1 // SPDX-License-Identifier: GPL-2.0-only
3 * linux/mm/memory_hotplug.c
8 #include <linux/stddef.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>
40 #include <asm/tlbflush.h>
47 * memory_hotplug.memmap_on_memory parameter
49 static bool memmap_on_memory __ro_after_init;
50 #ifdef CONFIG_MHP_MEMMAP_ON_MEMORY
51 module_param(memmap_on_memory, bool, 0444);
52 MODULE_PARM_DESC(memmap_on_memory, "Enable memmap on memory for memory hotplug");
56 * online_page_callback contains pointer to current page onlining function.
57 * Initially it is generic_online_page(). If it is required it could be
58 * changed by calling set_online_page_callback() for callback registration
59 * and restore_online_page_callback() for generic callback restore.
62 static online_page_callback_t online_page_callback = generic_online_page;
63 static DEFINE_MUTEX(online_page_callback_lock);
65 DEFINE_STATIC_PERCPU_RWSEM(mem_hotplug_lock);
67 void get_online_mems(void)
69 percpu_down_read(&mem_hotplug_lock);
72 void put_online_mems(void)
74 percpu_up_read(&mem_hotplug_lock);
77 bool movable_node_enabled = false;
79 #ifndef CONFIG_MEMORY_HOTPLUG_DEFAULT_ONLINE
80 int mhp_default_online_type = MMOP_OFFLINE;
82 int mhp_default_online_type = MMOP_ONLINE;
85 static int __init setup_memhp_default_state(char *str)
87 const int online_type = mhp_online_type_from_str(str);
90 mhp_default_online_type = online_type;
94 __setup("memhp_default_state=", setup_memhp_default_state);
96 void mem_hotplug_begin(void)
99 percpu_down_write(&mem_hotplug_lock);
102 void mem_hotplug_done(void)
104 percpu_up_write(&mem_hotplug_lock);
108 u64 max_mem_size = U64_MAX;
110 /* add this memory to iomem resource */
111 static struct resource *register_memory_resource(u64 start, u64 size,
112 const char *resource_name)
114 struct resource *res;
115 unsigned long flags = IORESOURCE_SYSTEM_RAM | IORESOURCE_BUSY;
117 if (strcmp(resource_name, "System RAM"))
118 flags |= IORESOURCE_SYSRAM_DRIVER_MANAGED;
120 if (!mhp_range_allowed(start, size, true))
121 return ERR_PTR(-E2BIG);
124 * Make sure value parsed from 'mem=' only restricts memory adding
125 * while booting, so that memory hotplug won't be impacted. Please
126 * refer to document of 'mem=' in kernel-parameters.txt for more
129 if (start + size > max_mem_size && system_state < SYSTEM_RUNNING)
130 return ERR_PTR(-E2BIG);
133 * Request ownership of the new memory range. This might be
134 * a child of an existing resource that was present but
135 * not marked as busy.
137 res = __request_region(&iomem_resource, start, size,
138 resource_name, flags);
141 pr_debug("Unable to reserve System RAM region: %016llx->%016llx\n",
142 start, start + size);
143 return ERR_PTR(-EEXIST);
148 static void release_memory_resource(struct resource *res)
152 release_resource(res);
156 #ifdef CONFIG_MEMORY_HOTPLUG_SPARSE
157 void get_page_bootmem(unsigned long info, struct page *page,
160 page->freelist = (void *)type;
161 SetPagePrivate(page);
162 set_page_private(page, info);
166 void put_page_bootmem(struct page *page)
170 type = (unsigned long) page->freelist;
171 BUG_ON(type < MEMORY_HOTPLUG_MIN_BOOTMEM_TYPE ||
172 type > MEMORY_HOTPLUG_MAX_BOOTMEM_TYPE);
174 if (page_ref_dec_return(page) == 1) {
175 page->freelist = NULL;
176 ClearPagePrivate(page);
177 set_page_private(page, 0);
178 INIT_LIST_HEAD(&page->lru);
179 free_reserved_page(page);
183 #ifdef CONFIG_HAVE_BOOTMEM_INFO_NODE
184 #ifndef CONFIG_SPARSEMEM_VMEMMAP
185 static void register_page_bootmem_info_section(unsigned long start_pfn)
187 unsigned long mapsize, section_nr, i;
188 struct mem_section *ms;
189 struct page *page, *memmap;
190 struct mem_section_usage *usage;
192 section_nr = pfn_to_section_nr(start_pfn);
193 ms = __nr_to_section(section_nr);
195 /* Get section's memmap address */
196 memmap = sparse_decode_mem_map(ms->section_mem_map, section_nr);
199 * Get page for the memmap's phys address
200 * XXX: need more consideration for sparse_vmemmap...
202 page = virt_to_page(memmap);
203 mapsize = sizeof(struct page) * PAGES_PER_SECTION;
204 mapsize = PAGE_ALIGN(mapsize) >> PAGE_SHIFT;
206 /* remember memmap's page */
207 for (i = 0; i < mapsize; i++, page++)
208 get_page_bootmem(section_nr, page, SECTION_INFO);
211 page = virt_to_page(usage);
213 mapsize = PAGE_ALIGN(mem_section_usage_size()) >> PAGE_SHIFT;
215 for (i = 0; i < mapsize; i++, page++)
216 get_page_bootmem(section_nr, page, MIX_SECTION_INFO);
219 #else /* CONFIG_SPARSEMEM_VMEMMAP */
220 static void register_page_bootmem_info_section(unsigned long start_pfn)
222 unsigned long mapsize, section_nr, i;
223 struct mem_section *ms;
224 struct page *page, *memmap;
225 struct mem_section_usage *usage;
227 section_nr = pfn_to_section_nr(start_pfn);
228 ms = __nr_to_section(section_nr);
230 memmap = sparse_decode_mem_map(ms->section_mem_map, section_nr);
232 register_page_bootmem_memmap(section_nr, memmap, PAGES_PER_SECTION);
235 page = virt_to_page(usage);
237 mapsize = PAGE_ALIGN(mem_section_usage_size()) >> PAGE_SHIFT;
239 for (i = 0; i < mapsize; i++, page++)
240 get_page_bootmem(section_nr, page, MIX_SECTION_INFO);
242 #endif /* !CONFIG_SPARSEMEM_VMEMMAP */
244 void __init register_page_bootmem_info_node(struct pglist_data *pgdat)
246 unsigned long i, pfn, end_pfn, nr_pages;
247 int node = pgdat->node_id;
250 nr_pages = PAGE_ALIGN(sizeof(struct pglist_data)) >> PAGE_SHIFT;
251 page = virt_to_page(pgdat);
253 for (i = 0; i < nr_pages; i++, page++)
254 get_page_bootmem(node, page, NODE_INFO);
256 pfn = pgdat->node_start_pfn;
257 end_pfn = pgdat_end_pfn(pgdat);
259 /* register section info */
260 for (; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
262 * Some platforms can assign the same pfn to multiple nodes - on
263 * node0 as well as nodeN. To avoid registering a pfn against
264 * multiple nodes we check that this pfn does not already
265 * reside in some other nodes.
267 if (pfn_valid(pfn) && (early_pfn_to_nid(pfn) == node))
268 register_page_bootmem_info_section(pfn);
271 #endif /* CONFIG_HAVE_BOOTMEM_INFO_NODE */
273 static int check_pfn_span(unsigned long pfn, unsigned long nr_pages,
277 * Disallow all operations smaller than a sub-section and only
278 * allow operations smaller than a section for
279 * SPARSEMEM_VMEMMAP. Note that check_hotplug_memory_range()
280 * enforces a larger memory_block_size_bytes() granularity for
281 * memory that will be marked online, so this check should only
282 * fire for direct arch_{add,remove}_memory() users outside of
283 * add_memory_resource().
