1 // SPDX-License-Identifier: GPL-2.0-only
3 * linux/mm/memory_hotplug.c
8 #include <linux/stddef.h>
10 #include <linux/sched/signal.h>
11 #include <linux/swap.h>
12 #include <linux/interrupt.h>
13 #include <linux/pagemap.h>
14 #include <linux/compiler.h>
15 #include <linux/export.h>
16 #include <linux/pagevec.h>
17 #include <linux/writeback.h>
18 #include <linux/slab.h>
19 #include <linux/sysctl.h>
20 #include <linux/cpu.h>
21 #include <linux/memory.h>
22 #include <linux/memremap.h>
23 #include <linux/memory_hotplug.h>
24 #include <linux/highmem.h>
25 #include <linux/vmalloc.h>
26 #include <linux/ioport.h>
27 #include <linux/delay.h>
28 #include <linux/migrate.h>
29 #include <linux/page-isolation.h>
30 #include <linux/pfn.h>
31 #include <linux/suspend.h>
32 #include <linux/mm_inline.h>
33 #include <linux/firmware-map.h>
34 #include <linux/stop_machine.h>
35 #include <linux/hugetlb.h>
36 #include <linux/memblock.h>
37 #include <linux/compaction.h>
38 #include <linux/rmap.h>
40 #include <asm/tlbflush.h>
47 * memory_hotplug.memmap_on_memory parameter
49 static bool memmap_on_memory __ro_after_init;
50 #ifdef CONFIG_MHP_MEMMAP_ON_MEMORY
51 module_param(memmap_on_memory, bool, 0444);
52 MODULE_PARM_DESC(memmap_on_memory, "Enable memmap on memory for memory hotplug");
56 ONLINE_POLICY_CONTIG_ZONES = 0,
57 ONLINE_POLICY_AUTO_MOVABLE,
60 const char *online_policy_to_str[] = {
61 [ONLINE_POLICY_CONTIG_ZONES] = "contig-zones",
62 [ONLINE_POLICY_AUTO_MOVABLE] = "auto-movable",
65 static int set_online_policy(const char *val, const struct kernel_param *kp)
67 int ret = sysfs_match_string(online_policy_to_str, val);
71 *((int *)kp->arg) = ret;
75 static int get_online_policy(char *buffer, const struct kernel_param *kp)
77 return sprintf(buffer, "%s\n", online_policy_to_str[*((int *)kp->arg)]);
81 * memory_hotplug.online_policy: configure online behavior when onlining without
82 * specifying a zone (MMOP_ONLINE)
84 * "contig-zones": keep zone contiguous
85 * "auto-movable": online memory to ZONE_MOVABLE if the configuration
86 * (auto_movable_ratio, auto_movable_numa_aware) allows for it
88 static int online_policy __read_mostly = ONLINE_POLICY_CONTIG_ZONES;
89 static const struct kernel_param_ops online_policy_ops = {
90 .set = set_online_policy,
91 .get = get_online_policy,
93 module_param_cb(online_policy, &online_policy_ops, &online_policy, 0644);
94 MODULE_PARM_DESC(online_policy,
95 "Set the online policy (\"contig-zones\", \"auto-movable\") "
96 "Default: \"contig-zones\"");
99 * memory_hotplug.auto_movable_ratio: specify maximum MOVABLE:KERNEL ratio
101 * The ratio represent an upper limit and the kernel might decide to not
102 * online some memory to ZONE_MOVABLE -- e.g., because hotplugged KERNEL memory
103 * doesn't allow for more MOVABLE memory.
105 static unsigned int auto_movable_ratio __read_mostly = 301;
106 module_param(auto_movable_ratio, uint, 0644);
107 MODULE_PARM_DESC(auto_movable_ratio,
108 "Set the maximum ratio of MOVABLE:KERNEL memory in the system "
109 "in percent for \"auto-movable\" online policy. Default: 301");
112 * memory_hotplug.auto_movable_numa_aware: consider numa node stats
115 static bool auto_movable_numa_aware __read_mostly = true;
116 module_param(auto_movable_numa_aware, bool, 0644);
117 MODULE_PARM_DESC(auto_movable_numa_aware,
118 "Consider numa node stats in addition to global stats in "
119 "\"auto-movable\" online policy. Default: true");
120 #endif /* CONFIG_NUMA */
123 * online_page_callback contains pointer to current page onlining function.
124 * Initially it is generic_online_page(). If it is required it could be
125 * changed by calling set_online_page_callback() for callback registration
126 * and restore_online_page_callback() for generic callback restore.
129 static online_page_callback_t online_page_callback = generic_online_page;
130 static DEFINE_MUTEX(online_page_callback_lock);
132 DEFINE_STATIC_PERCPU_RWSEM(mem_hotplug_lock);
134 void get_online_mems(void)
136 percpu_down_read(&mem_hotplug_lock);
139 void put_online_mems(void)
141 percpu_up_read(&mem_hotplug_lock);
144 bool movable_node_enabled = false;
146 #ifndef CONFIG_MEMORY_HOTPLUG_DEFAULT_ONLINE
147 int mhp_default_online_type = MMOP_OFFLINE;
149 int mhp_default_online_type = MMOP_ONLINE;
152 static int __init setup_memhp_default_state(char *str)
154 const int online_type = mhp_online_type_from_str(str);
156 if (online_type >= 0)
157 mhp_default_online_type = online_type;
161 __setup("memhp_default_state=", setup_memhp_default_state);
163 void mem_hotplug_begin(void)
166 percpu_down_write(&mem_hotplug_lock);
169 void mem_hotplug_done(void)
171 percpu_up_write(&mem_hotplug_lock);
175 u64 max_mem_size = U64_MAX;
177 /* add this memory to iomem resource */
178 static struct resource *register_memory_resource(u64 start, u64 size,
179 const char *resource_name)
181 struct resource *res;
182 unsigned long flags = IORESOURCE_SYSTEM_RAM | IORESOURCE_BUSY;
184 if (strcmp(resource_name, "System RAM"))
185 flags |= IORESOURCE_SYSRAM_DRIVER_MANAGED;
187 if (!mhp_range_allowed(start, size, true))
188 return ERR_PTR(-E2BIG);
191 * Make sure value parsed from 'mem=' only restricts memory adding
192 * while booting, so that memory hotplug won't be impacted. Please
193 * refer to document of 'mem=' in kernel-parameters.txt for more
196 if (start + size > max_mem_size && system_state < SYSTEM_RUNNING)
197 return ERR_PTR(-E2BIG);
200 * Request ownership of the new memory range. This might be
201 * a child of an existing resource that was present but
202 * not marked as busy.
204 res = __request_region(&iomem_resource, start, size,
205 resource_name, flags);
208 pr_debug("Unable to reserve System RAM region: %016llx->%016llx\n",
209 start, start + size);
210 return ERR_PTR(-EEXIST);
215 static void release_memory_resource(struct resource *res)
219 release_resource(res);
223 #ifdef CONFIG_MEMORY_HOTPLUG_SPARSE
224 static int check_pfn_span(unsigned long pfn, unsigned long nr_pages,
228 * Disallow all operations smaller than a sub-section and only
229 * allow operations smaller than a section for
230 * SPARSEMEM_VMEMMAP. Note that check_hotplug_memory_range()
231 * enforces a larger memory_block_size_bytes() granularity for
232 * memory that will be marked online, so this check should only
233 * fire for direct arch_{add,remove}_memory() users outside of
234 * add_memory_resource().
236 unsigned long min_align;
238 if (IS_ENABLED(CONFIG_SPARSEMEM_VMEMMAP))
239 min_align = PAGES_PER_SUBSECTION;
241 min_align = PAGES_PER_SECTION;
242 if (!IS_ALIGNED(pfn, min_align)
243 || !IS_ALIGNED(nr_pages, min_align)) {
244 WARN(1, "Misaligned __%s_pages start: %#lx end: #%lx\n",
245 reason, pfn, pfn + nr_pages - 1);
252 * Return page for the valid pfn only if the page is online. All pfn
253 * walkers which rely on the fully initialized page->flags and others
254 * should use this rather than pfn_valid && pfn_to_page
256 struct page *pfn_to_online_page(unsigned long pfn)
258 unsigned long nr = pfn_to_section_nr(pfn);
259 struct dev_pagemap *pgmap;
260 struct mem_section *ms;
262 if (nr >= NR_MEM_SECTIONS)
265 ms = __nr_to_section(nr);
266 if (!online_section(ms))
270 * Save some code text when online_section() +
271 * pfn_section_valid() are sufficient.
273 if (IS_ENABLED(CONFIG_HAVE_ARCH_PFN_VALID) && !pfn_valid(pfn))
276 if (!pfn_section_valid(ms, pfn))
279 if (!online_device_section(ms))
280 return pfn_to_page(pfn);
283 * Slowpath: when ZONE_DEVICE collides with
284 * ZONE_{NORMAL,MOVABLE} within the same section some pfns in
285 * the section may be 'offline' but 'valid'. Only
286 * get_dev_pagemap() can determine sub-section online status.
