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>
44 #ifdef CONFIG_MHP_MEMMAP_ON_MEMORY
46 * memory_hotplug.memmap_on_memory parameter
48 static bool memmap_on_memory __ro_after_init;
49 module_param(memmap_on_memory, bool, 0444);
50 MODULE_PARM_DESC(memmap_on_memory, "Enable memmap on memory for memory hotplug");
52 static inline bool mhp_memmap_on_memory(void)
54 return memmap_on_memory;
57 static inline bool mhp_memmap_on_memory(void)
64 ONLINE_POLICY_CONTIG_ZONES = 0,
65 ONLINE_POLICY_AUTO_MOVABLE,
68 static const char * const online_policy_to_str[] = {
69 [ONLINE_POLICY_CONTIG_ZONES] = "contig-zones",
70 [ONLINE_POLICY_AUTO_MOVABLE] = "auto-movable",
73 static int set_online_policy(const char *val, const struct kernel_param *kp)
75 int ret = sysfs_match_string(online_policy_to_str, val);
79 *((int *)kp->arg) = ret;
83 static int get_online_policy(char *buffer, const struct kernel_param *kp)
85 return sprintf(buffer, "%s\n", online_policy_to_str[*((int *)kp->arg)]);
89 * memory_hotplug.online_policy: configure online behavior when onlining without
90 * specifying a zone (MMOP_ONLINE)
92 * "contig-zones": keep zone contiguous
93 * "auto-movable": online memory to ZONE_MOVABLE if the configuration
94 * (auto_movable_ratio, auto_movable_numa_aware) allows for it
96 static int online_policy __read_mostly = ONLINE_POLICY_CONTIG_ZONES;
97 static const struct kernel_param_ops online_policy_ops = {
98 .set = set_online_policy,
99 .get = get_online_policy,
101 module_param_cb(online_policy, &online_policy_ops, &online_policy, 0644);
102 MODULE_PARM_DESC(online_policy,
103 "Set the online policy (\"contig-zones\", \"auto-movable\") "
104 "Default: \"contig-zones\"");
107 * memory_hotplug.auto_movable_ratio: specify maximum MOVABLE:KERNEL ratio
109 * The ratio represent an upper limit and the kernel might decide to not
110 * online some memory to ZONE_MOVABLE -- e.g., because hotplugged KERNEL memory
111 * doesn't allow for more MOVABLE memory.
113 static unsigned int auto_movable_ratio __read_mostly = 301;
114 module_param(auto_movable_ratio, uint, 0644);
115 MODULE_PARM_DESC(auto_movable_ratio,
116 "Set the maximum ratio of MOVABLE:KERNEL memory in the system "
117 "in percent for \"auto-movable\" online policy. Default: 301");
120 * memory_hotplug.auto_movable_numa_aware: consider numa node stats
123 static bool auto_movable_numa_aware __read_mostly = true;
124 module_param(auto_movable_numa_aware, bool, 0644);
125 MODULE_PARM_DESC(auto_movable_numa_aware,
126 "Consider numa node stats in addition to global stats in "
127 "\"auto-movable\" online policy. Default: true");
128 #endif /* CONFIG_NUMA */
131 * online_page_callback contains pointer to current page onlining function.
132 * Initially it is generic_online_page(). If it is required it could be
133 * changed by calling set_online_page_callback() for callback registration
134 * and restore_online_page_callback() for generic callback restore.
137 static online_page_callback_t online_page_callback = generic_online_page;
138 static DEFINE_MUTEX(online_page_callback_lock);
140 DEFINE_STATIC_PERCPU_RWSEM(mem_hotplug_lock);
142 void get_online_mems(void)
144 percpu_down_read(&mem_hotplug_lock);
147 void put_online_mems(void)
149 percpu_up_read(&mem_hotplug_lock);
152 bool movable_node_enabled = false;
154 #ifndef CONFIG_MEMORY_HOTPLUG_DEFAULT_ONLINE
155 int mhp_default_online_type = MMOP_OFFLINE;
157 int mhp_default_online_type = MMOP_ONLINE;
160 static int __init setup_memhp_default_state(char *str)
162 const int online_type = mhp_online_type_from_str(str);
164 if (online_type >= 0)
165 mhp_default_online_type = online_type;
169 __setup("memhp_default_state=", setup_memhp_default_state);
171 void mem_hotplug_begin(void)
174 percpu_down_write(&mem_hotplug_lock);
177 void mem_hotplug_done(void)
179 percpu_up_write(&mem_hotplug_lock);
183 u64 max_mem_size = U64_MAX;
185 /* add this memory to iomem resource */
186 static struct resource *register_memory_resource(u64 start, u64 size,
187 const char *resource_name)
189 struct resource *res;
190 unsigned long flags = IORESOURCE_SYSTEM_RAM | IORESOURCE_BUSY;
192 if (strcmp(resource_name, "System RAM"))
193 flags |= IORESOURCE_SYSRAM_DRIVER_MANAGED;
195 if (!mhp_range_allowed(start, size, true))
196 return ERR_PTR(-E2BIG);
199 * Make sure value parsed from 'mem=' only restricts memory adding
200 * while booting, so that memory hotplug won't be impacted. Please
201 * refer to document of 'mem=' in kernel-parameters.txt for more
204 if (start + size > max_mem_size && system_state < SYSTEM_RUNNING)
205 return ERR_PTR(-E2BIG);
208 * Request ownership of the new memory range. This might be
209 * a child of an existing resource that was present but
210 * not marked as busy.
212 res = __request_region(&iomem_resource, start, size,
213 resource_name, flags);
216 pr_debug("Unable to reserve System RAM region: %016llx->%016llx\n",
217 start, start + size);
218 return ERR_PTR(-EEXIST);
223 static void release_memory_resource(struct resource *res)
227 release_resource(res);
231 static int check_pfn_span(unsigned long pfn, unsigned long nr_pages)
234 * Disallow all operations smaller than a sub-section and only
235 * allow operations smaller than a section for
236 * SPARSEMEM_VMEMMAP. Note that check_hotplug_memory_range()
237 * enforces a larger memory_block_size_bytes() granularity for
238 * memory that will be marked online, so this check should only
239 * fire for direct arch_{add,remove}_memory() users outside of
240 * add_memory_resource().
242 unsigned long min_align;
244 if (IS_ENABLED(CONFIG_SPARSEMEM_VMEMMAP))
245 min_align = PAGES_PER_SUBSECTION;
247 min_align = PAGES_PER_SECTION;
248 if (!IS_ALIGNED(pfn | nr_pages, min_align))
254 * Return page for the valid pfn only if the page is online. All pfn
255 * walkers which rely on the fully initialized page->flags and others
256 * should use this rather than pfn_valid && pfn_to_page
258 struct page *pfn_to_online_page(unsigned long pfn)
260 unsigned long nr = pfn_to_section_nr(pfn);
261 struct dev_pagemap *pgmap;
262 struct mem_section *ms;
264 if (nr >= NR_MEM_SECTIONS)
267 ms = __nr_to_section(nr);
268 if (!online_section(ms))
272 * Save some code text when online_section() +
273 * pfn_section_valid() are sufficient.
275 if (IS_ENABLED(CONFIG_HAVE_ARCH_PFN_VALID) && !pfn_valid(pfn))
278 if (!pfn_section_valid(ms, pfn))
281 if (!online_device_section(ms))
282 return pfn_to_page(pfn);
285 * Slowpath: when ZONE_DEVICE collides with
286 * ZONE_{NORMAL,MOVABLE} within the same section some pfns in
287 * the section may be 'offline' but 'valid'. Only
288 * get_dev_pagemap() can determine sub-section online status.
