1 // SPDX-License-Identifier: GPL-2.0-only
3 * linux/mm/memory_hotplug.c
8 #include <linux/stddef.h>
10 #include <linux/sched/signal.h>
11 #include <linux/swap.h>
12 #include <linux/interrupt.h>
13 #include <linux/pagemap.h>
14 #include <linux/compiler.h>
15 #include <linux/export.h>
16 #include <linux/pagevec.h>
17 #include <linux/writeback.h>
18 #include <linux/slab.h>
19 #include <linux/sysctl.h>
20 #include <linux/cpu.h>
21 #include <linux/memory.h>
22 #include <linux/memremap.h>
23 #include <linux/memory_hotplug.h>
24 #include <linux/highmem.h>
25 #include <linux/vmalloc.h>
26 #include <linux/ioport.h>
27 #include <linux/delay.h>
28 #include <linux/migrate.h>
29 #include <linux/page-isolation.h>
30 #include <linux/pfn.h>
31 #include <linux/suspend.h>
32 #include <linux/mm_inline.h>
33 #include <linux/firmware-map.h>
34 #include <linux/stop_machine.h>
35 #include <linux/hugetlb.h>
36 #include <linux/memblock.h>
37 #include <linux/compaction.h>
38 #include <linux/rmap.h>
40 #include <asm/tlbflush.h>
47 * memory_hotplug.memmap_on_memory parameter
49 static bool memmap_on_memory __ro_after_init
;
50 #ifdef CONFIG_MHP_MEMMAP_ON_MEMORY
51 module_param(memmap_on_memory
, bool, 0444);
52 MODULE_PARM_DESC(memmap_on_memory
, "Enable memmap on memory for memory hotplug");
56 ONLINE_POLICY_CONTIG_ZONES
= 0,
57 ONLINE_POLICY_AUTO_MOVABLE
,
60 const char *online_policy_to_str
[] = {
61 [ONLINE_POLICY_CONTIG_ZONES
] = "contig-zones",
62 [ONLINE_POLICY_AUTO_MOVABLE
] = "auto-movable",
65 static int set_online_policy(const char *val
, const struct kernel_param
*kp
)
67 int ret
= sysfs_match_string(online_policy_to_str
, val
);
71 *((int *)kp
->arg
) = ret
;
75 static int get_online_policy(char *buffer
, const struct kernel_param
*kp
)
77 return sprintf(buffer
, "%s\n", online_policy_to_str
[*((int *)kp
->arg
)]);
81 * memory_hotplug.online_policy: configure online behavior when onlining without
82 * specifying a zone (MMOP_ONLINE)
84 * "contig-zones": keep zone contiguous
85 * "auto-movable": online memory to ZONE_MOVABLE if the configuration
86 * (auto_movable_ratio, auto_movable_numa_aware) allows for it
88 static int online_policy __read_mostly
= ONLINE_POLICY_CONTIG_ZONES
;
89 static const struct kernel_param_ops online_policy_ops
= {
90 .set
= set_online_policy
,
91 .get
= get_online_policy
,
93 module_param_cb(online_policy
, &online_policy_ops
, &online_policy
, 0644);
94 MODULE_PARM_DESC(online_policy
,
95 "Set the online policy (\"contig-zones\", \"auto-movable\") "
96 "Default: \"contig-zones\"");
99 * memory_hotplug.auto_movable_ratio: specify maximum MOVABLE:KERNEL ratio
101 * The ratio represent an upper limit and the kernel might decide to not
102 * online some memory to ZONE_MOVABLE -- e.g., because hotplugged KERNEL memory
103 * doesn't allow for more MOVABLE memory.
105 static unsigned int auto_movable_ratio __read_mostly
= 301;
106 module_param(auto_movable_ratio
, uint
, 0644);
107 MODULE_PARM_DESC(auto_movable_ratio
,
108 "Set the maximum ratio of MOVABLE:KERNEL memory in the system "
109 "in percent for \"auto-movable\" online policy. Default: 301");
112 * memory_hotplug.auto_movable_numa_aware: consider numa node stats
115 static bool auto_movable_numa_aware __read_mostly
= true;
116 module_param(auto_movable_numa_aware
, bool, 0644);
117 MODULE_PARM_DESC(auto_movable_numa_aware
,
118 "Consider numa node stats in addition to global stats in "
119 "\"auto-movable\" online policy. Default: true");
120 #endif /* CONFIG_NUMA */
123 * online_page_callback contains pointer to current page onlining function.
