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>
39 #include <asm/tlbflush.h>
46 * memory_hotplug.memmap_on_memory parameter
48 static bool memmap_on_memory __ro_after_init;
49 #ifdef CONFIG_MHP_MEMMAP_ON_MEMORY
50 module_param(memmap_on_memory, bool, 0444);
51 MODULE_PARM_DESC(memmap_on_memory, "Enable memmap on memory for memory hotplug");
55 ONLINE_POLICY_CONTIG_ZONES = 0,
56 ONLINE_POLICY_AUTO_MOVABLE,
59 static const char * const online_policy_to_str[] = {
60 [ONLINE_POLICY_CONTIG_ZONES] = "contig-zones",
61 [ONLINE_POLICY_AUTO_MOVABLE] = "auto-movable",
64 static int set_online_policy(const char *val, const struct kernel_param *kp)
66 int ret = sysfs_match_string(online_policy_to_str, val);
70 *((int *)kp->arg) = ret;
74 static int get_online_policy(char *buffer, const struct kernel_param *kp)
76 return sprintf(buffer, "%s\n", online_policy_to_str[*((int *)kp->arg)]);
80 * memory_hotplug.online_policy: configure online behavior when onlining without
81 * specifying a zone (MMOP_ONLINE)
83 * "contig-zones": keep zone contiguous
84 * "auto-movable": online memory to ZONE_MOVABLE if the configuration
85 * (auto_movable_ratio, auto_movable_numa_aware) allows for it
87 static int online_policy __read_mostly = ONLINE_POLICY_CONTIG_ZONES;
88 static const struct kernel_param_ops online_policy_ops = {
89 .set = set_online_policy,
90 .get = get_online_policy,
92 module_param_cb(online_policy, &online_policy_ops, &online_policy, 0644);
93 MODULE_PARM_DESC(online_policy,
94 "Set the online policy (\"contig-zones\", \"auto-movable\") "
95 "Default: \"contig-zones\"");
98 * memory_hotplug.auto_movable_ratio: specify maximum MOVABLE:KERNEL ratio
100 * The ratio represent an upper limit and the kernel might decide to not
101 * online some memory to ZONE_MOVABLE -- e.g., because hotplugged KERNEL memory
102 * doesn't allow for more MOVABLE memory.
104 static unsigned int auto_movable_ratio __read_mostly = 301;
105 module_param(auto_movable_ratio, uint, 0644);
106 MODULE_PARM_DESC(auto_movable_ratio,
107 "Set the maximum ratio of MOVABLE:KERNEL memory in the system "
108 "in percent for \"auto-movable\" online policy. Default: 301");
111 * memory_hotplug.auto_movable_numa_aware: consider numa node stats
114 static bool auto_movable_numa_aware __read_mostly = true;
115 module_param(auto_movable_numa_aware, bool, 0644);
116 MODULE_PARM_DESC(auto_movable_numa_aware,
117 "Consider numa node stats in addition to global stats in "
118 "\"auto-movable\" online policy. Default: true");
119 #endif /* CONFIG_NUMA */
122 * online_page_callback contains pointer to current page onlining function.
123 * Initially it is generic_online_page(). If it is required it could be
124 * changed by calling set_online_page_callback() for callback registration
125 * and restore_online_page_callback() for generic callback restore.
128 static online_page_callback_t online_page_callback = generic_online_page;
129 static DEFINE_MUTEX(online_page_callback_lock);
131 DEFINE_STATIC_PERCPU_RWSEM(mem_hotplug_lock);
133 void get_online_mems(void)
135 percpu_down_read(&mem_hotplug_lock);
138 void put_online_mems(void)
140 percpu_up_read(&mem_hotplug_lock);
143 bool movable_node_enabled = false;
145 #ifndef CONFIG_MEMORY_HOTPLUG_DEFAULT_ONLINE
146 int mhp_default_online_type = MMOP_OFFLINE;
148 int mhp_default_online_type = MMOP_ONLINE;
151 static int __init setup_memhp_default_state(char *str)
153 const int online_type = mhp_online_type_from_str(str);
155 if (online_type >= 0)
156 mhp_default_online_type = online_type;
160 __setup("memhp_default_state=", setup_memhp_default_state);
162 void mem_hotplug_begin(void)
165 percpu_down_write(&mem_hotplug_lock);
168 void mem_hotplug_done(void)
170 percpu_up_write(&mem_hotplug_lock);
174 u64 max_mem_size = U64_MAX;
176 /* add this memory to iomem resource */
177 static struct resource *register_memory_resource(u64 start, u64 size,
178 const char *resource_name)
180 struct resource *res;
181 unsigned long flags = IORESOURCE_SYSTEM_RAM | IORESOURCE_BUSY;
183 if (strcmp(resource_name, "System RAM"))
184 flags |= IORESOURCE_SYSRAM_DRIVER_MANAGED;
186 if (!mhp_range_allowed(start, size, true))
187 return ERR_PTR(-E2BIG);
190 * Make sure value parsed from 'mem=' only restricts memory adding
191 * while booting, so that memory hotplug won't be impacted. Please
192 * refer to document of 'mem=' in kernel-parameters.txt for more
195 if (start + size > max_mem_size && system_state < SYSTEM_RUNNING)
196 return ERR_PTR(-E2BIG);
199 * Request ownership of the new memory range. This might be
200 * a child of an existing resource that was present but
201 * not marked as busy.
203 res = __request_region(&iomem_resource, start, size,
204 resource_name, flags);
207 pr_debug("Unable to reserve System RAM region: %016llx->%016llx\n",
208 start, start + size);
209 return ERR_PTR(-EEXIST);
214 static void release_memory_resource(struct resource *res)
218 release_resource(res);
222 static int check_pfn_span(unsigned long pfn, unsigned long nr_pages,
226 * Disallow all operations smaller than a sub-section and only
227 * allow operations smaller than a section for
228 * SPARSEMEM_VMEMMAP. Note that check_hotplug_memory_range()
229 * enforces a larger memory_block_size_bytes() granularity for
230 * memory that will be marked online, so this check should only
231 * fire for direct arch_{add,remove}_memory() users outside of
232 * add_memory_resource().
234 unsigned long min_align;
236 if (IS_ENABLED(CONFIG_SPARSEMEM_VMEMMAP))
237 min_align = PAGES_PER_SUBSECTION;
239 min_align = PAGES_PER_SECTION;
240 if (!IS_ALIGNED(pfn, min_align)
241 || !IS_ALIGNED(nr_pages, min_align)) {
242 WARN(1, "Misaligned __%s_pages start: %#lx end: #%lx\n",
243 reason, pfn, pfn + nr_pages - 1);
250 * Return page for the valid pfn only if the page is online. All pfn
251 * walkers which rely on the fully initialized page->flags and others
252 * should use this rather than pfn_valid && pfn_to_page
254 struct page *pfn_to_online_page(unsigned long pfn)
256 unsigned long nr = pfn_to_section_nr(pfn);
257 struct dev_pagemap *pgmap;
258 struct mem_section *ms;
260 if (nr >= NR_MEM_SECTIONS)
263 ms = __nr_to_section(nr);
264 if (!online_section(ms))
268 * Save some code text when online_section() +
269 * pfn_section_valid() are sufficient.
271 if (IS_ENABLED(CONFIG_HAVE_ARCH_PFN_VALID) && !pfn_valid(pfn))
274 if (!pfn_section_valid(ms, pfn))
277 if (!online_device_section(ms))
278 return pfn_to_page(pfn);
281 * Slowpath: when ZONE_DEVICE collides with
282 * ZONE_{NORMAL,MOVABLE} within the same section some pfns in
283 * the section may be 'offline' but 'valid'. Only
284 * get_dev_pagemap() can determine sub-section online status.
