1 // SPDX-License-Identifier: GPL-2.0-only
3 * linux/mm/memory_hotplug.c
8 #include <linux/stddef.h>
10 #include <linux/sched/signal.h>
11 #include <linux/swap.h>
12 #include <linux/interrupt.h>
13 #include <linux/pagemap.h>
14 #include <linux/compiler.h>
15 #include <linux/export.h>
16 #include <linux/writeback.h>
17 #include <linux/slab.h>
18 #include <linux/sysctl.h>
19 #include <linux/cpu.h>
20 #include <linux/memory.h>
21 #include <linux/memremap.h>
22 #include <linux/memory_hotplug.h>
23 #include <linux/vmalloc.h>
24 #include <linux/ioport.h>
25 #include <linux/delay.h>
26 #include <linux/migrate.h>
27 #include <linux/page-isolation.h>
28 #include <linux/pfn.h>
29 #include <linux/suspend.h>
30 #include <linux/mm_inline.h>
31 #include <linux/firmware-map.h>
32 #include <linux/stop_machine.h>
33 #include <linux/hugetlb.h>
34 #include <linux/memblock.h>
35 #include <linux/compaction.h>
36 #include <linux/rmap.h>
37 #include <linux/module.h>
39 #include <asm/tlbflush.h>
45 MEMMAP_ON_MEMORY_DISABLE = 0,
46 MEMMAP_ON_MEMORY_ENABLE,
47 MEMMAP_ON_MEMORY_FORCE,
50 static int memmap_mode __read_mostly = MEMMAP_ON_MEMORY_DISABLE;
52 static inline unsigned long memory_block_memmap_size(void)
54 return PHYS_PFN(memory_block_size_bytes()) * sizeof(struct page);
57 static inline unsigned long memory_block_memmap_on_memory_pages(void)
59 unsigned long nr_pages = PFN_UP(memory_block_memmap_size());
62 * In "forced" memmap_on_memory mode, we add extra pages to align the
63 * vmemmap size to cover full pageblocks. That way, we can add memory
64 * even if the vmemmap size is not properly aligned, however, we might waste
67 if (memmap_mode == MEMMAP_ON_MEMORY_FORCE)
68 return pageblock_align(nr_pages);
72 #ifdef CONFIG_MHP_MEMMAP_ON_MEMORY
74 * memory_hotplug.memmap_on_memory parameter
76 static int set_memmap_mode(const char *val, const struct kernel_param *kp)
81 if (sysfs_streq(val, "force") || sysfs_streq(val, "FORCE")) {
82 mode = MEMMAP_ON_MEMORY_FORCE;
84 ret = kstrtobool(val, &enabled);
88 mode = MEMMAP_ON_MEMORY_ENABLE;
90 mode = MEMMAP_ON_MEMORY_DISABLE;
92 *((int *)kp->arg) = mode;
93 if (mode == MEMMAP_ON_MEMORY_FORCE) {
94 unsigned long memmap_pages = memory_block_memmap_on_memory_pages();
96 pr_info_once("Memory hotplug will waste %ld pages in each memory block\n",
97 memmap_pages - PFN_UP(memory_block_memmap_size()));
102 static int get_memmap_mode(char *buffer, const struct kernel_param *kp)
104 if (*((int *)kp->arg) == MEMMAP_ON_MEMORY_FORCE)
105 return sprintf(buffer, "force\n");
106 return param_get_bool(buffer, kp);
109 static const struct kernel_param_ops memmap_mode_ops = {
110 .set = set_memmap_mode,
111 .get = get_memmap_mode,
113 module_param_cb(memmap_on_memory, &memmap_mode_ops, &memmap_mode, 0444);
114 MODULE_PARM_DESC(memmap_on_memory, "Enable memmap on memory for memory hotplug\n"
115 "With value \"force\" it could result in memory wastage due "
116 "to memmap size limitations (Y/N/force)");
118 static inline bool mhp_memmap_on_memory(void)
120 return memmap_mode != MEMMAP_ON_MEMORY_DISABLE;
123 static inline bool mhp_memmap_on_memory(void)
130 ONLINE_POLICY_CONTIG_ZONES = 0,
131 ONLINE_POLICY_AUTO_MOVABLE,
134 static const char * const online_policy_to_str[] = {
135 [ONLINE_POLICY_CONTIG_ZONES] = "contig-zones",
136 [ONLINE_POLICY_AUTO_MOVABLE] = "auto-movable",
139 static int set_online_policy(const char *val, const struct kernel_param *kp)
141 int ret = sysfs_match_string(online_policy_to_str, val);
145 *((int *)kp->arg) = ret;
149 static int get_online_policy(char *buffer, const struct kernel_param *kp)
151 return sprintf(buffer, "%s\n", online_policy_to_str[*((int *)kp->arg)]);
155 * memory_hotplug.online_policy: configure online behavior when onlining without
156 * specifying a zone (MMOP_ONLINE)
158 * "contig-zones": keep zone contiguous
159 * "auto-movable": online memory to ZONE_MOVABLE if the configuration
160 * (auto_movable_ratio, auto_movable_numa_aware) allows for it
162 static int online_policy __read_mostly = ONLINE_POLICY_CONTIG_ZONES;
163 static const struct kernel_param_ops online_policy_ops = {
164 .set = set_online_policy,
165 .get = get_online_policy,
167 module_param_cb(online_policy, &online_policy_ops, &online_policy, 0644);
168 MODULE_PARM_DESC(online_policy,
169 "Set the online policy (\"contig-zones\", \"auto-movable\") "
170 "Default: \"contig-zones\"");
173 * memory_hotplug.auto_movable_ratio: specify maximum MOVABLE:KERNEL ratio
175 * The ratio represent an upper limit and the kernel might decide to not
176 * online some memory to ZONE_MOVABLE -- e.g., because hotplugged KERNEL memory
177 * doesn't allow for more MOVABLE memory.
179 static unsigned int auto_movable_ratio __read_mostly = 301;
180 module_param(auto_movable_ratio, uint, 0644);
181 MODULE_PARM_DESC(auto_movable_ratio,
182 "Set the maximum ratio of MOVABLE:KERNEL memory in the system "
183 "in percent for \"auto-movable\" online policy. Default: 301");
186 * memory_hotplug.auto_movable_numa_aware: consider numa node stats
189 static bool auto_movable_numa_aware __read_mostly = true;
190 module_param(auto_movable_numa_aware, bool, 0644);
191 MODULE_PARM_DESC(auto_movable_numa_aware,
192 "Consider numa node stats in addition to global stats in "
193 "\"auto-movable\" online policy. Default: true");
194 #endif /* CONFIG_NUMA */
197 * online_page_callback contains pointer to current page onlining function.
198 * Initially it is generic_online_page(). If it is required it could be
199 * changed by calling set_online_page_callback() for callback registration
200 * and restore_online_page_callback() for generic callback restore.
203 static online_page_callback_t online_page_callback = generic_online_page;
204 static DEFINE_MUTEX(online_page_callback_lock);
206 DEFINE_STATIC_PERCPU_RWSEM(mem_hotplug_lock);
208 void get_online_mems(void)
210 percpu_down_read(&mem_hotplug_lock);
213 void put_online_mems(void)
215 percpu_up_read(&mem_hotplug_lock);
218 bool movable_node_enabled = false;
220 #ifndef CONFIG_MEMORY_HOTPLUG_DEFAULT_ONLINE
221 int mhp_default_online_type = MMOP_OFFLINE;
223 int mhp_default_online_type = MMOP_ONLINE;
226 static int __init setup_memhp_default_state(char *str)
228 const int online_type = mhp_online_type_from_str(str);
230 if (online_type >= 0)
231 mhp_default_online_type = online_type;
235 __setup("memhp_default_state=", setup_memhp_default_state);
237 void mem_hotplug_begin(void)
240 percpu_down_write(&mem_hotplug_lock);
243 void mem_hotplug_done(void)
245 percpu_up_write(&mem_hotplug_lock);
249 u64 max_mem_size = U64_MAX;
251 /* add this memory to iomem resource */
252 static struct resource *register_memory_resource(u64 start, u64 size,
253 const char *resource_name)
255 struct resource *res;
256 unsigned long flags = IORESOURCE_SYSTEM_RAM | IORESOURCE_BUSY;
258 if (strcmp(resource_name, "System RAM"))
259 flags |= IORESOURCE_SYSRAM_DRIVER_MANAGED;
261 if (!mhp_range_allowed(start, size, true))
262 return ERR_PTR(-E2BIG);
265 * Make sure value parsed from 'mem=' only restricts memory adding
266 * while booting, so that memory hotplug won't be impacted. Please
267 * refer to document of 'mem=' in kernel-parameters.txt for more
270 if (start + size > max_mem_size && system_state < SYSTEM_RUNNING)
271 return ERR_PTR(-E2BIG);
274 * Request ownership of the new memory range. This might be
275 * a child of an existing resource that was present but
276 * not marked as busy.
