Linux Audio

Check our new training course

Loading...
v5.4
   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
  52static void generic_online_page(struct page *page, unsigned int order);
  53
  54static online_page_callback_t online_page_callback = generic_online_page;
  55static DEFINE_MUTEX(online_page_callback_lock);
  56
  57DEFINE_STATIC_PERCPU_RWSEM(mem_hotplug_lock);
  58
  59void get_online_mems(void)
  60{
  61	percpu_down_read(&mem_hotplug_lock);
  62}
  63
  64void put_online_mems(void)
  65{
  66	percpu_up_read(&mem_hotplug_lock);
  67}
  68
  69bool movable_node_enabled = false;
  70
  71#ifndef CONFIG_MEMORY_HOTPLUG_DEFAULT_ONLINE
  72bool memhp_auto_online;
  73#else
  74bool memhp_auto_online = true;
  75#endif
  76EXPORT_SYMBOL_GPL(memhp_auto_online);
  77
  78static int __init setup_memhp_default_state(char *str)
  79{
  80	if (!strcmp(str, "online"))
  81		memhp_auto_online = true;
  82	else if (!strcmp(str, "offline"))
  83		memhp_auto_online = false;
  84
  85	return 1;
  86}
  87__setup("memhp_default_state=", setup_memhp_default_state);
  88
  89void mem_hotplug_begin(void)
  90{
  91	cpus_read_lock();
  92	percpu_down_write(&mem_hotplug_lock);
  93}
  94
  95void mem_hotplug_done(void)
  96{
  97	percpu_up_write(&mem_hotplug_lock);
  98	cpus_read_unlock();
  99}
 100
 101u64 max_mem_size = U64_MAX;
 102
 103/* add this memory to iomem resource */
 104static struct resource *register_memory_resource(u64 start, u64 size)
 
 105{
 106	struct resource *res;
 107	unsigned long flags =  IORESOURCE_SYSTEM_RAM | IORESOURCE_BUSY;
 108	char *resource_name = "System RAM";
 109
 110	if (start + size > max_mem_size)
 
 
 
 
 
 
 
 
 
 
 
 
 111		return ERR_PTR(-E2BIG);
 112
 113	/*
 114	 * Request ownership of the new memory range.  This might be
 115	 * a child of an existing resource that was present but
 116	 * not marked as busy.
 117	 */
 118	res = __request_region(&iomem_resource, start, size,
 119			       resource_name, flags);
 120
 121	if (!res) {
 122		pr_debug("Unable to reserve System RAM region: %016llx->%016llx\n",
 123				start, start + size);
 124		return ERR_PTR(-EEXIST);
 125	}
 126	return res;
 127}
 128
 129static void release_memory_resource(struct resource *res)
 130{
 131	if (!res)
 132		return;
 133	release_resource(res);
 134	kfree(res);
 135}
 136
 137#ifdef CONFIG_MEMORY_HOTPLUG_SPARSE
 138void get_page_bootmem(unsigned long info,  struct page *page,
 139		      unsigned long type)
 140{
 141	page->freelist = (void *)type;
 142	SetPagePrivate(page);
 143	set_page_private(page, info);
 144	page_ref_inc(page);
 145}
 146
 147void put_page_bootmem(struct page *page)
 148{
 149	unsigned long type;
 150
 151	type = (unsigned long) page->freelist;
 152	BUG_ON(type < MEMORY_HOTPLUG_MIN_BOOTMEM_TYPE ||
 153	       type > MEMORY_HOTPLUG_MAX_BOOTMEM_TYPE);
 154
 155	if (page_ref_dec_return(page) == 1) {
 156		page->freelist = NULL;
 157		ClearPagePrivate(page);
 158		set_page_private(page, 0);
 159		INIT_LIST_HEAD(&page->lru);
 160		free_reserved_page(page);
 161	}
 162}
 163
 164#ifdef CONFIG_HAVE_BOOTMEM_INFO_NODE
 165#ifndef CONFIG_SPARSEMEM_VMEMMAP
 166static void register_page_bootmem_info_section(unsigned long start_pfn)
 167{
 168	unsigned long mapsize, section_nr, i;
 169	struct mem_section *ms;
 170	struct page *page, *memmap;
 171	struct mem_section_usage *usage;
 172
 173	section_nr = pfn_to_section_nr(start_pfn);
 174	ms = __nr_to_section(section_nr);
 175
 176	/* Get section's memmap address */
 177	memmap = sparse_decode_mem_map(ms->section_mem_map, section_nr);
 178
 179	/*
 180	 * Get page for the memmap's phys address
 181	 * XXX: need more consideration for sparse_vmemmap...
 182	 */
 183	page = virt_to_page(memmap);
 184	mapsize = sizeof(struct page) * PAGES_PER_SECTION;
 185	mapsize = PAGE_ALIGN(mapsize) >> PAGE_SHIFT;
 186
 187	/* remember memmap's page */
 188	for (i = 0; i < mapsize; i++, page++)
 189		get_page_bootmem(section_nr, page, SECTION_INFO);
 190
 191	usage = ms->usage;
 192	page = virt_to_page(usage);
 193
 194	mapsize = PAGE_ALIGN(mem_section_usage_size()) >> PAGE_SHIFT;
 195
 196	for (i = 0; i < mapsize; i++, page++)
 197		get_page_bootmem(section_nr, page, MIX_SECTION_INFO);
 198
 199}
 200#else /* CONFIG_SPARSEMEM_VMEMMAP */
 201static void register_page_bootmem_info_section(unsigned long start_pfn)
 202{
 203	unsigned long mapsize, section_nr, i;
 204	struct mem_section *ms;
 205	struct page *page, *memmap;
 206	struct mem_section_usage *usage;
 207
 208	section_nr = pfn_to_section_nr(start_pfn);
 209	ms = __nr_to_section(section_nr);
 210
 211	memmap = sparse_decode_mem_map(ms->section_mem_map, section_nr);
 212
 213	register_page_bootmem_memmap(section_nr, memmap, PAGES_PER_SECTION);
 214
 215	usage = ms->usage;
 216	page = virt_to_page(usage);
 217
 218	mapsize = PAGE_ALIGN(mem_section_usage_size()) >> PAGE_SHIFT;
 219
 220	for (i = 0; i < mapsize; i++, page++)
 221		get_page_bootmem(section_nr, page, MIX_SECTION_INFO);
 222}
 223#endif /* !CONFIG_SPARSEMEM_VMEMMAP */
 224
 225void __init register_page_bootmem_info_node(struct pglist_data *pgdat)
 226{
 227	unsigned long i, pfn, end_pfn, nr_pages;
 228	int node = pgdat->node_id;
 229	struct page *page;
 230
 231	nr_pages = PAGE_ALIGN(sizeof(struct pglist_data)) >> PAGE_SHIFT;
 232	page = virt_to_page(pgdat);
 233
 234	for (i = 0; i < nr_pages; i++, page++)
 235		get_page_bootmem(node, page, NODE_INFO);
 236
 237	pfn = pgdat->node_start_pfn;
 238	end_pfn = pgdat_end_pfn(pgdat);
 239
 240	/* register section info */
 241	for (; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
 242		/*
 243		 * Some platforms can assign the same pfn to multiple nodes - on
 244		 * node0 as well as nodeN.  To avoid registering a pfn against
 245		 * multiple nodes we check that this pfn does not already
 246		 * reside in some other nodes.
 247		 */
 248		if (pfn_valid(pfn) && (early_pfn_to_nid(pfn) == node))
 249			register_page_bootmem_info_section(pfn);
 250	}
 251}
 252#endif /* CONFIG_HAVE_BOOTMEM_INFO_NODE */
 253
 254static int check_pfn_span(unsigned long pfn, unsigned long nr_pages,
 255		const char *reason)
 256{
 257	/*
 258	 * Disallow all operations smaller than a sub-section and only
 259	 * allow operations smaller than a section for
 260	 * SPARSEMEM_VMEMMAP. Note that check_hotplug_memory_range()
 261	 * enforces a larger memory_block_size_bytes() granularity for
 262	 * memory that will be marked online, so this check should only
 263	 * fire for direct arch_{add,remove}_memory() users outside of
 264	 * add_memory_resource().
 265	 */
 266	unsigned long min_align;
 267
 268	if (IS_ENABLED(CONFIG_SPARSEMEM_VMEMMAP))
 269		min_align = PAGES_PER_SUBSECTION;
 270	else
 271		min_align = PAGES_PER_SECTION;
 272	if (!IS_ALIGNED(pfn, min_align)
 273			|| !IS_ALIGNED(nr_pages, min_align)) {
 274		WARN(1, "Misaligned __%s_pages start: %#lx end: #%lx\n",
 275				reason, pfn, pfn + nr_pages - 1);
 276		return -EINVAL;
 277	}
 278	return 0;
 279}
 280
 281/*
 282 * Reasonably generic function for adding memory.  It is
 283 * expected that archs that support memory hotplug will
 284 * call this function after deciding the zone to which to
 285 * add the new pages.
 286 */
 287int __ref __add_pages(int nid, unsigned long pfn, unsigned long nr_pages,
 288		struct mhp_restrictions *restrictions)
 289{
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 290	int err;
 291	unsigned long nr, start_sec, end_sec;
 292	struct vmem_altmap *altmap = restrictions->altmap;
 
 
 
 
 293
 294	if (altmap) {
 295		/*
 296		 * Validate altmap is within bounds of the total request
 297		 */
 298		if (altmap->base_pfn != pfn
 299				|| vmem_altmap_offset(altmap) > nr_pages) {
 300			pr_warn_once("memory add fail, invalid altmap\n");
 301			return -EINVAL;
 302		}
 303		altmap->alloc = 0;
 304	}
 305
 306	err = check_pfn_span(pfn, nr_pages, "add");
 307	if (err)
 308		return err;
 309
 310	start_sec = pfn_to_section_nr(pfn);
 311	end_sec = pfn_to_section_nr(pfn + nr_pages - 1);
 312	for (nr = start_sec; nr <= end_sec; nr++) {
 313		unsigned long pfns;
 314
 315		pfns = min(nr_pages, PAGES_PER_SECTION
 316				- (pfn & ~PAGE_SECTION_MASK));
 317		err = sparse_add_section(nid, pfn, pfns, altmap);
 318		if (err)
 319			break;
 320		pfn += pfns;
 321		nr_pages -= pfns;
 322		cond_resched();
 323	}
 324	vmemmap_populate_print_last();
 325	return err;
 326}
 327
 328/* find the smallest valid pfn in the range [start_pfn, end_pfn) */
 329static unsigned long find_smallest_section_pfn(int nid, struct zone *zone,
 330				     unsigned long start_pfn,
 331				     unsigned long end_pfn)
 332{
 333	for (; start_pfn < end_pfn; start_pfn += PAGES_PER_SUBSECTION) {
 334		if (unlikely(!pfn_to_online_page(start_pfn)))
 335			continue;
 336
 337		if (unlikely(pfn_to_nid(start_pfn) != nid))
 338			continue;
 339
 340		if (zone && zone != page_zone(pfn_to_page(start_pfn)))
 341			continue;
 342
 343		return start_pfn;
 344	}
 345
 346	return 0;
 347}
 348
 349/* find the biggest valid pfn in the range [start_pfn, end_pfn). */
 350static unsigned long find_biggest_section_pfn(int nid, struct zone *zone,
 351				    unsigned long start_pfn,
 352				    unsigned long end_pfn)
 353{
 354	unsigned long pfn;
 355
 356	/* pfn is the end pfn of a memory section. */
 357	pfn = end_pfn - 1;
 358	for (; pfn >= start_pfn; pfn -= PAGES_PER_SUBSECTION) {
 359		if (unlikely(!pfn_to_online_page(pfn)))
 360			continue;
 361
 362		if (unlikely(pfn_to_nid(pfn) != nid))
 363			continue;
 364
 365		if (zone && zone != page_zone(pfn_to_page(pfn)))
 366			continue;
 367
 368		return pfn;
 369	}
 370
 371	return 0;
 372}
 373
 374static void shrink_zone_span(struct zone *zone, unsigned long start_pfn,
 375			     unsigned long end_pfn)
 376{
 377	unsigned long zone_start_pfn = zone->zone_start_pfn;
 378	unsigned long z = zone_end_pfn(zone); /* zone_end_pfn namespace clash */
 379	unsigned long zone_end_pfn = z;
 380	unsigned long pfn;
 381	int nid = zone_to_nid(zone);
 382
 383	zone_span_writelock(zone);
 384	if (zone_start_pfn == start_pfn) {
 385		/*
 386		 * If the section is smallest section in the zone, it need
 387		 * shrink zone->zone_start_pfn and zone->zone_spanned_pages.
 388		 * In this case, we find second smallest valid mem_section
 389		 * for shrinking zone.
 390		 */
 391		pfn = find_smallest_section_pfn(nid, zone, end_pfn,
 392						zone_end_pfn);
 393		if (pfn) {
 
 394			zone->zone_start_pfn = pfn;
 395			zone->spanned_pages = zone_end_pfn - pfn;
 
 
 396		}
 397	} else if (zone_end_pfn == end_pfn) {
 398		/*
 399		 * If the section is biggest section in the zone, it need
 400		 * shrink zone->spanned_pages.
 401		 * In this case, we find second biggest valid mem_section for
 402		 * shrinking zone.
 403		 */
 404		pfn = find_biggest_section_pfn(nid, zone, zone_start_pfn,
 405					       start_pfn);
 406		if (pfn)
 407			zone->spanned_pages = pfn - zone_start_pfn + 1;
 408	}
 409
 410	/*
 411	 * The section is not biggest or smallest mem_section in the zone, it
 412	 * only creates a hole in the zone. So in this case, we need not
 413	 * change the zone. But perhaps, the zone has only hole data. Thus
 414	 * it check the zone has only hole or not.
 415	 */
 416	pfn = zone_start_pfn;
 417	for (; pfn < zone_end_pfn; pfn += PAGES_PER_SUBSECTION) {
 418		if (unlikely(!pfn_to_online_page(pfn)))
 419			continue;
 420
 421		if (page_zone(pfn_to_page(pfn)) != zone)
 422			continue;
 423
 424		/* Skip range to be removed */
 425		if (pfn >= start_pfn && pfn < end_pfn)
 426			continue;
 427
 428		/* If we find valid section, we have nothing to do */
 429		zone_span_writeunlock(zone);
 430		return;
 431	}
 432
 433	/* The zone has no valid section */
 434	zone->zone_start_pfn = 0;
 435	zone->spanned_pages = 0;
 436	zone_span_writeunlock(zone);
 437}
 438
 439static void update_pgdat_span(struct pglist_data *pgdat)
 440{
 441	unsigned long node_start_pfn = 0, node_end_pfn = 0;
 442	struct zone *zone;
 443
 444	for (zone = pgdat->node_zones;
 445	     zone < pgdat->node_zones + MAX_NR_ZONES; zone++) {
 446		unsigned long zone_end_pfn = zone->zone_start_pfn +
 447					     zone->spanned_pages;
 448
 449		/* No need to lock the zones, they can't change. */
 450		if (!zone->spanned_pages)
 451			continue;
 452		if (!node_end_pfn) {
 453			node_start_pfn = zone->zone_start_pfn;
 454			node_end_pfn = zone_end_pfn;
 455			continue;
 456		}
 457
 458		if (zone_end_pfn > node_end_pfn)
 459			node_end_pfn = zone_end_pfn;
 460		if (zone->zone_start_pfn < node_start_pfn)
 461			node_start_pfn = zone->zone_start_pfn;
 462	}
 463
 464	pgdat->node_start_pfn = node_start_pfn;
 465	pgdat->node_spanned_pages = node_end_pfn - node_start_pfn;
 466}
 467
 468static void __remove_zone(struct zone *zone, unsigned long start_pfn,
 469		unsigned long nr_pages)
 
 470{
 
 471	struct pglist_data *pgdat = zone->zone_pgdat;
 472	unsigned long flags;
 
 
 
 
 
 
 
 
 
 
 
 473
 474#ifdef CONFIG_ZONE_DEVICE
 475	/*
 476	 * Zone shrinking code cannot properly deal with ZONE_DEVICE. So
 477	 * we will not try to shrink the zones - which is okay as
 478	 * set_zone_contiguous() cannot deal with ZONE_DEVICE either way.
 479	 */
 480	if (zone_idx(zone) == ZONE_DEVICE)
 481		return;
 482#endif
 483
 484	pgdat_resize_lock(zone->zone_pgdat, &flags);
 
