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/vmalloc.h>
25 #include <linux/ioport.h>
26 #include <linux/delay.h>
27 #include <linux/migrate.h>
28 #include <linux/page-isolation.h>
29 #include <linux/pfn.h>
30 #include <linux/suspend.h>
31 #include <linux/mm_inline.h>
32 #include <linux/firmware-map.h>
33 #include <linux/stop_machine.h>
34 #include <linux/hugetlb.h>
35 #include <linux/memblock.h>
36 #include <linux/compaction.h>
37 #include <linux/rmap.h>
38 #include <linux/module.h>
40 #include <asm/tlbflush.h>
45 #ifdef CONFIG_MHP_MEMMAP_ON_MEMORY
47 * memory_hotplug.memmap_on_memory parameter
49 static bool memmap_on_memory __ro_after_init;
50 module_param(memmap_on_memory, bool, 0444);
51 MODULE_PARM_DESC(memmap_on_memory, "Enable memmap on memory for memory hotplug");
53 static inline bool mhp_memmap_on_memory(void)
55 return memmap_on_memory;
58 static inline bool mhp_memmap_on_memory(void)
65 ONLINE_POLICY_CONTIG_ZONES = 0,
66 ONLINE_POLICY_AUTO_MOVABLE,
69 static const char * const online_policy_to_str[] = {
70 [ONLINE_POLICY_CONTIG_ZONES] = "contig-zones",
71 [ONLINE_POLICY_AUTO_MOVABLE] = "auto-movable",
74 static int set_online_policy(const char *val, const struct kernel_param *kp)
76 int ret = sysfs_match_string(online_policy_to_str, val);
80 *((int *)kp->arg) = ret;
84 static int get_online_policy(char *buffer, const struct kernel_param *kp)
86 return sprintf(buffer, "%s\n", online_policy_to_str[*((int *)kp->arg)]);
90 * memory_hotplug.online_policy: configure online behavior when onlining without
91 * specifying a zone (MMOP_ONLINE)
93 * "contig-zones": keep zone contiguous
94 * "auto-movable": online memory to ZONE_MOVABLE if the configuration
95 * (auto_movable_ratio, auto_movable_numa_aware) allows for it
97 static int online_policy __read_mostly = ONLINE_POLICY_CONTIG_ZONES;
98 static const struct kernel_param_ops online_policy_ops = {
99 .set = set_online_policy,
100 .get = get_online_policy,
102 module_param_cb(online_policy, &online_policy_ops, &online_policy, 0644);
103 MODULE_PARM_DESC(online_policy,
104 "Set the online policy (\"contig-zones\", \"auto-movable\") "
105 "Default: \"contig-zones\"");
108 * memory_hotplug.auto_movable_ratio: specify maximum MOVABLE:KERNEL ratio
110 * The ratio represent an upper limit and the kernel might decide to not
111 * online some memory to ZONE_MOVABLE -- e.g., because hotplugged KERNEL memory
112 * doesn't allow for more MOVABLE memory.
114 static unsigned int auto_movable_ratio __read_mostly = 301;
115 module_param(auto_movable_ratio, uint, 0644);
116 MODULE_PARM_DESC(auto_movable_ratio,
117 "Set the maximum ratio of MOVABLE:KERNEL memory in the system "
118 "in percent for \"auto-movable\" online policy. Default: 301");
121 * memory_hotplug.auto_movable_numa_aware: consider numa node stats
124 static bool auto_movable_numa_aware __read_mostly = true;
125 module_param(auto_movable_numa_aware, bool, 0644);
126 MODULE_PARM_DESC(auto_movable_numa_aware,
127 "Consider numa node stats in addition to global stats in "
128 "\"auto-movable\" online policy. Default: true");
129 #endif /* CONFIG_NUMA */
132 * online_page_callback contains pointer to current page onlining function.
133 * Initially it is generic_online_page(). If it is required it could be
134 * changed by calling set_online_page_callback() for callback registration
135 * and restore_online_page_callback() for generic callback restore.
138 static online_page_callback_t online_page_callback = generic_online_page;
139 static DEFINE_MUTEX(online_page_callback_lock);
141 DEFINE_STATIC_PERCPU_RWSEM(mem_hotplug_lock);
143 void get_online_mems(void)
145 percpu_down_read(&mem_hotplug_lock);
148 void put_online_mems(void)
150 percpu_up_read(&mem_hotplug_lock);
153 bool movable_node_enabled = false;
155 #ifndef CONFIG_MEMORY_HOTPLUG_DEFAULT_ONLINE
156 int mhp_default_online_type = MMOP_OFFLINE;
158 int mhp_default_online_type = MMOP_ONLINE;
161 static int __init setup_memhp_default_state(char *str)
163 const int online_type = mhp_online_type_from_str(str);
165 if (online_type >= 0)
166 mhp_default_online_type = online_type;
170 __setup("memhp_default_state=", setup_memhp_default_state);
172 void mem_hotplug_begin(void)
175 percpu_down_write(&mem_hotplug_lock);
178 void mem_hotplug_done(void)
180 percpu_up_write(&mem_hotplug_lock);
184 u64 max_mem_size = U64_MAX;
186 /* add this memory to iomem resource */
187 static struct resource *register_memory_resource(u64 start, u64 size,
188 const char *resource_name)
190 struct resource *res;
191 unsigned long flags = IORESOURCE_SYSTEM_RAM | IORESOURCE_BUSY;
193 if (strcmp(resource_name, "System RAM"))
194 flags |= IORESOURCE_SYSRAM_DRIVER_MANAGED;
196 if (!mhp_range_allowed(start, size, true))
197 return ERR_PTR(-E2BIG);
200 * Make sure value parsed from 'mem=' only restricts memory adding
201 * while booting, so that memory hotplug won't be impacted. Please
202 * refer to document of 'mem=' in kernel-parameters.txt for more
205 if (start + size > max_mem_size && system_state < SYSTEM_RUNNING)
206 return ERR_PTR(-E2BIG);
209 * Request ownership of the new memory range. This might be
210 * a child of an existing resource that was present but
211 * not marked as busy.
213 res = __request_region(&iomem_resource, start, size,
214 resource_name, flags);
217 pr_debug("Unable to reserve System RAM region: %016llx->%016llx\n",
218 start, start + size);
219 return ERR_PTR(-EEXIST);
224 static void release_memory_resource(struct resource *res)
228 release_resource(res);
232 static int check_pfn_span(unsigned long pfn, unsigned long nr_pages)
235 * Disallow all operations smaller than a sub-section and only
236 * allow operations smaller than a section for
237 * SPARSEMEM_VMEMMAP. Note that check_hotplug_memory_range()
238 * enforces a larger memory_block_size_bytes() granularity for
239 * memory that will be marked online, so this check should only
240 * fire for direct arch_{add,remove}_memory() users outside of
241 * add_memory_resource().
243 unsigned long min_align;
245 if (IS_ENABLED(CONFIG_SPARSEMEM_VMEMMAP))
246 min_align = PAGES_PER_SUBSECTION;
248 min_align = PAGES_PER_SECTION;
249 if (!IS_ALIGNED(pfn | nr_pages, min_align))
255 * Return page for the valid pfn only if the page is online. All pfn
256 * walkers which rely on the fully initialized page->flags and others
257 * should use this rather than pfn_valid && pfn_to_page
259 struct page *pfn_to_online_page(unsigned long pfn)
261 unsigned long nr = pfn_to_section_nr(pfn);
262 struct dev_pagemap *pgmap;
263 struct mem_section *ms;
265 if (nr >= NR_MEM_SECTIONS)
268 ms = __nr_to_section(nr);
269 if (!online_section(ms))
273 * Save some code text when online_section() +
274 * pfn_section_valid() are sufficient.
276 if (IS_ENABLED(CONFIG_HAVE_ARCH_PFN_VALID) && !pfn_valid(pfn))
279 if (!pfn_section_valid(ms, pfn))
282 if (!online_device_section(ms))
283 return pfn_to_page(pfn);
286 * Slowpath: when ZONE_DEVICE collides with
287 * ZONE_{NORMAL,MOVABLE} within the same section some pfns in
288 * the section may be 'offline' but 'valid'. Only
289 * get_dev_pagemap() can determine sub-section online status.