290 pgmap = get_dev_pagemap(pfn, NULL);
291 put_dev_pagemap(pgmap);
293 /* The presence of a pgmap indicates ZONE_DEVICE offline pfn */
297 return pfn_to_page(pfn);
299 EXPORT_SYMBOL_GPL(pfn_to_online_page);
301 int __ref __add_pages(int nid, unsigned long pfn, unsigned long nr_pages,
302 struct mhp_params *params)
304 const unsigned long end_pfn = pfn + nr_pages;
305 unsigned long cur_nr_pages;
307 struct vmem_altmap *altmap = params->altmap;
309 if (WARN_ON_ONCE(!pgprot_val(params->pgprot)))
312 VM_BUG_ON(!mhp_range_allowed(PFN_PHYS(pfn), nr_pages * PAGE_SIZE, false));
316 * Validate altmap is within bounds of the total request
318 if (altmap->base_pfn != pfn
319 || vmem_altmap_offset(altmap) > nr_pages) {
320 pr_warn_once("memory add fail, invalid altmap\n");
326 if (check_pfn_span(pfn, nr_pages)) {
327 WARN(1, "Misaligned %s start: %#lx end: %#lx\n", __func__, pfn, pfn + nr_pages - 1);
331 for (; pfn < end_pfn; pfn += cur_nr_pages) {
332 /* Select all remaining pages up to the next section boundary */
333 cur_nr_pages = min(end_pfn - pfn,
334 SECTION_ALIGN_UP(pfn + 1) - pfn);
335 err = sparse_add_section(nid, pfn, cur_nr_pages, altmap,
341 vmemmap_populate_print_last();
345 /* find the smallest valid pfn in the range [start_pfn, end_pfn) */
346 static unsigned long find_smallest_section_pfn(int nid, struct zone *zone,
347 unsigned long start_pfn,
348 unsigned long end_pfn)
350 for (; start_pfn < end_pfn; start_pfn += PAGES_PER_SUBSECTION) {
351 if (unlikely(!pfn_to_online_page(start_pfn)))
354 if (unlikely(pfn_to_nid(start_pfn) != nid))
357 if (zone != page_zone(pfn_to_page(start_pfn)))
366 /* find the biggest valid pfn in the range [start_pfn, end_pfn). */
367 static unsigned long find_biggest_section_pfn(int nid, struct zone *zone,
368 unsigned long start_pfn,
369 unsigned long end_pfn)
373 /* pfn is the end pfn of a memory section. */
375 for (; pfn >= start_pfn; pfn -= PAGES_PER_SUBSECTION) {
376 if (unlikely(!pfn_to_online_page(pfn)))
379 if (unlikely(pfn_to_nid(pfn) != nid))
382 if (zone != page_zone(pfn_to_page(pfn)))
391 static void shrink_zone_span(struct zone *zone, unsigned long start_pfn,
392 unsigned long end_pfn)
395 int nid = zone_to_nid(zone);
397 if (zone->zone_start_pfn == start_pfn) {
399 * If the section is smallest section in the zone, it need
400 * shrink zone->zone_start_pfn and zone->zone_spanned_pages.
401 * In this case, we find second smallest valid mem_section
402 * for shrinking zone.
404 pfn = find_smallest_section_pfn(nid, zone, end_pfn,
407 zone->spanned_pages = zone_end_pfn(zone) - pfn;
408 zone->zone_start_pfn = pfn;
410 zone->zone_start_pfn = 0;
411 zone->spanned_pages = 0;
413 } else if (zone_end_pfn(zone) == end_pfn) {
415 * If the section is biggest section in the zone, it need
416 * shrink zone->spanned_pages.
417 * In this case, we find second biggest valid mem_section for
420 pfn = find_biggest_section_pfn(nid, zone, zone->zone_start_pfn,
423 zone->spanned_pages = pfn - zone->zone_start_pfn + 1;
425 zone->zone_start_pfn = 0;
426 zone->spanned_pages = 0;
431 static void update_pgdat_span(struct pglist_data *pgdat)
433 unsigned long node_start_pfn = 0, node_end_pfn = 0;
436 for (zone = pgdat->node_zones;
437 zone < pgdat->node_zones + MAX_NR_ZONES; zone++) {
438 unsigned long end_pfn = zone_end_pfn(zone);
440 /* No need to lock the zones, they can't change. */
441 if (!zone->spanned_pages)
444 node_start_pfn = zone->zone_start_pfn;
445 node_end_pfn = end_pfn;
449 if (end_pfn > node_end_pfn)
450 node_end_pfn = end_pfn;
451 if (zone->zone_start_pfn < node_start_pfn)
452 node_start_pfn = zone->zone_start_pfn;
455 pgdat->node_start_pfn = node_start_pfn;
456 pgdat->node_spanned_pages = node_end_pfn - node_start_pfn;
459 void __ref remove_pfn_range_from_zone(struct zone *zone,
460 unsigned long start_pfn,
461 unsigned long nr_pages)
463 const unsigned long end_pfn = start_pfn + nr_pages;
464 struct pglist_data *pgdat = zone->zone_pgdat;
465 unsigned long pfn, cur_nr_pages;
467 /* Poison struct pages because they are now uninitialized again. */
468 for (pfn = start_pfn; pfn < end_pfn; pfn += cur_nr_pages) {
471 /* Select all remaining pages up to the next section boundary */
473 min(end_pfn - pfn, SECTION_ALIGN_UP(pfn + 1) - pfn);
474 page_init_poison(pfn_to_page(pfn),
475 sizeof(struct page) * cur_nr_pages);
479 * Zone shrinking code cannot properly deal with ZONE_DEVICE. So
480 * we will not try to shrink the zones - which is okay as
481 * set_zone_contiguous() cannot deal with ZONE_DEVICE either way.
483 if (zone_is_zone_device(zone))
486 clear_zone_contiguous(zone);
488 shrink_zone_span(zone, start_pfn, start_pfn + nr_pages);
489 update_pgdat_span(pgdat);
491 set_zone_contiguous(zone);
495 * __remove_pages() - remove sections of pages
496 * @pfn: starting pageframe (must be aligned to start of a section)
497 * @nr_pages: number of pages to remove (must be multiple of section size)
498 * @altmap: alternative device page map or %NULL if default memmap is used
500 * Generic helper function to remove section mappings and sysfs entries
501 * for the section of the memory we are removing. Caller needs to make
502 * sure that pages are marked reserved and zones are adjust properly by
503 * calling offline_pages().
505 void __remove_pages(unsigned long pfn, unsigned long nr_pages,
506 struct vmem_altmap *altmap)
508 const unsigned long end_pfn = pfn + nr_pages;
509 unsigned long cur_nr_pages;
511 if (check_pfn_span(pfn, nr_pages)) {
512 WARN(1, "Misaligned %s start: %#lx end: %#lx\n", __func__, pfn, pfn + nr_pages - 1);
516 for (; pfn < end_pfn; pfn += cur_nr_pages) {
518 /* Select all remaining pages up to the next section boundary */
519 cur_nr_pages = min(end_pfn - pfn,
520 SECTION_ALIGN_UP(pfn + 1) - pfn);
521 sparse_remove_section(pfn, cur_nr_pages, altmap);
525 int set_online_page_callback(online_page_callback_t callback)
530 mutex_lock(&online_page_callback_lock);
532 if (online_page_callback == generic_online_page) {
533 online_page_callback = callback;
537 mutex_unlock(&online_page_callback_lock);
542 EXPORT_SYMBOL_GPL(set_online_page_callback);
544 int restore_online_page_callback(online_page_callback_t callback)
549 mutex_lock(&online_page_callback_lock);
551 if (online_page_callback == callback) {
552 online_page_callback = generic_online_page;
556 mutex_unlock(&online_page_callback_lock);
561 EXPORT_SYMBOL_GPL(restore_online_page_callback);
563 void generic_online_page(struct page *page, unsigned int order)
566 * Freeing the page with debug_pagealloc enabled will try to unmap it,
567 * so we should map it first. This is better than introducing a special
568 * case in page freeing fast path.
570 debug_pagealloc_map_pages(page, 1 << order);
571 __free_pages_core(page, order);
572 totalram_pages_add(1UL << order);
574 EXPORT_SYMBOL_GPL(generic_online_page);
576 static void online_pages_range(unsigned long start_pfn, unsigned long nr_pages)
578 const unsigned long end_pfn = start_pfn + nr_pages;
582 * Online the pages in MAX_ORDER aligned chunks. The callback might
583 * decide to not expose all pages to the buddy (e.g., expose them
584 * later). We account all pages as being online and belonging to this
586 * When using memmap_on_memory, the range might not be aligned to
587 * MAX_ORDER_NR_PAGES - 1, but pageblock aligned. __ffs() will detect
588 * this and the first chunk to online will be pageblock_nr_pages.
590 for (pfn = start_pfn; pfn < end_pfn;) {
594 * Free to online pages in the largest chunks alignment allows.
596 * __ffs() behaviour is undefined for 0. start == 0 is
597 * MAX_ORDER-aligned, Set order to MAX_ORDER for the case.