285 unsigned long min_align;
287 if (IS_ENABLED(CONFIG_SPARSEMEM_VMEMMAP))
288 min_align = PAGES_PER_SUBSECTION;
290 min_align = PAGES_PER_SECTION;
291 if (!IS_ALIGNED(pfn, min_align)
292 || !IS_ALIGNED(nr_pages, min_align)) {
293 WARN(1, "Misaligned __%s_pages start: %#lx end: #%lx\n",
294 reason, pfn, pfn + nr_pages - 1);
301 * Return page for the valid pfn only if the page is online. All pfn
302 * walkers which rely on the fully initialized page->flags and others
303 * should use this rather than pfn_valid && pfn_to_page
305 struct page *pfn_to_online_page(unsigned long pfn)
307 unsigned long nr = pfn_to_section_nr(pfn);
308 struct dev_pagemap *pgmap;
309 struct mem_section *ms;
311 if (nr >= NR_MEM_SECTIONS)
314 ms = __nr_to_section(nr);
315 if (!online_section(ms))
319 * Save some code text when online_section() +
320 * pfn_section_valid() are sufficient.
322 if (IS_ENABLED(CONFIG_HAVE_ARCH_PFN_VALID) && !pfn_valid(pfn))
325 if (!pfn_section_valid(ms, pfn))
328 if (!online_device_section(ms))
329 return pfn_to_page(pfn);
332 * Slowpath: when ZONE_DEVICE collides with
333 * ZONE_{NORMAL,MOVABLE} within the same section some pfns in
334 * the section may be 'offline' but 'valid'. Only
335 * get_dev_pagemap() can determine sub-section online status.
337 pgmap = get_dev_pagemap(pfn, NULL);
338 put_dev_pagemap(pgmap);
340 /* The presence of a pgmap indicates ZONE_DEVICE offline pfn */
344 return pfn_to_page(pfn);
346 EXPORT_SYMBOL_GPL(pfn_to_online_page);
349 * Reasonably generic function for adding memory. It is
350 * expected that archs that support memory hotplug will
351 * call this function after deciding the zone to which to
354 int __ref __add_pages(int nid, unsigned long pfn, unsigned long nr_pages,
355 struct mhp_params *params)
357 const unsigned long end_pfn = pfn + nr_pages;
358 unsigned long cur_nr_pages;
360 struct vmem_altmap *altmap = params->altmap;
362 if (WARN_ON_ONCE(!params->pgprot.pgprot))
365 VM_BUG_ON(!mhp_range_allowed(PFN_PHYS(pfn), nr_pages * PAGE_SIZE, false));
369 * Validate altmap is within bounds of the total request
371 if (altmap->base_pfn != pfn
372 || vmem_altmap_offset(altmap) > nr_pages) {
373 pr_warn_once("memory add fail, invalid altmap\n");
379 err = check_pfn_span(pfn, nr_pages, "add");
383 for (; pfn < end_pfn; pfn += cur_nr_pages) {
384 /* Select all remaining pages up to the next section boundary */
385 cur_nr_pages = min(end_pfn - pfn,
386 SECTION_ALIGN_UP(pfn + 1) - pfn);
387 err = sparse_add_section(nid, pfn, cur_nr_pages, altmap);
392 vmemmap_populate_print_last();
396 /* find the smallest valid pfn in the range [start_pfn, end_pfn) */
397 static unsigned long find_smallest_section_pfn(int nid, struct zone *zone,
398 unsigned long start_pfn,
399 unsigned long end_pfn)
401 for (; start_pfn < end_pfn; start_pfn += PAGES_PER_SUBSECTION) {
402 if (unlikely(!pfn_to_online_page(start_pfn)))
405 if (unlikely(pfn_to_nid(start_pfn) != nid))
408 if (zone != page_zone(pfn_to_page(start_pfn)))
417 /* find the biggest valid pfn in the range [start_pfn, end_pfn). */
418 static unsigned long find_biggest_section_pfn(int nid, struct zone *zone,
419 unsigned long start_pfn,
420 unsigned long end_pfn)
424 /* pfn is the end pfn of a memory section. */
426 for (; pfn >= start_pfn; pfn -= PAGES_PER_SUBSECTION) {
427 if (unlikely(!pfn_to_online_page(pfn)))
430 if (unlikely(pfn_to_nid(pfn) != nid))
433 if (zone != page_zone(pfn_to_page(pfn)))
442 static void shrink_zone_span(struct zone *zone, unsigned long start_pfn,
443 unsigned long end_pfn)
446 int nid = zone_to_nid(zone);
448 zone_span_writelock(zone);
449 if (zone->zone_start_pfn == start_pfn) {
451 * If the section is smallest section in the zone, it need
452 * shrink zone->zone_start_pfn and zone->zone_spanned_pages.
453 * In this case, we find second smallest valid mem_section
454 * for shrinking zone.
456 pfn = find_smallest_section_pfn(nid, zone, end_pfn,
459 zone->spanned_pages = zone_end_pfn(zone) - pfn;
460 zone->zone_start_pfn = pfn;
462 zone->zone_start_pfn = 0;
463 zone->spanned_pages = 0;
465 } else if (zone_end_pfn(zone) == end_pfn) {
467 * If the section is biggest section in the zone, it need
468 * shrink zone->spanned_pages.
469 * In this case, we find second biggest valid mem_section for
472 pfn = find_biggest_section_pfn(nid, zone, zone->zone_start_pfn,
475 zone->spanned_pages = pfn - zone->zone_start_pfn + 1;
477 zone->zone_start_pfn = 0;
478 zone->spanned_pages = 0;
481 zone_span_writeunlock(zone);
484 static void update_pgdat_span(struct pglist_data *pgdat)
486 unsigned long node_start_pfn = 0, node_end_pfn = 0;
489 for (zone = pgdat->node_zones;
490 zone < pgdat->node_zones + MAX_NR_ZONES; zone++) {
491 unsigned long end_pfn = zone_end_pfn(zone);
493 /* No need to lock the zones, they can't change. */
494 if (!zone->spanned_pages)
497 node_start_pfn = zone->zone_start_pfn;
498 node_end_pfn = end_pfn;
502 if (end_pfn > node_end_pfn)
503 node_end_pfn = end_pfn;
504 if (zone->zone_start_pfn < node_start_pfn)
505 node_start_pfn = zone->zone_start_pfn;
508 pgdat->node_start_pfn = node_start_pfn;
509 pgdat->node_spanned_pages = node_end_pfn - node_start_pfn;
512 void __ref remove_pfn_range_from_zone(struct zone *zone,
513 unsigned long start_pfn,
514 unsigned long nr_pages)
516 const unsigned long end_pfn = start_pfn + nr_pages;
517 struct pglist_data *pgdat = zone->zone_pgdat;
518 unsigned long pfn, cur_nr_pages, flags;
520 /* Poison struct pages because they are now uninitialized again. */
521 for (pfn = start_pfn; pfn < end_pfn; pfn += cur_nr_pages) {
524 /* Select all remaining pages up to the next section boundary */
526 min(end_pfn - pfn, SECTION_ALIGN_UP(pfn + 1) - pfn);
527 page_init_poison(pfn_to_page(pfn),
528 sizeof(struct page) * cur_nr_pages);
531 #ifdef CONFIG_ZONE_DEVICE
533 * Zone shrinking code cannot properly deal with ZONE_DEVICE. So
534 * we will not try to shrink the zones - which is okay as
535 * set_zone_contiguous() cannot deal with ZONE_DEVICE either way.
537 if (zone_idx(zone) == ZONE_DEVICE)
541 clear_zone_contiguous(zone);
543 pgdat_resize_lock(zone->zone_pgdat, &flags);
544 shrink_zone_span(zone, start_pfn, start_pfn + nr_pages);
545 update_pgdat_span(pgdat);
546 pgdat_resize_unlock(zone->zone_pgdat, &flags);
548 set_zone_contiguous(zone);
551 static void __remove_section(unsigned long pfn, unsigned long nr_pages,
552 unsigned long map_offset,
553 struct vmem_altmap *altmap)
555 struct mem_section *ms = __pfn_to_section(pfn);
557 if (WARN_ON_ONCE(!valid_section(ms)))
560 sparse_remove_section(ms, pfn, nr_pages, map_offset, altmap);
564 * __remove_pages() - remove sections of pages
565 * @pfn: starting pageframe (must be aligned to start of a section)
566 * @nr_pages: number of pages to remove (must be multiple of section size)
567 * @altmap: alternative device page map or %NULL if default memmap is used
569 * Generic helper function to remove section mappings and sysfs entries
570 * for the section of the memory we are removing. Caller needs to make
571 * sure that pages are marked reserved and zones are adjust properly by
572 * calling offline_pages().
