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/writeback.h>
17 #include <linux/slab.h>
18 #include <linux/sysctl.h>
19 #include <linux/cpu.h>
20 #include <linux/memory.h>
21 #include <linux/memremap.h>
22 #include <linux/memory_hotplug.h>
23 #include <linux/vmalloc.h>
24 #include <linux/ioport.h>
25 #include <linux/delay.h>
26 #include <linux/migrate.h>
27 #include <linux/page-isolation.h>
28 #include <linux/pfn.h>
29 #include <linux/suspend.h>
30 #include <linux/mm_inline.h>
31 #include <linux/firmware-map.h>
32 #include <linux/stop_machine.h>
33 #include <linux/hugetlb.h>
34 #include <linux/memblock.h>
35 #include <linux/compaction.h>
36 #include <linux/rmap.h>
37 #include <linux/module.h>
39 #include <asm/tlbflush.h>
45 MEMMAP_ON_MEMORY_DISABLE = 0,
46 MEMMAP_ON_MEMORY_ENABLE,
47 MEMMAP_ON_MEMORY_FORCE,
50 static int memmap_mode __read_mostly = MEMMAP_ON_MEMORY_DISABLE;
52 static inline unsigned long memory_block_memmap_size(void)
54 return PHYS_PFN(memory_block_size_bytes()) * sizeof(struct page);
57 static inline unsigned long memory_block_memmap_on_memory_pages(void)
59 unsigned long nr_pages = PFN_UP(memory_block_memmap_size());
62 * In "forced" memmap_on_memory mode, we add extra pages to align the
63 * vmemmap size to cover full pageblocks. That way, we can add memory
64 * even if the vmemmap size is not properly aligned, however, we might waste
67 if (memmap_mode == MEMMAP_ON_MEMORY_FORCE)
68 return pageblock_align(nr_pages);
72 #ifdef CONFIG_MHP_MEMMAP_ON_MEMORY
74 * memory_hotplug.memmap_on_memory parameter
76 static int set_memmap_mode(const char *val, const struct kernel_param *kp)
81 if (sysfs_streq(val, "force") || sysfs_streq(val, "FORCE")) {
82 mode = MEMMAP_ON_MEMORY_FORCE;
84 ret = kstrtobool(val, &enabled);
88 mode = MEMMAP_ON_MEMORY_ENABLE;
90 mode = MEMMAP_ON_MEMORY_DISABLE;
92 *((int *)kp->arg) = mode;
93 if (mode == MEMMAP_ON_MEMORY_FORCE) {
94 unsigned long memmap_pages = memory_block_memmap_on_memory_pages();
96 pr_info_once("Memory hotplug will waste %ld pages in each memory block\n",
97 memmap_pages - PFN_UP(memory_block_memmap_size()));
102 static int get_memmap_mode(char *buffer, const struct kernel_param *kp)
104 int mode = *((int *)kp->arg);
106 if (mode == MEMMAP_ON_MEMORY_FORCE)
107 return sprintf(buffer, "force\n");
108 return sprintf(buffer, "%c\n", mode ? 'Y' : 'N');
111 static const struct kernel_param_ops memmap_mode_ops = {
112 .set = set_memmap_mode,
113 .get = get_memmap_mode,
115 module_param_cb(memmap_on_memory, &memmap_mode_ops, &memmap_mode, 0444);
116 MODULE_PARM_DESC(memmap_on_memory, "Enable memmap on memory for memory hotplug\n"
117 "With value \"force\" it could result in memory wastage due "
118 "to memmap size limitations (Y/N/force)");
120 static inline bool mhp_memmap_on_memory(void)
122 return memmap_mode != MEMMAP_ON_MEMORY_DISABLE;
125 static inline bool mhp_memmap_on_memory(void)
132 ONLINE_POLICY_CONTIG_ZONES = 0,
133 ONLINE_POLICY_AUTO_MOVABLE,
136 static const char * const online_policy_to_str[] = {
137 [ONLINE_POLICY_CONTIG_ZONES] = "contig-zones",
138 [ONLINE_POLICY_AUTO_MOVABLE] = "auto-movable",
141 static int set_online_policy(const char *val, const struct kernel_param *kp)
143 int ret = sysfs_match_string(online_policy_to_str, val);
147 *((int *)kp->arg) = ret;
151 static int get_online_policy(char *buffer, const struct kernel_param *kp)
153 return sprintf(buffer, "%s\n", online_policy_to_str[*((int *)kp->arg)]);
157 * memory_hotplug.online_policy: configure online behavior when onlining without
158 * specifying a zone (MMOP_ONLINE)
160 * "contig-zones": keep zone contiguous
161 * "auto-movable": online memory to ZONE_MOVABLE if the configuration
162 * (auto_movable_ratio, auto_movable_numa_aware) allows for it
164 static int online_policy __read_mostly = ONLINE_POLICY_CONTIG_ZONES;
165 static const struct kernel_param_ops online_policy_ops = {
166 .set = set_online_policy,
167 .get = get_online_policy,
169 module_param_cb(online_policy, &online_policy_ops, &online_policy, 0644);
170 MODULE_PARM_DESC(online_policy,
171 "Set the online policy (\"contig-zones\", \"auto-movable\") "
172 "Default: \"contig-zones\"");
175 * memory_hotplug.auto_movable_ratio: specify maximum MOVABLE:KERNEL ratio
177 * The ratio represent an upper limit and the kernel might decide to not
178 * online some memory to ZONE_MOVABLE -- e.g., because hotplugged KERNEL memory
179 * doesn't allow for more MOVABLE memory.
181 static unsigned int auto_movable_ratio __read_mostly = 301;
182 module_param(auto_movable_ratio, uint, 0644);
183 MODULE_PARM_DESC(auto_movable_ratio,
184 "Set the maximum ratio of MOVABLE:KERNEL memory in the system "
185 "in percent for \"auto-movable\" online policy. Default: 301");
188 * memory_hotplug.auto_movable_numa_aware: consider numa node stats
191 static bool auto_movable_numa_aware __read_mostly = true;
192 module_param(auto_movable_numa_aware, bool, 0644);
193 MODULE_PARM_DESC(auto_movable_numa_aware,
194 "Consider numa node stats in addition to global stats in "
195 "\"auto-movable\" online policy. Default: true");
196 #endif /* CONFIG_NUMA */
199 * online_page_callback contains pointer to current page onlining function.
200 * Initially it is generic_online_page(). If it is required it could be
201 * changed by calling set_online_page_callback() for callback registration
202 * and restore_online_page_callback() for generic callback restore.
205 static online_page_callback_t online_page_callback = generic_online_page;
206 static DEFINE_MUTEX(online_page_callback_lock);
208 DEFINE_STATIC_PERCPU_RWSEM(mem_hotplug_lock);
210 void get_online_mems(void)
212 percpu_down_read(&mem_hotplug_lock);
215 void put_online_mems(void)
217 percpu_up_read(&mem_hotplug_lock);
220 bool movable_node_enabled = false;
222 #ifndef CONFIG_MEMORY_HOTPLUG_DEFAULT_ONLINE
223 int mhp_default_online_type = MMOP_OFFLINE;
225 int mhp_default_online_type = MMOP_ONLINE;
228 static int __init setup_memhp_default_state(char *str)
230 const int online_type = mhp_online_type_from_str(str);
232 if (online_type >= 0)
233 mhp_default_online_type = online_type;
237 __setup("memhp_default_state=", setup_memhp_default_state);
239 void mem_hotplug_begin(void)
242 percpu_down_write(&mem_hotplug_lock);
245 void mem_hotplug_done(void)
247 percpu_up_write(&mem_hotplug_lock);
251 u64 max_mem_size = U64_MAX;
253 /* add this memory to iomem resource */
254 static struct resource *register_memory_resource(u64 start, u64 size,
255 const char *resource_name)
257 struct resource *res;
258 unsigned long flags = IORESOURCE_SYSTEM_RAM | IORESOURCE_BUSY;
260 if (strcmp(resource_name, "System RAM"))
261 flags |= IORESOURCE_SYSRAM_DRIVER_MANAGED;
263 if (!mhp_range_allowed(start, size, true))
264 return ERR_PTR(-E2BIG);
267 * Make sure value parsed from 'mem=' only restricts memory adding
268 * while booting, so that memory hotplug won't be impacted. Please
269 * refer to document of 'mem=' in kernel-parameters.txt for more
272 if (start + size > max_mem_size && system_state < SYSTEM_RUNNING)
273 return ERR_PTR(-E2BIG);
276 * Request ownership of the new memory range. This might be
277 * a child of an existing resource that was present but
278 * not marked as busy.
280 res = __request_region(&iomem_resource, start, size,
281 resource_name, flags);
284 pr_debug("Unable to reserve System RAM region: %016llx->%016llx\n",
285 start, start + size);
286 return ERR_PTR(-EEXIST);
291 static void release_memory_resource(struct resource *res)
295 release_resource(res);
299 static int check_pfn_span(unsigned long pfn, unsigned long nr_pages)
302 * Disallow all operations smaller than a sub-section and only
303 * allow operations smaller than a section for
304 * SPARSEMEM_VMEMMAP. Note that check_hotplug_memory_range()
305 * enforces a larger memory_block_size_bytes() granularity for
306 * memory that will be marked online, so this check should only
307 * fire for direct arch_{add,remove}_memory() users outside of
308 * add_memory_resource().