291 pgmap = get_dev_pagemap(pfn, NULL);
292 put_dev_pagemap(pgmap);
294 /* The presence of a pgmap indicates ZONE_DEVICE offline pfn */
298 return pfn_to_page(pfn);
300 EXPORT_SYMBOL_GPL(pfn_to_online_page);
302 int __ref __add_pages(int nid, unsigned long pfn, unsigned long nr_pages,
303 struct mhp_params *params)
305 const unsigned long end_pfn = pfn + nr_pages;
306 unsigned long cur_nr_pages;
308 struct vmem_altmap *altmap = params->altmap;
310 if (WARN_ON_ONCE(!pgprot_val(params->pgprot)))
313 VM_BUG_ON(!mhp_range_allowed(PFN_PHYS(pfn), nr_pages * PAGE_SIZE, false));
317 * Validate altmap is within bounds of the total request
319 if (altmap->base_pfn != pfn
320 || vmem_altmap_offset(altmap) > nr_pages) {
321 pr_warn_once("memory add fail, invalid altmap\n");
327 if (check_pfn_span(pfn, nr_pages)) {
328 WARN(1, "Misaligned %s start: %#lx end: #%lx\n", __func__, pfn, pfn + nr_pages - 1);
332 for (; pfn < end_pfn; pfn += cur_nr_pages) {
333 /* Select all remaining pages up to the next section boundary */
334 cur_nr_pages = min(end_pfn - pfn,
335 SECTION_ALIGN_UP(pfn + 1) - pfn);
336 err = sparse_add_section(nid, pfn, cur_nr_pages, altmap,
342 vmemmap_populate_print_last();
346 /* find the smallest valid pfn in the range [start_pfn, end_pfn) */
347 static unsigned long find_smallest_section_pfn(int nid, struct zone *zone,
348 unsigned long start_pfn,
349 unsigned long end_pfn)
351 for (; start_pfn < end_pfn; start_pfn += PAGES_PER_SUBSECTION) {
352 if (unlikely(!pfn_to_online_page(start_pfn)))
355 if (unlikely(pfn_to_nid(start_pfn) != nid))
358 if (zone != page_zone(pfn_to_page(start_pfn)))
367 /* find the biggest valid pfn in the range [start_pfn, end_pfn). */
368 static unsigned long find_biggest_section_pfn(int nid, struct zone *zone,
369 unsigned long start_pfn,
370 unsigned long end_pfn)
374 /* pfn is the end pfn of a memory section. */
376 for (; pfn >= start_pfn; pfn -= PAGES_PER_SUBSECTION) {
377 if (unlikely(!pfn_to_online_page(pfn)))
380 if (unlikely(pfn_to_nid(pfn) != nid))
383 if (zone != page_zone(pfn_to_page(pfn)))
392 static void shrink_zone_span(struct zone *zone, unsigned long start_pfn,
393 unsigned long end_pfn)
396 int nid = zone_to_nid(zone);
398 if (zone->zone_start_pfn == start_pfn) {
400 * If the section is smallest section in the zone, it need
401 * shrink zone->zone_start_pfn and zone->zone_spanned_pages.
402 * In this case, we find second smallest valid mem_section
403 * for shrinking zone.
405 pfn = find_smallest_section_pfn(nid, zone, end_pfn,
408 zone->spanned_pages = zone_end_pfn(zone) - pfn;
409 zone->zone_start_pfn = pfn;
411 zone->zone_start_pfn = 0;
412 zone->spanned_pages = 0;
414 } else if (zone_end_pfn(zone) == end_pfn) {
416 * If the section is biggest section in the zone, it need
417 * shrink zone->spanned_pages.
418 * In this case, we find second biggest valid mem_section for
421 pfn = find_biggest_section_pfn(nid, zone, zone->zone_start_pfn,
424 zone->spanned_pages = pfn - zone->zone_start_pfn + 1;
426 zone->zone_start_pfn = 0;
427 zone->spanned_pages = 0;
432 static void update_pgdat_span(struct pglist_data *pgdat)
434 unsigned long node_start_pfn = 0, node_end_pfn = 0;
437 for (zone = pgdat->node_zones;
438 zone < pgdat->node_zones + MAX_NR_ZONES; zone++) {
439 unsigned long end_pfn = zone_end_pfn(zone);
441 /* No need to lock the zones, they can't change. */
442 if (!zone->spanned_pages)
445 node_start_pfn = zone->zone_start_pfn;
446 node_end_pfn = end_pfn;
450 if (end_pfn > node_end_pfn)
451 node_end_pfn = end_pfn;
452 if (zone->zone_start_pfn < node_start_pfn)
453 node_start_pfn = zone->zone_start_pfn;
456 pgdat->node_start_pfn = node_start_pfn;
457 pgdat->node_spanned_pages = node_end_pfn - node_start_pfn;
460 void __ref remove_pfn_range_from_zone(struct zone *zone,
461 unsigned long start_pfn,
462 unsigned long nr_pages)
464 const unsigned long end_pfn = start_pfn + nr_pages;
465 struct pglist_data *pgdat = zone->zone_pgdat;
466 unsigned long pfn, cur_nr_pages;
468 /* Poison struct pages because they are now uninitialized again. */
469 for (pfn = start_pfn; pfn < end_pfn; pfn += cur_nr_pages) {
472 /* Select all remaining pages up to the next section boundary */
474 min(end_pfn - pfn, SECTION_ALIGN_UP(pfn + 1) - pfn);
475 page_init_poison(pfn_to_page(pfn),
476 sizeof(struct page) * cur_nr_pages);
480 * Zone shrinking code cannot properly deal with ZONE_DEVICE. So
481 * we will not try to shrink the zones - which is okay as
482 * set_zone_contiguous() cannot deal with ZONE_DEVICE either way.
484 if (zone_is_zone_device(zone))
487 clear_zone_contiguous(zone);
489 shrink_zone_span(zone, start_pfn, start_pfn + nr_pages);
490 update_pgdat_span(pgdat);
492 set_zone_contiguous(zone);
495 static void __remove_section(unsigned long pfn, unsigned long nr_pages,
496 unsigned long map_offset,
497 struct vmem_altmap *altmap)
499 struct mem_section *ms = __pfn_to_section(pfn);
501 if (WARN_ON_ONCE(!valid_section(ms)))
504 sparse_remove_section(ms, pfn, nr_pages, map_offset, altmap);
508 * __remove_pages() - remove sections of pages
509 * @pfn: starting pageframe (must be aligned to start of a section)
510 * @nr_pages: number of pages to remove (must be multiple of section size)
511 * @altmap: alternative device page map or %NULL if default memmap is used
513 * Generic helper function to remove section mappings and sysfs entries
514 * for the section of the memory we are removing. Caller needs to make
515 * sure that pages are marked reserved and zones are adjust properly by
516 * calling offline_pages().
518 void __remove_pages(unsigned long pfn, unsigned long nr_pages,
519 struct vmem_altmap *altmap)
521 const unsigned long end_pfn = pfn + nr_pages;
522 unsigned long cur_nr_pages;
523 unsigned long map_offset = 0;
525 map_offset = vmem_altmap_offset(altmap);
527 if (check_pfn_span(pfn, nr_pages)) {
528 WARN(1, "Misaligned %s start: %#lx end: #%lx\n", __func__, pfn, pfn + nr_pages - 1);
532 for (; pfn < end_pfn; pfn += cur_nr_pages) {
534 /* Select all remaining pages up to the next section boundary */
535 cur_nr_pages = min(end_pfn - pfn,
536 SECTION_ALIGN_UP(pfn + 1) - pfn);
537 __remove_section(pfn, cur_nr_pages, map_offset, altmap);
542 int set_online_page_callback(online_page_callback_t callback)
547 mutex_lock(&online_page_callback_lock);
549 if (online_page_callback == generic_online_page) {
550 online_page_callback = callback;
554 mutex_unlock(&online_page_callback_lock);
559 EXPORT_SYMBOL_GPL(set_online_page_callback);
561 int restore_online_page_callback(online_page_callback_t callback)
566 mutex_lock(&online_page_callback_lock);
568 if (online_page_callback == callback) {
569 online_page_callback = generic_online_page;
573 mutex_unlock(&online_page_callback_lock);
578 EXPORT_SYMBOL_GPL(restore_online_page_callback);
580 void generic_online_page(struct page *page, unsigned int order)
583 * Freeing the page with debug_pagealloc enabled will try to unmap it,
584 * so we should map it first. This is better than introducing a special
585 * case in page freeing fast path.
587 debug_pagealloc_map_pages(page, 1 << order);
588 __free_pages_core(page, order);
589 totalram_pages_add(1UL << order);
591 EXPORT_SYMBOL_GPL(generic_online_page);
593 static void online_pages_range(unsigned long start_pfn, unsigned long nr_pages)
595 const unsigned long end_pfn = start_pfn + nr_pages;
599 * Online the pages in MAX_ORDER aligned chunks. The callback might
600 * decide to not expose all pages to the buddy (e.g., expose them
601 * later). We account all pages as being online and belonging to this
603 * When using memmap_on_memory, the range might not be aligned to
604 * MAX_ORDER_NR_PAGES - 1, but pageblock aligned. __ffs() will detect
605 * this and the first chunk to online will be pageblock_nr_pages.