286 pgmap = get_dev_pagemap(pfn, NULL);
287 put_dev_pagemap(pgmap);
289 /* The presence of a pgmap indicates ZONE_DEVICE offline pfn */
293 return pfn_to_page(pfn);
295 EXPORT_SYMBOL_GPL(pfn_to_online_page);
298 * Reasonably generic function for adding memory. It is
299 * expected that archs that support memory hotplug will
300 * call this function after deciding the zone to which to
303 int __ref __add_pages(int nid, unsigned long pfn, unsigned long nr_pages,
304 struct mhp_params *params)
306 const unsigned long end_pfn = pfn + nr_pages;
307 unsigned long cur_nr_pages;
309 struct vmem_altmap *altmap = params->altmap;
311 if (WARN_ON_ONCE(!params->pgprot.pgprot))
314 VM_BUG_ON(!mhp_range_allowed(PFN_PHYS(pfn), nr_pages * PAGE_SIZE, false));
318 * Validate altmap is within bounds of the total request
320 if (altmap->base_pfn != pfn
321 || vmem_altmap_offset(altmap) > nr_pages) {
322 pr_warn_once("memory add fail, invalid altmap\n");
328 err = check_pfn_span(pfn, nr_pages, "add");
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);
341 vmemmap_populate_print_last();
345 /* find the smallest valid pfn in the range [start_pfn, end_pfn) */
346 static unsigned long find_smallest_section_pfn(int nid, struct zone *zone,
347 unsigned long start_pfn,
348 unsigned long end_pfn)
350 for (; start_pfn < end_pfn; start_pfn += PAGES_PER_SUBSECTION) {
351 if (unlikely(!pfn_to_online_page(start_pfn)))
354 if (unlikely(pfn_to_nid(start_pfn) != nid))
357 if (zone != page_zone(pfn_to_page(start_pfn)))
366 /* find the biggest valid pfn in the range [start_pfn, end_pfn). */
367 static unsigned long find_biggest_section_pfn(int nid, struct zone *zone,
368 unsigned long start_pfn,
369 unsigned long end_pfn)
373 /* pfn is the end pfn of a memory section. */
375 for (; pfn >= start_pfn; pfn -= PAGES_PER_SUBSECTION) {
376 if (unlikely(!pfn_to_online_page(pfn)))
379 if (unlikely(pfn_to_nid(pfn) != nid))
382 if (zone != page_zone(pfn_to_page(pfn)))
391 static void shrink_zone_span(struct zone *zone, unsigned long start_pfn,
392 unsigned long end_pfn)
395 int nid = zone_to_nid(zone);
397 if (zone->zone_start_pfn == start_pfn) {
399 * If the section is smallest section in the zone, it need
400 * shrink zone->zone_start_pfn and zone->zone_spanned_pages.
401 * In this case, we find second smallest valid mem_section
402 * for shrinking zone.
404 pfn = find_smallest_section_pfn(nid, zone, end_pfn,
407 zone->spanned_pages = zone_end_pfn(zone) - pfn;
408 zone->zone_start_pfn = pfn;
410 zone->zone_start_pfn = 0;
411 zone->spanned_pages = 0;
413 } else if (zone_end_pfn(zone) == end_pfn) {
415 * If the section is biggest section in the zone, it need
416 * shrink zone->spanned_pages.
417 * In this case, we find second biggest valid mem_section for
420 pfn = find_biggest_section_pfn(nid, zone, zone->zone_start_pfn,
423 zone->spanned_pages = pfn - zone->zone_start_pfn + 1;
425 zone->zone_start_pfn = 0;
426 zone->spanned_pages = 0;
431 static void update_pgdat_span(struct pglist_data *pgdat)
433 unsigned long node_start_pfn = 0, node_end_pfn = 0;
436 for (zone = pgdat->node_zones;
437 zone < pgdat->node_zones + MAX_NR_ZONES; zone++) {
438 unsigned long end_pfn = zone_end_pfn(zone);
440 /* No need to lock the zones, they can't change. */
441 if (!zone->spanned_pages)
444 node_start_pfn = zone->zone_start_pfn;
445 node_end_pfn = end_pfn;
449 if (end_pfn > node_end_pfn)
450 node_end_pfn = end_pfn;
451 if (zone->zone_start_pfn < node_start_pfn)
452 node_start_pfn = zone->zone_start_pfn;
455 pgdat->node_start_pfn = node_start_pfn;
456 pgdat->node_spanned_pages = node_end_pfn - node_start_pfn;
459 void __ref remove_pfn_range_from_zone(struct zone *zone,
460 unsigned long start_pfn,
461 unsigned long nr_pages)
463 const unsigned long end_pfn = start_pfn + nr_pages;
464 struct pglist_data *pgdat = zone->zone_pgdat;
465 unsigned long pfn, cur_nr_pages;
467 /* Poison struct pages because they are now uninitialized again. */
468 for (pfn = start_pfn; pfn < end_pfn; pfn += cur_nr_pages) {
471 /* Select all remaining pages up to the next section boundary */
473 min(end_pfn - pfn, SECTION_ALIGN_UP(pfn + 1) - pfn);
474 page_init_poison(pfn_to_page(pfn),
475 sizeof(struct page) * cur_nr_pages);
479 * Zone shrinking code cannot properly deal with ZONE_DEVICE. So
480 * we will not try to shrink the zones - which is okay as
481 * set_zone_contiguous() cannot deal with ZONE_DEVICE either way.
483 if (zone_is_zone_device(zone))
486 clear_zone_contiguous(zone);
488 shrink_zone_span(zone, start_pfn, start_pfn + nr_pages);
489 update_pgdat_span(pgdat);
491 set_zone_contiguous(zone);
494 static void __remove_section(unsigned long pfn, unsigned long nr_pages,
495 unsigned long map_offset,
496 struct vmem_altmap *altmap)
498 struct mem_section *ms = __pfn_to_section(pfn);
500 if (WARN_ON_ONCE(!valid_section(ms)))
503 sparse_remove_section(ms, pfn, nr_pages, map_offset, altmap);
507 * __remove_pages() - remove sections of pages
508 * @pfn: starting pageframe (must be aligned to start of a section)
509 * @nr_pages: number of pages to remove (must be multiple of section size)
510 * @altmap: alternative device page map or %NULL if default memmap is used
512 * Generic helper function to remove section mappings and sysfs entries
513 * for the section of the memory we are removing. Caller needs to make
514 * sure that pages are marked reserved and zones are adjust properly by
515 * calling offline_pages().
517 void __remove_pages(unsigned long pfn, unsigned long nr_pages,
518 struct vmem_altmap *altmap)
520 const unsigned long end_pfn = pfn + nr_pages;
521 unsigned long cur_nr_pages;
522 unsigned long map_offset = 0;
524 map_offset = vmem_altmap_offset(altmap);
526 if (check_pfn_span(pfn, nr_pages, "remove"))
529 for (; pfn < end_pfn; pfn += cur_nr_pages) {
531 /* Select all remaining pages up to the next section boundary */
532 cur_nr_pages = min(end_pfn - pfn,
533 SECTION_ALIGN_UP(pfn + 1) - pfn);
534 __remove_section(pfn, cur_nr_pages, map_offset, altmap);
539 int set_online_page_callback(online_page_callback_t callback)
544 mutex_lock(&online_page_callback_lock);
546 if (online_page_callback == generic_online_page) {
547 online_page_callback = callback;
551 mutex_unlock(&online_page_callback_lock);
556 EXPORT_SYMBOL_GPL(set_online_page_callback);
558 int restore_online_page_callback(online_page_callback_t callback)
563 mutex_lock(&online_page_callback_lock);
565 if (online_page_callback == callback) {
566 online_page_callback = generic_online_page;
570 mutex_unlock(&online_page_callback_lock);
575 EXPORT_SYMBOL_GPL(restore_online_page_callback);
577 void generic_online_page(struct page *page, unsigned int order)
580 * Freeing the page with debug_pagealloc enabled will try to unmap it,
581 * so we should map it first. This is better than introducing a special
582 * case in page freeing fast path.
584 debug_pagealloc_map_pages(page, 1 << order);
585 __free_pages_core(page, order);
586 totalram_pages_add(1UL << order);
588 EXPORT_SYMBOL_GPL(generic_online_page);
590 static void online_pages_range(unsigned long start_pfn, unsigned long nr_pages)
592 const unsigned long end_pfn = start_pfn + nr_pages;
596 * Online the pages in MAX_ORDER - 1 aligned chunks. The callback might
597 * decide to not expose all pages to the buddy (e.g., expose them
598 * later). We account all pages as being online and belonging to this
600 * When using memmap_on_memory, the range might not be aligned to
601 * MAX_ORDER_NR_PAGES - 1, but pageblock aligned. __ffs() will detect
602 * this and the first chunk to online will be pageblock_nr_pages.