574 void __remove_pages(unsigned long pfn, unsigned long nr_pages,
575 struct vmem_altmap *altmap)
577 const unsigned long end_pfn = pfn + nr_pages;
578 unsigned long cur_nr_pages;
579 unsigned long map_offset = 0;
581 map_offset = vmem_altmap_offset(altmap);
583 if (check_pfn_span(pfn, nr_pages, "remove"))
586 for (; pfn < end_pfn; pfn += cur_nr_pages) {
588 /* Select all remaining pages up to the next section boundary */
589 cur_nr_pages = min(end_pfn - pfn,
590 SECTION_ALIGN_UP(pfn + 1) - pfn);
591 __remove_section(pfn, cur_nr_pages, map_offset, altmap);
596 int set_online_page_callback(online_page_callback_t callback)
601 mutex_lock(&online_page_callback_lock);
603 if (online_page_callback == generic_online_page) {
604 online_page_callback = callback;
608 mutex_unlock(&online_page_callback_lock);
613 EXPORT_SYMBOL_GPL(set_online_page_callback);
615 int restore_online_page_callback(online_page_callback_t callback)
620 mutex_lock(&online_page_callback_lock);
622 if (online_page_callback == callback) {
623 online_page_callback = generic_online_page;
627 mutex_unlock(&online_page_callback_lock);
632 EXPORT_SYMBOL_GPL(restore_online_page_callback);
634 void generic_online_page(struct page *page, unsigned int order)
637 * Freeing the page with debug_pagealloc enabled will try to unmap it,
638 * so we should map it first. This is better than introducing a special
639 * case in page freeing fast path.
641 debug_pagealloc_map_pages(page, 1 << order);
642 __free_pages_core(page, order);
643 totalram_pages_add(1UL << order);
644 #ifdef CONFIG_HIGHMEM
645 if (PageHighMem(page))
646 totalhigh_pages_add(1UL << order);
649 EXPORT_SYMBOL_GPL(generic_online_page);
651 static void online_pages_range(unsigned long start_pfn, unsigned long nr_pages)
653 const unsigned long end_pfn = start_pfn + nr_pages;
657 * Online the pages in MAX_ORDER - 1 aligned chunks. The callback might
658 * decide to not expose all pages to the buddy (e.g., expose them
659 * later). We account all pages as being online and belonging to this
661 * When using memmap_on_memory, the range might not be aligned to
662 * MAX_ORDER_NR_PAGES - 1, but pageblock aligned. __ffs() will detect
663 * this and the first chunk to online will be pageblock_nr_pages.
665 for (pfn = start_pfn; pfn < end_pfn;) {
666 int order = min(MAX_ORDER - 1UL, __ffs(pfn));
668 (*online_page_callback)(pfn_to_page(pfn), order);
669 pfn += (1UL << order);
672 /* mark all involved sections as online */
673 online_mem_sections(start_pfn, end_pfn);
676 /* check which state of node_states will be changed when online memory */
677 static void node_states_check_changes_online(unsigned long nr_pages,
678 struct zone *zone, struct memory_notify *arg)
680 int nid = zone_to_nid(zone);
682 arg->status_change_nid = NUMA_NO_NODE;
683 arg->status_change_nid_normal = NUMA_NO_NODE;
684 arg->status_change_nid_high = NUMA_NO_NODE;
686 if (!node_state(nid, N_MEMORY))
687 arg->status_change_nid = nid;
688 if (zone_idx(zone) <= ZONE_NORMAL && !node_state(nid, N_NORMAL_MEMORY))
689 arg->status_change_nid_normal = nid;
690 #ifdef CONFIG_HIGHMEM
691 if (zone_idx(zone) <= ZONE_HIGHMEM && !node_state(nid, N_HIGH_MEMORY))
692 arg->status_change_nid_high = nid;
696 static void node_states_set_node(int node, struct memory_notify *arg)
698 if (arg->status_change_nid_normal >= 0)
699 node_set_state(node, N_NORMAL_MEMORY);
701 if (arg->status_change_nid_high >= 0)
702 node_set_state(node, N_HIGH_MEMORY);
704 if (arg->status_change_nid >= 0)
705 node_set_state(node, N_MEMORY);
708 static void __meminit resize_zone_range(struct zone *zone, unsigned long start_pfn,
709 unsigned long nr_pages)
711 unsigned long old_end_pfn = zone_end_pfn(zone);
713 if (zone_is_empty(zone) || start_pfn < zone->zone_start_pfn)
714 zone->zone_start_pfn = start_pfn;
716 zone->spanned_pages = max(start_pfn + nr_pages, old_end_pfn) - zone->zone_start_pfn;
719 static void __meminit resize_pgdat_range(struct pglist_data *pgdat, unsigned long start_pfn,
720 unsigned long nr_pages)
722 unsigned long old_end_pfn = pgdat_end_pfn(pgdat);
724 if (!pgdat->node_spanned_pages || start_pfn < pgdat->node_start_pfn)
725 pgdat->node_start_pfn = start_pfn;
727 pgdat->node_spanned_pages = max(start_pfn + nr_pages, old_end_pfn) - pgdat->node_start_pfn;
731 static void section_taint_zone_device(unsigned long pfn)
733 struct mem_section *ms = __pfn_to_section(pfn);
735 ms->section_mem_map |= SECTION_TAINT_ZONE_DEVICE;
739 * Associate the pfn range with the given zone, initializing the memmaps
740 * and resizing the pgdat/zone data to span the added pages. After this
741 * call, all affected pages are PG_reserved.
743 * All aligned pageblocks are initialized to the specified migratetype
744 * (usually MIGRATE_MOVABLE). Besides setting the migratetype, no related
745 * zone stats (e.g., nr_isolate_pageblock) are touched.
747 void __ref move_pfn_range_to_zone(struct zone *zone, unsigned long start_pfn,
748 unsigned long nr_pages,
749 struct vmem_altmap *altmap, int migratetype)
751 struct pglist_data *pgdat = zone->zone_pgdat;
752 int nid = pgdat->node_id;
755 clear_zone_contiguous(zone);
757 /* TODO Huh pgdat is irqsave while zone is not. It used to be like that before */
758 pgdat_resize_lock(pgdat, &flags);
759 zone_span_writelock(zone);
760 if (zone_is_empty(zone))
761 init_currently_empty_zone(zone, start_pfn, nr_pages);
762 resize_zone_range(zone, start_pfn, nr_pages);
763 zone_span_writeunlock(zone);
764 resize_pgdat_range(pgdat, start_pfn, nr_pages);
765 pgdat_resize_unlock(pgdat, &flags);
768 * Subsection population requires care in pfn_to_online_page().
769 * Set the taint to enable the slow path detection of
770 * ZONE_DEVICE pages in an otherwise ZONE_{NORMAL,MOVABLE}
773 if (zone_is_zone_device(zone)) {
774 if (!IS_ALIGNED(start_pfn, PAGES_PER_SECTION))
775 section_taint_zone_device(start_pfn);
776 if (!IS_ALIGNED(start_pfn + nr_pages, PAGES_PER_SECTION))
777 section_taint_zone_device(start_pfn + nr_pages);
781 * TODO now we have a visible range of pages which are not associated
782 * with their zone properly. Not nice but set_pfnblock_flags_mask
783 * expects the zone spans the pfn range. All the pages in the range
784 * are reserved so nobody should be touching them so we should be safe
786 memmap_init_range(nr_pages, nid, zone_idx(zone), start_pfn, 0,
787 MEMINIT_HOTPLUG, altmap, migratetype);
789 set_zone_contiguous(zone);
793 * Returns a default kernel memory zone for the given pfn range.
794 * If no kernel zone covers this pfn range it will automatically go
795 * to the ZONE_NORMAL.
