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 static int check_pfn_span(unsigned long pfn, unsigned long nr_pages,
161 * Disallow all operations smaller than a sub-section and only
162 * allow operations smaller than a section for
163 * SPARSEMEM_VMEMMAP. Note that check_hotplug_memory_range()
164 * enforces a larger memory_block_size_bytes() granularity for
165 * memory that will be marked online, so this check should only
166 * fire for direct arch_{add,remove}_memory() users outside of
167 * add_memory_resource().
169 unsigned long min_align;
171 if (IS_ENABLED(CONFIG_SPARSEMEM_VMEMMAP))
172 min_align = PAGES_PER_SUBSECTION;
174 min_align = PAGES_PER_SECTION;
175 if (!IS_ALIGNED(pfn, min_align)
176 || !IS_ALIGNED(nr_pages, min_align)) {
177 WARN(1, "Misaligned __%s_pages start: %#lx end: #%lx\n",
178 reason, pfn, pfn + nr_pages - 1);
185 * Return page for the valid pfn only if the page is online. All pfn
186 * walkers which rely on the fully initialized page->flags and others
187 * should use this rather than pfn_valid && pfn_to_page
189 struct page *pfn_to_online_page(unsigned long pfn)
191 unsigned long nr = pfn_to_section_nr(pfn);
192 struct dev_pagemap *pgmap;
193 struct mem_section *ms;
195 if (nr >= NR_MEM_SECTIONS)
198 ms = __nr_to_section(nr);
199 if (!online_section(ms))
203 * Save some code text when online_section() +
204 * pfn_section_valid() are sufficient.
206 if (IS_ENABLED(CONFIG_HAVE_ARCH_PFN_VALID) && !pfn_valid(pfn))
209 if (!pfn_section_valid(ms, pfn))
212 if (!online_device_section(ms))
213 return pfn_to_page(pfn);
216 * Slowpath: when ZONE_DEVICE collides with
217 * ZONE_{NORMAL,MOVABLE} within the same section some pfns in
218 * the section may be 'offline' but 'valid'. Only
219 * get_dev_pagemap() can determine sub-section online status.
221 pgmap = get_dev_pagemap(pfn, NULL);
222 put_dev_pagemap(pgmap);
224 /* The presence of a pgmap indicates ZONE_DEVICE offline pfn */
228 return pfn_to_page(pfn);
230 EXPORT_SYMBOL_GPL(pfn_to_online_page);
233 * Reasonably generic function for adding memory. It is
234 * expected that archs that support memory hotplug will
235 * call this function after deciding the zone to which to
238 int __ref __add_pages(int nid, unsigned long pfn, unsigned long nr_pages,
239 struct mhp_params *params)
241 const unsigned long end_pfn = pfn + nr_pages;
242 unsigned long cur_nr_pages;
244 struct vmem_altmap *altmap = params->altmap;
246 if (WARN_ON_ONCE(!params->pgprot.pgprot))
249 VM_BUG_ON(!mhp_range_allowed(PFN_PHYS(pfn), nr_pages * PAGE_SIZE, false));
253 * Validate altmap is within bounds of the total request
255 if (altmap->base_pfn != pfn
256 || vmem_altmap_offset(altmap) > nr_pages) {
257 pr_warn_once("memory add fail, invalid altmap\n");
263 err = check_pfn_span(pfn, nr_pages, "add");
267 for (; pfn < end_pfn; pfn += cur_nr_pages) {
268 /* Select all remaining pages up to the next section boundary */
269 cur_nr_pages = min(end_pfn - pfn,
270 SECTION_ALIGN_UP(pfn + 1) - pfn);
271 err = sparse_add_section(nid, pfn, cur_nr_pages, altmap);
276 vmemmap_populate_print_last();
280 /* find the smallest valid pfn in the range [start_pfn, end_pfn) */
281 static unsigned long find_smallest_section_pfn(int nid, struct zone *zone,
282 unsigned long start_pfn,
283 unsigned long end_pfn)
285 for (; start_pfn < end_pfn; start_pfn += PAGES_PER_SUBSECTION) {
286 if (unlikely(!pfn_to_online_page(start_pfn)))
289 if (unlikely(pfn_to_nid(start_pfn) != nid))
292 if (zone != page_zone(pfn_to_page(start_pfn)))
301 /* find the biggest valid pfn in the range [start_pfn, end_pfn). */
302 static unsigned long find_biggest_section_pfn(int nid, struct zone *zone,
303 unsigned long start_pfn,
304 unsigned long end_pfn)
308 /* pfn is the end pfn of a memory section. */
310 for (; pfn >= start_pfn; pfn -= PAGES_PER_SUBSECTION) {
311 if (unlikely(!pfn_to_online_page(pfn)))
314 if (unlikely(pfn_to_nid(pfn) != nid))
317 if (zone != page_zone(pfn_to_page(pfn)))
326 static void shrink_zone_span(struct zone *zone, unsigned long start_pfn,
327 unsigned long end_pfn)
330 int nid = zone_to_nid(zone);
332 if (zone->zone_start_pfn == start_pfn) {
334 * If the section is smallest section in the zone, it need
335 * shrink zone->zone_start_pfn and zone->zone_spanned_pages.
336 * In this case, we find second smallest valid mem_section
337 * for shrinking zone.
339 pfn = find_smallest_section_pfn(nid, zone, end_pfn,
342 zone->spanned_pages = zone_end_pfn(zone) - pfn;
343 zone->zone_start_pfn = pfn;
345 zone->zone_start_pfn = 0;
346 zone->spanned_pages = 0;
348 } else if (zone_end_pfn(zone) == end_pfn) {
350 * If the section is biggest section in the zone, it need
351 * shrink zone->spanned_pages.
352 * In this case, we find second biggest valid mem_section for
355 pfn = find_biggest_section_pfn(nid, zone, zone->zone_start_pfn,
358 zone->spanned_pages = pfn - zone->zone_start_pfn + 1;
360 zone->zone_start_pfn = 0;
361 zone->spanned_pages = 0;
366 static void update_pgdat_span(struct pglist_data *pgdat)
368 unsigned long node_start_pfn = 0, node_end_pfn = 0;
371 for (zone = pgdat->node_zones;
372 zone < pgdat->node_zones + MAX_NR_ZONES; zone++) {
373 unsigned long end_pfn = zone_end_pfn(zone);
375 /* No need to lock the zones, they can't change. */
376 if (!zone->spanned_pages)
379 node_start_pfn = zone->zone_start_pfn;
380 node_end_pfn = end_pfn;
384 if (end_pfn > node_end_pfn)
385 node_end_pfn = end_pfn;
386 if (zone->zone_start_pfn < node_start_pfn)
387 node_start_pfn = zone->zone_start_pfn;
390 pgdat->node_start_pfn = node_start_pfn;
391 pgdat->node_spanned_pages = node_end_pfn - node_start_pfn;
394 void __ref remove_pfn_range_from_zone(struct zone *zone,
395 unsigned long start_pfn,
396 unsigned long nr_pages)
398 const unsigned long end_pfn = start_pfn + nr_pages;
399 struct pglist_data *pgdat = zone->zone_pgdat;
400 unsigned long pfn, cur_nr_pages;
402 /* Poison struct pages because they are now uninitialized again. */
403 for (pfn = start_pfn; pfn < end_pfn; pfn += cur_nr_pages) {
406 /* Select all remaining pages up to the next section boundary */
408 min(end_pfn - pfn, SECTION_ALIGN_UP(pfn + 1) - pfn);
409 page_init_poison(pfn_to_page(pfn),
410 sizeof(struct page) * cur_nr_pages);
413 #ifdef CONFIG_ZONE_DEVICE
415 * Zone shrinking code cannot properly deal with ZONE_DEVICE. So
416 * we will not try to shrink the zones - which is okay as
417 * set_zone_contiguous() cannot deal with ZONE_DEVICE either way.