310 unsigned long min_align;
312 if (IS_ENABLED(CONFIG_SPARSEMEM_VMEMMAP))
313 min_align = PAGES_PER_SUBSECTION;
315 min_align = PAGES_PER_SECTION;
316 if (!IS_ALIGNED(pfn | nr_pages, min_align))
322 * Return page for the valid pfn only if the page is online. All pfn
323 * walkers which rely on the fully initialized page->flags and others
324 * should use this rather than pfn_valid && pfn_to_page
326 struct page *pfn_to_online_page(unsigned long pfn)
328 unsigned long nr = pfn_to_section_nr(pfn);
329 struct dev_pagemap *pgmap;
330 struct mem_section *ms;
332 if (nr >= NR_MEM_SECTIONS)
335 ms = __nr_to_section(nr);
336 if (!online_section(ms))
340 * Save some code text when online_section() +
341 * pfn_section_valid() are sufficient.
343 if (IS_ENABLED(CONFIG_HAVE_ARCH_PFN_VALID) && !pfn_valid(pfn))
346 if (!pfn_section_valid(ms, pfn))
349 if (!online_device_section(ms))
350 return pfn_to_page(pfn);
353 * Slowpath: when ZONE_DEVICE collides with
354 * ZONE_{NORMAL,MOVABLE} within the same section some pfns in
355 * the section may be 'offline' but 'valid'. Only
356 * get_dev_pagemap() can determine sub-section online status.
358 pgmap = get_dev_pagemap(pfn, NULL);
359 put_dev_pagemap(pgmap);
361 /* The presence of a pgmap indicates ZONE_DEVICE offline pfn */
365 return pfn_to_page(pfn);
367 EXPORT_SYMBOL_GPL(pfn_to_online_page);
369 int __ref __add_pages(int nid, unsigned long pfn, unsigned long nr_pages,
370 struct mhp_params *params)
372 const unsigned long end_pfn = pfn + nr_pages;
373 unsigned long cur_nr_pages;
375 struct vmem_altmap *altmap = params->altmap;
377 if (WARN_ON_ONCE(!pgprot_val(params->pgprot)))
380 VM_BUG_ON(!mhp_range_allowed(PFN_PHYS(pfn), nr_pages * PAGE_SIZE, false));
384 * Validate altmap is within bounds of the total request
386 if (altmap->base_pfn != pfn
387 || vmem_altmap_offset(altmap) > nr_pages) {
388 pr_warn_once("memory add fail, invalid altmap\n");
394 if (check_pfn_span(pfn, nr_pages)) {
395 WARN(1, "Misaligned %s start: %#lx end: %#lx\n", __func__, pfn, pfn + nr_pages - 1);
399 for (; pfn < end_pfn; pfn += cur_nr_pages) {
400 /* Select all remaining pages up to the next section boundary */
401 cur_nr_pages = min(end_pfn - pfn,
402 SECTION_ALIGN_UP(pfn + 1) - pfn);
403 err = sparse_add_section(nid, pfn, cur_nr_pages, altmap,
409 vmemmap_populate_print_last();
413 /* find the smallest valid pfn in the range [start_pfn, end_pfn) */
414 static unsigned long find_smallest_section_pfn(int nid, struct zone *zone,
415 unsigned long start_pfn,
416 unsigned long end_pfn)
418 for (; start_pfn < end_pfn; start_pfn += PAGES_PER_SUBSECTION) {
419 if (unlikely(!pfn_to_online_page(start_pfn)))
422 if (unlikely(pfn_to_nid(start_pfn) != nid))
425 if (zone != page_zone(pfn_to_page(start_pfn)))
434 /* find the biggest valid pfn in the range [start_pfn, end_pfn). */
435 static unsigned long find_biggest_section_pfn(int nid, struct zone *zone,
436 unsigned long start_pfn,
437 unsigned long end_pfn)
441 /* pfn is the end pfn of a memory section. */
443 for (; pfn >= start_pfn; pfn -= PAGES_PER_SUBSECTION) {
444 if (unlikely(!pfn_to_online_page(pfn)))
447 if (unlikely(pfn_to_nid(pfn) != nid))
450 if (zone != page_zone(pfn_to_page(pfn)))
459 static void shrink_zone_span(struct zone *zone, unsigned long start_pfn,
460 unsigned long end_pfn)
463 int nid = zone_to_nid(zone);
465 if (zone->zone_start_pfn == start_pfn) {
467 * If the section is smallest section in the zone, it need
468 * shrink zone->zone_start_pfn and zone->zone_spanned_pages.
469 * In this case, we find second smallest valid mem_section
470 * for shrinking zone.
472 pfn = find_smallest_section_pfn(nid, zone, end_pfn,
475 zone->spanned_pages = zone_end_pfn(zone) - pfn;
476 zone->zone_start_pfn = pfn;
478 zone->zone_start_pfn = 0;
479 zone->spanned_pages = 0;
481 } else if (zone_end_pfn(zone) == end_pfn) {
483 * If the section is biggest section in the zone, it need
484 * shrink zone->spanned_pages.
485 * In this case, we find second biggest valid mem_section for
488 pfn = find_biggest_section_pfn(nid, zone, zone->zone_start_pfn,
491 zone->spanned_pages = pfn - zone->zone_start_pfn + 1;
493 zone->zone_start_pfn = 0;
494 zone->spanned_pages = 0;
499 static void update_pgdat_span(struct pglist_data *pgdat)
501 unsigned long node_start_pfn = 0, node_end_pfn = 0;
504 for (zone = pgdat->node_zones;
505 zone < pgdat->node_zones + MAX_NR_ZONES; zone++) {
506 unsigned long end_pfn = zone_end_pfn(zone);
508 /* No need to lock the zones, they can't change. */
509 if (!zone->spanned_pages)
512 node_start_pfn = zone->zone_start_pfn;
513 node_end_pfn = end_pfn;
517 if (end_pfn > node_end_pfn)
518 node_end_pfn = end_pfn;
519 if (zone->zone_start_pfn < node_start_pfn)
520 node_start_pfn = zone->zone_start_pfn;
523 pgdat->node_start_pfn = node_start_pfn;
524 pgdat->node_spanned_pages = node_end_pfn - node_start_pfn;
527 void __ref remove_pfn_range_from_zone(struct zone *zone,
528 unsigned long start_pfn,
529 unsigned long nr_pages)
531 const unsigned long end_pfn = start_pfn + nr_pages;
532 struct pglist_data *pgdat = zone->zone_pgdat;
533 unsigned long pfn, cur_nr_pages;
535 /* Poison struct pages because they are now uninitialized again. */
536 for (pfn = start_pfn; pfn < end_pfn; pfn += cur_nr_pages) {
539 /* Select all remaining pages up to the next section boundary */
541 min(end_pfn - pfn, SECTION_ALIGN_UP(pfn + 1) - pfn);
542 page_init_poison(pfn_to_page(pfn),
543 sizeof(struct page) * cur_nr_pages);
547 * Zone shrinking code cannot properly deal with ZONE_DEVICE. So
548 * we will not try to shrink the zones - which is okay as
549 * set_zone_contiguous() cannot deal with ZONE_DEVICE either way.
551 if (zone_is_zone_device(zone))
554 clear_zone_contiguous(zone);
556 shrink_zone_span(zone, start_pfn, start_pfn + nr_pages);
557 update_pgdat_span(pgdat);
559 set_zone_contiguous(zone);
563 * __remove_pages() - remove sections of pages
564 * @pfn: starting pageframe (must be aligned to start of a section)
565 * @nr_pages: number of pages to remove (must be multiple of section size)
566 * @altmap: alternative device page map or %NULL if default memmap is used
568 * Generic helper function to remove section mappings and sysfs entries
569 * for the section of the memory we are removing. Caller needs to make
570 * sure that pages are marked reserved and zones are adjust properly by
571 * calling offline_pages().
573 void __remove_pages(unsigned long pfn, unsigned long nr_pages,
574 struct vmem_altmap *altmap)
576 const unsigned long end_pfn = pfn + nr_pages;
577 unsigned long cur_nr_pages;
579 if (check_pfn_span(pfn, nr_pages)) {
580 WARN(1, "Misaligned %s start: %#lx end: %#lx\n", __func__, pfn, pfn + nr_pages - 1);
584 for (; pfn < end_pfn; pfn += cur_nr_pages) {
586 /* Select all remaining pages up to the next section boundary */
587 cur_nr_pages = min(end_pfn - pfn,
588 SECTION_ALIGN_UP(pfn + 1) - pfn);
589 sparse_remove_section(pfn, cur_nr_pages, altmap);
593 int set_online_page_callback(online_page_callback_t callback)
598 mutex_lock(&online_page_callback_lock);
600 if (online_page_callback == generic_online_page) {
601 online_page_callback = callback;
605 mutex_unlock(&online_page_callback_lock);
610 EXPORT_SYMBOL_GPL(set_online_page_callback);
612 int restore_online_page_callback(online_page_callback_t callback)
617 mutex_lock(&online_page_callback_lock);
619 if (online_page_callback == callback) {
620 online_page_callback = generic_online_page;
624 mutex_unlock(&online_page_callback_lock);
629 EXPORT_SYMBOL_GPL(restore_online_page_callback);
631 void generic_online_page(struct page *page, unsigned int order)
634 * Freeing the page with debug_pagealloc enabled will try to unmap it,
635 * so we should map it first. This is better than introducing a special
636 * case in page freeing fast path.