604 for (pfn = start_pfn; pfn < end_pfn;) {
605 int order = min(MAX_ORDER - 1UL, __ffs(pfn));
607 (*online_page_callback)(pfn_to_page(pfn), order);
608 pfn += (1UL << order);
611 /* mark all involved sections as online */
612 online_mem_sections(start_pfn, end_pfn);
615 /* check which state of node_states will be changed when online memory */
616 static void node_states_check_changes_online(unsigned long nr_pages,
617 struct zone *zone, struct memory_notify *arg)
619 int nid = zone_to_nid(zone);
621 arg->status_change_nid = NUMA_NO_NODE;
622 arg->status_change_nid_normal = NUMA_NO_NODE;
624 if (!node_state(nid, N_MEMORY))
625 arg->status_change_nid = nid;
626 if (zone_idx(zone) <= ZONE_NORMAL && !node_state(nid, N_NORMAL_MEMORY))
627 arg->status_change_nid_normal = nid;
630 static void node_states_set_node(int node, struct memory_notify *arg)
632 if (arg->status_change_nid_normal >= 0)
633 node_set_state(node, N_NORMAL_MEMORY);
635 if (arg->status_change_nid >= 0)
636 node_set_state(node, N_MEMORY);
639 static void __meminit resize_zone_range(struct zone *zone, unsigned long start_pfn,
640 unsigned long nr_pages)
642 unsigned long old_end_pfn = zone_end_pfn(zone);
644 if (zone_is_empty(zone) || start_pfn < zone->zone_start_pfn)
645 zone->zone_start_pfn = start_pfn;
647 zone->spanned_pages = max(start_pfn + nr_pages, old_end_pfn) - zone->zone_start_pfn;
650 static void __meminit resize_pgdat_range(struct pglist_data *pgdat, unsigned long start_pfn,
651 unsigned long nr_pages)
653 unsigned long old_end_pfn = pgdat_end_pfn(pgdat);
655 if (!pgdat->node_spanned_pages || start_pfn < pgdat->node_start_pfn)
656 pgdat->node_start_pfn = start_pfn;
658 pgdat->node_spanned_pages = max(start_pfn + nr_pages, old_end_pfn) - pgdat->node_start_pfn;
662 static void section_taint_zone_device(unsigned long pfn)
664 struct mem_section *ms = __pfn_to_section(pfn);
666 ms->section_mem_map |= SECTION_TAINT_ZONE_DEVICE;
670 * Associate the pfn range with the given zone, initializing the memmaps
671 * and resizing the pgdat/zone data to span the added pages. After this
672 * call, all affected pages are PG_reserved.
674 * All aligned pageblocks are initialized to the specified migratetype
675 * (usually MIGRATE_MOVABLE). Besides setting the migratetype, no related
676 * zone stats (e.g., nr_isolate_pageblock) are touched.
678 void __ref move_pfn_range_to_zone(struct zone *zone, unsigned long start_pfn,
679 unsigned long nr_pages,
680 struct vmem_altmap *altmap, int migratetype)
682 struct pglist_data *pgdat = zone->zone_pgdat;
683 int nid = pgdat->node_id;
685 clear_zone_contiguous(zone);
687 if (zone_is_empty(zone))
688 init_currently_empty_zone(zone, start_pfn, nr_pages);
689 resize_zone_range(zone, start_pfn, nr_pages);
690 resize_pgdat_range(pgdat, start_pfn, nr_pages);
693 * Subsection population requires care in pfn_to_online_page().
694 * Set the taint to enable the slow path detection of
695 * ZONE_DEVICE pages in an otherwise ZONE_{NORMAL,MOVABLE}
698 if (zone_is_zone_device(zone)) {
699 if (!IS_ALIGNED(start_pfn, PAGES_PER_SECTION))
700 section_taint_zone_device(start_pfn);
701 if (!IS_ALIGNED(start_pfn + nr_pages, PAGES_PER_SECTION))
702 section_taint_zone_device(start_pfn + nr_pages);
706 * TODO now we have a visible range of pages which are not associated
707 * with their zone properly. Not nice but set_pfnblock_flags_mask
708 * expects the zone spans the pfn range. All the pages in the range
709 * are reserved so nobody should be touching them so we should be safe
711 memmap_init_range(nr_pages, nid, zone_idx(zone), start_pfn, 0,
712 MEMINIT_HOTPLUG, altmap, migratetype);
714 set_zone_contiguous(zone);
717 struct auto_movable_stats {
718 unsigned long kernel_early_pages;
719 unsigned long movable_pages;
722 static void auto_movable_stats_account_zone(struct auto_movable_stats *stats,
725 if (zone_idx(zone) == ZONE_MOVABLE) {
726 stats->movable_pages += zone->present_pages;
728 stats->kernel_early_pages += zone->present_early_pages;
731 * CMA pages (never on hotplugged memory) behave like
734 stats->movable_pages += zone->cma_pages;
735 stats->kernel_early_pages -= zone->cma_pages;
736 #endif /* CONFIG_CMA */
739 struct auto_movable_group_stats {
740 unsigned long movable_pages;
741 unsigned long req_kernel_early_pages;
744 static int auto_movable_stats_account_group(struct memory_group *group,
747 const int ratio = READ_ONCE(auto_movable_ratio);
748 struct auto_movable_group_stats *stats = arg;
752 * We don't support modifying the config while the auto-movable online
753 * policy is already enabled. Just avoid the division by zero below.
759 * Calculate how many early kernel pages this group requires to
760 * satisfy the configured zone ratio.
762 pages = group->present_movable_pages * 100 / ratio;
763 pages -= group->present_kernel_pages;
766 stats->req_kernel_early_pages += pages;
767 stats->movable_pages += group->present_movable_pages;
771 static bool auto_movable_can_online_movable(int nid, struct memory_group *group,
772 unsigned long nr_pages)
774 unsigned long kernel_early_pages, movable_pages;
775 struct auto_movable_group_stats group_stats = {};
776 struct auto_movable_stats stats = {};
777 pg_data_t *pgdat = NODE_DATA(nid);
781 /* Walk all relevant zones and collect MOVABLE vs. KERNEL stats. */
782 if (nid == NUMA_NO_NODE) {
783 /* TODO: cache values */
784 for_each_populated_zone(zone)
785 auto_movable_stats_account_zone(&stats, zone);
787 for (i = 0; i < MAX_NR_ZONES; i++) {
788 zone = pgdat->node_zones + i;
789 if (populated_zone(zone))
790 auto_movable_stats_account_zone(&stats, zone);
794 kernel_early_pages = stats.kernel_early_pages;
795 movable_pages = stats.movable_pages;
798 * Kernel memory inside dynamic memory group allows for more MOVABLE
799 * memory within the same group. Remove the effect of all but the
800 * current group from the stats.
802 walk_dynamic_memory_groups(nid, auto_movable_stats_account_group,
803 group, &group_stats);
804 if (kernel_early_pages <= group_stats.req_kernel_early_pages)
806 kernel_early_pages -= group_stats.req_kernel_early_pages;
807 movable_pages -= group_stats.movable_pages;
809 if (group && group->is_dynamic)
810 kernel_early_pages += group->present_kernel_pages;
813 * Test if we could online the given number of pages to ZONE_MOVABLE
814 * and still stay in the configured ratio.
816 movable_pages += nr_pages;
817 return movable_pages <= (auto_movable_ratio * kernel_early_pages) / 100;
821 * Returns a default kernel memory zone for the given pfn range.
822 * If no kernel zone covers this pfn range it will automatically go
823 * to the ZONE_NORMAL.
825 static struct zone *default_kernel_zone_for_pfn(int nid, unsigned long start_pfn,
826 unsigned long nr_pages)
828 struct pglist_data *pgdat = NODE_DATA(nid);
831 for (zid = 0; zid <= ZONE_NORMAL; zid++) {
832 struct zone *zone = &pgdat->node_zones[zid];
834 if (zone_intersects(zone, start_pfn, nr_pages))
838 return &pgdat->node_zones[ZONE_NORMAL];
842 * Determine to which zone to online memory dynamically based on user
843 * configuration and system stats. We care about the following ratio:
847 * Whereby MOVABLE is memory in ZONE_MOVABLE and KERNEL is memory in
848 * one of the kernel zones. CMA pages inside one of the kernel zones really
849 * behaves like ZONE_MOVABLE, so we treat them accordingly.
851 * We don't allow for hotplugged memory in a KERNEL zone to increase the
852 * amount of MOVABLE memory we can have, so we end up with:
854 * MOVABLE : KERNEL_EARLY
856 * Whereby KERNEL_EARLY is memory in one of the kernel zones, available sinze
857 * boot. We base our calculation on KERNEL_EARLY internally, because:
859 * a) Hotplugged memory in one of the kernel zones can sometimes still get
860 * hotunplugged, especially when hot(un)plugging individual memory blocks.
861 * There is no coordination across memory devices, therefore "automatic"
862 * hotunplugging, as implemented in hypervisors, could result in zone
864 * b) Early/boot memory in one of the kernel zones can usually not get
865 * hotunplugged again (e.g., no firmware interface to unplug, fragmented
866 * with unmovable allocations). While there are corner cases where it might
867 * still work, it is barely relevant in practice.
869 * Exceptions are dynamic memory groups, which allow for more MOVABLE
870 * memory within the same memory group -- because in that case, there is
871 * coordination within the single memory device managed by a single driver.
873 * We rely on "present pages" instead of "managed pages", as the latter is
874 * highly unreliable and dynamic in virtualized environments, and does not
875 * consider boot time allocations. For example, memory ballooning adjusts the
876 * managed pages when inflating/deflating the balloon, and balloon compaction
877 * can even migrate inflated pages between zones.
879 * Using "present pages" is better but some things to keep in mind are:
881 * a) Some memblock allocations, such as for the crashkernel area, are
882 * effectively unused by the kernel, yet they account to "present pages".
883 * Fortunately, these allocations are comparatively small in relevant setups
884 * (e.g., fraction of system memory).