419 if (zone_idx(zone) == ZONE_DEVICE)
423 clear_zone_contiguous(zone);
425 shrink_zone_span(zone, start_pfn, start_pfn + nr_pages);
426 update_pgdat_span(pgdat);
428 set_zone_contiguous(zone);
431 static void __remove_section(unsigned long pfn, unsigned long nr_pages,
432 unsigned long map_offset,
433 struct vmem_altmap *altmap)
435 struct mem_section *ms = __pfn_to_section(pfn);
437 if (WARN_ON_ONCE(!valid_section(ms)))
440 sparse_remove_section(ms, pfn, nr_pages, map_offset, altmap);
444 * __remove_pages() - remove sections of pages
445 * @pfn: starting pageframe (must be aligned to start of a section)
446 * @nr_pages: number of pages to remove (must be multiple of section size)
447 * @altmap: alternative device page map or %NULL if default memmap is used
449 * Generic helper function to remove section mappings and sysfs entries
450 * for the section of the memory we are removing. Caller needs to make
451 * sure that pages are marked reserved and zones are adjust properly by
452 * calling offline_pages().
454 void __remove_pages(unsigned long pfn, unsigned long nr_pages,
455 struct vmem_altmap *altmap)
457 const unsigned long end_pfn = pfn + nr_pages;
458 unsigned long cur_nr_pages;
459 unsigned long map_offset = 0;
461 map_offset = vmem_altmap_offset(altmap);
463 if (check_pfn_span(pfn, nr_pages, "remove"))
466 for (; pfn < end_pfn; pfn += cur_nr_pages) {
468 /* Select all remaining pages up to the next section boundary */
469 cur_nr_pages = min(end_pfn - pfn,
470 SECTION_ALIGN_UP(pfn + 1) - pfn);
471 __remove_section(pfn, cur_nr_pages, map_offset, altmap);
476 int set_online_page_callback(online_page_callback_t callback)
481 mutex_lock(&online_page_callback_lock);
483 if (online_page_callback == generic_online_page) {
484 online_page_callback = callback;
488 mutex_unlock(&online_page_callback_lock);
493 EXPORT_SYMBOL_GPL(set_online_page_callback);
495 int restore_online_page_callback(online_page_callback_t callback)
500 mutex_lock(&online_page_callback_lock);
502 if (online_page_callback == callback) {
503 online_page_callback = generic_online_page;
507 mutex_unlock(&online_page_callback_lock);
512 EXPORT_SYMBOL_GPL(restore_online_page_callback);
514 void generic_online_page(struct page *page, unsigned int order)
517 * Freeing the page with debug_pagealloc enabled will try to unmap it,
518 * so we should map it first. This is better than introducing a special
519 * case in page freeing fast path.
521 debug_pagealloc_map_pages(page, 1 << order);
522 __free_pages_core(page, order);
523 totalram_pages_add(1UL << order);
524 #ifdef CONFIG_HIGHMEM
525 if (PageHighMem(page))
526 totalhigh_pages_add(1UL << order);
529 EXPORT_SYMBOL_GPL(generic_online_page);
531 static void online_pages_range(unsigned long start_pfn, unsigned long nr_pages)
533 const unsigned long end_pfn = start_pfn + nr_pages;
537 * Online the pages in MAX_ORDER - 1 aligned chunks. The callback might
538 * decide to not expose all pages to the buddy (e.g., expose them
539 * later). We account all pages as being online and belonging to this
541 * When using memmap_on_memory, the range might not be aligned to
542 * MAX_ORDER_NR_PAGES - 1, but pageblock aligned. __ffs() will detect
543 * this and the first chunk to online will be pageblock_nr_pages.
545 for (pfn = start_pfn; pfn < end_pfn;) {
546 int order = min(MAX_ORDER - 1UL, __ffs(pfn));
548 (*online_page_callback)(pfn_to_page(pfn), order);
549 pfn += (1UL << order);
552 /* mark all involved sections as online */
553 online_mem_sections(start_pfn, end_pfn);
556 /* check which state of node_states will be changed when online memory */
557 static void node_states_check_changes_online(unsigned long nr_pages,
558 struct zone *zone, struct memory_notify *arg)
560 int nid = zone_to_nid(zone);
562 arg->status_change_nid = NUMA_NO_NODE;
563 arg->status_change_nid_normal = NUMA_NO_NODE;
564 arg->status_change_nid_high = NUMA_NO_NODE;
566 if (!node_state(nid, N_MEMORY))
567 arg->status_change_nid = nid;
568 if (zone_idx(zone) <= ZONE_NORMAL && !node_state(nid, N_NORMAL_MEMORY))
569 arg->status_change_nid_normal = nid;
570 #ifdef CONFIG_HIGHMEM
571 if (zone_idx(zone) <= ZONE_HIGHMEM && !node_state(nid, N_HIGH_MEMORY))
572 arg->status_change_nid_high = nid;
576 static void node_states_set_node(int node, struct memory_notify *arg)
578 if (arg->status_change_nid_normal >= 0)
579 node_set_state(node, N_NORMAL_MEMORY);
581 if (arg->status_change_nid_high >= 0)
582 node_set_state(node, N_HIGH_MEMORY);
584 if (arg->status_change_nid >= 0)
585 node_set_state(node, N_MEMORY);
588 static void __meminit resize_zone_range(struct zone *zone, unsigned long start_pfn,
589 unsigned long nr_pages)
591 unsigned long old_end_pfn = zone_end_pfn(zone);
593 if (zone_is_empty(zone) || start_pfn < zone->zone_start_pfn)
594 zone->zone_start_pfn = start_pfn;
596 zone->spanned_pages = max(start_pfn + nr_pages, old_end_pfn) - zone->zone_start_pfn;
599 static void __meminit resize_pgdat_range(struct pglist_data *pgdat, unsigned long start_pfn,
600 unsigned long nr_pages)
602 unsigned long old_end_pfn = pgdat_end_pfn(pgdat);
604 if (!pgdat->node_spanned_pages || start_pfn < pgdat->node_start_pfn)
605 pgdat->node_start_pfn = start_pfn;
607 pgdat->node_spanned_pages = max(start_pfn + nr_pages, old_end_pfn) - pgdat->node_start_pfn;
611 static void section_taint_zone_device(unsigned long pfn)
613 struct mem_section *ms = __pfn_to_section(pfn);
615 ms->section_mem_map |= SECTION_TAINT_ZONE_DEVICE;
619 * Associate the pfn range with the given zone, initializing the memmaps
620 * and resizing the pgdat/zone data to span the added pages. After this
621 * call, all affected pages are PG_reserved.