638 debug_pagealloc_map_pages(page, 1 << order);
639 __free_pages_core(page, order);
640 totalram_pages_add(1UL << order);
642 EXPORT_SYMBOL_GPL(generic_online_page);
644 static void online_pages_range(unsigned long start_pfn, unsigned long nr_pages)
646 const unsigned long end_pfn = start_pfn + nr_pages;
650 * Online the pages in MAX_PAGE_ORDER aligned chunks. The callback might
651 * decide to not expose all pages to the buddy (e.g., expose them
652 * later). We account all pages as being online and belonging to this
654 * When using memmap_on_memory, the range might not be aligned to
655 * MAX_ORDER_NR_PAGES - 1, but pageblock aligned. __ffs() will detect
656 * this and the first chunk to online will be pageblock_nr_pages.
658 for (pfn = start_pfn; pfn < end_pfn;) {
662 * Free to online pages in the largest chunks alignment allows.
664 * __ffs() behaviour is undefined for 0. start == 0 is
665 * MAX_PAGE_ORDER-aligned, Set order to MAX_PAGE_ORDER for
669 order = min_t(int, MAX_PAGE_ORDER, __ffs(pfn));
671 order = MAX_PAGE_ORDER;
673 (*online_page_callback)(pfn_to_page(pfn), order);
674 pfn += (1UL << order);
677 /* mark all involved sections as online */
678 online_mem_sections(start_pfn, end_pfn);
681 /* check which state of node_states will be changed when online memory */
682 static void node_states_check_changes_online(unsigned long nr_pages,
683 struct zone *zone, struct memory_notify *arg)
685 int nid = zone_to_nid(zone);
687 arg->status_change_nid = NUMA_NO_NODE;
688 arg->status_change_nid_normal = NUMA_NO_NODE;
690 if (!node_state(nid, N_MEMORY))
691 arg->status_change_nid = nid;
692 if (zone_idx(zone) <= ZONE_NORMAL && !node_state(nid, N_NORMAL_MEMORY))
693 arg->status_change_nid_normal = nid;
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 >= 0)
702 node_set_state(node, N_MEMORY);
705 static void __meminit resize_zone_range(struct zone *zone, unsigned long start_pfn,
706 unsigned long nr_pages)
708 unsigned long old_end_pfn = zone_end_pfn(zone);
710 if (zone_is_empty(zone) || start_pfn < zone->zone_start_pfn)
711 zone->zone_start_pfn = start_pfn;
713 zone->spanned_pages = max(start_pfn + nr_pages, old_end_pfn) - zone->zone_start_pfn;
716 static void __meminit resize_pgdat_range(struct pglist_data *pgdat, unsigned long start_pfn,
717 unsigned long nr_pages)
719 unsigned long old_end_pfn = pgdat_end_pfn(pgdat);
721 if (!pgdat->node_spanned_pages || start_pfn < pgdat->node_start_pfn)
722 pgdat->node_start_pfn = start_pfn;
724 pgdat->node_spanned_pages = max(start_pfn + nr_pages, old_end_pfn) - pgdat->node_start_pfn;
728 #ifdef CONFIG_ZONE_DEVICE
729 static void section_taint_zone_device(unsigned long pfn)
731 struct mem_section *ms = __pfn_to_section(pfn);
733 ms->section_mem_map |= SECTION_TAINT_ZONE_DEVICE;
736 static inline void section_taint_zone_device(unsigned long pfn)
742 * Associate the pfn range with the given zone, initializing the memmaps
743 * and resizing the pgdat/zone data to span the added pages. After this
744 * call, all affected pages are PG_reserved.
746 * All aligned pageblocks are initialized to the specified migratetype
747 * (usually MIGRATE_MOVABLE). Besides setting the migratetype, no related
748 * zone stats (e.g., nr_isolate_pageblock) are touched.
750 void __ref move_pfn_range_to_zone(struct zone *zone, unsigned long start_pfn,
751 unsigned long nr_pages,
752 struct vmem_altmap *altmap, int migratetype)
754 struct pglist_data *pgdat = zone->zone_pgdat;
755 int nid = pgdat->node_id;
757 clear_zone_contiguous(zone);
759 if (zone_is_empty(zone))
760 init_currently_empty_zone(zone, start_pfn, nr_pages);
761 resize_zone_range(zone, start_pfn, nr_pages);
762 resize_pgdat_range(pgdat, start_pfn, nr_pages);
765 * Subsection population requires care in pfn_to_online_page().
766 * Set the taint to enable the slow path detection of
767 * ZONE_DEVICE pages in an otherwise ZONE_{NORMAL,MOVABLE}
770 if (zone_is_zone_device(zone)) {
771 if (!IS_ALIGNED(start_pfn, PAGES_PER_SECTION))
772 section_taint_zone_device(start_pfn);
773 if (!IS_ALIGNED(start_pfn + nr_pages, PAGES_PER_SECTION))
774 section_taint_zone_device(start_pfn + nr_pages);
778 * TODO now we have a visible range of pages which are not associated
779 * with their zone properly. Not nice but set_pfnblock_flags_mask
780 * expects the zone spans the pfn range. All the pages in the range
781 * are reserved so nobody should be touching them so we should be safe
783 memmap_init_range(nr_pages, nid, zone_idx(zone), start_pfn, 0,
784 MEMINIT_HOTPLUG, altmap, migratetype);
786 set_zone_contiguous(zone);
789 struct auto_movable_stats {
790 unsigned long kernel_early_pages;
791 unsigned long movable_pages;
794 static void auto_movable_stats_account_zone(struct auto_movable_stats *stats,
797 if (zone_idx(zone) == ZONE_MOVABLE) {
798 stats->movable_pages += zone->present_pages;
800 stats->kernel_early_pages += zone->present_early_pages;
803 * CMA pages (never on hotplugged memory) behave like
806 stats->movable_pages += zone->cma_pages;
807 stats->kernel_early_pages -= zone->cma_pages;
808 #endif /* CONFIG_CMA */
811 struct auto_movable_group_stats {
812 unsigned long movable_pages;
813 unsigned long req_kernel_early_pages;
816 static int auto_movable_stats_account_group(struct memory_group *group,
819 const int ratio = READ_ONCE(auto_movable_ratio);
820 struct auto_movable_group_stats *stats = arg;
824 * We don't support modifying the config while the auto-movable online
825 * policy is already enabled. Just avoid the division by zero below.
831 * Calculate how many early kernel pages this group requires to
832 * satisfy the configured zone ratio.
834 pages = group->present_movable_pages * 100 / ratio;
835 pages -= group->present_kernel_pages;
838 stats->req_kernel_early_pages += pages;
839 stats->movable_pages += group->present_movable_pages;
843 static bool auto_movable_can_online_movable(int nid, struct memory_group *group,
844 unsigned long nr_pages)
846 unsigned long kernel_early_pages, movable_pages;
847 struct auto_movable_group_stats group_stats = {};
848 struct auto_movable_stats stats = {};
849 pg_data_t *pgdat = NODE_DATA(nid);
853 /* Walk all relevant zones and collect MOVABLE vs. KERNEL stats. */
854 if (nid == NUMA_NO_NODE) {
855 /* TODO: cache values */
856 for_each_populated_zone(zone)
857 auto_movable_stats_account_zone(&stats, zone);
859 for (i = 0; i < MAX_NR_ZONES; i++) {
860 zone = pgdat->node_zones + i;
861 if (populated_zone(zone))
862 auto_movable_stats_account_zone(&stats, zone);
866 kernel_early_pages = stats.kernel_early_pages;
867 movable_pages = stats.movable_pages;
870 * Kernel memory inside dynamic memory group allows for more MOVABLE
871 * memory within the same group. Remove the effect of all but the
872 * current group from the stats.
874 walk_dynamic_memory_groups(nid, auto_movable_stats_account_group,
875 group, &group_stats);
876 if (kernel_early_pages <= group_stats.req_kernel_early_pages)
878 kernel_early_pages -= group_stats.req_kernel_early_pages;
879 movable_pages -= group_stats.movable_pages;
881 if (group && group->is_dynamic)
882 kernel_early_pages += group->present_kernel_pages;
885 * Test if we could online the given number of pages to ZONE_MOVABLE
886 * and still stay in the configured ratio.
888 movable_pages += nr_pages;
889 return movable_pages <= (auto_movable_ratio * kernel_early_pages) / 100;
893 * Returns a default kernel memory zone for the given pfn range.
894 * If no kernel zone covers this pfn range it will automatically go
895 * to the ZONE_NORMAL.
897 static struct zone *default_kernel_zone_for_pfn(int nid, unsigned long start_pfn,
898 unsigned long nr_pages)
900 struct pglist_data *pgdat = NODE_DATA(nid);
903 for (zid = 0; zid < ZONE_NORMAL; zid++) {
904 struct zone *zone = &pgdat->node_zones[zid];
906 if (zone_intersects(zone, start_pfn, nr_pages))
910 return &pgdat->node_zones[ZONE_NORMAL];
914 * Determine to which zone to online memory dynamically based on user
915 * configuration and system stats. We care about the following ratio:
919 * Whereby MOVABLE is memory in ZONE_MOVABLE and KERNEL is memory in
920 * one of the kernel zones. CMA pages inside one of the kernel zones really
921 * behaves like ZONE_MOVABLE, so we treat them accordingly.