885 * b) Some hotplugged memory blocks in virtualized environments, esecially
886 * hotplugged by virtio-mem, look like they are completely present, however,
887 * only parts of the memory block are actually currently usable.
888 * "present pages" is an upper limit that can get reached at runtime. As
889 * we base our calculations on KERNEL_EARLY, this is not an issue.
891 static struct zone *auto_movable_zone_for_pfn(int nid,
892 struct memory_group *group,
894 unsigned long nr_pages)
896 unsigned long online_pages = 0, max_pages, end_pfn;
899 if (!auto_movable_ratio)
902 if (group && !group->is_dynamic) {
903 max_pages = group->s.max_pages;
904 online_pages = group->present_movable_pages;
906 /* If anything is !MOVABLE online the rest !MOVABLE. */
907 if (group->present_kernel_pages)
909 } else if (!group || group->d.unit_pages == nr_pages) {
910 max_pages = nr_pages;
912 max_pages = group->d.unit_pages;
914 * Take a look at all online sections in the current unit.
915 * We can safely assume that all pages within a section belong
916 * to the same zone, because dynamic memory groups only deal
917 * with hotplugged memory.
919 pfn = ALIGN_DOWN(pfn, group->d.unit_pages);
920 end_pfn = pfn + group->d.unit_pages;
921 for (; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
922 page = pfn_to_online_page(pfn);
925 /* If anything is !MOVABLE online the rest !MOVABLE. */
926 if (page_zonenum(page) != ZONE_MOVABLE)
928 online_pages += PAGES_PER_SECTION;
933 * Online MOVABLE if we could *currently* online all remaining parts
934 * MOVABLE. We expect to (add+) online them immediately next, so if
935 * nobody interferes, all will be MOVABLE if possible.
937 nr_pages = max_pages - online_pages;
938 if (!auto_movable_can_online_movable(NUMA_NO_NODE, group, nr_pages))
942 if (auto_movable_numa_aware &&
943 !auto_movable_can_online_movable(nid, group, nr_pages))
945 #endif /* CONFIG_NUMA */
947 return &NODE_DATA(nid)->node_zones[ZONE_MOVABLE];
949 return default_kernel_zone_for_pfn(nid, pfn, nr_pages);
952 static inline struct zone *default_zone_for_pfn(int nid, unsigned long start_pfn,
953 unsigned long nr_pages)
955 struct zone *kernel_zone = default_kernel_zone_for_pfn(nid, start_pfn,
957 struct zone *movable_zone = &NODE_DATA(nid)->node_zones[ZONE_MOVABLE];
958 bool in_kernel = zone_intersects(kernel_zone, start_pfn, nr_pages);
959 bool in_movable = zone_intersects(movable_zone, start_pfn, nr_pages);
962 * We inherit the existing zone in a simple case where zones do not
963 * overlap in the given range
965 if (in_kernel ^ in_movable)
966 return (in_kernel) ? kernel_zone : movable_zone;
969 * If the range doesn't belong to any zone or two zones overlap in the
970 * given range then we use movable zone only if movable_node is
971 * enabled because we always online to a kernel zone by default.
973 return movable_node_enabled ? movable_zone : kernel_zone;
976 struct zone *zone_for_pfn_range(int online_type, int nid,
977 struct memory_group *group, unsigned long start_pfn,
978 unsigned long nr_pages)
980 if (online_type == MMOP_ONLINE_KERNEL)
981 return default_kernel_zone_for_pfn(nid, start_pfn, nr_pages);
983 if (online_type == MMOP_ONLINE_MOVABLE)
984 return &NODE_DATA(nid)->node_zones[ZONE_MOVABLE];
986 if (online_policy == ONLINE_POLICY_AUTO_MOVABLE)
987 return auto_movable_zone_for_pfn(nid, group, start_pfn, nr_pages);
989 return default_zone_for_pfn(nid, start_pfn, nr_pages);
993 * This function should only be called by memory_block_{online,offline},
994 * and {online,offline}_pages.
996 void adjust_present_page_count(struct page *page, struct memory_group *group,
999 struct zone *zone = page_zone(page);
1000 const bool movable = zone_idx(zone) == ZONE_MOVABLE;
1003 * We only support onlining/offlining/adding/removing of complete
1004 * memory blocks; therefore, either all is either early or hotplugged.
1006 if (early_section(__pfn_to_section(page_to_pfn(page))))
1007 zone->present_early_pages += nr_pages;
1008 zone->present_pages += nr_pages;
1009 zone->zone_pgdat->node_present_pages += nr_pages;
1011 if (group && movable)
1012 group->present_movable_pages += nr_pages;
1013 else if (group && !movable)
1014 group->present_kernel_pages += nr_pages;
1017 int mhp_init_memmap_on_memory(unsigned long pfn, unsigned long nr_pages,
1020 unsigned long end_pfn = pfn + nr_pages;
1023 ret = kasan_add_zero_shadow(__va(PFN_PHYS(pfn)), PFN_PHYS(nr_pages));
1027 move_pfn_range_to_zone(zone, pfn, nr_pages, NULL, MIGRATE_UNMOVABLE);
1030 * It might be that the vmemmap_pages fully span sections. If that is
1031 * the case, mark those sections online here as otherwise they will be
1034 if (nr_pages >= PAGES_PER_SECTION)
1035 online_mem_sections(pfn, ALIGN_DOWN(end_pfn, PAGES_PER_SECTION));
1040 void mhp_deinit_memmap_on_memory(unsigned long pfn, unsigned long nr_pages)
1042 unsigned long end_pfn = pfn + nr_pages;
1045 * It might be that the vmemmap_pages fully span sections. If that is
1046 * the case, mark those sections offline here as otherwise they will be
1049 if (nr_pages >= PAGES_PER_SECTION)
1050 offline_mem_sections(pfn, ALIGN_DOWN(end_pfn, PAGES_PER_SECTION));
1053 * The pages associated with this vmemmap have been offlined, so
1054 * we can reset its state here.
1056 remove_pfn_range_from_zone(page_zone(pfn_to_page(pfn)), pfn, nr_pages);
1057 kasan_remove_zero_shadow(__va(PFN_PHYS(pfn)), PFN_PHYS(nr_pages));
1060 int __ref online_pages(unsigned long pfn, unsigned long nr_pages,
1061 struct zone *zone, struct memory_group *group)
1063 unsigned long flags;
1064 int need_zonelists_rebuild = 0;
1065 const int nid = zone_to_nid(zone);
1067 struct memory_notify arg;
1070 * {on,off}lining is constrained to full memory sections (or more
1071 * precisely to memory blocks from the user space POV).
1072 * memmap_on_memory is an exception because it reserves initial part
1073 * of the physical memory space for vmemmaps. That space is pageblock
1076 if (WARN_ON_ONCE(!nr_pages ||
1077 !IS_ALIGNED(pfn, pageblock_nr_pages) ||
1078 !IS_ALIGNED(pfn + nr_pages, PAGES_PER_SECTION)))
1081 mem_hotplug_begin();
1083 /* associate pfn range with the zone */
1084 move_pfn_range_to_zone(zone, pfn, nr_pages, NULL, MIGRATE_ISOLATE);
1086 arg.start_pfn = pfn;
1087 arg.nr_pages = nr_pages;
1088 node_states_check_changes_online(nr_pages, zone, &arg);
1090 ret = memory_notify(MEM_GOING_ONLINE, &arg);
1091 ret = notifier_to_errno(ret);
1093 goto failed_addition;
1096 * Fixup the number of isolated pageblocks before marking the sections
1097 * onlining, such that undo_isolate_page_range() works correctly.
1099 spin_lock_irqsave(&zone->lock, flags);
1100 zone->nr_isolate_pageblock += nr_pages / pageblock_nr_pages;
1101 spin_unlock_irqrestore(&zone->lock, flags);
1104 * If this zone is not populated, then it is not in zonelist.
1105 * This means the page allocator ignores this zone.
1106 * So, zonelist must be updated after online.
1108 if (!populated_zone(zone)) {
1109 need_zonelists_rebuild = 1;
1110 setup_zone_pageset(zone);
1113 online_pages_range(pfn, nr_pages);
1114 adjust_present_page_count(pfn_to_page(pfn), group, nr_pages);
1116 node_states_set_node(nid, &arg);
1117 if (need_zonelists_rebuild)
1118 build_all_zonelists(NULL);
1120 /* Basic onlining is complete, allow allocation of onlined pages. */
1121 undo_isolate_page_range(pfn, pfn + nr_pages, MIGRATE_MOVABLE);
1124 * Freshly onlined pages aren't shuffled (e.g., all pages are placed to
1125 * the tail of the freelist when undoing isolation). Shuffle the whole
1126 * zone to make sure the just onlined pages are properly distributed
1127 * across the whole freelist - to create an initial shuffle.