607 for (pfn = start_pfn; pfn < end_pfn;) {
608 int order = min_t(int, MAX_ORDER, __ffs(pfn));
610 (*online_page_callback)(pfn_to_page(pfn), order);
611 pfn += (1UL << order);
614 /* mark all involved sections as online */
615 online_mem_sections(start_pfn, end_pfn);
618 /* check which state of node_states will be changed when online memory */
619 static void node_states_check_changes_online(unsigned long nr_pages,
620 struct zone *zone, struct memory_notify *arg)
622 int nid = zone_to_nid(zone);
624 arg->status_change_nid = NUMA_NO_NODE;
625 arg->status_change_nid_normal = NUMA_NO_NODE;
627 if (!node_state(nid, N_MEMORY))
628 arg->status_change_nid = nid;
629 if (zone_idx(zone) <= ZONE_NORMAL && !node_state(nid, N_NORMAL_MEMORY))
630 arg->status_change_nid_normal = nid;
633 static void node_states_set_node(int node, struct memory_notify *arg)
635 if (arg->status_change_nid_normal >= 0)
636 node_set_state(node, N_NORMAL_MEMORY);
638 if (arg->status_change_nid >= 0)
639 node_set_state(node, N_MEMORY);
642 static void __meminit resize_zone_range(struct zone *zone, unsigned long start_pfn,
643 unsigned long nr_pages)
645 unsigned long old_end_pfn = zone_end_pfn(zone);
647 if (zone_is_empty(zone) || start_pfn < zone->zone_start_pfn)
648 zone->zone_start_pfn = start_pfn;
650 zone->spanned_pages = max(start_pfn + nr_pages, old_end_pfn) - zone->zone_start_pfn;
653 static void __meminit resize_pgdat_range(struct pglist_data *pgdat, unsigned long start_pfn,
654 unsigned long nr_pages)
656 unsigned long old_end_pfn = pgdat_end_pfn(pgdat);
658 if (!pgdat->node_spanned_pages || start_pfn < pgdat->node_start_pfn)
659 pgdat->node_start_pfn = start_pfn;
661 pgdat->node_spanned_pages = max(start_pfn + nr_pages, old_end_pfn) - pgdat->node_start_pfn;
665 #ifdef CONFIG_ZONE_DEVICE
666 static void section_taint_zone_device(unsigned long pfn)
668 struct mem_section *ms = __pfn_to_section(pfn);
670 ms->section_mem_map |= SECTION_TAINT_ZONE_DEVICE;
673 static inline void section_taint_zone_device(unsigned long pfn)
679 * Associate the pfn range with the given zone, initializing the memmaps
680 * and resizing the pgdat/zone data to span the added pages. After this
681 * call, all affected pages are PG_reserved.
683 * All aligned pageblocks are initialized to the specified migratetype
684 * (usually MIGRATE_MOVABLE). Besides setting the migratetype, no related
685 * zone stats (e.g., nr_isolate_pageblock) are touched.
687 void __ref move_pfn_range_to_zone(struct zone *zone, unsigned long start_pfn,
688 unsigned long nr_pages,
689 struct vmem_altmap *altmap, int migratetype)
691 struct pglist_data *pgdat = zone->zone_pgdat;
692 int nid = pgdat->node_id;
694 clear_zone_contiguous(zone);
696 if (zone_is_empty(zone))
697 init_currently_empty_zone(zone, start_pfn, nr_pages);
698 resize_zone_range(zone, start_pfn, nr_pages);
699 resize_pgdat_range(pgdat, start_pfn, nr_pages);
702 * Subsection population requires care in pfn_to_online_page().
703 * Set the taint to enable the slow path detection of
704 * ZONE_DEVICE pages in an otherwise ZONE_{NORMAL,MOVABLE}
707 if (zone_is_zone_device(zone)) {
708 if (!IS_ALIGNED(start_pfn, PAGES_PER_SECTION))
709 section_taint_zone_device(start_pfn);
710 if (!IS_ALIGNED(start_pfn + nr_pages, PAGES_PER_SECTION))
711 section_taint_zone_device(start_pfn + nr_pages);
715 * TODO now we have a visible range of pages which are not associated
716 * with their zone properly. Not nice but set_pfnblock_flags_mask
717 * expects the zone spans the pfn range. All the pages in the range
718 * are reserved so nobody should be touching them so we should be safe
720 memmap_init_range(nr_pages, nid, zone_idx(zone), start_pfn, 0,
721 MEMINIT_HOTPLUG, altmap, migratetype);
723 set_zone_contiguous(zone);
726 struct auto_movable_stats {
727 unsigned long kernel_early_pages;
728 unsigned long movable_pages;
731 static void auto_movable_stats_account_zone(struct auto_movable_stats *stats,
734 if (zone_idx(zone) == ZONE_MOVABLE) {
735 stats->movable_pages += zone->present_pages;
737 stats->kernel_early_pages += zone->present_early_pages;
740 * CMA pages (never on hotplugged memory) behave like
743 stats->movable_pages += zone->cma_pages;
744 stats->kernel_early_pages -= zone->cma_pages;
745 #endif /* CONFIG_CMA */
748 struct auto_movable_group_stats {
749 unsigned long movable_pages;
750 unsigned long req_kernel_early_pages;
753 static int auto_movable_stats_account_group(struct memory_group *group,
756 const int ratio = READ_ONCE(auto_movable_ratio);
757 struct auto_movable_group_stats *stats = arg;
761 * We don't support modifying the config while the auto-movable online
762 * policy is already enabled. Just avoid the division by zero below.
768 * Calculate how many early kernel pages this group requires to
769 * satisfy the configured zone ratio.
771 pages = group->present_movable_pages * 100 / ratio;
772 pages -= group->present_kernel_pages;
775 stats->req_kernel_early_pages += pages;
776 stats->movable_pages += group->present_movable_pages;
780 static bool auto_movable_can_online_movable(int nid, struct memory_group *group,
781 unsigned long nr_pages)
783 unsigned long kernel_early_pages, movable_pages;
784 struct auto_movable_group_stats group_stats = {};
785 struct auto_movable_stats stats = {};
786 pg_data_t *pgdat = NODE_DATA(nid);
790 /* Walk all relevant zones and collect MOVABLE vs. KERNEL stats. */
791 if (nid == NUMA_NO_NODE) {
792 /* TODO: cache values */
793 for_each_populated_zone(zone)
794 auto_movable_stats_account_zone(&stats, zone);
796 for (i = 0; i < MAX_NR_ZONES; i++) {
797 zone = pgdat->node_zones + i;
798 if (populated_zone(zone))
799 auto_movable_stats_account_zone(&stats, zone);
803 kernel_early_pages = stats.kernel_early_pages;
804 movable_pages = stats.movable_pages;
807 * Kernel memory inside dynamic memory group allows for more MOVABLE
808 * memory within the same group. Remove the effect of all but the
809 * current group from the stats.
811 walk_dynamic_memory_groups(nid, auto_movable_stats_account_group,
812 group, &group_stats);
813 if (kernel_early_pages <= group_stats.req_kernel_early_pages)
815 kernel_early_pages -= group_stats.req_kernel_early_pages;
816 movable_pages -= group_stats.movable_pages;
818 if (group && group->is_dynamic)
819 kernel_early_pages += group->present_kernel_pages;
822 * Test if we could online the given number of pages to ZONE_MOVABLE
823 * and still stay in the configured ratio.
825 movable_pages += nr_pages;
826 return movable_pages <= (auto_movable_ratio * kernel_early_pages) / 100;
830 * Returns a default kernel memory zone for the given pfn range.
831 * If no kernel zone covers this pfn range it will automatically go
832 * to the ZONE_NORMAL.
834 static struct zone *default_kernel_zone_for_pfn(int nid, unsigned long start_pfn,
835 unsigned long nr_pages)
837 struct pglist_data *pgdat = NODE_DATA(nid);
840 for (zid = 0; zid < ZONE_NORMAL; zid++) {
841 struct zone *zone = &pgdat->node_zones[zid];
843 if (zone_intersects(zone, start_pfn, nr_pages))
847 return &pgdat->node_zones[ZONE_NORMAL];
851 * Determine to which zone to online memory dynamically based on user
852 * configuration and system stats. We care about the following ratio:
856 * Whereby MOVABLE is memory in ZONE_MOVABLE and KERNEL is memory in
857 * one of the kernel zones. CMA pages inside one of the kernel zones really
858 * behaves like ZONE_MOVABLE, so we treat them accordingly.