923 * We don't allow for hotplugged memory in a KERNEL zone to increase the
924 * amount of MOVABLE memory we can have, so we end up with:
926 * MOVABLE : KERNEL_EARLY
928 * Whereby KERNEL_EARLY is memory in one of the kernel zones, available sinze
929 * boot. We base our calculation on KERNEL_EARLY internally, because:
931 * a) Hotplugged memory in one of the kernel zones can sometimes still get
932 * hotunplugged, especially when hot(un)plugging individual memory blocks.
933 * There is no coordination across memory devices, therefore "automatic"
934 * hotunplugging, as implemented in hypervisors, could result in zone
936 * b) Early/boot memory in one of the kernel zones can usually not get
937 * hotunplugged again (e.g., no firmware interface to unplug, fragmented
938 * with unmovable allocations). While there are corner cases where it might
939 * still work, it is barely relevant in practice.
941 * Exceptions are dynamic memory groups, which allow for more MOVABLE
942 * memory within the same memory group -- because in that case, there is
943 * coordination within the single memory device managed by a single driver.
945 * We rely on "present pages" instead of "managed pages", as the latter is
946 * highly unreliable and dynamic in virtualized environments, and does not
947 * consider boot time allocations. For example, memory ballooning adjusts the
948 * managed pages when inflating/deflating the balloon, and balloon compaction
949 * can even migrate inflated pages between zones.
951 * Using "present pages" is better but some things to keep in mind are:
953 * a) Some memblock allocations, such as for the crashkernel area, are
954 * effectively unused by the kernel, yet they account to "present pages".
955 * Fortunately, these allocations are comparatively small in relevant setups
956 * (e.g., fraction of system memory).
957 * b) Some hotplugged memory blocks in virtualized environments, esecially
958 * hotplugged by virtio-mem, look like they are completely present, however,
959 * only parts of the memory block are actually currently usable.
960 * "present pages" is an upper limit that can get reached at runtime. As
961 * we base our calculations on KERNEL_EARLY, this is not an issue.
963 static struct zone *auto_movable_zone_for_pfn(int nid,
964 struct memory_group *group,
966 unsigned long nr_pages)
968 unsigned long online_pages = 0, max_pages, end_pfn;
971 if (!auto_movable_ratio)
974 if (group && !group->is_dynamic) {
975 max_pages = group->s.max_pages;
976 online_pages = group->present_movable_pages;
978 /* If anything is !MOVABLE online the rest !MOVABLE. */
979 if (group->present_kernel_pages)
981 } else if (!group || group->d.unit_pages == nr_pages) {
982 max_pages = nr_pages;
984 max_pages = group->d.unit_pages;
986 * Take a look at all online sections in the current unit.
987 * We can safely assume that all pages within a section belong
988 * to the same zone, because dynamic memory groups only deal
989 * with hotplugged memory.
991 pfn = ALIGN_DOWN(pfn, group->d.unit_pages);
992 end_pfn = pfn + group->d.unit_pages;
993 for (; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
994 page = pfn_to_online_page(pfn);
997 /* If anything is !MOVABLE online the rest !MOVABLE. */
998 if (!is_zone_movable_page(page))
1000 online_pages += PAGES_PER_SECTION;
1005 * Online MOVABLE if we could *currently* online all remaining parts
1006 * MOVABLE. We expect to (add+) online them immediately next, so if
1007 * nobody interferes, all will be MOVABLE if possible.
1009 nr_pages = max_pages - online_pages;
1010 if (!auto_movable_can_online_movable(NUMA_NO_NODE, group, nr_pages))
1014 if (auto_movable_numa_aware &&
1015 !auto_movable_can_online_movable(nid, group, nr_pages))
1017 #endif /* CONFIG_NUMA */
1019 return &NODE_DATA(nid)->node_zones[ZONE_MOVABLE];
1021 return default_kernel_zone_for_pfn(nid, pfn, nr_pages);
1024 static inline struct zone *default_zone_for_pfn(int nid, unsigned long start_pfn,
1025 unsigned long nr_pages)
1027 struct zone *kernel_zone = default_kernel_zone_for_pfn(nid, start_pfn,
1029 struct zone *movable_zone = &NODE_DATA(nid)->node_zones[ZONE_MOVABLE];
1030 bool in_kernel = zone_intersects(kernel_zone, start_pfn, nr_pages);
1031 bool in_movable = zone_intersects(movable_zone, start_pfn, nr_pages);
1034 * We inherit the existing zone in a simple case where zones do not
1035 * overlap in the given range
1037 if (in_kernel ^ in_movable)
1038 return (in_kernel) ? kernel_zone : movable_zone;
1041 * If the range doesn't belong to any zone or two zones overlap in the
1042 * given range then we use movable zone only if movable_node is
1043 * enabled because we always online to a kernel zone by default.
1045 return movable_node_enabled ? movable_zone : kernel_zone;
1048 struct zone *zone_for_pfn_range(int online_type, int nid,
1049 struct memory_group *group, unsigned long start_pfn,
1050 unsigned long nr_pages)
1052 if (online_type == MMOP_ONLINE_KERNEL)
1053 return default_kernel_zone_for_pfn(nid, start_pfn, nr_pages);
1055 if (online_type == MMOP_ONLINE_MOVABLE)
1056 return &NODE_DATA(nid)->node_zones[ZONE_MOVABLE];
1058 if (online_policy == ONLINE_POLICY_AUTO_MOVABLE)
1059 return auto_movable_zone_for_pfn(nid, group, start_pfn, nr_pages);
1061 return default_zone_for_pfn(nid, start_pfn, nr_pages);
1065 * This function should only be called by memory_block_{online,offline},
1066 * and {online,offline}_pages.
1068 void adjust_present_page_count(struct page *page, struct memory_group *group,
1071 struct zone *zone = page_zone(page);
1072 const bool movable = zone_idx(zone) == ZONE_MOVABLE;
1075 * We only support onlining/offlining/adding/removing of complete
1076 * memory blocks; therefore, either all is either early or hotplugged.
1078 if (early_section(__pfn_to_section(page_to_pfn(page))))
1079 zone->present_early_pages += nr_pages;
1080 zone->present_pages += nr_pages;
1081 zone->zone_pgdat->node_present_pages += nr_pages;
1083 if (group && movable)
1084 group->present_movable_pages += nr_pages;
1085 else if (group && !movable)
1086 group->present_kernel_pages += nr_pages;
1089 int mhp_init_memmap_on_memory(unsigned long pfn, unsigned long nr_pages,
1090 struct zone *zone, bool mhp_off_inaccessible)
1092 unsigned long end_pfn = pfn + nr_pages;
1095 ret = kasan_add_zero_shadow(__va(PFN_PHYS(pfn)), PFN_PHYS(nr_pages));
1100 * Memory block is accessible at this stage and hence poison the struct
1101 * pages now. If the memory block is accessible during memory hotplug
1102 * addition phase, then page poisining is already performed in
1103 * sparse_add_section().
1105 if (mhp_off_inaccessible)
1106 page_init_poison(pfn_to_page(pfn), sizeof(struct page) * nr_pages);
1108 move_pfn_range_to_zone(zone, pfn, nr_pages, NULL, MIGRATE_UNMOVABLE);
1110 for (i = 0; i < nr_pages; i++)
1111 SetPageVmemmapSelfHosted(pfn_to_page(pfn + i));
1114 * It might be that the vmemmap_pages fully span sections. If that is
1115 * the case, mark those sections online here as otherwise they will be
1118 if (nr_pages >= PAGES_PER_SECTION)
1119 online_mem_sections(pfn, ALIGN_DOWN(end_pfn, PAGES_PER_SECTION));
1124 void mhp_deinit_memmap_on_memory(unsigned long pfn, unsigned long nr_pages)
1126 unsigned long end_pfn = pfn + nr_pages;
1129 * It might be that the vmemmap_pages fully span sections. If that is
1130 * the case, mark those sections offline here as otherwise they will be
1133 if (nr_pages >= PAGES_PER_SECTION)
1134 offline_mem_sections(pfn, ALIGN_DOWN(end_pfn, PAGES_PER_SECTION));
1137 * The pages associated with this vmemmap have been offlined, so
1138 * we can reset its state here.
1140 remove_pfn_range_from_zone(page_zone(pfn_to_page(pfn)), pfn, nr_pages);
1141 kasan_remove_zero_shadow(__va(PFN_PHYS(pfn)), PFN_PHYS(nr_pages));
1145 * Must be called with mem_hotplug_lock in write mode.
1147 int __ref online_pages(unsigned long pfn, unsigned long nr_pages,
1148 struct zone *zone, struct memory_group *group)
1150 unsigned long flags;
1151 int need_zonelists_rebuild = 0;
1152 const int nid = zone_to_nid(zone);
1154 struct memory_notify arg;
1157 * {on,off}lining is constrained to full memory sections (or more
1158 * precisely to memory blocks from the user space POV).