1131 /* reinitialise watermarks and update pcp limits */
1132 init_per_zone_wmark_min();
1137 writeback_set_ratelimit();
1139 memory_notify(MEM_ONLINE, &arg);
1144 pr_debug("online_pages [mem %#010llx-%#010llx] failed\n",
1145 (unsigned long long) pfn << PAGE_SHIFT,
1146 (((unsigned long long) pfn + nr_pages) << PAGE_SHIFT) - 1);
1147 memory_notify(MEM_CANCEL_ONLINE, &arg);
1148 remove_pfn_range_from_zone(zone, pfn, nr_pages);
1153 static void reset_node_present_pages(pg_data_t *pgdat)
1157 for (z = pgdat->node_zones; z < pgdat->node_zones + MAX_NR_ZONES; z++)
1158 z->present_pages = 0;
1160 pgdat->node_present_pages = 0;
1163 /* we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG */
1164 static pg_data_t __ref *hotadd_new_pgdat(int nid)
1166 struct pglist_data *pgdat;
1168 pgdat = NODE_DATA(nid);
1170 pgdat = arch_alloc_nodedata(nid);
1174 pgdat->per_cpu_nodestats =
1175 alloc_percpu(struct per_cpu_nodestat);
1176 arch_refresh_nodedata(nid, pgdat);
1180 * Reset the nr_zones, order and highest_zoneidx before reuse.
1181 * Note that kswapd will init kswapd_highest_zoneidx properly
1182 * when it starts in the near future.
1184 pgdat->nr_zones = 0;
1185 pgdat->kswapd_order = 0;
1186 pgdat->kswapd_highest_zoneidx = 0;
1187 for_each_online_cpu(cpu) {
1188 struct per_cpu_nodestat *p;
1190 p = per_cpu_ptr(pgdat->per_cpu_nodestats, cpu);
1191 memset(p, 0, sizeof(*p));
1195 /* we can use NODE_DATA(nid) from here */
1196 pgdat->node_id = nid;
1197 pgdat->node_start_pfn = 0;
1199 /* init node's zones as empty zones, we don't have any present pages.*/
1200 free_area_init_core_hotplug(nid);
1203 * The node we allocated has no zone fallback lists. For avoiding
1204 * to access not-initialized zonelist, build here.
1206 build_all_zonelists(pgdat);
1209 * When memory is hot-added, all the memory is in offline state. So
1210 * clear all zones' present_pages because they will be updated in
1211 * online_pages() and offline_pages().
1213 reset_node_managed_pages(pgdat);
1214 reset_node_present_pages(pgdat);
1219 static void rollback_node_hotadd(int nid)
1221 pg_data_t *pgdat = NODE_DATA(nid);
1223 arch_refresh_nodedata(nid, NULL);
1224 free_percpu(pgdat->per_cpu_nodestats);
1225 arch_free_nodedata(pgdat);
1230 * __try_online_node - online a node if offlined
1232 * @set_node_online: Whether we want to online the node
1233 * called by cpu_up() to online a node without onlined memory.
1236 * 1 -> a new node has been allocated
1237 * 0 -> the node is already online
1238 * -ENOMEM -> the node could not be allocated
1240 static int __try_online_node(int nid, bool set_node_online)
1245 if (node_online(nid))
1248 pgdat = hotadd_new_pgdat(nid);
1250 pr_err("Cannot online node %d due to NULL pgdat\n", nid);
1255 if (set_node_online) {
1256 node_set_online(nid);
1257 ret = register_one_node(nid);
1265 * Users of this function always want to online/register the node
1267 int try_online_node(int nid)
1271 mem_hotplug_begin();
1272 ret = __try_online_node(nid, true);
1277 static int check_hotplug_memory_range(u64 start, u64 size)
1279 /* memory range must be block size aligned */
1280 if (!size || !IS_ALIGNED(start, memory_block_size_bytes()) ||
1281 !IS_ALIGNED(size, memory_block_size_bytes())) {
1282 pr_err("Block size [%#lx] unaligned hotplug range: start %#llx, size %#llx",
1283 memory_block_size_bytes(), start, size);
1290 static int online_memory_block(struct memory_block *mem, void *arg)
1292 mem->online_type = mhp_default_online_type;
1293 return device_online(&mem->dev);
1296 bool mhp_supports_memmap_on_memory(unsigned long size)
1298 unsigned long nr_vmemmap_pages = size / PAGE_SIZE;
1299 unsigned long vmemmap_size = nr_vmemmap_pages * sizeof(struct page);
1300 unsigned long remaining_size = size - vmemmap_size;
1303 * Besides having arch support and the feature enabled at runtime, we
1304 * need a few more assumptions to hold true:
1306 * a) We span a single memory block: memory onlining/offlinin;g happens
1307 * in memory block granularity. We don't want the vmemmap of online
1308 * memory blocks to reside on offline memory blocks. In the future,
1309 * we might want to support variable-sized memory blocks to make the
1310 * feature more versatile.
1312 * b) The vmemmap pages span complete PMDs: We don't want vmemmap code
1313 * to populate memory from the altmap for unrelated parts (i.e.,
1314 * other memory blocks)
1316 * c) The vmemmap pages (and thereby the pages that will be exposed to
1317 * the buddy) have to cover full pageblocks: memory onlining/offlining
1318 * code requires applicable ranges to be page-aligned, for example, to
1319 * set the migratetypes properly.
1321 * TODO: Although we have a check here to make sure that vmemmap pages
1322 * fully populate a PMD, it is not the right place to check for
1323 * this. A much better solution involves improving vmemmap code
1324 * to fallback to base pages when trying to populate vmemmap using
1325 * altmap as an alternative source of memory, and we do not exactly
1326 * populate a single PMD.
1328 return memmap_on_memory &&
1329 !hugetlb_free_vmemmap_enabled &&
1330 IS_ENABLED(CONFIG_MHP_MEMMAP_ON_MEMORY) &&
1331 size == memory_block_size_bytes() &&
1332 IS_ALIGNED(vmemmap_size, PMD_SIZE) &&
1333 IS_ALIGNED(remaining_size, (pageblock_nr_pages << PAGE_SHIFT));
1337 * NOTE: The caller must call lock_device_hotplug() to serialize hotplug
1338 * and online/offline operations (triggered e.g. by sysfs).
1340 * we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG
1342 int __ref add_memory_resource(int nid, struct resource *res, mhp_t mhp_flags)
1344 struct mhp_params params = { .pgprot = pgprot_mhp(PAGE_KERNEL) };
1345 enum memblock_flags memblock_flags = MEMBLOCK_NONE;
1346 struct vmem_altmap mhp_altmap = {};
1347 struct memory_group *group = NULL;
1349 bool new_node = false;
1353 size = resource_size(res);
1355 ret = check_hotplug_memory_range(start, size);
1359 if (mhp_flags & MHP_NID_IS_MGID) {
1360 group = memory_group_find_by_id(nid);
1366 if (!node_possible(nid)) {
1367 WARN(1, "node %d was absent from the node_possible_map\n", nid);
1371 mem_hotplug_begin();
1373 if (IS_ENABLED(CONFIG_ARCH_KEEP_MEMBLOCK)) {
1374 if (res->flags & IORESOURCE_SYSRAM_DRIVER_MANAGED)
1375 memblock_flags = MEMBLOCK_DRIVER_MANAGED;
1376 ret = memblock_add_node(start, size, nid, memblock_flags);
1378 goto error_mem_hotplug_end;
1381 ret = __try_online_node(nid, false);
1387 * Self hosted memmap array
1389 if (mhp_flags & MHP_MEMMAP_ON_MEMORY) {
1390 if (!mhp_supports_memmap_on_memory(size)) {
1394 mhp_altmap.free = PHYS_PFN(size);
1395 mhp_altmap.base_pfn = PHYS_PFN(start);
1396 params.altmap = &mhp_altmap;
1399 /* call arch's memory hotadd */
1400 ret = arch_add_memory(nid, start, size, ¶ms);
1404 /* create memory block devices after memory was added */
1405 ret = create_memory_block_devices(start, size, mhp_altmap.alloc,
1408 arch_remove_memory(start, size, NULL);
1413 /* If sysfs file of new node can't be created, cpu on the node
1414 * can't be hot-added. There is no rollback way now.
1415 * So, check by BUG_ON() to catch it reluctantly..
1416 * We online node here. We can't roll back from here.
1418 node_set_online(nid);
1419 ret = __register_one_node(nid);
1423 /* link memory sections under this node.*/
1424 link_mem_sections(nid, PFN_DOWN(start), PFN_UP(start + size - 1),
1427 /* create new memmap entry */
1428 if (!strcmp(res->name, "System RAM"))
1429 firmware_map_add_hotplug(start, start + size, "System RAM");
1431 /* device_online() will take the lock when calling online_pages() */
1435 * In case we're allowed to merge the resource, flag it and trigger
1436 * merging now that adding succeeded.