600 order = min_t(int, MAX_ORDER, __ffs(pfn));
604 (*online_page_callback)(pfn_to_page(pfn), order);
605 pfn += (1UL << order);
608 /* mark all involved sections as online */
609 online_mem_sections(start_pfn, end_pfn);
612 /* check which state of node_states will be changed when online memory */
613 static void node_states_check_changes_online(unsigned long nr_pages,
614 struct zone *zone, struct memory_notify *arg)
616 int nid = zone_to_nid(zone);
618 arg->status_change_nid = NUMA_NO_NODE;
619 arg->status_change_nid_normal = NUMA_NO_NODE;
621 if (!node_state(nid, N_MEMORY))
622 arg->status_change_nid = nid;
623 if (zone_idx(zone) <= ZONE_NORMAL && !node_state(nid, N_NORMAL_MEMORY))
624 arg->status_change_nid_normal = nid;
627 static void node_states_set_node(int node, struct memory_notify *arg)
629 if (arg->status_change_nid_normal >= 0)
630 node_set_state(node, N_NORMAL_MEMORY);
632 if (arg->status_change_nid >= 0)
633 node_set_state(node, N_MEMORY);
636 static void __meminit resize_zone_range(struct zone *zone, unsigned long start_pfn,
637 unsigned long nr_pages)
639 unsigned long old_end_pfn = zone_end_pfn(zone);
641 if (zone_is_empty(zone) || start_pfn < zone->zone_start_pfn)
642 zone->zone_start_pfn = start_pfn;
644 zone->spanned_pages = max(start_pfn + nr_pages, old_end_pfn) - zone->zone_start_pfn;
647 static void __meminit resize_pgdat_range(struct pglist_data *pgdat, unsigned long start_pfn,
648 unsigned long nr_pages)
650 unsigned long old_end_pfn = pgdat_end_pfn(pgdat);
652 if (!pgdat->node_spanned_pages || start_pfn < pgdat->node_start_pfn)
653 pgdat->node_start_pfn = start_pfn;
655 pgdat->node_spanned_pages = max(start_pfn + nr_pages, old_end_pfn) - pgdat->node_start_pfn;
659 #ifdef CONFIG_ZONE_DEVICE
660 static void section_taint_zone_device(unsigned long pfn)
662 struct mem_section *ms = __pfn_to_section(pfn);
664 ms->section_mem_map |= SECTION_TAINT_ZONE_DEVICE;
667 static inline void section_taint_zone_device(unsigned long pfn)
673 * Associate the pfn range with the given zone, initializing the memmaps
674 * and resizing the pgdat/zone data to span the added pages. After this
675 * call, all affected pages are PG_reserved.
677 * All aligned pageblocks are initialized to the specified migratetype
678 * (usually MIGRATE_MOVABLE). Besides setting the migratetype, no related
679 * zone stats (e.g., nr_isolate_pageblock) are touched.
681 void __ref move_pfn_range_to_zone(struct zone *zone, unsigned long start_pfn,
682 unsigned long nr_pages,
683 struct vmem_altmap *altmap, int migratetype)
685 struct pglist_data *pgdat = zone->zone_pgdat;
686 int nid = pgdat->node_id;
688 clear_zone_contiguous(zone);
690 if (zone_is_empty(zone))
691 init_currently_empty_zone(zone, start_pfn, nr_pages);
692 resize_zone_range(zone, start_pfn, nr_pages);
693 resize_pgdat_range(pgdat, start_pfn, nr_pages);
696 * Subsection population requires care in pfn_to_online_page().
697 * Set the taint to enable the slow path detection of
698 * ZONE_DEVICE pages in an otherwise ZONE_{NORMAL,MOVABLE}
701 if (zone_is_zone_device(zone)) {
702 if (!IS_ALIGNED(start_pfn, PAGES_PER_SECTION))
703 section_taint_zone_device(start_pfn);
704 if (!IS_ALIGNED(start_pfn + nr_pages, PAGES_PER_SECTION))
705 section_taint_zone_device(start_pfn + nr_pages);
709 * TODO now we have a visible range of pages which are not associated
710 * with their zone properly. Not nice but set_pfnblock_flags_mask
711 * expects the zone spans the pfn range. All the pages in the range
712 * are reserved so nobody should be touching them so we should be safe
714 memmap_init_range(nr_pages, nid, zone_idx(zone), start_pfn, 0,
715 MEMINIT_HOTPLUG, altmap, migratetype);
717 set_zone_contiguous(zone);
720 struct auto_movable_stats {
721 unsigned long kernel_early_pages;
722 unsigned long movable_pages;
725 static void auto_movable_stats_account_zone(struct auto_movable_stats *stats,
728 if (zone_idx(zone) == ZONE_MOVABLE) {
729 stats->movable_pages += zone->present_pages;
731 stats->kernel_early_pages += zone->present_early_pages;
734 * CMA pages (never on hotplugged memory) behave like
737 stats->movable_pages += zone->cma_pages;
738 stats->kernel_early_pages -= zone->cma_pages;
739 #endif /* CONFIG_CMA */
742 struct auto_movable_group_stats {
743 unsigned long movable_pages;
744 unsigned long req_kernel_early_pages;
747 static int auto_movable_stats_account_group(struct memory_group *group,
750 const int ratio = READ_ONCE(auto_movable_ratio);
751 struct auto_movable_group_stats *stats = arg;
755 * We don't support modifying the config while the auto-movable online
756 * policy is already enabled. Just avoid the division by zero below.
762 * Calculate how many early kernel pages this group requires to
763 * satisfy the configured zone ratio.
765 pages = group->present_movable_pages * 100 / ratio;
766 pages -= group->present_kernel_pages;
769 stats->req_kernel_early_pages += pages;
770 stats->movable_pages += group->present_movable_pages;
774 static bool auto_movable_can_online_movable(int nid, struct memory_group *group,
775 unsigned long nr_pages)
777 unsigned long kernel_early_pages, movable_pages;
778 struct auto_movable_group_stats group_stats = {};
779 struct auto_movable_stats stats = {};
780 pg_data_t *pgdat = NODE_DATA(nid);
784 /* Walk all relevant zones and collect MOVABLE vs. KERNEL stats. */
785 if (nid == NUMA_NO_NODE) {
786 /* TODO: cache values */
787 for_each_populated_zone(zone)
788 auto_movable_stats_account_zone(&stats, zone);
790 for (i = 0; i < MAX_NR_ZONES; i++) {
791 zone = pgdat->node_zones + i;
792 if (populated_zone(zone))
793 auto_movable_stats_account_zone(&stats, zone);
797 kernel_early_pages = stats.kernel_early_pages;
798 movable_pages = stats.movable_pages;
801 * Kernel memory inside dynamic memory group allows for more MOVABLE
802 * memory within the same group. Remove the effect of all but the
803 * current group from the stats.
805 walk_dynamic_memory_groups(nid, auto_movable_stats_account_group,
806 group, &group_stats);
807 if (kernel_early_pages <= group_stats.req_kernel_early_pages)
809 kernel_early_pages -= group_stats.req_kernel_early_pages;
810 movable_pages -= group_stats.movable_pages;
812 if (group && group->is_dynamic)
813 kernel_early_pages += group->present_kernel_pages;
816 * Test if we could online the given number of pages to ZONE_MOVABLE
817 * and still stay in the configured ratio.
819 movable_pages += nr_pages;
820 return movable_pages <= (auto_movable_ratio * kernel_early_pages) / 100;
824 * Returns a default kernel memory zone for the given pfn range.
825 * If no kernel zone covers this pfn range it will automatically go
826 * to the ZONE_NORMAL.
828 static struct zone *default_kernel_zone_for_pfn(int nid, unsigned long start_pfn,
829 unsigned long nr_pages)
831 struct pglist_data *pgdat = NODE_DATA(nid);
834 for (zid = 0; zid < ZONE_NORMAL; zid++) {
835 struct zone *zone = &pgdat->node_zones[zid];
837 if (zone_intersects(zone, start_pfn, nr_pages))
841 return &pgdat->node_zones[ZONE_NORMAL];
845 * Determine to which zone to online memory dynamically based on user
846 * configuration and system stats. We care about the following ratio:
850 * Whereby MOVABLE is memory in ZONE_MOVABLE and KERNEL is memory in
851 * one of the kernel zones. CMA pages inside one of the kernel zones really
852 * behaves like ZONE_MOVABLE, so we treat them accordingly.
854 * We don't allow for hotplugged memory in a KERNEL zone to increase the
855 * amount of MOVABLE memory we can have, so we end up with:
857 * MOVABLE : KERNEL_EARLY
859 * Whereby KERNEL_EARLY is memory in one of the kernel zones, available sinze
860 * boot. We base our calculation on KERNEL_EARLY internally, because:
862 * a) Hotplugged memory in one of the kernel zones can sometimes still get
863 * hotunplugged, especially when hot(un)plugging individual memory blocks.
864 * There is no coordination across memory devices, therefore "automatic"
865 * hotunplugging, as implemented in hypervisors, could result in zone
867 * b) Early/boot memory in one of the kernel zones can usually not get
868 * hotunplugged again (e.g., no firmware interface to unplug, fragmented
869 * with unmovable allocations). While there are corner cases where it might
870 * still work, it is barely relevant in practice.