278 res = __request_region(&iomem_resource, start, size,
279 resource_name, flags);
282 pr_debug("Unable to reserve System RAM region: %016llx->%016llx\n",
283 start, start + size);
284 return ERR_PTR(-EEXIST);
289 static void release_memory_resource(struct resource *res)
293 release_resource(res);
297 static int check_pfn_span(unsigned long pfn, unsigned long nr_pages)
300 * Disallow all operations smaller than a sub-section and only
301 * allow operations smaller than a section for
302 * SPARSEMEM_VMEMMAP. Note that check_hotplug_memory_range()
303 * enforces a larger memory_block_size_bytes() granularity for
304 * memory that will be marked online, so this check should only
305 * fire for direct arch_{add,remove}_memory() users outside of
306 * add_memory_resource().
308 unsigned long min_align;
310 if (IS_ENABLED(CONFIG_SPARSEMEM_VMEMMAP))
311 min_align = PAGES_PER_SUBSECTION;
313 min_align = PAGES_PER_SECTION;
314 if (!IS_ALIGNED(pfn | nr_pages, min_align))
320 * Return page for the valid pfn only if the page is online. All pfn
321 * walkers which rely on the fully initialized page->flags and others
322 * should use this rather than pfn_valid && pfn_to_page
324 struct page *pfn_to_online_page(unsigned long pfn)
326 unsigned long nr = pfn_to_section_nr(pfn);
327 struct dev_pagemap *pgmap;
328 struct mem_section *ms;
330 if (nr >= NR_MEM_SECTIONS)
333 ms = __nr_to_section(nr);
334 if (!online_section(ms))
338 * Save some code text when online_section() +
339 * pfn_section_valid() are sufficient.
341 if (IS_ENABLED(CONFIG_HAVE_ARCH_PFN_VALID) && !pfn_valid(pfn))
344 if (!pfn_section_valid(ms, pfn))
347 if (!online_device_section(ms))
348 return pfn_to_page(pfn);
351 * Slowpath: when ZONE_DEVICE collides with
352 * ZONE_{NORMAL,MOVABLE} within the same section some pfns in
353 * the section may be 'offline' but 'valid'. Only
354 * get_dev_pagemap() can determine sub-section online status.
356 pgmap = get_dev_pagemap(pfn, NULL);
357 put_dev_pagemap(pgmap);
359 /* The presence of a pgmap indicates ZONE_DEVICE offline pfn */
363 return pfn_to_page(pfn);
365 EXPORT_SYMBOL_GPL(pfn_to_online_page);
367 int __ref __add_pages(int nid, unsigned long pfn, unsigned long nr_pages,
368 struct mhp_params *params)
370 const unsigned long end_pfn = pfn + nr_pages;
371 unsigned long cur_nr_pages;
373 struct vmem_altmap *altmap = params->altmap;
375 if (WARN_ON_ONCE(!pgprot_val(params->pgprot)))
378 VM_BUG_ON(!mhp_range_allowed(PFN_PHYS(pfn), nr_pages * PAGE_SIZE, false));
382 * Validate altmap is within bounds of the total request
384 if (altmap->base_pfn != pfn
385 || vmem_altmap_offset(altmap) > nr_pages) {
386 pr_warn_once("memory add fail, invalid altmap\n");
392 if (check_pfn_span(pfn, nr_pages)) {
393 WARN(1, "Misaligned %s start: %#lx end: %#lx\n", __func__, pfn, pfn + nr_pages - 1);
397 for (; pfn < end_pfn; pfn += cur_nr_pages) {
398 /* Select all remaining pages up to the next section boundary */
399 cur_nr_pages = min(end_pfn - pfn,
400 SECTION_ALIGN_UP(pfn + 1) - pfn);
401 err = sparse_add_section(nid, pfn, cur_nr_pages, altmap,
407 vmemmap_populate_print_last();
411 /* find the smallest valid pfn in the range [start_pfn, end_pfn) */
412 static unsigned long find_smallest_section_pfn(int nid, struct zone *zone,
413 unsigned long start_pfn,
414 unsigned long end_pfn)
416 for (; start_pfn < end_pfn; start_pfn += PAGES_PER_SUBSECTION) {
417 if (unlikely(!pfn_to_online_page(start_pfn)))
420 if (unlikely(pfn_to_nid(start_pfn) != nid))
423 if (zone != page_zone(pfn_to_page(start_pfn)))
432 /* find the biggest valid pfn in the range [start_pfn, end_pfn). */
433 static unsigned long find_biggest_section_pfn(int nid, struct zone *zone,
434 unsigned long start_pfn,
435 unsigned long end_pfn)
439 /* pfn is the end pfn of a memory section. */
441 for (; pfn >= start_pfn; pfn -= PAGES_PER_SUBSECTION) {
442 if (unlikely(!pfn_to_online_page(pfn)))
445 if (unlikely(pfn_to_nid(pfn) != nid))
448 if (zone != page_zone(pfn_to_page(pfn)))
457 static void shrink_zone_span(struct zone *zone, unsigned long start_pfn,
458 unsigned long end_pfn)
461 int nid = zone_to_nid(zone);
463 if (zone->zone_start_pfn == start_pfn) {
465 * If the section is smallest section in the zone, it need
466 * shrink zone->zone_start_pfn and zone->zone_spanned_pages.
467 * In this case, we find second smallest valid mem_section
468 * for shrinking zone.
470 pfn = find_smallest_section_pfn(nid, zone, end_pfn,
473 zone->spanned_pages = zone_end_pfn(zone) - pfn;
474 zone->zone_start_pfn = pfn;
476 zone->zone_start_pfn = 0;
477 zone->spanned_pages = 0;
479 } else if (zone_end_pfn(zone) == end_pfn) {
481 * If the section is biggest section in the zone, it need
482 * shrink zone->spanned_pages.
483 * In this case, we find second biggest valid mem_section for
486 pfn = find_biggest_section_pfn(nid, zone, zone->zone_start_pfn,
489 zone->spanned_pages = pfn - zone->zone_start_pfn + 1;
491 zone->zone_start_pfn = 0;
492 zone->spanned_pages = 0;
497 static void update_pgdat_span(struct pglist_data *pgdat)
499 unsigned long node_start_pfn = 0, node_end_pfn = 0;
502 for (zone = pgdat->node_zones;
503 zone < pgdat->node_zones + MAX_NR_ZONES; zone++) {
504 unsigned long end_pfn = zone_end_pfn(zone);
506 /* No need to lock the zones, they can't change. */
507 if (!zone->spanned_pages)
510 node_start_pfn = zone->zone_start_pfn;
511 node_end_pfn = end_pfn;
515 if (end_pfn > node_end_pfn)
516 node_end_pfn = end_pfn;
517 if (zone->zone_start_pfn < node_start_pfn)
518 node_start_pfn = zone->zone_start_pfn;
521 pgdat->node_start_pfn = node_start_pfn;
522 pgdat->node_spanned_pages = node_end_pfn - node_start_pfn;
525 void __ref remove_pfn_range_from_zone(struct zone *zone,
526 unsigned long start_pfn,
527 unsigned long nr_pages)
529 const unsigned long end_pfn = start_pfn + nr_pages;
530 struct pglist_data *pgdat = zone->zone_pgdat;
531 unsigned long pfn, cur_nr_pages;
533 /* Poison struct pages because they are now uninitialized again. */
534 for (pfn = start_pfn; pfn < end_pfn; pfn += cur_nr_pages) {
537 /* Select all remaining pages up to the next section boundary */
539 min(end_pfn - pfn, SECTION_ALIGN_UP(pfn + 1) - pfn);
540 page_init_poison(pfn_to_page(pfn),
541 sizeof(struct page) * cur_nr_pages);
545 * Zone shrinking code cannot properly deal with ZONE_DEVICE. So
546 * we will not try to shrink the zones - which is okay as
547 * set_zone_contiguous() cannot deal with ZONE_DEVICE either way.
549 if (zone_is_zone_device(zone))
552 clear_zone_contiguous(zone);
554 shrink_zone_span(zone, start_pfn, start_pfn + nr_pages);
555 update_pgdat_span(pgdat);
557 set_zone_contiguous(zone);
561 * __remove_pages() - remove sections of pages
562 * @pfn: starting pageframe (must be aligned to start of a section)
563 * @nr_pages: number of pages to remove (must be multiple of section size)
564 * @altmap: alternative device page map or %NULL if default memmap is used
566 * Generic helper function to remove section mappings and sysfs entries
567 * for the section of the memory we are removing. Caller needs to make
568 * sure that pages are marked reserved and zones are adjust properly by
569 * calling offline_pages().