623 * All aligned pageblocks are initialized to the specified migratetype
624 * (usually MIGRATE_MOVABLE). Besides setting the migratetype, no related
625 * zone stats (e.g., nr_isolate_pageblock) are touched.
627 void __ref move_pfn_range_to_zone(struct zone *zone, unsigned long start_pfn,
628 unsigned long nr_pages,
629 struct vmem_altmap *altmap, int migratetype)
631 struct pglist_data *pgdat = zone->zone_pgdat;
632 int nid = pgdat->node_id;
634 clear_zone_contiguous(zone);
636 if (zone_is_empty(zone))
637 init_currently_empty_zone(zone, start_pfn, nr_pages);
638 resize_zone_range(zone, start_pfn, nr_pages);
639 resize_pgdat_range(pgdat, start_pfn, nr_pages);
642 * Subsection population requires care in pfn_to_online_page().
643 * Set the taint to enable the slow path detection of
644 * ZONE_DEVICE pages in an otherwise ZONE_{NORMAL,MOVABLE}
647 if (zone_is_zone_device(zone)) {
648 if (!IS_ALIGNED(start_pfn, PAGES_PER_SECTION))
649 section_taint_zone_device(start_pfn);
650 if (!IS_ALIGNED(start_pfn + nr_pages, PAGES_PER_SECTION))
651 section_taint_zone_device(start_pfn + nr_pages);
655 * TODO now we have a visible range of pages which are not associated
656 * with their zone properly. Not nice but set_pfnblock_flags_mask
657 * expects the zone spans the pfn range. All the pages in the range
658 * are reserved so nobody should be touching them so we should be safe
660 memmap_init_range(nr_pages, nid, zone_idx(zone), start_pfn, 0,
661 MEMINIT_HOTPLUG, altmap, migratetype);
663 set_zone_contiguous(zone);
667 * Returns a default kernel memory zone for the given pfn range.
668 * If no kernel zone covers this pfn range it will automatically go
669 * to the ZONE_NORMAL.
671 static struct zone *default_kernel_zone_for_pfn(int nid, unsigned long start_pfn,
672 unsigned long nr_pages)
674 struct pglist_data *pgdat = NODE_DATA(nid);
677 for (zid = 0; zid <= ZONE_NORMAL; zid++) {
678 struct zone *zone = &pgdat->node_zones[zid];
680 if (zone_intersects(zone, start_pfn, nr_pages))
684 return &pgdat->node_zones[ZONE_NORMAL];
687 static inline struct zone *default_zone_for_pfn(int nid, unsigned long start_pfn,
688 unsigned long nr_pages)
690 struct zone *kernel_zone = default_kernel_zone_for_pfn(nid, start_pfn,
692 struct zone *movable_zone = &NODE_DATA(nid)->node_zones[ZONE_MOVABLE];
693 bool in_kernel = zone_intersects(kernel_zone, start_pfn, nr_pages);
694 bool in_movable = zone_intersects(movable_zone, start_pfn, nr_pages);
697 * We inherit the existing zone in a simple case where zones do not
698 * overlap in the given range
700 if (in_kernel ^ in_movable)
701 return (in_kernel) ? kernel_zone : movable_zone;
704 * If the range doesn't belong to any zone or two zones overlap in the
705 * given range then we use movable zone only if movable_node is
706 * enabled because we always online to a kernel zone by default.
708 return movable_node_enabled ? movable_zone : kernel_zone;
711 struct zone *zone_for_pfn_range(int online_type, int nid, unsigned start_pfn,
712 unsigned long nr_pages)
714 if (online_type == MMOP_ONLINE_KERNEL)
715 return default_kernel_zone_for_pfn(nid, start_pfn, nr_pages);
717 if (online_type == MMOP_ONLINE_MOVABLE)
718 return &NODE_DATA(nid)->node_zones[ZONE_MOVABLE];
720 return default_zone_for_pfn(nid, start_pfn, nr_pages);
724 * This function should only be called by memory_block_{online,offline},
725 * and {online,offline}_pages.
727 void adjust_present_page_count(struct zone *zone, long nr_pages)
729 zone->present_pages += nr_pages;
730 zone->zone_pgdat->node_present_pages += nr_pages;
733 int mhp_init_memmap_on_memory(unsigned long pfn, unsigned long nr_pages,
736 unsigned long end_pfn = pfn + nr_pages;
739 ret = kasan_add_zero_shadow(__va(PFN_PHYS(pfn)), PFN_PHYS(nr_pages));
743 move_pfn_range_to_zone(zone, pfn, nr_pages, NULL, MIGRATE_UNMOVABLE);
746 * It might be that the vmemmap_pages fully span sections. If that is
747 * the case, mark those sections online here as otherwise they will be
750 if (nr_pages >= PAGES_PER_SECTION)
751 online_mem_sections(pfn, ALIGN_DOWN(end_pfn, PAGES_PER_SECTION));
756 void mhp_deinit_memmap_on_memory(unsigned long pfn, unsigned long nr_pages)
758 unsigned long end_pfn = pfn + nr_pages;
761 * It might be that the vmemmap_pages fully span sections. If that is
762 * the case, mark those sections offline here as otherwise they will be
765 if (nr_pages >= PAGES_PER_SECTION)
766 offline_mem_sections(pfn, ALIGN_DOWN(end_pfn, PAGES_PER_SECTION));
769 * The pages associated with this vmemmap have been offlined, so
770 * we can reset its state here.
772 remove_pfn_range_from_zone(page_zone(pfn_to_page(pfn)), pfn, nr_pages);
773 kasan_remove_zero_shadow(__va(PFN_PHYS(pfn)), PFN_PHYS(nr_pages));
776 int __ref online_pages(unsigned long pfn, unsigned long nr_pages, struct zone *zone)
779 int need_zonelists_rebuild = 0;
780 const int nid = zone_to_nid(zone);
782 struct memory_notify arg;
785 * {on,off}lining is constrained to full memory sections (or more
786 * precisely to memory blocks from the user space POV).
787 * memmap_on_memory is an exception because it reserves initial part
788 * of the physical memory space for vmemmaps. That space is pageblock
791 if (WARN_ON_ONCE(!nr_pages ||
792 !IS_ALIGNED(pfn, pageblock_nr_pages) ||
793 !IS_ALIGNED(pfn + nr_pages, PAGES_PER_SECTION)))
798 /* associate pfn range with the zone */
799 move_pfn_range_to_zone(zone, pfn, nr_pages, NULL, MIGRATE_ISOLATE);
802 arg.nr_pages = nr_pages;
803 node_states_check_changes_online(nr_pages, zone, &arg);
805 ret = memory_notify(MEM_GOING_ONLINE, &arg);
806 ret = notifier_to_errno(ret);
808 goto failed_addition;
811 * Fixup the number of isolated pageblocks before marking the sections
812 * onlining, such that undo_isolate_page_range() works correctly.
814 spin_lock_irqsave(&zone->lock, flags);
815 zone->nr_isolate_pageblock += nr_pages / pageblock_nr_pages;
816 spin_unlock_irqrestore(&zone->lock, flags);
819 * If this zone is not populated, then it is not in zonelist.
820 * This means the page allocator ignores this zone.
821 * So, zonelist must be updated after online.