1159 * memmap_on_memory is an exception because it reserves initial part
1160 * of the physical memory space for vmemmaps. That space is pageblock
1163 if (WARN_ON_ONCE(!nr_pages || !pageblock_aligned(pfn) ||
1164 !IS_ALIGNED(pfn + nr_pages, PAGES_PER_SECTION)))
1168 /* associate pfn range with the zone */
1169 move_pfn_range_to_zone(zone, pfn, nr_pages, NULL, MIGRATE_ISOLATE);
1171 arg.start_pfn = pfn;
1172 arg.nr_pages = nr_pages;
1173 node_states_check_changes_online(nr_pages, zone, &arg);
1175 ret = memory_notify(MEM_GOING_ONLINE, &arg);
1176 ret = notifier_to_errno(ret);
1178 goto failed_addition;
1181 * Fixup the number of isolated pageblocks before marking the sections
1182 * onlining, such that undo_isolate_page_range() works correctly.
1184 spin_lock_irqsave(&zone->lock, flags);
1185 zone->nr_isolate_pageblock += nr_pages / pageblock_nr_pages;
1186 spin_unlock_irqrestore(&zone->lock, flags);
1189 * If this zone is not populated, then it is not in zonelist.
1190 * This means the page allocator ignores this zone.
1191 * So, zonelist must be updated after online.
1193 if (!populated_zone(zone)) {
1194 need_zonelists_rebuild = 1;
1195 setup_zone_pageset(zone);
1198 online_pages_range(pfn, nr_pages);
1199 adjust_present_page_count(pfn_to_page(pfn), group, nr_pages);
1201 node_states_set_node(nid, &arg);
1202 if (need_zonelists_rebuild)
1203 build_all_zonelists(NULL);
1205 /* Basic onlining is complete, allow allocation of onlined pages. */
1206 undo_isolate_page_range(pfn, pfn + nr_pages, MIGRATE_MOVABLE);
1209 * Freshly onlined pages aren't shuffled (e.g., all pages are placed to
1210 * the tail of the freelist when undoing isolation). Shuffle the whole
1211 * zone to make sure the just onlined pages are properly distributed
1212 * across the whole freelist - to create an initial shuffle.
1216 /* reinitialise watermarks and update pcp limits */
1217 init_per_zone_wmark_min();
1222 writeback_set_ratelimit();
1224 memory_notify(MEM_ONLINE, &arg);
1228 pr_debug("online_pages [mem %#010llx-%#010llx] failed\n",
1229 (unsigned long long) pfn << PAGE_SHIFT,
1230 (((unsigned long long) pfn + nr_pages) << PAGE_SHIFT) - 1);
1231 memory_notify(MEM_CANCEL_ONLINE, &arg);
1232 remove_pfn_range_from_zone(zone, pfn, nr_pages);
1236 /* we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG */
1237 static pg_data_t __ref *hotadd_init_pgdat(int nid)
1239 struct pglist_data *pgdat;
1242 * NODE_DATA is preallocated (free_area_init) but its internal
1243 * state is not allocated completely. Add missing pieces.
1244 * Completely offline nodes stay around and they just need
1247 pgdat = NODE_DATA(nid);
1249 /* init node's zones as empty zones, we don't have any present pages.*/
1250 free_area_init_core_hotplug(pgdat);
1253 * The node we allocated has no zone fallback lists. For avoiding
1254 * to access not-initialized zonelist, build here.
1256 build_all_zonelists(pgdat);
1262 * __try_online_node - online a node if offlined
1264 * @set_node_online: Whether we want to online the node
1265 * called by cpu_up() to online a node without onlined memory.
1268 * 1 -> a new node has been allocated
1269 * 0 -> the node is already online
1270 * -ENOMEM -> the node could not be allocated
1272 static int __try_online_node(int nid, bool set_node_online)
1277 if (node_online(nid))
1280 pgdat = hotadd_init_pgdat(nid);
1282 pr_err("Cannot online node %d due to NULL pgdat\n", nid);
1287 if (set_node_online) {
1288 node_set_online(nid);
1289 ret = register_one_node(nid);
1297 * Users of this function always want to online/register the node
1299 int try_online_node(int nid)
1303 mem_hotplug_begin();
1304 ret = __try_online_node(nid, true);
1309 static int check_hotplug_memory_range(u64 start, u64 size)
1311 /* memory range must be block size aligned */
1312 if (!size || !IS_ALIGNED(start, memory_block_size_bytes()) ||
1313 !IS_ALIGNED(size, memory_block_size_bytes())) {
1314 pr_err("Block size [%#lx] unaligned hotplug range: start %#llx, size %#llx",
1315 memory_block_size_bytes(), start, size);
1322 static int online_memory_block(struct memory_block *mem, void *arg)
1324 mem->online_type = mhp_default_online_type;
1325 return device_online(&mem->dev);
1328 #ifndef arch_supports_memmap_on_memory
1329 static inline bool arch_supports_memmap_on_memory(unsigned long vmemmap_size)
1332 * As default, we want the vmemmap to span a complete PMD such that we
1333 * can map the vmemmap using a single PMD if supported by the
1336 return IS_ALIGNED(vmemmap_size, PMD_SIZE);
1340 bool mhp_supports_memmap_on_memory(void)
1342 unsigned long vmemmap_size = memory_block_memmap_size();
1343 unsigned long memmap_pages = memory_block_memmap_on_memory_pages();
1346 * Besides having arch support and the feature enabled at runtime, we
1347 * need a few more assumptions to hold true:
1349 * a) The vmemmap pages span complete PMDs: We don't want vmemmap code
1350 * to populate memory from the altmap for unrelated parts (i.e.,
1351 * other memory blocks)
1353 * b) The vmemmap pages (and thereby the pages that will be exposed to
1354 * the buddy) have to cover full pageblocks: memory onlining/offlining
1355 * code requires applicable ranges to be page-aligned, for example, to
1356 * set the migratetypes properly.
1358 * TODO: Although we have a check here to make sure that vmemmap pages
1359 * fully populate a PMD, it is not the right place to check for
1360 * this. A much better solution involves improving vmemmap code
1361 * to fallback to base pages when trying to populate vmemmap using
1362 * altmap as an alternative source of memory, and we do not exactly
1363 * populate a single PMD.
1365 if (!mhp_memmap_on_memory())
1369 * Make sure the vmemmap allocation is fully contained
1370 * so that we always allocate vmemmap memory from altmap area.
1372 if (!IS_ALIGNED(vmemmap_size, PAGE_SIZE))
1376 * start pfn should be pageblock_nr_pages aligned for correctly
1377 * setting migrate types
1379 if (!pageblock_aligned(memmap_pages))
1382 if (memmap_pages == PHYS_PFN(memory_block_size_bytes()))
1383 /* No effective hotplugged memory doesn't make sense. */
1386 return arch_supports_memmap_on_memory(vmemmap_size);
1388 EXPORT_SYMBOL_GPL(mhp_supports_memmap_on_memory);
1390 static void __ref remove_memory_blocks_and_altmaps(u64 start, u64 size)
1392 unsigned long memblock_size = memory_block_size_bytes();
1396 * For memmap_on_memory, the altmaps were added on a per-memblock
1397 * basis; we have to process each individual memory block.
1399 for (cur_start = start; cur_start < start + size;
1400 cur_start += memblock_size) {
1401 struct vmem_altmap *altmap = NULL;
1402 struct memory_block *mem;
1404 mem = find_memory_block(pfn_to_section_nr(PFN_DOWN(cur_start)));
1405 if (WARN_ON_ONCE(!mem))
1408 altmap = mem->altmap;
1411 remove_memory_block_devices(cur_start, memblock_size);
1413 arch_remove_memory(cur_start, memblock_size, altmap);
1415 /* Verify that all vmemmap pages have actually been freed. */
1416 WARN(altmap->alloc, "Altmap not fully unmapped");
1421 static int create_altmaps_and_memory_blocks(int nid, struct memory_group *group,
1422 u64 start, u64 size, mhp_t mhp_flags)
1424 unsigned long memblock_size = memory_block_size_bytes();
1428 for (cur_start = start; cur_start < start + size;
1429 cur_start += memblock_size) {
1430 struct mhp_params params = { .pgprot =
1431 pgprot_mhp(PAGE_KERNEL) };
1432 struct vmem_altmap mhp_altmap = {
1433 .base_pfn = PHYS_PFN(cur_start),
1434 .end_pfn = PHYS_PFN(cur_start + memblock_size - 1),
1437 mhp_altmap.free = memory_block_memmap_on_memory_pages();
1438 if (mhp_flags & MHP_OFFLINE_INACCESSIBLE)
1439 mhp_altmap.inaccessible = true;
1440 params.altmap = kmemdup(&mhp_altmap, sizeof(struct vmem_altmap),
1442 if (!params.altmap) {
1447 /* call arch's memory hotadd */
1448 ret = arch_add_memory(nid, cur_start, memblock_size, ¶ms);
1450 kfree(params.altmap);
1454 /* create memory block devices after memory was added */
1455 ret = create_memory_block_devices(cur_start, memblock_size,
1456 params.altmap, group);
1458 arch_remove_memory(cur_start, memblock_size, NULL);
1459 kfree(params.altmap);
1466 if (ret && cur_start != start)
1467 remove_memory_blocks_and_altmaps(start, cur_start - start);
1472 * NOTE: The caller must call lock_device_hotplug() to serialize hotplug
1473 * and online/offline operations (triggered e.g. by sysfs).