872 * Exceptions are dynamic memory groups, which allow for more MOVABLE
873 * memory within the same memory group -- because in that case, there is
874 * coordination within the single memory device managed by a single driver.
876 * We rely on "present pages" instead of "managed pages", as the latter is
877 * highly unreliable and dynamic in virtualized environments, and does not
878 * consider boot time allocations. For example, memory ballooning adjusts the
879 * managed pages when inflating/deflating the balloon, and balloon compaction
880 * can even migrate inflated pages between zones.
882 * Using "present pages" is better but some things to keep in mind are:
884 * a) Some memblock allocations, such as for the crashkernel area, are
885 * effectively unused by the kernel, yet they account to "present pages".
886 * Fortunately, these allocations are comparatively small in relevant setups
887 * (e.g., fraction of system memory).
888 * b) Some hotplugged memory blocks in virtualized environments, esecially
889 * hotplugged by virtio-mem, look like they are completely present, however,
890 * only parts of the memory block are actually currently usable.
891 * "present pages" is an upper limit that can get reached at runtime. As
892 * we base our calculations on KERNEL_EARLY, this is not an issue.
894 static struct zone *auto_movable_zone_for_pfn(int nid,
895 struct memory_group *group,
897 unsigned long nr_pages)
899 unsigned long online_pages = 0, max_pages, end_pfn;
902 if (!auto_movable_ratio)
905 if (group && !group->is_dynamic) {
906 max_pages = group->s.max_pages;
907 online_pages = group->present_movable_pages;
909 /* If anything is !MOVABLE online the rest !MOVABLE. */
910 if (group->present_kernel_pages)
912 } else if (!group || group->d.unit_pages == nr_pages) {
913 max_pages = nr_pages;
915 max_pages = group->d.unit_pages;
917 * Take a look at all online sections in the current unit.
918 * We can safely assume that all pages within a section belong
919 * to the same zone, because dynamic memory groups only deal
920 * with hotplugged memory.
922 pfn = ALIGN_DOWN(pfn, group->d.unit_pages);
923 end_pfn = pfn + group->d.unit_pages;
924 for (; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
925 page = pfn_to_online_page(pfn);
928 /* If anything is !MOVABLE online the rest !MOVABLE. */
929 if (!is_zone_movable_page(page))
931 online_pages += PAGES_PER_SECTION;
936 * Online MOVABLE if we could *currently* online all remaining parts
937 * MOVABLE. We expect to (add+) online them immediately next, so if
938 * nobody interferes, all will be MOVABLE if possible.
940 nr_pages = max_pages - online_pages;
941 if (!auto_movable_can_online_movable(NUMA_NO_NODE, group, nr_pages))
945 if (auto_movable_numa_aware &&
946 !auto_movable_can_online_movable(nid, group, nr_pages))
948 #endif /* CONFIG_NUMA */
950 return &NODE_DATA(nid)->node_zones[ZONE_MOVABLE];
952 return default_kernel_zone_for_pfn(nid, pfn, nr_pages);
955 static inline struct zone *default_zone_for_pfn(int nid, unsigned long start_pfn,
956 unsigned long nr_pages)
958 struct zone *kernel_zone = default_kernel_zone_for_pfn(nid, start_pfn,
960 struct zone *movable_zone = &NODE_DATA(nid)->node_zones[ZONE_MOVABLE];
961 bool in_kernel = zone_intersects(kernel_zone, start_pfn, nr_pages);
962 bool in_movable = zone_intersects(movable_zone, start_pfn, nr_pages);
965 * We inherit the existing zone in a simple case where zones do not
966 * overlap in the given range
968 if (in_kernel ^ in_movable)
969 return (in_kernel) ? kernel_zone : movable_zone;
972 * If the range doesn't belong to any zone or two zones overlap in the
973 * given range then we use movable zone only if movable_node is
974 * enabled because we always online to a kernel zone by default.
976 return movable_node_enabled ? movable_zone : kernel_zone;
979 struct zone *zone_for_pfn_range(int online_type, int nid,
980 struct memory_group *group, unsigned long start_pfn,
981 unsigned long nr_pages)
983 if (online_type == MMOP_ONLINE_KERNEL)
984 return default_kernel_zone_for_pfn(nid, start_pfn, nr_pages);
986 if (online_type == MMOP_ONLINE_MOVABLE)
987 return &NODE_DATA(nid)->node_zones[ZONE_MOVABLE];
989 if (online_policy == ONLINE_POLICY_AUTO_MOVABLE)
990 return auto_movable_zone_for_pfn(nid, group, start_pfn, nr_pages);
992 return default_zone_for_pfn(nid, start_pfn, nr_pages);
996 * This function should only be called by memory_block_{online,offline},
997 * and {online,offline}_pages.
999 void adjust_present_page_count(struct page *page, struct memory_group *group,
1002 struct zone *zone = page_zone(page);
1003 const bool movable = zone_idx(zone) == ZONE_MOVABLE;
1006 * We only support onlining/offlining/adding/removing of complete
1007 * memory blocks; therefore, either all is either early or hotplugged.
1009 if (early_section(__pfn_to_section(page_to_pfn(page))))
1010 zone->present_early_pages += nr_pages;
1011 zone->present_pages += nr_pages;
1012 zone->zone_pgdat->node_present_pages += nr_pages;
1014 if (group && movable)
1015 group->present_movable_pages += nr_pages;
1016 else if (group && !movable)
1017 group->present_kernel_pages += nr_pages;
1020 int mhp_init_memmap_on_memory(unsigned long pfn, unsigned long nr_pages,
1023 unsigned long end_pfn = pfn + nr_pages;
1026 ret = kasan_add_zero_shadow(__va(PFN_PHYS(pfn)), PFN_PHYS(nr_pages));
1030 move_pfn_range_to_zone(zone, pfn, nr_pages, NULL, MIGRATE_UNMOVABLE);
1032 for (i = 0; i < nr_pages; i++)
1033 SetPageVmemmapSelfHosted(pfn_to_page(pfn + i));
1036 * It might be that the vmemmap_pages fully span sections. If that is
1037 * the case, mark those sections online here as otherwise they will be
1040 if (nr_pages >= PAGES_PER_SECTION)
1041 online_mem_sections(pfn, ALIGN_DOWN(end_pfn, PAGES_PER_SECTION));
1046 void mhp_deinit_memmap_on_memory(unsigned long pfn, unsigned long nr_pages)
1048 unsigned long end_pfn = pfn + nr_pages;
1051 * It might be that the vmemmap_pages fully span sections. If that is
1052 * the case, mark those sections offline here as otherwise they will be
1055 if (nr_pages >= PAGES_PER_SECTION)
1056 offline_mem_sections(pfn, ALIGN_DOWN(end_pfn, PAGES_PER_SECTION));
1059 * The pages associated with this vmemmap have been offlined, so
1060 * we can reset its state here.
1062 remove_pfn_range_from_zone(page_zone(pfn_to_page(pfn)), pfn, nr_pages);
1063 kasan_remove_zero_shadow(__va(PFN_PHYS(pfn)), PFN_PHYS(nr_pages));
1066 int __ref online_pages(unsigned long pfn, unsigned long nr_pages,
1067 struct zone *zone, struct memory_group *group)
1069 unsigned long flags;
1070 int need_zonelists_rebuild = 0;
1071 const int nid = zone_to_nid(zone);
1073 struct memory_notify arg;
1076 * {on,off}lining is constrained to full memory sections (or more
1077 * precisely to memory blocks from the user space POV).
1078 * memmap_on_memory is an exception because it reserves initial part
1079 * of the physical memory space for vmemmaps. That space is pageblock
1082 if (WARN_ON_ONCE(!nr_pages || !pageblock_aligned(pfn) ||
1083 !IS_ALIGNED(pfn + nr_pages, PAGES_PER_SECTION)))
1086 mem_hotplug_begin();
1088 /* associate pfn range with the zone */
1089 move_pfn_range_to_zone(zone, pfn, nr_pages, NULL, MIGRATE_ISOLATE);
1091 arg.start_pfn = pfn;
1092 arg.nr_pages = nr_pages;
1093 node_states_check_changes_online(nr_pages, zone, &arg);
1095 ret = memory_notify(MEM_GOING_ONLINE, &arg);
1096 ret = notifier_to_errno(ret);
1098 goto failed_addition;
1101 * Fixup the number of isolated pageblocks before marking the sections
1102 * onlining, such that undo_isolate_page_range() works correctly.