860 * We don't allow for hotplugged memory in a KERNEL zone to increase the
861 * amount of MOVABLE memory we can have, so we end up with:
863 * MOVABLE : KERNEL_EARLY
865 * Whereby KERNEL_EARLY is memory in one of the kernel zones, available sinze
866 * boot. We base our calculation on KERNEL_EARLY internally, because:
868 * a) Hotplugged memory in one of the kernel zones can sometimes still get
869 * hotunplugged, especially when hot(un)plugging individual memory blocks.
870 * There is no coordination across memory devices, therefore "automatic"
871 * hotunplugging, as implemented in hypervisors, could result in zone
873 * b) Early/boot memory in one of the kernel zones can usually not get
874 * hotunplugged again (e.g., no firmware interface to unplug, fragmented
875 * with unmovable allocations). While there are corner cases where it might
876 * still work, it is barely relevant in practice.
878 * Exceptions are dynamic memory groups, which allow for more MOVABLE
879 * memory within the same memory group -- because in that case, there is
880 * coordination within the single memory device managed by a single driver.
882 * We rely on "present pages" instead of "managed pages", as the latter is
883 * highly unreliable and dynamic in virtualized environments, and does not
884 * consider boot time allocations. For example, memory ballooning adjusts the
885 * managed pages when inflating/deflating the balloon, and balloon compaction
886 * can even migrate inflated pages between zones.
888 * Using "present pages" is better but some things to keep in mind are:
890 * a) Some memblock allocations, such as for the crashkernel area, are
891 * effectively unused by the kernel, yet they account to "present pages".
892 * Fortunately, these allocations are comparatively small in relevant setups
893 * (e.g., fraction of system memory).
894 * b) Some hotplugged memory blocks in virtualized environments, esecially
895 * hotplugged by virtio-mem, look like they are completely present, however,
896 * only parts of the memory block are actually currently usable.
897 * "present pages" is an upper limit that can get reached at runtime. As
898 * we base our calculations on KERNEL_EARLY, this is not an issue.
900 static struct zone *auto_movable_zone_for_pfn(int nid,
901 struct memory_group *group,
903 unsigned long nr_pages)
905 unsigned long online_pages = 0, max_pages, end_pfn;
908 if (!auto_movable_ratio)
911 if (group && !group->is_dynamic) {
912 max_pages = group->s.max_pages;
913 online_pages = group->present_movable_pages;
915 /* If anything is !MOVABLE online the rest !MOVABLE. */
916 if (group->present_kernel_pages)
918 } else if (!group || group->d.unit_pages == nr_pages) {
919 max_pages = nr_pages;
921 max_pages = group->d.unit_pages;
923 * Take a look at all online sections in the current unit.
924 * We can safely assume that all pages within a section belong
925 * to the same zone, because dynamic memory groups only deal
926 * with hotplugged memory.
928 pfn = ALIGN_DOWN(pfn, group->d.unit_pages);
929 end_pfn = pfn + group->d.unit_pages;
930 for (; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
931 page = pfn_to_online_page(pfn);
934 /* If anything is !MOVABLE online the rest !MOVABLE. */
935 if (!is_zone_movable_page(page))
937 online_pages += PAGES_PER_SECTION;
942 * Online MOVABLE if we could *currently* online all remaining parts
943 * MOVABLE. We expect to (add+) online them immediately next, so if
944 * nobody interferes, all will be MOVABLE if possible.
946 nr_pages = max_pages - online_pages;
947 if (!auto_movable_can_online_movable(NUMA_NO_NODE, group, nr_pages))
951 if (auto_movable_numa_aware &&
952 !auto_movable_can_online_movable(nid, group, nr_pages))
954 #endif /* CONFIG_NUMA */
956 return &NODE_DATA(nid)->node_zones[ZONE_MOVABLE];
958 return default_kernel_zone_for_pfn(nid, pfn, nr_pages);
961 static inline struct zone *default_zone_for_pfn(int nid, unsigned long start_pfn,
962 unsigned long nr_pages)
964 struct zone *kernel_zone = default_kernel_zone_for_pfn(nid, start_pfn,
966 struct zone *movable_zone = &NODE_DATA(nid)->node_zones[ZONE_MOVABLE];
967 bool in_kernel = zone_intersects(kernel_zone, start_pfn, nr_pages);
968 bool in_movable = zone_intersects(movable_zone, start_pfn, nr_pages);
971 * We inherit the existing zone in a simple case where zones do not
972 * overlap in the given range
974 if (in_kernel ^ in_movable)
975 return (in_kernel) ? kernel_zone : movable_zone;
978 * If the range doesn't belong to any zone or two zones overlap in the
979 * given range then we use movable zone only if movable_node is
980 * enabled because we always online to a kernel zone by default.
982 return movable_node_enabled ? movable_zone : kernel_zone;
985 struct zone *zone_for_pfn_range(int online_type, int nid,
986 struct memory_group *group, unsigned long start_pfn,
987 unsigned long nr_pages)
989 if (online_type == MMOP_ONLINE_KERNEL)
990 return default_kernel_zone_for_pfn(nid, start_pfn, nr_pages);
992 if (online_type == MMOP_ONLINE_MOVABLE)
993 return &NODE_DATA(nid)->node_zones[ZONE_MOVABLE];
995 if (online_policy == ONLINE_POLICY_AUTO_MOVABLE)
996 return auto_movable_zone_for_pfn(nid, group, start_pfn, nr_pages);
998 return default_zone_for_pfn(nid, start_pfn, nr_pages);
1002 * This function should only be called by memory_block_{online,offline},
1003 * and {online,offline}_pages.
1005 void adjust_present_page_count(struct page *page, struct memory_group *group,
1008 struct zone *zone = page_zone(page);
1009 const bool movable = zone_idx(zone) == ZONE_MOVABLE;
1012 * We only support onlining/offlining/adding/removing of complete
1013 * memory blocks; therefore, either all is either early or hotplugged.
1015 if (early_section(__pfn_to_section(page_to_pfn(page))))
1016 zone->present_early_pages += nr_pages;
1017 zone->present_pages += nr_pages;
1018 zone->zone_pgdat->node_present_pages += nr_pages;
1020 if (group && movable)
1021 group->present_movable_pages += nr_pages;
1022 else if (group && !movable)
1023 group->present_kernel_pages += nr_pages;
1026 int mhp_init_memmap_on_memory(unsigned long pfn, unsigned long nr_pages,
1029 unsigned long end_pfn = pfn + nr_pages;
1032 ret = kasan_add_zero_shadow(__va(PFN_PHYS(pfn)), PFN_PHYS(nr_pages));
1036 move_pfn_range_to_zone(zone, pfn, nr_pages, NULL, MIGRATE_UNMOVABLE);
1038 for (i = 0; i < nr_pages; i++)
1039 SetPageVmemmapSelfHosted(pfn_to_page(pfn + i));
1042 * It might be that the vmemmap_pages fully span sections. If that is
1043 * the case, mark those sections online here as otherwise they will be
1046 if (nr_pages >= PAGES_PER_SECTION)
1047 online_mem_sections(pfn, ALIGN_DOWN(end_pfn, PAGES_PER_SECTION));
1052 void mhp_deinit_memmap_on_memory(unsigned long pfn, unsigned long nr_pages)
1054 unsigned long end_pfn = pfn + nr_pages;
1057 * It might be that the vmemmap_pages fully span sections. If that is
1058 * the case, mark those sections offline here as otherwise they will be
1061 if (nr_pages >= PAGES_PER_SECTION)
1062 offline_mem_sections(pfn, ALIGN_DOWN(end_pfn, PAGES_PER_SECTION));
1065 * The pages associated with this vmemmap have been offlined, so
1066 * we can reset its state here.
1068 remove_pfn_range_from_zone(page_zone(pfn_to_page(pfn)), pfn, nr_pages);
1069 kasan_remove_zero_shadow(__va(PFN_PHYS(pfn)), PFN_PHYS(nr_pages));
1072 int __ref online_pages(unsigned long pfn, unsigned long nr_pages,
1073 struct zone *zone, struct memory_group *group)
1075 unsigned long flags;
1076 int need_zonelists_rebuild = 0;
1077 const int nid = zone_to_nid(zone);
1079 struct memory_notify arg;
1082 * {on,off}lining is constrained to full memory sections (or more
1083 * precisely to memory blocks from the user space POV).
1084 * memmap_on_memory is an exception because it reserves initial part
1085 * of the physical memory space for vmemmaps. That space is pageblock
1088 if (WARN_ON_ONCE(!nr_pages || !pageblock_aligned(pfn) ||
1089 !IS_ALIGNED(pfn + nr_pages, PAGES_PER_SECTION)))
1092 mem_hotplug_begin();
1094 /* associate pfn range with the zone */
1095 move_pfn_range_to_zone(zone, pfn, nr_pages, NULL, MIGRATE_ISOLATE);
1097 arg.start_pfn = pfn;
1098 arg.nr_pages = nr_pages;
1099 node_states_check_changes_online(nr_pages, zone, &arg);
1101 ret = memory_notify(MEM_GOING_ONLINE, &arg);
1102 ret = notifier_to_errno(ret);
1104 goto failed_addition;
1107 * Fixup the number of isolated pageblocks before marking the sections
1108 * onlining, such that undo_isolate_page_range() works correctly.