823 if (!populated_zone(zone)) {
824 need_zonelists_rebuild = 1;
825 setup_zone_pageset(zone);
828 online_pages_range(pfn, nr_pages);
829 adjust_present_page_count(zone, nr_pages);
831 node_states_set_node(nid, &arg);
832 if (need_zonelists_rebuild)
833 build_all_zonelists(NULL);
835 /* Basic onlining is complete, allow allocation of onlined pages. */
836 undo_isolate_page_range(pfn, pfn + nr_pages, MIGRATE_MOVABLE);
839 * Freshly onlined pages aren't shuffled (e.g., all pages are placed to
840 * the tail of the freelist when undoing isolation). Shuffle the whole
841 * zone to make sure the just onlined pages are properly distributed
842 * across the whole freelist - to create an initial shuffle.
846 /* reinitialise watermarks and update pcp limits */
847 init_per_zone_wmark_min();
852 writeback_set_ratelimit();
854 memory_notify(MEM_ONLINE, &arg);
859 pr_debug("online_pages [mem %#010llx-%#010llx] failed\n",
860 (unsigned long long) pfn << PAGE_SHIFT,
861 (((unsigned long long) pfn + nr_pages) << PAGE_SHIFT) - 1);
862 memory_notify(MEM_CANCEL_ONLINE, &arg);
863 remove_pfn_range_from_zone(zone, pfn, nr_pages);
867 #endif /* CONFIG_MEMORY_HOTPLUG_SPARSE */
869 static void reset_node_present_pages(pg_data_t *pgdat)
873 for (z = pgdat->node_zones; z < pgdat->node_zones + MAX_NR_ZONES; z++)
874 z->present_pages = 0;
876 pgdat->node_present_pages = 0;
879 /* we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG */
880 static pg_data_t __ref *hotadd_new_pgdat(int nid)
882 struct pglist_data *pgdat;
884 pgdat = NODE_DATA(nid);
886 pgdat = arch_alloc_nodedata(nid);
890 pgdat->per_cpu_nodestats =
891 alloc_percpu(struct per_cpu_nodestat);
892 arch_refresh_nodedata(nid, pgdat);
896 * Reset the nr_zones, order and highest_zoneidx before reuse.
897 * Note that kswapd will init kswapd_highest_zoneidx properly
898 * when it starts in the near future.
901 pgdat->kswapd_order = 0;
902 pgdat->kswapd_highest_zoneidx = 0;
903 for_each_online_cpu(cpu) {
904 struct per_cpu_nodestat *p;
906 p = per_cpu_ptr(pgdat->per_cpu_nodestats, cpu);
907 memset(p, 0, sizeof(*p));
911 /* we can use NODE_DATA(nid) from here */
912 pgdat->node_id = nid;
913 pgdat->node_start_pfn = 0;
915 /* init node's zones as empty zones, we don't have any present pages.*/
916 free_area_init_core_hotplug(nid);
919 * The node we allocated has no zone fallback lists. For avoiding
920 * to access not-initialized zonelist, build here.
922 build_all_zonelists(pgdat);
925 * When memory is hot-added, all the memory is in offline state. So
926 * clear all zones' present_pages because they will be updated in
927 * online_pages() and offline_pages().
929 reset_node_managed_pages(pgdat);
930 reset_node_present_pages(pgdat);
935 static void rollback_node_hotadd(int nid)
937 pg_data_t *pgdat = NODE_DATA(nid);
939 arch_refresh_nodedata(nid, NULL);
940 free_percpu(pgdat->per_cpu_nodestats);
941 arch_free_nodedata(pgdat);
946 * __try_online_node - online a node if offlined
948 * @set_node_online: Whether we want to online the node
949 * called by cpu_up() to online a node without onlined memory.
952 * 1 -> a new node has been allocated
953 * 0 -> the node is already online
954 * -ENOMEM -> the node could not be allocated
956 static int __try_online_node(int nid, bool set_node_online)
961 if (node_online(nid))
964 pgdat = hotadd_new_pgdat(nid);
966 pr_err("Cannot online node %d due to NULL pgdat\n", nid);
971 if (set_node_online) {
972 node_set_online(nid);
973 ret = register_one_node(nid);
981 * Users of this function always want to online/register the node
983 int try_online_node(int nid)
988 ret = __try_online_node(nid, true);
993 static int check_hotplug_memory_range(u64 start, u64 size)
995 /* memory range must be block size aligned */
996 if (!size || !IS_ALIGNED(start, memory_block_size_bytes()) ||
997 !IS_ALIGNED(size, memory_block_size_bytes())) {
998 pr_err("Block size [%#lx] unaligned hotplug range: start %#llx, size %#llx",
999 memory_block_size_bytes(), start, size);
1006 static int online_memory_block(struct memory_block *mem, void *arg)
1008 mem->online_type = mhp_default_online_type;
1009 return device_online(&mem->dev);
1012 bool mhp_supports_memmap_on_memory(unsigned long size)
1014 unsigned long nr_vmemmap_pages = size / PAGE_SIZE;
1015 unsigned long vmemmap_size = nr_vmemmap_pages * sizeof(struct page);
1016 unsigned long remaining_size = size - vmemmap_size;
1019 * Besides having arch support and the feature enabled at runtime, we
1020 * need a few more assumptions to hold true:
1022 * a) We span a single memory block: memory onlining/offlinin;g happens
1023 * in memory block granularity. We don't want the vmemmap of online
1024 * memory blocks to reside on offline memory blocks. In the future,
1025 * we might want to support variable-sized memory blocks to make the
1026 * feature more versatile.
1028 * b) The vmemmap pages span complete PMDs: We don't want vmemmap code
1029 * to populate memory from the altmap for unrelated parts (i.e.,
1030 * other memory blocks)
1032 * c) The vmemmap pages (and thereby the pages that will be exposed to
1033 * the buddy) have to cover full pageblocks: memory onlining/offlining
1034 * code requires applicable ranges to be page-aligned, for example, to
1035 * set the migratetypes properly.
1037 * TODO: Although we have a check here to make sure that vmemmap pages
1038 * fully populate a PMD, it is not the right place to check for
1039 * this. A much better solution involves improving vmemmap code
1040 * to fallback to base pages when trying to populate vmemmap using
1041 * altmap as an alternative source of memory, and we do not exactly
1042 * populate a single PMD.
1044 return memmap_on_memory &&
1045 !hugetlb_free_vmemmap_enabled &&
1046 IS_ENABLED(CONFIG_MHP_MEMMAP_ON_MEMORY) &&
1047 size == memory_block_size_bytes() &&
1048 IS_ALIGNED(vmemmap_size, PMD_SIZE) &&
1049 IS_ALIGNED(remaining_size, (pageblock_nr_pages << PAGE_SHIFT));
1053 * NOTE: The caller must call lock_device_hotplug() to serialize hotplug
1054 * and online/offline operations (triggered e.g. by sysfs).
