 return -EBUSY; 2140 } 2141 return 0; 2142} 2143 2144static int count_memory_range_altmaps_cb(struct memory_block *mem, void *arg) 2145{ 2146 u64 *num_altmaps = (u64 *)arg; 2147 2148 if (mem->altmap) 2149 *num_altmaps += 1; 2150 2151 return 0; 2152} 2153 2154static int check_cpu_on_node(int nid) 2155{ 2156 int cpu; 2157 2158 for_each_present_cpu(cpu) { 2159 if (cpu_to_node(cpu) == nid) 2160 /* 2161 * the cpu on this node isn't removed, and we can't 2162 * offline this node. 2163 */ 2164 return -EBUSY; 2165 } 2166 2167 return 0; 2168} 2169 2170static int check_no_memblock_for_node_cb(struct memory_block *mem, void *arg) 2171{ 2172 int nid = *(int *)arg; 2173 2174 /* 2175 * If a memory block belongs to multiple nodes, the stored nid is not 2176 * reliable. However, such blocks are always online (e.g., cannot get 2177 * offlined) and, therefore, are still spanned by the node. 2178 */ 2179 return mem->nid == nid ? -EEXIST : 0; 2180} 2181 2182/** 2183 * try_offline_node 2184 * @nid: the node ID 2185 * 2186 * Offline a node if all memory sections and cpus of the node are removed. 2187 * 2188 * NOTE: The caller must call lock_device_hotplug() to serialize hotplug 2189 * and online/offline operations before this call. 2190 */ 2191void try_offline_node(int nid) 2192{ 2193 int rc; 2194 2195 /* 2196 * If the node still spans pages (especially ZONE_DEVICE), don't 2197 * offline it. A node spans memory after move_pfn_range_to_zone(), 2198 * e.g., after the memory block was onlined. 2199 */ 2200 if (node_spanned_pages(nid)) 2201 return; 2202 2203 /* 2204 * Especially offline memory blocks might not be spanned by the 2205 * node. They will get spanned by the node once they get onlined. 2206 * However, they link to the node in sysfs and can get onlined later. 2207 */ 2208 rc = for_each_memory_block(&nid, check_no_memblock_for_node_cb); 2209 if (rc) 2210 return; 2211 2212 if (check_cpu_on_node(nid)) 2213 return; 2214 2215 /* 2216 * all memory/cpu of this node are removed, we can offline this 2217 * node now. 2218 */ 2219 node_set_offline(nid); 2220 unregister_one_node(nid); 2221} 2222EXPORT_SYMBOL(try_offline_node); 2223 2224static int memory_blocks_have_altmaps(u64 start, u64 size) 2225{ 2226 u64 num_memblocks = size / memory_block_size_bytes(); 2227 u64 num_altmaps = 0; 2228 2229 if (!mhp_memmap_on_memory()) 2230 return 0; 2231 2232 walk_memory_blocks(start, size, &num_altmaps, 2233 count_memory_range_altmaps_cb); 2234 2235 if (num_altmaps == 0) 2236 return 0; 2237 2238 if (WARN_ON_ONCE(num_memblocks != num_altmaps)) 2239 return -EINVAL; 2240 2241 return 1; 2242} 2243 2244static int __ref try_remove_memory(u64 start, u64 size) 2245{ 2246 int rc, nid = NUMA_NO_NODE; 2247 2248 BUG_ON(check_hotplug_memory_range(start, size)); 2249 2250 /* 2251 * All memory blocks must be offlined before removing memory. Check 2252 * whether all memory blocks in question are offline and return error 2253 * if this is not the case. 2254 * 2255 * While at it, determine the nid. Note that if we'd have mixed nodes, 2256 * we'd only try to offline the last determined one -- which is good 2257 * enough for the cases we care about. 2258 */ 2259 rc = walk_memory_blocks(start, size, &nid, check_memblock_offlined_cb); 2260 if (rc) 2261 return rc; 2262 2263 /* remove memmap entry */ 2264 firmware_map_remove(start, start + size, "System RAM"); 2265 2266 mem_hotplug_begin(); 2267 2268 rc = memory_blocks_have_altmaps(start, size); 2269 if (rc < 0) { 2270 mem_hotplug_done(); 2271 return rc; 2272 } else if (!rc) { 2273 /* 2274 * Memory block device removal under the device_hotplug_lock is 2275 * a barrier against racing online attempts. 2276 * No altmaps present, do the removal directly 2277 */ 2278 remove_memory_block_devices(start, size); 2279 arch_remove_memory(start, size, NULL); 2280 } else { 2281 /* all memblocks in the range have altmaps */ 2282 remove_memory_blocks_and_altmaps(start, size); 2283 } 2284 2285 if (IS_ENABLED(CONFIG_ARCH_KEEP_MEMBLOCK)) { 2286 memblock_phys_free(start, size); 2287 memblock_remove(start, size); 2288 } 2289 2290 release_mem_region_adjustable(start, size); 2291 2292 if (nid != NUMA_NO_NODE) 2293 try_offline_node(nid); 2294 2295 mem_hotplug_done(); 2296 return 0; 2297} 2298 2299/** 2300 * __remove_memory - Remove memory if every memory block is offline 2301 * @start: physical address of the region to remove 2302 * @size: size of the region to remove 2303 * 2304 * NOTE: The caller must call lock_device_hotplug() to serialize hotplug 2305 * and online/offline operations before this call, as required by 2306 * try_offline_node(). 