797 static struct zone *default_kernel_zone_for_pfn(int nid, unsigned long start_pfn,
798 unsigned long nr_pages)
800 struct pglist_data *pgdat = NODE_DATA(nid);
803 for (zid = 0; zid <= ZONE_NORMAL; zid++) {
804 struct zone *zone = &pgdat->node_zones[zid];
806 if (zone_intersects(zone, start_pfn, nr_pages))
810 return &pgdat->node_zones[ZONE_NORMAL];
813 static inline struct zone *default_zone_for_pfn(int nid, unsigned long start_pfn,
814 unsigned long nr_pages)
816 struct zone *kernel_zone = default_kernel_zone_for_pfn(nid, start_pfn,
818 struct zone *movable_zone = &NODE_DATA(nid)->node_zones[ZONE_MOVABLE];
819 bool in_kernel = zone_intersects(kernel_zone, start_pfn, nr_pages);
820 bool in_movable = zone_intersects(movable_zone, start_pfn, nr_pages);
823 * We inherit the existing zone in a simple case where zones do not
824 * overlap in the given range
826 if (in_kernel ^ in_movable)
827 return (in_kernel) ? kernel_zone : movable_zone;
830 * If the range doesn't belong to any zone or two zones overlap in the
831 * given range then we use movable zone only if movable_node is
832 * enabled because we always online to a kernel zone by default.
834 return movable_node_enabled ? movable_zone : kernel_zone;
837 struct zone *zone_for_pfn_range(int online_type, int nid, unsigned start_pfn,
838 unsigned long nr_pages)
840 if (online_type == MMOP_ONLINE_KERNEL)
841 return default_kernel_zone_for_pfn(nid, start_pfn, nr_pages);
843 if (online_type == MMOP_ONLINE_MOVABLE)
844 return &NODE_DATA(nid)->node_zones[ZONE_MOVABLE];
846 return default_zone_for_pfn(nid, start_pfn, nr_pages);
850 * This function should only be called by memory_block_{online,offline},
851 * and {online,offline}_pages.
853 void adjust_present_page_count(struct zone *zone, long nr_pages)
857 zone->present_pages += nr_pages;
858 pgdat_resize_lock(zone->zone_pgdat, &flags);
859 zone->zone_pgdat->node_present_pages += nr_pages;
860 pgdat_resize_unlock(zone->zone_pgdat, &flags);
863 int mhp_init_memmap_on_memory(unsigned long pfn, unsigned long nr_pages,
866 unsigned long end_pfn = pfn + nr_pages;
869 ret = kasan_add_zero_shadow(__va(PFN_PHYS(pfn)), PFN_PHYS(nr_pages));
873 move_pfn_range_to_zone(zone, pfn, nr_pages, NULL, MIGRATE_UNMOVABLE);
876 * It might be that the vmemmap_pages fully span sections. If that is
877 * the case, mark those sections online here as otherwise they will be
880 if (nr_pages >= PAGES_PER_SECTION)
881 online_mem_sections(pfn, ALIGN_DOWN(end_pfn, PAGES_PER_SECTION));
886 void mhp_deinit_memmap_on_memory(unsigned long pfn, unsigned long nr_pages)
888 unsigned long end_pfn = pfn + nr_pages;
891 * It might be that the vmemmap_pages fully span sections. If that is
892 * the case, mark those sections offline here as otherwise they will be
895 if (nr_pages >= PAGES_PER_SECTION)
896 offline_mem_sections(pfn, ALIGN_DOWN(end_pfn, PAGES_PER_SECTION));
899 * The pages associated with this vmemmap have been offlined, so
900 * we can reset its state here.
902 remove_pfn_range_from_zone(page_zone(pfn_to_page(pfn)), pfn, nr_pages);
903 kasan_remove_zero_shadow(__va(PFN_PHYS(pfn)), PFN_PHYS(nr_pages));
906 int __ref online_pages(unsigned long pfn, unsigned long nr_pages, struct zone *zone)
909 int need_zonelists_rebuild = 0;
910 const int nid = zone_to_nid(zone);
912 struct memory_notify arg;
915 * {on,off}lining is constrained to full memory sections (or more
916 * precisly to memory blocks from the user space POV).
917 * memmap_on_memory is an exception because it reserves initial part
918 * of the physical memory space for vmemmaps. That space is pageblock
921 if (WARN_ON_ONCE(!nr_pages ||
922 !IS_ALIGNED(pfn, pageblock_nr_pages) ||
923 !IS_ALIGNED(pfn + nr_pages, PAGES_PER_SECTION)))
928 /* associate pfn range with the zone */
929 move_pfn_range_to_zone(zone, pfn, nr_pages, NULL, MIGRATE_ISOLATE);
932 arg.nr_pages = nr_pages;
933 node_states_check_changes_online(nr_pages, zone, &arg);
935 ret = memory_notify(MEM_GOING_ONLINE, &arg);
936 ret = notifier_to_errno(ret);
938 goto failed_addition;
941 * Fixup the number of isolated pageblocks before marking the sections
942 * onlining, such that undo_isolate_page_range() works correctly.
944 spin_lock_irqsave(&zone->lock, flags);
945 zone->nr_isolate_pageblock += nr_pages / pageblock_nr_pages;
946 spin_unlock_irqrestore(&zone->lock, flags);
949 * If this zone is not populated, then it is not in zonelist.
950 * This means the page allocator ignores this zone.
951 * So, zonelist must be updated after online.
953 if (!populated_zone(zone)) {
954 need_zonelists_rebuild = 1;
955 setup_zone_pageset(zone);
958 online_pages_range(pfn, nr_pages);
959 adjust_present_page_count(zone, nr_pages);
961 node_states_set_node(nid, &arg);
962 if (need_zonelists_rebuild)
963 build_all_zonelists(NULL);
964 zone_pcp_update(zone);
966 /* Basic onlining is complete, allow allocation of onlined pages. */
967 undo_isolate_page_range(pfn, pfn + nr_pages, MIGRATE_MOVABLE);
970 * Freshly onlined pages aren't shuffled (e.g., all pages are placed to
971 * the tail of the freelist when undoing isolation). Shuffle the whole
972 * zone to make sure the just onlined pages are properly distributed
973 * across the whole freelist - to create an initial shuffle.
977 init_per_zone_wmark_min();
982 writeback_set_ratelimit();
984 memory_notify(MEM_ONLINE, &arg);
989 pr_debug("online_pages [mem %#010llx-%#010llx] failed\n",
990 (unsigned long long) pfn << PAGE_SHIFT,
991 (((unsigned long long) pfn + nr_pages) << PAGE_SHIFT) - 1);
992 memory_notify(MEM_CANCEL_ONLINE, &arg);
993 remove_pfn_range_from_zone(zone, pfn, nr_pages);
997 #endif /* CONFIG_MEMORY_HOTPLUG_SPARSE */
999 static void reset_node_present_pages(pg_data_t *pgdat)
1003 for (z = pgdat->node_zones; z < pgdat->node_zones + MAX_NR_ZONES; z++)
1004 z->present_pages = 0;
1006 pgdat->node_present_pages = 0;
1009 /* we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG */
1010 static pg_data_t __ref *hotadd_new_pgdat(int nid)
1012 struct pglist_data *pgdat;
1014 pgdat = NODE_DATA(nid);
1016 pgdat = arch_alloc_nodedata(nid);
1020 pgdat->per_cpu_nodestats =
1021 alloc_percpu(struct per_cpu_nodestat);
1022 arch_refresh_nodedata(nid, pgdat);
1026 * Reset the nr_zones, order and highest_zoneidx before reuse.
1027 * Note that kswapd will init kswapd_highest_zoneidx properly
1028 * when it starts in the near future.
1030 pgdat->nr_zones = 0;
1031 pgdat->kswapd_order = 0;
1032 pgdat->kswapd_highest_zoneidx = 0;
1033 for_each_online_cpu(cpu) {
1034 struct per_cpu_nodestat *p;
1036 p = per_cpu_ptr(pgdat->per_cpu_nodestats, cpu);
1037 memset(p, 0, sizeof(*p));
1041 /* we can use NODE_DATA(nid) from here */
1042 pgdat->node_id = nid;
1043 pgdat->node_start_pfn = 0;
1045 /* init node's zones as empty zones, we don't have any present pages.*/
1046 free_area_init_core_hotplug(nid);
1049 * The node we allocated has no zone fallback lists. For avoiding
1050 * to access not-initialized zonelist, build here.
1052 build_all_zonelists(pgdat);
1055 * When memory is hot-added, all the memory is in offline state. So
1056 * clear all zones' present_pages because they will be updated in
1057 * online_pages() and offline_pages().