571 void __remove_pages(unsigned long pfn, unsigned long nr_pages,
572 struct vmem_altmap *altmap)
574 const unsigned long end_pfn = pfn + nr_pages;
575 unsigned long cur_nr_pages;
577 if (check_pfn_span(pfn, nr_pages)) {
578 WARN(1, "Misaligned %s start: %#lx end: %#lx\n", __func__, pfn, pfn + nr_pages - 1);
582 for (; pfn < end_pfn; pfn += cur_nr_pages) {
584 /* Select all remaining pages up to the next section boundary */
585 cur_nr_pages = min(end_pfn - pfn,
586 SECTION_ALIGN_UP(pfn + 1) - pfn);
587 sparse_remove_section(pfn, cur_nr_pages, altmap);
591 int set_online_page_callback(online_page_callback_t callback)
596 mutex_lock(&online_page_callback_lock);
598 if (online_page_callback == generic_online_page) {
599 online_page_callback = callback;
603 mutex_unlock(&online_page_callback_lock);
608 EXPORT_SYMBOL_GPL(set_online_page_callback);
610 int restore_online_page_callback(online_page_callback_t callback)
615 mutex_lock(&online_page_callback_lock);
617 if (online_page_callback == callback) {
618 online_page_callback = generic_online_page;
622 mutex_unlock(&online_page_callback_lock);
627 EXPORT_SYMBOL_GPL(restore_online_page_callback);
629 void generic_online_page(struct page *page, unsigned int order)
632 * Freeing the page with debug_pagealloc enabled will try to unmap it,
633 * so we should map it first. This is better than introducing a special
634 * case in page freeing fast path.
636 debug_pagealloc_map_pages(page, 1 << order);
637 __free_pages_core(page, order);
638 totalram_pages_add(1UL << order);
640 EXPORT_SYMBOL_GPL(generic_online_page);
642 static void online_pages_range(unsigned long start_pfn, unsigned long nr_pages)
644 const unsigned long end_pfn = start_pfn + nr_pages;
648 * Online the pages in MAX_ORDER aligned chunks. The callback might
649 * decide to not expose all pages to the buddy (e.g., expose them
650 * later). We account all pages as being online and belonging to this
652 * When using memmap_on_memory, the range might not be aligned to
653 * MAX_ORDER_NR_PAGES - 1, but pageblock aligned. __ffs() will detect
654 * this and the first chunk to online will be pageblock_nr_pages.
656 for (pfn = start_pfn; pfn < end_pfn;) {
660 * Free to online pages in the largest chunks alignment allows.
662 * __ffs() behaviour is undefined for 0. start == 0 is
663 * MAX_ORDER-aligned, Set order to MAX_ORDER for the case.
666 order = min_t(int, MAX_ORDER, __ffs(pfn));
670 (*online_page_callback)(pfn_to_page(pfn), order);
671 pfn += (1UL << order);
674 /* mark all involved sections as online */
675 online_mem_sections(start_pfn, end_pfn);
678 /* check which state of node_states will be changed when online memory */
679 static void node_states_check_changes_online(unsigned long nr_pages,
680 struct zone *zone, struct memory_notify *arg)
682 int nid = zone_to_nid(zone);
684 arg->status_change_nid = NUMA_NO_NODE;
685 arg->status_change_nid_normal = NUMA_NO_NODE;
687 if (!node_state(nid, N_MEMORY))
688 arg->status_change_nid = nid;
689 if (zone_idx(zone) <= ZONE_NORMAL && !node_state(nid, N_NORMAL_MEMORY))
690 arg->status_change_nid_normal = nid;
693 static void node_states_set_node(int node, struct memory_notify *arg)
695 if (arg->status_change_nid_normal >= 0)
696 node_set_state(node, N_NORMAL_MEMORY);
698 if (arg->status_change_nid >= 0)
699 node_set_state(node, N_MEMORY);
702 static void __meminit resize_zone_range(struct zone *zone, unsigned long start_pfn,
703 unsigned long nr_pages)
705 unsigned long old_end_pfn = zone_end_pfn(zone);
707 if (zone_is_empty(zone) || start_pfn < zone->zone_start_pfn)
708 zone->zone_start_pfn = start_pfn;
710 zone->spanned_pages = max(start_pfn + nr_pages, old_end_pfn) - zone->zone_start_pfn;
713 static void __meminit resize_pgdat_range(struct pglist_data *pgdat, unsigned long start_pfn,
714 unsigned long nr_pages)
716 unsigned long old_end_pfn = pgdat_end_pfn(pgdat);
718 if (!pgdat->node_spanned_pages || start_pfn < pgdat->node_start_pfn)
719 pgdat->node_start_pfn = start_pfn;
721 pgdat->node_spanned_pages = max(start_pfn + nr_pages, old_end_pfn) - pgdat->node_start_pfn;
725 #ifdef CONFIG_ZONE_DEVICE
726 static void section_taint_zone_device(unsigned long pfn)
728 struct mem_section *ms = __pfn_to_section(pfn);
730 ms->section_mem_map |= SECTION_TAINT_ZONE_DEVICE;
733 static inline void section_taint_zone_device(unsigned long pfn)
739 * Associate the pfn range with the given zone, initializing the memmaps
740 * and resizing the pgdat/zone data to span the added pages. After this
741 * call, all affected pages are PG_reserved.
743 * All aligned pageblocks are initialized to the specified migratetype
744 * (usually MIGRATE_MOVABLE). Besides setting the migratetype, no related
745 * zone stats (e.g., nr_isolate_pageblock) are touched.
747 void __ref move_pfn_range_to_zone(struct zone *zone, unsigned long start_pfn,
748 unsigned long nr_pages,
749 struct vmem_altmap *altmap, int migratetype)
751 struct pglist_data *pgdat = zone->zone_pgdat;
752 int nid = pgdat->node_id;
754 clear_zone_contiguous(zone);
756 if (zone_is_empty(zone))
757 init_currently_empty_zone(zone, start_pfn, nr_pages);
758 resize_zone_range(zone, start_pfn, nr_pages);
759 resize_pgdat_range(pgdat, start_pfn, nr_pages);
762 * Subsection population requires care in pfn_to_online_page().
763 * Set the taint to enable the slow path detection of
764 * ZONE_DEVICE pages in an otherwise ZONE_{NORMAL,MOVABLE}
767 if (zone_is_zone_device(zone)) {
768 if (!IS_ALIGNED(start_pfn, PAGES_PER_SECTION))
769 section_taint_zone_device(start_pfn);
770 if (!IS_ALIGNED(start_pfn + nr_pages, PAGES_PER_SECTION))
771 section_taint_zone_device(start_pfn + nr_pages);
775 * TODO now we have a visible range of pages which are not associated
776 * with their zone properly. Not nice but set_pfnblock_flags_mask
777 * expects the zone spans the pfn range. All the pages in the range
778 * are reserved so nobody should be touching them so we should be safe
780 memmap_init_range(nr_pages, nid, zone_idx(zone), start_pfn, 0,
781 MEMINIT_HOTPLUG, altmap, migratetype);
783 set_zone_contiguous(zone);
786 struct auto_movable_stats {
787 unsigned long kernel_early_pages;
788 unsigned long movable_pages;
791 static void auto_movable_stats_account_zone(struct auto_movable_stats *stats,
794 if (zone_idx(zone) == ZONE_MOVABLE) {
795 stats->movable_pages += zone->present_pages;
797 stats->kernel_early_pages += zone->present_early_pages;
800 * CMA pages (never on hotplugged memory) behave like
803 stats->movable_pages += zone->cma_pages;
804 stats->kernel_early_pages -= zone->cma_pages;
805 #endif /* CONFIG_CMA */
808 struct auto_movable_group_stats {
809 unsigned long movable_pages;
810 unsigned long req_kernel_early_pages;
813 static int auto_movable_stats_account_group(struct memory_group *group,
816 const int ratio = READ_ONCE(auto_movable_ratio);
817 struct auto_movable_group_stats *stats = arg;
821 * We don't support modifying the config while the auto-movable online
822 * policy is already enabled. Just avoid the division by zero below.
828 * Calculate how many early kernel pages this group requires to
829 * satisfy the configured zone ratio.
831 pages = group->present_movable_pages * 100 / ratio;
832 pages -= group->present_kernel_pages;
835 stats->req_kernel_early_pages += pages;
836 stats->movable_pages += group->present_movable_pages;
840 static bool auto_movable_can_online_movable(int nid, struct memory_group *group,
841 unsigned long nr_pages)
843 unsigned long kernel_early_pages, movable_pages;
844 struct auto_movable_group_stats group_stats = {};
845 struct auto_movable_stats stats = {};
846 pg_data_t *pgdat = NODE_DATA(nid);
850 /* Walk all relevant zones and collect MOVABLE vs. KERNEL stats. */
851 if (nid == NUMA_NO_NODE) {
852 /* TODO: cache values */
853 for_each_populated_zone(zone)
854 auto_movable_stats_account_zone(&stats, zone);
856 for (i = 0; i < MAX_NR_ZONES; i++) {
857 zone = pgdat->node_zones + i;
858 if (populated_zone(zone))
859 auto_movable_stats_account_zone(&stats, zone);
863 kernel_early_pages = stats.kernel_early_pages;
864 movable_pages = stats.movable_pages;
867 * Kernel memory inside dynamic memory group allows for more MOVABLE
868 * memory within the same group. Remove the effect of all but the
869 * current group from the stats.
871 walk_dynamic_memory_groups(nid, auto_movable_stats_account_group,
872 group, &group_stats);
873 if (kernel_early_pages <= group_stats.req_kernel_early_pages)
875 kernel_early_pages -= group_stats.req_kernel_early_pages;
876 movable_pages -= group_stats.movable_pages;
878 if (group && group->is_dynamic)
879 kernel_early_pages += group->present_kernel_pages;
882 * Test if we could online the given number of pages to ZONE_MOVABLE
883 * and still stay in the configured ratio.