288 pgmap = get_dev_pagemap(pfn, NULL);
289 put_dev_pagemap(pgmap);
291 /* The presence of a pgmap indicates ZONE_DEVICE offline pfn */
295 return pfn_to_page(pfn);
297 EXPORT_SYMBOL_GPL(pfn_to_online_page);
300 * Reasonably generic function for adding memory. It is
301 * expected that archs that support memory hotplug will
302 * call this function after deciding the zone to which to
305 int __ref __add_pages(int nid, unsigned long pfn, unsigned long nr_pages,
306 struct mhp_params *params)
308 const unsigned long end_pfn = pfn + nr_pages;
309 unsigned long cur_nr_pages;
311 struct vmem_altmap *altmap = params->altmap;
313 if (WARN_ON_ONCE(!params->pgprot.pgprot))
316 VM_BUG_ON(!mhp_range_allowed(PFN_PHYS(pfn), nr_pages * PAGE_SIZE, false));
320 * Validate altmap is within bounds of the total request
322 if (altmap->base_pfn != pfn
323 || vmem_altmap_offset(altmap) > nr_pages) {
324 pr_warn_once("memory add fail, invalid altmap\n");
330 err = check_pfn_span(pfn, nr_pages, "add");
334 for (; pfn < end_pfn; pfn += cur_nr_pages) {
335 /* Select all remaining pages up to the next section boundary */
336 cur_nr_pages = min(end_pfn - pfn,
337 SECTION_ALIGN_UP(pfn + 1) - pfn);
338 err = sparse_add_section(nid, pfn, cur_nr_pages, altmap);
343 vmemmap_populate_print_last();
347 /* find the smallest valid pfn in the range [start_pfn, end_pfn) */
348 static unsigned long find_smallest_section_pfn(int nid, struct zone *zone,
349 unsigned long start_pfn,
350 unsigned long end_pfn)
352 for (; start_pfn < end_pfn; start_pfn += PAGES_PER_SUBSECTION) {
353 if (unlikely(!pfn_to_online_page(start_pfn)))
356 if (unlikely(pfn_to_nid(start_pfn) != nid))
359 if (zone != page_zone(pfn_to_page(start_pfn)))
368 /* find the biggest valid pfn in the range [start_pfn, end_pfn). */
369 static unsigned long find_biggest_section_pfn(int nid, struct zone *zone,
370 unsigned long start_pfn,
371 unsigned long end_pfn)
375 /* pfn is the end pfn of a memory section. */
377 for (; pfn >= start_pfn; pfn -= PAGES_PER_SUBSECTION) {
378 if (unlikely(!pfn_to_online_page(pfn)))
381 if (unlikely(pfn_to_nid(pfn) != nid))
384 if (zone != page_zone(pfn_to_page(pfn)))
393 static void shrink_zone_span(struct zone *zone, unsigned long start_pfn,
394 unsigned long end_pfn)
397 int nid = zone_to_nid(zone);
399 if (zone->zone_start_pfn == start_pfn) {
401 * If the section is smallest section in the zone, it need
402 * shrink zone->zone_start_pfn and zone->zone_spanned_pages.
403 * In this case, we find second smallest valid mem_section
404 * for shrinking zone.
406 pfn = find_smallest_section_pfn(nid, zone, end_pfn,
409 zone->spanned_pages = zone_end_pfn(zone) - pfn;
410 zone->zone_start_pfn = pfn;
412 zone->zone_start_pfn = 0;
413 zone->spanned_pages = 0;
415 } else if (zone_end_pfn(zone) == end_pfn) {
417 * If the section is biggest section in the zone, it need
418 * shrink zone->spanned_pages.
419 * In this case, we find second biggest valid mem_section for
422 pfn = find_biggest_section_pfn(nid, zone, zone->zone_start_pfn,
425 zone->spanned_pages = pfn - zone->zone_start_pfn + 1;
427 zone->zone_start_pfn = 0;
428 zone->spanned_pages = 0;
433 static void update_pgdat_span(struct pglist_data *pgdat)
435 unsigned long node_start_pfn = 0, node_end_pfn = 0;
438 for (zone = pgdat->node_zones;
439 zone < pgdat->node_zones + MAX_NR_ZONES; zone++) {
440 unsigned long end_pfn = zone_end_pfn(zone);
442 /* No need to lock the zones, they can't change. */
443 if (!zone->spanned_pages)
446 node_start_pfn = zone->zone_start_pfn;
447 node_end_pfn = end_pfn;
451 if (end_pfn > node_end_pfn)
452 node_end_pfn = end_pfn;
453 if (zone->zone_start_pfn < node_start_pfn)
454 node_start_pfn = zone->zone_start_pfn;
457 pgdat->node_start_pfn = node_start_pfn;
458 pgdat->node_spanned_pages = node_end_pfn - node_start_pfn;
461 void __ref remove_pfn_range_from_zone(struct zone *zone,
462 unsigned long start_pfn,
463 unsigned long nr_pages)
465 const unsigned long end_pfn = start_pfn + nr_pages;
466 struct pglist_data *pgdat = zone->zone_pgdat;
467 unsigned long pfn, cur_nr_pages;
469 /* Poison struct pages because they are now uninitialized again. */
470 for (pfn = start_pfn; pfn < end_pfn; pfn += cur_nr_pages) {
473 /* Select all remaining pages up to the next section boundary */
475 min(end_pfn - pfn, SECTION_ALIGN_UP(pfn + 1) - pfn);
476 page_init_poison(pfn_to_page(pfn),
477 sizeof(struct page) * cur_nr_pages);
480 #ifdef CONFIG_ZONE_DEVICE
482 * Zone shrinking code cannot properly deal with ZONE_DEVICE. So
483 * we will not try to shrink the zones - which is okay as
484 * set_zone_contiguous() cannot deal with ZONE_DEVICE either way.
486 if (zone_idx(zone) == ZONE_DEVICE)
490 clear_zone_contiguous(zone);
492 shrink_zone_span(zone, start_pfn, start_pfn + nr_pages);
493 update_pgdat_span(pgdat);
495 set_zone_contiguous(zone);
498 static void __remove_section(unsigned long pfn, unsigned long nr_pages,
499 unsigned long map_offset,
500 struct vmem_altmap *altmap)
502 struct mem_section *ms = __pfn_to_section(pfn);
504 if (WARN_ON_ONCE(!valid_section(ms)))
507 sparse_remove_section(ms, pfn, nr_pages, map_offset, altmap);
511 * __remove_pages() - remove sections of pages
512 * @pfn: starting pageframe (must be aligned to start of a section)
513 * @nr_pages: number of pages to remove (must be multiple of section size)
514 * @altmap: alternative device page map or %NULL if default memmap is used
516 * Generic helper function to remove section mappings and sysfs entries
517 * for the section of the memory we are removing. Caller needs to make
518 * sure that pages are marked reserved and zones are adjust properly by
519 * calling offline_pages().
521 void __remove_pages(unsigned long pfn, unsigned long nr_pages,
522 struct vmem_altmap *altmap)
524 const unsigned long end_pfn = pfn + nr_pages;
525 unsigned long cur_nr_pages;
526 unsigned long map_offset = 0;
528 map_offset = vmem_altmap_offset(altmap);
530 if (check_pfn_span(pfn, nr_pages, "remove"))
533 for (; pfn < end_pfn; pfn += cur_nr_pages) {
535 /* Select all remaining pages up to the next section boundary */
536 cur_nr_pages = min(end_pfn - pfn,
537 SECTION_ALIGN_UP(pfn + 1) - pfn);
538 __remove_section(pfn, cur_nr_pages, map_offset, altmap);
543 int set_online_page_callback(online_page_callback_t callback)
548 mutex_lock(&online_page_callback_lock);
550 if (online_page_callback == generic_online_page) {
551 online_page_callback = callback;
555 mutex_unlock(&online_page_callback_lock);
560 EXPORT_SYMBOL_GPL(set_online_page_callback);
562 int restore_online_page_callback(online_page_callback_t callback)
567 mutex_lock(&online_page_callback_lock);
569 if (online_page_callback == callback) {
570 online_page_callback = generic_online_page;
574 mutex_unlock(&online_page_callback_lock);
579 EXPORT_SYMBOL_GPL(restore_online_page_callback);
581 void generic_online_page(struct page *page, unsigned int order)
584 * Freeing the page with debug_pagealloc enabled will try to unmap it,
585 * so we should map it first. This is better than introducing a special
586 * case in page freeing fast path.
588 debug_pagealloc_map_pages(page, 1 << order);
589 __free_pages_core(page, order);
590 totalram_pages_add(1UL << order);
591 #ifdef CONFIG_HIGHMEM
592 if (PageHighMem(page))
593 totalhigh_pages_add(1UL << order);
596 EXPORT_SYMBOL_GPL(generic_online_page);
598 static void online_pages_range(unsigned long start_pfn, unsigned long nr_pages)
600 const unsigned long end_pfn = start_pfn + nr_pages;
604 * Online the pages in MAX_ORDER - 1 aligned chunks. The callback might
605 * decide to not expose all pages to the buddy (e.g., expose them
606 * later). We account all pages as being online and belonging to this
608 * When using memmap_on_memory, the range might not be aligned to
609 * MAX_ORDER_NR_PAGES - 1, but pageblock aligned. __ffs() will detect
610 * this and the first chunk to online will be pageblock_nr_pages.