1059 reset_node_managed_pages(pgdat);
1060 reset_node_present_pages(pgdat);
1065 static void rollback_node_hotadd(int nid)
1067 pg_data_t *pgdat = NODE_DATA(nid);
1069 arch_refresh_nodedata(nid, NULL);
1070 free_percpu(pgdat->per_cpu_nodestats);
1071 arch_free_nodedata(pgdat);
1076 * try_online_node - online a node if offlined
1078 * @set_node_online: Whether we want to online the node
1079 * called by cpu_up() to online a node without onlined memory.
1082 * 1 -> a new node has been allocated
1083 * 0 -> the node is already online
1084 * -ENOMEM -> the node could not be allocated
1086 static int __try_online_node(int nid, bool set_node_online)
1091 if (node_online(nid))
1094 pgdat = hotadd_new_pgdat(nid);
1096 pr_err("Cannot online node %d due to NULL pgdat\n", nid);
1101 if (set_node_online) {
1102 node_set_online(nid);
1103 ret = register_one_node(nid);
1111 * Users of this function always want to online/register the node
1113 int try_online_node(int nid)
1117 mem_hotplug_begin();
1118 ret = __try_online_node(nid, true);
1123 static int check_hotplug_memory_range(u64 start, u64 size)
1125 /* memory range must be block size aligned */
1126 if (!size || !IS_ALIGNED(start, memory_block_size_bytes()) ||
1127 !IS_ALIGNED(size, memory_block_size_bytes())) {
1128 pr_err("Block size [%#lx] unaligned hotplug range: start %#llx, size %#llx",
1129 memory_block_size_bytes(), start, size);
1136 static int online_memory_block(struct memory_block *mem, void *arg)
1138 mem->online_type = mhp_default_online_type;
1139 return device_online(&mem->dev);
1142 bool mhp_supports_memmap_on_memory(unsigned long size)
1144 unsigned long nr_vmemmap_pages = size / PAGE_SIZE;
1145 unsigned long vmemmap_size = nr_vmemmap_pages * sizeof(struct page);
1146 unsigned long remaining_size = size - vmemmap_size;
1149 * Besides having arch support and the feature enabled at runtime, we
1150 * need a few more assumptions to hold true:
1152 * a) We span a single memory block: memory onlining/offlinin;g happens
1153 * in memory block granularity. We don't want the vmemmap of online
1154 * memory blocks to reside on offline memory blocks. In the future,
1155 * we might want to support variable-sized memory blocks to make the
1156 * feature more versatile.
1158 * b) The vmemmap pages span complete PMDs: We don't want vmemmap code
1159 * to populate memory from the altmap for unrelated parts (i.e.,
1160 * other memory blocks)
1162 * c) The vmemmap pages (and thereby the pages that will be exposed to
1163 * the buddy) have to cover full pageblocks: memory onlining/offlining
1164 * code requires applicable ranges to be page-aligned, for example, to
1165 * set the migratetypes properly.
1167 * TODO: Although we have a check here to make sure that vmemmap pages
1168 * fully populate a PMD, it is not the right place to check for
1169 * this. A much better solution involves improving vmemmap code
1170 * to fallback to base pages when trying to populate vmemmap using
1171 * altmap as an alternative source of memory, and we do not exactly
1172 * populate a single PMD.
1174 return memmap_on_memory &&
1175 IS_ENABLED(CONFIG_MHP_MEMMAP_ON_MEMORY) &&
1176 size == memory_block_size_bytes() &&
1177 IS_ALIGNED(vmemmap_size, PMD_SIZE) &&
1178 IS_ALIGNED(remaining_size, (pageblock_nr_pages << PAGE_SHIFT));
1182 * NOTE: The caller must call lock_device_hotplug() to serialize hotplug
1183 * and online/offline operations (triggered e.g. by sysfs).
1185 * we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG
1187 int __ref add_memory_resource(int nid, struct resource *res, mhp_t mhp_flags)
1189 struct mhp_params params = { .pgprot = pgprot_mhp(PAGE_KERNEL) };
1190 struct vmem_altmap mhp_altmap = {};
1192 bool new_node = false;
1196 size = resource_size(res);
1198 ret = check_hotplug_memory_range(start, size);
1202 if (!node_possible(nid)) {
1203 WARN(1, "node %d was absent from the node_possible_map\n", nid);
1207 mem_hotplug_begin();
1209 if (IS_ENABLED(CONFIG_ARCH_KEEP_MEMBLOCK))
1210 memblock_add_node(start, size, nid);
1212 ret = __try_online_node(nid, false);
1218 * Self hosted memmap array
1220 if (mhp_flags & MHP_MEMMAP_ON_MEMORY) {
1221 if (!mhp_supports_memmap_on_memory(size)) {
1225 mhp_altmap.free = PHYS_PFN(size);
1226 mhp_altmap.base_pfn = PHYS_PFN(start);
1227 params.altmap = &mhp_altmap;
1230 /* call arch's memory hotadd */
1231 ret = arch_add_memory(nid, start, size, ¶ms);
1235 /* create memory block devices after memory was added */
1236 ret = create_memory_block_devices(start, size, mhp_altmap.alloc);
1238 arch_remove_memory(nid, start, size, NULL);
1243 /* If sysfs file of new node can't be created, cpu on the node
1244 * can't be hot-added. There is no rollback way now.
1245 * So, check by BUG_ON() to catch it reluctantly..
1246 * We online node here. We can't roll back from here.
1248 node_set_online(nid);
1249 ret = __register_one_node(nid);
1253 /* link memory sections under this node.*/
1254 link_mem_sections(nid, PFN_DOWN(start), PFN_UP(start + size - 1),
1257 /* create new memmap entry */
1258 if (!strcmp(res->name, "System RAM"))
1259 firmware_map_add_hotplug(start, start + size, "System RAM");
1261 /* device_online() will take the lock when calling online_pages() */
1265 * In case we're allowed to merge the resource, flag it and trigger
1266 * merging now that adding succeeded.
1268 if (mhp_flags & MHP_MERGE_RESOURCE)
1269 merge_system_ram_resource(res);
1271 /* online pages if requested */
1272 if (mhp_default_online_type != MMOP_OFFLINE)
1273 walk_memory_blocks(start, size, NULL, online_memory_block);
1277 /* rollback pgdat allocation and others */
1279 rollback_node_hotadd(nid);
1280 if (IS_ENABLED(CONFIG_ARCH_KEEP_MEMBLOCK))
1281 memblock_remove(start, size);
1286 /* requires device_hotplug_lock, see add_memory_resource() */
1287 int __ref __add_memory(int nid, u64 start, u64 size, mhp_t mhp_flags)
1289 struct resource *res;
1292 res = register_memory_resource(start, size, "System RAM");
1294 return PTR_ERR(res);
1296 ret = add_memory_resource(nid, res, mhp_flags);
1298 release_memory_resource(res);
1302 int add_memory(int nid, u64 start, u64 size, mhp_t mhp_flags)
1306 lock_device_hotplug();
1307 rc = __add_memory(nid, start, size, mhp_flags);
1308 unlock_device_hotplug();
1312 EXPORT_SYMBOL_GPL(add_memory);
1315 * Add special, driver-managed memory to the system as system RAM. Such
1316 * memory is not exposed via the raw firmware-provided memmap as system
1317 * RAM, instead, it is detected and added by a driver - during cold boot,
1318 * after a reboot, and after kexec.
1320 * Reasons why this memory should not be used for the initial memmap of a
1321 * kexec kernel or for placing kexec images:
1322 * - The booting kernel is in charge of determining how this memory will be
1323 * used (e.g., use persistent memory as system RAM)
1324 * - Coordination with a hypervisor is required before this memory
1325 * can be used (e.g., inaccessible parts).
1327 * For this memory, no entries in /sys/firmware/memmap ("raw firmware-provided
1328 * memory map") are created. Also, the created memory resource is flagged
1329 * with IORESOURCE_SYSRAM_DRIVER_MANAGED, so in-kernel users can special-case
1330 * this memory as well (esp., not place kexec images onto it).
1332 * The resource_name (visible via /proc/iomem) has to have the format
1333 * "System RAM ($DRIVER)".