1056 * we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG
1058 int __ref add_memory_resource(int nid, struct resource *res, mhp_t mhp_flags)
1060 struct mhp_params params = { .pgprot = pgprot_mhp(PAGE_KERNEL) };
1061 struct vmem_altmap mhp_altmap = {};
1063 bool new_node = false;
1067 size = resource_size(res);
1069 ret = check_hotplug_memory_range(start, size);
1073 if (!node_possible(nid)) {
1074 WARN(1, "node %d was absent from the node_possible_map\n", nid);
1078 mem_hotplug_begin();
1080 if (IS_ENABLED(CONFIG_ARCH_KEEP_MEMBLOCK))
1081 memblock_add_node(start, size, nid);
1083 ret = __try_online_node(nid, false);
1089 * Self hosted memmap array
1091 if (mhp_flags & MHP_MEMMAP_ON_MEMORY) {
1092 if (!mhp_supports_memmap_on_memory(size)) {
1096 mhp_altmap.free = PHYS_PFN(size);
1097 mhp_altmap.base_pfn = PHYS_PFN(start);
1098 params.altmap = &mhp_altmap;
1101 /* call arch's memory hotadd */
1102 ret = arch_add_memory(nid, start, size, ¶ms);
1106 /* create memory block devices after memory was added */
1107 ret = create_memory_block_devices(start, size, mhp_altmap.alloc);
1109 arch_remove_memory(nid, start, size, NULL);
1114 /* If sysfs file of new node can't be created, cpu on the node
1115 * can't be hot-added. There is no rollback way now.
1116 * So, check by BUG_ON() to catch it reluctantly..
1117 * We online node here. We can't roll back from here.
1119 node_set_online(nid);
1120 ret = __register_one_node(nid);
1124 /* link memory sections under this node.*/
1125 link_mem_sections(nid, PFN_DOWN(start), PFN_UP(start + size - 1),
1128 /* create new memmap entry */
1129 if (!strcmp(res->name, "System RAM"))
1130 firmware_map_add_hotplug(start, start + size, "System RAM");
1132 /* device_online() will take the lock when calling online_pages() */
1136 * In case we're allowed to merge the resource, flag it and trigger
1137 * merging now that adding succeeded.
1139 if (mhp_flags & MHP_MERGE_RESOURCE)
1140 merge_system_ram_resource(res);
1142 /* online pages if requested */
1143 if (mhp_default_online_type != MMOP_OFFLINE)
1144 walk_memory_blocks(start, size, NULL, online_memory_block);
1148 /* rollback pgdat allocation and others */
1150 rollback_node_hotadd(nid);
1151 if (IS_ENABLED(CONFIG_ARCH_KEEP_MEMBLOCK))
1152 memblock_remove(start, size);
1157 /* requires device_hotplug_lock, see add_memory_resource() */
1158 int __ref __add_memory(int nid, u64 start, u64 size, mhp_t mhp_flags)
1160 struct resource *res;
1163 res = register_memory_resource(start, size, "System RAM");
1165 return PTR_ERR(res);
1167 ret = add_memory_resource(nid, res, mhp_flags);
1169 release_memory_resource(res);
1173 int add_memory(int nid, u64 start, u64 size, mhp_t mhp_flags)
1177 lock_device_hotplug();
1178 rc = __add_memory(nid, start, size, mhp_flags);
1179 unlock_device_hotplug();
1183 EXPORT_SYMBOL_GPL(add_memory);
1186 * Add special, driver-managed memory to the system as system RAM. Such
1187 * memory is not exposed via the raw firmware-provided memmap as system
1188 * RAM, instead, it is detected and added by a driver - during cold boot,
1189 * after a reboot, and after kexec.
1191 * Reasons why this memory should not be used for the initial memmap of a
1192 * kexec kernel or for placing kexec images:
1193 * - The booting kernel is in charge of determining how this memory will be
1194 * used (e.g., use persistent memory as system RAM)
1195 * - Coordination with a hypervisor is required before this memory
1196 * can be used (e.g., inaccessible parts).
1198 * For this memory, no entries in /sys/firmware/memmap ("raw firmware-provided
1199 * memory map") are created. Also, the created memory resource is flagged
1200 * with IORESOURCE_SYSRAM_DRIVER_MANAGED, so in-kernel users can special-case
1201 * this memory as well (esp., not place kexec images onto it).
1203 * The resource_name (visible via /proc/iomem) has to have the format
1204 * "System RAM ($DRIVER)".
1206 int add_memory_driver_managed(int nid, u64 start, u64 size,
1207 const char *resource_name, mhp_t mhp_flags)
1209 struct resource *res;
1212 if (!resource_name ||
1213 strstr(resource_name, "System RAM (") != resource_name ||
1214 resource_name[strlen(resource_name) - 1] != ')')
1217 lock_device_hotplug();
1219 res = register_memory_resource(start, size, resource_name);
1225 rc = add_memory_resource(nid, res, mhp_flags);
1227 release_memory_resource(res);
1230 unlock_device_hotplug();
1233 EXPORT_SYMBOL_GPL(add_memory_driver_managed);
1236 * Platforms should define arch_get_mappable_range() that provides
1237 * maximum possible addressable physical memory range for which the
1238 * linear mapping could be created. The platform returned address
1239 * range must adhere to these following semantics.
1241 * - range.start <= range.end
1242 * - Range includes both end points [range.start..range.end]
1244 * There is also a fallback definition provided here, allowing the
1245 * entire possible physical address range in case any platform does
1246 * not define arch_get_mappable_range().
1248 struct range __weak arch_get_mappable_range(void)
1250 struct range mhp_range = {
1257 struct range mhp_get_pluggable_range(bool need_mapping)
1259 const u64 max_phys = (1ULL << MAX_PHYSMEM_BITS) - 1;
1260 struct range mhp_range;
1263 mhp_range = arch_get_mappable_range();
1264 if (mhp_range.start > max_phys) {
1265 mhp_range.start = 0;
1268 mhp_range.end = min_t(u64, mhp_range.end, max_phys);
1270 mhp_range.start = 0;
1271 mhp_range.end = max_phys;
1275 EXPORT_SYMBOL_GPL(mhp_get_pluggable_range);
1277 bool mhp_range_allowed(u64 start, u64 size, bool need_mapping)
1279 struct range mhp_range = mhp_get_pluggable_range(need_mapping);
1280 u64 end = start + size;
1282 if (start < end && start >= mhp_range.start && (end - 1) <= mhp_range.end)
1285 pr_warn("Hotplug memory [%#llx-%#llx] exceeds maximum addressable range [%#llx-%#llx]\n",
1286 start, end, mhp_range.start, mhp_range.end);
1290 #ifdef CONFIG_MEMORY_HOTREMOVE
1292 * Confirm all pages in a range [start, end) belong to the same zone (skipping
1293 * memory holes). When true, return the zone.
1295 struct zone *test_pages_in_a_zone(unsigned long start_pfn,
1296 unsigned long end_pfn)
1298 unsigned long pfn, sec_end_pfn;
1299 struct zone *zone = NULL;
1302 for (pfn = start_pfn, sec_end_pfn = SECTION_ALIGN_UP(start_pfn + 1);
1304 pfn = sec_end_pfn, sec_end_pfn += PAGES_PER_SECTION) {
1305 /* Make sure the memory section is present first */
1306 if (!present_section_nr(pfn_to_section_nr(pfn)))
1308 for (; pfn < sec_end_pfn && pfn < end_pfn;
1309 pfn += MAX_ORDER_NR_PAGES) {
1311 /* This is just a CONFIG_HOLES_IN_ZONE check.*/
1312 while ((i < MAX_ORDER_NR_PAGES) &&
1313 !pfn_valid_within(pfn + i))
1315 if (i == MAX_ORDER_NR_PAGES || pfn + i >= end_pfn)
1317 /* Check if we got outside of the zone */
1318 if (zone && !zone_spans_pfn(zone, pfn + i))
1320 page = pfn_to_page(pfn + i);
1321 if (zone && page_zone(page) != zone)
1323 zone = page_zone(page);
1331 * Scan pfn range [start,end) to find movable/migratable pages (LRU pages,
1332 * non-lru movable pages and hugepages). Will skip over most unmovable
1333 * pages (esp., pages that can be skipped when offlining), but bail out on
1334 * definitely unmovable pages.