885 movable_pages += nr_pages;
886 return movable_pages <= (auto_movable_ratio * kernel_early_pages) / 100;
890 * Returns a default kernel memory zone for the given pfn range.
891 * If no kernel zone covers this pfn range it will automatically go
892 * to the ZONE_NORMAL.
894 static struct zone *default_kernel_zone_for_pfn(int nid, unsigned long start_pfn,
895 unsigned long nr_pages)
897 struct pglist_data *pgdat = NODE_DATA(nid);
900 for (zid = 0; zid < ZONE_NORMAL; zid++) {
901 struct zone *zone = &pgdat->node_zones[zid];
903 if (zone_intersects(zone, start_pfn, nr_pages))
907 return &pgdat->node_zones[ZONE_NORMAL];
911 * Determine to which zone to online memory dynamically based on user
912 * configuration and system stats. We care about the following ratio:
916 * Whereby MOVABLE is memory in ZONE_MOVABLE and KERNEL is memory in
917 * one of the kernel zones. CMA pages inside one of the kernel zones really
918 * behaves like ZONE_MOVABLE, so we treat them accordingly.
920 * We don't allow for hotplugged memory in a KERNEL zone to increase the
921 * amount of MOVABLE memory we can have, so we end up with:
923 * MOVABLE : KERNEL_EARLY
925 * Whereby KERNEL_EARLY is memory in one of the kernel zones, available sinze
926 * boot. We base our calculation on KERNEL_EARLY internally, because:
928 * a) Hotplugged memory in one of the kernel zones can sometimes still get
929 * hotunplugged, especially when hot(un)plugging individual memory blocks.
930 * There is no coordination across memory devices, therefore "automatic"
931 * hotunplugging, as implemented in hypervisors, could result in zone
933 * b) Early/boot memory in one of the kernel zones can usually not get
934 * hotunplugged again (e.g., no firmware interface to unplug, fragmented
935 * with unmovable allocations). While there are corner cases where it might
936 * still work, it is barely relevant in practice.
938 * Exceptions are dynamic memory groups, which allow for more MOVABLE
939 * memory within the same memory group -- because in that case, there is
940 * coordination within the single memory device managed by a single driver.
942 * We rely on "present pages" instead of "managed pages", as the latter is
943 * highly unreliable and dynamic in virtualized environments, and does not
944 * consider boot time allocations. For example, memory ballooning adjusts the
945 * managed pages when inflating/deflating the balloon, and balloon compaction
946 * can even migrate inflated pages between zones.
948 * Using "present pages" is better but some things to keep in mind are:
950 * a) Some memblock allocations, such as for the crashkernel area, are
951 * effectively unused by the kernel, yet they account to "present pages".
952 * Fortunately, these allocations are comparatively small in relevant setups
953 * (e.g., fraction of system memory).
954 * b) Some hotplugged memory blocks in virtualized environments, esecially
955 * hotplugged by virtio-mem, look like they are completely present, however,
956 * only parts of the memory block are actually currently usable.
957 * "present pages" is an upper limit that can get reached at runtime. As
958 * we base our calculations on KERNEL_EARLY, this is not an issue.
960 static struct zone *auto_movable_zone_for_pfn(int nid,
961 struct memory_group *group,
963 unsigned long nr_pages)
965 unsigned long online_pages = 0, max_pages, end_pfn;
968 if (!auto_movable_ratio)
971 if (group && !group->is_dynamic) {
972 max_pages = group->s.max_pages;
973 online_pages = group->present_movable_pages;
975 /* If anything is !MOVABLE online the rest !MOVABLE. */
976 if (group->present_kernel_pages)
978 } else if (!group || group->d.unit_pages == nr_pages) {
979 max_pages = nr_pages;
981 max_pages = group->d.unit_pages;
983 * Take a look at all online sections in the current unit.
984 * We can safely assume that all pages within a section belong
985 * to the same zone, because dynamic memory groups only deal
986 * with hotplugged memory.
988 pfn = ALIGN_DOWN(pfn, group->d.unit_pages);
989 end_pfn = pfn + group->d.unit_pages;
990 for (; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
991 page = pfn_to_online_page(pfn);
994 /* If anything is !MOVABLE online the rest !MOVABLE. */
995 if (!is_zone_movable_page(page))
997 online_pages += PAGES_PER_SECTION;
1002 * Online MOVABLE if we could *currently* online all remaining parts
1003 * MOVABLE. We expect to (add+) online them immediately next, so if
1004 * nobody interferes, all will be MOVABLE if possible.
1006 nr_pages = max_pages - online_pages;
1007 if (!auto_movable_can_online_movable(NUMA_NO_NODE, group, nr_pages))
1011 if (auto_movable_numa_aware &&
1012 !auto_movable_can_online_movable(nid, group, nr_pages))
1014 #endif /* CONFIG_NUMA */
1016 return &NODE_DATA(nid)->node_zones[ZONE_MOVABLE];
1018 return default_kernel_zone_for_pfn(nid, pfn, nr_pages);
1021 static inline struct zone *default_zone_for_pfn(int nid, unsigned long start_pfn,
1022 unsigned long nr_pages)
1024 struct zone *kernel_zone = default_kernel_zone_for_pfn(nid, start_pfn,
1026 struct zone *movable_zone = &NODE_DATA(nid)->node_zones[ZONE_MOVABLE];
1027 bool in_kernel = zone_intersects(kernel_zone, start_pfn, nr_pages);
1028 bool in_movable = zone_intersects(movable_zone, start_pfn, nr_pages);
1031 * We inherit the existing zone in a simple case where zones do not
1032 * overlap in the given range
1034 if (in_kernel ^ in_movable)
1035 return (in_kernel) ? kernel_zone : movable_zone;
1038 * If the range doesn't belong to any zone or two zones overlap in the
1039 * given range then we use movable zone only if movable_node is
1040 * enabled because we always online to a kernel zone by default.
1042 return movable_node_enabled ? movable_zone : kernel_zone;
1045 struct zone *zone_for_pfn_range(int online_type, int nid,
1046 struct memory_group *group, unsigned long start_pfn,
1047 unsigned long nr_pages)
1049 if (online_type == MMOP_ONLINE_KERNEL)
1050 return default_kernel_zone_for_pfn(nid, start_pfn, nr_pages);
1052 if (online_type == MMOP_ONLINE_MOVABLE)
1053 return &NODE_DATA(nid)->node_zones[ZONE_MOVABLE];
1055 if (online_policy == ONLINE_POLICY_AUTO_MOVABLE)
1056 return auto_movable_zone_for_pfn(nid, group, start_pfn, nr_pages);
1058 return default_zone_for_pfn(nid, start_pfn, nr_pages);
1062 * This function should only be called by memory_block_{online,offline},
1063 * and {online,offline}_pages.
1065 void adjust_present_page_count(struct page *page, struct memory_group *group,
1068 struct zone *zone = page_zone(page);
1069 const bool movable = zone_idx(zone) == ZONE_MOVABLE;
1072 * We only support onlining/offlining/adding/removing of complete
1073 * memory blocks; therefore, either all is either early or hotplugged.
1075 if (early_section(__pfn_to_section(page_to_pfn(page))))
1076 zone->present_early_pages += nr_pages;
1077 zone->present_pages += nr_pages;
1078 zone->zone_pgdat->node_present_pages += nr_pages;
1080 if (group && movable)
1081 group->present_movable_pages += nr_pages;
1082 else if (group && !movable)
1083 group->present_kernel_pages += nr_pages;
1086 int mhp_init_memmap_on_memory(unsigned long pfn, unsigned long nr_pages,
1089 unsigned long end_pfn = pfn + nr_pages;
1092 ret = kasan_add_zero_shadow(__va(PFN_PHYS(pfn)), PFN_PHYS(nr_pages));
1096 move_pfn_range_to_zone(zone, pfn, nr_pages, NULL, MIGRATE_UNMOVABLE);
1098 for (i = 0; i < nr_pages; i++)
1099 SetPageVmemmapSelfHosted(pfn_to_page(pfn + i));
1102 * It might be that the vmemmap_pages fully span sections. If that is
1103 * the case, mark those sections online here as otherwise they will be
1106 if (nr_pages >= PAGES_PER_SECTION)
1107 online_mem_sections(pfn, ALIGN_DOWN(end_pfn, PAGES_PER_SECTION));
1112 void mhp_deinit_memmap_on_memory(unsigned long pfn, unsigned long nr_pages)
1114 unsigned long end_pfn = pfn + nr_pages;
1117 * It might be that the vmemmap_pages fully span sections. If that is
1118 * the case, mark those sections offline here as otherwise they will be
1121 if (nr_pages >= PAGES_PER_SECTION)
1122 offline_mem_sections(pfn, ALIGN_DOWN(end_pfn, PAGES_PER_SECTION));
1125 * The pages associated with this vmemmap have been offlined, so
1126 * we can reset its state here.