1335 int add_memory_driver_managed(int nid, u64 start, u64 size,
1336 const char *resource_name, mhp_t mhp_flags)
1338 struct resource *res;
1341 if (!resource_name ||
1342 strstr(resource_name, "System RAM (") != resource_name ||
1343 resource_name[strlen(resource_name) - 1] != ')')
1346 lock_device_hotplug();
1348 res = register_memory_resource(start, size, resource_name);
1354 rc = add_memory_resource(nid, res, mhp_flags);
1356 release_memory_resource(res);
1359 unlock_device_hotplug();
1362 EXPORT_SYMBOL_GPL(add_memory_driver_managed);
1365 * Platforms should define arch_get_mappable_range() that provides
1366 * maximum possible addressable physical memory range for which the
1367 * linear mapping could be created. The platform returned address
1368 * range must adhere to these following semantics.
1370 * - range.start <= range.end
1371 * - Range includes both end points [range.start..range.end]
1373 * There is also a fallback definition provided here, allowing the
1374 * entire possible physical address range in case any platform does
1375 * not define arch_get_mappable_range().
1377 struct range __weak arch_get_mappable_range(void)
1379 struct range mhp_range = {
1386 struct range mhp_get_pluggable_range(bool need_mapping)
1388 const u64 max_phys = (1ULL << MAX_PHYSMEM_BITS) - 1;
1389 struct range mhp_range;
1392 mhp_range = arch_get_mappable_range();
1393 if (mhp_range.start > max_phys) {
1394 mhp_range.start = 0;
1397 mhp_range.end = min_t(u64, mhp_range.end, max_phys);
1399 mhp_range.start = 0;
1400 mhp_range.end = max_phys;
1404 EXPORT_SYMBOL_GPL(mhp_get_pluggable_range);
1406 bool mhp_range_allowed(u64 start, u64 size, bool need_mapping)
1408 struct range mhp_range = mhp_get_pluggable_range(need_mapping);
1409 u64 end = start + size;
1411 if (start < end && start >= mhp_range.start && (end - 1) <= mhp_range.end)
1414 pr_warn("Hotplug memory [%#llx-%#llx] exceeds maximum addressable range [%#llx-%#llx]\n",
1415 start, end, mhp_range.start, mhp_range.end);
1419 #ifdef CONFIG_MEMORY_HOTREMOVE
1421 * Confirm all pages in a range [start, end) belong to the same zone (skipping
1422 * memory holes). When true, return the zone.
1424 struct zone *test_pages_in_a_zone(unsigned long start_pfn,
1425 unsigned long end_pfn)
1427 unsigned long pfn, sec_end_pfn;
1428 struct zone *zone = NULL;
1431 for (pfn = start_pfn, sec_end_pfn = SECTION_ALIGN_UP(start_pfn + 1);
1433 pfn = sec_end_pfn, sec_end_pfn += PAGES_PER_SECTION) {
1434 /* Make sure the memory section is present first */
1435 if (!present_section_nr(pfn_to_section_nr(pfn)))
1437 for (; pfn < sec_end_pfn && pfn < end_pfn;
1438 pfn += MAX_ORDER_NR_PAGES) {
1440 /* This is just a CONFIG_HOLES_IN_ZONE check.*/
1441 while ((i < MAX_ORDER_NR_PAGES) &&
1442 !pfn_valid_within(pfn + i))
1444 if (i == MAX_ORDER_NR_PAGES || pfn + i >= end_pfn)
1446 /* Check if we got outside of the zone */
1447 if (zone && !zone_spans_pfn(zone, pfn + i))
1449 page = pfn_to_page(pfn + i);
1450 if (zone && page_zone(page) != zone)
1452 zone = page_zone(page);
1460 * Scan pfn range [start,end) to find movable/migratable pages (LRU pages,
1461 * non-lru movable pages and hugepages). Will skip over most unmovable
1462 * pages (esp., pages that can be skipped when offlining), but bail out on
1463 * definitely unmovable pages.
1466 * 0 in case a movable page is found and movable_pfn was updated.
1467 * -ENOENT in case no movable page was found.
1468 * -EBUSY in case a definitely unmovable page was found.
1470 static int scan_movable_pages(unsigned long start, unsigned long end,
1471 unsigned long *movable_pfn)
1475 for (pfn = start; pfn < end; pfn++) {
1476 struct page *page, *head;
1479 if (!pfn_valid(pfn))
1481 page = pfn_to_page(pfn);
1484 if (__PageMovable(page))
1488 * PageOffline() pages that are not marked __PageMovable() and
1489 * have a reference count > 0 (after MEM_GOING_OFFLINE) are
1490 * definitely unmovable. If their reference count would be 0,
1491 * they could at least be skipped when offlining memory.
1493 if (PageOffline(page) && page_count(page))
1496 if (!PageHuge(page))
1498 head = compound_head(page);
1500 * This test is racy as we hold no reference or lock. The
1501 * hugetlb page could have been free'ed and head is no longer
1502 * a hugetlb page before the following check. In such unlikely
1503 * cases false positives and negatives are possible. Calling
1504 * code must deal with these scenarios.
1506 if (HPageMigratable(head))
1508 skip = compound_nr(head) - (page - head);
1518 do_migrate_range(unsigned long start_pfn, unsigned long end_pfn)
1521 struct page *page, *head;
1525 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
1526 if (!pfn_valid(pfn))
1528 page = pfn_to_page(pfn);
1529 head = compound_head(page);
1531 if (PageHuge(page)) {
1532 pfn = page_to_pfn(head) + compound_nr(head) - 1;
1533 isolate_huge_page(head, &source);
1535 } else if (PageTransHuge(page))
1536 pfn = page_to_pfn(head) + thp_nr_pages(page) - 1;
1539 * HWPoison pages have elevated reference counts so the migration would
1540 * fail on them. It also doesn't make any sense to migrate them in the
1541 * first place. Still try to unmap such a page in case it is still mapped
1542 * (e.g. current hwpoison implementation doesn't unmap KSM pages but keep
1543 * the unmap as the catch all safety net).
1545 if (PageHWPoison(page)) {
1546 if (WARN_ON(PageLRU(page)))
1547 isolate_lru_page(page);
1548 if (page_mapped(page))
1549 try_to_unmap(page, TTU_IGNORE_MLOCK);
1553 if (!get_page_unless_zero(page))
1556 * We can skip free pages. And we can deal with pages on
1557 * LRU and non-lru movable pages.
1560 ret = isolate_lru_page(page);
1562 ret = isolate_movable_page(page, ISOLATE_UNEVICTABLE);
1563 if (!ret) { /* Success */
1564 list_add_tail(&page->lru, &source);
1565 if (!__PageMovable(page))
1566 inc_node_page_state(page, NR_ISOLATED_ANON +
1567 page_is_file_lru(page));
1570 pr_warn("failed to isolate pfn %lx\n", pfn);
1571 dump_page(page, "isolation failed");
1575 if (!list_empty(&source)) {
1576 nodemask_t nmask = node_states[N_MEMORY];
1577 struct migration_target_control mtc = {
1579 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
1583 * We have checked that migration range is on a single zone so
1584 * we can use the nid of the first page to all the others.
1586 mtc.nid = page_to_nid(list_first_entry(&source, struct page, lru));
1589 * try to allocate from a different node but reuse this node
1590 * if there are no other online nodes to be used (e.g. we are
1591 * offlining a part of the only existing node)
1593 node_clear(mtc.nid, nmask);
1594 if (nodes_empty(nmask))
1595 node_set(mtc.nid, nmask);
1596 ret = migrate_pages(&source, alloc_migration_target, NULL,
1597 (unsigned long)&mtc, MIGRATE_SYNC, MR_MEMORY_HOTPLUG);
1599 list_for_each_entry(page, &source, lru) {
1600 pr_warn("migrating pfn %lx failed ret:%d ",
1601 page_to_pfn(page), ret);
1602 dump_page(page, "migration failure");
1604 putback_movable_pages(&source);
1611 static int __init cmdline_parse_movable_node(char *p)
1613 movable_node_enabled = true;
1616 early_param("movable_node", cmdline_parse_movable_node);
1618 /* check which state of node_states will be changed when offline memory */
1619 static void node_states_check_changes_offline(unsigned long nr_pages,
1620 struct zone *zone, struct memory_notify *arg)
1622 struct pglist_data *pgdat = zone->zone_pgdat;
1623 unsigned long present_pages = 0;
1626 arg->status_change_nid = NUMA_NO_NODE;
1627 arg->status_change_nid_normal = NUMA_NO_NODE;
1628 arg->status_change_nid_high = NUMA_NO_NODE;
1631 * Check whether node_states[N_NORMAL_MEMORY] will be changed.