1110 spin_lock_irqsave(&zone->lock, flags);
1111 zone->nr_isolate_pageblock += nr_pages / pageblock_nr_pages;
1112 spin_unlock_irqrestore(&zone->lock, flags);
1115 * If this zone is not populated, then it is not in zonelist.
1116 * This means the page allocator ignores this zone.
1117 * So, zonelist must be updated after online.
1119 if (!populated_zone(zone)) {
1120 need_zonelists_rebuild = 1;
1121 setup_zone_pageset(zone);
1124 online_pages_range(pfn, nr_pages);
1125 adjust_present_page_count(pfn_to_page(pfn), group, nr_pages);
1127 node_states_set_node(nid, &arg);
1128 if (need_zonelists_rebuild)
1129 build_all_zonelists(NULL);
1131 /* Basic onlining is complete, allow allocation of onlined pages. */
1132 undo_isolate_page_range(pfn, pfn + nr_pages, MIGRATE_MOVABLE);
1135 * Freshly onlined pages aren't shuffled (e.g., all pages are placed to
1136 * the tail of the freelist when undoing isolation). Shuffle the whole
1137 * zone to make sure the just onlined pages are properly distributed
1138 * across the whole freelist - to create an initial shuffle.
1142 /* reinitialise watermarks and update pcp limits */
1143 init_per_zone_wmark_min();
1148 writeback_set_ratelimit();
1150 memory_notify(MEM_ONLINE, &arg);
1155 pr_debug("online_pages [mem %#010llx-%#010llx] failed\n",
1156 (unsigned long long) pfn << PAGE_SHIFT,
1157 (((unsigned long long) pfn + nr_pages) << PAGE_SHIFT) - 1);
1158 memory_notify(MEM_CANCEL_ONLINE, &arg);
1159 remove_pfn_range_from_zone(zone, pfn, nr_pages);
1164 static void reset_node_present_pages(pg_data_t *pgdat)
1168 for (z = pgdat->node_zones; z < pgdat->node_zones + MAX_NR_ZONES; z++)
1169 z->present_pages = 0;
1171 pgdat->node_present_pages = 0;
1174 /* we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG */
1175 static pg_data_t __ref *hotadd_init_pgdat(int nid)
1177 struct pglist_data *pgdat;
1180 * NODE_DATA is preallocated (free_area_init) but its internal
1181 * state is not allocated completely. Add missing pieces.
1182 * Completely offline nodes stay around and they just need
1185 pgdat = NODE_DATA(nid);
1187 /* init node's zones as empty zones, we don't have any present pages.*/
1188 free_area_init_core_hotplug(pgdat);
1191 * The node we allocated has no zone fallback lists. For avoiding
1192 * to access not-initialized zonelist, build here.
1194 build_all_zonelists(pgdat);
1197 * When memory is hot-added, all the memory is in offline state. So
1198 * clear all zones' present_pages because they will be updated in
1199 * online_pages() and offline_pages().
1200 * TODO: should be in free_area_init_core_hotplug?
1202 reset_node_managed_pages(pgdat);
1203 reset_node_present_pages(pgdat);
1209 * __try_online_node - online a node if offlined
1211 * @set_node_online: Whether we want to online the node
1212 * called by cpu_up() to online a node without onlined memory.
1215 * 1 -> a new node has been allocated
1216 * 0 -> the node is already online
1217 * -ENOMEM -> the node could not be allocated
1219 static int __try_online_node(int nid, bool set_node_online)
1224 if (node_online(nid))
1227 pgdat = hotadd_init_pgdat(nid);
1229 pr_err("Cannot online node %d due to NULL pgdat\n", nid);
1234 if (set_node_online) {
1235 node_set_online(nid);
1236 ret = register_one_node(nid);
1244 * Users of this function always want to online/register the node
1246 int try_online_node(int nid)
1250 mem_hotplug_begin();
1251 ret = __try_online_node(nid, true);
1256 static int check_hotplug_memory_range(u64 start, u64 size)
1258 /* memory range must be block size aligned */
1259 if (!size || !IS_ALIGNED(start, memory_block_size_bytes()) ||
1260 !IS_ALIGNED(size, memory_block_size_bytes())) {
1261 pr_err("Block size [%#lx] unaligned hotplug range: start %#llx, size %#llx",
1262 memory_block_size_bytes(), start, size);
1269 static int online_memory_block(struct memory_block *mem, void *arg)
1271 mem->online_type = mhp_default_online_type;
1272 return device_online(&mem->dev);
1275 bool mhp_supports_memmap_on_memory(unsigned long size)
1277 unsigned long nr_vmemmap_pages = size / PAGE_SIZE;
1278 unsigned long vmemmap_size = nr_vmemmap_pages * sizeof(struct page);
1279 unsigned long remaining_size = size - vmemmap_size;
1282 * Besides having arch support and the feature enabled at runtime, we
1283 * need a few more assumptions to hold true:
1285 * a) We span a single memory block: memory onlining/offlinin;g happens
1286 * in memory block granularity. We don't want the vmemmap of online
1287 * memory blocks to reside on offline memory blocks. In the future,
1288 * we might want to support variable-sized memory blocks to make the
1289 * feature more versatile.
1291 * b) The vmemmap pages span complete PMDs: We don't want vmemmap code
1292 * to populate memory from the altmap for unrelated parts (i.e.,
1293 * other memory blocks)
1295 * c) The vmemmap pages (and thereby the pages that will be exposed to
1296 * the buddy) have to cover full pageblocks: memory onlining/offlining
1297 * code requires applicable ranges to be page-aligned, for example, to
1298 * set the migratetypes properly.
1300 * TODO: Although we have a check here to make sure that vmemmap pages
1301 * fully populate a PMD, it is not the right place to check for
1302 * this. A much better solution involves improving vmemmap code
1303 * to fallback to base pages when trying to populate vmemmap using
1304 * altmap as an alternative source of memory, and we do not exactly
1305 * populate a single PMD.
1307 return mhp_memmap_on_memory() &&
1308 size == memory_block_size_bytes() &&
1309 IS_ALIGNED(vmemmap_size, PMD_SIZE) &&
1310 IS_ALIGNED(remaining_size, (pageblock_nr_pages << PAGE_SHIFT));
1314 * NOTE: The caller must call lock_device_hotplug() to serialize hotplug
1315 * and online/offline operations (triggered e.g. by sysfs).
1317 * we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG
1319 int __ref add_memory_resource(int nid, struct resource *res, mhp_t mhp_flags)
1321 struct mhp_params params = { .pgprot = pgprot_mhp(PAGE_KERNEL) };
1322 enum memblock_flags memblock_flags = MEMBLOCK_NONE;
1323 struct vmem_altmap mhp_altmap = {};
1324 struct memory_group *group = NULL;
1326 bool new_node = false;
1330 size = resource_size(res);
1332 ret = check_hotplug_memory_range(start, size);
1336 if (mhp_flags & MHP_NID_IS_MGID) {
1337 group = memory_group_find_by_id(nid);
1343 if (!node_possible(nid)) {
1344 WARN(1, "node %d was absent from the node_possible_map\n", nid);
1348 mem_hotplug_begin();
1350 if (IS_ENABLED(CONFIG_ARCH_KEEP_MEMBLOCK)) {
1351 if (res->flags & IORESOURCE_SYSRAM_DRIVER_MANAGED)
1352 memblock_flags = MEMBLOCK_DRIVER_MANAGED;
1353 ret = memblock_add_node(start, size, nid, memblock_flags);
1355 goto error_mem_hotplug_end;
1358 ret = __try_online_node(nid, false);
1364 * Self hosted memmap array
1366 if (mhp_flags & MHP_MEMMAP_ON_MEMORY) {
1367 if (!mhp_supports_memmap_on_memory(size)) {
1371 mhp_altmap.free = PHYS_PFN(size);
1372 mhp_altmap.base_pfn = PHYS_PFN(start);
1373 params.altmap = &mhp_altmap;
1376 /* call arch's memory hotadd */
1377 ret = arch_add_memory(nid, start, size, ¶ms);
1381 /* create memory block devices after memory was added */
1382 ret = create_memory_block_devices(start, size, mhp_altmap.alloc,
1385 arch_remove_memory(start, size, NULL);
1390 /* If sysfs file of new node can't be created, cpu on the node
1391 * can't be hot-added. There is no rollback way now.