1128 remove_pfn_range_from_zone(page_zone(pfn_to_page(pfn)), pfn, nr_pages);
1129 kasan_remove_zero_shadow(__va(PFN_PHYS(pfn)), PFN_PHYS(nr_pages));
1132 int __ref online_pages(unsigned long pfn, unsigned long nr_pages,
1133 struct zone *zone, struct memory_group *group)
1135 unsigned long flags;
1136 int need_zonelists_rebuild = 0;
1137 const int nid = zone_to_nid(zone);
1139 struct memory_notify arg;
1142 * {on,off}lining is constrained to full memory sections (or more
1143 * precisely to memory blocks from the user space POV).
1144 * memmap_on_memory is an exception because it reserves initial part
1145 * of the physical memory space for vmemmaps. That space is pageblock
1148 if (WARN_ON_ONCE(!nr_pages || !pageblock_aligned(pfn) ||
1149 !IS_ALIGNED(pfn + nr_pages, PAGES_PER_SECTION)))
1152 mem_hotplug_begin();
1154 /* associate pfn range with the zone */
1155 move_pfn_range_to_zone(zone, pfn, nr_pages, NULL, MIGRATE_ISOLATE);
1157 arg.start_pfn = pfn;
1158 arg.nr_pages = nr_pages;
1159 node_states_check_changes_online(nr_pages, zone, &arg);
1161 ret = memory_notify(MEM_GOING_ONLINE, &arg);
1162 ret = notifier_to_errno(ret);
1164 goto failed_addition;
1167 * Fixup the number of isolated pageblocks before marking the sections
1168 * onlining, such that undo_isolate_page_range() works correctly.
1170 spin_lock_irqsave(&zone->lock, flags);
1171 zone->nr_isolate_pageblock += nr_pages / pageblock_nr_pages;
1172 spin_unlock_irqrestore(&zone->lock, flags);
1175 * If this zone is not populated, then it is not in zonelist.
1176 * This means the page allocator ignores this zone.
1177 * So, zonelist must be updated after online.
1179 if (!populated_zone(zone)) {
1180 need_zonelists_rebuild = 1;
1181 setup_zone_pageset(zone);
1184 online_pages_range(pfn, nr_pages);
1185 adjust_present_page_count(pfn_to_page(pfn), group, nr_pages);
1187 node_states_set_node(nid, &arg);
1188 if (need_zonelists_rebuild)
1189 build_all_zonelists(NULL);
1191 /* Basic onlining is complete, allow allocation of onlined pages. */
1192 undo_isolate_page_range(pfn, pfn + nr_pages, MIGRATE_MOVABLE);
1195 * Freshly onlined pages aren't shuffled (e.g., all pages are placed to
1196 * the tail of the freelist when undoing isolation). Shuffle the whole
1197 * zone to make sure the just onlined pages are properly distributed
1198 * across the whole freelist - to create an initial shuffle.
1202 /* reinitialise watermarks and update pcp limits */
1203 init_per_zone_wmark_min();
1208 writeback_set_ratelimit();
1210 memory_notify(MEM_ONLINE, &arg);
1215 pr_debug("online_pages [mem %#010llx-%#010llx] failed\n",
1216 (unsigned long long) pfn << PAGE_SHIFT,
1217 (((unsigned long long) pfn + nr_pages) << PAGE_SHIFT) - 1);
1218 memory_notify(MEM_CANCEL_ONLINE, &arg);
1219 remove_pfn_range_from_zone(zone, pfn, nr_pages);
1224 /* we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG */
1225 static pg_data_t __ref *hotadd_init_pgdat(int nid)
1227 struct pglist_data *pgdat;
1230 * NODE_DATA is preallocated (free_area_init) but its internal
1231 * state is not allocated completely. Add missing pieces.
1232 * Completely offline nodes stay around and they just need
1235 pgdat = NODE_DATA(nid);
1237 /* init node's zones as empty zones, we don't have any present pages.*/
1238 free_area_init_core_hotplug(pgdat);
1241 * The node we allocated has no zone fallback lists. For avoiding
1242 * to access not-initialized zonelist, build here.
1244 build_all_zonelists(pgdat);
1250 * __try_online_node - online a node if offlined
1252 * @set_node_online: Whether we want to online the node
1253 * called by cpu_up() to online a node without onlined memory.
1256 * 1 -> a new node has been allocated
1257 * 0 -> the node is already online
1258 * -ENOMEM -> the node could not be allocated
1260 static int __try_online_node(int nid, bool set_node_online)
1265 if (node_online(nid))
1268 pgdat = hotadd_init_pgdat(nid);
1270 pr_err("Cannot online node %d due to NULL pgdat\n", nid);
1275 if (set_node_online) {
1276 node_set_online(nid);
1277 ret = register_one_node(nid);
1285 * Users of this function always want to online/register the node
1287 int try_online_node(int nid)
1291 mem_hotplug_begin();
1292 ret = __try_online_node(nid, true);
1297 static int check_hotplug_memory_range(u64 start, u64 size)
1299 /* memory range must be block size aligned */
1300 if (!size || !IS_ALIGNED(start, memory_block_size_bytes()) ||
1301 !IS_ALIGNED(size, memory_block_size_bytes())) {
1302 pr_err("Block size [%#lx] unaligned hotplug range: start %#llx, size %#llx",
1303 memory_block_size_bytes(), start, size);
1310 static int online_memory_block(struct memory_block *mem, void *arg)
1312 mem->online_type = mhp_default_online_type;
1313 return device_online(&mem->dev);
1316 #ifndef arch_supports_memmap_on_memory
1317 static inline bool arch_supports_memmap_on_memory(unsigned long vmemmap_size)
1320 * As default, we want the vmemmap to span a complete PMD such that we
1321 * can map the vmemmap using a single PMD if supported by the
1324 return IS_ALIGNED(vmemmap_size, PMD_SIZE);
1328 static bool mhp_supports_memmap_on_memory(unsigned long size)
1330 unsigned long vmemmap_size = memory_block_memmap_size();
1331 unsigned long memmap_pages = memory_block_memmap_on_memory_pages();
1334 * Besides having arch support and the feature enabled at runtime, we
1335 * need a few more assumptions to hold true:
1337 * a) We span a single memory block: memory onlining/offlinin;g happens
1338 * in memory block granularity. We don't want the vmemmap of online
1339 * memory blocks to reside on offline memory blocks. In the future,
1340 * we might want to support variable-sized memory blocks to make the
1341 * feature more versatile.
1343 * b) The vmemmap pages span complete PMDs: We don't want vmemmap code
1344 * to populate memory from the altmap for unrelated parts (i.e.,
1345 * other memory blocks)
1347 * c) The vmemmap pages (and thereby the pages that will be exposed to
1348 * the buddy) have to cover full pageblocks: memory onlining/offlining
1349 * code requires applicable ranges to be page-aligned, for example, to
1350 * set the migratetypes properly.
1352 * TODO: Although we have a check here to make sure that vmemmap pages
1353 * fully populate a PMD, it is not the right place to check for
1354 * this. A much better solution involves improving vmemmap code
1355 * to fallback to base pages when trying to populate vmemmap using
1356 * altmap as an alternative source of memory, and we do not exactly
1357 * populate a single PMD.
1359 if (!mhp_memmap_on_memory() || size != memory_block_size_bytes())
1363 * Make sure the vmemmap allocation is fully contained
1364 * so that we always allocate vmemmap memory from altmap area.
1366 if (!IS_ALIGNED(vmemmap_size, PAGE_SIZE))
1370 * start pfn should be pageblock_nr_pages aligned for correctly
1371 * setting migrate types
1373 if (!pageblock_aligned(memmap_pages))
1376 if (memmap_pages == PHYS_PFN(memory_block_size_bytes()))
1377 /* No effective hotplugged memory doesn't make sense. */
1380 return arch_supports_memmap_on_memory(vmemmap_size);
1384 * NOTE: The caller must call lock_device_hotplug() to serialize hotplug
1385 * and online/offline operations (triggered e.g. by sysfs).