1632 * If the memory to be offline is within the range
1633 * [0..ZONE_NORMAL], and it is the last present memory there,
1634 * the zones in that range will become empty after the offlining,
1635 * thus we can determine that we need to clear the node from
1636 * node_states[N_NORMAL_MEMORY].
1638 for (zt = 0; zt <= ZONE_NORMAL; zt++)
1639 present_pages += pgdat->node_zones[zt].present_pages;
1640 if (zone_idx(zone) <= ZONE_NORMAL && nr_pages >= present_pages)
1641 arg->status_change_nid_normal = zone_to_nid(zone);
1643 #ifdef CONFIG_HIGHMEM
1645 * node_states[N_HIGH_MEMORY] contains nodes which
1646 * have normal memory or high memory.
1647 * Here we add the present_pages belonging to ZONE_HIGHMEM.
1648 * If the zone is within the range of [0..ZONE_HIGHMEM), and
1649 * we determine that the zones in that range become empty,
1650 * we need to clear the node for N_HIGH_MEMORY.
1652 present_pages += pgdat->node_zones[ZONE_HIGHMEM].present_pages;
1653 if (zone_idx(zone) <= ZONE_HIGHMEM && nr_pages >= present_pages)
1654 arg->status_change_nid_high = zone_to_nid(zone);
1658 * We have accounted the pages from [0..ZONE_NORMAL), and
1659 * in case of CONFIG_HIGHMEM the pages from ZONE_HIGHMEM
1661 * Here we count the possible pages from ZONE_MOVABLE.
1662 * If after having accounted all the pages, we see that the nr_pages
1663 * to be offlined is over or equal to the accounted pages,
1664 * we know that the node will become empty, and so, we can clear
1665 * it for N_MEMORY as well.
1667 present_pages += pgdat->node_zones[ZONE_MOVABLE].present_pages;
1669 if (nr_pages >= present_pages)
1670 arg->status_change_nid = zone_to_nid(zone);
1673 static void node_states_clear_node(int node, struct memory_notify *arg)
1675 if (arg->status_change_nid_normal >= 0)
1676 node_clear_state(node, N_NORMAL_MEMORY);
1678 if (arg->status_change_nid_high >= 0)
1679 node_clear_state(node, N_HIGH_MEMORY);
1681 if (arg->status_change_nid >= 0)
1682 node_clear_state(node, N_MEMORY);
1685 static int count_system_ram_pages_cb(unsigned long start_pfn,
1686 unsigned long nr_pages, void *data)
1688 unsigned long *nr_system_ram_pages = data;
1690 *nr_system_ram_pages += nr_pages;
1694 int __ref offline_pages(unsigned long start_pfn, unsigned long nr_pages)
1696 const unsigned long end_pfn = start_pfn + nr_pages;
1697 unsigned long pfn, system_ram_pages = 0;
1698 unsigned long flags;
1700 struct memory_notify arg;
1705 * {on,off}lining is constrained to full memory sections (or more
1706 * precisly to memory blocks from the user space POV).
1707 * memmap_on_memory is an exception because it reserves initial part
1708 * of the physical memory space for vmemmaps. That space is pageblock
1711 if (WARN_ON_ONCE(!nr_pages ||
1712 !IS_ALIGNED(start_pfn, pageblock_nr_pages) ||
1713 !IS_ALIGNED(start_pfn + nr_pages, PAGES_PER_SECTION)))
1716 mem_hotplug_begin();
1719 * Don't allow to offline memory blocks that contain holes.
1720 * Consequently, memory blocks with holes can never get onlined
1721 * via the hotplug path - online_pages() - as hotplugged memory has
1722 * no holes. This way, we e.g., don't have to worry about marking
1723 * memory holes PG_reserved, don't need pfn_valid() checks, and can
1724 * avoid using walk_system_ram_range() later.
1726 walk_system_ram_range(start_pfn, nr_pages, &system_ram_pages,
1727 count_system_ram_pages_cb);
1728 if (system_ram_pages != nr_pages) {
1730 reason = "memory holes";
1731 goto failed_removal;
1734 /* This makes hotplug much easier...and readable.
1735 we assume this for now. .*/
1736 zone = test_pages_in_a_zone(start_pfn, end_pfn);
1739 reason = "multizone range";
1740 goto failed_removal;
1742 node = zone_to_nid(zone);
1745 * Disable pcplists so that page isolation cannot race with freeing
1746 * in a way that pages from isolated pageblock are left on pcplists.
1748 zone_pcp_disable(zone);
1749 lru_cache_disable();
1751 /* set above range as isolated */
1752 ret = start_isolate_page_range(start_pfn, end_pfn,
1754 MEMORY_OFFLINE | REPORT_FAILURE);
1756 reason = "failure to isolate range";
1757 goto failed_removal_pcplists_disabled;
1760 arg.start_pfn = start_pfn;
1761 arg.nr_pages = nr_pages;
1762 node_states_check_changes_offline(nr_pages, zone, &arg);
1764 ret = memory_notify(MEM_GOING_OFFLINE, &arg);
1765 ret = notifier_to_errno(ret);
1767 reason = "notifier failure";
1768 goto failed_removal_isolated;
1774 if (signal_pending(current)) {
1776 reason = "signal backoff";
1777 goto failed_removal_isolated;
1782 ret = scan_movable_pages(pfn, end_pfn, &pfn);
1785 * TODO: fatal migration failures should bail
1788 do_migrate_range(pfn, end_pfn);
1792 if (ret != -ENOENT) {
1793 reason = "unmovable page";
1794 goto failed_removal_isolated;
1798 * Dissolve free hugepages in the memory block before doing
1799 * offlining actually in order to make hugetlbfs's object
1800 * counting consistent.
1802 ret = dissolve_free_huge_pages(start_pfn, end_pfn);
1804 reason = "failure to dissolve huge pages";
1805 goto failed_removal_isolated;
1808 ret = test_pages_isolated(start_pfn, end_pfn, MEMORY_OFFLINE);
1812 /* Mark all sections offline and remove free pages from the buddy. */
1813 __offline_isolated_pages(start_pfn, end_pfn);
1814 pr_debug("Offlined Pages %ld\n", nr_pages);
1817 * The memory sections are marked offline, and the pageblock flags
1818 * effectively stale; nobody should be touching them. Fixup the number
1819 * of isolated pageblocks, memory onlining will properly revert this.
1821 spin_lock_irqsave(&zone->lock, flags);
1822 zone->nr_isolate_pageblock -= nr_pages / pageblock_nr_pages;
1823 spin_unlock_irqrestore(&zone->lock, flags);
1826 zone_pcp_enable(zone);
1828 /* removal success */
1829 adjust_managed_page_count(pfn_to_page(start_pfn), -nr_pages);
1830 adjust_present_page_count(zone, -nr_pages);
1832 init_per_zone_wmark_min();
1834 if (!populated_zone(zone)) {
1835 zone_pcp_reset(zone);
1836 build_all_zonelists(NULL);
1838 zone_pcp_update(zone);
1840 node_states_clear_node(node, &arg);
1841 if (arg.status_change_nid >= 0) {
1843 kcompactd_stop(node);
1846 writeback_set_ratelimit();
1848 memory_notify(MEM_OFFLINE, &arg);
1849 remove_pfn_range_from_zone(zone, start_pfn, nr_pages);
1853 failed_removal_isolated:
1854 undo_isolate_page_range(start_pfn, end_pfn, MIGRATE_MOVABLE);
1855 memory_notify(MEM_CANCEL_OFFLINE, &arg);
1856 failed_removal_pcplists_disabled:
1857 zone_pcp_enable(zone);
1859 pr_debug("memory offlining [mem %#010llx-%#010llx] failed due to %s\n",
1860 (unsigned long long) start_pfn << PAGE_SHIFT,
1861 ((unsigned long long) end_pfn << PAGE_SHIFT) - 1,
1863 /* pushback to free area */
1868 static int check_memblock_offlined_cb(struct memory_block *mem, void *arg)
1870 int ret = !is_memblock_offlined(mem);
1872 if (unlikely(ret)) {
1873 phys_addr_t beginpa, endpa;
1875 beginpa = PFN_PHYS(section_nr_to_pfn(mem->start_section_nr));
1876 endpa = beginpa + memory_block_size_bytes() - 1;
1877 pr_warn("removing memory fails, because memory [%pa-%pa] is onlined\n",
1885 static int get_nr_vmemmap_pages_cb(struct memory_block *mem, void *arg)
1888 * If not set, continue with the next block.