 485	shrink_zone_span(zone, start_pfn, start_pfn + nr_pages);
 486	update_pgdat_span(pgdat);
 487	pgdat_resize_unlock(zone->zone_pgdat, &flags);
 488}
 489
 490static void __remove_section(struct zone *zone, unsigned long pfn,
 491		unsigned long nr_pages, unsigned long map_offset,
 492		struct vmem_altmap *altmap)
 493{
 494	struct mem_section *ms = __nr_to_section(pfn_to_section_nr(pfn));
 495
 496	if (WARN_ON_ONCE(!valid_section(ms)))
 497		return;
 498
 499	__remove_zone(zone, pfn, nr_pages);
 500	sparse_remove_section(ms, pfn, nr_pages, map_offset, altmap);
 501}
 502
 503/**
 504 * __remove_pages() - remove sections of pages from a zone
 505 * @zone: zone from which pages need to be removed
 506 * @pfn: starting pageframe (must be aligned to start of a section)
 507 * @nr_pages: number of pages to remove (must be multiple of section size)
 508 * @altmap: alternative device page map or %NULL if default memmap is used
 509 *
 510 * Generic helper function to remove section mappings and sysfs entries
 511 * for the section of the memory we are removing. Caller needs to make
 512 * sure that pages are marked reserved and zones are adjust properly by
 513 * calling offline_pages().
 514 */
 515void __remove_pages(struct zone *zone, unsigned long pfn,
 516		    unsigned long nr_pages, struct vmem_altmap *altmap)
 517{
 518	unsigned long map_offset = 0;
 519	unsigned long nr, start_sec, end_sec;
 520
 521	map_offset = vmem_altmap_offset(altmap);
 522
 523	clear_zone_contiguous(zone);
 524
 525	if (check_pfn_span(pfn, nr_pages, "remove"))
 526		return;
 
 527
 528	start_sec = pfn_to_section_nr(pfn);
 529	end_sec = pfn_to_section_nr(pfn + nr_pages - 1);
 530	for (nr = start_sec; nr <= end_sec; nr++) {
 531		unsigned long pfns;
 532
 533		cond_resched();
 534		pfns = min(nr_pages, PAGES_PER_SECTION
 535				- (pfn & ~PAGE_SECTION_MASK));
 536		__remove_section(zone, pfn, pfns, map_offset, altmap);
 537		pfn += pfns;
 538		nr_pages -= pfns;
 539		map_offset = 0;
 540	}
 541
 542	set_zone_contiguous(zone);
 543}
 544
 545int set_online_page_callback(online_page_callback_t callback)
 546{
 547	int rc = -EINVAL;
 548
 549	get_online_mems();
 550	mutex_lock(&online_page_callback_lock);
 551
 552	if (online_page_callback == generic_online_page) {
 553		online_page_callback = callback;
 554		rc = 0;
 555	}
 556
 557	mutex_unlock(&online_page_callback_lock);
 558	put_online_mems();
 559
 560	return rc;
 561}
 562EXPORT_SYMBOL_GPL(set_online_page_callback);
 563
 564int restore_online_page_callback(online_page_callback_t callback)
 565{
 566	int rc = -EINVAL;
 567
 568	get_online_mems();
 569	mutex_lock(&online_page_callback_lock);
 570
 571	if (online_page_callback == callback) {
 572		online_page_callback = generic_online_page;
 573		rc = 0;
 574	}
 575
 576	mutex_unlock(&online_page_callback_lock);
 577	put_online_mems();
 578
 579	return rc;
 580}
 581EXPORT_SYMBOL_GPL(restore_online_page_callback);
 582
 583void __online_page_set_limits(struct page *page)
 584{
 585}
 586EXPORT_SYMBOL_GPL(__online_page_set_limits);
 587
 588void __online_page_increment_counters(struct page *page)
 589{
 590	adjust_managed_page_count(page, 1);
 591}
 592EXPORT_SYMBOL_GPL(__online_page_increment_counters);
 593
 594void __online_page_free(struct page *page)
 595{
 596	__free_reserved_page(page);
 597}
 598EXPORT_SYMBOL_GPL(__online_page_free);
 599
 600static void generic_online_page(struct page *page, unsigned int order)
 601{
 602	kernel_map_pages(page, 1 << order, 1);
 603	__free_pages_core(page, order);
 604	totalram_pages_add(1UL << order);
 605#ifdef CONFIG_HIGHMEM
 606	if (PageHighMem(page))
 607		totalhigh_pages_add(1UL << order);
 608#endif
 609}
 
 610
 611static int online_pages_range(unsigned long start_pfn, unsigned long nr_pages,
 612			void *arg)
 613{
 614	const unsigned long end_pfn = start_pfn + nr_pages;
 615	unsigned long pfn;
 616	int order;
 617
 618	/*
 619	 * Online the pages. The callback might decide to keep some pages
 620	 * PG_reserved (to add them to the buddy later), but we still account
 621	 * them as being online/belonging to this zone ("present").
 622	 */
 623	for (pfn = start_pfn; pfn < end_pfn; pfn += 1ul << order) {
 624		order = min(MAX_ORDER - 1, get_order(PFN_PHYS(end_pfn - pfn)));
 625		/* __free_pages_core() wants pfns to be aligned to the order */
 626		if (WARN_ON_ONCE(!IS_ALIGNED(pfn, 1ul << order)))
 627			order = 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 628		(*online_page_callback)(pfn_to_page(pfn), order);
 
 629	}
 630
 631	/* mark all involved sections as online */
 632	online_mem_sections(start_pfn, end_pfn);
 633
 634	*(unsigned long *)arg += nr_pages;
 635	return 0;
 636}
 637
 638/* check which state of node_states will be changed when online memory */
 639static void node_states_check_changes_online(unsigned long nr_pages,
 640	struct zone *zone, struct memory_notify *arg)
 641{
 642	int nid = zone_to_nid(zone);
 643
 644	arg->status_change_nid = NUMA_NO_NODE;
 645	arg->status_change_nid_normal = NUMA_NO_NODE;
 646	arg->status_change_nid_high = NUMA_NO_NODE;
 647
 648	if (!node_state(nid, N_MEMORY))
 649		arg->status_change_nid = nid;
 650	if (zone_idx(zone) <= ZONE_NORMAL && !node_state(nid, N_NORMAL_MEMORY))
 651		arg->status_change_nid_normal = nid;
 652#ifdef CONFIG_HIGHMEM
 653	if (zone_idx(zone) <= ZONE_HIGHMEM && !node_state(nid, N_HIGH_MEMORY))
 654		arg->status_change_nid_high = nid;
 655#endif
 656}
 657
 658static void node_states_set_node(int node, struct memory_notify *arg)
 659{
 660	if (arg->status_change_nid_normal >= 0)
 661		node_set_state(node, N_NORMAL_MEMORY);
 662
 663	if (arg->status_change_nid_high >= 0)
 664		node_set_state(node, N_HIGH_MEMORY);
 665
 666	if (arg->status_change_nid >= 0)
 667		node_set_state(node, N_MEMORY);
 668}
 669
 670static void __meminit resize_zone_range(struct zone *zone, unsigned long start_pfn,
 671		unsigned long nr_pages)
 672{
 673	unsigned long old_end_pfn = zone_end_pfn(zone);
 674
 675	if (zone_is_empty(zone) || start_pfn < zone->zone_start_pfn)
 676		zone->zone_start_pfn = start_pfn;
 677
 678	zone->spanned_pages = max(start_pfn + nr_pages, old_end_pfn) - zone->zone_start_pfn;
 679}
 680
 681static void __meminit resize_pgdat_range(struct pglist_data *pgdat, unsigned long start_pfn,
 682                                     unsigned long nr_pages)
 683{
 684	unsigned long old_end_pfn = pgdat_end_pfn(pgdat);
 685
 686	if (!pgdat->node_spanned_pages || start_pfn < pgdat->node_start_pfn)
 687		pgdat->node_start_pfn = start_pfn;
 688
 689	pgdat->node_spanned_pages = max(start_pfn + nr_pages, old_end_pfn) - pgdat->node_start_pfn;
 690
 691}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 692/*
 693 * Associate the pfn range with the given zone, initializing the memmaps
 694 * and resizing the pgdat/zone data to span the added pages. After this
 695 * call, all affected pages are PG_reserved.
 
 
 
 
 696 */
 697void __ref move_pfn_range_to_zone(struct zone *zone, unsigned long start_pfn,
 698		unsigned long nr_pages, struct vmem_altmap *altmap)
 
 699{
 700	struct pglist_data *pgdat = zone->zone_pgdat;
 701	int nid = pgdat->node_id;
 702	unsigned long flags;
 703
 704	clear_zone_contiguous(zone);
 705
 706	/* TODO Huh pgdat is irqsave while zone is not. It used to be like that before */
 707	pgdat_resize_lock(pgdat, &flags);
 708	zone_span_writelock(zone);
 709	if (zone_is_empty(zone))
 710		init_currently_empty_zone(zone, start_pfn, nr_pages);
 711	resize_zone_range(zone, start_pfn, nr_pages);
 712	zone_span_writeunlock(zone);
 713	resize_pgdat_range(pgdat, start_pfn, nr_pages);
 714	pgdat_resize_unlock(pgdat, &flags);
 
 
 
 
 
 
 
 
 
 
 
 
 715
 716	/*
 717	 * TODO now we have a visible range of pages which are not associated
 718	 * with their zone properly. Not nice but set_pfnblock_flags_mask
 719	 * expects the zone spans the pfn range. All the pages in the range
 720	 * are reserved so nobody should be touching them so we should be safe
 721	 */
 722	memmap_init_zone(nr_pages, nid, zone_idx(zone), start_pfn,
 723			MEMMAP_HOTPLUG, altmap);
 724
 725	set_zone_contiguous(zone);
 726}
 727
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 728/*
 729 * Returns a default kernel memory zone for the given pfn range.
 730 * If no kernel zone covers this pfn range it will automatically go
 731 * to the ZONE_NORMAL.
 732 */
 733static struct zone *default_kernel_zone_for_pfn(int nid, unsigned long start_pfn,
 734		unsigned long nr_pages)
 735{
 736	struct pglist_data *pgdat = NODE_DATA(nid);
 737	int zid;
 738
 739	for (zid = 0; zid <= ZONE_NORMAL; zid++) {
 740		struct zone *zone = &pgdat->node_zones[zid];
 741
 742		if (zone_intersects(zone, start_pfn, nr_pages))
 743			return zone;
 744	}
 745
 746	return &pgdat->node_zones[ZONE_NORMAL];
 747}
 748
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 749static inline struct zone *default_zone_for_pfn(int nid, unsigned long start_pfn,
 750		unsigned long nr_pages)
 751{
 752	struct zone *kernel_zone = default_kernel_zone_for_pfn(nid, start_pfn,
 753			nr_pages);
 754	struct zone *movable_zone = &NODE_DATA(nid)->node_zones[ZONE_MOVABLE];
 755	bool in_kernel = zone_intersects(kernel_zone, start_pfn, nr_pages);
 756	bool in_movable = zone_intersects(movable_zone, start_pfn, nr_pages);
 757
 758	/*
 759	 * We inherit the existing zone in a simple case where zones do not
 760	 * overlap in the given range
 761	 */
 762	if (in_kernel ^ in_movable)
 763		return (in_kernel) ? kernel_zone : movable_zone;
 764
 765	/*
 766	 * If the range doesn't belong to any zone or two zones overlap in the
 767	 * given range then we use movable zone only if movable_node is
 768	 * enabled because we always online to a kernel zone by default.
 769	 */
 770	return movable_node_enabled ? movable_zone : kernel_zone;
 771}
 772
 773struct zone * zone_for_pfn_range(int online_type, int nid, unsigned start_pfn,
 
 774		unsigned long nr_pages)
 775{
 776	if (online_type == MMOP_ONLINE_KERNEL)
 777		return default_kernel_zone_for_pfn(nid, start_pfn, nr_pages);
 778
 779	if (online_type == MMOP_ONLINE_MOVABLE)
 780		return &NODE_DATA(nid)->node_zones[ZONE_MOVABLE];
 781
 
 
 
 782	return default_zone_for_pfn(nid, start_pfn, nr_pages);
 783}
 784
 785int __ref online_pages(unsigned long pfn, unsigned long nr_pages, int online_type)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 786{
 787	unsigned long flags;
 788	unsigned long onlined_pages = 0;
 789	struct zone *zone;
 790	int need_zonelists_rebuild = 0;
 791	int nid;
 792	int ret;
 793	struct memory_notify arg;
 794	struct memory_block *mem;
 795
 796	mem_hotplug_begin();
 797
 798	/*
 799	 * We can't use pfn_to_nid() because nid might be stored in struct page
 800	 * which is not yet initialized. Instead, we find nid from memory block.
 
 
 
 801	 */
 802	mem = find_memory_block(__pfn_to_section(pfn));
 803	nid = mem->nid;
 804	put_device(&mem->dev);
 
 805
 806	/* associate pfn range with the zone */
 807	zone = zone_for_pfn_range(online_type, nid, pfn, nr_pages);
 808	move_pfn_range_to_zone(zone, pfn, nr_pages, NULL);
 809
 810	arg.start_pfn = pfn;
 811	arg.nr_pages = nr_pages;
 812	node_states_check_changes_online(nr_pages, zone, &arg);
 813
 814	ret = memory_notify(MEM_GOING_ONLINE, &arg);
 815	ret = notifier_to_errno(ret);
 816	if (ret)
 817		goto failed_addition;
 818
 819	/*
 
 
 
 
 
 
 
 
 820	 * If this zone is not populated, then it is not in zonelist.
 821	 * This means the page allocator ignores this zone.
 822	 * So, zonelist must be updated after online.
 823	 */
 824	if (!populated_zone(zone)) {
 825		need_zonelists_rebuild = 1;
 826		setup_zone_pageset(zone);
 827	}
 828
 829	ret = walk_system_ram_range(pfn, nr_pages, &onlined_pages,
 830		online_pages_range);
 831	if (ret) {
 832		/* not a single memory resource was applicable */
 833		if (need_zonelists_rebuild)
 834			zone_pcp_reset(zone);
 835		goto failed_addition;
 836	}
 837
 838	zone->present_pages += onlined_pages;
 839
 840	pgdat_resize_lock(zone->zone_pgdat, &flags);
 841	zone->zone_pgdat->node_present_pages += onlined_pages;
 842	pgdat_resize_unlock(zone->zone_pgdat, &flags);
 843
 844	shuffle_zone(zone);
 845
 846	node_states_set_node(nid, &arg);
 847	if (need_zonelists_rebuild)
 848		build_all_zonelists(NULL);
 849	else
 850		zone_pcp_update(zone);
 851
 
 
 
 
 
 
 
 
 
 
 
 
 852	init_per_zone_wmark_min();
 853
 854	kswapd_run(nid);
 855	kcompactd_run(nid);
 856
 857	vm_total_pages = nr_free_pagecache_pages();
 858
 859	writeback_set_ratelimit();
 860
 861	memory_notify(MEM_ONLINE, &arg);
 862	mem_hotplug_done();
 863	return 0;
 864
 865failed_addition:
 866	pr_debug("online_pages [mem %#010llx-%#010llx] failed\n",
 867		 (unsigned long long) pfn << PAGE_SHIFT,
 868		 (((unsigned long long) pfn + nr_pages) << PAGE_SHIFT) - 1);
 869	memory_notify(MEM_CANCEL_ONLINE, &arg);
 870	mem_hotplug_done();
 871	return ret;
 872}
 873#endif /* CONFIG_MEMORY_HOTPLUG_SPARSE */
 874
 875static void reset_node_present_pages(pg_data_t *pgdat)
 876{
 877	struct zone *z;
 878
 879	for (z = pgdat->node_zones; z < pgdat->node_zones + MAX_NR_ZONES; z++)
 880		z->present_pages = 0;
 881
 882	pgdat->node_present_pages = 0;
 883}
 884
 885/* we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG */
 886static pg_data_t __ref *hotadd_new_pgdat(int nid, u64 start)
 887{
 888	struct pglist_data *pgdat;
 889	unsigned long start_pfn = PFN_DOWN(start);
 890
 
 
 
 
 