612 for (pfn = start_pfn; pfn < end_pfn;) {
613 int order = min(MAX_ORDER - 1UL, __ffs(pfn));
615 (*online_page_callback)(pfn_to_page(pfn), order);
616 pfn += (1UL << order);
619 /* mark all involved sections as online */
620 online_mem_sections(start_pfn, end_pfn);
623 /* check which state of node_states will be changed when online memory */
624 static void node_states_check_changes_online(unsigned long nr_pages,
625 struct zone *zone, struct memory_notify *arg)
627 int nid = zone_to_nid(zone);
629 arg->status_change_nid = NUMA_NO_NODE;
630 arg->status_change_nid_normal = NUMA_NO_NODE;
631 arg->status_change_nid_high = NUMA_NO_NODE;
633 if (!node_state(nid, N_MEMORY))
634 arg->status_change_nid = nid;
635 if (zone_idx(zone) <= ZONE_NORMAL && !node_state(nid, N_NORMAL_MEMORY))
636 arg->status_change_nid_normal = nid;
637 #ifdef CONFIG_HIGHMEM
638 if (zone_idx(zone) <= ZONE_HIGHMEM && !node_state(nid, N_HIGH_MEMORY))
639 arg->status_change_nid_high = nid;
643 static void node_states_set_node(int node, struct memory_notify *arg)
645 if (arg->status_change_nid_normal >= 0)
646 node_set_state(node, N_NORMAL_MEMORY);
648 if (arg->status_change_nid_high >= 0)
649 node_set_state(node, N_HIGH_MEMORY);
651 if (arg->status_change_nid >= 0)
652 node_set_state(node, N_MEMORY);
655 static void __meminit resize_zone_range(struct zone *zone, unsigned long start_pfn,
656 unsigned long nr_pages)
658 unsigned long old_end_pfn = zone_end_pfn(zone);
660 if (zone_is_empty(zone) || start_pfn < zone->zone_start_pfn)
661 zone->zone_start_pfn = start_pfn;
663 zone->spanned_pages = max(start_pfn + nr_pages, old_end_pfn) - zone->zone_start_pfn;
666 static void __meminit resize_pgdat_range(struct pglist_data *pgdat, unsigned long start_pfn,
667 unsigned long nr_pages)
669 unsigned long old_end_pfn = pgdat_end_pfn(pgdat);
671 if (!pgdat->node_spanned_pages || start_pfn < pgdat->node_start_pfn)
672 pgdat->node_start_pfn = start_pfn;
674 pgdat->node_spanned_pages = max(start_pfn + nr_pages, old_end_pfn) - pgdat->node_start_pfn;
678 static void section_taint_zone_device(unsigned long pfn)
680 struct mem_section *ms = __pfn_to_section(pfn);
682 ms->section_mem_map |= SECTION_TAINT_ZONE_DEVICE;
686 * Associate the pfn range with the given zone, initializing the memmaps
687 * and resizing the pgdat/zone data to span the added pages. After this
688 * call, all affected pages are PG_reserved.
690 * All aligned pageblocks are initialized to the specified migratetype
691 * (usually MIGRATE_MOVABLE). Besides setting the migratetype, no related
692 * zone stats (e.g., nr_isolate_pageblock) are touched.
694 void __ref move_pfn_range_to_zone(struct zone *zone, unsigned long start_pfn,
695 unsigned long nr_pages,
696 struct vmem_altmap *altmap, int migratetype)
698 struct pglist_data *pgdat = zone->zone_pgdat;
699 int nid = pgdat->node_id;
701 clear_zone_contiguous(zone);
703 if (zone_is_empty(zone))
704 init_currently_empty_zone(zone, start_pfn, nr_pages);
705 resize_zone_range(zone, start_pfn, nr_pages);
706 resize_pgdat_range(pgdat, start_pfn, nr_pages);
709 * Subsection population requires care in pfn_to_online_page().
710 * Set the taint to enable the slow path detection of
711 * ZONE_DEVICE pages in an otherwise ZONE_{NORMAL,MOVABLE}
714 if (zone_is_zone_device(zone)) {
715 if (!IS_ALIGNED(start_pfn, PAGES_PER_SECTION))
716 section_taint_zone_device(start_pfn);
717 if (!IS_ALIGNED(start_pfn + nr_pages, PAGES_PER_SECTION))
718 section_taint_zone_device(start_pfn + nr_pages);
722 * TODO now we have a visible range of pages which are not associated
723 * with their zone properly. Not nice but set_pfnblock_flags_mask
724 * expects the zone spans the pfn range. All the pages in the range
725 * are reserved so nobody should be touching them so we should be safe
727 memmap_init_range(nr_pages, nid, zone_idx(zone), start_pfn, 0,
728 MEMINIT_HOTPLUG, altmap, migratetype);
730 set_zone_contiguous(zone);
733 struct auto_movable_stats {
734 unsigned long kernel_early_pages;
735 unsigned long movable_pages;
738 static void auto_movable_stats_account_zone(struct auto_movable_stats *stats,
741 if (zone_idx(zone) == ZONE_MOVABLE) {
742 stats->movable_pages += zone->present_pages;
744 stats->kernel_early_pages += zone->present_early_pages;
747 * CMA pages (never on hotplugged memory) behave like
750 stats->movable_pages += zone->cma_pages;
751 stats->kernel_early_pages -= zone->cma_pages;
752 #endif /* CONFIG_CMA */
756 static bool auto_movable_can_online_movable(int nid, unsigned long nr_pages)
758 struct auto_movable_stats stats = {};
759 unsigned long kernel_early_pages, movable_pages;
760 pg_data_t *pgdat = NODE_DATA(nid);
764 /* Walk all relevant zones and collect MOVABLE vs. KERNEL stats. */
765 if (nid == NUMA_NO_NODE) {
766 /* TODO: cache values */
767 for_each_populated_zone(zone)
768 auto_movable_stats_account_zone(&stats, zone);
770 for (i = 0; i < MAX_NR_ZONES; i++) {
771 zone = pgdat->node_zones + i;
772 if (populated_zone(zone))
773 auto_movable_stats_account_zone(&stats, zone);
777 kernel_early_pages = stats.kernel_early_pages;
778 movable_pages = stats.movable_pages;
781 * Test if we could online the given number of pages to ZONE_MOVABLE
782 * and still stay in the configured ratio.
784 movable_pages += nr_pages;
785 return movable_pages <= (auto_movable_ratio * kernel_early_pages) / 100;
789 * Returns a default kernel memory zone for the given pfn range.
790 * If no kernel zone covers this pfn range it will automatically go
791 * to the ZONE_NORMAL.
793 static struct zone *default_kernel_zone_for_pfn(int nid, unsigned long start_pfn,
794 unsigned long nr_pages)
796 struct pglist_data *pgdat = NODE_DATA(nid);
799 for (zid = 0; zid <= ZONE_NORMAL; zid++) {
800 struct zone *zone = &pgdat->node_zones[zid];
802 if (zone_intersects(zone, start_pfn, nr_pages))
806 return &pgdat->node_zones[ZONE_NORMAL];
810 * Determine to which zone to online memory dynamically based on user
811 * configuration and system stats. We care about the following ratio:
815 * Whereby MOVABLE is memory in ZONE_MOVABLE and KERNEL is memory in
816 * one of the kernel zones. CMA pages inside one of the kernel zones really
817 * behaves like ZONE_MOVABLE, so we treat them accordingly.
819 * We don't allow for hotplugged memory in a KERNEL zone to increase the
820 * amount of MOVABLE memory we can have, so we end up with:
822 * MOVABLE : KERNEL_EARLY
824 * Whereby KERNEL_EARLY is memory in one of the kernel zones, available sinze
825 * boot. We base our calculation on KERNEL_EARLY internally, because:
827 * a) Hotplugged memory in one of the kernel zones can sometimes still get
828 * hotunplugged, especially when hot(un)plugging individual memory blocks.
829 * There is no coordination across memory devices, therefore "automatic"
830 * hotunplugging, as implemented in hypervisors, could result in zone
832 * b) Early/boot memory in one of the kernel zones can usually not get
833 * hotunplugged again (e.g., no firmware interface to unplug, fragmented
834 * with unmovable allocations). While there are corner cases where it might
835 * still work, it is barely relevant in practice.
837 * We rely on "present pages" instead of "managed pages", as the latter is
838 * highly unreliable and dynamic in virtualized environments, and does not
839 * consider boot time allocations. For example, memory ballooning adjusts the
840 * managed pages when inflating/deflating the balloon, and balloon compaction
841 * can even migrate inflated pages between zones.