1438 if (mhp_flags & MHP_MERGE_RESOURCE)
1439 merge_system_ram_resource(res);
1441 /* online pages if requested */
1442 if (mhp_default_online_type != MMOP_OFFLINE)
1443 walk_memory_blocks(start, size, NULL, online_memory_block);
1447 /* rollback pgdat allocation and others */
1449 rollback_node_hotadd(nid);
1450 if (IS_ENABLED(CONFIG_ARCH_KEEP_MEMBLOCK))
1451 memblock_remove(start, size);
1452 error_mem_hotplug_end:
1457 /* requires device_hotplug_lock, see add_memory_resource() */
1458 int __ref __add_memory(int nid, u64 start, u64 size, mhp_t mhp_flags)
1460 struct resource *res;
1463 res = register_memory_resource(start, size, "System RAM");
1465 return PTR_ERR(res);
1467 ret = add_memory_resource(nid, res, mhp_flags);
1469 release_memory_resource(res);
1473 int add_memory(int nid, u64 start, u64 size, mhp_t mhp_flags)
1477 lock_device_hotplug();
1478 rc = __add_memory(nid, start, size, mhp_flags);
1479 unlock_device_hotplug();
1483 EXPORT_SYMBOL_GPL(add_memory);
1486 * Add special, driver-managed memory to the system as system RAM. Such
1487 * memory is not exposed via the raw firmware-provided memmap as system
1488 * RAM, instead, it is detected and added by a driver - during cold boot,
1489 * after a reboot, and after kexec.
1491 * Reasons why this memory should not be used for the initial memmap of a
1492 * kexec kernel or for placing kexec images:
1493 * - The booting kernel is in charge of determining how this memory will be
1494 * used (e.g., use persistent memory as system RAM)
1495 * - Coordination with a hypervisor is required before this memory
1496 * can be used (e.g., inaccessible parts).
1498 * For this memory, no entries in /sys/firmware/memmap ("raw firmware-provided
1499 * memory map") are created. Also, the created memory resource is flagged
1500 * with IORESOURCE_SYSRAM_DRIVER_MANAGED, so in-kernel users can special-case
1501 * this memory as well (esp., not place kexec images onto it).
1503 * The resource_name (visible via /proc/iomem) has to have the format
1504 * "System RAM ($DRIVER)".
1506 int add_memory_driver_managed(int nid, u64 start, u64 size,
1507 const char *resource_name, mhp_t mhp_flags)
1509 struct resource *res;
1512 if (!resource_name ||
1513 strstr(resource_name, "System RAM (") != resource_name ||
1514 resource_name[strlen(resource_name) - 1] != ')')
1517 lock_device_hotplug();
1519 res = register_memory_resource(start, size, resource_name);
1525 rc = add_memory_resource(nid, res, mhp_flags);
1527 release_memory_resource(res);
1530 unlock_device_hotplug();
1533 EXPORT_SYMBOL_GPL(add_memory_driver_managed);
1536 * Platforms should define arch_get_mappable_range() that provides
1537 * maximum possible addressable physical memory range for which the
1538 * linear mapping could be created. The platform returned address
1539 * range must adhere to these following semantics.
1541 * - range.start <= range.end
1542 * - Range includes both end points [range.start..range.end]
1544 * There is also a fallback definition provided here, allowing the
1545 * entire possible physical address range in case any platform does
1546 * not define arch_get_mappable_range().
1548 struct range __weak arch_get_mappable_range(void)
1550 struct range mhp_range = {
1557 struct range mhp_get_pluggable_range(bool need_mapping)
1559 const u64 max_phys = (1ULL << MAX_PHYSMEM_BITS) - 1;
1560 struct range mhp_range;
1563 mhp_range = arch_get_mappable_range();
1564 if (mhp_range.start > max_phys) {
1565 mhp_range.start = 0;
1568 mhp_range.end = min_t(u64, mhp_range.end, max_phys);
1570 mhp_range.start = 0;
1571 mhp_range.end = max_phys;
1575 EXPORT_SYMBOL_GPL(mhp_get_pluggable_range);
1577 bool mhp_range_allowed(u64 start, u64 size, bool need_mapping)
1579 struct range mhp_range = mhp_get_pluggable_range(need_mapping);
1580 u64 end = start + size;
1582 if (start < end && start >= mhp_range.start && (end - 1) <= mhp_range.end)
1585 pr_warn("Hotplug memory [%#llx-%#llx] exceeds maximum addressable range [%#llx-%#llx]\n",
1586 start, end, mhp_range.start, mhp_range.end);
1590 #ifdef CONFIG_MEMORY_HOTREMOVE
1592 * Confirm all pages in a range [start, end) belong to the same zone (skipping
1593 * memory holes). When true, return the zone.
1595 struct zone *test_pages_in_a_zone(unsigned long start_pfn,
1596 unsigned long end_pfn)
1598 unsigned long pfn, sec_end_pfn;
1599 struct zone *zone = NULL;
1602 for (pfn = start_pfn, sec_end_pfn = SECTION_ALIGN_UP(start_pfn + 1);
1604 pfn = sec_end_pfn, sec_end_pfn += PAGES_PER_SECTION) {
1605 /* Make sure the memory section is present first */
1606 if (!present_section_nr(pfn_to_section_nr(pfn)))
1608 for (; pfn < sec_end_pfn && pfn < end_pfn;
1609 pfn += MAX_ORDER_NR_PAGES) {
1610 /* Check if we got outside of the zone */
1611 if (zone && !zone_spans_pfn(zone, pfn))
1613 page = pfn_to_page(pfn);
1614 if (zone && page_zone(page) != zone)
1616 zone = page_zone(page);
1624 * Scan pfn range [start,end) to find movable/migratable pages (LRU pages,
1625 * non-lru movable pages and hugepages). Will skip over most unmovable
1626 * pages (esp., pages that can be skipped when offlining), but bail out on
1627 * definitely unmovable pages.
1630 * 0 in case a movable page is found and movable_pfn was updated.
1631 * -ENOENT in case no movable page was found.
1632 * -EBUSY in case a definitely unmovable page was found.
1634 static int scan_movable_pages(unsigned long start, unsigned long end,
1635 unsigned long *movable_pfn)
1639 for (pfn = start; pfn < end; pfn++) {
1640 struct page *page, *head;
1643 if (!pfn_valid(pfn))
1645 page = pfn_to_page(pfn);
1648 if (__PageMovable(page))
1652 * PageOffline() pages that are not marked __PageMovable() and
1653 * have a reference count > 0 (after MEM_GOING_OFFLINE) are
1654 * definitely unmovable. If their reference count would be 0,
1655 * they could at least be skipped when offlining memory.
1657 if (PageOffline(page) && page_count(page))
1660 if (!PageHuge(page))
1662 head = compound_head(page);
1664 * This test is racy as we hold no reference or lock. The
1665 * hugetlb page could have been free'ed and head is no longer
1666 * a hugetlb page before the following check. In such unlikely
1667 * cases false positives and negatives are possible. Calling
1668 * code must deal with these scenarios.
1670 if (HPageMigratable(head))
1672 skip = compound_nr(head) - (page - head);
1682 do_migrate_range(unsigned long start_pfn, unsigned long end_pfn)
1685 struct page *page, *head;
1688 static DEFINE_RATELIMIT_STATE(migrate_rs, DEFAULT_RATELIMIT_INTERVAL,
1689 DEFAULT_RATELIMIT_BURST);
1691 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
1692 if (!pfn_valid(pfn))
1694 page = pfn_to_page(pfn);
1695 head = compound_head(page);
1697 if (PageHuge(page)) {
1698 pfn = page_to_pfn(head) + compound_nr(head) - 1;
1699 isolate_huge_page(head, &source);
1701 } else if (PageTransHuge(page))
1702 pfn = page_to_pfn(head) + thp_nr_pages(page) - 1;
1705 * HWPoison pages have elevated reference counts so the migration would
1706 * fail on them. It also doesn't make any sense to migrate them in the
1707 * first place. Still try to unmap such a page in case it is still mapped
1708 * (e.g. current hwpoison implementation doesn't unmap KSM pages but keep
1709 * the unmap as the catch all safety net).
1711 if (PageHWPoison(page)) {
1712 if (WARN_ON(PageLRU(page)))
1713 isolate_lru_page(page);
1714 if (page_mapped(page))
1715 try_to_unmap(page, TTU_IGNORE_MLOCK);
1719 if (!get_page_unless_zero(page))
1722 * We can skip free pages. And we can deal with pages on
1723 * LRU and non-lru movable pages.