1104 spin_lock_irqsave(&zone->lock, flags);
1105 zone->nr_isolate_pageblock += nr_pages / pageblock_nr_pages;
1106 spin_unlock_irqrestore(&zone->lock, flags);
1109 * If this zone is not populated, then it is not in zonelist.
1110 * This means the page allocator ignores this zone.
1111 * So, zonelist must be updated after online.
1113 if (!populated_zone(zone)) {
1114 need_zonelists_rebuild = 1;
1115 setup_zone_pageset(zone);
1118 online_pages_range(pfn, nr_pages);
1119 adjust_present_page_count(pfn_to_page(pfn), group, nr_pages);
1121 node_states_set_node(nid, &arg);
1122 if (need_zonelists_rebuild)
1123 build_all_zonelists(NULL);
1125 /* Basic onlining is complete, allow allocation of onlined pages. */
1126 undo_isolate_page_range(pfn, pfn + nr_pages, MIGRATE_MOVABLE);
1129 * Freshly onlined pages aren't shuffled (e.g., all pages are placed to
1130 * the tail of the freelist when undoing isolation). Shuffle the whole
1131 * zone to make sure the just onlined pages are properly distributed
1132 * across the whole freelist - to create an initial shuffle.
1136 /* reinitialise watermarks and update pcp limits */
1137 init_per_zone_wmark_min();
1142 writeback_set_ratelimit();
1144 memory_notify(MEM_ONLINE, &arg);
1149 pr_debug("online_pages [mem %#010llx-%#010llx] failed\n",
1150 (unsigned long long) pfn << PAGE_SHIFT,
1151 (((unsigned long long) pfn + nr_pages) << PAGE_SHIFT) - 1);
1152 memory_notify(MEM_CANCEL_ONLINE, &arg);
1153 remove_pfn_range_from_zone(zone, pfn, nr_pages);
1158 /* we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG */
1159 static pg_data_t __ref *hotadd_init_pgdat(int nid)
1161 struct pglist_data *pgdat;
1164 * NODE_DATA is preallocated (free_area_init) but its internal
1165 * state is not allocated completely. Add missing pieces.
1166 * Completely offline nodes stay around and they just need
1169 pgdat = NODE_DATA(nid);
1171 /* init node's zones as empty zones, we don't have any present pages.*/
1172 free_area_init_core_hotplug(pgdat);
1175 * The node we allocated has no zone fallback lists. For avoiding
1176 * to access not-initialized zonelist, build here.
1178 build_all_zonelists(pgdat);
1184 * __try_online_node - online a node if offlined
1186 * @set_node_online: Whether we want to online the node
1187 * called by cpu_up() to online a node without onlined memory.
1190 * 1 -> a new node has been allocated
1191 * 0 -> the node is already online
1192 * -ENOMEM -> the node could not be allocated
1194 static int __try_online_node(int nid, bool set_node_online)
1199 if (node_online(nid))
1202 pgdat = hotadd_init_pgdat(nid);
1204 pr_err("Cannot online node %d due to NULL pgdat\n", nid);
1209 if (set_node_online) {
1210 node_set_online(nid);
1211 ret = register_one_node(nid);
1219 * Users of this function always want to online/register the node
1221 int try_online_node(int nid)
1225 mem_hotplug_begin();
1226 ret = __try_online_node(nid, true);
1231 static int check_hotplug_memory_range(u64 start, u64 size)
1233 /* memory range must be block size aligned */
1234 if (!size || !IS_ALIGNED(start, memory_block_size_bytes()) ||
1235 !IS_ALIGNED(size, memory_block_size_bytes())) {
1236 pr_err("Block size [%#lx] unaligned hotplug range: start %#llx, size %#llx",
1237 memory_block_size_bytes(), start, size);
1244 static int online_memory_block(struct memory_block *mem, void *arg)
1246 mem->online_type = mhp_default_online_type;
1247 return device_online(&mem->dev);
1250 static inline unsigned long memory_block_memmap_size(void)
1252 return PHYS_PFN(memory_block_size_bytes()) * sizeof(struct page);
1255 #ifndef arch_supports_memmap_on_memory
1256 static inline bool arch_supports_memmap_on_memory(unsigned long vmemmap_size)
1259 * As default, we want the vmemmap to span a complete PMD such that we
1260 * can map the vmemmap using a single PMD if supported by the
1263 return IS_ALIGNED(vmemmap_size, PMD_SIZE);
1267 static bool mhp_supports_memmap_on_memory(unsigned long size)
1269 unsigned long vmemmap_size = memory_block_memmap_size();
1270 unsigned long remaining_size = size - vmemmap_size;
1273 * Besides having arch support and the feature enabled at runtime, we
1274 * need a few more assumptions to hold true:
1276 * a) We span a single memory block: memory onlining/offlinin;g happens
1277 * in memory block granularity. We don't want the vmemmap of online
1278 * memory blocks to reside on offline memory blocks. In the future,
1279 * we might want to support variable-sized memory blocks to make the
1280 * feature more versatile.
1282 * b) The vmemmap pages span complete PMDs: We don't want vmemmap code
1283 * to populate memory from the altmap for unrelated parts (i.e.,
1284 * other memory blocks)
1286 * c) The vmemmap pages (and thereby the pages that will be exposed to
1287 * the buddy) have to cover full pageblocks: memory onlining/offlining
1288 * code requires applicable ranges to be page-aligned, for example, to
1289 * set the migratetypes properly.
1291 * TODO: Although we have a check here to make sure that vmemmap pages
1292 * fully populate a PMD, it is not the right place to check for
1293 * this. A much better solution involves improving vmemmap code
1294 * to fallback to base pages when trying to populate vmemmap using
1295 * altmap as an alternative source of memory, and we do not exactly
1296 * populate a single PMD.
1298 return mhp_memmap_on_memory() &&
1299 size == memory_block_size_bytes() &&
1300 IS_ALIGNED(remaining_size, (pageblock_nr_pages << PAGE_SHIFT)) &&
1301 arch_supports_memmap_on_memory(vmemmap_size);
1305 * NOTE: The caller must call lock_device_hotplug() to serialize hotplug
1306 * and online/offline operations (triggered e.g. by sysfs).
1308 * we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG
1310 int __ref add_memory_resource(int nid, struct resource *res, mhp_t mhp_flags)
1312 struct mhp_params params = { .pgprot = pgprot_mhp(PAGE_KERNEL) };
1313 enum memblock_flags memblock_flags = MEMBLOCK_NONE;
1314 struct vmem_altmap mhp_altmap = {};
1315 struct memory_group *group = NULL;
1317 bool new_node = false;
1321 size = resource_size(res);
1323 ret = check_hotplug_memory_range(start, size);
1327 if (mhp_flags & MHP_NID_IS_MGID) {
1328 group = memory_group_find_by_id(nid);
1334 if (!node_possible(nid)) {
1335 WARN(1, "node %d was absent from the node_possible_map\n", nid);
1339 mem_hotplug_begin();
1341 if (IS_ENABLED(CONFIG_ARCH_KEEP_MEMBLOCK)) {
1342 if (res->flags & IORESOURCE_SYSRAM_DRIVER_MANAGED)
1343 memblock_flags = MEMBLOCK_DRIVER_MANAGED;
1344 ret = memblock_add_node(start, size, nid, memblock_flags);
1346 goto error_mem_hotplug_end;
1349 ret = __try_online_node(nid, false);
1355 * Self hosted memmap array
1357 if (mhp_flags & MHP_MEMMAP_ON_MEMORY) {
1358 if (mhp_supports_memmap_on_memory(size)) {
1359 mhp_altmap.free = PHYS_PFN(size);
1360 mhp_altmap.base_pfn = PHYS_PFN(start);
1361 params.altmap = &mhp_altmap;
1363 /* fallback to not using altmap */
1366 /* call arch's memory hotadd */
1367 ret = arch_add_memory(nid, start, size, ¶ms);
1371 /* create memory block devices after memory was added */
1372 ret = create_memory_block_devices(start, size, mhp_altmap.alloc,
1375 arch_remove_memory(start, size, NULL);
1380 /* If sysfs file of new node can't be created, cpu on the node
1381 * can't be hot-added. There is no rollback way now.
1382 * So, check by BUG_ON() to catch it reluctantly..
1383 * We online node here. We can't roll back from here.
1385 node_set_online(nid);
1386 ret = __register_one_node(nid);
1390 register_memory_blocks_under_node(nid, PFN_DOWN(start),
1391 PFN_UP(start + size - 1),
1394 /* create new memmap entry */
1395 if (!strcmp(res->name, "System RAM"))
1396 firmware_map_add_hotplug(start, start + size, "System RAM");
1398 /* device_online() will take the lock when calling online_pages() */
1402 * In case we're allowed to merge the resource, flag it and trigger
1403 * merging now that adding succeeded.