124 * Initially it is generic_online_page(). If it is required it could be
125 * changed by calling set_online_page_callback() for callback registration
126 * and restore_online_page_callback() for generic callback restore.
129 static online_page_callback_t online_page_callback
= generic_online_page
;
130 static DEFINE_MUTEX(online_page_callback_lock
);
132 DEFINE_STATIC_PERCPU_RWSEM(mem_hotplug_lock
);
134 void get_online_mems(void)
136 percpu_down_read(&mem_hotplug_lock
);
139 void put_online_mems(void)
141 percpu_up_read(&mem_hotplug_lock
);
144 bool movable_node_enabled
= false;
146 #ifndef CONFIG_MEMORY_HOTPLUG_DEFAULT_ONLINE
147 int mhp_default_online_type
= MMOP_OFFLINE
;
149 int mhp_default_online_type
= MMOP_ONLINE
;
152 static int __init
setup_memhp_default_state(char *str
)
154 const int online_type
= mhp_online_type_from_str(str
);
156 if (online_type
>= 0)
157 mhp_default_online_type
= online_type
;
161 __setup("memhp_default_state=", setup_memhp_default_state
);
163 void mem_hotplug_begin(void)
166 percpu_down_write(&mem_hotplug_lock
);
169 void mem_hotplug_done(void)
171 percpu_up_write(&mem_hotplug_lock
);
175 u64 max_mem_size
= U64_MAX
;
177 /* add this memory to iomem resource */
178 static struct resource
*register_memory_resource(u64 start
, u64 size
,
179 const char *resource_name
)
181 struct resource
*res
;
182 unsigned long flags
= IORESOURCE_SYSTEM_RAM
| IORESOURCE_BUSY
;
184 if (strcmp(resource_name
, "System RAM"))
185 flags
|= IORESOURCE_SYSRAM_DRIVER_MANAGED
;
187 if (!mhp_range_allowed(start
, size
, true))
188 return ERR_PTR(-E2BIG
);
191 * Make sure value parsed from 'mem=' only restricts memory adding
192 * while booting, so that memory hotplug won't be impacted. Please
193 * refer to document of 'mem=' in kernel-parameters.txt for more
196 if (start
+ size
> max_mem_size
&& system_state
< SYSTEM_RUNNING
)
197 return ERR_PTR(-E2BIG
);
200 * Request ownership of the new memory range. This might be
201 * a child of an existing resource that was present but
202 * not marked as busy.
204 res
= __request_region(&iomem_resource
, start
, size
,
205 resource_name
, flags
);
208 pr_debug("Unable to reserve System RAM region: %016llx->%016llx\n",
209 start
, start
+ size
);
210 return ERR_PTR(-EEXIST
);
215 static void release_memory_resource(struct resource
*res
)
219 release_resource(res
);
223 #ifdef CONFIG_MEMORY_HOTPLUG_SPARSE
224 static int check_pfn_span(unsigned long pfn
, unsigned long nr_pages
,
228 * Disallow all operations smaller than a sub-section and only
229 * allow operations smaller than a section for
230 * SPARSEMEM_VMEMMAP. Note that check_hotplug_memory_range()
231 * enforces a larger memory_block_size_bytes() granularity for
232 * memory that will be marked online, so this check should only
233 * fire for direct arch_{add,remove}_memory() users outside of
234 * add_memory_resource().
236 unsigned long min_align
;
238 if (IS_ENABLED(CONFIG_SPARSEMEM_VMEMMAP
))
239 min_align
= PAGES_PER_SUBSECTION
;
241 min_align
= PAGES_PER_SECTION
;
242 if (!IS_ALIGNED(pfn
, min_align
)
243 || !IS_ALIGNED(nr_pages
, min_align
)) {
244 WARN(1, "Misaligned __%s_pages start: %#lx end: #%lx\n",
245 reason
, pfn
, pfn
+ nr_pages
- 1);
252 * Return page for the valid pfn only if the page is online. All pfn
253 * walkers which rely on the fully initialized page->flags and others
254 * should use this rather than pfn_valid && pfn_to_page
256 struct page
*pfn_to_online_page(unsigned long pfn
)
258 unsigned long nr
= pfn_to_section_nr(pfn
);
259 struct dev_pagemap
*pgmap
;
260 struct mem_section
*ms
;
262 if (nr
>= NR_MEM_SECTIONS
)
265 ms
= __nr_to_section(nr
);
266 if (!online_section(ms
))
270 * Save some code text when online_section() +
271 * pfn_section_valid() are sufficient.