 
 891	pgdat = NODE_DATA(nid);
 892	if (!pgdat) {
 893		pgdat = arch_alloc_nodedata(nid);
 894		if (!pgdat)
 895			return NULL;
 896
 897		pgdat->per_cpu_nodestats =
 898			alloc_percpu(struct per_cpu_nodestat);
 899		arch_refresh_nodedata(nid, pgdat);
 900	} else {
 901		int cpu;
 902		/*
 903		 * Reset the nr_zones, order and classzone_idx before reuse.
 904		 * Note that kswapd will init kswapd_classzone_idx properly
 905		 * when it starts in the near future.
 906		 */
 907		pgdat->nr_zones = 0;
 908		pgdat->kswapd_order = 0;
 909		pgdat->kswapd_classzone_idx = 0;
 910		for_each_online_cpu(cpu) {
 911			struct per_cpu_nodestat *p;
 912
 913			p = per_cpu_ptr(pgdat->per_cpu_nodestats, cpu);
 914			memset(p, 0, sizeof(*p));
 915		}
 916	}
 917
 918	/* we can use NODE_DATA(nid) from here */
 919
 920	pgdat->node_id = nid;
 921	pgdat->node_start_pfn = start_pfn;
 922
 923	/* init node's zones as empty zones, we don't have any present pages.*/
 924	free_area_init_core_hotplug(nid);
 925
 926	/*
 927	 * The node we allocated has no zone fallback lists. For avoiding
 928	 * to access not-initialized zonelist, build here.
 929	 */
 930	build_all_zonelists(pgdat);
 931
 932	/*
 933	 * When memory is hot-added, all the memory is in offline state. So
 934	 * clear all zones' present_pages because they will be updated in
 935	 * online_pages() and offline_pages().
 936	 */
 937	reset_node_managed_pages(pgdat);
 938	reset_node_present_pages(pgdat);
 939
 940	return pgdat;
 941}
 942
 943static void rollback_node_hotadd(int nid)
 944{
 945	pg_data_t *pgdat = NODE_DATA(nid);
 946
 947	arch_refresh_nodedata(nid, NULL);
 948	free_percpu(pgdat->per_cpu_nodestats);
 949	arch_free_nodedata(pgdat);
 950}
 951
 952
 953/**
 954 * try_online_node - online a node if offlined
 955 * @nid: the node ID
 956 * @start: start addr of the node
 957 * @set_node_online: Whether we want to online the node
 958 * called by cpu_up() to online a node without onlined memory.
 959 *
 960 * Returns:
 961 * 1 -> a new node has been allocated
 962 * 0 -> the node is already online
 963 * -ENOMEM -> the node could not be allocated
 964 */
 965static int __try_online_node(int nid, u64 start, bool set_node_online)
 966{
 967	pg_data_t *pgdat;
 968	int ret = 1;
 969
 970	if (node_online(nid))
 971		return 0;
 972
 973	pgdat = hotadd_new_pgdat(nid, start);
 974	if (!pgdat) {
 975		pr_err("Cannot online node %d due to NULL pgdat\n", nid);
 976		ret = -ENOMEM;
 977		goto out;
 978	}
 979
 980	if (set_node_online) {
 981		node_set_online(nid);
 982		ret = register_one_node(nid);
 983		BUG_ON(ret);
 984	}
 985out:
 986	return ret;
 987}
 988
 989/*
 990 * Users of this function always want to online/register the node
 991 */
 992int try_online_node(int nid)
 993{
 994	int ret;
 995
 996	mem_hotplug_begin();
 997	ret =  __try_online_node(nid, 0, true);
 998	mem_hotplug_done();
 999	return ret;
1000}
1001
1002static int check_hotplug_memory_range(u64 start, u64 size)
1003{
1004	/* memory range must be block size aligned */
1005	if (!size || !IS_ALIGNED(start, memory_block_size_bytes()) ||
1006	    !IS_ALIGNED(size, memory_block_size_bytes())) {
1007		pr_err("Block size [%#lx] unaligned hotplug range: start %#llx, size %#llx",
1008		       memory_block_size_bytes(), start, size);
1009		return -EINVAL;
1010	}
1011
1012	return 0;
1013}
1014
1015static int online_memory_block(struct memory_block *mem, void *arg)
1016{
 
1017	return device_online(&mem->dev);
1018}
1019
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1020/*
1021 * NOTE: The caller must call lock_device_hotplug() to serialize hotplug
1022 * and online/offline operations (triggered e.g. by sysfs).
1023 *
1024 * we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG
1025 */
1026int __ref add_memory_resource(int nid, struct resource *res)
1027{
1028	struct mhp_restrictions restrictions = {};
 
 
1029	u64 start, size;
1030	bool new_node = false;
1031	int ret;
1032
1033	start = res->start;
1034	size = resource_size(res);
1035
1036	ret = check_hotplug_memory_range(start, size);
1037	if (ret)
1038		return ret;
1039
 
 
 
 
 
 
 
 
 
 
 
 
1040	mem_hotplug_begin();
1041
1042	/*
1043	 * Add new range to memblock so that when hotadd_new_pgdat() is called
1044	 * to allocate new pgdat, get_pfn_range_for_nid() will be able to find
1045	 * this new range and calculate total pages correctly.  The range will
1046	 * be removed at hot-remove time.
1047	 */
1048	memblock_add_node(start, size, nid);
1049
1050	ret = __try_online_node(nid, start, false);
1051	if (ret < 0)
1052		goto error;
1053	new_node = ret;
1054
1055	/* call arch's memory hotadd */
1056	ret = arch_add_memory(nid, start, size, &restrictions);
1057	if (ret < 0)
1058		goto error;
 
 
 
 
 
 
 
 
1059
1060	/* create memory block devices after memory was added */
1061	ret = create_memory_block_devices(start, size);
1062	if (ret) {
1063		arch_remove_memory(nid, start, size, NULL);
1064		goto error;
 
1065	}
1066
1067	if (new_node) {
1068		/* If sysfs file of new node can't be created, cpu on the node
1069		 * can't be hot-added. There is no rollback way now.
1070		 * So, check by BUG_ON() to catch it reluctantly..
1071		 * We online node here. We can't roll back from here.
1072		 */
1073		node_set_online(nid);
1074		ret = __register_one_node(nid);
1075		BUG_ON(ret);
1076	}
1077
1078	/* link memory sections under this node.*/
1079	ret = link_mem_sections(nid, PFN_DOWN(start), PFN_UP(start + size - 1));
1080	BUG_ON(ret);
1081
1082	/* create new memmap entry */
1083	firmware_map_add_hotplug(start, start + size, "System RAM");
 
1084
1085	/* device_online() will take the lock when calling online_pages() */
1086	mem_hotplug_done();
1087
 
 
 
 
 
 
 
1088	/* online pages if requested */
1089	if (memhp_auto_online)
1090		walk_memory_blocks(start, size, NULL, online_memory_block);
1091
1092	return ret;
1093error:
1094	/* rollback pgdat allocation and others */
1095	if (new_node)
1096		rollback_node_hotadd(nid);
1097	memblock_remove(start, size);
1098	mem_hotplug_done();
1099	return ret;
1100}
1101
1102/* requires device_hotplug_lock, see add_memory_resource() */
1103int __ref __add_memory(int nid, u64 start, u64 size)
1104{
1105	struct resource *res;
1106	int ret;
1107
1108	res = register_memory_resource(start, size);
1109	if (IS_ERR(res))
1110		return PTR_ERR(res);
1111
1112	ret = add_memory_resource(nid, res);
1113	if (ret < 0)
1114		release_memory_resource(res);
1115	return ret;
1116}
1117
1118int add_memory(int nid, u64 start, u64 size)
1119{
1120	int rc;
1121
1122	lock_device_hotplug();
1123	rc = __add_memory(nid, start, size);
1124	unlock_device_hotplug();
1125
1126	return rc;
1127}
1128EXPORT_SYMBOL_GPL(add_memory);
1129
1130#ifdef CONFIG_MEMORY_HOTREMOVE
1131/*
1132 * A free page on the buddy free lists (not the per-cpu lists) has PageBuddy
1133 * set and the size of the free page is given by page_order(). Using this,
1134 * the function determines if the pageblock contains only free pages.
1135 * Due to buddy contraints, a free page at least the size of a pageblock will
1136 * be located at the start of the pageblock
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1137 */
1138static inline int pageblock_free(struct page *page)
 
1139{
1140	return PageBuddy(page) && page_order(page) >= pageblock_order;
1141}
1142
1143/* Return the pfn of the start of the next active pageblock after a given pfn */
1144static unsigned long next_active_pageblock(unsigned long pfn)
1145{
1146	struct page *page = pfn_to_page(pfn);
1147
1148	/* Ensure the starting page is pageblock-aligned */
1149	BUG_ON(pfn & (pageblock_nr_pages - 1));
1150
1151	/* If the entire pageblock is free, move to the end of free page */
1152	if (pageblock_free(page)) {
1153		int order;
1154		/* be careful. we don't have locks, page_order can be changed.*/
1155		order = page_order(page);
1156		if ((order < MAX_ORDER) && (order >= pageblock_order))
1157			return pfn + (1 << order);
1158	}
1159
1160	return pfn + pageblock_nr_pages;
 
 
 
 
 
 
1161}
 
1162
1163static bool is_pageblock_removable_nolock(unsigned long pfn)
 
 
 
 
 
 
 
 
 
 
 
 
 
1164{
1165	struct page *page = pfn_to_page(pfn);
1166	struct zone *zone;
1167
1168	/*
1169	 * We have to be careful here because we are iterating over memory
1170	 * sections which are not zone aware so we might end up outside of
1171	 * the zone but still within the section.
1172	 * We have to take care about the node as well. If the node is offline
1173	 * its NODE_DATA will be NULL - see page_zone.
1174	 */
1175	if (!node_online(page_to_nid(page)))
1176		return false;
1177
1178	zone = page_zone(page);
1179	pfn = page_to_pfn(page);
1180	if (!zone_spans_pfn(zone, pfn))
1181		return false;
1182
1183	return !has_unmovable_pages(zone, page, 0, MIGRATE_MOVABLE, SKIP_HWPOISON);
1184}
1185
1186/* Checks if this range of memory is likely to be hot-removable. */
1187bool is_mem_section_removable(unsigned long start_pfn, unsigned long nr_pages)
1188{
1189	unsigned long end_pfn, pfn;
 
1190
1191	end_pfn = min(start_pfn + nr_pages,
1192			zone_end_pfn(page_zone(pfn_to_page(start_pfn))));
1193
1194	/* Check the starting page of each pageblock within the range */
1195	for (pfn = start_pfn; pfn < end_pfn; pfn = next_active_pageblock(pfn)) {
1196		if (!is_pageblock_removable_nolock(pfn))
1197			return false;
1198		cond_resched();
 
 
1199	}
1200
1201	/* All pageblocks in the memory block are likely to be hot-removable */
1202	return true;
1203}
 
1204
1205/*
1206 * Confirm all pages in a range [start, end) belong to the same zone.
1207 * When true, return its valid [start, end).
1208 */
1209int test_pages_in_a_zone(unsigned long start_pfn, unsigned long end_pfn,
1210			 unsigned long *valid_start, unsigned long *valid_end)
1211{
1212	unsigned long pfn, sec_end_pfn;
1213	unsigned long start, end;
1214	struct zone *zone = NULL;
1215	struct page *page;
1216	int i;
1217	for (pfn = start_pfn, sec_end_pfn = SECTION_ALIGN_UP(start_pfn + 1);
1218	     pfn < end_pfn;
1219	     pfn = sec_end_pfn, sec_end_pfn += PAGES_PER_SECTION) {
1220		/* Make sure the memory section is present first */
1221		if (!present_section_nr(pfn_to_section_nr(pfn)))
1222			continue;
1223		for (; pfn < sec_end_pfn && pfn < end_pfn;
1224		     pfn += MAX_ORDER_NR_PAGES) {
1225			i = 0;
1226			/* This is just a CONFIG_HOLES_IN_ZONE check.*/
1227			while ((i < MAX_ORDER_NR_PAGES) &&
1228				!pfn_valid_within(pfn + i))
1229				i++;
1230			if (i == MAX_ORDER_NR_PAGES || pfn + i >= end_pfn)
1231				continue;
1232			/* Check if we got outside of the zone */
1233			if (zone && !zone_spans_pfn(zone, pfn + i))
1234				return 0;
1235			page = pfn_to_page(pfn + i);
1236			if (zone && page_zone(page) != zone)
1237				return 0;
1238			if (!zone)
1239				start = pfn + i;
1240			zone = page_zone(page);
1241			end = pfn + MAX_ORDER_NR_PAGES;
1242		}
1243	}
1244
1245	if (zone) {
1246		*valid_start = start;
1247		*valid_end = min(end, end_pfn);
1248		return 1;
1249	} else {
1250		return 0;
1251	}
1252}
1253
 
1254/*
1255 * Scan pfn range [start,end) to find movable/migratable pages (LRU pages,
1256 * non-lru movable pages and hugepages). We scan pfn because it's much
1257 * easier than scanning over linked list. This function returns the pfn
1258 * of the first found movable page if it's found, otherwise 0.
 
 
 
 
 
1259 */
1260static unsigned long scan_movable_pages(unsigned long start, unsigned long end)
 
1261{
1262	unsigned long pfn;
1263
1264	for (pfn = start; pfn < end; pfn++) {
1265		struct page *page, *head;
1266		unsigned long skip;
1267
1268		if (!pfn_valid(pfn))
1269			continue;
1270		page = pfn_to_page(pfn);
1271		if (PageLRU(page))
1272			return pfn;
1273		if (__PageMovable(page))
1274			return pfn;
 
 
 
 
 
 
 
 
 
1275
1276		if (!PageHuge(page))
1277			continue;
1278		head = compound_head(page);
1279		if (page_huge_active(head))
1280			return pfn;
1281		skip = compound_nr(head) - (page - head);
1282		pfn += skip - 1;
1283	}
 
 
 
 
 
 
 
 
 
1284	return 0;
1285}
1286
1287static struct page *new_node_page(struct page *page, unsigned long private)
1288{
1289	int nid = page_to_nid(page);
1290	nodemask_t nmask = node_states[N_MEMORY];
1291
1292	/*
1293	 * try to allocate from a different node but reuse this node if there
1294	 * are no other online nodes to be used (e.g. we are offlining a part
1295	 * of the only existing node)
1296	 */
1297	node_clear(nid, nmask);
1298	if (nodes_empty(nmask))
1299		node_set(nid, nmask);
1300
1301	return new_page_nodemask(page, nid, &nmask);
1302}
1303
1304static int
1305do_migrate_range(unsigned long start_pfn, unsigned long end_pfn)
1306{
 
1307	unsigned long pfn;
1308	struct page *page;
1309	int ret = 0;
1310	LIST_HEAD(source);
 
 
1311
1312	for (pfn = start_pfn; pfn < end_pfn; pfn++) {
 
 
1313		if (!pfn_valid(pfn))
1314			continue;
1315		page = pfn_to_page(pfn);
1316
1317		if (PageHuge(page)) {
1318			struct page *head = compound_head(page);
1319			pfn = page_to_pfn(head) + compound_nr(head) - 1;
1320			isolate_huge_page(head, &source);
1321			continue;
1322		} else if (PageTransHuge(page))
1323			pfn = page_to_pfn(compound_head(page))
1324				+ hpage_nr_pages(page) - 1;
1325
1326		/*
1327		 * HWPoison pages have elevated reference counts so the migration would
1328		 * fail on them. It also doesn't make any sense to migrate them in the
1329		 * first place. Still try to unmap such a page in case it is still mapped
1330		 * (e.g. current hwpoison implementation doesn't unmap KSM pages but keep
1331		 * the unmap as the catch all safety net).
1332		 */
1333		if (PageHWPoison(page)) {
1334			if (WARN_ON(PageLRU(page)))
1335				isolate_lru_page(page);
1336			if (page_mapped(page))
1337				try_to_unmap(page, TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS);
1338			continue;
1339		}
1340
1341		if (!get_page_unless_zero(page))
1342			continue;
1343		/*
1344		 * We can skip free pages. And we can deal with pages on
1345		 * LRU and non-lru movable pages.
1346		 */
1347		if (PageLRU(page))
1348			ret = isolate_lru_page(page);
1349		else
1350			ret = isolate_movable_page(page, ISOLATE_UNEVICTABLE);
1351		if (!ret) { /* Success */
1352			list_add_tail(&page->lru, &source);
1353			if (!__PageMovable(page))
1354				inc_node_page_state(page, NR_ISOLATED_ANON +
1355						    page_is_file_cache(page));
1356
1357		} else {
1358			pr_warn("failed to isolate pfn %lx\n", pfn);
1359			dump_page(page, "isolation failed");
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1360		}
1361		put_page(page);
 
1362	}
1363	if (!list_empty(&source)) {
1364		/* Allocate a new page from the nearest neighbor node */
1365		ret = migrate_pages(&source, new_node_page, NULL, 0,
1366					MIGRATE_SYNC, MR_MEMORY_HOTPLUG);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1367		if (ret) {
1368			list_for_each_entry(page, &source, lru) {
1369				pr_warn("migrating pfn %lx failed ret:%d ",
1370				       page_to_pfn(page), ret);
1371				dump_page(page, "migration failure");
 
 
 