843 * Using "present pages" is better but some things to keep in mind are:
845 * a) Some memblock allocations, such as for the crashkernel area, are
846 * effectively unused by the kernel, yet they account to "present pages".
847 * Fortunately, these allocations are comparatively small in relevant setups
848 * (e.g., fraction of system memory).
849 * b) Some hotplugged memory blocks in virtualized environments, esecially
850 * hotplugged by virtio-mem, look like they are completely present, however,
851 * only parts of the memory block are actually currently usable.
852 * "present pages" is an upper limit that can get reached at runtime. As
853 * we base our calculations on KERNEL_EARLY, this is not an issue.
855 static struct zone *auto_movable_zone_for_pfn(int nid,
856 struct memory_group *group,
858 unsigned long nr_pages)
860 unsigned long online_pages = 0, max_pages, end_pfn;
863 if (!auto_movable_ratio)
866 if (group && !group->is_dynamic) {
867 max_pages = group->s.max_pages;
868 online_pages = group->present_movable_pages;
870 /* If anything is !MOVABLE online the rest !MOVABLE. */
871 if (group->present_kernel_pages)
873 } else if (!group || group->d.unit_pages == nr_pages) {
874 max_pages = nr_pages;
876 max_pages = group->d.unit_pages;
878 * Take a look at all online sections in the current unit.
879 * We can safely assume that all pages within a section belong
880 * to the same zone, because dynamic memory groups only deal
881 * with hotplugged memory.
883 pfn = ALIGN_DOWN(pfn, group->d.unit_pages);
884 end_pfn = pfn + group->d.unit_pages;
885 for (; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
886 page = pfn_to_online_page(pfn);
889 /* If anything is !MOVABLE online the rest !MOVABLE. */
890 if (page_zonenum(page) != ZONE_MOVABLE)
892 online_pages += PAGES_PER_SECTION;
897 * Online MOVABLE if we could *currently* online all remaining parts
898 * MOVABLE. We expect to (add+) online them immediately next, so if
899 * nobody interferes, all will be MOVABLE if possible.
901 nr_pages = max_pages - online_pages;
902 if (!auto_movable_can_online_movable(NUMA_NO_NODE, nr_pages))
906 if (auto_movable_numa_aware &&
907 !auto_movable_can_online_movable(nid, nr_pages))
909 #endif /* CONFIG_NUMA */
911 return &NODE_DATA(nid)->node_zones[ZONE_MOVABLE];
913 return default_kernel_zone_for_pfn(nid, pfn, nr_pages);
916 static inline struct zone *default_zone_for_pfn(int nid, unsigned long start_pfn,
917 unsigned long nr_pages)
919 struct zone *kernel_zone = default_kernel_zone_for_pfn(nid, start_pfn,
921 struct zone *movable_zone = &NODE_DATA(nid)->node_zones[ZONE_MOVABLE];
922 bool in_kernel = zone_intersects(kernel_zone, start_pfn, nr_pages);
923 bool in_movable = zone_intersects(movable_zone, start_pfn, nr_pages);
926 * We inherit the existing zone in a simple case where zones do not
927 * overlap in the given range
929 if (in_kernel ^ in_movable)
930 return (in_kernel) ? kernel_zone : movable_zone;
933 * If the range doesn't belong to any zone or two zones overlap in the
934 * given range then we use movable zone only if movable_node is
935 * enabled because we always online to a kernel zone by default.
937 return movable_node_enabled ? movable_zone : kernel_zone;
940 struct zone *zone_for_pfn_range(int online_type, int nid,
941 struct memory_group *group, unsigned long start_pfn,
942 unsigned long nr_pages)
944 if (online_type == MMOP_ONLINE_KERNEL)
945 return default_kernel_zone_for_pfn(nid, start_pfn, nr_pages);
947 if (online_type == MMOP_ONLINE_MOVABLE)
948 return &NODE_DATA(nid)->node_zones[ZONE_MOVABLE];
950 if (online_policy == ONLINE_POLICY_AUTO_MOVABLE)
951 return auto_movable_zone_for_pfn(nid, group, start_pfn, nr_pages);
953 return default_zone_for_pfn(nid, start_pfn, nr_pages);
957 * This function should only be called by memory_block_{online,offline},
958 * and {online,offline}_pages.
960 void adjust_present_page_count(struct page *page, struct memory_group *group,
963 struct zone *zone = page_zone(page);
964 const bool movable = zone_idx(zone) == ZONE_MOVABLE;
967 * We only support onlining/offlining/adding/removing of complete
968 * memory blocks; therefore, either all is either early or hotplugged.
970 if (early_section(__pfn_to_section(page_to_pfn(page))))
971 zone->present_early_pages += nr_pages;
972 zone->present_pages += nr_pages;
973 zone->zone_pgdat->node_present_pages += nr_pages;
975 if (group && movable)
976 group->present_movable_pages += nr_pages;
977 else if (group && !movable)
978 group->present_kernel_pages += nr_pages;
981 int mhp_init_memmap_on_memory(unsigned long pfn, unsigned long nr_pages,
984 unsigned long end_pfn = pfn + nr_pages;
987 ret = kasan_add_zero_shadow(__va(PFN_PHYS(pfn)), PFN_PHYS(nr_pages));
991 move_pfn_range_to_zone(zone, pfn, nr_pages, NULL, MIGRATE_UNMOVABLE);
994 * It might be that the vmemmap_pages fully span sections. If that is
995 * the case, mark those sections online here as otherwise they will be
998 if (nr_pages >= PAGES_PER_SECTION)
999 online_mem_sections(pfn, ALIGN_DOWN(end_pfn, PAGES_PER_SECTION));
1004 void mhp_deinit_memmap_on_memory(unsigned long pfn, unsigned long nr_pages)
1006 unsigned long end_pfn = pfn + nr_pages;
1009 * It might be that the vmemmap_pages fully span sections. If that is
1010 * the case, mark those sections offline here as otherwise they will be
1013 if (nr_pages >= PAGES_PER_SECTION)
1014 offline_mem_sections(pfn, ALIGN_DOWN(end_pfn, PAGES_PER_SECTION));
1017 * The pages associated with this vmemmap have been offlined, so
1018 * we can reset its state here.
1020 remove_pfn_range_from_zone(page_zone(pfn_to_page(pfn)), pfn, nr_pages);
1021 kasan_remove_zero_shadow(__va(PFN_PHYS(pfn)), PFN_PHYS(nr_pages));
1024 int __ref online_pages(unsigned long pfn, unsigned long nr_pages,
1025 struct zone *zone, struct memory_group *group)
1027 unsigned long flags;
1028 int need_zonelists_rebuild = 0;
1029 const int nid = zone_to_nid(zone);
1031 struct memory_notify arg;
1034 * {on,off}lining is constrained to full memory sections (or more
1035 * precisely to memory blocks from the user space POV).
1036 * memmap_on_memory is an exception because it reserves initial part
1037 * of the physical memory space for vmemmaps. That space is pageblock
1040 if (WARN_ON_ONCE(!nr_pages ||
1041 !IS_ALIGNED(pfn, pageblock_nr_pages) ||
1042 !IS_ALIGNED(pfn + nr_pages, PAGES_PER_SECTION)))
1045 mem_hotplug_begin();
1047 /* associate pfn range with the zone */
1048 move_pfn_range_to_zone(zone, pfn, nr_pages, NULL, MIGRATE_ISOLATE);
1050 arg.start_pfn = pfn;
1051 arg.nr_pages = nr_pages;
1052 node_states_check_changes_online(nr_pages, zone, &arg);
1054 ret = memory_notify(MEM_GOING_ONLINE, &arg);
1055 ret = notifier_to_errno(ret);
1057 goto failed_addition;
1060 * Fixup the number of isolated pageblocks before marking the sections
1061 * onlining, such that undo_isolate_page_range() works correctly.
1063 spin_lock_irqsave(&zone->lock, flags);
1064 zone->nr_isolate_pageblock += nr_pages / pageblock_nr_pages;
1065 spin_unlock_irqrestore(&zone->lock, flags);
1068 * If this zone is not populated, then it is not in zonelist.
1069 * This means the page allocator ignores this zone.
1070 * So, zonelist must be updated after online.
1072 if (!populated_zone(zone)) {
1073 need_zonelists_rebuild = 1;
1074 setup_zone_pageset(zone);
1077 online_pages_range(pfn, nr_pages);
1078 adjust_present_page_count(pfn_to_page(pfn), group, nr_pages);
1080 node_states_set_node(nid, &arg);
1081 if (need_zonelists_rebuild)
1082 build_all_zonelists(NULL);
1084 /* Basic onlining is complete, allow allocation of onlined pages. */
1085 undo_isolate_page_range(pfn, pfn + nr_pages, MIGRATE_MOVABLE);
1088 * Freshly onlined pages aren't shuffled (e.g., all pages are placed to
1089 * the tail of the freelist when undoing isolation). Shuffle the whole
1090 * zone to make sure the just onlined pages are properly distributed
1091 * across the whole freelist - to create an initial shuffle.