1392 * So, check by BUG_ON() to catch it reluctantly..
1393 * We online node here. We can't roll back from here.
1395 node_set_online(nid);
1396 ret = __register_one_node(nid);
1400 register_memory_blocks_under_node(nid, PFN_DOWN(start),
1401 PFN_UP(start + size - 1),
1404 /* create new memmap entry */
1405 if (!strcmp(res->name, "System RAM"))
1406 firmware_map_add_hotplug(start, start + size, "System RAM");
1408 /* device_online() will take the lock when calling online_pages() */
1412 * In case we're allowed to merge the resource, flag it and trigger
1413 * merging now that adding succeeded.
1415 if (mhp_flags & MHP_MERGE_RESOURCE)
1416 merge_system_ram_resource(res);
1418 /* online pages if requested */
1419 if (mhp_default_online_type != MMOP_OFFLINE)
1420 walk_memory_blocks(start, size, NULL, online_memory_block);
1424 if (IS_ENABLED(CONFIG_ARCH_KEEP_MEMBLOCK))
1425 memblock_remove(start, size);
1426 error_mem_hotplug_end:
1431 /* requires device_hotplug_lock, see add_memory_resource() */
1432 int __ref __add_memory(int nid, u64 start, u64 size, mhp_t mhp_flags)
1434 struct resource *res;
1437 res = register_memory_resource(start, size, "System RAM");
1439 return PTR_ERR(res);
1441 ret = add_memory_resource(nid, res, mhp_flags);
1443 release_memory_resource(res);
1447 int add_memory(int nid, u64 start, u64 size, mhp_t mhp_flags)
1451 lock_device_hotplug();
1452 rc = __add_memory(nid, start, size, mhp_flags);
1453 unlock_device_hotplug();
1457 EXPORT_SYMBOL_GPL(add_memory);
1460 * Add special, driver-managed memory to the system as system RAM. Such
1461 * memory is not exposed via the raw firmware-provided memmap as system
1462 * RAM, instead, it is detected and added by a driver - during cold boot,
1463 * after a reboot, and after kexec.
1465 * Reasons why this memory should not be used for the initial memmap of a
1466 * kexec kernel or for placing kexec images:
1467 * - The booting kernel is in charge of determining how this memory will be
1468 * used (e.g., use persistent memory as system RAM)
1469 * - Coordination with a hypervisor is required before this memory
1470 * can be used (e.g., inaccessible parts).
1472 * For this memory, no entries in /sys/firmware/memmap ("raw firmware-provided
1473 * memory map") are created. Also, the created memory resource is flagged
1474 * with IORESOURCE_SYSRAM_DRIVER_MANAGED, so in-kernel users can special-case
1475 * this memory as well (esp., not place kexec images onto it).
1477 * The resource_name (visible via /proc/iomem) has to have the format
1478 * "System RAM ($DRIVER)".
1480 int add_memory_driver_managed(int nid, u64 start, u64 size,
1481 const char *resource_name, mhp_t mhp_flags)
1483 struct resource *res;
1486 if (!resource_name ||
1487 strstr(resource_name, "System RAM (") != resource_name ||
1488 resource_name[strlen(resource_name) - 1] != ')')
1491 lock_device_hotplug();
1493 res = register_memory_resource(start, size, resource_name);
1499 rc = add_memory_resource(nid, res, mhp_flags);
1501 release_memory_resource(res);
1504 unlock_device_hotplug();
1507 EXPORT_SYMBOL_GPL(add_memory_driver_managed);
1510 * Platforms should define arch_get_mappable_range() that provides
1511 * maximum possible addressable physical memory range for which the
1512 * linear mapping could be created. The platform returned address
1513 * range must adhere to these following semantics.
1515 * - range.start <= range.end
1516 * - Range includes both end points [range.start..range.end]
1518 * There is also a fallback definition provided here, allowing the
1519 * entire possible physical address range in case any platform does
1520 * not define arch_get_mappable_range().
1522 struct range __weak arch_get_mappable_range(void)
1524 struct range mhp_range = {
1531 struct range mhp_get_pluggable_range(bool need_mapping)
1533 const u64 max_phys = (1ULL << MAX_PHYSMEM_BITS) - 1;
1534 struct range mhp_range;
1537 mhp_range = arch_get_mappable_range();
1538 if (mhp_range.start > max_phys) {
1539 mhp_range.start = 0;
1542 mhp_range.end = min_t(u64, mhp_range.end, max_phys);
1544 mhp_range.start = 0;
1545 mhp_range.end = max_phys;
1549 EXPORT_SYMBOL_GPL(mhp_get_pluggable_range);
1551 bool mhp_range_allowed(u64 start, u64 size, bool need_mapping)
1553 struct range mhp_range = mhp_get_pluggable_range(need_mapping);
1554 u64 end = start + size;
1556 if (start < end && start >= mhp_range.start && (end - 1) <= mhp_range.end)
1559 pr_warn("Hotplug memory [%#llx-%#llx] exceeds maximum addressable range [%#llx-%#llx]\n",
1560 start, end, mhp_range.start, mhp_range.end);
1564 #ifdef CONFIG_MEMORY_HOTREMOVE
1566 * Scan pfn range [start,end) to find movable/migratable pages (LRU pages,
1567 * non-lru movable pages and hugepages). Will skip over most unmovable
1568 * pages (esp., pages that can be skipped when offlining), but bail out on
1569 * definitely unmovable pages.
1572 * 0 in case a movable page is found and movable_pfn was updated.
1573 * -ENOENT in case no movable page was found.
1574 * -EBUSY in case a definitely unmovable page was found.
1576 static int scan_movable_pages(unsigned long start, unsigned long end,
1577 unsigned long *movable_pfn)
1581 for (pfn = start; pfn < end; pfn++) {
1582 struct page *page, *head;
1585 if (!pfn_valid(pfn))
1587 page = pfn_to_page(pfn);
1590 if (__PageMovable(page))
1594 * PageOffline() pages that are not marked __PageMovable() and
1595 * have a reference count > 0 (after MEM_GOING_OFFLINE) are
1596 * definitely unmovable. If their reference count would be 0,
1597 * they could at least be skipped when offlining memory.
1599 if (PageOffline(page) && page_count(page))
1602 if (!PageHuge(page))
1604 head = compound_head(page);
1606 * This test is racy as we hold no reference or lock. The
1607 * hugetlb page could have been free'ed and head is no longer
1608 * a hugetlb page before the following check. In such unlikely
1609 * cases false positives and negatives are possible. Calling
1610 * code must deal with these scenarios.
1612 if (HPageMigratable(head))
1614 skip = compound_nr(head) - (page - head);
1623 static void do_migrate_range(unsigned long start_pfn, unsigned long end_pfn)
1626 struct page *page, *head;
1628 static DEFINE_RATELIMIT_STATE(migrate_rs, DEFAULT_RATELIMIT_INTERVAL,
1629 DEFAULT_RATELIMIT_BURST);
1631 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
1632 struct folio *folio;
1635 if (!pfn_valid(pfn))
1637 page = pfn_to_page(pfn);
1638 folio = page_folio(page);
1639 head = &folio->page;
1641 if (PageHuge(page)) {
1642 pfn = page_to_pfn(head) + compound_nr(head) - 1;
1643 isolate_hugetlb(folio, &source);
1645 } else if (PageTransHuge(page))
1646 pfn = page_to_pfn(head) + thp_nr_pages(page) - 1;
1649 * HWPoison pages have elevated reference counts so the migration would
1650 * fail on them. It also doesn't make any sense to migrate them in the
1651 * first place. Still try to unmap such a page in case it is still mapped
1652 * (e.g. current hwpoison implementation doesn't unmap KSM pages but keep
1653 * the unmap as the catch all safety net).
1655 if (PageHWPoison(page)) {
1656 if (WARN_ON(folio_test_lru(folio)))
1657 folio_isolate_lru(folio);
1658 if (folio_mapped(folio))
1659 try_to_unmap(folio, TTU_IGNORE_MLOCK);
1663 if (!get_page_unless_zero(page))
1666 * We can skip free pages. And we can deal with pages on
1667 * LRU and non-lru movable pages.
1670 isolated = isolate_lru_page(page);
1672 isolated = isolate_movable_page(page, ISOLATE_UNEVICTABLE);
1674 list_add_tail(&page->lru, &source);
1675 if (!__PageMovable(page))
1676 inc_node_page_state(page, NR_ISOLATED_ANON +
1677 page_is_file_lru(page));
1680 if (__ratelimit(&migrate_rs)) {
1681 pr_warn("failed to isolate pfn %lx\n", pfn);
1682 dump_page(page, "isolation failed");
1687 if (!list_empty(&source)) {
1688 nodemask_t nmask = node_states[N_MEMORY];
1689 struct migration_target_control mtc = {
1691 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
1696 * We have checked that migration range is on a single zone so
1697 * we can use the nid of the first page to all the others.