1372			}
1373			putback_movable_pages(&source);
1374		}
1375	}
1376
1377	return ret;
1378}
1379
1380/*
1381 * remove from free_area[] and mark all as Reserved.
1382 */
1383static int
1384offline_isolated_pages_cb(unsigned long start, unsigned long nr_pages,
1385			void *data)
1386{
1387	unsigned long *offlined_pages = (unsigned long *)data;
1388
1389	*offlined_pages += __offline_isolated_pages(start, start + nr_pages);
1390	return 0;
1391}
1392
1393/*
1394 * Check all pages in range, recoreded as memory resource, are isolated.
1395 */
1396static int
1397check_pages_isolated_cb(unsigned long start_pfn, unsigned long nr_pages,
1398			void *data)
1399{
1400	return test_pages_isolated(start_pfn, start_pfn + nr_pages, true);
1401}
1402
1403static int __init cmdline_parse_movable_node(char *p)
1404{
1405#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
1406	movable_node_enabled = true;
1407#else
1408	pr_warn("movable_node parameter depends on CONFIG_HAVE_MEMBLOCK_NODE_MAP to work properly\n");
1409#endif
1410	return 0;
1411}
1412early_param("movable_node", cmdline_parse_movable_node);
1413
1414/* check which state of node_states will be changed when offline memory */
1415static void node_states_check_changes_offline(unsigned long nr_pages,
1416		struct zone *zone, struct memory_notify *arg)
1417{
1418	struct pglist_data *pgdat = zone->zone_pgdat;
1419	unsigned long present_pages = 0;
1420	enum zone_type zt;
1421
1422	arg->status_change_nid = NUMA_NO_NODE;
1423	arg->status_change_nid_normal = NUMA_NO_NODE;
1424	arg->status_change_nid_high = NUMA_NO_NODE;
1425
1426	/*
1427	 * Check whether node_states[N_NORMAL_MEMORY] will be changed.
1428	 * If the memory to be offline is within the range
1429	 * [0..ZONE_NORMAL], and it is the last present memory there,
1430	 * the zones in that range will become empty after the offlining,
1431	 * thus we can determine that we need to clear the node from
1432	 * node_states[N_NORMAL_MEMORY].
1433	 */
1434	for (zt = 0; zt <= ZONE_NORMAL; zt++)
1435		present_pages += pgdat->node_zones[zt].present_pages;
1436	if (zone_idx(zone) <= ZONE_NORMAL && nr_pages >= present_pages)
1437		arg->status_change_nid_normal = zone_to_nid(zone);
1438
1439#ifdef CONFIG_HIGHMEM
1440	/*
1441	 * node_states[N_HIGH_MEMORY] contains nodes which
1442	 * have normal memory or high memory.
1443	 * Here we add the present_pages belonging to ZONE_HIGHMEM.
1444	 * If the zone is within the range of [0..ZONE_HIGHMEM), and
1445	 * we determine that the zones in that range become empty,
1446	 * we need to clear the node for N_HIGH_MEMORY.
1447	 */
1448	present_pages += pgdat->node_zones[ZONE_HIGHMEM].present_pages;
1449	if (zone_idx(zone) <= ZONE_HIGHMEM && nr_pages >= present_pages)
1450		arg->status_change_nid_high = zone_to_nid(zone);
1451#endif
1452
1453	/*
1454	 * We have accounted the pages from [0..ZONE_NORMAL), and
1455	 * in case of CONFIG_HIGHMEM the pages from ZONE_HIGHMEM
1456	 * as well.
1457	 * Here we count the possible pages from ZONE_MOVABLE.
1458	 * If after having accounted all the pages, we see that the nr_pages
1459	 * to be offlined is over or equal to the accounted pages,
1460	 * we know that the node will become empty, and so, we can clear
1461	 * it for N_MEMORY as well.
1462	 */
1463	present_pages += pgdat->node_zones[ZONE_MOVABLE].present_pages;
1464
1465	if (nr_pages >= present_pages)
1466		arg->status_change_nid = zone_to_nid(zone);
1467}
1468
1469static void node_states_clear_node(int node, struct memory_notify *arg)
1470{
1471	if (arg->status_change_nid_normal >= 0)
1472		node_clear_state(node, N_NORMAL_MEMORY);
1473
1474	if (arg->status_change_nid_high >= 0)
1475		node_clear_state(node, N_HIGH_MEMORY);
1476
1477	if (arg->status_change_nid >= 0)
1478		node_clear_state(node, N_MEMORY);
1479}
1480
1481static int __ref __offline_pages(unsigned long start_pfn,
1482		  unsigned long end_pfn)
1483{
1484	unsigned long pfn, nr_pages;
1485	unsigned long offlined_pages = 0;
1486	int ret, node, nr_isolate_pageblock;
 
 
 
 
 
 
 
 
 
 
 
 
1487	unsigned long flags;
1488	unsigned long valid_start, valid_end;
1489	struct zone *zone;
1490	struct memory_notify arg;
1491	char *reason;
 
1492
1493	mem_hotplug_begin();
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1494
1495	/* This makes hotplug much easier...and readable.
1496	   we assume this for now. .*/
1497	if (!test_pages_in_a_zone(start_pfn, end_pfn, &valid_start,
1498				  &valid_end)) {
 
 
 
1499		ret = -EINVAL;
1500		reason = "multizone range";
1501		goto failed_removal;
1502	}
1503
1504	zone = page_zone(pfn_to_page(valid_start));
1505	node = zone_to_nid(zone);
1506	nr_pages = end_pfn - start_pfn;
 
 
 
1507
1508	/* set above range as isolated */
1509	ret = start_isolate_page_range(start_pfn, end_pfn,
1510				       MIGRATE_MOVABLE,
1511				       SKIP_HWPOISON | REPORT_FAILURE);
1512	if (ret < 0) {
 
1513		reason = "failure to isolate range";
1514		goto failed_removal;
1515	}
1516	nr_isolate_pageblock = ret;
1517
1518	arg.start_pfn = start_pfn;
1519	arg.nr_pages = nr_pages;
1520	node_states_check_changes_offline(nr_pages, zone, &arg);
1521
1522	ret = memory_notify(MEM_GOING_OFFLINE, &arg);
1523	ret = notifier_to_errno(ret);
1524	if (ret) {
1525		reason = "notifier failure";
1526		goto failed_removal_isolated;
1527	}
1528
1529	do {
1530		for (pfn = start_pfn; pfn;) {
 
 
 
 
 
 
1531			if (signal_pending(current)) {
1532				ret = -EINTR;
1533				reason = "signal backoff";
1534				goto failed_removal_isolated;
1535			}
1536
1537			cond_resched();
1538			lru_add_drain_all();
1539
1540			pfn = scan_movable_pages(pfn, end_pfn);
1541			if (pfn) {
1542				/*
1543				 * TODO: fatal migration failures should bail
1544				 * out
1545				 */
1546				do_migrate_range(pfn, end_pfn);
1547			}
 
 
 
 
 
1548		}
1549
1550		/*
1551		 * Dissolve free hugepages in the memory block before doing
1552		 * offlining actually in order to make hugetlbfs's object
1553		 * counting consistent.
1554		 */
1555		ret = dissolve_free_huge_pages(start_pfn, end_pfn);
1556		if (ret) {
1557			reason = "failure to dissolve huge pages";
1558			goto failed_removal_isolated;
1559		}
1560		/* check again */
1561		ret = walk_system_ram_range(start_pfn, end_pfn - start_pfn,
1562					    NULL, check_pages_isolated_cb);
1563	} while (ret);
1564
1565	/* Ok, all of our target is isolated.
1566	   We cannot do rollback at this point. */
1567	walk_system_ram_range(start_pfn, end_pfn - start_pfn,
1568			      &offlined_pages, offline_isolated_pages_cb);
1569	pr_info("Offlined Pages %ld\n", offlined_pages);
1570	/*
1571	 * Onlining will reset pagetype flags and makes migrate type
1572	 * MOVABLE, so just need to decrease the number of isolated
1573	 * pageblocks zone counter here.
1574	 */
1575	spin_lock_irqsave(&zone->lock, flags);
1576	zone->nr_isolate_pageblock -= nr_isolate_pageblock;
1577	spin_unlock_irqrestore(&zone->lock, flags);
1578
1579	/* removal success */
1580	adjust_managed_page_count(pfn_to_page(start_pfn), -offlined_pages);
1581	zone->present_pages -= offlined_pages;
1582
1583	pgdat_resize_lock(zone->zone_pgdat, &flags);
1584	zone->zone_pgdat->node_present_pages -= offlined_pages;
1585	pgdat_resize_unlock(zone->zone_pgdat, &flags);
1586
 
1587	init_per_zone_wmark_min();
1588
 
 
 
 
 
1589	if (!populated_zone(zone)) {
1590		zone_pcp_reset(zone);
1591		build_all_zonelists(NULL);
1592	} else
1593		zone_pcp_update(zone);
1594
1595	node_states_clear_node(node, &arg);
1596	if (arg.status_change_nid >= 0) {
1597		kswapd_stop(node);
1598		kcompactd_stop(node);
 
1599	}
1600
1601	vm_total_pages = nr_free_pagecache_pages();
1602	writeback_set_ratelimit();
1603
1604	memory_notify(MEM_OFFLINE, &arg);
1605	mem_hotplug_done();
1606	return 0;
1607
1608failed_removal_isolated:
 
1609	undo_isolate_page_range(start_pfn, end_pfn, MIGRATE_MOVABLE);
1610	memory_notify(MEM_CANCEL_OFFLINE, &arg);
 
 
 
1611failed_removal:
1612	pr_debug("memory offlining [mem %#010llx-%#010llx] failed due to %s\n",
1613		 (unsigned long long) start_pfn << PAGE_SHIFT,
1614		 ((unsigned long long) end_pfn << PAGE_SHIFT) - 1,
1615		 reason);
1616	/* pushback to free area */
1617	mem_hotplug_done();
1618	return ret;
1619}
1620
1621int offline_pages(unsigned long start_pfn, unsigned long nr_pages)
1622{
1623	return __offline_pages(start_pfn, start_pfn + nr_pages);
1624}
1625
1626static int check_memblock_offlined_cb(struct memory_block *mem, void *arg)
1627{
1628	int ret = !is_memblock_offlined(mem);
1629
1630	if (unlikely(ret)) {
 
1631		phys_addr_t beginpa, endpa;
1632
1633		beginpa = PFN_PHYS(section_nr_to_pfn(mem->start_section_nr));
1634		endpa = beginpa + memory_block_size_bytes() - 1;
1635		pr_warn("removing memory fails, because memory [%pa-%pa] is onlined\n",
1636			&beginpa, &endpa);
1637
1638		return -EBUSY;
1639	}
1640	return 0;
1641}
1642
1643static int check_cpu_on_node(pg_data_t *pgdat)
 
 
 
 
 
 
 
 
 
 
1644{
1645	int cpu;
1646
1647	for_each_present_cpu(cpu) {
1648		if (cpu_to_node(cpu) == pgdat->node_id)
1649			/*
1650			 * the cpu on this node isn't removed, and we can't
1651			 * offline this node.
1652			 */
1653			return -EBUSY;
1654	}
1655
1656	return 0;
1657}
1658
1659static int check_no_memblock_for_node_cb(struct memory_block *mem, void *arg)
1660{
1661	int nid = *(int *)arg;
1662
1663	/*
1664	 * If a memory block belongs to multiple nodes, the stored nid is not
1665	 * reliable. However, such blocks are always online (e.g., cannot get
1666	 * offlined) and, therefore, are still spanned by the node.
1667	 */
1668	return mem->nid == nid ? -EEXIST : 0;
1669}
1670
1671/**
1672 * try_offline_node
1673 * @nid: the node ID
1674 *
1675 * Offline a node if all memory sections and cpus of the node are removed.
1676 *
1677 * NOTE: The caller must call lock_device_hotplug() to serialize hotplug
1678 * and online/offline operations before this call.
1679 */
1680void try_offline_node(int nid)
1681{
1682	pg_data_t *pgdat = NODE_DATA(nid);
1683	int rc;
1684
1685	/*
1686	 * If the node still spans pages (especially ZONE_DEVICE), don't
1687	 * offline it. A node spans memory after move_pfn_range_to_zone(),
1688	 * e.g., after the memory block was onlined.
1689	 */
1690	if (pgdat->node_spanned_pages)
1691		return;
1692
1693	/*
1694	 * Especially offline memory blocks might not be spanned by the
1695	 * node. They will get spanned by the node once they get onlined.
1696	 * However, they link to the node in sysfs and can get onlined later.
1697	 */
1698	rc = for_each_memory_block(&nid, check_no_memblock_for_node_cb);
1699	if (rc)
1700		return;
1701
1702	if (check_cpu_on_node(pgdat))
1703		return;
1704
1705	/*
1706	 * all memory/cpu of this node are removed, we can offline this
1707	 * node now.
1708	 */
1709	node_set_offline(nid);
1710	unregister_one_node(nid);
1711}
1712EXPORT_SYMBOL(try_offline_node);
1713
1714static void __release_memory_resource(resource_size_t start,
1715				      resource_size_t size)
1716{
1717	int ret;
 
1718
1719	/*
1720	 * When removing memory in the same granularity as it was added,
1721	 * this function never fails. It might only fail if resources
1722	 * have to be adjusted or split. We'll ignore the error, as
1723	 * removing of memory cannot fail.
1724	 */
1725	ret = release_mem_region_adjustable(&iomem_resource, start, size);
1726	if (ret) {
1727		resource_size_t endres = start + size - 1;
1728
1729		pr_warn("Unable to release resource <%pa-%pa> (%d)\n",
1730			&start, &endres, ret);
1731	}
 
 
 
 
 
 
 
1732}
1733
1734static int __ref try_remove_memory(int nid, u64 start, u64 size)
1735{
1736	int rc = 0;
1737
1738	BUG_ON(check_hotplug_memory_range(start, size));
1739
1740	mem_hotplug_begin();
1741
1742	/*
1743	 * All memory blocks must be offlined before removing memory.  Check
1744	 * whether all memory blocks in question are offline and return error
1745	 * if this is not the case.
 
 
 
 
1746	 */
1747	rc = walk_memory_blocks(start, size, NULL, check_memblock_offlined_cb);
1748	if (rc)
1749		goto done;
1750
1751	/* remove memmap entry */
1752	firmware_map_remove(start, start + size, "System RAM");
1753	memblock_free(start, size);
1754	memblock_remove(start, size);
1755
1756	/* remove memory block devices before removing memory */
1757	remove_memory_block_devices(start, size);
1758
1759	arch_remove_memory(nid, start, size, NULL);
1760	__release_memory_resource(start, size);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1761
1762	try_offline_node(nid);
 
1763
1764done:
1765	mem_hotplug_done();
1766	return rc;
1767}
1768
1769/**
1770 * remove_memory
1771 * @nid: the node ID
1772 * @start: physical address of the region to remove
1773 * @size: size of the region to remove
1774 *
1775 * NOTE: The caller must call lock_device_hotplug() to serialize hotplug
1776 * and online/offline operations before this call, as required by
1777 * try_offline_node().
1778 */
1779void __remove_memory(int nid, u64 start, u64 size)
1780{
1781
1782	/*
1783	 * trigger BUG() if some memory is not offlined prior to calling this
1784	 * function
1785	 */
1786	if (try_remove_memory(nid, start, size))
1787		BUG();
1788}
1789
1790/*
1791 * Remove memory if every memory block is offline, otherwise return -EBUSY is
1792 * some memory is not offline
1793 */
1794int remove_memory(int nid, u64 start, u64 size)
1795{
1796	int rc;
1797
1798	lock_device_hotplug();
1799	rc  = try_remove_memory(nid, start, size);
1800	unlock_device_hotplug();
1801
1802	return rc;
1803}
1804EXPORT_SYMBOL_GPL(remove_memory);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1805#endif /* CONFIG_MEMORY_HOTREMOVE */
v6.13.7
   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 __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 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
 631/* we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG */
 632void generic_online_page(struct page *page, unsigned int order)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 633{
 634	__free_pages_core(page, order, MEMINIT_HOTPLUG);
 
 
 
 
 
 
 635}
 636EXPORT_SYMBOL_GPL(generic_online_page);
 637
 638static void online_pages_range(unsigned long start_pfn, unsigned long nr_pages)
 
 639{
 640	const unsigned long end_pfn = start_pfn + nr_pages;
 641	unsigned long pfn;
 
 642
 643	/*
 644	 * Online the pages in MAX_PAGE_ORDER aligned chunks. The callback might
 645	 * decide to not expose all pages to the buddy (e.g., expose them
 646	 * later). We account all pages as being online and belonging to this
 647	 * zone ("present").
 648	 * When using memmap_on_memory, the range might not be aligned to
 649	 * MAX_ORDER_NR_PAGES - 1, but pageblock aligned. __ffs() will detect
 650	 * this and the first chunk to online will be pageblock_nr_pages.
 651	 */
 652	for (pfn = start_pfn; pfn < end_pfn;) {
 653		int order;
 654
 655		/*
 656		 * Free to online pages in the largest chunks alignment allows.
 657		 *
 658		 * __ffs() behaviour is undefined for 0. start == 0 is
 659		 * MAX_PAGE_ORDER-aligned, Set order to MAX_PAGE_ORDER for
 660		 * the case.
 661		 */
 662		if (pfn)
 663			order = min_t(int, MAX_PAGE_ORDER, __ffs(pfn));
 664		else
 665			order = MAX_PAGE_ORDER;
 666
 667		(*online_page_callback)(pfn_to_page(pfn), order);
 668		pfn += (1UL << order);
 669	}
 670
 671	/* mark all involved sections as online */
 672	online_mem_sections(start_pfn, end_pfn);
 