273 if (IS_ENABLED(CONFIG_HAVE_ARCH_PFN_VALID
) && !pfn_valid(pfn
))
276 if (!pfn_section_valid(ms
, pfn
))
279 if (!online_device_section(ms
))
280 return pfn_to_page(pfn
);
283 * Slowpath: when ZONE_DEVICE collides with
284 * ZONE_{NORMAL,MOVABLE} within the same section some pfns in
285 * the section may be 'offline' but 'valid'. Only
286 * get_dev_pagemap() can determine sub-section online status.
288 pgmap
= get_dev_pagemap(pfn
, NULL
);
289 put_dev_pagemap(pgmap
);
291 /* The presence of a pgmap indicates ZONE_DEVICE offline pfn */
295 return pfn_to_page(pfn
);
297 EXPORT_SYMBOL_GPL(pfn_to_online_page
);
300 * Reasonably generic function for adding memory. It is
301 * expected that archs that support memory hotplug will
302 * call this function after deciding the zone to which to
305 int __ref
__add_pages(int nid
, unsigned long pfn
, unsigned long nr_pages
,
306 struct mhp_params
*params
)
308 const unsigned long end_pfn
= pfn
+ nr_pages
;
309 unsigned long cur_nr_pages
;
311 struct vmem_altmap
*altmap
= params
->altmap
;
313 if (WARN_ON_ONCE(!params
->pgprot
.pgprot
))
316 VM_BUG_ON(!mhp_range_allowed(PFN_PHYS(pfn
), nr_pages
* PAGE_SIZE
, false));
320 * Validate altmap is within bounds of the total request
322 if (altmap
->base_pfn
!= pfn
323 || vmem_altmap_offset(altmap
) > nr_pages
) {
324 pr_warn_once("memory add fail, invalid altmap\n");
330 err
= check_pfn_span(pfn
, nr_pages
, "add");
334 for (; pfn
< end_pfn
; pfn
+= cur_nr_pages
) {
335 /* Select all remaining pages up to the next section boundary */
336 cur_nr_pages
= min(end_pfn
- pfn
,
337 SECTION_ALIGN_UP(pfn
+ 1) - pfn
);
338 err
= sparse_add_section(nid
, pfn
, cur_nr_pages
, altmap
);
343 vmemmap_populate_print_last();
347 /* find the smallest valid pfn in the range [start_pfn, end_pfn) */
348 static unsigned long find_smallest_section_pfn(int nid
, struct zone
*zone
,
349 unsigned long start_pfn
,
350 unsigned long end_pfn
)
352 for (; start_pfn
< end_pfn
; start_pfn
+= PAGES_PER_SUBSECTION
) {
353 if (unlikely(!pfn_to_online_page(start_pfn
)))
356 if (unlikely(pfn_to_nid(start_pfn
) != nid
))
359 if (zone
!= page_zone(pfn_to_page(start_pfn
)))
368 /* find the biggest valid pfn in the range [start_pfn, end_pfn). */
369 static unsigned long find_biggest_section_pfn(int nid
, struct zone
*zone
,
370 unsigned long start_pfn
,
371 unsigned long end_pfn
)
375 /* pfn is the end pfn of a memory section. */
377 for (; pfn
>= start_pfn
; pfn
-= PAGES_PER_SUBSECTION
) {
378 if (unlikely(!pfn_to_online_page(pfn
)))
381 if (unlikely(pfn_to_nid(pfn
) != nid
))
384 if (zone
!= page_zone(pfn_to_page(pfn
)))
393 static void shrink_zone_span(struct zone
*zone
, unsigned long start_pfn
,
394 unsigned long end_pfn
)
397 int nid
= zone_to_nid(zone
);
399 if (zone
->zone_start_pfn
== start_pfn
) {
401 * If the section is smallest section in the zone, it need
402 * shrink zone->zone_start_pfn and zone->zone_spanned_pages.