1387 * we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG
1389 int __ref add_memory_resource(int nid, struct resource *res, mhp_t mhp_flags)
1391 struct mhp_params params = { .pgprot = pgprot_mhp(PAGE_KERNEL) };
1392 enum memblock_flags memblock_flags = MEMBLOCK_NONE;
1393 struct vmem_altmap mhp_altmap = {
1394 .base_pfn = PHYS_PFN(res->start),
1395 .end_pfn = PHYS_PFN(res->end),
1397 struct memory_group *group = NULL;
1399 bool new_node = false;
1403 size = resource_size(res);
1405 ret = check_hotplug_memory_range(start, size);
1409 if (mhp_flags & MHP_NID_IS_MGID) {
1410 group = memory_group_find_by_id(nid);
1416 if (!node_possible(nid)) {
1417 WARN(1, "node %d was absent from the node_possible_map\n", nid);
1421 mem_hotplug_begin();
1423 if (IS_ENABLED(CONFIG_ARCH_KEEP_MEMBLOCK)) {
1424 if (res->flags & IORESOURCE_SYSRAM_DRIVER_MANAGED)
1425 memblock_flags = MEMBLOCK_DRIVER_MANAGED;
1426 ret = memblock_add_node(start, size, nid, memblock_flags);
1428 goto error_mem_hotplug_end;
1431 ret = __try_online_node(nid, false);
1437 * Self hosted memmap array
1439 if (mhp_flags & MHP_MEMMAP_ON_MEMORY) {
1440 if (mhp_supports_memmap_on_memory(size)) {
1441 mhp_altmap.free = memory_block_memmap_on_memory_pages();
1442 params.altmap = kmalloc(sizeof(struct vmem_altmap), GFP_KERNEL);
1443 if (!params.altmap) {
1448 memcpy(params.altmap, &mhp_altmap, sizeof(mhp_altmap));
1450 /* fallback to not using altmap */
1453 /* call arch's memory hotadd */
1454 ret = arch_add_memory(nid, start, size, ¶ms);
1458 /* create memory block devices after memory was added */
1459 ret = create_memory_block_devices(start, size, params.altmap, group);
1461 arch_remove_memory(start, size, NULL);
1466 /* If sysfs file of new node can't be created, cpu on the node
1467 * can't be hot-added. There is no rollback way now.
1468 * So, check by BUG_ON() to catch it reluctantly..
1469 * We online node here. We can't roll back from here.
1471 node_set_online(nid);
1472 ret = __register_one_node(nid);
1476 register_memory_blocks_under_node(nid, PFN_DOWN(start),
1477 PFN_UP(start + size - 1),
1480 /* create new memmap entry */
1481 if (!strcmp(res->name, "System RAM"))
1482 firmware_map_add_hotplug(start, start + size, "System RAM");
1484 /* device_online() will take the lock when calling online_pages() */
1488 * In case we're allowed to merge the resource, flag it and trigger
1489 * merging now that adding succeeded.
1491 if (mhp_flags & MHP_MERGE_RESOURCE)
1492 merge_system_ram_resource(res);
1494 /* online pages if requested */
1495 if (mhp_default_online_type != MMOP_OFFLINE)
1496 walk_memory_blocks(start, size, NULL, online_memory_block);
1500 kfree(params.altmap);
1502 if (IS_ENABLED(CONFIG_ARCH_KEEP_MEMBLOCK))
1503 memblock_remove(start, size);
1504 error_mem_hotplug_end:
1509 /* requires device_hotplug_lock, see add_memory_resource() */
1510 int __ref __add_memory(int nid, u64 start, u64 size, mhp_t mhp_flags)
1512 struct resource *res;
1515 res = register_memory_resource(start, size, "System RAM");
1517 return PTR_ERR(res);
1519 ret = add_memory_resource(nid, res, mhp_flags);
1521 release_memory_resource(res);
1525 int add_memory(int nid, u64 start, u64 size, mhp_t mhp_flags)
1529 lock_device_hotplug();
1530 rc = __add_memory(nid, start, size, mhp_flags);
1531 unlock_device_hotplug();
1535 EXPORT_SYMBOL_GPL(add_memory);
1538 * Add special, driver-managed memory to the system as system RAM. Such
1539 * memory is not exposed via the raw firmware-provided memmap as system
1540 * RAM, instead, it is detected and added by a driver - during cold boot,
1541 * after a reboot, and after kexec.
1543 * Reasons why this memory should not be used for the initial memmap of a
1544 * kexec kernel or for placing kexec images:
1545 * - The booting kernel is in charge of determining how this memory will be
1546 * used (e.g., use persistent memory as system RAM)
1547 * - Coordination with a hypervisor is required before this memory
1548 * can be used (e.g., inaccessible parts).
1550 * For this memory, no entries in /sys/firmware/memmap ("raw firmware-provided
1551 * memory map") are created. Also, the created memory resource is flagged
1552 * with IORESOURCE_SYSRAM_DRIVER_MANAGED, so in-kernel users can special-case
1553 * this memory as well (esp., not place kexec images onto it).
1555 * The resource_name (visible via /proc/iomem) has to have the format
1556 * "System RAM ($DRIVER)".
1558 int add_memory_driver_managed(int nid, u64 start, u64 size,
1559 const char *resource_name, mhp_t mhp_flags)
1561 struct resource *res;
1564 if (!resource_name ||
1565 strstr(resource_name, "System RAM (") != resource_name ||
1566 resource_name[strlen(resource_name) - 1] != ')')
1569 lock_device_hotplug();
1571 res = register_memory_resource(start, size, resource_name);
1577 rc = add_memory_resource(nid, res, mhp_flags);
1579 release_memory_resource(res);
1582 unlock_device_hotplug();
1585 EXPORT_SYMBOL_GPL(add_memory_driver_managed);
1588 * Platforms should define arch_get_mappable_range() that provides
1589 * maximum possible addressable physical memory range for which the
1590 * linear mapping could be created. The platform returned address
1591 * range must adhere to these following semantics.
1593 * - range.start <= range.end
1594 * - Range includes both end points [range.start..range.end]
1596 * There is also a fallback definition provided here, allowing the
1597 * entire possible physical address range in case any platform does
1598 * not define arch_get_mappable_range().
1600 struct range __weak arch_get_mappable_range(void)
1602 struct range mhp_range = {
1609 struct range mhp_get_pluggable_range(bool need_mapping)
1611 const u64 max_phys = (1ULL << MAX_PHYSMEM_BITS) - 1;
1612 struct range mhp_range;
1615 mhp_range = arch_get_mappable_range();
1616 if (mhp_range.start > max_phys) {
1617 mhp_range.start = 0;
1620 mhp_range.end = min_t(u64, mhp_range.end, max_phys);
1622 mhp_range.start = 0;
1623 mhp_range.end = max_phys;
1627 EXPORT_SYMBOL_GPL(mhp_get_pluggable_range);
1629 bool mhp_range_allowed(u64 start, u64 size, bool need_mapping)
1631 struct range mhp_range = mhp_get_pluggable_range(need_mapping);
1632 u64 end = start + size;
1634 if (start < end && start >= mhp_range.start && (end - 1) <= mhp_range.end)
1637 pr_warn("Hotplug memory [%#llx-%#llx] exceeds maximum addressable range [%#llx-%#llx]\n",
1638 start, end, mhp_range.start, mhp_range.end);
1642 #ifdef CONFIG_MEMORY_HOTREMOVE
1644 * Scan pfn range [start,end) to find movable/migratable pages (LRU pages,
1645 * non-lru movable pages and hugepages). Will skip over most unmovable
1646 * pages (esp., pages that can be skipped when offlining), but bail out on
1647 * definitely unmovable pages.
1650 * 0 in case a movable page is found and movable_pfn was updated.
1651 * -ENOENT in case no movable page was found.
1652 * -EBUSY in case a definitely unmovable page was found.
1654 static int scan_movable_pages(unsigned long start, unsigned long end,
1655 unsigned long *movable_pfn)
1659 for (pfn = start; pfn < end; pfn++) {
1660 struct page *page, *head;
1663 if (!pfn_valid(pfn))
1665 page = pfn_to_page(pfn);
1668 if (__PageMovable(page))
1672 * PageOffline() pages that are not marked __PageMovable() and
1673 * have a reference count > 0 (after MEM_GOING_OFFLINE) are
1674 * definitely unmovable. If their reference count would be 0,
1675 * they could at least be skipped when offlining memory.
1677 if (PageOffline(page) && page_count(page))
1680 if (!PageHuge(page))
1682 head = compound_head(page);
1684 * This test is racy as we hold no reference or lock. The
1685 * hugetlb page could have been free'ed and head is no longer
1686 * a hugetlb page before the following check. In such unlikely
1687 * cases false positives and negatives are possible. Calling
1688 * code must deal with these scenarios.
1690 if (HPageMigratable(head))
1692 skip = compound_nr(head) - (page - head);
1701 static void do_migrate_range(unsigned long start_pfn, unsigned long end_pfn)
1704 struct page *page, *head;
1706 static DEFINE_RATELIMIT_STATE(migrate_rs, DEFAULT_RATELIMIT_INTERVAL,
1707 DEFAULT_RATELIMIT_BURST);
1709 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
1710 struct folio *folio;
1713 if (!pfn_valid(pfn))
1715 page = pfn_to_page(pfn);
1716 folio = page_folio(page);
1717 head = &folio->page;
1719 if (PageHuge(page)) {
1720 pfn = page_to_pfn(head) + compound_nr(head) - 1;
1721 isolate_hugetlb(folio, &source);
1723 } else if (PageTransHuge(page))
1724 pfn = page_to_pfn(head) + thp_nr_pages(page) - 1;
1727 * HWPoison pages have elevated reference counts so the migration would
1728 * fail on them. It also doesn't make any sense to migrate them in the
1729 * first place. Still try to unmap such a page in case it is still mapped
1730 * (e.g. current hwpoison implementation doesn't unmap KSM pages but keep
1731 * the unmap as the catch all safety net).
1733 if (PageHWPoison(page)) {
1734 if (WARN_ON(folio_test_lru(folio)))
1735 folio_isolate_lru(folio);
1736 if (folio_mapped(folio))
1737 try_to_unmap(folio, TTU_IGNORE_MLOCK);
1741 if (!get_page_unless_zero(page))
1744 * We can skip free pages. And we can deal with pages on
1745 * LRU and non-lru movable pages.