 
 
 673}
 674
 675/* check which state of node_states will be changed when online memory */
 676static void node_states_check_changes_online(unsigned long nr_pages,
 677	struct zone *zone, struct memory_notify *arg)
 678{
 679	int nid = zone_to_nid(zone);
 680
 681	arg->status_change_nid = NUMA_NO_NODE;
 682	arg->status_change_nid_normal = NUMA_NO_NODE;
 
 683
 684	if (!node_state(nid, N_MEMORY))
 685		arg->status_change_nid = nid;
 686	if (zone_idx(zone) <= ZONE_NORMAL && !node_state(nid, N_NORMAL_MEMORY))
 687		arg->status_change_nid_normal = nid;
 
 
 
 
 688}
 689
 690static void node_states_set_node(int node, struct memory_notify *arg)
 691{
 692	if (arg->status_change_nid_normal >= 0)
 693		node_set_state(node, N_NORMAL_MEMORY);
 694
 
 
 
 695	if (arg->status_change_nid >= 0)
 696		node_set_state(node, N_MEMORY);
 697}
 698
 699static void __meminit resize_zone_range(struct zone *zone, unsigned long start_pfn,
 700		unsigned long nr_pages)
 701{
 702	unsigned long old_end_pfn = zone_end_pfn(zone);
 703
 704	if (zone_is_empty(zone) || start_pfn < zone->zone_start_pfn)
 705		zone->zone_start_pfn = start_pfn;
 706
 707	zone->spanned_pages = max(start_pfn + nr_pages, old_end_pfn) - zone->zone_start_pfn;
 708}
 709
 710static void __meminit resize_pgdat_range(struct pglist_data *pgdat, unsigned long start_pfn,
 711                                     unsigned long nr_pages)
 712{
 713	unsigned long old_end_pfn = pgdat_end_pfn(pgdat);
 714
 715	if (!pgdat->node_spanned_pages || start_pfn < pgdat->node_start_pfn)
 716		pgdat->node_start_pfn = start_pfn;
 717
 718	pgdat->node_spanned_pages = max(start_pfn + nr_pages, old_end_pfn) - pgdat->node_start_pfn;
 719
 720}
 721
 722#ifdef CONFIG_ZONE_DEVICE
 723static void section_taint_zone_device(unsigned long pfn)
 724{
 725	struct mem_section *ms = __pfn_to_section(pfn);
 726
 727	ms->section_mem_map |= SECTION_TAINT_ZONE_DEVICE;
 728}
 729#else
 730static inline void section_taint_zone_device(unsigned long pfn)
 731{
 732}
 733#endif
 734
 735/*
 736 * Associate the pfn range with the given zone, initializing the memmaps
 737 * and resizing the pgdat/zone data to span the added pages. After this
 738 * call, all affected pages are PageOffline().
 739 *
 740 * All aligned pageblocks are initialized to the specified migratetype
 741 * (usually MIGRATE_MOVABLE). Besides setting the migratetype, no related
 742 * zone stats (e.g., nr_isolate_pageblock) are touched.
 743 */
 744void move_pfn_range_to_zone(struct zone *zone, unsigned long start_pfn,
 745				  unsigned long nr_pages,
 746				  struct vmem_altmap *altmap, int migratetype)
 747{
 748	struct pglist_data *pgdat = zone->zone_pgdat;
 749	int nid = pgdat->node_id;
 
 750
 751	clear_zone_contiguous(zone);
 752
 
 
 
 753	if (zone_is_empty(zone))
 754		init_currently_empty_zone(zone, start_pfn, nr_pages);
 755	resize_zone_range(zone, start_pfn, nr_pages);
 
 756	resize_pgdat_range(pgdat, start_pfn, nr_pages);
 757
 758	/*
 759	 * Subsection population requires care in pfn_to_online_page().
 760	 * Set the taint to enable the slow path detection of
 761	 * ZONE_DEVICE pages in an otherwise  ZONE_{NORMAL,MOVABLE}
 762	 * section.
 763	 */
 764	if (zone_is_zone_device(zone)) {
 765		if (!IS_ALIGNED(start_pfn, PAGES_PER_SECTION))
 766			section_taint_zone_device(start_pfn);
 767		if (!IS_ALIGNED(start_pfn + nr_pages, PAGES_PER_SECTION))
 768			section_taint_zone_device(start_pfn + nr_pages);
 769	}
 770
 771	/*
 772	 * TODO now we have a visible range of pages which are not associated
 773	 * with their zone properly. Not nice but set_pfnblock_flags_mask
 774	 * expects the zone spans the pfn range. All the pages in the range
 775	 * are reserved so nobody should be touching them so we should be safe
 776	 */
 777	memmap_init_range(nr_pages, nid, zone_idx(zone), start_pfn, 0,
 778			 MEMINIT_HOTPLUG, altmap, migratetype);
 779
 780	set_zone_contiguous(zone);
 781}
 782
 783struct auto_movable_stats {
 784	unsigned long kernel_early_pages;
 785	unsigned long movable_pages;
 786};
 787
 788static void auto_movable_stats_account_zone(struct auto_movable_stats *stats,
 789					    struct zone *zone)
 790{
 791	if (zone_idx(zone) == ZONE_MOVABLE) {
 792		stats->movable_pages += zone->present_pages;
 793	} else {
 794		stats->kernel_early_pages += zone->present_early_pages;
 795#ifdef CONFIG_CMA
 796		/*
 797		 * CMA pages (never on hotplugged memory) behave like
 798		 * ZONE_MOVABLE.
 799		 */
 800		stats->movable_pages += zone->cma_pages;
 801		stats->kernel_early_pages -= zone->cma_pages;
 802#endif /* CONFIG_CMA */
 803	}
 804}
 805struct auto_movable_group_stats {
 806	unsigned long movable_pages;
 807	unsigned long req_kernel_early_pages;
 808};
 809
 810static int auto_movable_stats_account_group(struct memory_group *group,
 811					   void *arg)
 812{
 813	const int ratio = READ_ONCE(auto_movable_ratio);
 814	struct auto_movable_group_stats *stats = arg;
 815	long pages;
 816
 817	/*
 818	 * We don't support modifying the config while the auto-movable online
 819	 * policy is already enabled. Just avoid the division by zero below.
 820	 */
 821	if (!ratio)
 822		return 0;
 823
 824	/*
 825	 * Calculate how many early kernel pages this group requires to
 826	 * satisfy the configured zone ratio.
 827	 */
 828	pages = group->present_movable_pages * 100 / ratio;
 829	pages -= group->present_kernel_pages;
 830
 831	if (pages > 0)
 832		stats->req_kernel_early_pages += pages;
 833	stats->movable_pages += group->present_movable_pages;
 834	return 0;
 835}
 836
 837static bool auto_movable_can_online_movable(int nid, struct memory_group *group,
 838					    unsigned long nr_pages)
 839{
 840	unsigned long kernel_early_pages, movable_pages;
 841	struct auto_movable_group_stats group_stats = {};
 842	struct auto_movable_stats stats = {};
 843	struct zone *zone;
 844	int i;
 845
 846	/* Walk all relevant zones and collect MOVABLE vs. KERNEL stats. */
 847	if (nid == NUMA_NO_NODE) {
 848		/* TODO: cache values */
 849		for_each_populated_zone(zone)
 850			auto_movable_stats_account_zone(&stats, zone);
 851	} else {
 852		for (i = 0; i < MAX_NR_ZONES; i++) {
 853			pg_data_t *pgdat = NODE_DATA(nid);
 854
 855			zone = pgdat->node_zones + i;
 856			if (populated_zone(zone))
 857				auto_movable_stats_account_zone(&stats, zone);
 858		}
 859	}
 860
 861	kernel_early_pages = stats.kernel_early_pages;
 862	movable_pages = stats.movable_pages;
 863
 864	/*
 865	 * Kernel memory inside dynamic memory group allows for more MOVABLE
 866	 * memory within the same group. Remove the effect of all but the
 867	 * current group from the stats.
 868	 */
 869	walk_dynamic_memory_groups(nid, auto_movable_stats_account_group,
 870				   group, &group_stats);
 871	if (kernel_early_pages <= group_stats.req_kernel_early_pages)
 872		return false;
 873	kernel_early_pages -= group_stats.req_kernel_early_pages;
 874	movable_pages -= group_stats.movable_pages;
 875
 876	if (group && group->is_dynamic)
 877		kernel_early_pages += group->present_kernel_pages;
 878
 879	/*
 880	 * Test if we could online the given number of pages to ZONE_MOVABLE
 881	 * and still stay in the configured ratio.
 882	 */
 883	movable_pages += nr_pages;
 884	return movable_pages <= (auto_movable_ratio * kernel_early_pages) / 100;
 885}
 886
 887/*
 888 * Returns a default kernel memory zone for the given pfn range.
 889 * If no kernel zone covers this pfn range it will automatically go
 890 * to the ZONE_NORMAL.
 891 */
 892static struct zone *default_kernel_zone_for_pfn(int nid, unsigned long start_pfn,
 893		unsigned long nr_pages)
 894{
 895	struct pglist_data *pgdat = NODE_DATA(nid);
 896	int zid;
 897
 898	for (zid = 0; zid < ZONE_NORMAL; zid++) {
 899		struct zone *zone = &pgdat->node_zones[zid];
 900
 901		if (zone_intersects(zone, start_pfn, nr_pages))
 902			return zone;
 903	}
 904
 905	return &pgdat->node_zones[ZONE_NORMAL];
 906}
 907
 908/*
 909 * Determine to which zone to online memory dynamically based on user
 910 * configuration and system stats. We care about the following ratio:
 911 *
 912 *   MOVABLE : KERNEL
 913 *
 914 * Whereby MOVABLE is memory in ZONE_MOVABLE and KERNEL is memory in
 915 * one of the kernel zones. CMA pages inside one of the kernel zones really
 916 * behaves like ZONE_MOVABLE, so we treat them accordingly.
 917 *
 918 * We don't allow for hotplugged memory in a KERNEL zone to increase the
 919 * amount of MOVABLE memory we can have, so we end up with:
 920 *
 921 *   MOVABLE : KERNEL_EARLY
 922 *
 923 * Whereby KERNEL_EARLY is memory in one of the kernel zones, available sinze
 924 * boot. We base our calculation on KERNEL_EARLY internally, because:
 925 *
 926 * a) Hotplugged memory in one of the kernel zones can sometimes still get
 927 *    hotunplugged, especially when hot(un)plugging individual memory blocks.
 928 *    There is no coordination across memory devices, therefore "automatic"
 929 *    hotunplugging, as implemented in hypervisors, could result in zone
 930 *    imbalances.
 931 * b) Early/boot memory in one of the kernel zones can usually not get
 932 *    hotunplugged again (e.g., no firmware interface to unplug, fragmented
 933 *    with unmovable allocations). While there are corner cases where it might
 934 *    still work, it is barely relevant in practice.
 935 *
 936 * Exceptions are dynamic memory groups, which allow for more MOVABLE
 937 * memory within the same memory group -- because in that case, there is
 938 * coordination within the single memory device managed by a single driver.
 939 *
 940 * We rely on "present pages" instead of "managed pages", as the latter is
 941 * highly unreliable and dynamic in virtualized environments, and does not
 942 * consider boot time allocations. For example, memory ballooning adjusts the
 943 * managed pages when inflating/deflating the balloon, and balloon compaction
 944 * can even migrate inflated pages between zones.
 945 *
 946 * Using "present pages" is better but some things to keep in mind are:
 947 *
 948 * a) Some memblock allocations, such as for the crashkernel area, are
 949 *    effectively unused by the kernel, yet they account to "present pages".
 950 *    Fortunately, these allocations are comparatively small in relevant setups
 951 *    (e.g., fraction of system memory).
 952 * b) Some hotplugged memory blocks in virtualized environments, esecially
 953 *    hotplugged by virtio-mem, look like they are completely present, however,
 954 *    only parts of the memory block are actually currently usable.
 955 *    "present pages" is an upper limit that can get reached at runtime. As
 956 *    we base our calculations on KERNEL_EARLY, this is not an issue.
 957 */
 958static struct zone *auto_movable_zone_for_pfn(int nid,
 959					      struct memory_group *group,
 960					      unsigned long pfn,
 961					      unsigned long nr_pages)
 962{
 963	unsigned long online_pages = 0, max_pages, end_pfn;
 964	struct page *page;
 965
 966	if (!auto_movable_ratio)
 967		goto kernel_zone;
 968
 969	if (group && !group->is_dynamic) {
 970		max_pages = group->s.max_pages;
 971		online_pages = group->present_movable_pages;
 972
 973		/* If anything is !MOVABLE online the rest !MOVABLE. */
 974		if (group->present_kernel_pages)
 975			goto kernel_zone;
 976	} else if (!group || group->d.unit_pages == nr_pages) {
 977		max_pages = nr_pages;
 978	} else {
 979		max_pages = group->d.unit_pages;
 980		/*
 981		 * Take a look at all online sections in the current unit.
 982		 * We can safely assume that all pages within a section belong
 983		 * to the same zone, because dynamic memory groups only deal
 984		 * with hotplugged memory.
 985		 */
 986		pfn = ALIGN_DOWN(pfn, group->d.unit_pages);
 987		end_pfn = pfn + group->d.unit_pages;
 988		for (; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
 989			page = pfn_to_online_page(pfn);
 990			if (!page)
 991				continue;
 992			/* If anything is !MOVABLE online the rest !MOVABLE. */
 993			if (!is_zone_movable_page(page))
 994				goto kernel_zone;
 995			online_pages += PAGES_PER_SECTION;
 996		}
 997	}
 998
 999	/*
1000	 * Online MOVABLE if we could *currently* online all remaining parts
1001	 * MOVABLE. We expect to (add+) online them immediately next, so if
1002	 * nobody interferes, all will be MOVABLE if possible.
1003	 */
1004	nr_pages = max_pages - online_pages;
1005	if (!auto_movable_can_online_movable(NUMA_NO_NODE, group, nr_pages))
1006		goto kernel_zone;
1007
1008#ifdef CONFIG_NUMA
1009	if (auto_movable_numa_aware &&
1010	    !auto_movable_can_online_movable(nid, group, nr_pages))
1011		goto kernel_zone;
1012#endif /* CONFIG_NUMA */
1013
1014	return &NODE_DATA(nid)->node_zones[ZONE_MOVABLE];
1015kernel_zone:
1016	return default_kernel_zone_for_pfn(nid, pfn, nr_pages);
1017}
1018
1019static inline struct zone *default_zone_for_pfn(int nid, unsigned long start_pfn,
1020		unsigned long nr_pages)
1021{
1022	struct zone *kernel_zone = default_kernel_zone_for_pfn(nid, start_pfn,
1023			nr_pages);
1024	struct zone *movable_zone = &NODE_DATA(nid)->node_zones[ZONE_MOVABLE];
1025	bool in_kernel = zone_intersects(kernel_zone, start_pfn, nr_pages);
1026	bool in_movable = zone_intersects(movable_zone, start_pfn, nr_pages);
1027
1028	/*
1029	 * We inherit the existing zone in a simple case where zones do not
1030	 * overlap in the given range
1031	 */
1032	if (in_kernel ^ in_movable)
1033		return (in_kernel) ? kernel_zone : movable_zone;
1034
1035	/*
1036	 * If the range doesn't belong to any zone or two zones overlap in the
1037	 * given range then we use movable zone only if movable_node is
1038	 * enabled because we always online to a kernel zone by default.
1039	 */
1040	return movable_node_enabled ? movable_zone : kernel_zone;
1041}
1042
1043struct zone *zone_for_pfn_range(int online_type, int nid,
1044		struct memory_group *group, unsigned long start_pfn,
1045		unsigned long nr_pages)
1046{
1047	if (online_type == MMOP_ONLINE_KERNEL)
1048		return default_kernel_zone_for_pfn(nid, start_pfn, nr_pages);
1049
1050	if (online_type == MMOP_ONLINE_MOVABLE)
1051		return &NODE_DATA(nid)->node_zones[ZONE_MOVABLE];
1052
1053	if (online_policy == ONLINE_POLICY_AUTO_MOVABLE)
1054		return auto_movable_zone_for_pfn(nid, group, start_pfn, nr_pages);
1055
1056	return default_zone_for_pfn(nid, start_pfn, nr_pages);
1057}
1058
1059/*
1060 * This function should only be called by memory_block_{online,offline},
1061 * and {online,offline}_pages.
1062 */
1063void adjust_present_page_count(struct page *page, struct memory_group *group,
1064			       long nr_pages)
1065{
1066	struct zone *zone = page_zone(page);
1067	const bool movable = zone_idx(zone) == ZONE_MOVABLE;
1068
1069	/*
1070	 * We only support onlining/offlining/adding/removing of complete
1071	 * memory blocks; therefore, either all is either early or hotplugged.
1072	 */
1073	if (early_section(__pfn_to_section(page_to_pfn(page))))
1074		zone->present_early_pages += nr_pages;
1075	zone->present_pages += nr_pages;
1076	zone->zone_pgdat->node_present_pages += nr_pages;
1077
1078	if (group && movable)
1079		group->present_movable_pages += nr_pages;
1080	else if (group && !movable)
1081		group->present_kernel_pages += nr_pages;
1082}
1083
1084int mhp_init_memmap_on_memory(unsigned long pfn, unsigned long nr_pages,
1085			      struct zone *zone, bool mhp_off_inaccessible)
1086{
1087	unsigned long end_pfn = pfn + nr_pages;
1088	int ret, i;
1089
1090	ret = kasan_add_zero_shadow(__va(PFN_PHYS(pfn)), PFN_PHYS(nr_pages));
1091	if (ret)
1092		return ret;
1093
1094	/*
1095	 * Memory block is accessible at this stage and hence poison the struct
1096	 * pages now.  If the memory block is accessible during memory hotplug
1097	 * addition phase, then page poisining is already performed in
1098	 * sparse_add_section().
1099	 */
1100	if (mhp_off_inaccessible)
1101		page_init_poison(pfn_to_page(pfn), sizeof(struct page) * nr_pages);
1102
1103	move_pfn_range_to_zone(zone, pfn, nr_pages, NULL, MIGRATE_UNMOVABLE);
1104
1105	for (i = 0; i < nr_pages; i++) {
1106		struct page *page = pfn_to_page(pfn + i);
1107
1108		__ClearPageOffline(page);
1109		SetPageVmemmapSelfHosted(page);
1110	}
1111
1112	/*
1113	 * It might be that the vmemmap_pages fully span sections. If that is
1114	 * the case, mark those sections online here as otherwise they will be
1115	 * left offline.
1116	 */
1117	if (nr_pages >= PAGES_PER_SECTION)
1118	        online_mem_sections(pfn, ALIGN_DOWN(end_pfn, PAGES_PER_SECTION));
1119
1120	return ret;
1121}
1122
1123void mhp_deinit_memmap_on_memory(unsigned long pfn, unsigned long nr_pages)
1124{
1125	unsigned long end_pfn = pfn + nr_pages;
1126
1127	/*
1128	 * It might be that the vmemmap_pages fully span sections. If that is
1129	 * the case, mark those sections offline here as otherwise they will be
1130	 * left online.
1131	 */
1132	if (nr_pages >= PAGES_PER_SECTION)
1133		offline_mem_sections(pfn, ALIGN_DOWN(end_pfn, PAGES_PER_SECTION));
1134
1135        /*
1136	 * The pages associated with this vmemmap have been offlined, so
1137	 * we can reset its state here.
1138	 */
1139	remove_pfn_range_from_zone(page_zone(pfn_to_page(pfn)), pfn, nr_pages);
1140	kasan_remove_zero_shadow(__va(PFN_PHYS(pfn)), PFN_PHYS(nr_pages));
1141}
1142
1143/*
1144 * Must be called with mem_hotplug_lock in write mode.
1145 */
1146int online_pages(unsigned long pfn, unsigned long nr_pages,
1147		       struct zone *zone, struct memory_group *group)
1148{
1149	unsigned long flags;
 