1475 * we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG
1477 int __ref add_memory_resource(int nid, struct resource *res, mhp_t mhp_flags)
1479 struct mhp_params params = { .pgprot = pgprot_mhp(PAGE_KERNEL) };
1480 enum memblock_flags memblock_flags = MEMBLOCK_NONE;
1481 struct memory_group *group = NULL;
1483 bool new_node = false;
1487 size = resource_size(res);
1489 ret = check_hotplug_memory_range(start, size);
1493 if (mhp_flags & MHP_NID_IS_MGID) {
1494 group = memory_group_find_by_id(nid);
1500 if (!node_possible(nid)) {
1501 WARN(1, "node %d was absent from the node_possible_map\n", nid);
1505 mem_hotplug_begin();
1507 if (IS_ENABLED(CONFIG_ARCH_KEEP_MEMBLOCK)) {
1508 if (res->flags & IORESOURCE_SYSRAM_DRIVER_MANAGED)
1509 memblock_flags = MEMBLOCK_DRIVER_MANAGED;
1510 ret = memblock_add_node(start, size, nid, memblock_flags);
1512 goto error_mem_hotplug_end;
1515 ret = __try_online_node(nid, false);
1521 * Self hosted memmap array
1523 if ((mhp_flags & MHP_MEMMAP_ON_MEMORY) &&
1524 mhp_supports_memmap_on_memory()) {
1525 ret = create_altmaps_and_memory_blocks(nid, group, start, size, mhp_flags);
1529 ret = arch_add_memory(nid, start, size, ¶ms);
1533 /* create memory block devices after memory was added */
1534 ret = create_memory_block_devices(start, size, NULL, group);
1536 arch_remove_memory(start, size, params.altmap);
1542 /* If sysfs file of new node can't be created, cpu on the node
1543 * can't be hot-added. There is no rollback way now.
1544 * So, check by BUG_ON() to catch it reluctantly..
1545 * We online node here. We can't roll back from here.
1547 node_set_online(nid);
1548 ret = __register_one_node(nid);
1552 register_memory_blocks_under_node(nid, PFN_DOWN(start),
1553 PFN_UP(start + size - 1),
1556 /* create new memmap entry */
1557 if (!strcmp(res->name, "System RAM"))
1558 firmware_map_add_hotplug(start, start + size, "System RAM");
1560 /* device_online() will take the lock when calling online_pages() */
1564 * In case we're allowed to merge the resource, flag it and trigger
1565 * merging now that adding succeeded.
1567 if (mhp_flags & MHP_MERGE_RESOURCE)
1568 merge_system_ram_resource(res);
1570 /* online pages if requested */
1571 if (mhp_default_online_type != MMOP_OFFLINE)
1572 walk_memory_blocks(start, size, NULL, online_memory_block);
1576 if (IS_ENABLED(CONFIG_ARCH_KEEP_MEMBLOCK))
1577 memblock_remove(start, size);
1578 error_mem_hotplug_end:
1583 /* requires device_hotplug_lock, see add_memory_resource() */
1584 int __ref __add_memory(int nid, u64 start, u64 size, mhp_t mhp_flags)
1586 struct resource *res;
1589 res = register_memory_resource(start, size, "System RAM");
1591 return PTR_ERR(res);
1593 ret = add_memory_resource(nid, res, mhp_flags);
1595 release_memory_resource(res);
1599 int add_memory(int nid, u64 start, u64 size, mhp_t mhp_flags)
1603 lock_device_hotplug();
1604 rc = __add_memory(nid, start, size, mhp_flags);
1605 unlock_device_hotplug();
1609 EXPORT_SYMBOL_GPL(add_memory);
1612 * Add special, driver-managed memory to the system as system RAM. Such
1613 * memory is not exposed via the raw firmware-provided memmap as system
1614 * RAM, instead, it is detected and added by a driver - during cold boot,
1615 * after a reboot, and after kexec.
1617 * Reasons why this memory should not be used for the initial memmap of a
1618 * kexec kernel or for placing kexec images:
1619 * - The booting kernel is in charge of determining how this memory will be
1620 * used (e.g., use persistent memory as system RAM)
1621 * - Coordination with a hypervisor is required before this memory
1622 * can be used (e.g., inaccessible parts).
1624 * For this memory, no entries in /sys/firmware/memmap ("raw firmware-provided
1625 * memory map") are created. Also, the created memory resource is flagged
1626 * with IORESOURCE_SYSRAM_DRIVER_MANAGED, so in-kernel users can special-case
1627 * this memory as well (esp., not place kexec images onto it).
1629 * The resource_name (visible via /proc/iomem) has to have the format
1630 * "System RAM ($DRIVER)".
1632 int add_memory_driver_managed(int nid, u64 start, u64 size,
1633 const char *resource_name, mhp_t mhp_flags)
1635 struct resource *res;
1638 if (!resource_name ||
1639 strstr(resource_name, "System RAM (") != resource_name ||
1640 resource_name[strlen(resource_name) - 1] != ')')
1643 lock_device_hotplug();
1645 res = register_memory_resource(start, size, resource_name);
1651 rc = add_memory_resource(nid, res, mhp_flags);
1653 release_memory_resource(res);
1656 unlock_device_hotplug();
1659 EXPORT_SYMBOL_GPL(add_memory_driver_managed);
1662 * Platforms should define arch_get_mappable_range() that provides
1663 * maximum possible addressable physical memory range for which the
1664 * linear mapping could be created. The platform returned address
1665 * range must adhere to these following semantics.
1667 * - range.start <= range.end
1668 * - Range includes both end points [range.start..range.end]
1670 * There is also a fallback definition provided here, allowing the
1671 * entire possible physical address range in case any platform does
1672 * not define arch_get_mappable_range().
1674 struct range __weak arch_get_mappable_range(void)
1676 struct range mhp_range = {
1683 struct range mhp_get_pluggable_range(bool need_mapping)
1685 const u64 max_phys = (1ULL << MAX_PHYSMEM_BITS) - 1;
1686 struct range mhp_range;
1689 mhp_range = arch_get_mappable_range();
1690 if (mhp_range.start > max_phys) {
1691 mhp_range.start = 0;
1694 mhp_range.end = min_t(u64, mhp_range.end, max_phys);
1696 mhp_range.start = 0;
1697 mhp_range.end = max_phys;
1701 EXPORT_SYMBOL_GPL(mhp_get_pluggable_range);
1703 bool mhp_range_allowed(u64 start, u64 size, bool need_mapping)
1705 struct range mhp_range = mhp_get_pluggable_range(need_mapping);
1706 u64 end = start + size;
1708 if (start < end && start >= mhp_range.start && (end - 1) <= mhp_range.end)
1711 pr_warn("Hotplug memory [%#llx-%#llx] exceeds maximum addressable range [%#llx-%#llx]\n",
1712 start, end, mhp_range.start, mhp_range.end);
1716 #ifdef CONFIG_MEMORY_HOTREMOVE
1718 * Scan pfn range [start,end) to find movable/migratable pages (LRU pages,
1719 * non-lru movable pages and hugepages). Will skip over most unmovable
1720 * pages (esp., pages that can be skipped when offlining), but bail out on
1721 * definitely unmovable pages.
1724 * 0 in case a movable page is found and movable_pfn was updated.
1725 * -ENOENT in case no movable page was found.
1726 * -EBUSY in case a definitely unmovable page was found.
1728 static int scan_movable_pages(unsigned long start, unsigned long end,
1729 unsigned long *movable_pfn)
1733 for (pfn = start; pfn < end; pfn++) {
1734 struct page *page, *head;
1737 if (!pfn_valid(pfn))
1739 page = pfn_to_page(pfn);
1742 if (__PageMovable(page))
1746 * PageOffline() pages that are not marked __PageMovable() and
1747 * have a reference count > 0 (after MEM_GOING_OFFLINE) are
1748 * definitely unmovable. If their reference count would be 0,
1749 * they could at least be skipped when offlining memory.
1751 if (PageOffline(page) && page_count(page))
1754 if (!PageHuge(page))
1756 head = compound_head(page);
1758 * This test is racy as we hold no reference or lock. The
1759 * hugetlb page could have been free'ed and head is no longer
1760 * a hugetlb page before the following check. In such unlikely
1761 * cases false positives and negatives are possible. Calling
1762 * code must deal with these scenarios.
1764 if (HPageMigratable(head))
1766 skip = compound_nr(head) - (pfn - page_to_pfn(head));
1775 static void do_migrate_range(unsigned long start_pfn, unsigned long end_pfn)
1778 struct page *page, *head;
1780 static DEFINE_RATELIMIT_STATE(migrate_rs, DEFAULT_RATELIMIT_INTERVAL,
1781 DEFAULT_RATELIMIT_BURST);
1783 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
1784 struct folio *folio;
1787 if (!pfn_valid(pfn))
1789 page = pfn_to_page(pfn);
1790 folio = page_folio(page);
1791 head = &folio->page;
1793 if (PageHuge(page)) {
1794 pfn = page_to_pfn(head) + compound_nr(head) - 1;
1795 isolate_hugetlb(folio, &source);
1797 } else if (PageTransHuge(page))
1798 pfn = page_to_pfn(head) + thp_nr_pages(page) - 1;
1801 * HWPoison pages have elevated reference counts so the migration would
1802 * fail on them. It also doesn't make any sense to migrate them in the
1803 * first place. Still try to unmap such a page in case it is still mapped
1804 * (e.g. current hwpoison implementation doesn't unmap KSM pages but keep
1805 * the unmap as the catch all safety net).