403 * In this case, we find second smallest valid mem_section
404 * for shrinking zone.
406 pfn
= find_smallest_section_pfn(nid
, zone
, end_pfn
,
409 zone
->spanned_pages
= zone_end_pfn(zone
) - pfn
;
410 zone
->zone_start_pfn
= pfn
;
412 zone
->zone_start_pfn
= 0;
413 zone
->spanned_pages
= 0;
415 } else if (zone_end_pfn(zone
) == end_pfn
) {
417 * If the section is biggest section in the zone, it need
418 * shrink zone->spanned_pages.
419 * In this case, we find second biggest valid mem_section for
422 pfn
= find_biggest_section_pfn(nid
, zone
, zone
->zone_start_pfn
,
425 zone
->spanned_pages
= pfn
- zone
->zone_start_pfn
+ 1;
427 zone
->zone_start_pfn
= 0;
428 zone
->spanned_pages
= 0;
433 static void update_pgdat_span(struct pglist_data
*pgdat
)
435 unsigned long node_start_pfn
= 0, node_end_pfn
= 0;
438 for (zone
= pgdat
->node_zones
;
439 zone
< pgdat
->node_zones
+ MAX_NR_ZONES
; zone
++) {
440 unsigned long end_pfn
= zone_end_pfn(zone
);
442 /* No need to lock the zones, they can't change. */
443 if (!zone
->spanned_pages
)
446 node_start_pfn
= zone
->zone_start_pfn
;
447 node_end_pfn
= end_pfn
;
451 if (end_pfn
> node_end_pfn
)
452 node_end_pfn
= end_pfn
;
453 if (zone
->zone_start_pfn
< node_start_pfn
)
454 node_start_pfn
= zone
->zone_start_pfn
;
457 pgdat
->node_start_pfn
= node_start_pfn
;
458 pgdat
->node_spanned_pages
= node_end_pfn
- node_start_pfn
;
461 void __ref
remove_pfn_range_from_zone(struct zone
*zone
,
462 unsigned long start_pfn
,
463 unsigned long nr_pages
)
465 const unsigned long end_pfn
= start_pfn
+ nr_pages
;
466 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
467 unsigned long pfn
, cur_nr_pages
;
469 /* Poison struct pages because they are now uninitialized again. */
470 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
+= cur_nr_pages
) {
473 /* Select all remaining pages up to the next section boundary */
475 min(end_pfn
- pfn
, SECTION_ALIGN_UP(pfn
+ 1) - pfn
);
476 page_init_poison(pfn_to_page(pfn
),
477 sizeof(struct page
) * cur_nr_pages
);
481 * Zone shrinking code cannot properly deal with ZONE_DEVICE. So
482 * we will not try to shrink the zones - which is okay as
483 * set_zone_contiguous() cannot deal with ZONE_DEVICE either way.
485 if (zone_is_zone_device(zone
))
488 clear_zone_contiguous(zone
);
490 shrink_zone_span(zone
, start_pfn
, start_pfn
+ nr_pages
);
491 update_pgdat_span(pgdat
);
493 set_zone_contiguous(zone
);
496 static void __remove_section(unsigned long pfn
, unsigned long nr_pages
,
497 unsigned long map_offset
,
498 struct vmem_altmap
*altmap
)
500 struct mem_section
*ms
= __pfn_to_section(pfn
);
502 if (WARN_ON_ONCE(!valid_section(ms
)))
505 sparse_remove_section(ms
, pfn
, nr_pages
, map_offset
, altmap
);
509 * __remove_pages() - remove sections of pages
510 * @pfn: starting pageframe (must be aligned to start of a section)
511 * @nr_pages: number of pages to remove (must be multiple of section size)
512 * @altmap: alternative device page map or %NULL if default memmap is used
514 * Generic helper function to remove section mappings and sysfs entries
515 * for the section of the memory we are removing. Caller needs to make
516 * sure that pages are marked reserved and zones are adjust properly by
517 * calling offline_pages().