1890 return mem->nr_vmemmap_pages;
1893 static int check_cpu_on_node(pg_data_t *pgdat)
1897 for_each_present_cpu(cpu) {
1898 if (cpu_to_node(cpu) == pgdat->node_id)
1900 * the cpu on this node isn't removed, and we can't
1901 * offline this node.
1909 static int check_no_memblock_for_node_cb(struct memory_block *mem, void *arg)
1911 int nid = *(int *)arg;
1914 * If a memory block belongs to multiple nodes, the stored nid is not
1915 * reliable. However, such blocks are always online (e.g., cannot get
1916 * offlined) and, therefore, are still spanned by the node.
1918 return mem->nid == nid ? -EEXIST : 0;
1925 * Offline a node if all memory sections and cpus of the node are removed.
1927 * NOTE: The caller must call lock_device_hotplug() to serialize hotplug
1928 * and online/offline operations before this call.
1930 void try_offline_node(int nid)
1932 pg_data_t *pgdat = NODE_DATA(nid);
1936 * If the node still spans pages (especially ZONE_DEVICE), don't
1937 * offline it. A node spans memory after move_pfn_range_to_zone(),
1938 * e.g., after the memory block was onlined.
1940 if (pgdat->node_spanned_pages)
1944 * Especially offline memory blocks might not be spanned by the
1945 * node. They will get spanned by the node once they get onlined.
1946 * However, they link to the node in sysfs and can get onlined later.
1948 rc = for_each_memory_block(&nid, check_no_memblock_for_node_cb);
1952 if (check_cpu_on_node(pgdat))
1956 * all memory/cpu of this node are removed, we can offline this
1959 node_set_offline(nid);
1960 unregister_one_node(nid);
1962 EXPORT_SYMBOL(try_offline_node);
1964 static int __ref try_remove_memory(int nid, u64 start, u64 size)
1967 struct vmem_altmap mhp_altmap = {};
1968 struct vmem_altmap *altmap = NULL;
1969 unsigned long nr_vmemmap_pages;
1971 BUG_ON(check_hotplug_memory_range(start, size));
1974 * All memory blocks must be offlined before removing memory. Check
1975 * whether all memory blocks in question are offline and return error
1976 * if this is not the case.
1978 rc = walk_memory_blocks(start, size, NULL, check_memblock_offlined_cb);
1983 * We only support removing memory added with MHP_MEMMAP_ON_MEMORY in
1984 * the same granularity it was added - a single memory block.
1986 if (memmap_on_memory) {
1987 nr_vmemmap_pages = walk_memory_blocks(start, size, NULL,
1988 get_nr_vmemmap_pages_cb);
1989 if (nr_vmemmap_pages) {
1990 if (size != memory_block_size_bytes()) {
1991 pr_warn("Refuse to remove %#llx - %#llx,"
1992 "wrong granularity\n",
1993 start, start + size);
1998 * Let remove_pmd_table->free_hugepage_table do the
1999 * right thing if we used vmem_altmap when hot-adding
2002 mhp_altmap.alloc = nr_vmemmap_pages;
2003 altmap = &mhp_altmap;
2007 /* remove memmap entry */
2008 firmware_map_remove(start, start + size, "System RAM");
2011 * Memory block device removal under the device_hotplug_lock is
2012 * a barrier against racing online attempts.
2014 remove_memory_block_devices(start, size);
2016 mem_hotplug_begin();
2018 arch_remove_memory(nid, start, size, altmap);
2020 if (IS_ENABLED(CONFIG_ARCH_KEEP_MEMBLOCK)) {
2021 memblock_free(start, size);
2022 memblock_remove(start, size);
2025 release_mem_region_adjustable(start, size);
2027 try_offline_node(nid);
2036 * @start: physical address of the region to remove
2037 * @size: size of the region to remove
2039 * NOTE: The caller must call lock_device_hotplug() to serialize hotplug
2040 * and online/offline operations before this call, as required by
2041 * try_offline_node().
2043 void __remove_memory(int nid, u64 start, u64 size)
2047 * trigger BUG() if some memory is not offlined prior to calling this
2050 if (try_remove_memory(nid, start, size))
2055 * Remove memory if every memory block is offline, otherwise return -EBUSY is
2056 * some memory is not offline
2058 int remove_memory(int nid, u64 start, u64 size)
2062 lock_device_hotplug();
2063 rc = try_remove_memory(nid, start, size);
2064 unlock_device_hotplug();
2068 EXPORT_SYMBOL_GPL(remove_memory);
2070 static int try_offline_memory_block(struct memory_block *mem, void *arg)
2072 uint8_t online_type = MMOP_ONLINE_KERNEL;
2073 uint8_t **online_types = arg;
2078 * Sense the online_type via the zone of the memory block. Offlining
2079 * with multiple zones within one memory block will be rejected
2080 * by offlining code ... so we don't care about that.
2082 page = pfn_to_online_page(section_nr_to_pfn(mem->start_section_nr));
2083 if (page && zone_idx(page_zone(page)) == ZONE_MOVABLE)
2084 online_type = MMOP_ONLINE_MOVABLE;
2086 rc = device_offline(&mem->dev);
2088 * Default is MMOP_OFFLINE - change it only if offlining succeeded,
2089 * so try_reonline_memory_block() can do the right thing.
2092 **online_types = online_type;
2095 /* Ignore if already offline. */
2096 return rc < 0 ? rc : 0;
2099 static int try_reonline_memory_block(struct memory_block *mem, void *arg)
2101 uint8_t **online_types = arg;
2104 if (**online_types != MMOP_OFFLINE) {
2105 mem->online_type = **online_types;
2106 rc = device_online(&mem->dev);
2108 pr_warn("%s: Failed to re-online memory: %d",
2112 /* Continue processing all remaining memory blocks. */
2118 * Try to offline and remove memory. Might take a long time to finish in case
2119 * memory is still in use. Primarily useful for memory devices that logically
2120 * unplugged all memory (so it's no longer in use) and want to offline + remove
2123 int offline_and_remove_memory(int nid, u64 start, u64 size)
2125 const unsigned long mb_count = size / memory_block_size_bytes();
2126 uint8_t *online_types, *tmp;
2129 if (!IS_ALIGNED(start, memory_block_size_bytes()) ||
2130 !IS_ALIGNED(size, memory_block_size_bytes()) || !size)
2134 * We'll remember the old online type of each memory block, so we can
2135 * try to revert whatever we did when offlining one memory block fails
2136 * after offlining some others succeeded.
2138 online_types = kmalloc_array(mb_count, sizeof(*online_types),
2143 * Initialize all states to MMOP_OFFLINE, so when we abort processing in
2144 * try_offline_memory_block(), we'll skip all unprocessed blocks in
2145 * try_reonline_memory_block().
2147 memset(online_types, MMOP_OFFLINE, mb_count);
2149 lock_device_hotplug();
2152 rc = walk_memory_blocks(start, size, &tmp, try_offline_memory_block);
2155 * In case we succeeded to offline all memory, remove it.
2156 * This cannot fail as it cannot get onlined in the meantime.
2159 rc = try_remove_memory(nid, start, size);
2161 pr_err("%s: Failed to remove memory: %d", __func__, rc);
2165 * Rollback what we did. While memory onlining might theoretically fail
2166 * (nacked by a notifier), it barely ever happens.
2170 walk_memory_blocks(start, size, &tmp,
2171 try_reonline_memory_block);
2173 unlock_device_hotplug();
2175 kfree(online_types);
2178 EXPORT_SYMBOL_GPL(offline_and_remove_memory);
2179 #endif /* CONFIG_MEMORY_HOTREMOVE */