1726 ret = isolate_lru_page(page);
1728 ret = isolate_movable_page(page, ISOLATE_UNEVICTABLE);
1729 if (!ret) { /* Success */
1730 list_add_tail(&page->lru, &source);
1731 if (!__PageMovable(page))
1732 inc_node_page_state(page, NR_ISOLATED_ANON +
1733 page_is_file_lru(page));
1736 if (__ratelimit(&migrate_rs)) {
1737 pr_warn("failed to isolate pfn %lx\n", pfn);
1738 dump_page(page, "isolation failed");
1743 if (!list_empty(&source)) {
1744 nodemask_t nmask = node_states[N_MEMORY];
1745 struct migration_target_control mtc = {
1747 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
1751 * We have checked that migration range is on a single zone so
1752 * we can use the nid of the first page to all the others.
1754 mtc.nid = page_to_nid(list_first_entry(&source, struct page, lru));
1757 * try to allocate from a different node but reuse this node
1758 * if there are no other online nodes to be used (e.g. we are
1759 * offlining a part of the only existing node)
1761 node_clear(mtc.nid, nmask);
1762 if (nodes_empty(nmask))
1763 node_set(mtc.nid, nmask);
1764 ret = migrate_pages(&source, alloc_migration_target, NULL,
1765 (unsigned long)&mtc, MIGRATE_SYNC, MR_MEMORY_HOTPLUG, NULL);
1767 list_for_each_entry(page, &source, lru) {
1768 if (__ratelimit(&migrate_rs)) {
1769 pr_warn("migrating pfn %lx failed ret:%d\n",
1770 page_to_pfn(page), ret);
1771 dump_page(page, "migration failure");
1774 putback_movable_pages(&source);
1781 static int __init cmdline_parse_movable_node(char *p)
1783 movable_node_enabled = true;
1786 early_param("movable_node", cmdline_parse_movable_node);
1788 /* check which state of node_states will be changed when offline memory */
1789 static void node_states_check_changes_offline(unsigned long nr_pages,
1790 struct zone *zone, struct memory_notify *arg)
1792 struct pglist_data *pgdat = zone->zone_pgdat;
1793 unsigned long present_pages = 0;
1796 arg->status_change_nid = NUMA_NO_NODE;
1797 arg->status_change_nid_normal = NUMA_NO_NODE;
1800 * Check whether node_states[N_NORMAL_MEMORY] will be changed.
1801 * If the memory to be offline is within the range
1802 * [0..ZONE_NORMAL], and it is the last present memory there,
1803 * the zones in that range will become empty after the offlining,
1804 * thus we can determine that we need to clear the node from
1805 * node_states[N_NORMAL_MEMORY].
1807 for (zt = 0; zt <= ZONE_NORMAL; zt++)
1808 present_pages += pgdat->node_zones[zt].present_pages;
1809 if (zone_idx(zone) <= ZONE_NORMAL && nr_pages >= present_pages)
1810 arg->status_change_nid_normal = zone_to_nid(zone);
1813 * We have accounted the pages from [0..ZONE_NORMAL); ZONE_HIGHMEM
1814 * does not apply as we don't support 32bit.
1815 * Here we count the possible pages from ZONE_MOVABLE.
1816 * If after having accounted all the pages, we see that the nr_pages
1817 * to be offlined is over or equal to the accounted pages,
1818 * we know that the node will become empty, and so, we can clear
1819 * it for N_MEMORY as well.
1821 present_pages += pgdat->node_zones[ZONE_MOVABLE].present_pages;
1823 if (nr_pages >= present_pages)
1824 arg->status_change_nid = zone_to_nid(zone);
1827 static void node_states_clear_node(int node, struct memory_notify *arg)
1829 if (arg->status_change_nid_normal >= 0)
1830 node_clear_state(node, N_NORMAL_MEMORY);
1832 if (arg->status_change_nid >= 0)
1833 node_clear_state(node, N_MEMORY);
1836 static int count_system_ram_pages_cb(unsigned long start_pfn,
1837 unsigned long nr_pages, void *data)
1839 unsigned long *nr_system_ram_pages = data;
1841 *nr_system_ram_pages += nr_pages;
1845 int __ref offline_pages(unsigned long start_pfn, unsigned long nr_pages,
1846 struct memory_group *group)
1848 const unsigned long end_pfn = start_pfn + nr_pages;
1849 unsigned long pfn, system_ram_pages = 0;
1850 unsigned long flags;
1852 struct memory_notify arg;
1857 * {on,off}lining is constrained to full memory sections (or more
1858 * precisely to memory blocks from the user space POV).
1859 * memmap_on_memory is an exception because it reserves initial part
1860 * of the physical memory space for vmemmaps. That space is pageblock
1863 if (WARN_ON_ONCE(!nr_pages ||
1864 !IS_ALIGNED(start_pfn, pageblock_nr_pages) ||
1865 !IS_ALIGNED(start_pfn + nr_pages, PAGES_PER_SECTION)))
1868 mem_hotplug_begin();
1871 * Don't allow to offline memory blocks that contain holes.
1872 * Consequently, memory blocks with holes can never get onlined
1873 * via the hotplug path - online_pages() - as hotplugged memory has
1874 * no holes. This way, we e.g., don't have to worry about marking
1875 * memory holes PG_reserved, don't need pfn_valid() checks, and can
1876 * avoid using walk_system_ram_range() later.
1878 walk_system_ram_range(start_pfn, nr_pages, &system_ram_pages,
1879 count_system_ram_pages_cb);
1880 if (system_ram_pages != nr_pages) {
1882 reason = "memory holes";
1883 goto failed_removal;
1886 /* This makes hotplug much easier...and readable.
1887 we assume this for now. .*/
1888 zone = test_pages_in_a_zone(start_pfn, end_pfn);
1891 reason = "multizone range";
1892 goto failed_removal;
1894 node = zone_to_nid(zone);
1897 * Disable pcplists so that page isolation cannot race with freeing
1898 * in a way that pages from isolated pageblock are left on pcplists.
1900 zone_pcp_disable(zone);
1901 lru_cache_disable();
1903 /* set above range as isolated */
1904 ret = start_isolate_page_range(start_pfn, end_pfn,
1906 MEMORY_OFFLINE | REPORT_FAILURE);
1908 reason = "failure to isolate range";
1909 goto failed_removal_pcplists_disabled;
1912 arg.start_pfn = start_pfn;
1913 arg.nr_pages = nr_pages;
1914 node_states_check_changes_offline(nr_pages, zone, &arg);
1916 ret = memory_notify(MEM_GOING_OFFLINE, &arg);
1917 ret = notifier_to_errno(ret);
1919 reason = "notifier failure";
1920 goto failed_removal_isolated;
1926 if (signal_pending(current)) {
1928 reason = "signal backoff";
1929 goto failed_removal_isolated;
1934 ret = scan_movable_pages(pfn, end_pfn, &pfn);
1937 * TODO: fatal migration failures should bail
1940 do_migrate_range(pfn, end_pfn);
1944 if (ret != -ENOENT) {
1945 reason = "unmovable page";
1946 goto failed_removal_isolated;
1950 * Dissolve free hugepages in the memory block before doing
1951 * offlining actually in order to make hugetlbfs's object
1952 * counting consistent.
1954 ret = dissolve_free_huge_pages(start_pfn, end_pfn);
1956 reason = "failure to dissolve huge pages";
1957 goto failed_removal_isolated;
1960 ret = test_pages_isolated(start_pfn, end_pfn, MEMORY_OFFLINE);
1964 /* Mark all sections offline and remove free pages from the buddy. */
1965 __offline_isolated_pages(start_pfn, end_pfn);
1966 pr_debug("Offlined Pages %ld\n", nr_pages);
1969 * The memory sections are marked offline, and the pageblock flags
1970 * effectively stale; nobody should be touching them. Fixup the number
1971 * of isolated pageblocks, memory onlining will properly revert this.
1973 spin_lock_irqsave(&zone->lock, flags);
1974 zone->nr_isolate_pageblock -= nr_pages / pageblock_nr_pages;
1975 spin_unlock_irqrestore(&zone->lock, flags);
1978 zone_pcp_enable(zone);
1980 /* removal success */
1981 adjust_managed_page_count(pfn_to_page(start_pfn), -nr_pages);
1982 adjust_present_page_count(pfn_to_page(start_pfn), group, -nr_pages);
1984 /* reinitialise watermarks and update pcp limits */
1985 init_per_zone_wmark_min();
1987 if (!populated_zone(zone)) {
1988 zone_pcp_reset(zone);
1989 build_all_zonelists(NULL);
1992 node_states_clear_node(node, &arg);
1993 if (arg.status_change_nid >= 0) {
1995 kcompactd_stop(node);
1998 writeback_set_ratelimit();
2000 memory_notify(MEM_OFFLINE, &arg);
2001 remove_pfn_range_from_zone(zone, start_pfn, nr_pages);
2005 failed_removal_isolated:
2006 undo_isolate_page_range(start_pfn, end_pfn, MIGRATE_MOVABLE);
2007 memory_notify(MEM_CANCEL_OFFLINE, &arg);
2008 failed_removal_pcplists_disabled:
2010 zone_pcp_enable(zone);
2012 pr_debug("memory offlining [mem %#010llx-%#010llx] failed due to %s\n",
2013 (unsigned long long) start_pfn << PAGE_SHIFT,
2014 ((unsigned long long) end_pfn << PAGE_SHIFT) - 1,
2016 /* pushback to free area */
2021 static int check_memblock_offlined_cb(struct memory_block *mem, void *arg)
2023 int ret = !is_memblock_offlined(mem);
2027 if (unlikely(ret)) {
2028 phys_addr_t beginpa, endpa;
2030 beginpa = PFN_PHYS(section_nr_to_pfn(mem->start_section_nr));
2031 endpa = beginpa + memory_block_size_bytes() - 1;
2032 pr_warn("removing memory fails, because memory [%pa-%pa] is onlined\n",
2040 static int get_nr_vmemmap_pages_cb(struct memory_block *mem, void *arg)
2043 * If not set, continue with the next block.