519 void __remove_pages(unsigned long pfn
, unsigned long nr_pages
,
520 struct vmem_altmap
*altmap
)
522 const unsigned long end_pfn
= pfn
+ nr_pages
;
523 unsigned long cur_nr_pages
;
524 unsigned long map_offset
= 0;
526 map_offset
= vmem_altmap_offset(altmap
);
528 if (check_pfn_span(pfn
, nr_pages
, "remove"))
531 for (; pfn
< end_pfn
; pfn
+= cur_nr_pages
) {
533 /* Select all remaining pages up to the next section boundary */
534 cur_nr_pages
= min(end_pfn
- pfn
,
535 SECTION_ALIGN_UP(pfn
+ 1) - pfn
);
536 __remove_section(pfn
, cur_nr_pages
, map_offset
, altmap
);
541 int set_online_page_callback(online_page_callback_t callback
)
546 mutex_lock(&online_page_callback_lock
);
548 if (online_page_callback
== generic_online_page
) {
549 online_page_callback
= callback
;
553 mutex_unlock(&online_page_callback_lock
);
558 EXPORT_SYMBOL_GPL(set_online_page_callback
);
560 int restore_online_page_callback(online_page_callback_t callback
)
565 mutex_lock(&online_page_callback_lock
);
567 if (online_page_callback
== callback
) {
568 online_page_callback
= generic_online_page
;
572 mutex_unlock(&online_page_callback_lock
);
577 EXPORT_SYMBOL_GPL(restore_online_page_callback
);
579 void generic_online_page(struct page
*page
, unsigned int order
)
582 * Freeing the page with debug_pagealloc enabled will try to unmap it,
583 * so we should map it first. This is better than introducing a special
584 * case in page freeing fast path.
586 debug_pagealloc_map_pages(page
, 1 << order
);
587 __free_pages_core(page
, order
);
588 totalram_pages_add(1UL << order
);
589 #ifdef CONFIG_HIGHMEM
590 if (PageHighMem(page
))
591 totalhigh_pages_add(1UL << order
);
594 EXPORT_SYMBOL_GPL(generic_online_page
);
596 static void online_pages_range(unsigned long start_pfn
, unsigned long nr_pages
)
598 const unsigned long end_pfn
= start_pfn
+ nr_pages
;
602 * Online the pages in MAX_ORDER - 1 aligned chunks. The callback might
603 * decide to not expose all pages to the buddy (e.g., expose them
604 * later). We account all pages as being online and belonging to this
606 * When using memmap_on_memory, the range might not be aligned to
607 * MAX_ORDER_NR_PAGES - 1, but pageblock aligned. __ffs() will detect
608 * this and the first chunk to online will be pageblock_nr_pages.
610 for (pfn
= start_pfn
; pfn
< end_pfn
;) {
611 int order
= min(MAX_ORDER
- 1UL, __ffs(pfn
));
613 (*online_page_callback
)(pfn_to_page(pfn
), order
);
614 pfn
+= (1UL << order
);
617 /* mark all involved sections as online */
618 online_mem_sections(start_pfn
, end_pfn
);
621 /* check which state of node_states will be changed when online memory */
622 static void node_states_check_changes_online(unsigned long nr_pages
,
623 struct zone
*zone
, struct memory_notify
*arg
)
625 int nid
= zone_to_nid(zone
);
627 arg
->status_change_nid
= NUMA_NO_NODE
;
628 arg
->status_change_nid_normal
= NUMA_NO_NODE
;
629 arg
->status_change_nid_high
= NUMA_NO_NODE
;
631 if (!node_state(nid
, N_MEMORY
))
632 arg
->status_change_nid
= nid
;
633 if (zone_idx(zone
) <= ZONE_NORMAL
&& !node_state(nid
, N_NORMAL_MEMORY
))
634 arg
->status_change_nid_normal
= nid
;
635 #ifdef CONFIG_HIGHMEM
636 if (zone_idx(zone
) <= ZONE_HIGHMEM
&& !node_state(nid
, N_HIGH_MEMORY
))
637 arg
->status_change_nid_high
= nid
;
641 static void node_states_set_node(int node
, struct memory_notify
*arg
)
643 if (arg
->status_change_nid_normal
>= 0)
644 node_set_state(node
, N_NORMAL_MEMORY
);
646 if (arg
->status_change_nid_high
>= 0)
647 node_set_state(node
, N_HIGH_MEMORY
);
649 if (arg
->status_change_nid
>= 0)
650 node_set_state(node
, N_MEMORY
);
653 static void __meminit