 
1150	int need_zonelists_rebuild = 0;
1151	const int nid = zone_to_nid(zone);
1152	int ret;
1153	struct memory_notify arg;
 
 
 
1154
1155	/*
1156	 * {on,off}lining is constrained to full memory sections (or more
1157	 * precisely to memory blocks from the user space POV).
1158	 * memmap_on_memory is an exception because it reserves initial part
1159	 * of the physical memory space for vmemmaps. That space is pageblock
1160	 * aligned.
1161	 */
1162	if (WARN_ON_ONCE(!nr_pages || !pageblock_aligned(pfn) ||
1163			 !IS_ALIGNED(pfn + nr_pages, PAGES_PER_SECTION)))
1164		return -EINVAL;
1165
1166
1167	/* associate pfn range with the zone */
1168	move_pfn_range_to_zone(zone, pfn, nr_pages, NULL, MIGRATE_ISOLATE);
 
1169
1170	arg.start_pfn = pfn;
1171	arg.nr_pages = nr_pages;
1172	node_states_check_changes_online(nr_pages, zone, &arg);
1173
1174	ret = memory_notify(MEM_GOING_ONLINE, &arg);
1175	ret = notifier_to_errno(ret);
1176	if (ret)
1177		goto failed_addition;
1178
1179	/*
1180	 * Fixup the number of isolated pageblocks before marking the sections
1181	 * onlining, such that undo_isolate_page_range() works correctly.
1182	 */
1183	spin_lock_irqsave(&zone->lock, flags);
1184	zone->nr_isolate_pageblock += nr_pages / pageblock_nr_pages;
1185	spin_unlock_irqrestore(&zone->lock, flags);
1186
1187	/*
1188	 * If this zone is not populated, then it is not in zonelist.
1189	 * This means the page allocator ignores this zone.
1190	 * So, zonelist must be updated after online.
1191	 */
1192	if (!populated_zone(zone)) {
1193		need_zonelists_rebuild = 1;
1194		setup_zone_pageset(zone);
1195	}
1196
1197	online_pages_range(pfn, nr_pages);
1198	adjust_present_page_count(pfn_to_page(pfn), group, nr_pages);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1199
1200	node_states_set_node(nid, &arg);
1201	if (need_zonelists_rebuild)
1202		build_all_zonelists(NULL);
 
 
1203
1204	/* Basic onlining is complete, allow allocation of onlined pages. */
1205	undo_isolate_page_range(pfn, pfn + nr_pages, MIGRATE_MOVABLE);
1206
1207	/*
1208	 * Freshly onlined pages aren't shuffled (e.g., all pages are placed to
1209	 * the tail of the freelist when undoing isolation). Shuffle the whole
1210	 * zone to make sure the just onlined pages are properly distributed
1211	 * across the whole freelist - to create an initial shuffle.
1212	 */
1213	shuffle_zone(zone);
1214
1215	/* reinitialise watermarks and update pcp limits */
1216	init_per_zone_wmark_min();
1217
1218	kswapd_run(nid);
1219	kcompactd_run(nid);
1220
 
 
1221	writeback_set_ratelimit();
1222
1223	memory_notify(MEM_ONLINE, &arg);
 
1224	return 0;
1225
1226failed_addition:
1227	pr_debug("online_pages [mem %#010llx-%#010llx] failed\n",
1228		 (unsigned long long) pfn << PAGE_SHIFT,
1229		 (((unsigned long long) pfn + nr_pages) << PAGE_SHIFT) - 1);
1230	memory_notify(MEM_CANCEL_ONLINE, &arg);
1231	remove_pfn_range_from_zone(zone, pfn, nr_pages);
1232	return ret;
1233}
 
 
 
 
 
 
 
 
 
 
 
1234
1235/* we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG */
1236static pg_data_t *hotadd_init_pgdat(int nid)
1237{
1238	struct pglist_data *pgdat;
 
1239
1240	/*
1241	 * NODE_DATA is preallocated (free_area_init) but its internal
1242	 * state is not allocated completely. Add missing pieces.
1243	 * Completely offline nodes stay around and they just need
1244	 * reintialization.
1245	 */
1246	pgdat = NODE_DATA(nid);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1247
1248	/* init node's zones as empty zones, we don't have any present pages.*/
1249	free_area_init_core_hotplug(pgdat);
1250
1251	/*
1252	 * The node we allocated has no zone fallback lists. For avoiding
1253	 * to access not-initialized zonelist, build here.
1254	 */
1255	build_all_zonelists(pgdat);
1256
 
 
 
 
 
 
 
 
1257	return pgdat;
1258}
1259
1260/*
1261 * __try_online_node - online a node if offlined
 
 
 
 
 
 
 
 
 
 
1262 * @nid: the node ID
 
1263 * @set_node_online: Whether we want to online the node
1264 * called by cpu_up() to online a node without onlined memory.
1265 *
1266 * Returns:
1267 * 1 -> a new node has been allocated
1268 * 0 -> the node is already online
1269 * -ENOMEM -> the node could not be allocated
1270 */
1271static int __try_online_node(int nid, bool set_node_online)
1272{
1273	pg_data_t *pgdat;
1274	int ret = 1;
1275
1276	if (node_online(nid))
1277		return 0;
1278
1279	pgdat = hotadd_init_pgdat(nid);
1280	if (!pgdat) {
1281		pr_err("Cannot online node %d due to NULL pgdat\n", nid);
1282		ret = -ENOMEM;
1283		goto out;
1284	}
1285
1286	if (set_node_online) {
1287		node_set_online(nid);
1288		ret = register_one_node(nid);
1289		BUG_ON(ret);
1290	}
1291out:
1292	return ret;
1293}
1294
1295/*
1296 * Users of this function always want to online/register the node
1297 */
1298int try_online_node(int nid)
1299{
1300	int ret;
1301
1302	mem_hotplug_begin();
1303	ret =  __try_online_node(nid, true);
1304	mem_hotplug_done();
1305	return ret;
1306}
1307
1308static int check_hotplug_memory_range(u64 start, u64 size)
1309{
1310	/* memory range must be block size aligned */
1311	if (!size || !IS_ALIGNED(start, memory_block_size_bytes()) ||
1312	    !IS_ALIGNED(size, memory_block_size_bytes())) {
1313		pr_err("Block size [%#lx] unaligned hotplug range: start %#llx, size %#llx",
1314		       memory_block_size_bytes(), start, size);
1315		return -EINVAL;
1316	}
1317
1318	return 0;
1319}
1320
1321static int online_memory_block(struct memory_block *mem, void *arg)
1322{
1323	mem->online_type = mhp_default_online_type;
1324	return device_online(&mem->dev);
1325}
1326
1327#ifndef arch_supports_memmap_on_memory
1328static inline bool arch_supports_memmap_on_memory(unsigned long vmemmap_size)
1329{
1330	/*
1331	 * As default, we want the vmemmap to span a complete PMD such that we
1332	 * can map the vmemmap using a single PMD if supported by the
1333	 * architecture.
1334	 */
1335	return IS_ALIGNED(vmemmap_size, PMD_SIZE);
1336}
1337#endif
1338
1339bool mhp_supports_memmap_on_memory(void)
1340{
1341	unsigned long vmemmap_size = memory_block_memmap_size();
1342	unsigned long memmap_pages = memory_block_memmap_on_memory_pages();
1343
1344	/*
1345	 * Besides having arch support and the feature enabled at runtime, we
1346	 * need a few more assumptions to hold true:
1347	 *
1348	 * a) The vmemmap pages span complete PMDs: We don't want vmemmap code
1349	 *    to populate memory from the altmap for unrelated parts (i.e.,
1350	 *    other memory blocks)
1351	 *
1352	 * b) The vmemmap pages (and thereby the pages that will be exposed to
1353	 *    the buddy) have to cover full pageblocks: memory onlining/offlining
1354	 *    code requires applicable ranges to be page-aligned, for example, to
1355	 *    set the migratetypes properly.
1356	 *
1357	 * TODO: Although we have a check here to make sure that vmemmap pages
1358	 *       fully populate a PMD, it is not the right place to check for
1359	 *       this. A much better solution involves improving vmemmap code
1360	 *       to fallback to base pages when trying to populate vmemmap using
1361	 *       altmap as an alternative source of memory, and we do not exactly
1362	 *       populate a single PMD.
1363	 */
1364	if (!mhp_memmap_on_memory())
1365		return false;
1366
1367	/*
1368	 * Make sure the vmemmap allocation is fully contained
1369	 * so that we always allocate vmemmap memory from altmap area.
1370	 */
1371	if (!IS_ALIGNED(vmemmap_size, PAGE_SIZE))
1372		return false;
1373
1374	/*
1375	 * start pfn should be pageblock_nr_pages aligned for correctly
1376	 * setting migrate types
1377	 */
1378	if (!pageblock_aligned(memmap_pages))
1379		return false;
1380
1381	if (memmap_pages == PHYS_PFN(memory_block_size_bytes()))
1382		/* No effective hotplugged memory doesn't make sense. */
1383		return false;
1384
1385	return arch_supports_memmap_on_memory(vmemmap_size);
1386}
1387EXPORT_SYMBOL_GPL(mhp_supports_memmap_on_memory);
1388
1389static void remove_memory_blocks_and_altmaps(u64 start, u64 size)
1390{
1391	unsigned long memblock_size = memory_block_size_bytes();
1392	u64 cur_start;
1393
1394	/*
1395	 * For memmap_on_memory, the altmaps were added on a per-memblock
1396	 * basis; we have to process each individual memory block.
1397	 */
1398	for (cur_start = start; cur_start < start + size;
1399	     cur_start += memblock_size) {
1400		struct vmem_altmap *altmap = NULL;
1401		struct memory_block *mem;
1402
1403		mem = find_memory_block(pfn_to_section_nr(PFN_DOWN(cur_start)));
1404		if (WARN_ON_ONCE(!mem))
1405			continue;
1406
1407		altmap = mem->altmap;
1408		mem->altmap = NULL;
1409
1410		remove_memory_block_devices(cur_start, memblock_size);
1411
1412		arch_remove_memory(cur_start, memblock_size, altmap);
1413
1414		/* Verify that all vmemmap pages have actually been freed. */
1415		WARN(altmap->alloc, "Altmap not fully unmapped");
1416		kfree(altmap);
1417	}
1418}
1419
1420static int create_altmaps_and_memory_blocks(int nid, struct memory_group *group,
1421					    u64 start, u64 size, mhp_t mhp_flags)
1422{
1423	unsigned long memblock_size = memory_block_size_bytes();
1424	u64 cur_start;
1425	int ret;
1426
1427	for (cur_start = start; cur_start < start + size;
1428	     cur_start += memblock_size) {
1429		struct mhp_params params = { .pgprot =
1430						     pgprot_mhp(PAGE_KERNEL) };
1431		struct vmem_altmap mhp_altmap = {
1432			.base_pfn = PHYS_PFN(cur_start),
1433			.end_pfn = PHYS_PFN(cur_start + memblock_size - 1),
1434		};
1435
1436		mhp_altmap.free = memory_block_memmap_on_memory_pages();
1437		if (mhp_flags & MHP_OFFLINE_INACCESSIBLE)
1438			mhp_altmap.inaccessible = true;
1439		params.altmap = kmemdup(&mhp_altmap, sizeof(struct vmem_altmap),
1440					GFP_KERNEL);
1441		if (!params.altmap) {
1442			ret = -ENOMEM;
1443			goto out;
1444		}
1445
1446		/* call arch's memory hotadd */
1447		ret = arch_add_memory(nid, cur_start, memblock_size, &params);
1448		if (ret < 0) {
1449			kfree(params.altmap);
1450			goto out;
1451		}
1452
1453		/* create memory block devices after memory was added */
1454		ret = create_memory_block_devices(cur_start, memblock_size,
1455						  params.altmap, group);
1456		if (ret) {
1457			arch_remove_memory(cur_start, memblock_size, NULL);
1458			kfree(params.altmap);
1459			goto out;
1460		}
1461	}
1462
1463	return 0;
1464out:
1465	if (ret && cur_start != start)
1466		remove_memory_blocks_and_altmaps(start, cur_start - start);
1467	return ret;
1468}
1469
1470/*
1471 * NOTE: The caller must call lock_device_hotplug() to serialize hotplug
1472 * and online/offline operations (triggered e.g. by sysfs).
1473 *
1474 * we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG
1475 */
1476int add_memory_resource(int nid, struct resource *res, mhp_t mhp_flags)
1477{
1478	struct mhp_params params = { .pgprot = pgprot_mhp(PAGE_KERNEL) };
1479	enum memblock_flags memblock_flags = MEMBLOCK_NONE;
1480	struct memory_group *group = NULL;
1481	u64 start, size;
1482	bool new_node = false;
1483	int ret;
1484
1485	start = res->start;
1486	size = resource_size(res);
1487
1488	ret = check_hotplug_memory_range(start, size);
1489	if (ret)
1490		return ret;
1491
1492	if (mhp_flags & MHP_NID_IS_MGID) {
1493		group = memory_group_find_by_id(nid);
1494		if (!group)
1495			return -EINVAL;
1496		nid = group->nid;
1497	}
1498
1499	if (!node_possible(nid)) {
1500		WARN(1, "node %d was absent from the node_possible_map\n", nid);
1501		return -EINVAL;
1502	}
1503
1504	mem_hotplug_begin();
1505
1506	if (IS_ENABLED(CONFIG_ARCH_KEEP_MEMBLOCK)) {
1507		if (res->flags & IORESOURCE_SYSRAM_DRIVER_MANAGED)
1508			memblock_flags = MEMBLOCK_DRIVER_MANAGED;
1509		ret = memblock_add_node(start, size, nid, memblock_flags);
1510		if (ret)
1511			goto error_mem_hotplug_end;
1512	}
1513
1514	ret = __try_online_node(nid, false);
1515	if (ret < 0)
1516		goto error;
1517	new_node = ret;
1518
1519	/*
1520	 * Self hosted memmap array
1521	 */
1522	if ((mhp_flags & MHP_MEMMAP_ON_MEMORY) &&
1523	    mhp_supports_memmap_on_memory()) {
1524		ret = create_altmaps_and_memory_blocks(nid, group, start, size, mhp_flags);
1525		if (ret)
1526			goto error;
1527	} else {
1528		ret = arch_add_memory(nid, start, size, &params);
1529		if (ret < 0)
1530			goto error;
1531
1532		/* create memory block devices after memory was added */
1533		ret = create_memory_block_devices(start, size, NULL, group);
1534		if (ret) {
1535			arch_remove_memory(start, size, params.altmap);
1536			goto error;
1537		}
1538	}
1539
1540	if (new_node) {
1541		/* If sysfs file of new node can't be created, cpu on the node
1542		 * can't be hot-added. There is no rollback way now.
1543		 * So, check by BUG_ON() to catch it reluctantly..
1544		 * We online node here. We can't roll back from here.
1545		 */
1546		node_set_online(nid);
1547		ret = __register_one_node(nid);
1548		BUG_ON(ret);
1549	}
1550
1551	register_memory_blocks_under_node(nid, PFN_DOWN(start),
1552					  PFN_UP(start + size - 1),
1553					  MEMINIT_HOTPLUG);
1554
1555	/* create new memmap entry */
1556	if (!strcmp(res->name, "System RAM"))
1557		firmware_map_add_hotplug(start, start + size, "System RAM");
1558
1559	/* device_online() will take the lock when calling online_pages() */
1560	mem_hotplug_done();
1561
1562	/*
1563	 * In case we're allowed to merge the resource, flag it and trigger
1564	 * merging now that adding succeeded.
1565	 */
1566	if (mhp_flags & MHP_MERGE_RESOURCE)
1567		merge_system_ram_resource(res);
1568
1569	/* online pages if requested */
1570	if (mhp_default_online_type != MMOP_OFFLINE)
1571		walk_memory_blocks(start, size, NULL, online_memory_block);
1572
1573	return ret;
1574error:
1575	if (IS_ENABLED(CONFIG_ARCH_KEEP_MEMBLOCK))
1576		memblock_remove(start, size);
1577error_mem_hotplug_end:
 