1095 /* reinitialise watermarks and update pcp limits */
1096 init_per_zone_wmark_min();
1101 writeback_set_ratelimit();
1103 memory_notify(MEM_ONLINE, &arg);
1108 pr_debug("online_pages [mem %#010llx-%#010llx] failed\n",
1109 (unsigned long long) pfn << PAGE_SHIFT,
1110 (((unsigned long long) pfn + nr_pages) << PAGE_SHIFT) - 1);
1111 memory_notify(MEM_CANCEL_ONLINE, &arg);
1112 remove_pfn_range_from_zone(zone, pfn, nr_pages);
1116 #endif /* CONFIG_MEMORY_HOTPLUG_SPARSE */
1118 static void reset_node_present_pages(pg_data_t *pgdat)
1122 for (z = pgdat->node_zones; z < pgdat->node_zones + MAX_NR_ZONES; z++)
1123 z->present_pages = 0;
1125 pgdat->node_present_pages = 0;
1128 /* we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG */
1129 static pg_data_t __ref *hotadd_new_pgdat(int nid)
1131 struct pglist_data *pgdat;
1133 pgdat = NODE_DATA(nid);
1135 pgdat = arch_alloc_nodedata(nid);
1139 pgdat->per_cpu_nodestats =
1140 alloc_percpu(struct per_cpu_nodestat);
1141 arch_refresh_nodedata(nid, pgdat);
1145 * Reset the nr_zones, order and highest_zoneidx before reuse.
1146 * Note that kswapd will init kswapd_highest_zoneidx properly
1147 * when it starts in the near future.
1149 pgdat->nr_zones = 0;
1150 pgdat->kswapd_order = 0;
1151 pgdat->kswapd_highest_zoneidx = 0;
1152 for_each_online_cpu(cpu) {
1153 struct per_cpu_nodestat *p;
1155 p = per_cpu_ptr(pgdat->per_cpu_nodestats, cpu);
1156 memset(p, 0, sizeof(*p));
1160 /* we can use NODE_DATA(nid) from here */
1161 pgdat->node_id = nid;
1162 pgdat->node_start_pfn = 0;
1164 /* init node's zones as empty zones, we don't have any present pages.*/
1165 free_area_init_core_hotplug(nid);
1168 * The node we allocated has no zone fallback lists. For avoiding
1169 * to access not-initialized zonelist, build here.
1171 build_all_zonelists(pgdat);
1174 * When memory is hot-added, all the memory is in offline state. So
1175 * clear all zones' present_pages because they will be updated in
1176 * online_pages() and offline_pages().
1178 reset_node_managed_pages(pgdat);
1179 reset_node_present_pages(pgdat);
1184 static void rollback_node_hotadd(int nid)
1186 pg_data_t *pgdat = NODE_DATA(nid);
1188 arch_refresh_nodedata(nid, NULL);
1189 free_percpu(pgdat->per_cpu_nodestats);
1190 arch_free_nodedata(pgdat);
1195 * __try_online_node - online a node if offlined
1197 * @set_node_online: Whether we want to online the node
1198 * called by cpu_up() to online a node without onlined memory.
1201 * 1 -> a new node has been allocated
1202 * 0 -> the node is already online
1203 * -ENOMEM -> the node could not be allocated
1205 static int __try_online_node(int nid, bool set_node_online)
1210 if (node_online(nid))
1213 pgdat = hotadd_new_pgdat(nid);
1215 pr_err("Cannot online node %d due to NULL pgdat\n", nid);
1220 if (set_node_online) {
1221 node_set_online(nid);
1222 ret = register_one_node(nid);
1230 * Users of this function always want to online/register the node
1232 int try_online_node(int nid)
1236 mem_hotplug_begin();
1237 ret = __try_online_node(nid, true);
1242 static int check_hotplug_memory_range(u64 start, u64 size)
1244 /* memory range must be block size aligned */
1245 if (!size || !IS_ALIGNED(start, memory_block_size_bytes()) ||
1246 !IS_ALIGNED(size, memory_block_size_bytes())) {
1247 pr_err("Block size [%#lx] unaligned hotplug range: start %#llx, size %#llx",
1248 memory_block_size_bytes(), start, size);
1255 static int online_memory_block(struct memory_block *mem, void *arg)
1257 mem->online_type = mhp_default_online_type;
1258 return device_online(&mem->dev);
1261 bool mhp_supports_memmap_on_memory(unsigned long size)
1263 unsigned long nr_vmemmap_pages = size / PAGE_SIZE;
1264 unsigned long vmemmap_size = nr_vmemmap_pages * sizeof(struct page);
1265 unsigned long remaining_size = size - vmemmap_size;
1268 * Besides having arch support and the feature enabled at runtime, we
1269 * need a few more assumptions to hold true:
1271 * a) We span a single memory block: memory onlining/offlinin;g happens
1272 * in memory block granularity. We don't want the vmemmap of online
1273 * memory blocks to reside on offline memory blocks. In the future,
1274 * we might want to support variable-sized memory blocks to make the
1275 * feature more versatile.
1277 * b) The vmemmap pages span complete PMDs: We don't want vmemmap code
1278 * to populate memory from the altmap for unrelated parts (i.e.,
1279 * other memory blocks)
1281 * c) The vmemmap pages (and thereby the pages that will be exposed to
1282 * the buddy) have to cover full pageblocks: memory onlining/offlining
1283 * code requires applicable ranges to be page-aligned, for example, to
1284 * set the migratetypes properly.
1286 * TODO: Although we have a check here to make sure that vmemmap pages
1287 * fully populate a PMD, it is not the right place to check for
1288 * this. A much better solution involves improving vmemmap code
1289 * to fallback to base pages when trying to populate vmemmap using
1290 * altmap as an alternative source of memory, and we do not exactly
1291 * populate a single PMD.
1293 return memmap_on_memory &&
1294 !hugetlb_free_vmemmap_enabled &&
1295 IS_ENABLED(CONFIG_MHP_MEMMAP_ON_MEMORY) &&
1296 size == memory_block_size_bytes() &&
1297 IS_ALIGNED(vmemmap_size, PMD_SIZE) &&
1298 IS_ALIGNED(remaining_size, (pageblock_nr_pages << PAGE_SHIFT));
1302 * NOTE: The caller must call lock_device_hotplug() to serialize hotplug
1303 * and online/offline operations (triggered e.g. by sysfs).
1305 * we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG
1307 int __ref add_memory_resource(int nid, struct resource *res, mhp_t mhp_flags)
1309 struct mhp_params params = { .pgprot = pgprot_mhp(PAGE_KERNEL) };
1310 struct vmem_altmap mhp_altmap = {};
1311 struct memory_group *group = NULL;
1313 bool new_node = false;
1317 size = resource_size(res);
1319 ret = check_hotplug_memory_range(start, size);
1323 if (mhp_flags & MHP_NID_IS_MGID) {
1324 group = memory_group_find_by_id(nid);
1330 if (!node_possible(nid)) {
1331 WARN(1, "node %d was absent from the node_possible_map\n", nid);
1335 mem_hotplug_begin();
1337 if (IS_ENABLED(CONFIG_ARCH_KEEP_MEMBLOCK))
1338 memblock_add_node(start, size, nid);
1340 ret = __try_online_node(nid, false);
1346 * Self hosted memmap array
1348 if (mhp_flags & MHP_MEMMAP_ON_MEMORY) {
1349 if (!mhp_supports_memmap_on_memory(size)) {
1353 mhp_altmap.free = PHYS_PFN(size);
1354 mhp_altmap.base_pfn = PHYS_PFN(start);
1355 params.altmap = &mhp_altmap;
1358 /* call arch's memory hotadd */
1359 ret = arch_add_memory(nid, start, size, ¶ms);
1363 /* create memory block devices after memory was added */
1364 ret = create_memory_block_devices(start, size, mhp_altmap.alloc,
1367 arch_remove_memory(start, size, NULL);
1372 /* If sysfs file of new node can't be created, cpu on the node
1373 * can't be hot-added. There is no rollback way now.
1374 * So, check by BUG_ON() to catch it reluctantly..
1375 * We online node here. We can't roll back from here.
1377 node_set_online(nid);
1378 ret = __register_one_node(nid);
1382 /* link memory sections under this node.*/
1383 link_mem_sections(nid, PFN_DOWN(start), PFN_UP(start + size - 1),
1386 /* create new memmap entry */
1387 if (!strcmp(res->name, "System RAM"))
1388 firmware_map_add_hotplug(start, start + size, "System RAM");
1390 /* device_online() will take the lock when calling online_pages() */
1394 * In case we're allowed to merge the resource, flag it and trigger
1395 * merging now that adding succeeded.