1405 if (mhp_flags & MHP_MERGE_RESOURCE)
1406 merge_system_ram_resource(res);
1408 /* online pages if requested */
1409 if (mhp_default_online_type != MMOP_OFFLINE)
1410 walk_memory_blocks(start, size, NULL, online_memory_block);
1414 if (IS_ENABLED(CONFIG_ARCH_KEEP_MEMBLOCK))
1415 memblock_remove(start, size);
1416 error_mem_hotplug_end:
1421 /* requires device_hotplug_lock, see add_memory_resource() */
1422 int __ref __add_memory(int nid, u64 start, u64 size, mhp_t mhp_flags)
1424 struct resource *res;
1427 res = register_memory_resource(start, size, "System RAM");
1429 return PTR_ERR(res);
1431 ret = add_memory_resource(nid, res, mhp_flags);
1433 release_memory_resource(res);
1437 int add_memory(int nid, u64 start, u64 size, mhp_t mhp_flags)
1441 lock_device_hotplug();
1442 rc = __add_memory(nid, start, size, mhp_flags);
1443 unlock_device_hotplug();
1447 EXPORT_SYMBOL_GPL(add_memory);
1450 * Add special, driver-managed memory to the system as system RAM. Such
1451 * memory is not exposed via the raw firmware-provided memmap as system
1452 * RAM, instead, it is detected and added by a driver - during cold boot,
1453 * after a reboot, and after kexec.
1455 * Reasons why this memory should not be used for the initial memmap of a
1456 * kexec kernel or for placing kexec images:
1457 * - The booting kernel is in charge of determining how this memory will be
1458 * used (e.g., use persistent memory as system RAM)
1459 * - Coordination with a hypervisor is required before this memory
1460 * can be used (e.g., inaccessible parts).
1462 * For this memory, no entries in /sys/firmware/memmap ("raw firmware-provided
1463 * memory map") are created. Also, the created memory resource is flagged
1464 * with IORESOURCE_SYSRAM_DRIVER_MANAGED, so in-kernel users can special-case
1465 * this memory as well (esp., not place kexec images onto it).
1467 * The resource_name (visible via /proc/iomem) has to have the format
1468 * "System RAM ($DRIVER)".
1470 int add_memory_driver_managed(int nid, u64 start, u64 size,
1471 const char *resource_name, mhp_t mhp_flags)
1473 struct resource *res;
1476 if (!resource_name ||
1477 strstr(resource_name, "System RAM (") != resource_name ||
1478 resource_name[strlen(resource_name) - 1] != ')')
1481 lock_device_hotplug();
1483 res = register_memory_resource(start, size, resource_name);
1489 rc = add_memory_resource(nid, res, mhp_flags);
1491 release_memory_resource(res);
1494 unlock_device_hotplug();
1497 EXPORT_SYMBOL_GPL(add_memory_driver_managed);
1500 * Platforms should define arch_get_mappable_range() that provides
1501 * maximum possible addressable physical memory range for which the
1502 * linear mapping could be created. The platform returned address
1503 * range must adhere to these following semantics.
1505 * - range.start <= range.end
1506 * - Range includes both end points [range.start..range.end]
1508 * There is also a fallback definition provided here, allowing the
1509 * entire possible physical address range in case any platform does
1510 * not define arch_get_mappable_range().
1512 struct range __weak arch_get_mappable_range(void)
1514 struct range mhp_range = {
1521 struct range mhp_get_pluggable_range(bool need_mapping)
1523 const u64 max_phys = (1ULL << MAX_PHYSMEM_BITS) - 1;
1524 struct range mhp_range;
1527 mhp_range = arch_get_mappable_range();
1528 if (mhp_range.start > max_phys) {
1529 mhp_range.start = 0;
1532 mhp_range.end = min_t(u64, mhp_range.end, max_phys);
1534 mhp_range.start = 0;
1535 mhp_range.end = max_phys;
1539 EXPORT_SYMBOL_GPL(mhp_get_pluggable_range);
1541 bool mhp_range_allowed(u64 start, u64 size, bool need_mapping)
1543 struct range mhp_range = mhp_get_pluggable_range(need_mapping);
1544 u64 end = start + size;
1546 if (start < end && start >= mhp_range.start && (end - 1) <= mhp_range.end)
1549 pr_warn("Hotplug memory [%#llx-%#llx] exceeds maximum addressable range [%#llx-%#llx]\n",
1550 start, end, mhp_range.start, mhp_range.end);
1554 #ifdef CONFIG_MEMORY_HOTREMOVE
1556 * Scan pfn range [start,end) to find movable/migratable pages (LRU pages,
1557 * non-lru movable pages and hugepages). Will skip over most unmovable
1558 * pages (esp., pages that can be skipped when offlining), but bail out on
1559 * definitely unmovable pages.
1562 * 0 in case a movable page is found and movable_pfn was updated.
1563 * -ENOENT in case no movable page was found.
1564 * -EBUSY in case a definitely unmovable page was found.
1566 static int scan_movable_pages(unsigned long start, unsigned long end,
1567 unsigned long *movable_pfn)
1571 for (pfn = start; pfn < end; pfn++) {
1572 struct page *page, *head;
1575 if (!pfn_valid(pfn))
1577 page = pfn_to_page(pfn);
1580 if (__PageMovable(page))
1584 * PageOffline() pages that are not marked __PageMovable() and
1585 * have a reference count > 0 (after MEM_GOING_OFFLINE) are
1586 * definitely unmovable. If their reference count would be 0,
1587 * they could at least be skipped when offlining memory.
1589 if (PageOffline(page) && page_count(page))
1592 if (!PageHuge(page))
1594 head = compound_head(page);
1596 * This test is racy as we hold no reference or lock. The
1597 * hugetlb page could have been free'ed and head is no longer
1598 * a hugetlb page before the following check. In such unlikely
1599 * cases false positives and negatives are possible. Calling
1600 * code must deal with these scenarios.
1602 if (HPageMigratable(head))
1604 skip = compound_nr(head) - (page - head);
1613 static void do_migrate_range(unsigned long start_pfn, unsigned long end_pfn)
1616 struct page *page, *head;
1618 static DEFINE_RATELIMIT_STATE(migrate_rs, DEFAULT_RATELIMIT_INTERVAL,
1619 DEFAULT_RATELIMIT_BURST);
1621 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
1622 struct folio *folio;
1625 if (!pfn_valid(pfn))
1627 page = pfn_to_page(pfn);
1628 folio = page_folio(page);
1629 head = &folio->page;
1631 if (PageHuge(page)) {
1632 pfn = page_to_pfn(head) + compound_nr(head) - 1;
1633 isolate_hugetlb(folio, &source);
1635 } else if (PageTransHuge(page))
1636 pfn = page_to_pfn(head) + thp_nr_pages(page) - 1;
1639 * HWPoison pages have elevated reference counts so the migration would
1640 * fail on them. It also doesn't make any sense to migrate them in the
1641 * first place. Still try to unmap such a page in case it is still mapped
1642 * (e.g. current hwpoison implementation doesn't unmap KSM pages but keep
1643 * the unmap as the catch all safety net).
1645 if (PageHWPoison(page)) {
1646 if (WARN_ON(folio_test_lru(folio)))
1647 folio_isolate_lru(folio);
1648 if (folio_mapped(folio))
1649 try_to_unmap(folio, TTU_IGNORE_MLOCK);
1653 if (!get_page_unless_zero(page))
1656 * We can skip free pages. And we can deal with pages on
1657 * LRU and non-lru movable pages.
1660 isolated = isolate_lru_page(page);
1662 isolated = isolate_movable_page(page, ISOLATE_UNEVICTABLE);
1664 list_add_tail(&page->lru, &source);
1665 if (!__PageMovable(page))
1666 inc_node_page_state(page, NR_ISOLATED_ANON +
1667 page_is_file_lru(page));
1670 if (__ratelimit(&migrate_rs)) {
1671 pr_warn("failed to isolate pfn %lx\n", pfn);
1672 dump_page(page, "isolation failed");
1677 if (!list_empty(&source)) {
1678 nodemask_t nmask = node_states[N_MEMORY];
1679 struct migration_target_control mtc = {
1681 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
1686 * We have checked that migration range is on a single zone so
1687 * we can use the nid of the first page to all the others.