1337 * 0 in case a movable page is found and movable_pfn was updated.
1338 * -ENOENT in case no movable page was found.
1339 * -EBUSY in case a definitely unmovable page was found.
1341 static int scan_movable_pages(unsigned long start, unsigned long end,
1342 unsigned long *movable_pfn)
1346 for (pfn = start; pfn < end; pfn++) {
1347 struct page *page, *head;
1350 if (!pfn_valid(pfn))
1352 page = pfn_to_page(pfn);
1355 if (__PageMovable(page))
1359 * PageOffline() pages that are not marked __PageMovable() and
1360 * have a reference count > 0 (after MEM_GOING_OFFLINE) are
1361 * definitely unmovable. If their reference count would be 0,
1362 * they could at least be skipped when offlining memory.
1364 if (PageOffline(page) && page_count(page))
1367 if (!PageHuge(page))
1369 head = compound_head(page);
1371 * This test is racy as we hold no reference or lock. The
1372 * hugetlb page could have been free'ed and head is no longer
1373 * a hugetlb page before the following check. In such unlikely
1374 * cases false positives and negatives are possible. Calling
1375 * code must deal with these scenarios.
1377 if (HPageMigratable(head))
1379 skip = compound_nr(head) - (page - head);
1389 do_migrate_range(unsigned long start_pfn, unsigned long end_pfn)
1392 struct page *page, *head;
1395 static DEFINE_RATELIMIT_STATE(migrate_rs, DEFAULT_RATELIMIT_INTERVAL,
1396 DEFAULT_RATELIMIT_BURST);
1398 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
1399 if (!pfn_valid(pfn))
1401 page = pfn_to_page(pfn);
1402 head = compound_head(page);
1404 if (PageHuge(page)) {
1405 pfn = page_to_pfn(head) + compound_nr(head) - 1;
1406 isolate_huge_page(head, &source);
1408 } else if (PageTransHuge(page))
1409 pfn = page_to_pfn(head) + thp_nr_pages(page) - 1;
1412 * HWPoison pages have elevated reference counts so the migration would
1413 * fail on them. It also doesn't make any sense to migrate them in the
1414 * first place. Still try to unmap such a page in case it is still mapped
1415 * (e.g. current hwpoison implementation doesn't unmap KSM pages but keep
1416 * the unmap as the catch all safety net).
1418 if (PageHWPoison(page)) {
1419 if (WARN_ON(PageLRU(page)))
1420 isolate_lru_page(page);
1421 if (page_mapped(page))
1422 try_to_unmap(page, TTU_IGNORE_MLOCK);
1426 if (!get_page_unless_zero(page))
1429 * We can skip free pages. And we can deal with pages on
1430 * LRU and non-lru movable pages.
1433 ret = isolate_lru_page(page);
1435 ret = isolate_movable_page(page, ISOLATE_UNEVICTABLE);
1436 if (!ret) { /* Success */
1437 list_add_tail(&page->lru, &source);
1438 if (!__PageMovable(page))
1439 inc_node_page_state(page, NR_ISOLATED_ANON +
1440 page_is_file_lru(page));
1443 if (__ratelimit(&migrate_rs)) {
1444 pr_warn("failed to isolate pfn %lx\n", pfn);
1445 dump_page(page, "isolation failed");
1450 if (!list_empty(&source)) {
1451 nodemask_t nmask = node_states[N_MEMORY];
1452 struct migration_target_control mtc = {
1454 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
1458 * We have checked that migration range is on a single zone so
1459 * we can use the nid of the first page to all the others.
1461 mtc.nid = page_to_nid(list_first_entry(&source, struct page, lru));
1464 * try to allocate from a different node but reuse this node
1465 * if there are no other online nodes to be used (e.g. we are
1466 * offlining a part of the only existing node)
1468 node_clear(mtc.nid, nmask);
1469 if (nodes_empty(nmask))
1470 node_set(mtc.nid, nmask);
1471 ret = migrate_pages(&source, alloc_migration_target, NULL,
1472 (unsigned long)&mtc, MIGRATE_SYNC, MR_MEMORY_HOTPLUG);
1474 list_for_each_entry(page, &source, lru) {
1475 if (__ratelimit(&migrate_rs)) {
1476 pr_warn("migrating pfn %lx failed ret:%d\n",
1477 page_to_pfn(page), ret);
1478 dump_page(page, "migration failure");
1481 putback_movable_pages(&source);
1488 static int __init cmdline_parse_movable_node(char *p)
1490 movable_node_enabled = true;
1493 early_param("movable_node", cmdline_parse_movable_node);
1495 /* check which state of node_states will be changed when offline memory */
1496 static void node_states_check_changes_offline(unsigned long nr_pages,
1497 struct zone *zone, struct memory_notify *arg)
1499 struct pglist_data *pgdat = zone->zone_pgdat;
1500 unsigned long present_pages = 0;
1503 arg->status_change_nid = NUMA_NO_NODE;
1504 arg->status_change_nid_normal = NUMA_NO_NODE;
1505 arg->status_change_nid_high = NUMA_NO_NODE;
1508 * Check whether node_states[N_NORMAL_MEMORY] will be changed.
1509 * If the memory to be offline is within the range
1510 * [0..ZONE_NORMAL], and it is the last present memory there,
1511 * the zones in that range will become empty after the offlining,
1512 * thus we can determine that we need to clear the node from
1513 * node_states[N_NORMAL_MEMORY].
1515 for (zt = 0; zt <= ZONE_NORMAL; zt++)
1516 present_pages += pgdat->node_zones[zt].present_pages;
1517 if (zone_idx(zone) <= ZONE_NORMAL && nr_pages >= present_pages)
1518 arg->status_change_nid_normal = zone_to_nid(zone);
1520 #ifdef CONFIG_HIGHMEM
1522 * node_states[N_HIGH_MEMORY] contains nodes which
1523 * have normal memory or high memory.
1524 * Here we add the present_pages belonging to ZONE_HIGHMEM.
1525 * If the zone is within the range of [0..ZONE_HIGHMEM), and
1526 * we determine that the zones in that range become empty,
1527 * we need to clear the node for N_HIGH_MEMORY.
1529 present_pages += pgdat->node_zones[ZONE_HIGHMEM].present_pages;
1530 if (zone_idx(zone) <= ZONE_HIGHMEM && nr_pages >= present_pages)
1531 arg->status_change_nid_high = zone_to_nid(zone);
1535 * We have accounted the pages from [0..ZONE_NORMAL), and
1536 * in case of CONFIG_HIGHMEM the pages from ZONE_HIGHMEM
1538 * Here we count the possible pages from ZONE_MOVABLE.
1539 * If after having accounted all the pages, we see that the nr_pages
1540 * to be offlined is over or equal to the accounted pages,
1541 * we know that the node will become empty, and so, we can clear
1542 * it for N_MEMORY as well.