1397 if (mhp_flags & MHP_MERGE_RESOURCE)
1398 merge_system_ram_resource(res);
1400 /* online pages if requested */
1401 if (mhp_default_online_type != MMOP_OFFLINE)
1402 walk_memory_blocks(start, size, NULL, online_memory_block);
1406 /* rollback pgdat allocation and others */
1408 rollback_node_hotadd(nid);
1409 if (IS_ENABLED(CONFIG_ARCH_KEEP_MEMBLOCK))
1410 memblock_remove(start, size);
1415 /* requires device_hotplug_lock, see add_memory_resource() */
1416 int __ref __add_memory(int nid, u64 start, u64 size, mhp_t mhp_flags)
1418 struct resource *res;
1421 res = register_memory_resource(start, size, "System RAM");
1423 return PTR_ERR(res);
1425 ret = add_memory_resource(nid, res, mhp_flags);
1427 release_memory_resource(res);
1431 int add_memory(int nid, u64 start, u64 size, mhp_t mhp_flags)
1435 lock_device_hotplug();
1436 rc = __add_memory(nid, start, size, mhp_flags);
1437 unlock_device_hotplug();
1441 EXPORT_SYMBOL_GPL(add_memory);
1444 * Add special, driver-managed memory to the system as system RAM. Such
1445 * memory is not exposed via the raw firmware-provided memmap as system
1446 * RAM, instead, it is detected and added by a driver - during cold boot,
1447 * after a reboot, and after kexec.
1449 * Reasons why this memory should not be used for the initial memmap of a
1450 * kexec kernel or for placing kexec images:
1451 * - The booting kernel is in charge of determining how this memory will be
1452 * used (e.g., use persistent memory as system RAM)
1453 * - Coordination with a hypervisor is required before this memory
1454 * can be used (e.g., inaccessible parts).
1456 * For this memory, no entries in /sys/firmware/memmap ("raw firmware-provided
1457 * memory map") are created. Also, the created memory resource is flagged
1458 * with IORESOURCE_SYSRAM_DRIVER_MANAGED, so in-kernel users can special-case
1459 * this memory as well (esp., not place kexec images onto it).
1461 * The resource_name (visible via /proc/iomem) has to have the format
1462 * "System RAM ($DRIVER)".
1464 int add_memory_driver_managed(int nid, u64 start, u64 size,
1465 const char *resource_name, mhp_t mhp_flags)
1467 struct resource *res;
1470 if (!resource_name ||
1471 strstr(resource_name, "System RAM (") != resource_name ||
1472 resource_name[strlen(resource_name) - 1] != ')')
1475 lock_device_hotplug();
1477 res = register_memory_resource(start, size, resource_name);
1483 rc = add_memory_resource(nid, res, mhp_flags);
1485 release_memory_resource(res);
1488 unlock_device_hotplug();
1491 EXPORT_SYMBOL_GPL(add_memory_driver_managed);
1494 * Platforms should define arch_get_mappable_range() that provides
1495 * maximum possible addressable physical memory range for which the
1496 * linear mapping could be created. The platform returned address
1497 * range must adhere to these following semantics.
1499 * - range.start <= range.end
1500 * - Range includes both end points [range.start..range.end]
1502 * There is also a fallback definition provided here, allowing the
1503 * entire possible physical address range in case any platform does
1504 * not define arch_get_mappable_range().
1506 struct range __weak arch_get_mappable_range(void)
1508 struct range mhp_range = {
1515 struct range mhp_get_pluggable_range(bool need_mapping)
1517 const u64 max_phys = (1ULL << MAX_PHYSMEM_BITS) - 1;
1518 struct range mhp_range;
1521 mhp_range = arch_get_mappable_range();
1522 if (mhp_range.start > max_phys) {
1523 mhp_range.start = 0;
1526 mhp_range.end = min_t(u64, mhp_range.end, max_phys);
1528 mhp_range.start = 0;
1529 mhp_range.end = max_phys;
1533 EXPORT_SYMBOL_GPL(mhp_get_pluggable_range);
1535 bool mhp_range_allowed(u64 start, u64 size, bool need_mapping)
1537 struct range mhp_range = mhp_get_pluggable_range(need_mapping);
1538 u64 end = start + size;
1540 if (start < end && start >= mhp_range.start && (end - 1) <= mhp_range.end)
1543 pr_warn("Hotplug memory [%#llx-%#llx] exceeds maximum addressable range [%#llx-%#llx]\n",
1544 start, end, mhp_range.start, mhp_range.end);
1548 #ifdef CONFIG_MEMORY_HOTREMOVE
1550 * Confirm all pages in a range [start, end) belong to the same zone (skipping
1551 * memory holes). When true, return the zone.
1553 struct zone *test_pages_in_a_zone(unsigned long start_pfn,
1554 unsigned long end_pfn)
1556 unsigned long pfn, sec_end_pfn;
1557 struct zone *zone = NULL;
1560 for (pfn = start_pfn, sec_end_pfn = SECTION_ALIGN_UP(start_pfn + 1);
1562 pfn = sec_end_pfn, sec_end_pfn += PAGES_PER_SECTION) {
1563 /* Make sure the memory section is present first */
1564 if (!present_section_nr(pfn_to_section_nr(pfn)))
1566 for (; pfn < sec_end_pfn && pfn < end_pfn;
1567 pfn += MAX_ORDER_NR_PAGES) {
1568 /* Check if we got outside of the zone */
1569 if (zone && !zone_spans_pfn(zone, pfn))
1571 page = pfn_to_page(pfn);
1572 if (zone && page_zone(page) != zone)
1574 zone = page_zone(page);
1582 * Scan pfn range [start,end) to find movable/migratable pages (LRU pages,
1583 * non-lru movable pages and hugepages). Will skip over most unmovable
1584 * pages (esp., pages that can be skipped when offlining), but bail out on
1585 * definitely unmovable pages.
1588 * 0 in case a movable page is found and movable_pfn was updated.
1589 * -ENOENT in case no movable page was found.
1590 * -EBUSY in case a definitely unmovable page was found.
1592 static int scan_movable_pages(unsigned long start, unsigned long end,
1593 unsigned long *movable_pfn)
1597 for (pfn = start; pfn < end; pfn++) {
1598 struct page *page, *head;
1601 if (!pfn_valid(pfn))
1603 page = pfn_to_page(pfn);
1606 if (__PageMovable(page))
1610 * PageOffline() pages that are not marked __PageMovable() and
1611 * have a reference count > 0 (after MEM_GOING_OFFLINE) are
1612 * definitely unmovable. If their reference count would be 0,
1613 * they could at least be skipped when offlining memory.
1615 if (PageOffline(page) && page_count(page))
1618 if (!PageHuge(page))
1620 head = compound_head(page);
1622 * This test is racy as we hold no reference or lock. The
1623 * hugetlb page could have been free'ed and head is no longer
1624 * a hugetlb page before the following check. In such unlikely
1625 * cases false positives and negatives are possible. Calling
1626 * code must deal with these scenarios.
1628 if (HPageMigratable(head))
1630 skip = compound_nr(head) - (page - head);
1640 do_migrate_range(unsigned long start_pfn, unsigned long end_pfn)
1643 struct page *page, *head;
1646 static DEFINE_RATELIMIT_STATE(migrate_rs, DEFAULT_RATELIMIT_INTERVAL,
1647 DEFAULT_RATELIMIT_BURST);
1649 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
1650 if (!pfn_valid(pfn))
1652 page = pfn_to_page(pfn);
1653 head = compound_head(page);
1655 if (PageHuge(page)) {
1656 pfn = page_to_pfn(head) + compound_nr(head) - 1;
1657 isolate_huge_page(head, &source);
1659 } else if (PageTransHuge(page))
1660 pfn = page_to_pfn(head) + thp_nr_pages(page) - 1;
1663 * HWPoison pages have elevated reference counts so the migration would
1664 * fail on them. It also doesn't make any sense to migrate them in the
1665 * first place. Still try to unmap such a page in case it is still mapped
1666 * (e.g. current hwpoison implementation doesn't unmap KSM pages but keep
1667 * the unmap as the catch all safety net).
1669 if (PageHWPoison(page)) {
1670 if (WARN_ON(PageLRU(page)))
1671 isolate_lru_page(page);
1672 if (page_mapped(page))
1673 try_to_unmap(page, TTU_IGNORE_MLOCK);
1677 if (!get_page_unless_zero(page))
1680 * We can skip free pages. And we can deal with pages on
1681 * LRU and non-lru movable pages.
1684 ret = isolate_lru_page(page);
1686 ret = isolate_movable_page(page, ISOLATE_UNEVICTABLE);
1687 if (!ret) { /* Success */
1688 list_add_tail(&page->lru, &source);
1689 if (!__PageMovable(page))
1690 inc_node_page_state(page, NR_ISOLATED_ANON +
1691 page_is_file_lru(page));
1694 if (__ratelimit(&migrate_rs)) {
1695 pr_warn("failed to isolate pfn %lx\n", pfn);
1696 dump_page(page, "isolation failed");
1701 if (!list_empty(&source)) {
1702 nodemask_t nmask = node_states[N_MEMORY];
1703 struct migration_target_control mtc = {
1705 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
1709 * We have checked that migration range is on a single zone so
1710 * we can use the nid of the first page to all the others.