1689 mtc.nid = page_to_nid(list_first_entry(&source, struct page, lru));
1692 * try to allocate from a different node but reuse this node
1693 * if there are no other online nodes to be used (e.g. we are
1694 * offlining a part of the only existing node)
1696 node_clear(mtc.nid, nmask);
1697 if (nodes_empty(nmask))
1698 node_set(mtc.nid, nmask);
1699 ret = migrate_pages(&source, alloc_migration_target, NULL,
1700 (unsigned long)&mtc, MIGRATE_SYNC, MR_MEMORY_HOTPLUG, NULL);
1702 list_for_each_entry(page, &source, lru) {
1703 if (__ratelimit(&migrate_rs)) {
1704 pr_warn("migrating pfn %lx failed ret:%d\n",
1705 page_to_pfn(page), ret);
1706 dump_page(page, "migration failure");
1709 putback_movable_pages(&source);
1714 static int __init cmdline_parse_movable_node(char *p)
1716 movable_node_enabled = true;
1719 early_param("movable_node", cmdline_parse_movable_node);
1721 /* check which state of node_states will be changed when offline memory */
1722 static void node_states_check_changes_offline(unsigned long nr_pages,
1723 struct zone *zone, struct memory_notify *arg)
1725 struct pglist_data *pgdat = zone->zone_pgdat;
1726 unsigned long present_pages = 0;
1729 arg->status_change_nid = NUMA_NO_NODE;
1730 arg->status_change_nid_normal = NUMA_NO_NODE;
1733 * Check whether node_states[N_NORMAL_MEMORY] will be changed.
1734 * If the memory to be offline is within the range
1735 * [0..ZONE_NORMAL], and it is the last present memory there,
1736 * the zones in that range will become empty after the offlining,
1737 * thus we can determine that we need to clear the node from
1738 * node_states[N_NORMAL_MEMORY].
1740 for (zt = 0; zt <= ZONE_NORMAL; zt++)
1741 present_pages += pgdat->node_zones[zt].present_pages;
1742 if (zone_idx(zone) <= ZONE_NORMAL && nr_pages >= present_pages)
1743 arg->status_change_nid_normal = zone_to_nid(zone);
1746 * We have accounted the pages from [0..ZONE_NORMAL); ZONE_HIGHMEM
1747 * does not apply as we don't support 32bit.
1748 * Here we count the possible pages from ZONE_MOVABLE.
1749 * If after having accounted all the pages, we see that the nr_pages
1750 * to be offlined is over or equal to the accounted pages,
1751 * we know that the node will become empty, and so, we can clear
1752 * it for N_MEMORY as well.
1754 present_pages += pgdat->node_zones[ZONE_MOVABLE].present_pages;
1756 if (nr_pages >= present_pages)
1757 arg->status_change_nid = zone_to_nid(zone);
1760 static void node_states_clear_node(int node, struct memory_notify *arg)
1762 if (arg->status_change_nid_normal >= 0)
1763 node_clear_state(node, N_NORMAL_MEMORY);
1765 if (arg->status_change_nid >= 0)
1766 node_clear_state(node, N_MEMORY);
1769 static int count_system_ram_pages_cb(unsigned long start_pfn,
1770 unsigned long nr_pages, void *data)
1772 unsigned long *nr_system_ram_pages = data;
1774 *nr_system_ram_pages += nr_pages;
1778 int __ref offline_pages(unsigned long start_pfn, unsigned long nr_pages,
1779 struct zone *zone, struct memory_group *group)
1781 const unsigned long end_pfn = start_pfn + nr_pages;
1782 unsigned long pfn, system_ram_pages = 0;
1783 const int node = zone_to_nid(zone);
1784 unsigned long flags;
1785 struct memory_notify arg;
1790 * {on,off}lining is constrained to full memory sections (or more
1791 * precisely to memory blocks from the user space POV).
1792 * memmap_on_memory is an exception because it reserves initial part
1793 * of the physical memory space for vmemmaps. That space is pageblock
1796 if (WARN_ON_ONCE(!nr_pages || !pageblock_aligned(start_pfn) ||
1797 !IS_ALIGNED(start_pfn + nr_pages, PAGES_PER_SECTION)))
1800 mem_hotplug_begin();
1803 * Don't allow to offline memory blocks that contain holes.
1804 * Consequently, memory blocks with holes can never get onlined
1805 * via the hotplug path - online_pages() - as hotplugged memory has
1806 * no holes. This way, we e.g., don't have to worry about marking
1807 * memory holes PG_reserved, don't need pfn_valid() checks, and can
1808 * avoid using walk_system_ram_range() later.
1810 walk_system_ram_range(start_pfn, nr_pages, &system_ram_pages,
1811 count_system_ram_pages_cb);
1812 if (system_ram_pages != nr_pages) {
1814 reason = "memory holes";
1815 goto failed_removal;
1819 * We only support offlining of memory blocks managed by a single zone,
1820 * checked by calling code. This is just a sanity check that we might
1821 * want to remove in the future.
1823 if (WARN_ON_ONCE(page_zone(pfn_to_page(start_pfn)) != zone ||
1824 page_zone(pfn_to_page(end_pfn - 1)) != zone)) {
1826 reason = "multizone range";
1827 goto failed_removal;
1831 * Disable pcplists so that page isolation cannot race with freeing
1832 * in a way that pages from isolated pageblock are left on pcplists.
1834 zone_pcp_disable(zone);
1835 lru_cache_disable();
1837 /* set above range as isolated */
1838 ret = start_isolate_page_range(start_pfn, end_pfn,
1840 MEMORY_OFFLINE | REPORT_FAILURE,
1841 GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL);
1843 reason = "failure to isolate range";
1844 goto failed_removal_pcplists_disabled;
1847 arg.start_pfn = start_pfn;
1848 arg.nr_pages = nr_pages;
1849 node_states_check_changes_offline(nr_pages, zone, &arg);
1851 ret = memory_notify(MEM_GOING_OFFLINE, &arg);
1852 ret = notifier_to_errno(ret);
1854 reason = "notifier failure";
1855 goto failed_removal_isolated;
1862 * Historically we always checked for any signal and
1863 * can't limit it to fatal signals without eventually
1864 * breaking user space.
1866 if (signal_pending(current)) {
1868 reason = "signal backoff";
1869 goto failed_removal_isolated;
1874 ret = scan_movable_pages(pfn, end_pfn, &pfn);
1877 * TODO: fatal migration failures should bail
1880 do_migrate_range(pfn, end_pfn);
1884 if (ret != -ENOENT) {
1885 reason = "unmovable page";
1886 goto failed_removal_isolated;
1890 * Dissolve free hugepages in the memory block before doing
1891 * offlining actually in order to make hugetlbfs's object
1892 * counting consistent.
1894 ret = dissolve_free_huge_pages(start_pfn, end_pfn);
1896 reason = "failure to dissolve huge pages";
1897 goto failed_removal_isolated;
1900 ret = test_pages_isolated(start_pfn, end_pfn, MEMORY_OFFLINE);
1904 /* Mark all sections offline and remove free pages from the buddy. */
1905 __offline_isolated_pages(start_pfn, end_pfn);
1906 pr_debug("Offlined Pages %ld\n", nr_pages);
1909 * The memory sections are marked offline, and the pageblock flags
1910 * effectively stale; nobody should be touching them. Fixup the number
1911 * of isolated pageblocks, memory onlining will properly revert this.
1913 spin_lock_irqsave(&zone->lock, flags);
1914 zone->nr_isolate_pageblock -= nr_pages / pageblock_nr_pages;
1915 spin_unlock_irqrestore(&zone->lock, flags);
1918 zone_pcp_enable(zone);
1920 /* removal success */
1921 adjust_managed_page_count(pfn_to_page(start_pfn), -nr_pages);
1922 adjust_present_page_count(pfn_to_page(start_pfn), group, -nr_pages);
1924 /* reinitialise watermarks and update pcp limits */
1925 init_per_zone_wmark_min();
1927 if (!populated_zone(zone)) {
1928 zone_pcp_reset(zone);
1929 build_all_zonelists(NULL);
1932 node_states_clear_node(node, &arg);
1933 if (arg.status_change_nid >= 0) {
1934 kcompactd_stop(node);
1938 writeback_set_ratelimit();
1940 memory_notify(MEM_OFFLINE, &arg);
1941 remove_pfn_range_from_zone(zone, start_pfn, nr_pages);
1945 failed_removal_isolated:
1946 /* pushback to free area */
1947 undo_isolate_page_range(start_pfn, end_pfn, MIGRATE_MOVABLE);
1948 memory_notify(MEM_CANCEL_OFFLINE, &arg);
1949 failed_removal_pcplists_disabled:
1951 zone_pcp_enable(zone);
1953 pr_debug("memory offlining [mem %#010llx-%#010llx] failed due to %s\n",
1954 (unsigned long long) start_pfn << PAGE_SHIFT,
1955 ((unsigned long long) end_pfn << PAGE_SHIFT) - 1,
1961 static int check_memblock_offlined_cb(struct memory_block *mem, void *arg)
1966 if (unlikely(mem->state != MEM_OFFLINE)) {
1967 phys_addr_t beginpa, endpa;
1969 beginpa = PFN_PHYS(section_nr_to_pfn(mem->start_section_nr));
1970 endpa = beginpa + memory_block_size_bytes() - 1;
1971 pr_warn("removing memory fails, because memory [%pa-%pa] is onlined\n",
1979 static int get_nr_vmemmap_pages_cb(struct memory_block *mem, void *arg)
1982 * If not set, continue with the next block.