1699 mtc.nid = page_to_nid(list_first_entry(&source, struct page, lru));
1702 * try to allocate from a different node but reuse this node
1703 * if there are no other online nodes to be used (e.g. we are
1704 * offlining a part of the only existing node)
1706 node_clear(mtc.nid, nmask);
1707 if (nodes_empty(nmask))
1708 node_set(mtc.nid, nmask);
1709 ret = migrate_pages(&source, alloc_migration_target, NULL,
1710 (unsigned long)&mtc, MIGRATE_SYNC, MR_MEMORY_HOTPLUG, NULL);
1712 list_for_each_entry(page, &source, lru) {
1713 if (__ratelimit(&migrate_rs)) {
1714 pr_warn("migrating pfn %lx failed ret:%d\n",
1715 page_to_pfn(page), ret);
1716 dump_page(page, "migration failure");
1719 putback_movable_pages(&source);
1724 static int __init cmdline_parse_movable_node(char *p)
1726 movable_node_enabled = true;
1729 early_param("movable_node", cmdline_parse_movable_node);
1731 /* check which state of node_states will be changed when offline memory */
1732 static void node_states_check_changes_offline(unsigned long nr_pages,
1733 struct zone *zone, struct memory_notify *arg)
1735 struct pglist_data *pgdat = zone->zone_pgdat;
1736 unsigned long present_pages = 0;
1739 arg->status_change_nid = NUMA_NO_NODE;
1740 arg->status_change_nid_normal = NUMA_NO_NODE;
1743 * Check whether node_states[N_NORMAL_MEMORY] will be changed.
1744 * If the memory to be offline is within the range
1745 * [0..ZONE_NORMAL], and it is the last present memory there,
1746 * the zones in that range will become empty after the offlining,
1747 * thus we can determine that we need to clear the node from
1748 * node_states[N_NORMAL_MEMORY].
1750 for (zt = 0; zt <= ZONE_NORMAL; zt++)
1751 present_pages += pgdat->node_zones[zt].present_pages;
1752 if (zone_idx(zone) <= ZONE_NORMAL && nr_pages >= present_pages)
1753 arg->status_change_nid_normal = zone_to_nid(zone);
1756 * We have accounted the pages from [0..ZONE_NORMAL); ZONE_HIGHMEM
1757 * does not apply as we don't support 32bit.
1758 * Here we count the possible pages from ZONE_MOVABLE.
1759 * If after having accounted all the pages, we see that the nr_pages
1760 * to be offlined is over or equal to the accounted pages,
1761 * we know that the node will become empty, and so, we can clear
1762 * it for N_MEMORY as well.
1764 present_pages += pgdat->node_zones[ZONE_MOVABLE].present_pages;
1766 if (nr_pages >= present_pages)
1767 arg->status_change_nid = zone_to_nid(zone);
1770 static void node_states_clear_node(int node, struct memory_notify *arg)
1772 if (arg->status_change_nid_normal >= 0)
1773 node_clear_state(node, N_NORMAL_MEMORY);
1775 if (arg->status_change_nid >= 0)
1776 node_clear_state(node, N_MEMORY);
1779 static int count_system_ram_pages_cb(unsigned long start_pfn,
1780 unsigned long nr_pages, void *data)
1782 unsigned long *nr_system_ram_pages = data;
1784 *nr_system_ram_pages += nr_pages;
1788 int __ref offline_pages(unsigned long start_pfn, unsigned long nr_pages,
1789 struct zone *zone, struct memory_group *group)
1791 const unsigned long end_pfn = start_pfn + nr_pages;
1792 unsigned long pfn, system_ram_pages = 0;
1793 const int node = zone_to_nid(zone);
1794 unsigned long flags;
1795 struct memory_notify arg;
1800 * {on,off}lining is constrained to full memory sections (or more
1801 * precisely to memory blocks from the user space POV).
1802 * memmap_on_memory is an exception because it reserves initial part
1803 * of the physical memory space for vmemmaps. That space is pageblock
1806 if (WARN_ON_ONCE(!nr_pages || !pageblock_aligned(start_pfn) ||
1807 !IS_ALIGNED(start_pfn + nr_pages, PAGES_PER_SECTION)))
1810 mem_hotplug_begin();
1813 * Don't allow to offline memory blocks that contain holes.
1814 * Consequently, memory blocks with holes can never get onlined
1815 * via the hotplug path - online_pages() - as hotplugged memory has
1816 * no holes. This way, we e.g., don't have to worry about marking
1817 * memory holes PG_reserved, don't need pfn_valid() checks, and can
1818 * avoid using walk_system_ram_range() later.
1820 walk_system_ram_range(start_pfn, nr_pages, &system_ram_pages,
1821 count_system_ram_pages_cb);
1822 if (system_ram_pages != nr_pages) {
1824 reason = "memory holes";
1825 goto failed_removal;
1829 * We only support offlining of memory blocks managed by a single zone,
1830 * checked by calling code. This is just a sanity check that we might
1831 * want to remove in the future.
1833 if (WARN_ON_ONCE(page_zone(pfn_to_page(start_pfn)) != zone ||
1834 page_zone(pfn_to_page(end_pfn - 1)) != zone)) {
1836 reason = "multizone range";
1837 goto failed_removal;
1841 * Disable pcplists so that page isolation cannot race with freeing
1842 * in a way that pages from isolated pageblock are left on pcplists.
1844 zone_pcp_disable(zone);
1845 lru_cache_disable();
1847 /* set above range as isolated */
1848 ret = start_isolate_page_range(start_pfn, end_pfn,
1850 MEMORY_OFFLINE | REPORT_FAILURE,
1851 GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL);
1853 reason = "failure to isolate range";
1854 goto failed_removal_pcplists_disabled;
1857 arg.start_pfn = start_pfn;
1858 arg.nr_pages = nr_pages;
1859 node_states_check_changes_offline(nr_pages, zone, &arg);
1861 ret = memory_notify(MEM_GOING_OFFLINE, &arg);
1862 ret = notifier_to_errno(ret);
1864 reason = "notifier failure";
1865 goto failed_removal_isolated;
1871 if (signal_pending(current)) {
1873 reason = "signal backoff";
1874 goto failed_removal_isolated;
1879 ret = scan_movable_pages(pfn, end_pfn, &pfn);
1882 * TODO: fatal migration failures should bail
1885 do_migrate_range(pfn, end_pfn);
1889 if (ret != -ENOENT) {
1890 reason = "unmovable page";
1891 goto failed_removal_isolated;
1895 * Dissolve free hugepages in the memory block before doing
1896 * offlining actually in order to make hugetlbfs's object
1897 * counting consistent.
1899 ret = dissolve_free_huge_pages(start_pfn, end_pfn);
1901 reason = "failure to dissolve huge pages";
1902 goto failed_removal_isolated;
1905 ret = test_pages_isolated(start_pfn, end_pfn, MEMORY_OFFLINE);
1909 /* Mark all sections offline and remove free pages from the buddy. */
1910 __offline_isolated_pages(start_pfn, end_pfn);
1911 pr_debug("Offlined Pages %ld\n", nr_pages);
1914 * The memory sections are marked offline, and the pageblock flags
1915 * effectively stale; nobody should be touching them. Fixup the number
1916 * of isolated pageblocks, memory onlining will properly revert this.
1918 spin_lock_irqsave(&zone->lock, flags);
1919 zone->nr_isolate_pageblock -= nr_pages / pageblock_nr_pages;
1920 spin_unlock_irqrestore(&zone->lock, flags);
1923 zone_pcp_enable(zone);
1925 /* removal success */
1926 adjust_managed_page_count(pfn_to_page(start_pfn), -nr_pages);
1927 adjust_present_page_count(pfn_to_page(start_pfn), group, -nr_pages);
1929 /* reinitialise watermarks and update pcp limits */
1930 init_per_zone_wmark_min();
1932 if (!populated_zone(zone)) {
1933 zone_pcp_reset(zone);
1934 build_all_zonelists(NULL);
1937 node_states_clear_node(node, &arg);
1938 if (arg.status_change_nid >= 0) {
1939 kcompactd_stop(node);
1943 writeback_set_ratelimit();
1945 memory_notify(MEM_OFFLINE, &arg);
1946 remove_pfn_range_from_zone(zone, start_pfn, nr_pages);
1950 failed_removal_isolated:
1951 /* pushback to free area */
1952 undo_isolate_page_range(start_pfn, end_pfn, MIGRATE_MOVABLE);
1953 memory_notify(MEM_CANCEL_OFFLINE, &arg);
1954 failed_removal_pcplists_disabled:
1956 zone_pcp_enable(zone);
1958 pr_debug("memory offlining [mem %#010llx-%#010llx] failed due to %s\n",
1959 (unsigned long long) start_pfn << PAGE_SHIFT,
1960 ((unsigned long long) end_pfn << PAGE_SHIFT) - 1,
1966 static int check_memblock_offlined_cb(struct memory_block *mem, void *arg)
1971 if (unlikely(mem->state != MEM_OFFLINE)) {
1972 phys_addr_t beginpa, endpa;
1974 beginpa = PFN_PHYS(section_nr_to_pfn(mem->start_section_nr));
1975 endpa = beginpa + memory_block_size_bytes() - 1;
1976 pr_warn("removing memory fails, because memory [%pa-%pa] is onlined\n",
1984 static int get_nr_vmemmap_pages_cb(struct memory_block *mem, void *arg)
1987 * If not set, continue with the next block.