resize_zone_range(struct zone
*zone
, unsigned long start_pfn
,
654 unsigned long nr_pages
)
656 unsigned long old_end_pfn
= zone_end_pfn(zone
);
658 if (zone_is_empty(zone
) || start_pfn
< zone
->zone_start_pfn
)
659 zone
->zone_start_pfn
= start_pfn
;
661 zone
->spanned_pages
= max(start_pfn
+ nr_pages
, old_end_pfn
) - zone
->zone_start_pfn
;
664 static void __meminit
resize_pgdat_range(struct pglist_data
*pgdat
, unsigned long start_pfn
,
665 unsigned long nr_pages
)
667 unsigned long old_end_pfn
= pgdat_end_pfn(pgdat
);
669 if (!pgdat
->node_spanned_pages
|| start_pfn
< pgdat
->node_start_pfn
)
670 pgdat
->node_start_pfn
= start_pfn
;
672 pgdat
->node_spanned_pages
= max(start_pfn
+ nr_pages
, old_end_pfn
) - pgdat
->node_start_pfn
;
676 static void section_taint_zone_device(unsigned long pfn
)
678 struct mem_section
*ms
= __pfn_to_section(pfn
);
680 ms
->section_mem_map
|= SECTION_TAINT_ZONE_DEVICE
;
684 * Associate the pfn range with the given zone, initializing the memmaps
685 * and resizing the pgdat/zone data to span the added pages. After this
686 * call, all affected pages are PG_reserved.
688 * All aligned pageblocks are initialized to the specified migratetype
689 * (usually MIGRATE_MOVABLE). Besides setting the migratetype, no related
690 * zone stats (e.g., nr_isolate_pageblock) are touched.
692 void __ref
move_pfn_range_to_zone(struct zone
*zone
, unsigned long start_pfn
,
693 unsigned long nr_pages
,
694 struct vmem_altmap
*altmap
, int migratetype
)
696 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
697 int nid
= pgdat
->node_id
;
699 clear_zone_contiguous(zone
);
701 if (zone_is_empty(zone
))
702 init_currently_empty_zone(zone
, start_pfn
, nr_pages
);
703 resize_zone_range(zone
, start_pfn
, nr_pages
);
704 resize_pgdat_range(pgdat
, start_pfn
, nr_pages
);
707 * Subsection population requires care in pfn_to_online_page().
708 * Set the taint to enable the slow path detection of
709 * ZONE_DEVICE pages in an otherwise ZONE_{NORMAL,MOVABLE}
712 if (zone_is_zone_device(zone
)) {
713 if (!IS_ALIGNED(start_pfn
, PAGES_PER_SECTION
))
714 section_taint_zone_device(start_pfn
);
715 if (!IS_ALIGNED(start_pfn
+ nr_pages
, PAGES_PER_SECTION
))
716 section_taint_zone_device(start_pfn
+ nr_pages
);
720 * TODO now we have a visible range of pages which are not associated
721 * with their zone properly. Not nice but set_pfnblock_flags_mask
722 * expects the zone spans the pfn range. All the pages in the range
723 * are reserved so nobody should be touching them so we should be safe
725 memmap_init_range(nr_pages
, nid
, zone_idx(zone
), start_pfn
, 0,
726 MEMINIT_HOTPLUG
, altmap
, migratetype
);
728 set_zone_contiguous(zone
);
731 struct auto_movable_stats
{
732 unsigned long kernel_early_pages
;
733 unsigned long movable_pages
;
736 static void auto_movable_stats_account_zone(struct auto_movable_stats
*stats
,
739 if (zone_idx(zone
) == ZONE_MOVABLE
) {
740 stats
->movable_pages
+= zone
->present_pages
;
742 stats
->kernel_early_pages
+= zone
->present_early_pages
;
745 * CMA pages (never on hotplugged memory) behave like
748 stats
->movable_pages
+= zone
->cma_pages
;
749 stats
->kernel_early_pages
-= zone
->cma_pages
;
750 #endif /* CONFIG_CMA */
753 struct auto_movable_group_stats
{
754 unsigned long movable_pages
;
755 unsigned long req_kernel_early_pages
;
758 static int auto_movable_stats_account_group(struct memory_group
*group
,
761 const int ratio
= READ_ONCE(auto_movable_ratio
);
762 struct auto_movable_group_stats
*stats
= arg
;
766 * We don't support modifying the config while the auto-movable online
767 * policy is already enabled. Just avoid the division by zero below.