1984 return mem->nr_vmemmap_pages;
1987 static int check_cpu_on_node(int nid)
1991 for_each_present_cpu(cpu) {
1992 if (cpu_to_node(cpu) == nid)
1994 * the cpu on this node isn't removed, and we can't
1995 * offline this node.
2003 static int check_no_memblock_for_node_cb(struct memory_block *mem, void *arg)
2005 int nid = *(int *)arg;
2008 * If a memory block belongs to multiple nodes, the stored nid is not
2009 * reliable. However, such blocks are always online (e.g., cannot get
2010 * offlined) and, therefore, are still spanned by the node.
2012 return mem->nid == nid ? -EEXIST : 0;
2019 * Offline a node if all memory sections and cpus of the node are removed.
2021 * NOTE: The caller must call lock_device_hotplug() to serialize hotplug
2022 * and online/offline operations before this call.
2024 void try_offline_node(int nid)
2029 * If the node still spans pages (especially ZONE_DEVICE), don't
2030 * offline it. A node spans memory after move_pfn_range_to_zone(),
2031 * e.g., after the memory block was onlined.
2033 if (node_spanned_pages(nid))
2037 * Especially offline memory blocks might not be spanned by the
2038 * node. They will get spanned by the node once they get onlined.
2039 * However, they link to the node in sysfs and can get onlined later.
2041 rc = for_each_memory_block(&nid, check_no_memblock_for_node_cb);
2045 if (check_cpu_on_node(nid))
2049 * all memory/cpu of this node are removed, we can offline this
2052 node_set_offline(nid);
2053 unregister_one_node(nid);
2055 EXPORT_SYMBOL(try_offline_node);
2057 static int __ref try_remove_memory(u64 start, u64 size)
2059 struct vmem_altmap mhp_altmap = {};
2060 struct vmem_altmap *altmap = NULL;
2061 unsigned long nr_vmemmap_pages;
2062 int rc = 0, nid = NUMA_NO_NODE;
2064 BUG_ON(check_hotplug_memory_range(start, size));
2067 * All memory blocks must be offlined before removing memory. Check
2068 * whether all memory blocks in question are offline and return error
2069 * if this is not the case.
2071 * While at it, determine the nid. Note that if we'd have mixed nodes,
2072 * we'd only try to offline the last determined one -- which is good
2073 * enough for the cases we care about.
2075 rc = walk_memory_blocks(start, size, &nid, check_memblock_offlined_cb);
2080 * We only support removing memory added with MHP_MEMMAP_ON_MEMORY in
2081 * the same granularity it was added - a single memory block.
2083 if (mhp_memmap_on_memory()) {
2084 nr_vmemmap_pages = walk_memory_blocks(start, size, NULL,
2085 get_nr_vmemmap_pages_cb);
2086 if (nr_vmemmap_pages) {
2087 if (size != memory_block_size_bytes()) {
2088 pr_warn("Refuse to remove %#llx - %#llx,"
2089 "wrong granularity\n",
2090 start, start + size);
2095 * Let remove_pmd_table->free_hugepage_table do the
2096 * right thing if we used vmem_altmap when hot-adding
2099 mhp_altmap.alloc = nr_vmemmap_pages;
2100 altmap = &mhp_altmap;
2104 /* remove memmap entry */
2105 firmware_map_remove(start, start + size, "System RAM");
2108 * Memory block device removal under the device_hotplug_lock is
2109 * a barrier against racing online attempts.
2111 remove_memory_block_devices(start, size);
2113 mem_hotplug_begin();
2115 arch_remove_memory(start, size, altmap);
2117 if (IS_ENABLED(CONFIG_ARCH_KEEP_MEMBLOCK)) {
2118 memblock_phys_free(start, size);
2119 memblock_remove(start, size);
2122 release_mem_region_adjustable(start, size);
2124 if (nid != NUMA_NO_NODE)
2125 try_offline_node(nid);
2132 * __remove_memory - Remove memory if every memory block is offline
2133 * @start: physical address of the region to remove
2134 * @size: size of the region to remove
2136 * NOTE: The caller must call lock_device_hotplug() to serialize hotplug
2137 * and online/offline operations before this call, as required by
2138 * try_offline_node().
2140 void __remove_memory(u64 start, u64 size)
2144 * trigger BUG() if some memory is not offlined prior to calling this
2147 if (try_remove_memory(start, size))
2152 * Remove memory if every memory block is offline, otherwise return -EBUSY is
2153 * some memory is not offline
2155 int remove_memory(u64 start, u64 size)
2159 lock_device_hotplug();
2160 rc = try_remove_memory(start, size);
2161 unlock_device_hotplug();
2165 EXPORT_SYMBOL_GPL(remove_memory);
2167 static int try_offline_memory_block(struct memory_block *mem, void *arg)
2169 uint8_t online_type = MMOP_ONLINE_KERNEL;
2170 uint8_t **online_types = arg;
2175 * Sense the online_type via the zone of the memory block. Offlining
2176 * with multiple zones within one memory block will be rejected
2177 * by offlining code ... so we don't care about that.
2179 page = pfn_to_online_page(section_nr_to_pfn(mem->start_section_nr));
2180 if (page && zone_idx(page_zone(page)) == ZONE_MOVABLE)
2181 online_type = MMOP_ONLINE_MOVABLE;
2183 rc = device_offline(&mem->dev);
2185 * Default is MMOP_OFFLINE - change it only if offlining succeeded,
2186 * so try_reonline_memory_block() can do the right thing.
2189 **online_types = online_type;
2192 /* Ignore if already offline. */
2193 return rc < 0 ? rc : 0;
2196 static int try_reonline_memory_block(struct memory_block *mem, void *arg)
2198 uint8_t **online_types = arg;
2201 if (**online_types != MMOP_OFFLINE) {
2202 mem->online_type = **online_types;
2203 rc = device_online(&mem->dev);
2205 pr_warn("%s: Failed to re-online memory: %d",
2209 /* Continue processing all remaining memory blocks. */
2215 * Try to offline and remove memory. Might take a long time to finish in case
2216 * memory is still in use. Primarily useful for memory devices that logically
2217 * unplugged all memory (so it's no longer in use) and want to offline + remove
2220 int offline_and_remove_memory(u64 start, u64 size)
2222 const unsigned long mb_count = size / memory_block_size_bytes();
2223 uint8_t *online_types, *tmp;
2226 if (!IS_ALIGNED(start, memory_block_size_bytes()) ||
2227 !IS_ALIGNED(size, memory_block_size_bytes()) || !size)
2231 * We'll remember the old online type of each memory block, so we can
2232 * try to revert whatever we did when offlining one memory block fails
2233 * after offlining some others succeeded.
2235 online_types = kmalloc_array(mb_count, sizeof(*online_types),
2240 * Initialize all states to MMOP_OFFLINE, so when we abort processing in
2241 * try_offline_memory_block(), we'll skip all unprocessed blocks in
2242 * try_reonline_memory_block().
2244 memset(online_types, MMOP_OFFLINE, mb_count);
2246 lock_device_hotplug();
2249 rc = walk_memory_blocks(start, size, &tmp, try_offline_memory_block);
2252 * In case we succeeded to offline all memory, remove it.
2253 * This cannot fail as it cannot get onlined in the meantime.
2256 rc = try_remove_memory(start, size);
2258 pr_err("%s: Failed to remove memory: %d", __func__, rc);
2262 * Rollback what we did. While memory onlining might theoretically fail
2263 * (nacked by a notifier), it barely ever happens.
2267 walk_memory_blocks(start, size, &tmp,
2268 try_reonline_memory_block);
2270 unlock_device_hotplug();
2272 kfree(online_types);
2275 EXPORT_SYMBOL_GPL(offline_and_remove_memory);
2276 #endif /* CONFIG_MEMORY_HOTREMOVE */