2307 */ 2308void __remove_memory(u64 start, u64 size) 2309{ 2310 2311 /* 2312 * trigger BUG() if some memory is not offlined prior to calling this 2313 * function 2314 */ 2315 if (try_remove_memory(start, size)) 2316 BUG(); 2317} 2318 2319/* 2320 * Remove memory if every memory block is offline, otherwise return -EBUSY is 2321 * some memory is not offline 2322 */ 2323int remove_memory(u64 start, u64 size) 2324{ 2325 int rc; 2326 2327 lock_device_hotplug(); 2328 rc = try_remove_memory(start, size); 2329 unlock_device_hotplug(); 2330 2331 return rc; 2332} 2333EXPORT_SYMBOL_GPL(remove_memory); 2334 2335static int try_offline_memory_block(struct memory_block *mem, void *arg) 2336{ 2337 uint8_t online_type = MMOP_ONLINE_KERNEL; 2338 uint8_t **online_types = arg; 2339 struct page *page; 2340 int rc; 2341 2342 /* 2343 * Sense the online_type via the zone of the memory block. Offlining 2344 * with multiple zones within one memory block will be rejected 2345 * by offlining code ... so we don't care about that. 2346 */ 2347 page = pfn_to_online_page(section_nr_to_pfn(mem->start_section_nr)); 2348 if (page && zone_idx(page_zone(page)) == ZONE_MOVABLE) 2349 online_type = MMOP_ONLINE_MOVABLE; 2350 2351 rc = device_offline(&mem->dev); 2352 /* 2353 * Default is MMOP_OFFLINE - change it only if offlining succeeded, 2354 * so try_reonline_memory_block() can do the right thing. 2355 */ 2356 if (!rc) 2357 **online_types = online_type; 2358 2359 (*online_types)++; 2360 /* Ignore if already offline. */ 2361 return rc < 0 ? rc : 0; 2362} 2363 2364static int try_reonline_memory_block(struct memory_block *mem, void *arg) 2365{ 2366 uint8_t **online_types = arg; 2367 int rc; 2368 2369 if (**online_types != MMOP_OFFLINE) { 2370 mem->online_type = **online_types; 2371 rc = device_online(&mem->dev); 2372 if (rc < 0) 2373 pr_warn("%s: Failed to re-online memory: %d", 2374 __func__, rc); 2375 } 2376 2377 /* Continue processing all remaining memory blocks. */ 2378 (*online_types)++; 2379 return 0; 2380} 2381 2382/* 2383 * Try to offline and remove memory. Might take a long time to finish in case 2384 * memory is still in use. Primarily useful for memory devices that logically 2385 * unplugged all memory (so it's no longer in use) and want to offline + remove 2386 * that memory. 2387 */ 2388int offline_and_remove_memory(u64 start, u64 size) 2389{ 2390 const unsigned long mb_count = size / memory_block_size_bytes(); 2391 uint8_t *online_types, *tmp; 2392 int rc; 2393 2394 if (!IS_ALIGNED(start, memory_block_size_bytes()) || 2395 !IS_ALIGNED(size, memory_block_size_bytes()) || !size) 2396 return -EINVAL; 2397 2398 /* 2399 * We'll remember the old online type of each memory block, so we can 2400 * try to revert whatever we did when offlining one memory block fails 2401 * after offlining some others succeeded. 2402 */ 2403 online_types = kmalloc_array(mb_count, sizeof(*online_types), 2404 GFP_KERNEL); 2405 if (!online_types) 2406 return -ENOMEM; 2407 /* 2408 * Initialize all states to MMOP_OFFLINE, so when we abort processing in 2409 * try_offline_memory_block(), we'll skip all unprocessed blocks in 2410 * try_reonline_memory_block(). 2411 */ 2412 memset(online_types, MMOP_OFFLINE, mb_count); 2413 2414 lock_device_hotplug(); 2415 2416 tmp = online_types; 2417 rc = walk_memory_blocks(start, size, &tmp, try_offline_memory_block); 2418 2419 /* 2420 * In case we succeeded to offline all memory, remove it. 2421 * This cannot fail as it cannot get onlined in the meantime. 2422 */ 2423 if (!rc) { 2424 rc = try_remove_memory(start, size); 2425 if (rc) 2426 pr_err("%s: Failed to remove memory: %d", __func__, rc); 2427 } 2428 2429 /* 2430 * Rollback what we did. While memory onlining might theoretically fail 2431 * (nacked by a notifier), it barely ever happens. 2432 */ 2433 if (rc) { 2434 tmp = online_types; 2435 walk_memory_blocks(start, size, &tmp, 2436 try_reonline_memory_block); 2437 } 2438 unlock_device_hotplug(); 2439 2440 kfree(online_types); 2441 return rc; 2442} 2443EXPORT_SYMBOL_GPL(offline_and_remove_memory); 2444#endif /* CONFIG_MEMORY_HOTREMOVE */ 2445