1544 present_pages += pgdat->node_zones[ZONE_MOVABLE].present_pages;
1546 if (nr_pages >= present_pages)
1547 arg->status_change_nid = zone_to_nid(zone);
1550 static void node_states_clear_node(int node, struct memory_notify *arg)
1552 if (arg->status_change_nid_normal >= 0)
1553 node_clear_state(node, N_NORMAL_MEMORY);
1555 if (arg->status_change_nid_high >= 0)
1556 node_clear_state(node, N_HIGH_MEMORY);
1558 if (arg->status_change_nid >= 0)
1559 node_clear_state(node, N_MEMORY);
1562 static int count_system_ram_pages_cb(unsigned long start_pfn,
1563 unsigned long nr_pages, void *data)
1565 unsigned long *nr_system_ram_pages = data;
1567 *nr_system_ram_pages += nr_pages;
1571 int __ref offline_pages(unsigned long start_pfn, unsigned long nr_pages)
1573 const unsigned long end_pfn = start_pfn + nr_pages;
1574 unsigned long pfn, system_ram_pages = 0;
1575 unsigned long flags;
1577 struct memory_notify arg;
1582 * {on,off}lining is constrained to full memory sections (or more
1583 * precisely to memory blocks from the user space POV).
1584 * memmap_on_memory is an exception because it reserves initial part
1585 * of the physical memory space for vmemmaps. That space is pageblock
1588 if (WARN_ON_ONCE(!nr_pages ||
1589 !IS_ALIGNED(start_pfn, pageblock_nr_pages) ||
1590 !IS_ALIGNED(start_pfn + nr_pages, PAGES_PER_SECTION)))
1593 mem_hotplug_begin();
1596 * Don't allow to offline memory blocks that contain holes.
1597 * Consequently, memory blocks with holes can never get onlined
1598 * via the hotplug path - online_pages() - as hotplugged memory has
1599 * no holes. This way, we e.g., don't have to worry about marking
1600 * memory holes PG_reserved, don't need pfn_valid() checks, and can
1601 * avoid using walk_system_ram_range() later.
1603 walk_system_ram_range(start_pfn, nr_pages, &system_ram_pages,
1604 count_system_ram_pages_cb);
1605 if (system_ram_pages != nr_pages) {
1607 reason = "memory holes";
1608 goto failed_removal;
1611 /* This makes hotplug much easier...and readable.
1612 we assume this for now. .*/
1613 zone = test_pages_in_a_zone(start_pfn, end_pfn);
1616 reason = "multizone range";
1617 goto failed_removal;
1619 node = zone_to_nid(zone);
1622 * Disable pcplists so that page isolation cannot race with freeing
1623 * in a way that pages from isolated pageblock are left on pcplists.
1625 zone_pcp_disable(zone);
1626 lru_cache_disable();
1628 /* set above range as isolated */
1629 ret = start_isolate_page_range(start_pfn, end_pfn,
1631 MEMORY_OFFLINE | REPORT_FAILURE);
1633 reason = "failure to isolate range";
1634 goto failed_removal_pcplists_disabled;
1637 arg.start_pfn = start_pfn;
1638 arg.nr_pages = nr_pages;
1639 node_states_check_changes_offline(nr_pages, zone, &arg);
1641 ret = memory_notify(MEM_GOING_OFFLINE, &arg);
1642 ret = notifier_to_errno(ret);
1644 reason = "notifier failure";
1645 goto failed_removal_isolated;
1651 if (signal_pending(current)) {
1653 reason = "signal backoff";
1654 goto failed_removal_isolated;
1659 ret = scan_movable_pages(pfn, end_pfn, &pfn);
1662 * TODO: fatal migration failures should bail
1665 do_migrate_range(pfn, end_pfn);
1669 if (ret != -ENOENT) {
1670 reason = "unmovable page";
1671 goto failed_removal_isolated;
1675 * Dissolve free hugepages in the memory block before doing
1676 * offlining actually in order to make hugetlbfs's object
1677 * counting consistent.
1679 ret = dissolve_free_huge_pages(start_pfn, end_pfn);
1681 reason = "failure to dissolve huge pages";
1682 goto failed_removal_isolated;
1685 ret = test_pages_isolated(start_pfn, end_pfn, MEMORY_OFFLINE);
1689 /* Mark all sections offline and remove free pages from the buddy. */
1690 __offline_isolated_pages(start_pfn, end_pfn);
1691 pr_debug("Offlined Pages %ld\n", nr_pages);
1694 * The memory sections are marked offline, and the pageblock flags
1695 * effectively stale; nobody should be touching them. Fixup the number
1696 * of isolated pageblocks, memory onlining will properly revert this.
1698 spin_lock_irqsave(&zone->lock, flags);
1699 zone->nr_isolate_pageblock -= nr_pages / pageblock_nr_pages;
1700 spin_unlock_irqrestore(&zone->lock, flags);
1703 zone_pcp_enable(zone);
1705 /* removal success */
1706 adjust_managed_page_count(pfn_to_page(start_pfn), -nr_pages);
1707 adjust_present_page_count(zone, -nr_pages);
1709 /* reinitialise watermarks and update pcp limits */
1710 init_per_zone_wmark_min();
1712 if (!populated_zone(zone)) {
1713 zone_pcp_reset(zone);
1714 build_all_zonelists(NULL);
1717 node_states_clear_node(node, &arg);
1718 if (arg.status_change_nid >= 0) {
1720 kcompactd_stop(node);
1723 writeback_set_ratelimit();
1725 memory_notify(MEM_OFFLINE, &arg);
1726 remove_pfn_range_from_zone(zone, start_pfn, nr_pages);
1730 failed_removal_isolated:
1731 undo_isolate_page_range(start_pfn, end_pfn, MIGRATE_MOVABLE);
1732 memory_notify(MEM_CANCEL_OFFLINE, &arg);
1733 failed_removal_pcplists_disabled:
1735 zone_pcp_enable(zone);
1737 pr_debug("memory offlining [mem %#010llx-%#010llx] failed due to %s\n",
1738 (unsigned long long) start_pfn << PAGE_SHIFT,
1739 ((unsigned long long) end_pfn << PAGE_SHIFT) - 1,
1741 /* pushback to free area */
1746 static int check_memblock_offlined_cb(struct memory_block *mem, void *arg)
1748 int ret = !is_memblock_offlined(mem);
1750 if (unlikely(ret)) {
1751 phys_addr_t beginpa, endpa;
1753 beginpa = PFN_PHYS(section_nr_to_pfn(mem->start_section_nr));
1754 endpa = beginpa + memory_block_size_bytes() - 1;
1755 pr_warn("removing memory fails, because memory [%pa-%pa] is onlined\n",
1763 static int get_nr_vmemmap_pages_cb(struct memory_block *mem, void *arg)
1766 * If not set, continue with the next block.
1768 return mem->nr_vmemmap_pages;
1771 static int check_cpu_on_node(pg_data_t *pgdat)
1775 for_each_present_cpu(cpu) {
1776 if (cpu_to_node(cpu) == pgdat->node_id)
1778 * the cpu on this node isn't removed, and we can't
1779 * offline this node.