1578	mem_hotplug_done();
1579	return ret;
1580}
1581
1582/* requires device_hotplug_lock, see add_memory_resource() */
1583int __add_memory(int nid, u64 start, u64 size, mhp_t mhp_flags)
1584{
1585	struct resource *res;
1586	int ret;
1587
1588	res = register_memory_resource(start, size, "System RAM");
1589	if (IS_ERR(res))
1590		return PTR_ERR(res);
1591
1592	ret = add_memory_resource(nid, res, mhp_flags);
1593	if (ret < 0)
1594		release_memory_resource(res);
1595	return ret;
1596}
1597
1598int add_memory(int nid, u64 start, u64 size, mhp_t mhp_flags)
1599{
1600	int rc;
1601
1602	lock_device_hotplug();
1603	rc = __add_memory(nid, start, size, mhp_flags);
1604	unlock_device_hotplug();
1605
1606	return rc;
1607}
1608EXPORT_SYMBOL_GPL(add_memory);
1609
 
1610/*
1611 * Add special, driver-managed memory to the system as system RAM. Such
1612 * memory is not exposed via the raw firmware-provided memmap as system
1613 * RAM, instead, it is detected and added by a driver - during cold boot,
1614 * after a reboot, and after kexec.
1615 *
1616 * Reasons why this memory should not be used for the initial memmap of a
1617 * kexec kernel or for placing kexec images:
1618 * - The booting kernel is in charge of determining how this memory will be
1619 *   used (e.g., use persistent memory as system RAM)
1620 * - Coordination with a hypervisor is required before this memory
1621 *   can be used (e.g., inaccessible parts).
1622 *
1623 * For this memory, no entries in /sys/firmware/memmap ("raw firmware-provided
1624 * memory map") are created. Also, the created memory resource is flagged
1625 * with IORESOURCE_SYSRAM_DRIVER_MANAGED, so in-kernel users can special-case
1626 * this memory as well (esp., not place kexec images onto it).
1627 *
1628 * The resource_name (visible via /proc/iomem) has to have the format
1629 * "System RAM ($DRIVER)".
1630 */
1631int add_memory_driver_managed(int nid, u64 start, u64 size,
1632			      const char *resource_name, mhp_t mhp_flags)
1633{
1634	struct resource *res;
1635	int rc;
1636
1637	if (!resource_name ||
1638	    strstr(resource_name, "System RAM (") != resource_name ||
1639	    resource_name[strlen(resource_name) - 1] != ')')
1640		return -EINVAL;
1641
1642	lock_device_hotplug();
 
1643
1644	res = register_memory_resource(start, size, resource_name);
1645	if (IS_ERR(res)) {
1646		rc = PTR_ERR(res);
1647		goto out_unlock;
 
 
 
1648	}
1649
1650	rc = add_memory_resource(nid, res, mhp_flags);
1651	if (rc < 0)
1652		release_memory_resource(res);
1653
1654out_unlock:
1655	unlock_device_hotplug();
1656	return rc;
1657}
1658EXPORT_SYMBOL_GPL(add_memory_driver_managed);
1659
1660/*
1661 * Platforms should define arch_get_mappable_range() that provides
1662 * maximum possible addressable physical memory range for which the
1663 * linear mapping could be created. The platform returned address
1664 * range must adhere to these following semantics.
1665 *
1666 * - range.start <= range.end
1667 * - Range includes both end points [range.start..range.end]
1668 *
1669 * There is also a fallback definition provided here, allowing the
1670 * entire possible physical address range in case any platform does
1671 * not define arch_get_mappable_range().
1672 */
1673struct range __weak arch_get_mappable_range(void)
1674{
1675	struct range mhp_range = {
1676		.start = 0UL,
1677		.end = -1ULL,
1678	};
1679	return mhp_range;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1680}
1681
1682struct range mhp_get_pluggable_range(bool need_mapping)
 
1683{
1684	const u64 max_phys = DIRECT_MAP_PHYSMEM_END;
1685	struct range mhp_range;
1686
1687	if (need_mapping) {
1688		mhp_range = arch_get_mappable_range();
1689		if (mhp_range.start > max_phys) {
1690			mhp_range.start = 0;
1691			mhp_range.end = 0;
1692		}
1693		mhp_range.end = min_t(u64, mhp_range.end, max_phys);
1694	} else {
1695		mhp_range.start = 0;
1696		mhp_range.end = max_phys;
1697	}
1698	return mhp_range;
 
 
1699}
1700EXPORT_SYMBOL_GPL(mhp_get_pluggable_range);
1701
1702bool mhp_range_allowed(u64 start, u64 size, bool need_mapping)
 
 
 
 
 
1703{
1704	struct range mhp_range = mhp_get_pluggable_range(need_mapping);
1705	u64 end = start + size;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1706
1707	if (start < end && start >= mhp_range.start && (end - 1) <= mhp_range.end)
1708		return true;
1709
1710	pr_warn("Hotplug memory [%#llx-%#llx] exceeds maximum addressable range [%#llx-%#llx]\n",
1711		start, end, mhp_range.start, mhp_range.end);
1712	return false;
 
1713}
1714
1715#ifdef CONFIG_MEMORY_HOTREMOVE
1716/*
1717 * Scan pfn range [start,end) to find movable/migratable pages (LRU pages,
1718 * non-lru movable pages and hugepages). Will skip over most unmovable
1719 * pages (esp., pages that can be skipped when offlining), but bail out on
1720 * definitely unmovable pages.
1721 *
1722 * Returns:
1723 *	0 in case a movable page is found and movable_pfn was updated.
1724 *	-ENOENT in case no movable page was found.
1725 *	-EBUSY in case a definitely unmovable page was found.
1726 */
1727static int scan_movable_pages(unsigned long start, unsigned long end,
1728			      unsigned long *movable_pfn)
1729{
1730	unsigned long pfn;
1731
1732	for (pfn = start; pfn < end; pfn++) {
1733		struct page *page;
1734		struct folio *folio;
1735
1736		if (!pfn_valid(pfn))
1737			continue;
1738		page = pfn_to_page(pfn);
1739		if (PageLRU(page))
1740			goto found;
1741		if (__PageMovable(page))
1742			goto found;
1743
1744		/*
1745		 * PageOffline() pages that are not marked __PageMovable() and
1746		 * have a reference count > 0 (after MEM_GOING_OFFLINE) are
1747		 * definitely unmovable. If their reference count would be 0,
1748		 * they could at least be skipped when offlining memory.
1749		 */
1750		if (PageOffline(page) && page_count(page))
1751			return -EBUSY;
1752
1753		if (!PageHuge(page))
1754			continue;
1755		folio = page_folio(page);
1756		/*
1757		 * This test is racy as we hold no reference or lock.  The
1758		 * hugetlb page could have been free'ed and head is no longer
1759		 * a hugetlb page before the following check.  In such unlikely
1760		 * cases false positives and negatives are possible.  Calling
1761		 * code must deal with these scenarios.
1762		 */
1763		if (folio_test_hugetlb_migratable(folio))
1764			goto found;
1765		pfn |= folio_nr_pages(folio) - 1;
1766	}
1767	return -ENOENT;
1768found:
1769	*movable_pfn = pfn;
1770	return 0;
1771}
1772
1773static void do_migrate_range(unsigned long start_pfn, unsigned long end_pfn)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1774{
1775	struct folio *folio;
1776	unsigned long pfn;
 
 
1777	LIST_HEAD(source);
1778	static DEFINE_RATELIMIT_STATE(migrate_rs, DEFAULT_RATELIMIT_INTERVAL,
1779				      DEFAULT_RATELIMIT_BURST);
1780
1781	for (pfn = start_pfn; pfn < end_pfn; pfn++) {
1782		struct page *page;
1783
1784		if (!pfn_valid(pfn))
1785			continue;
1786		page = pfn_to_page(pfn);
1787		folio = page_folio(page);
 
 
 
 
 
 
 
 
1788
1789		/*
1790		 * No reference or lock is held on the folio, so it might
1791		 * be modified concurrently (e.g. split).  As such,
1792		 * folio_nr_pages() may read garbage.  This is fine as the outer
1793		 * loop will revisit the split folio later.
 
1794		 */
1795		if (folio_test_large(folio))
1796			pfn = folio_pfn(folio) + folio_nr_pages(folio) - 1;
 
 
 
 
 
1797
1798		if (!folio_try_get(folio))
1799			continue;
 
 
 
 
 
 
 
 
 
 
 
 
 
1800
1801		if (unlikely(page_folio(page) != folio))
1802			goto put_folio;
1803
1804		if (folio_test_hwpoison(folio) ||
1805		    (folio_test_large(folio) && folio_test_has_hwpoisoned(folio))) {
1806			if (WARN_ON(folio_test_lru(folio)))
1807				folio_isolate_lru(folio);
1808			if (folio_mapped(folio)) {
1809				folio_lock(folio);
1810				unmap_poisoned_folio(folio, pfn, false);
1811				folio_unlock(folio);
1812			}
1813
1814			goto put_folio;
1815		}
1816
1817		if (!isolate_folio_to_list(folio, &source)) {
1818			if (__ratelimit(&migrate_rs)) {
1819				pr_warn("failed to isolate pfn %lx\n",
1820					page_to_pfn(page));
1821				dump_page(page, "isolation failed");
1822			}
1823		}
1824put_folio:
1825		folio_put(folio);
1826	}
1827	if (!list_empty(&source)) {
1828		nodemask_t nmask = node_states[N_MEMORY];
1829		struct migration_target_control mtc = {
1830			.nmask = &nmask,
1831			.gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
1832			.reason = MR_MEMORY_HOTPLUG,
1833		};
1834		int ret;
1835
1836		/*
1837		 * We have checked that migration range is on a single zone so
1838		 * we can use the nid of the first page to all the others.
1839		 */
1840		mtc.nid = folio_nid(list_first_entry(&source, struct folio, lru));
1841
1842		/*
1843		 * try to allocate from a different node but reuse this node
1844		 * if there are no other online nodes to be used (e.g. we are
1845		 * offlining a part of the only existing node)
1846		 */
1847		node_clear(mtc.nid, nmask);
1848		if (nodes_empty(nmask))
1849			node_set(mtc.nid, nmask);
1850		ret = migrate_pages(&source, alloc_migration_target, NULL,
1851			(unsigned long)&mtc, MIGRATE_SYNC, MR_MEMORY_HOTPLUG, NULL);
1852		if (ret) {
1853			list_for_each_entry(folio, &source, lru) {
1854				if (__ratelimit(&migrate_rs)) {
1855					pr_warn("migrating pfn %lx failed ret:%d\n",
1856						folio_pfn(folio), ret);
1857					dump_page(&folio->page,
1858						  "migration failure");
1859				}
1860			}
1861			putback_movable_pages(&source);
1862		}
1863	}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1864}
1865
1866static int __init cmdline_parse_movable_node(char *p)
1867{
 
1868	movable_node_enabled = true;
 
 
 
1869	return 0;
1870}
1871early_param("movable_node", cmdline_parse_movable_node);
1872
1873/* check which state of node_states will be changed when offline memory */
1874static void node_states_check_changes_offline(unsigned long nr_pages,
1875		struct zone *zone, struct memory_notify *arg)
1876{
1877	struct pglist_data *pgdat = zone->zone_pgdat;
1878	unsigned long present_pages = 0;
1879	enum zone_type zt;
1880
1881	arg->status_change_nid = NUMA_NO_NODE;
1882	arg->status_change_nid_normal = NUMA_NO_NODE;
 
1883
1884	/*
1885	 * Check whether node_states[N_NORMAL_MEMORY] will be changed.
1886	 * If the memory to be offline is within the range
1887	 * [0..ZONE_NORMAL], and it is the last present memory there,
1888	 * the zones in that range will become empty after the offlining,
1889	 * thus we can determine that we need to clear the node from
1890	 * node_states[N_NORMAL_MEMORY].
1891	 */
1892	for (zt = 0; zt <= ZONE_NORMAL; zt++)
1893		present_pages += pgdat->node_zones[zt].present_pages;
1894	if (zone_idx(zone) <= ZONE_NORMAL && nr_pages >= present_pages)
1895		arg->status_change_nid_normal = zone_to_nid(zone);
1896
 
1897	/*
1898	 * We have accounted the pages from [0..ZONE_NORMAL); ZONE_HIGHMEM
1899	 * does not apply as we don't support 32bit.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1900	 * Here we count the possible pages from ZONE_MOVABLE.
1901	 * If after having accounted all the pages, we see that the nr_pages
1902	 * to be offlined is over or equal to the accounted pages,
1903	 * we know that the node will become empty, and so, we can clear
1904	 * it for N_MEMORY as well.
1905	 */
1906	present_pages += pgdat->node_zones[ZONE_MOVABLE].present_pages;
1907
1908	if (nr_pages >= present_pages)
1909		arg->status_change_nid = zone_to_nid(zone);
1910}
1911
1912static void node_states_clear_node(int node, struct memory_notify *arg)
1913{
1914	if (arg->status_change_nid_normal >= 0)
1915		node_clear_state(node, N_NORMAL_MEMORY);
1916
 
 
 
1917	if (arg->status_change_nid >= 0)
1918		node_clear_state(node, N_MEMORY);
1919}
1920
1921static int count_system_ram_pages_cb(unsigned long start_pfn,
1922				     unsigned long nr_pages, void *data)
1923{
1924	unsigned long *nr_system_ram_pages = data;
1925
1926	*nr_system_ram_pages += nr_pages;
1927	return 0;
1928}
1929
1930/*
1931 * Must be called with mem_hotplug_lock in write mode.
1932 */
1933int offline_pages(unsigned long start_pfn, unsigned long nr_pages,
1934			struct zone *zone, struct memory_group *group)
1935{
1936	const unsigned long end_pfn = start_pfn + nr_pages;
1937	unsigned long pfn, managed_pages, system_ram_pages = 0;
1938	const int node = zone_to_nid(zone);
1939	unsigned long flags;
 