2045 return mem->nr_vmemmap_pages;
2048 static int check_cpu_on_node(pg_data_t *pgdat)
2052 for_each_present_cpu(cpu) {
2053 if (cpu_to_node(cpu) == pgdat->node_id)
2055 * the cpu on this node isn't removed, and we can't
2056 * offline this node.
2064 static int check_no_memblock_for_node_cb(struct memory_block *mem, void *arg)
2066 int nid = *(int *)arg;
2069 * If a memory block belongs to multiple nodes, the stored nid is not
2070 * reliable. However, such blocks are always online (e.g., cannot get
2071 * offlined) and, therefore, are still spanned by the node.
2073 return mem->nid == nid ? -EEXIST : 0;
2080 * Offline a node if all memory sections and cpus of the node are removed.
2082 * NOTE: The caller must call lock_device_hotplug() to serialize hotplug
2083 * and online/offline operations before this call.
2085 void try_offline_node(int nid)
2087 pg_data_t *pgdat = NODE_DATA(nid);
2091 * If the node still spans pages (especially ZONE_DEVICE), don't
2092 * offline it. A node spans memory after move_pfn_range_to_zone(),
2093 * e.g., after the memory block was onlined.
2095 if (pgdat->node_spanned_pages)
2099 * Especially offline memory blocks might not be spanned by the
2100 * node. They will get spanned by the node once they get onlined.
2101 * However, they link to the node in sysfs and can get onlined later.
2103 rc = for_each_memory_block(&nid, check_no_memblock_for_node_cb);
2107 if (check_cpu_on_node(pgdat))
2111 * all memory/cpu of this node are removed, we can offline this
2114 node_set_offline(nid);
2115 unregister_one_node(nid);
2117 EXPORT_SYMBOL(try_offline_node);
2119 static int __ref try_remove_memory(u64 start, u64 size)
2121 struct vmem_altmap mhp_altmap = {};
2122 struct vmem_altmap *altmap = NULL;
2123 unsigned long nr_vmemmap_pages;
2124 int rc = 0, nid = NUMA_NO_NODE;
2126 BUG_ON(check_hotplug_memory_range(start, size));
2129 * All memory blocks must be offlined before removing memory. Check
2130 * whether all memory blocks in question are offline and return error
2131 * if this is not the case.
2133 * While at it, determine the nid. Note that if we'd have mixed nodes,
2134 * we'd only try to offline the last determined one -- which is good
2135 * enough for the cases we care about.
2137 rc = walk_memory_blocks(start, size, &nid, check_memblock_offlined_cb);
2142 * We only support removing memory added with MHP_MEMMAP_ON_MEMORY in
2143 * the same granularity it was added - a single memory block.
2145 if (memmap_on_memory) {
2146 nr_vmemmap_pages = walk_memory_blocks(start, size, NULL,
2147 get_nr_vmemmap_pages_cb);
2148 if (nr_vmemmap_pages) {
2149 if (size != memory_block_size_bytes()) {
2150 pr_warn("Refuse to remove %#llx - %#llx,"
2151 "wrong granularity\n",
2152 start, start + size);
2157 * Let remove_pmd_table->free_hugepage_table do the
2158 * right thing if we used vmem_altmap when hot-adding
2161 mhp_altmap.alloc = nr_vmemmap_pages;
2162 altmap = &mhp_altmap;
2166 /* remove memmap entry */
2167 firmware_map_remove(start, start + size, "System RAM");
2170 * Memory block device removal under the device_hotplug_lock is
2171 * a barrier against racing online attempts.
2173 remove_memory_block_devices(start, size);
2175 mem_hotplug_begin();
2177 arch_remove_memory(start, size, altmap);
2179 if (IS_ENABLED(CONFIG_ARCH_KEEP_MEMBLOCK)) {
2180 memblock_phys_free(start, size);
2181 memblock_remove(start, size);
2184 release_mem_region_adjustable(start, size);
2186 if (nid != NUMA_NO_NODE)
2187 try_offline_node(nid);
2194 * __remove_memory - Remove memory if every memory block is offline
2195 * @start: physical address of the region to remove
2196 * @size: size of the region to remove
2198 * NOTE: The caller must call lock_device_hotplug() to serialize hotplug
2199 * and online/offline operations before this call, as required by
2200 * try_offline_node().
2202 void __remove_memory(u64 start, u64 size)
2206 * trigger BUG() if some memory is not offlined prior to calling this
2209 if (try_remove_memory(start, size))
2214 * Remove memory if every memory block is offline, otherwise return -EBUSY is
2215 * some memory is not offline
2217 int remove_memory(u64 start, u64 size)
2221 lock_device_hotplug();
2222 rc = try_remove_memory(start, size);
2223 unlock_device_hotplug();
2227 EXPORT_SYMBOL_GPL(remove_memory);
2229 static int try_offline_memory_block(struct memory_block *mem, void *arg)
2231 uint8_t online_type = MMOP_ONLINE_KERNEL;
2232 uint8_t **online_types = arg;
2237 * Sense the online_type via the zone of the memory block. Offlining
2238 * with multiple zones within one memory block will be rejected
2239 * by offlining code ... so we don't care about that.
2241 page = pfn_to_online_page(section_nr_to_pfn(mem->start_section_nr));
2242 if (page && zone_idx(page_zone(page)) == ZONE_MOVABLE)
2243 online_type = MMOP_ONLINE_MOVABLE;
2245 rc = device_offline(&mem->dev);
2247 * Default is MMOP_OFFLINE - change it only if offlining succeeded,
2248 * so try_reonline_memory_block() can do the right thing.
2251 **online_types = online_type;
2254 /* Ignore if already offline. */
2255 return rc < 0 ? rc : 0;
2258 static int try_reonline_memory_block(struct memory_block *mem, void *arg)
2260 uint8_t **online_types = arg;
2263 if (**online_types != MMOP_OFFLINE) {
2264 mem->online_type = **online_types;
2265 rc = device_online(&mem->dev);
2267 pr_warn("%s: Failed to re-online memory: %d",
2271 /* Continue processing all remaining memory blocks. */
2277 * Try to offline and remove memory. Might take a long time to finish in case
2278 * memory is still in use. Primarily useful for memory devices that logically
2279 * unplugged all memory (so it's no longer in use) and want to offline + remove
2282 int offline_and_remove_memory(u64 start, u64 size)
2284 const unsigned long mb_count = size / memory_block_size_bytes();
2285 uint8_t *online_types, *tmp;
2288 if (!IS_ALIGNED(start, memory_block_size_bytes()) ||
2289 !IS_ALIGNED(size, memory_block_size_bytes()) || !size)
2293 * We'll remember the old online type of each memory block, so we can
2294 * try to revert whatever we did when offlining one memory block fails
2295 * after offlining some others succeeded.
2297 online_types = kmalloc_array(mb_count, sizeof(*online_types),
2302 * Initialize all states to MMOP_OFFLINE, so when we abort processing in
2303 * try_offline_memory_block(), we'll skip all unprocessed blocks in
2304 * try_reonline_memory_block().
2306 memset(online_types, MMOP_OFFLINE, mb_count);
2308 lock_device_hotplug();
2311 rc = walk_memory_blocks(start, size, &tmp, try_offline_memory_block);
2314 * In case we succeeded to offline all memory, remove it.
2315 * This cannot fail as it cannot get onlined in the meantime.
2318 rc = try_remove_memory(start, size);
2320 pr_err("%s: Failed to remove memory: %d", __func__, rc);
2324 * Rollback what we did. While memory onlining might theoretically fail
2325 * (nacked by a notifier), it barely ever happens.
2329 walk_memory_blocks(start, size, &tmp,
2330 try_reonline_memory_block);
2332 unlock_device_hotplug();
2334 kfree(online_types);
2337 EXPORT_SYMBOL_GPL(offline_and_remove_memory);
2338 #endif /* CONFIG_MEMORY_HOTREMOVE */