1712 mtc.nid = page_to_nid(list_first_entry(&source, struct page, lru));
1715 * try to allocate from a different node but reuse this node
1716 * if there are no other online nodes to be used (e.g. we are
1717 * offlining a part of the only existing node)
1719 node_clear(mtc.nid, nmask);
1720 if (nodes_empty(nmask))
1721 node_set(mtc.nid, nmask);
1722 ret = migrate_pages(&source, alloc_migration_target, NULL,
1723 (unsigned long)&mtc, MIGRATE_SYNC, MR_MEMORY_HOTPLUG);
1725 list_for_each_entry(page, &source, lru) {
1726 if (__ratelimit(&migrate_rs)) {
1727 pr_warn("migrating pfn %lx failed ret:%d\n",
1728 page_to_pfn(page), ret);
1729 dump_page(page, "migration failure");
1732 putback_movable_pages(&source);
1739 static int __init cmdline_parse_movable_node(char *p)
1741 movable_node_enabled = true;
1744 early_param("movable_node", cmdline_parse_movable_node);
1746 /* check which state of node_states will be changed when offline memory */
1747 static void node_states_check_changes_offline(unsigned long nr_pages,
1748 struct zone *zone, struct memory_notify *arg)
1750 struct pglist_data *pgdat = zone->zone_pgdat;
1751 unsigned long present_pages = 0;
1754 arg->status_change_nid = NUMA_NO_NODE;
1755 arg->status_change_nid_normal = NUMA_NO_NODE;
1756 arg->status_change_nid_high = NUMA_NO_NODE;
1759 * Check whether node_states[N_NORMAL_MEMORY] will be changed.
1760 * If the memory to be offline is within the range
1761 * [0..ZONE_NORMAL], and it is the last present memory there,
1762 * the zones in that range will become empty after the offlining,
1763 * thus we can determine that we need to clear the node from
1764 * node_states[N_NORMAL_MEMORY].
1766 for (zt = 0; zt <= ZONE_NORMAL; zt++)
1767 present_pages += pgdat->node_zones[zt].present_pages;
1768 if (zone_idx(zone) <= ZONE_NORMAL && nr_pages >= present_pages)
1769 arg->status_change_nid_normal = zone_to_nid(zone);
1771 #ifdef CONFIG_HIGHMEM
1773 * node_states[N_HIGH_MEMORY] contains nodes which
1774 * have normal memory or high memory.
1775 * Here we add the present_pages belonging to ZONE_HIGHMEM.
1776 * If the zone is within the range of [0..ZONE_HIGHMEM), and
1777 * we determine that the zones in that range become empty,
1778 * we need to clear the node for N_HIGH_MEMORY.
1780 present_pages += pgdat->node_zones[ZONE_HIGHMEM].present_pages;
1781 if (zone_idx(zone) <= ZONE_HIGHMEM && nr_pages >= present_pages)
1782 arg->status_change_nid_high = zone_to_nid(zone);
1786 * We have accounted the pages from [0..ZONE_NORMAL), and
1787 * in case of CONFIG_HIGHMEM the pages from ZONE_HIGHMEM
1789 * Here we count the possible pages from ZONE_MOVABLE.
1790 * If after having accounted all the pages, we see that the nr_pages
1791 * to be offlined is over or equal to the accounted pages,
1792 * we know that the node will become empty, and so, we can clear
1793 * it for N_MEMORY as well.
1795 present_pages += pgdat->node_zones[ZONE_MOVABLE].present_pages;
1797 if (nr_pages >= present_pages)
1798 arg->status_change_nid = zone_to_nid(zone);
1801 static void node_states_clear_node(int node, struct memory_notify *arg)
1803 if (arg->status_change_nid_normal >= 0)
1804 node_clear_state(node, N_NORMAL_MEMORY);
1806 if (arg->status_change_nid_high >= 0)
1807 node_clear_state(node, N_HIGH_MEMORY);
1809 if (arg->status_change_nid >= 0)
1810 node_clear_state(node, N_MEMORY);
1813 static int count_system_ram_pages_cb(unsigned long start_pfn,
1814 unsigned long nr_pages, void *data)
1816 unsigned long *nr_system_ram_pages = data;
1818 *nr_system_ram_pages += nr_pages;
1822 int __ref offline_pages(unsigned long start_pfn, unsigned long nr_pages,
1823 struct memory_group *group)
1825 const unsigned long end_pfn = start_pfn + nr_pages;
1826 unsigned long pfn, system_ram_pages = 0;
1827 unsigned long flags;
1829 struct memory_notify arg;
1834 * {on,off}lining is constrained to full memory sections (or more
1835 * precisely to memory blocks from the user space POV).
1836 * memmap_on_memory is an exception because it reserves initial part
1837 * of the physical memory space for vmemmaps. That space is pageblock
1840 if (WARN_ON_ONCE(!nr_pages ||
1841 !IS_ALIGNED(start_pfn, pageblock_nr_pages) ||
1842 !IS_ALIGNED(start_pfn + nr_pages, PAGES_PER_SECTION)))
1845 mem_hotplug_begin();
1848 * Don't allow to offline memory blocks that contain holes.
1849 * Consequently, memory blocks with holes can never get onlined
1850 * via the hotplug path - online_pages() - as hotplugged memory has
1851 * no holes. This way, we e.g., don't have to worry about marking
1852 * memory holes PG_reserved, don't need pfn_valid() checks, and can
1853 * avoid using walk_system_ram_range() later.
1855 walk_system_ram_range(start_pfn, nr_pages, &system_ram_pages,
1856 count_system_ram_pages_cb);
1857 if (system_ram_pages != nr_pages) {
1859 reason = "memory holes";
1860 goto failed_removal;
1863 /* This makes hotplug much easier...and readable.
1864 we assume this for now. .*/
1865 zone = test_pages_in_a_zone(start_pfn, end_pfn);
1868 reason = "multizone range";
1869 goto failed_removal;
1871 node = zone_to_nid(zone);
1874 * Disable pcplists so that page isolation cannot race with freeing
1875 * in a way that pages from isolated pageblock are left on pcplists.
1877 zone_pcp_disable(zone);
1878 lru_cache_disable();
1880 /* set above range as isolated */
1881 ret = start_isolate_page_range(start_pfn, end_pfn,
1883 MEMORY_OFFLINE | REPORT_FAILURE);
1885 reason = "failure to isolate range";
1886 goto failed_removal_pcplists_disabled;
1889 arg.start_pfn = start_pfn;
1890 arg.nr_pages = nr_pages;
1891 node_states_check_changes_offline(nr_pages, zone, &arg);
1893 ret = memory_notify(MEM_GOING_OFFLINE, &arg);
1894 ret = notifier_to_errno(ret);
1896 reason = "notifier failure";
1897 goto failed_removal_isolated;
1903 if (signal_pending(current)) {
1905 reason = "signal backoff";
1906 goto failed_removal_isolated;
1911 ret = scan_movable_pages(pfn, end_pfn, &pfn);
1914 * TODO: fatal migration failures should bail
1917 do_migrate_range(pfn, end_pfn);
1921 if (ret != -ENOENT) {
1922 reason = "unmovable page";
1923 goto failed_removal_isolated;
1927 * Dissolve free hugepages in the memory block before doing
1928 * offlining actually in order to make hugetlbfs's object
1929 * counting consistent.
1931 ret = dissolve_free_huge_pages(start_pfn, end_pfn);
1933 reason = "failure to dissolve huge pages";
1934 goto failed_removal_isolated;
1937 ret = test_pages_isolated(start_pfn, end_pfn, MEMORY_OFFLINE);
1941 /* Mark all sections offline and remove free pages from the buddy. */
1942 __offline_isolated_pages(start_pfn, end_pfn);
1943 pr_debug("Offlined Pages %ld\n", nr_pages);
1946 * The memory sections are marked offline, and the pageblock flags
1947 * effectively stale; nobody should be touching them. Fixup the number
1948 * of isolated pageblocks, memory onlining will properly revert this.