 
1940	struct memory_notify arg;
1941	char *reason;
1942	int ret;
1943
1944	/*
1945	 * {on,off}lining is constrained to full memory sections (or more
1946	 * precisely to memory blocks from the user space POV).
1947	 * memmap_on_memory is an exception because it reserves initial part
1948	 * of the physical memory space for vmemmaps. That space is pageblock
1949	 * aligned.
1950	 */
1951	if (WARN_ON_ONCE(!nr_pages || !pageblock_aligned(start_pfn) ||
1952			 !IS_ALIGNED(start_pfn + nr_pages, PAGES_PER_SECTION)))
1953		return -EINVAL;
1954
1955	/*
1956	 * Don't allow to offline memory blocks that contain holes.
1957	 * Consequently, memory blocks with holes can never get onlined
1958	 * via the hotplug path - online_pages() - as hotplugged memory has
1959	 * no holes. This way, we don't have to worry about memory holes,
1960	 * don't need pfn_valid() checks, and can avoid using
1961	 * walk_system_ram_range() later.
1962	 */
1963	walk_system_ram_range(start_pfn, nr_pages, &system_ram_pages,
1964			      count_system_ram_pages_cb);
1965	if (system_ram_pages != nr_pages) {
1966		ret = -EINVAL;
1967		reason = "memory holes";
1968		goto failed_removal;
1969	}
1970
1971	/*
1972	 * We only support offlining of memory blocks managed by a single zone,
1973	 * checked by calling code. This is just a sanity check that we might
1974	 * want to remove in the future.
1975	 */
1976	if (WARN_ON_ONCE(page_zone(pfn_to_page(start_pfn)) != zone ||
1977			 page_zone(pfn_to_page(end_pfn - 1)) != zone)) {
1978		ret = -EINVAL;
1979		reason = "multizone range";
1980		goto failed_removal;
1981	}
1982
1983	/*
1984	 * Disable pcplists so that page isolation cannot race with freeing
1985	 * in a way that pages from isolated pageblock are left on pcplists.
1986	 */
1987	zone_pcp_disable(zone);
1988	lru_cache_disable();
1989
1990	/* set above range as isolated */
1991	ret = start_isolate_page_range(start_pfn, end_pfn,
1992				       MIGRATE_MOVABLE,
1993				       MEMORY_OFFLINE | REPORT_FAILURE,
1994				       GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL);
1995	if (ret) {
1996		reason = "failure to isolate range";
1997		goto failed_removal_pcplists_disabled;
1998	}
 
1999
2000	arg.start_pfn = start_pfn;
2001	arg.nr_pages = nr_pages;
2002	node_states_check_changes_offline(nr_pages, zone, &arg);
2003
2004	ret = memory_notify(MEM_GOING_OFFLINE, &arg);
2005	ret = notifier_to_errno(ret);
2006	if (ret) {
2007		reason = "notifier failure";
2008		goto failed_removal_isolated;
2009	}
2010
2011	do {
2012		pfn = start_pfn;
2013		do {
2014			/*
2015			 * Historically we always checked for any signal and
2016			 * can't limit it to fatal signals without eventually
2017			 * breaking user space.
2018			 */
2019			if (signal_pending(current)) {
2020				ret = -EINTR;
2021				reason = "signal backoff";
2022				goto failed_removal_isolated;
2023			}
2024
2025			cond_resched();
 
2026
2027			ret = scan_movable_pages(pfn, end_pfn, &pfn);
2028			if (!ret) {
2029				/*
2030				 * TODO: fatal migration failures should bail
2031				 * out
2032				 */
2033				do_migrate_range(pfn, end_pfn);
2034			}
2035		} while (!ret);
2036
2037		if (ret != -ENOENT) {
2038			reason = "unmovable page";
2039			goto failed_removal_isolated;
2040		}
2041
2042		/*
2043		 * Dissolve free hugetlb folios in the memory block before doing
2044		 * offlining actually in order to make hugetlbfs's object
2045		 * counting consistent.
2046		 */
2047		ret = dissolve_free_hugetlb_folios(start_pfn, end_pfn);
2048		if (ret) {
2049			reason = "failure to dissolve huge pages";
2050			goto failed_removal_isolated;
2051		}
2052
2053		ret = test_pages_isolated(start_pfn, end_pfn, MEMORY_OFFLINE);
2054
2055	} while (ret);
2056
2057	/* Mark all sections offline and remove free pages from the buddy. */
2058	managed_pages = __offline_isolated_pages(start_pfn, end_pfn);
2059	pr_debug("Offlined Pages %ld\n", nr_pages);
2060
2061	/*
2062	 * The memory sections are marked offline, and the pageblock flags
2063	 * effectively stale; nobody should be touching them. Fixup the number
2064	 * of isolated pageblocks, memory onlining will properly revert this.
 
2065	 */
2066	spin_lock_irqsave(&zone->lock, flags);
2067	zone->nr_isolate_pageblock -= nr_pages / pageblock_nr_pages;
2068	spin_unlock_irqrestore(&zone->lock, flags);
2069
2070	lru_cache_enable();
2071	zone_pcp_enable(zone);
 
2072
2073	/* removal success */
2074	adjust_managed_page_count(pfn_to_page(start_pfn), -managed_pages);
2075	adjust_present_page_count(pfn_to_page(start_pfn), group, -nr_pages);
2076
2077	/* reinitialise watermarks and update pcp limits */
2078	init_per_zone_wmark_min();
2079
2080	/*
2081	 * Make sure to mark the node as memory-less before rebuilding the zone
2082	 * list. Otherwise this node would still appear in the fallback lists.
2083	 */
2084	node_states_clear_node(node, &arg);
2085	if (!populated_zone(zone)) {
2086		zone_pcp_reset(zone);
2087		build_all_zonelists(NULL);
2088	}
 
2089
 
2090	if (arg.status_change_nid >= 0) {
 
2091		kcompactd_stop(node);
2092		kswapd_stop(node);
2093	}
2094
 
2095	writeback_set_ratelimit();
2096
2097	memory_notify(MEM_OFFLINE, &arg);
2098	remove_pfn_range_from_zone(zone, start_pfn, nr_pages);
2099	return 0;
2100
2101failed_removal_isolated:
2102	/* pushback to free area */
2103	undo_isolate_page_range(start_pfn, end_pfn, MIGRATE_MOVABLE);
2104	memory_notify(MEM_CANCEL_OFFLINE, &arg);
2105failed_removal_pcplists_disabled:
2106	lru_cache_enable();
2107	zone_pcp_enable(zone);
2108failed_removal:
2109	pr_debug("memory offlining [mem %#010llx-%#010llx] failed due to %s\n",
2110		 (unsigned long long) start_pfn << PAGE_SHIFT,
2111		 ((unsigned long long) end_pfn << PAGE_SHIFT) - 1,
2112		 reason);
 
 
2113	return ret;
2114}
2115
 
 
 
 
 
2116static int check_memblock_offlined_cb(struct memory_block *mem, void *arg)
2117{
2118	int *nid = arg;
2119
2120	*nid = mem->nid;
2121	if (unlikely(mem->state != MEM_OFFLINE)) {
2122		phys_addr_t beginpa, endpa;
2123
2124		beginpa = PFN_PHYS(section_nr_to_pfn(mem->start_section_nr));
2125		endpa = beginpa + memory_block_size_bytes() - 1;
2126		pr_warn("removing memory fails, because memory [%pa-%pa] is onlined\n",
2127			&beginpa, &endpa);
2128
2129		return -EBUSY;
2130	}
2131	return 0;
2132}
2133
2134static int count_memory_range_altmaps_cb(struct memory_block *mem, void *arg)
2135{
2136	u64 *num_altmaps = (u64 *)arg;
2137
2138	if (mem->altmap)
2139		*num_altmaps += 1;
2140
2141	return 0;
2142}
2143
2144static int check_cpu_on_node(int nid)
2145{
2146	int cpu;
2147
2148	for_each_present_cpu(cpu) {
2149		if (cpu_to_node(cpu) == nid)
2150			/*
2151			 * the cpu on this node isn't removed, and we can't
2152			 * offline this node.
2153			 */
2154			return -EBUSY;
2155	}
2156
2157	return 0;
2158}
2159
2160static int check_no_memblock_for_node_cb(struct memory_block *mem, void *arg)
2161{
2162	int nid = *(int *)arg;
2163
2164	/*
2165	 * If a memory block belongs to multiple nodes, the stored nid is not
2166	 * reliable. However, such blocks are always online (e.g., cannot get
2167	 * offlined) and, therefore, are still spanned by the node.
2168	 */
2169	return mem->nid == nid ? -EEXIST : 0;
2170}
2171
2172/**
2173 * try_offline_node
2174 * @nid: the node ID
2175 *
2176 * Offline a node if all memory sections and cpus of the node are removed.
2177 *
2178 * NOTE: The caller must call lock_device_hotplug() to serialize hotplug
2179 * and online/offline operations before this call.
2180 */
2181void try_offline_node(int nid)
2182{
 
2183	int rc;
2184
2185	/*
2186	 * If the node still spans pages (especially ZONE_DEVICE), don't
2187	 * offline it. A node spans memory after move_pfn_range_to_zone(),
2188	 * e.g., after the memory block was onlined.
2189	 */
2190	if (node_spanned_pages(nid))
2191		return;
2192
2193	/*
2194	 * Especially offline memory blocks might not be spanned by the
2195	 * node. They will get spanned by the node once they get onlined.
2196	 * However, they link to the node in sysfs and can get onlined later.
2197	 */
2198	rc = for_each_memory_block(&nid, check_no_memblock_for_node_cb);
2199	if (rc)
2200		return;
2201
2202	if (check_cpu_on_node(nid))
2203		return;
2204
2205	/*
2206	 * all memory/cpu of this node are removed, we can offline this
2207	 * node now.
2208	 */
2209	node_set_offline(nid);
2210	unregister_one_node(nid);
2211}
2212EXPORT_SYMBOL(try_offline_node);
2213
2214static int memory_blocks_have_altmaps(u64 start, u64 size)
 
2215{
2216	u64 num_memblocks = size / memory_block_size_bytes();
2217	u64 num_altmaps = 0;
2218
2219	if (!mhp_memmap_on_memory())
2220		return 0;
 
 
 
 
 
 
 
2221
2222	walk_memory_blocks(start, size, &num_altmaps,
2223			   count_memory_range_altmaps_cb);
2224
2225	if (num_altmaps == 0)
2226		return 0;
2227
2228	if (WARN_ON_ONCE(num_memblocks != num_altmaps))
2229		return -EINVAL;
2230
2231	return 1;
2232}
2233
2234static int try_remove_memory(u64 start, u64 size)
2235{
2236	int rc, nid = NUMA_NO_NODE;
2237
2238	BUG_ON(check_hotplug_memory_range(start, size));
2239
 
 
2240	/*
2241	 * All memory blocks must be offlined before removing memory.  Check
2242	 * whether all memory blocks in question are offline and return error
2243	 * if this is not the case.
2244	 *
2245	 * While at it, determine the nid. Note that if we'd have mixed nodes,
2246	 * we'd only try to offline the last determined one -- which is good
2247	 * enough for the cases we care about.
2248	 */
2249	rc = walk_memory_blocks(start, size, &nid, check_memblock_offlined_cb);
2250	if (rc)
2251		return rc;
2252
2253	/* remove memmap entry */
2254	firmware_map_remove(start, start + size, "System RAM");
 
 
2255
2256	mem_hotplug_begin();
 
2257
2258	rc = memory_blocks_have_altmaps(start, size);
2259	if (rc < 0) {
2260		mem_hotplug_done();
2261		return rc;
2262	} else if (!rc) {
2263		/*
2264		 * Memory block device removal under the device_hotplug_lock is
2265		 * a barrier against racing online attempts.
2266		 * No altmaps present, do the removal directly
2267		 */
2268		remove_memory_block_devices(start, size);
2269		arch_remove_memory(start, size, NULL);
2270	} else {
2271		/* all memblocks in the range have altmaps */
2272		remove_memory_blocks_and_altmaps(start, size);
2273	}
2274
2275	if (IS_ENABLED(CONFIG_ARCH_KEEP_MEMBLOCK))
2276		memblock_remove(start, size);
2277
2278	release_mem_region_adjustable(start, size);
2279
2280	if (nid != NUMA_NO_NODE)
2281		try_offline_node(nid);
2282
 
2283	mem_hotplug_done();
2284	return 0;
2285}
2286
2287/**
2288 * __remove_memory - Remove memory if every memory block is offline
 
2289 * @start: physical address of the region to remove
2290 * @size: size of the region to remove
2291 *
2292 * NOTE: The caller must call lock_device_hotplug() to serialize hotplug
2293 * and online/offline operations before this call, as required by
2294 * try_offline_node().
2295 */
2296void __remove_memory(u64 start, u64 size)
2297{
2298
2299	/*
2300	 * trigger BUG() if some memory is not offlined prior to calling this
2301	 * function
2302	 */
2303	if (try_remove_memory(start, size))
2304		BUG();
2305}
2306
2307/*
2308 * Remove memory if every memory block is offline, otherwise return -EBUSY is
2309 * some memory is not offline
2310 */
2311int remove_memory(u64 start, u64 size)
2312{
2313	int rc;
2314
2315	lock_device_hotplug();
2316	rc = try_remove_memory(start, size);
2317	unlock_device_hotplug();
2318
2319	return rc;
2320}
2321EXPORT_SYMBOL_GPL(remove_memory);
2322
2323static int try_offline_memory_block(struct memory_block *mem, void *arg)
2324{
2325	uint8_t online_type = MMOP_ONLINE_KERNEL;
2326	uint8_t **online_types = arg;
2327	struct page *page;
2328	int rc;
2329
2330	/*
2331	 * Sense the online_type via the zone of the memory block. Offlining
2332	 * with multiple zones within one memory block will be rejected
2333	 * by offlining code ... so we don't care about that.
2334	 */
2335	page = pfn_to_online_page(section_nr_to_pfn(mem->start_section_nr));
2336	if (page && zone_idx(page_zone(page)) == ZONE_MOVABLE)
2337		online_type = MMOP_ONLINE_MOVABLE;
2338
2339	rc = device_offline(&mem->dev);
2340	/*
2341	 * Default is MMOP_OFFLINE - change it only if offlining succeeded,
2342	 * so try_reonline_memory_block() can do the right thing.
2343	 */
2344	if (!rc)
2345		**online_types = online_type;
2346
2347	(*online_types)++;
2348	/* Ignore if already offline. */
2349	return rc < 0 ? rc : 0;
2350}
2351
2352static int try_reonline_memory_block(struct memory_block *mem, void *arg)
2353{
2354	uint8_t **online_types = arg;
2355	int rc;
2356
2357	if (**online_types != MMOP_OFFLINE) {
2358		mem->online_type = **online_types;
2359		rc = device_online(&mem->dev);
2360		if (rc < 0)
2361			pr_warn("%s: Failed to re-online memory: %d",
2362				__func__, rc);
2363	}
2364
2365	/* Continue processing all remaining memory blocks. */
2366	(*online_types)++;
2367	return 0;
2368}
2369
2370/*
2371 * Try to offline and remove memory. Might take a long time to finish in case
2372 * memory is still in use. Primarily useful for memory devices that logically
2373 * unplugged all memory (so it's no longer in use) and want to offline + remove
2374 * that memory.
2375 */
2376int offline_and_remove_memory(u64 start, u64 size)
2377{
2378	const unsigned long mb_count = size / memory_block_size_bytes();
2379	uint8_t *online_types, *tmp;
2380	int rc;
2381
2382	if (!IS_ALIGNED(start, memory_block_size_bytes()) ||
2383	    !IS_ALIGNED(size, memory_block_size_bytes()) || !size)
2384		return -EINVAL;
2385
2386	/*
2387	 * We'll remember the old online type of each memory block, so we can
2388	 * try to revert whatever we did when offlining one memory block fails
2389	 * after offlining some others succeeded.
2390	 */
2391	online_types = kmalloc_array(mb_count, sizeof(*online_types),
2392				     GFP_KERNEL);
2393	if (!online_types)
2394		return -ENOMEM;
2395	/*
2396	 * Initialize all states to MMOP_OFFLINE, so when we abort processing in
2397	 * try_offline_memory_block(), we'll skip all unprocessed blocks in
2398	 * try_reonline_memory_block().
2399	 */
2400	memset(online_types, MMOP_OFFLINE, mb_count);
2401
2402	lock_device_hotplug();
2403
2404	tmp = online_types;
2405	rc = walk_memory_blocks(start, size, &tmp, try_offline_memory_block);
2406
2407	/*
2408	 * In case we succeeded to offline all memory, remove it.
2409	 * This cannot fail as it cannot get onlined in the meantime.
2410	 */
2411	if (!rc) {
2412		rc = try_remove_memory(start, size);
2413		if (rc)
2414			pr_err("%s: Failed to remove memory: %d", __func__, rc);
2415	}
2416
2417	/*
2418	 * Rollback what we did. While memory onlining might theoretically fail
2419	 * (nacked by a notifier), it barely ever happens.
2420	 */
2421	if (rc) {
2422		tmp = online_types;
2423		walk_memory_blocks(start, size, &tmp,
2424				   try_reonline_memory_block);
2425	}
2426	unlock_device_hotplug();
2427
2428	kfree(online_types);
2429	return rc;
2430}
2431EXPORT_SYMBOL_GPL(offline_and_remove_memory);
2432#endif /* CONFIG_MEMORY_HOTREMOVE */