1748 isolated = isolate_lru_page(page);
1750 isolated = isolate_movable_page(page, ISOLATE_UNEVICTABLE);
1752 list_add_tail(&page->lru, &source);
1753 if (!__PageMovable(page))
1754 inc_node_page_state(page, NR_ISOLATED_ANON +
1755 page_is_file_lru(page));
1758 if (__ratelimit(&migrate_rs)) {
1759 pr_warn("failed to isolate pfn %lx\n", pfn);
1760 dump_page(page, "isolation failed");
1765 if (!list_empty(&source)) {
1766 nodemask_t nmask = node_states[N_MEMORY];
1767 struct migration_target_control mtc = {
1769 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
1774 * We have checked that migration range is on a single zone so
1775 * we can use the nid of the first page to all the others.
1777 mtc.nid = page_to_nid(list_first_entry(&source, struct page, lru));
1780 * try to allocate from a different node but reuse this node
1781 * if there are no other online nodes to be used (e.g. we are
1782 * offlining a part of the only existing node)
1784 node_clear(mtc.nid, nmask);
1785 if (nodes_empty(nmask))
1786 node_set(mtc.nid, nmask);
1787 ret = migrate_pages(&source, alloc_migration_target, NULL,
1788 (unsigned long)&mtc, MIGRATE_SYNC, MR_MEMORY_HOTPLUG, NULL);
1790 list_for_each_entry(page, &source, lru) {
1791 if (__ratelimit(&migrate_rs)) {
1792 pr_warn("migrating pfn %lx failed ret:%d\n",
1793 page_to_pfn(page), ret);
1794 dump_page(page, "migration failure");
1797 putback_movable_pages(&source);
1802 static int __init cmdline_parse_movable_node(char *p)
1804 movable_node_enabled = true;
1807 early_param("movable_node", cmdline_parse_movable_node);
1809 /* check which state of node_states will be changed when offline memory */
1810 static void node_states_check_changes_offline(unsigned long nr_pages,
1811 struct zone *zone, struct memory_notify *arg)
1813 struct pglist_data *pgdat = zone->zone_pgdat;
1814 unsigned long present_pages = 0;
1817 arg->status_change_nid = NUMA_NO_NODE;
1818 arg->status_change_nid_normal = NUMA_NO_NODE;
1821 * Check whether node_states[N_NORMAL_MEMORY] will be changed.
1822 * If the memory to be offline is within the range
1823 * [0..ZONE_NORMAL], and it is the last present memory there,
1824 * the zones in that range will become empty after the offlining,
1825 * thus we can determine that we need to clear the node from
1826 * node_states[N_NORMAL_MEMORY].
1828 for (zt = 0; zt <= ZONE_NORMAL; zt++)
1829 present_pages += pgdat->node_zones[zt].present_pages;
1830 if (zone_idx(zone) <= ZONE_NORMAL && nr_pages >= present_pages)
1831 arg->status_change_nid_normal = zone_to_nid(zone);
1834 * We have accounted the pages from [0..ZONE_NORMAL); ZONE_HIGHMEM
1835 * does not apply as we don't support 32bit.
1836 * Here we count the possible pages from ZONE_MOVABLE.
1837 * If after having accounted all the pages, we see that the nr_pages
1838 * to be offlined is over or equal to the accounted pages,
1839 * we know that the node will become empty, and so, we can clear
1840 * it for N_MEMORY as well.
1842 present_pages += pgdat->node_zones[ZONE_MOVABLE].present_pages;
1844 if (nr_pages >= present_pages)
1845 arg->status_change_nid = zone_to_nid(zone);
1848 static void node_states_clear_node(int node, struct memory_notify *arg)
1850 if (arg->status_change_nid_normal >= 0)
1851 node_clear_state(node, N_NORMAL_MEMORY);
1853 if (arg->status_change_nid >= 0)
1854 node_clear_state(node, N_MEMORY);
1857 static int count_system_ram_pages_cb(unsigned long start_pfn,
1858 unsigned long nr_pages, void *data)
1860 unsigned long *nr_system_ram_pages = data;
1862 *nr_system_ram_pages += nr_pages;
1866 int __ref offline_pages(unsigned long start_pfn, unsigned long nr_pages,
1867 struct zone *zone, struct memory_group *group)
1869 const unsigned long end_pfn = start_pfn + nr_pages;
1870 unsigned long pfn, system_ram_pages = 0;
1871 const int node = zone_to_nid(zone);
1872 unsigned long flags;
1873 struct memory_notify arg;
1878 * {on,off}lining is constrained to full memory sections (or more
1879 * precisely to memory blocks from the user space POV).
1880 * memmap_on_memory is an exception because it reserves initial part
1881 * of the physical memory space for vmemmaps. That space is pageblock
1884 if (WARN_ON_ONCE(!nr_pages || !pageblock_aligned(start_pfn) ||
1885 !IS_ALIGNED(start_pfn + nr_pages, PAGES_PER_SECTION)))
1888 mem_hotplug_begin();
1891 * Don't allow to offline memory blocks that contain holes.
1892 * Consequently, memory blocks with holes can never get onlined
1893 * via the hotplug path - online_pages() - as hotplugged memory has
1894 * no holes. This way, we e.g., don't have to worry about marking
1895 * memory holes PG_reserved, don't need pfn_valid() checks, and can
1896 * avoid using walk_system_ram_range() later.
1898 walk_system_ram_range(start_pfn, nr_pages, &system_ram_pages,
1899 count_system_ram_pages_cb);
1900 if (system_ram_pages != nr_pages) {
1902 reason = "memory holes";
1903 goto failed_removal;
1907 * We only support offlining of memory blocks managed by a single zone,
1908 * checked by calling code. This is just a sanity check that we might
1909 * want to remove in the future.
1911 if (WARN_ON_ONCE(page_zone(pfn_to_page(start_pfn)) != zone ||
1912 page_zone(pfn_to_page(end_pfn - 1)) != zone)) {
1914 reason = "multizone range";
1915 goto failed_removal;
1919 * Disable pcplists so that page isolation cannot race with freeing
1920 * in a way that pages from isolated pageblock are left on pcplists.
1922 zone_pcp_disable(zone);
1923 lru_cache_disable();
1925 /* set above range as isolated */
1926 ret = start_isolate_page_range(start_pfn, end_pfn,
1928 MEMORY_OFFLINE | REPORT_FAILURE,
1929 GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL);
1931 reason = "failure to isolate range";
1932 goto failed_removal_pcplists_disabled;
1935 arg.start_pfn = start_pfn;
1936 arg.nr_pages = nr_pages;
1937 node_states_check_changes_offline(nr_pages, zone, &arg);
1939 ret = memory_notify(MEM_GOING_OFFLINE, &arg);
1940 ret = notifier_to_errno(ret);
1942 reason = "notifier failure";
1943 goto failed_removal_isolated;
1950 * Historically we always checked for any signal and
1951 * can't limit it to fatal signals without eventually
1952 * breaking user space.
1954 if (signal_pending(current)) {
1956 reason = "signal backoff";
1957 goto failed_removal_isolated;
1962 ret = scan_movable_pages(pfn, end_pfn, &pfn);
1965 * TODO: fatal migration failures should bail
1968 do_migrate_range(pfn, end_pfn);
1972 if (ret != -ENOENT) {
1973 reason = "unmovable page";
1974 goto failed_removal_isolated;
1978 * Dissolve free hugepages in the memory block before doing
1979 * offlining actually in order to make hugetlbfs's object
1980 * counting consistent.
1982 ret = dissolve_free_huge_pages(start_pfn, end_pfn);
1984 reason = "failure to dissolve huge pages";
1985 goto failed_removal_isolated;
1988 ret = test_pages_isolated(start_pfn, end_pfn, MEMORY_OFFLINE);
1992 /* Mark all sections offline and remove free pages from the buddy. */
1993 __offline_isolated_pages(start_pfn, end_pfn);
1994 pr_debug("Offlined Pages %ld\n", nr_pages);
1997 * The memory sections are marked offline, and the pageblock flags
1998 * effectively stale; nobody should be touching them. Fixup the number
1999 * of isolated pageblocks, memory onlining will properly revert this.