1950 spin_lock_irqsave(&zone->lock, flags);
1951 zone->nr_isolate_pageblock -= nr_pages / pageblock_nr_pages;
1952 spin_unlock_irqrestore(&zone->lock, flags);
1955 zone_pcp_enable(zone);
1957 /* removal success */
1958 adjust_managed_page_count(pfn_to_page(start_pfn), -nr_pages);
1959 adjust_present_page_count(pfn_to_page(start_pfn), group, -nr_pages);
1961 /* reinitialise watermarks and update pcp limits */
1962 init_per_zone_wmark_min();
1964 if (!populated_zone(zone)) {
1965 zone_pcp_reset(zone);
1966 build_all_zonelists(NULL);
1969 node_states_clear_node(node, &arg);
1970 if (arg.status_change_nid >= 0) {
1972 kcompactd_stop(node);
1975 writeback_set_ratelimit();
1977 memory_notify(MEM_OFFLINE, &arg);
1978 remove_pfn_range_from_zone(zone, start_pfn, nr_pages);
1982 failed_removal_isolated:
1983 undo_isolate_page_range(start_pfn, end_pfn, MIGRATE_MOVABLE);
1984 memory_notify(MEM_CANCEL_OFFLINE, &arg);
1985 failed_removal_pcplists_disabled:
1987 zone_pcp_enable(zone);
1989 pr_debug("memory offlining [mem %#010llx-%#010llx] failed due to %s\n",
1990 (unsigned long long) start_pfn << PAGE_SHIFT,
1991 ((unsigned long long) end_pfn << PAGE_SHIFT) - 1,
1993 /* pushback to free area */
1998 static int check_memblock_offlined_cb(struct memory_block *mem, void *arg)
2000 int ret = !is_memblock_offlined(mem);
2004 if (unlikely(ret)) {
2005 phys_addr_t beginpa, endpa;
2007 beginpa = PFN_PHYS(section_nr_to_pfn(mem->start_section_nr));
2008 endpa = beginpa + memory_block_size_bytes() - 1;
2009 pr_warn("removing memory fails, because memory [%pa-%pa] is onlined\n",
2017 static int get_nr_vmemmap_pages_cb(struct memory_block *mem, void *arg)
2020 * If not set, continue with the next block.
2022 return mem->nr_vmemmap_pages;
2025 static int check_cpu_on_node(pg_data_t *pgdat)
2029 for_each_present_cpu(cpu) {
2030 if (cpu_to_node(cpu) == pgdat->node_id)
2032 * the cpu on this node isn't removed, and we can't
2033 * offline this node.
2041 static int check_no_memblock_for_node_cb(struct memory_block *mem, void *arg)
2043 int nid = *(int *)arg;
2046 * If a memory block belongs to multiple nodes, the stored nid is not
2047 * reliable. However, such blocks are always online (e.g., cannot get
2048 * offlined) and, therefore, are still spanned by the node.
2050 return mem->nid == nid ? -EEXIST : 0;
2057 * Offline a node if all memory sections and cpus of the node are removed.
2059 * NOTE: The caller must call lock_device_hotplug() to serialize hotplug
2060 * and online/offline operations before this call.
2062 void try_offline_node(int nid)
2064 pg_data_t *pgdat = NODE_DATA(nid);
2068 * If the node still spans pages (especially ZONE_DEVICE), don't
2069 * offline it. A node spans memory after move_pfn_range_to_zone(),
2070 * e.g., after the memory block was onlined.
2072 if (pgdat->node_spanned_pages)
2076 * Especially offline memory blocks might not be spanned by the
2077 * node. They will get spanned by the node once they get onlined.
2078 * However, they link to the node in sysfs and can get onlined later.
2080 rc = for_each_memory_block(&nid, check_no_memblock_for_node_cb);
2084 if (check_cpu_on_node(pgdat))
2088 * all memory/cpu of this node are removed, we can offline this
2091 node_set_offline(nid);
2092 unregister_one_node(nid);
2094 EXPORT_SYMBOL(try_offline_node);
2096 static int __ref try_remove_memory(u64 start, u64 size)
2098 struct vmem_altmap mhp_altmap = {};
2099 struct vmem_altmap *altmap = NULL;
2100 unsigned long nr_vmemmap_pages;
2101 int rc = 0, nid = NUMA_NO_NODE;
2103 BUG_ON(check_hotplug_memory_range(start, size));
2106 * All memory blocks must be offlined before removing memory. Check
2107 * whether all memory blocks in question are offline and return error
2108 * if this is not the case.
2110 * While at it, determine the nid. Note that if we'd have mixed nodes,
2111 * we'd only try to offline the last determined one -- which is good
2112 * enough for the cases we care about.
2114 rc = walk_memory_blocks(start, size, &nid, check_memblock_offlined_cb);
2119 * We only support removing memory added with MHP_MEMMAP_ON_MEMORY in
2120 * the same granularity it was added - a single memory block.
2122 if (memmap_on_memory) {
2123 nr_vmemmap_pages = walk_memory_blocks(start, size, NULL,
2124 get_nr_vmemmap_pages_cb);
2125 if (nr_vmemmap_pages) {
2126 if (size != memory_block_size_bytes()) {
2127 pr_warn("Refuse to remove %#llx - %#llx,"
2128 "wrong granularity\n",
2129 start, start + size);
2134 * Let remove_pmd_table->free_hugepage_table do the
2135 * right thing if we used vmem_altmap when hot-adding
2138 mhp_altmap.alloc = nr_vmemmap_pages;
2139 altmap = &mhp_altmap;
2143 /* remove memmap entry */
2144 firmware_map_remove(start, start + size, "System RAM");
2147 * Memory block device removal under the device_hotplug_lock is
2148 * a barrier against racing online attempts.
2150 remove_memory_block_devices(start, size);
2152 mem_hotplug_begin();
2154 arch_remove_memory(start, size, altmap);
2156 if (IS_ENABLED(CONFIG_ARCH_KEEP_MEMBLOCK)) {
2157 memblock_free(start, size);
2158 memblock_remove(start, size);
2161 release_mem_region_adjustable(start, size);
2163 if (nid != NUMA_NO_NODE)
2164 try_offline_node(nid);
2171 * __remove_memory - Remove memory if every memory block is offline
2172 * @start: physical address of the region to remove
2173 * @size: size of the region to remove
2175 * NOTE: The caller must call lock_device_hotplug() to serialize hotplug
2176 * and online/offline operations before this call, as required by
2177 * try_offline_node().
2179 void __remove_memory(u64 start, u64 size)
2183 * trigger BUG() if some memory is not offlined prior to calling this
2186 if (try_remove_memory(start, size))
2191 * Remove memory if every memory block is offline, otherwise return -EBUSY is
2192 * some memory is not offline
2194 int remove_memory(u64 start, u64 size)
2198 lock_device_hotplug();
2199 rc = try_remove_memory(start, size);
2200 unlock_device_hotplug();
2204 EXPORT_SYMBOL_GPL(remove_memory);
2206 static int try_offline_memory_block(struct memory_block *mem, void *arg)
2208 uint8_t online_type = MMOP_ONLINE_KERNEL;
2209 uint8_t **online_types = arg;
2214 * Sense the online_type via the zone of the memory block. Offlining
2215 * with multiple zones within one memory block will be rejected
2216 * by offlining code ... so we don't care about that.
2218 page = pfn_to_online_page(section_nr_to_pfn(mem->start_section_nr));
2219 if (page && zone_idx(page_zone(page)) == ZONE_MOVABLE)
2220 online_type = MMOP_ONLINE_MOVABLE;
2222 rc = device_offline(&mem->dev);
2224 * Default is MMOP_OFFLINE - change it only if offlining succeeded,
2225 * so try_reonline_memory_block() can do the right thing.
2228 **online_types = online_type;
2231 /* Ignore if already offline. */
2232 return rc < 0 ? rc : 0;
2235 static int try_reonline_memory_block(struct memory_block *mem, void *arg)
2237 uint8_t **online_types = arg;
2240 if (**online_types != MMOP_OFFLINE) {
2241 mem->online_type = **online_types;
2242 rc = device_online(&mem->dev);
2244 pr_warn("%s: Failed to re-online memory: %d",
2248 /* Continue processing all remaining memory blocks. */
2254 * Try to offline and remove memory. Might take a long time to finish in case
2255 * memory is still in use. Primarily useful for memory devices that logically
2256 * unplugged all memory (so it's no longer in use) and want to offline + remove
2259 int offline_and_remove_memory(u64 start, u64 size)
2261 const unsigned long mb_count = size / memory_block_size_bytes();
2262 uint8_t *online_types, *tmp;
2265 if (!IS_ALIGNED(start, memory_block_size_bytes()) ||
2266 !IS_ALIGNED(size, memory_block_size_bytes()) || !size)
2270 * We'll remember the old online type of each memory block, so we can
2271 * try to revert whatever we did when offlining one memory block fails
2272 * after offlining some others succeeded.
2274 online_types = kmalloc_array(mb_count, sizeof(*online_types),
2279 * Initialize all states to MMOP_OFFLINE, so when we abort processing in
2280 * try_offline_memory_block(), we'll skip all unprocessed blocks in
2281 * try_reonline_memory_block().
2283 memset(online_types, MMOP_OFFLINE, mb_count);
2285 lock_device_hotplug();
2288 rc = walk_memory_blocks(start, size, &tmp, try_offline_memory_block);
2291 * In case we succeeded to offline all memory, remove it.
2292 * This cannot fail as it cannot get onlined in the meantime.
2295 rc = try_remove_memory(start, size);
2297 pr_err("%s: Failed to remove memory: %d", __func__, rc);
2301 * Rollback what we did. While memory onlining might theoretically fail
2302 * (nacked by a notifier), it barely ever happens.
2306 walk_memory_blocks(start, size, &tmp,
2307 try_reonline_memory_block);
2309 unlock_device_hotplug();
2311 kfree(online_types);
2314 EXPORT_SYMBOL_GPL(offline_and_remove_memory);
2315 #endif /* CONFIG_MEMORY_HOTREMOVE */