1807 if (PageHWPoison(page)) {
1808 if (WARN_ON(folio_test_lru(folio)))
1809 folio_isolate_lru(folio);
1810 if (folio_mapped(folio))
1811 try_to_unmap(folio, TTU_IGNORE_MLOCK);
1815 if (!get_page_unless_zero(page))
1818 * We can skip free pages. And we can deal with pages on
1819 * LRU and non-lru movable pages.
1822 isolated = isolate_lru_page(page);
1824 isolated = isolate_movable_page(page, ISOLATE_UNEVICTABLE);
1826 list_add_tail(&page->lru, &source);
1827 if (!__PageMovable(page))
1828 inc_node_page_state(page, NR_ISOLATED_ANON +
1829 page_is_file_lru(page));
1832 if (__ratelimit(&migrate_rs)) {
1833 pr_warn("failed to isolate pfn %lx\n", pfn);
1834 dump_page(page, "isolation failed");
1839 if (!list_empty(&source)) {
1840 nodemask_t nmask = node_states[N_MEMORY];
1841 struct migration_target_control mtc = {
1843 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
1848 * We have checked that migration range is on a single zone so
1849 * we can use the nid of the first page to all the others.
1851 mtc.nid = page_to_nid(list_first_entry(&source, struct page, lru));
1854 * try to allocate from a different node but reuse this node
1855 * if there are no other online nodes to be used (e.g. we are
1856 * offlining a part of the only existing node)
1858 node_clear(mtc.nid, nmask);
1859 if (nodes_empty(nmask))
1860 node_set(mtc.nid, nmask);
1861 ret = migrate_pages(&source, alloc_migration_target, NULL,
1862 (unsigned long)&mtc, MIGRATE_SYNC, MR_MEMORY_HOTPLUG, NULL);
1864 list_for_each_entry(page, &source, lru) {
1865 if (__ratelimit(&migrate_rs)) {
1866 pr_warn("migrating pfn %lx failed ret:%d\n",
1867 page_to_pfn(page), ret);
1868 dump_page(page, "migration failure");
1871 putback_movable_pages(&source);
1876 static int __init cmdline_parse_movable_node(char *p)
1878 movable_node_enabled = true;
1881 early_param("movable_node", cmdline_parse_movable_node);
1883 /* check which state of node_states will be changed when offline memory */
1884 static void node_states_check_changes_offline(unsigned long nr_pages,
1885 struct zone *zone, struct memory_notify *arg)
1887 struct pglist_data *pgdat = zone->zone_pgdat;
1888 unsigned long present_pages = 0;
1891 arg->status_change_nid = NUMA_NO_NODE;
1892 arg->status_change_nid_normal = NUMA_NO_NODE;
1895 * Check whether node_states[N_NORMAL_MEMORY] will be changed.
1896 * If the memory to be offline is within the range
1897 * [0..ZONE_NORMAL], and it is the last present memory there,
1898 * the zones in that range will become empty after the offlining,
1899 * thus we can determine that we need to clear the node from
1900 * node_states[N_NORMAL_MEMORY].
1902 for (zt = 0; zt <= ZONE_NORMAL; zt++)
1903 present_pages += pgdat->node_zones[zt].present_pages;
1904 if (zone_idx(zone) <= ZONE_NORMAL && nr_pages >= present_pages)
1905 arg->status_change_nid_normal = zone_to_nid(zone);
1908 * We have accounted the pages from [0..ZONE_NORMAL); ZONE_HIGHMEM
1909 * does not apply as we don't support 32bit.
1910 * Here we count the possible pages from ZONE_MOVABLE.
1911 * If after having accounted all the pages, we see that the nr_pages
1912 * to be offlined is over or equal to the accounted pages,
1913 * we know that the node will become empty, and so, we can clear
1914 * it for N_MEMORY as well.
1916 present_pages += pgdat->node_zones[ZONE_MOVABLE].present_pages;
1918 if (nr_pages >= present_pages)
1919 arg->status_change_nid = zone_to_nid(zone);
1922 static void node_states_clear_node(int node, struct memory_notify *arg)
1924 if (arg->status_change_nid_normal >= 0)
1925 node_clear_state(node, N_NORMAL_MEMORY);
1927 if (arg->status_change_nid >= 0)
1928 node_clear_state(node, N_MEMORY);
1931 static int count_system_ram_pages_cb(unsigned long start_pfn,
1932 unsigned long nr_pages, void *data)
1934 unsigned long *nr_system_ram_pages = data;
1936 *nr_system_ram_pages += nr_pages;
1941 * Must be called with mem_hotplug_lock in write mode.
1943 int __ref offline_pages(unsigned long start_pfn, unsigned long nr_pages,
1944 struct zone *zone, struct memory_group *group)
1946 const unsigned long end_pfn = start_pfn + nr_pages;
1947 unsigned long pfn, system_ram_pages = 0;
1948 const int node = zone_to_nid(zone);
1949 unsigned long flags;
1950 struct memory_notify arg;
1955 * {on,off}lining is constrained to full memory sections (or more
1956 * precisely to memory blocks from the user space POV).
1957 * memmap_on_memory is an exception because it reserves initial part
1958 * of the physical memory space for vmemmaps. That space is pageblock
1961 if (WARN_ON_ONCE(!nr_pages || !pageblock_aligned(start_pfn) ||
1962 !IS_ALIGNED(start_pfn + nr_pages, PAGES_PER_SECTION)))
1966 * Don't allow to offline memory blocks that contain holes.
1967 * Consequently, memory blocks with holes can never get onlined
1968 * via the hotplug path - online_pages() - as hotplugged memory has
1969 * no holes. This way, we e.g., don't have to worry about marking
1970 * memory holes PG_reserved, don't need pfn_valid() checks, and can
1971 * avoid using walk_system_ram_range() later.
1973 walk_system_ram_range(start_pfn, nr_pages, &system_ram_pages,
1974 count_system_ram_pages_cb);
1975 if (system_ram_pages != nr_pages) {
1977 reason = "memory holes";
1978 goto failed_removal;
1982 * We only support offlining of memory blocks managed by a single zone,
1983 * checked by calling code. This is just a sanity check that we might
1984 * want to remove in the future.
1986 if (WARN_ON_ONCE(page_zone(pfn_to_page(start_pfn)) != zone ||
1987 page_zone(pfn_to_page(end_pfn - 1)) != zone)) {
1989 reason = "multizone range";
1990 goto failed_removal;
1994 * Disable pcplists so that page isolation cannot race with freeing
1995 * in a way that pages from isolated pageblock are left on pcplists.
1997 zone_pcp_disable(zone);
1998 lru_cache_disable();
2000 /* set above range as isolated */
2001 ret = start_isolate_page_range(start_pfn, end_pfn,
2003 MEMORY_OFFLINE | REPORT_FAILURE,
2004 GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL);
2006 reason = "failure to isolate range";
2007 goto failed_removal_pcplists_disabled;
2010 arg.start_pfn = start_pfn;
2011 arg.nr_pages = nr_pages;
2012 node_states_check_changes_offline(nr_pages, zone, &arg);
2014 ret = memory_notify(MEM_GOING_OFFLINE, &arg);
2015 ret = notifier_to_errno(ret);
2017 reason = "notifier failure";
2018 goto failed_removal_isolated;
2025 * Historically we always checked for any signal and
2026 * can't limit it to fatal signals without eventually
2027 * breaking user space.
2029 if (signal_pending(current)) {
2031 reason = "signal backoff";
2032 goto failed_removal_isolated;
2037 ret = scan_movable_pages(pfn, end_pfn, &pfn);
2040 * TODO: fatal migration failures should bail
2043 do_migrate_range(pfn, end_pfn);
2047 if (ret != -ENOENT) {
2048 reason = "unmovable page";
2049 goto failed_removal_isolated;
2053 * Dissolve free hugepages in the memory block before doing
2054 * offlining actually in order to make hugetlbfs's object
2055 * counting consistent.
2057 ret = dissolve_free_huge_pages(start_pfn, end_pfn);
2059 reason = "failure to dissolve huge pages";
2060 goto failed_removal_isolated;
2063 ret = test_pages_isolated(start_pfn, end_pfn, MEMORY_OFFLINE);
2067 /* Mark all sections offline and remove free pages from the buddy. */
2068 __offline_isolated_pages(start_pfn, end_pfn);
2069 pr_debug("Offlined Pages %ld\n", nr_pages);
2072 * The memory sections are marked offline, and the pageblock flags
2073 * effectively stale; nobody should be touching them. Fixup the number
2074 * of isolated pageblocks, memory onlining will properly revert this.
2076 spin_lock_irqsave(&zone->lock, flags);
2077 zone->nr_isolate_pageblock -= nr_pages / pageblock_nr_pages;
2078 spin_unlock_irqrestore(&zone->lock, flags);
2081 zone_pcp_enable(zone);
2083 /* removal success */
2084 adjust_managed_page_count(pfn_to_page(start_pfn), -nr_pages);
2085 adjust_present_page_count(pfn_to_page(start_pfn), group, -nr_pages);
2087 /* reinitialise watermarks and update pcp limits */
2088 init_per_zone_wmark_min();
2091 * Make sure to mark the node as memory-less before rebuilding the zone
2092 * list. Otherwise this node would still appear in the fallback lists.