2001 spin_lock_irqsave(&zone->lock, flags);
2002 zone->nr_isolate_pageblock -= nr_pages / pageblock_nr_pages;
2003 spin_unlock_irqrestore(&zone->lock, flags);
2006 zone_pcp_enable(zone);
2008 /* removal success */
2009 adjust_managed_page_count(pfn_to_page(start_pfn), -nr_pages);
2010 adjust_present_page_count(pfn_to_page(start_pfn), group, -nr_pages);
2012 /* reinitialise watermarks and update pcp limits */
2013 init_per_zone_wmark_min();
2015 if (!populated_zone(zone)) {
2016 zone_pcp_reset(zone);
2017 build_all_zonelists(NULL);
2020 node_states_clear_node(node, &arg);
2021 if (arg.status_change_nid >= 0) {
2022 kcompactd_stop(node);
2026 writeback_set_ratelimit();
2028 memory_notify(MEM_OFFLINE, &arg);
2029 remove_pfn_range_from_zone(zone, start_pfn, nr_pages);
2033 failed_removal_isolated:
2034 /* pushback to free area */
2035 undo_isolate_page_range(start_pfn, end_pfn, MIGRATE_MOVABLE);
2036 memory_notify(MEM_CANCEL_OFFLINE, &arg);
2037 failed_removal_pcplists_disabled:
2039 zone_pcp_enable(zone);
2041 pr_debug("memory offlining [mem %#010llx-%#010llx] failed due to %s\n",
2042 (unsigned long long) start_pfn << PAGE_SHIFT,
2043 ((unsigned long long) end_pfn << PAGE_SHIFT) - 1,
2049 static int check_memblock_offlined_cb(struct memory_block *mem, void *arg)
2054 if (unlikely(mem->state != MEM_OFFLINE)) {
2055 phys_addr_t beginpa, endpa;
2057 beginpa = PFN_PHYS(section_nr_to_pfn(mem->start_section_nr));
2058 endpa = beginpa + memory_block_size_bytes() - 1;
2059 pr_warn("removing memory fails, because memory [%pa-%pa] is onlined\n",
2067 static int test_has_altmap_cb(struct memory_block *mem, void *arg)
2069 struct memory_block **mem_ptr = (struct memory_block **)arg;
2071 * return the memblock if we have altmap
2072 * and break callback.
2081 static int check_cpu_on_node(int nid)
2085 for_each_present_cpu(cpu) {
2086 if (cpu_to_node(cpu) == nid)
2088 * the cpu on this node isn't removed, and we can't
2089 * offline this node.
2097 static int check_no_memblock_for_node_cb(struct memory_block *mem, void *arg)
2099 int nid = *(int *)arg;
2102 * If a memory block belongs to multiple nodes, the stored nid is not
2103 * reliable. However, such blocks are always online (e.g., cannot get
2104 * offlined) and, therefore, are still spanned by the node.
2106 return mem->nid == nid ? -EEXIST : 0;
2113 * Offline a node if all memory sections and cpus of the node are removed.
2115 * NOTE: The caller must call lock_device_hotplug() to serialize hotplug
2116 * and online/offline operations before this call.
2118 void try_offline_node(int nid)
2123 * If the node still spans pages (especially ZONE_DEVICE), don't
2124 * offline it. A node spans memory after move_pfn_range_to_zone(),
2125 * e.g., after the memory block was onlined.
2127 if (node_spanned_pages(nid))
2131 * Especially offline memory blocks might not be spanned by the
2132 * node. They will get spanned by the node once they get onlined.
2133 * However, they link to the node in sysfs and can get onlined later.
2135 rc = for_each_memory_block(&nid, check_no_memblock_for_node_cb);
2139 if (check_cpu_on_node(nid))
2143 * all memory/cpu of this node are removed, we can offline this
2146 node_set_offline(nid);
2147 unregister_one_node(nid);
2149 EXPORT_SYMBOL(try_offline_node);
2151 static int __ref try_remove_memory(u64 start, u64 size)
2153 struct memory_block *mem;
2154 int rc = 0, nid = NUMA_NO_NODE;
2155 struct vmem_altmap *altmap = NULL;
2157 BUG_ON(check_hotplug_memory_range(start, size));
2160 * All memory blocks must be offlined before removing memory. Check
2161 * whether all memory blocks in question are offline and return error
2162 * if this is not the case.
2164 * While at it, determine the nid. Note that if we'd have mixed nodes,
2165 * we'd only try to offline the last determined one -- which is good
2166 * enough for the cases we care about.
2168 rc = walk_memory_blocks(start, size, &nid, check_memblock_offlined_cb);
2173 * We only support removing memory added with MHP_MEMMAP_ON_MEMORY in
2174 * the same granularity it was added - a single memory block.
2176 if (mhp_memmap_on_memory()) {
2177 rc = walk_memory_blocks(start, size, &mem, test_has_altmap_cb);
2179 if (size != memory_block_size_bytes()) {
2180 pr_warn("Refuse to remove %#llx - %#llx,"
2181 "wrong granularity\n",
2182 start, start + size);
2185 altmap = mem->altmap;
2187 * Mark altmap NULL so that we can add a debug
2188 * check on memblock free.
2194 /* remove memmap entry */
2195 firmware_map_remove(start, start + size, "System RAM");
2198 * Memory block device removal under the device_hotplug_lock is
2199 * a barrier against racing online attempts.
2201 remove_memory_block_devices(start, size);
2203 mem_hotplug_begin();
2205 arch_remove_memory(start, size, altmap);
2207 /* Verify that all vmemmap pages have actually been freed. */
2209 WARN(altmap->alloc, "Altmap not fully unmapped");
2213 if (IS_ENABLED(CONFIG_ARCH_KEEP_MEMBLOCK)) {
2214 memblock_phys_free(start, size);
2215 memblock_remove(start, size);
2218 release_mem_region_adjustable(start, size);
2220 if (nid != NUMA_NO_NODE)
2221 try_offline_node(nid);
2228 * __remove_memory - Remove memory if every memory block is offline
2229 * @start: physical address of the region to remove
2230 * @size: size of the region to remove
2232 * NOTE: The caller must call lock_device_hotplug() to serialize hotplug
2233 * and online/offline operations before this call, as required by
2234 * try_offline_node().
2236 void __remove_memory(u64 start, u64 size)
2240 * trigger BUG() if some memory is not offlined prior to calling this
2243 if (try_remove_memory(start, size))
2248 * Remove memory if every memory block is offline, otherwise return -EBUSY is
2249 * some memory is not offline
2251 int remove_memory(u64 start, u64 size)
2255 lock_device_hotplug();
2256 rc = try_remove_memory(start, size);
2257 unlock_device_hotplug();
2261 EXPORT_SYMBOL_GPL(remove_memory);
2263 static int try_offline_memory_block(struct memory_block *mem, void *arg)
2265 uint8_t online_type = MMOP_ONLINE_KERNEL;
2266 uint8_t **online_types = arg;
2271 * Sense the online_type via the zone of the memory block. Offlining
2272 * with multiple zones within one memory block will be rejected
2273 * by offlining code ... so we don't care about that.
2275 page = pfn_to_online_page(section_nr_to_pfn(mem->start_section_nr));
2276 if (page && zone_idx(page_zone(page)) == ZONE_MOVABLE)
2277 online_type = MMOP_ONLINE_MOVABLE;
2279 rc = device_offline(&mem->dev);
2281 * Default is MMOP_OFFLINE - change it only if offlining succeeded,
2282 * so try_reonline_memory_block() can do the right thing.
2285 **online_types = online_type;
2288 /* Ignore if already offline. */
2289 return rc < 0 ? rc : 0;
2292 static int try_reonline_memory_block(struct memory_block *mem, void *arg)
2294 uint8_t **online_types = arg;
2297 if (**online_types != MMOP_OFFLINE) {
2298 mem->online_type = **online_types;
2299 rc = device_online(&mem->dev);
2301 pr_warn("%s: Failed to re-online memory: %d",
2305 /* Continue processing all remaining memory blocks. */
2311 * Try to offline and remove memory. Might take a long time to finish in case
2312 * memory is still in use. Primarily useful for memory devices that logically
2313 * unplugged all memory (so it's no longer in use) and want to offline + remove
2316 int offline_and_remove_memory(u64 start, u64 size)
2318 const unsigned long mb_count = size / memory_block_size_bytes();
2319 uint8_t *online_types, *tmp;
2322 if (!IS_ALIGNED(start, memory_block_size_bytes()) ||
2323 !IS_ALIGNED(size, memory_block_size_bytes()) || !size)
2327 * We'll remember the old online type of each memory block, so we can
2328 * try to revert whatever we did when offlining one memory block fails
2329 * after offlining some others succeeded.
2331 online_types = kmalloc_array(mb_count, sizeof(*online_types),
2336 * Initialize all states to MMOP_OFFLINE, so when we abort processing in
2337 * try_offline_memory_block(), we'll skip all unprocessed blocks in
2338 * try_reonline_memory_block().
2340 memset(online_types, MMOP_OFFLINE, mb_count);
2342 lock_device_hotplug();
2345 rc = walk_memory_blocks(start, size, &tmp, try_offline_memory_block);
2348 * In case we succeeded to offline all memory, remove it.
2349 * This cannot fail as it cannot get onlined in the meantime.
2352 rc = try_remove_memory(start, size);
2354 pr_err("%s: Failed to remove memory: %d", __func__, rc);
2358 * Rollback what we did. While memory onlining might theoretically fail
2359 * (nacked by a notifier), it barely ever happens.
2363 walk_memory_blocks(start, size, &tmp,
2364 try_reonline_memory_block);
2366 unlock_device_hotplug();
2368 kfree(online_types);
2371 EXPORT_SYMBOL_GPL(offline_and_remove_memory);
2372 #endif /* CONFIG_MEMORY_HOTREMOVE */