1989 return mem->nr_vmemmap_pages;
1992 static int check_cpu_on_node(int nid)
1996 for_each_present_cpu(cpu) {
1997 if (cpu_to_node(cpu) == nid)
1999 * the cpu on this node isn't removed, and we can't
2000 * offline this node.
2008 static int check_no_memblock_for_node_cb(struct memory_block *mem, void *arg)
2010 int nid = *(int *)arg;
2013 * If a memory block belongs to multiple nodes, the stored nid is not
2014 * reliable. However, such blocks are always online (e.g., cannot get
2015 * offlined) and, therefore, are still spanned by the node.
2017 return mem->nid == nid ? -EEXIST : 0;
2024 * Offline a node if all memory sections and cpus of the node are removed.
2026 * NOTE: The caller must call lock_device_hotplug() to serialize hotplug
2027 * and online/offline operations before this call.
2029 void try_offline_node(int nid)
2034 * If the node still spans pages (especially ZONE_DEVICE), don't
2035 * offline it. A node spans memory after move_pfn_range_to_zone(),
2036 * e.g., after the memory block was onlined.
2038 if (node_spanned_pages(nid))
2042 * Especially offline memory blocks might not be spanned by the
2043 * node. They will get spanned by the node once they get onlined.
2044 * However, they link to the node in sysfs and can get onlined later.
2046 rc = for_each_memory_block(&nid, check_no_memblock_for_node_cb);
2050 if (check_cpu_on_node(nid))
2054 * all memory/cpu of this node are removed, we can offline this
2057 node_set_offline(nid);
2058 unregister_one_node(nid);
2060 EXPORT_SYMBOL(try_offline_node);
2062 static int __ref try_remove_memory(u64 start, u64 size)
2064 struct vmem_altmap mhp_altmap = {};
2065 struct vmem_altmap *altmap = NULL;
2066 unsigned long nr_vmemmap_pages;
2067 int rc = 0, nid = NUMA_NO_NODE;
2069 BUG_ON(check_hotplug_memory_range(start, size));
2072 * All memory blocks must be offlined before removing memory. Check
2073 * whether all memory blocks in question are offline and return error
2074 * if this is not the case.
2076 * While at it, determine the nid. Note that if we'd have mixed nodes,
2077 * we'd only try to offline the last determined one -- which is good
2078 * enough for the cases we care about.
2080 rc = walk_memory_blocks(start, size, &nid, check_memblock_offlined_cb);
2085 * We only support removing memory added with MHP_MEMMAP_ON_MEMORY in
2086 * the same granularity it was added - a single memory block.
2088 if (mhp_memmap_on_memory()) {
2089 nr_vmemmap_pages = walk_memory_blocks(start, size, NULL,
2090 get_nr_vmemmap_pages_cb);
2091 if (nr_vmemmap_pages) {
2092 if (size != memory_block_size_bytes()) {
2093 pr_warn("Refuse to remove %#llx - %#llx,"
2094 "wrong granularity\n",
2095 start, start + size);
2100 * Let remove_pmd_table->free_hugepage_table do the
2101 * right thing if we used vmem_altmap when hot-adding
2104 mhp_altmap.alloc = nr_vmemmap_pages;
2105 altmap = &mhp_altmap;
2109 /* remove memmap entry */
2110 firmware_map_remove(start, start + size, "System RAM");
2113 * Memory block device removal under the device_hotplug_lock is
2114 * a barrier against racing online attempts.
2116 remove_memory_block_devices(start, size);
2118 mem_hotplug_begin();
2120 arch_remove_memory(start, size, altmap);
2122 if (IS_ENABLED(CONFIG_ARCH_KEEP_MEMBLOCK)) {
2123 memblock_phys_free(start, size);
2124 memblock_remove(start, size);
2127 release_mem_region_adjustable(start, size);
2129 if (nid != NUMA_NO_NODE)
2130 try_offline_node(nid);
2137 * __remove_memory - Remove memory if every memory block is offline
2138 * @start: physical address of the region to remove
2139 * @size: size of the region to remove
2141 * NOTE: The caller must call lock_device_hotplug() to serialize hotplug
2142 * and online/offline operations before this call, as required by
2143 * try_offline_node().
2145 void __remove_memory(u64 start, u64 size)
2149 * trigger BUG() if some memory is not offlined prior to calling this
2152 if (try_remove_memory(start, size))
2157 * Remove memory if every memory block is offline, otherwise return -EBUSY is
2158 * some memory is not offline
2160 int remove_memory(u64 start, u64 size)
2164 lock_device_hotplug();
2165 rc = try_remove_memory(start, size);
2166 unlock_device_hotplug();
2170 EXPORT_SYMBOL_GPL(remove_memory);
2172 static int try_offline_memory_block(struct memory_block *mem, void *arg)
2174 uint8_t online_type = MMOP_ONLINE_KERNEL;
2175 uint8_t **online_types = arg;
2180 * Sense the online_type via the zone of the memory block. Offlining
2181 * with multiple zones within one memory block will be rejected
2182 * by offlining code ... so we don't care about that.
2184 page = pfn_to_online_page(section_nr_to_pfn(mem->start_section_nr));
2185 if (page && zone_idx(page_zone(page)) == ZONE_MOVABLE)
2186 online_type = MMOP_ONLINE_MOVABLE;
2188 rc = device_offline(&mem->dev);
2190 * Default is MMOP_OFFLINE - change it only if offlining succeeded,
2191 * so try_reonline_memory_block() can do the right thing.
2194 **online_types = online_type;
2197 /* Ignore if already offline. */
2198 return rc < 0 ? rc : 0;
2201 static int try_reonline_memory_block(struct memory_block *mem, void *arg)
2203 uint8_t **online_types = arg;
2206 if (**online_types != MMOP_OFFLINE) {
2207 mem->online_type = **online_types;
2208 rc = device_online(&mem->dev);
2210 pr_warn("%s: Failed to re-online memory: %d",
2214 /* Continue processing all remaining memory blocks. */
2220 * Try to offline and remove memory. Might take a long time to finish in case
2221 * memory is still in use. Primarily useful for memory devices that logically
2222 * unplugged all memory (so it's no longer in use) and want to offline + remove
2225 int offline_and_remove_memory(u64 start, u64 size)
2227 const unsigned long mb_count = size / memory_block_size_bytes();
2228 uint8_t *online_types, *tmp;
2231 if (!IS_ALIGNED(start, memory_block_size_bytes()) ||
2232 !IS_ALIGNED(size, memory_block_size_bytes()) || !size)
2236 * We'll remember the old online type of each memory block, so we can
2237 * try to revert whatever we did when offlining one memory block fails
2238 * after offlining some others succeeded.
2240 online_types = kmalloc_array(mb_count, sizeof(*online_types),
2245 * Initialize all states to MMOP_OFFLINE, so when we abort processing in
2246 * try_offline_memory_block(), we'll skip all unprocessed blocks in
2247 * try_reonline_memory_block().
2249 memset(online_types, MMOP_OFFLINE, mb_count);
2251 lock_device_hotplug();
2254 rc = walk_memory_blocks(start, size, &tmp, try_offline_memory_block);
2257 * In case we succeeded to offline all memory, remove it.
2258 * This cannot fail as it cannot get onlined in the meantime.
2261 rc = try_remove_memory(start, size);
2263 pr_err("%s: Failed to remove memory: %d", __func__, rc);
2267 * Rollback what we did. While memory onlining might theoretically fail
2268 * (nacked by a notifier), it barely ever happens.
2272 walk_memory_blocks(start, size, &tmp,
2273 try_reonline_memory_block);
2275 unlock_device_hotplug();
2277 kfree(online_types);
2280 EXPORT_SYMBOL_GPL(offline_and_remove_memory);
2281 #endif /* CONFIG_MEMORY_HOTREMOVE */