2094 node_states_clear_node(node, &arg);
2095 if (!populated_zone(zone)) {
2096 zone_pcp_reset(zone);
2097 build_all_zonelists(NULL);
2100 if (arg.status_change_nid >= 0) {
2101 kcompactd_stop(node);
2105 writeback_set_ratelimit();
2107 memory_notify(MEM_OFFLINE, &arg);
2108 remove_pfn_range_from_zone(zone, start_pfn, nr_pages);
2111 failed_removal_isolated:
2112 /* pushback to free area */
2113 undo_isolate_page_range(start_pfn, end_pfn, MIGRATE_MOVABLE);
2114 memory_notify(MEM_CANCEL_OFFLINE, &arg);
2115 failed_removal_pcplists_disabled:
2117 zone_pcp_enable(zone);
2119 pr_debug("memory offlining [mem %#010llx-%#010llx] failed due to %s\n",
2120 (unsigned long long) start_pfn << PAGE_SHIFT,
2121 ((unsigned long long) end_pfn << PAGE_SHIFT) - 1,
2126 static int check_memblock_offlined_cb(struct memory_block *mem, void *arg)
2131 if (unlikely(mem->state != MEM_OFFLINE)) {
2132 phys_addr_t beginpa, endpa;
2134 beginpa = PFN_PHYS(section_nr_to_pfn(mem->start_section_nr));
2135 endpa = beginpa + memory_block_size_bytes() - 1;
2136 pr_warn("removing memory fails, because memory [%pa-%pa] is onlined\n",
2144 static int count_memory_range_altmaps_cb(struct memory_block *mem, void *arg)
2146 u64 *num_altmaps = (u64 *)arg;
2154 static int check_cpu_on_node(int nid)
2158 for_each_present_cpu(cpu) {
2159 if (cpu_to_node(cpu) == nid)
2161 * the cpu on this node isn't removed, and we can't
2162 * offline this node.
2170 static int check_no_memblock_for_node_cb(struct memory_block *mem, void *arg)
2172 int nid = *(int *)arg;
2175 * If a memory block belongs to multiple nodes, the stored nid is not
2176 * reliable. However, such blocks are always online (e.g., cannot get
2177 * offlined) and, therefore, are still spanned by the node.
2179 return mem->nid == nid ? -EEXIST : 0;
2186 * Offline a node if all memory sections and cpus of the node are removed.
2188 * NOTE: The caller must call lock_device_hotplug() to serialize hotplug
2189 * and online/offline operations before this call.
2191 void try_offline_node(int nid)
2196 * If the node still spans pages (especially ZONE_DEVICE), don't
2197 * offline it. A node spans memory after move_pfn_range_to_zone(),
2198 * e.g., after the memory block was onlined.
2200 if (node_spanned_pages(nid))
2204 * Especially offline memory blocks might not be spanned by the
2205 * node. They will get spanned by the node once they get onlined.
2206 * However, they link to the node in sysfs and can get onlined later.
2208 rc = for_each_memory_block(&nid, check_no_memblock_for_node_cb);
2212 if (check_cpu_on_node(nid))
2216 * all memory/cpu of this node are removed, we can offline this
2219 node_set_offline(nid);
2220 unregister_one_node(nid);
2222 EXPORT_SYMBOL(try_offline_node);
2224 static int memory_blocks_have_altmaps(u64 start, u64 size)
2226 u64 num_memblocks = size / memory_block_size_bytes();
2227 u64 num_altmaps = 0;
2229 if (!mhp_memmap_on_memory())
2232 walk_memory_blocks(start, size, &num_altmaps,
2233 count_memory_range_altmaps_cb);
2235 if (num_altmaps == 0)
2238 if (WARN_ON_ONCE(num_memblocks != num_altmaps))
2244 static int __ref try_remove_memory(u64 start, u64 size)
2246 int rc, nid = NUMA_NO_NODE;
2248 BUG_ON(check_hotplug_memory_range(start, size));
2251 * All memory blocks must be offlined before removing memory. Check
2252 * whether all memory blocks in question are offline and return error
2253 * if this is not the case.
2255 * While at it, determine the nid. Note that if we'd have mixed nodes,
2256 * we'd only try to offline the last determined one -- which is good
2257 * enough for the cases we care about.
2259 rc = walk_memory_blocks(start, size, &nid, check_memblock_offlined_cb);
2263 /* remove memmap entry */
2264 firmware_map_remove(start, start + size, "System RAM");
2266 mem_hotplug_begin();
2268 rc = memory_blocks_have_altmaps(start, size);
2274 * Memory block device removal under the device_hotplug_lock is
2275 * a barrier against racing online attempts.
2276 * No altmaps present, do the removal directly
2278 remove_memory_block_devices(start, size);
2279 arch_remove_memory(start, size, NULL);
2281 /* all memblocks in the range have altmaps */
2282 remove_memory_blocks_and_altmaps(start, size);
2285 if (IS_ENABLED(CONFIG_ARCH_KEEP_MEMBLOCK)) {
2286 memblock_phys_free(start, size);
2287 memblock_remove(start, size);
2290 release_mem_region_adjustable(start, size);
2292 if (nid != NUMA_NO_NODE)
2293 try_offline_node(nid);
2300 * __remove_memory - Remove memory if every memory block is offline
2301 * @start: physical address of the region to remove
2302 * @size: size of the region to remove
2304 * NOTE: The caller must call lock_device_hotplug() to serialize hotplug
2305 * and online/offline operations before this call, as required by
2306 * try_offline_node().
2308 void __remove_memory(u64 start, u64 size)
2312 * trigger BUG() if some memory is not offlined prior to calling this
2315 if (try_remove_memory(start, size))
2320 * Remove memory if every memory block is offline, otherwise return -EBUSY is
2321 * some memory is not offline
2323 int remove_memory(u64 start, u64 size)
2327 lock_device_hotplug();
2328 rc = try_remove_memory(start, size);
2329 unlock_device_hotplug();
2333 EXPORT_SYMBOL_GPL(remove_memory);
2335 static int try_offline_memory_block(struct memory_block *mem, void *arg)
2337 uint8_t online_type = MMOP_ONLINE_KERNEL;
2338 uint8_t **online_types = arg;
2343 * Sense the online_type via the zone of the memory block. Offlining
2344 * with multiple zones within one memory block will be rejected
2345 * by offlining code ... so we don't care about that.
2347 page = pfn_to_online_page(section_nr_to_pfn(mem->start_section_nr));
2348 if (page && zone_idx(page_zone(page)) == ZONE_MOVABLE)
2349 online_type = MMOP_ONLINE_MOVABLE;
2351 rc = device_offline(&mem->dev);
2353 * Default is MMOP_OFFLINE - change it only if offlining succeeded,
2354 * so try_reonline_memory_block() can do the right thing.
2357 **online_types = online_type;
2360 /* Ignore if already offline. */
2361 return rc < 0 ? rc : 0;
2364 static int try_reonline_memory_block(struct memory_block *mem, void *arg)
2366 uint8_t **online_types = arg;
2369 if (**online_types != MMOP_OFFLINE) {
2370 mem->online_type = **online_types;
2371 rc = device_online(&mem->dev);
2373 pr_warn("%s: Failed to re-online memory: %d",
2377 /* Continue processing all remaining memory blocks. */
2383 * Try to offline and remove memory. Might take a long time to finish in case
2384 * memory is still in use. Primarily useful for memory devices that logically
2385 * unplugged all memory (so it's no longer in use) and want to offline + remove
2388 int offline_and_remove_memory(u64 start, u64 size)
2390 const unsigned long mb_count = size / memory_block_size_bytes();
2391 uint8_t *online_types, *tmp;
2394 if (!IS_ALIGNED(start, memory_block_size_bytes()) ||
2395 !IS_ALIGNED(size, memory_block_size_bytes()) || !size)
2399 * We'll remember the old online type of each memory block, so we can
2400 * try to revert whatever we did when offlining one memory block fails
2401 * after offlining some others succeeded.
2403 online_types = kmalloc_array(mb_count, sizeof(*online_types),
2408 * Initialize all states to MMOP_OFFLINE, so when we abort processing in
2409 * try_offline_memory_block(), we'll skip all unprocessed blocks in
2410 * try_reonline_memory_block().
2412 memset(online_types, MMOP_OFFLINE, mb_count);
2414 lock_device_hotplug();
2417 rc = walk_memory_blocks(start, size, &tmp, try_offline_memory_block);
2420 * In case we succeeded to offline all memory, remove it.
2421 * This cannot fail as it cannot get onlined in the meantime.
2424 rc = try_remove_memory(start, size);
2426 pr_err("%s: Failed to remove memory: %d", __func__, rc);
2430 * Rollback what we did. While memory onlining might theoretically fail
2431 * (nacked by a notifier), it barely ever happens.
2435 walk_memory_blocks(start, size, &tmp,
2436 try_reonline_memory_block);
2438 unlock_device_hotplug();
2440 kfree(online_types);
2443 EXPORT_SYMBOL_GPL(offline_and_remove_memory);
2444 #endif /* CONFIG_MEMORY_HOTREMOVE */