1787 static int check_no_memblock_for_node_cb(struct memory_block *mem, void *arg)
1789 int nid = *(int *)arg;
1792 * If a memory block belongs to multiple nodes, the stored nid is not
1793 * reliable. However, such blocks are always online (e.g., cannot get
1794 * offlined) and, therefore, are still spanned by the node.
1796 return mem->nid == nid ? -EEXIST : 0;
1803 * Offline a node if all memory sections and cpus of the node are removed.
1805 * NOTE: The caller must call lock_device_hotplug() to serialize hotplug
1806 * and online/offline operations before this call.
1808 void try_offline_node(int nid)
1810 pg_data_t *pgdat = NODE_DATA(nid);
1814 * If the node still spans pages (especially ZONE_DEVICE), don't
1815 * offline it. A node spans memory after move_pfn_range_to_zone(),
1816 * e.g., after the memory block was onlined.
1818 if (pgdat->node_spanned_pages)
1822 * Especially offline memory blocks might not be spanned by the
1823 * node. They will get spanned by the node once they get onlined.
1824 * However, they link to the node in sysfs and can get onlined later.
1826 rc = for_each_memory_block(&nid, check_no_memblock_for_node_cb);
1830 if (check_cpu_on_node(pgdat))
1834 * all memory/cpu of this node are removed, we can offline this
1837 node_set_offline(nid);
1838 unregister_one_node(nid);
1840 EXPORT_SYMBOL(try_offline_node);
1842 static int __ref try_remove_memory(int nid, u64 start, u64 size)
1845 struct vmem_altmap mhp_altmap = {};
1846 struct vmem_altmap *altmap = NULL;
1847 unsigned long nr_vmemmap_pages;
1849 BUG_ON(check_hotplug_memory_range(start, size));
1852 * All memory blocks must be offlined before removing memory. Check
1853 * whether all memory blocks in question are offline and return error
1854 * if this is not the case.
1856 rc = walk_memory_blocks(start, size, NULL, check_memblock_offlined_cb);
1861 * We only support removing memory added with MHP_MEMMAP_ON_MEMORY in
1862 * the same granularity it was added - a single memory block.
1864 if (memmap_on_memory) {
1865 nr_vmemmap_pages = walk_memory_blocks(start, size, NULL,
1866 get_nr_vmemmap_pages_cb);
1867 if (nr_vmemmap_pages) {
1868 if (size != memory_block_size_bytes()) {
1869 pr_warn("Refuse to remove %#llx - %#llx,"
1870 "wrong granularity\n",
1871 start, start + size);
1876 * Let remove_pmd_table->free_hugepage_table do the
1877 * right thing if we used vmem_altmap when hot-adding
1880 mhp_altmap.alloc = nr_vmemmap_pages;
1881 altmap = &mhp_altmap;
1885 /* remove memmap entry */
1886 firmware_map_remove(start, start + size, "System RAM");
1889 * Memory block device removal under the device_hotplug_lock is
1890 * a barrier against racing online attempts.
1892 remove_memory_block_devices(start, size);
1894 mem_hotplug_begin();
1896 arch_remove_memory(nid, start, size, altmap);
1898 if (IS_ENABLED(CONFIG_ARCH_KEEP_MEMBLOCK)) {
1899 memblock_free(start, size);
1900 memblock_remove(start, size);
1903 release_mem_region_adjustable(start, size);
1905 try_offline_node(nid);
1912 * __remove_memory - Remove memory if every memory block is offline
1914 * @start: physical address of the region to remove
1915 * @size: size of the region to remove
1917 * NOTE: The caller must call lock_device_hotplug() to serialize hotplug
1918 * and online/offline operations before this call, as required by
1919 * try_offline_node().
1921 void __remove_memory(int nid, u64 start, u64 size)
1925 * trigger BUG() if some memory is not offlined prior to calling this
1928 if (try_remove_memory(nid, start, size))
1933 * Remove memory if every memory block is offline, otherwise return -EBUSY is
1934 * some memory is not offline
1936 int remove_memory(int nid, u64 start, u64 size)
1940 lock_device_hotplug();
1941 rc = try_remove_memory(nid, start, size);
1942 unlock_device_hotplug();
1946 EXPORT_SYMBOL_GPL(remove_memory);
1948 static int try_offline_memory_block(struct memory_block *mem, void *arg)
1950 uint8_t online_type = MMOP_ONLINE_KERNEL;
1951 uint8_t **online_types = arg;
1956 * Sense the online_type via the zone of the memory block. Offlining
1957 * with multiple zones within one memory block will be rejected
1958 * by offlining code ... so we don't care about that.
1960 page = pfn_to_online_page(section_nr_to_pfn(mem->start_section_nr));
1961 if (page && zone_idx(page_zone(page)) == ZONE_MOVABLE)
1962 online_type = MMOP_ONLINE_MOVABLE;
1964 rc = device_offline(&mem->dev);
1966 * Default is MMOP_OFFLINE - change it only if offlining succeeded,
1967 * so try_reonline_memory_block() can do the right thing.
1970 **online_types = online_type;
1973 /* Ignore if already offline. */
1974 return rc < 0 ? rc : 0;
1977 static int try_reonline_memory_block(struct memory_block *mem, void *arg)
1979 uint8_t **online_types = arg;
1982 if (**online_types != MMOP_OFFLINE) {
1983 mem->online_type = **online_types;
1984 rc = device_online(&mem->dev);
1986 pr_warn("%s: Failed to re-online memory: %d",
1990 /* Continue processing all remaining memory blocks. */
1996 * Try to offline and remove memory. Might take a long time to finish in case
1997 * memory is still in use. Primarily useful for memory devices that logically
1998 * unplugged all memory (so it's no longer in use) and want to offline + remove
2001 int offline_and_remove_memory(int nid, u64 start, u64 size)
2003 const unsigned long mb_count = size / memory_block_size_bytes();
2004 uint8_t *online_types, *tmp;
2007 if (!IS_ALIGNED(start, memory_block_size_bytes()) ||
2008 !IS_ALIGNED(size, memory_block_size_bytes()) || !size)
2012 * We'll remember the old online type of each memory block, so we can
2013 * try to revert whatever we did when offlining one memory block fails
2014 * after offlining some others succeeded.
2016 online_types = kmalloc_array(mb_count, sizeof(*online_types),
2021 * Initialize all states to MMOP_OFFLINE, so when we abort processing in
2022 * try_offline_memory_block(), we'll skip all unprocessed blocks in
2023 * try_reonline_memory_block().
2025 memset(online_types, MMOP_OFFLINE, mb_count);
2027 lock_device_hotplug();
2030 rc = walk_memory_blocks(start, size, &tmp, try_offline_memory_block);
2033 * In case we succeeded to offline all memory, remove it.
2034 * This cannot fail as it cannot get onlined in the meantime.
2037 rc = try_remove_memory(nid, start, size);
2039 pr_err("%s: Failed to remove memory: %d", __func__, rc);
2043 * Rollback what we did. While memory onlining might theoretically fail
2044 * (nacked by a notifier), it barely ever happens.
2048 walk_memory_blocks(start, size, &tmp,
2049 try_reonline_memory_block);
2051 unlock_device_hotplug();
2053 kfree(online_types);
2056 EXPORT_SYMBOL_GPL(offline_and_remove_memory);
2057 #endif /* CONFIG_MEMORY_HOTREMOVE */