773 * Calculate how many early kernel pages this group requires to
774 * satisfy the configured zone ratio.
776 pages
= group
->present_movable_pages
* 100 / ratio
;
777 pages
-= group
->present_kernel_pages
;
780 stats
->req_kernel_early_pages
+= pages
;
781 stats
->movable_pages
+= group
->present_movable_pages
;
785 static bool auto_movable_can_online_movable(int nid
, struct memory_group
*group
,
786 unsigned long nr_pages
)
788 unsigned long kernel_early_pages
, movable_pages
;
789 struct auto_movable_group_stats group_stats
= {};
790 struct auto_movable_stats stats
= {};
791 pg_data_t
*pgdat
= NODE_DATA(nid
);
795 /* Walk all relevant zones and collect MOVABLE vs. KERNEL stats. */
796 if (nid
== NUMA_NO_NODE
) {
797 /* TODO: cache values */
798 for_each_populated_zone(zone
)
799 auto_movable_stats_account_zone(&stats
, zone
);
801 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
802 zone
= pgdat
->node_zones
+ i
;
803 if (populated_zone(zone
))
804 auto_movable_stats_account_zone(&stats
, zone
);
808 kernel_early_pages
= stats
.kernel_early_pages
;
809 movable_pages
= stats
.movable_pages
;
812 * Kernel memory inside dynamic memory group allows for more MOVABLE
813 * memory within the same group. Remove the effect of all but the
814 * current group from the stats.
816 walk_dynamic_memory_groups(nid
, auto_movable_stats_account_group
,
817 group
, &group_stats
);
818 if (kernel_early_pages
<= group_stats
.req_kernel_early_pages
)
820 kernel_early_pages
-= group_stats
.req_kernel_early_pages
;
821 movable_pages
-= group_stats
.movable_pages
;
823 if (group
&& group
->is_dynamic
)
824 kernel_early_pages
+= group
->present_kernel_pages
;
827 * Test if we could online the given number of pages to ZONE_MOVABLE
828 * and still stay in the configured ratio.
830 movable_pages
+= nr_pages
;
831 return movable_pages
<= (auto_movable_ratio
* kernel_early_pages
) / 100;
835 * Returns a default kernel memory zone for the given pfn range.
836 * If no kernel zone covers this pfn range it will automatically go
837 * to the ZONE_NORMAL.
839 static struct zone
*default_kernel_zone_for_pfn(int nid
, unsigned long start_pfn
,
840 unsigned long nr_pages
)
842 struct pglist_data
*pgdat
= NODE_DATA(nid
);
845 for (zid
= 0; zid
<= ZONE_NORMAL
; zid
++) {
846 struct zone
*zone
= &pgdat
->node_zones
[zid
];
848 if (zone_intersects(zone
, start_pfn
, nr_pages
))
852 return &pgdat
->node_zones
[ZONE_NORMAL
];
856 * Determine to which zone to online memory dynamically based on user
857 * configuration and system stats. We care about the following ratio:
861 * Whereby MOVABLE is memory in ZONE_MOVABLE and KERNEL is memory in
862 * one of the kernel zones. CMA pages inside one of the kernel zones really
863 * behaves like ZONE_MOVABLE, so we treat them accordingly.
865 * We don't allow for hotplugged memory in a KERNEL zone to increase the
866 * amount of MOVABLE memory we can have, so we end up with:
868 * MOVABLE : KERNEL_EARLY
870 * Whereby KERNEL_EARLY is memory in one of the kernel zones, available sinze
871 * boot. We base our calculation on KERNEL_EARLY internally, because:
873 * a) Hotplugged memory in one of the kernel zones can sometimes still get
874 * hotunplugged, especially when hot(un)plugging individual memory blocks.
875 * There is no coordination across memory devices, therefore "automatic"
876 * hotunplugging, as implemented in hypervisors, could result in zone
878 * b) Early/boot memory in one of the kernel zones can usually not get
879 * hotunplugged again (e.g., no firmware interface to unplug, fragmented
880 * with unmovable allocations). While there are corner cases where it might
881 * still work, it is barely relevant in practice.
883 * Exceptions are dynamic memory groups, which allow for more MOVABLE
884 * memory within the same memory group -- because in that case, there is
885 * coordination within the single memory device managed by a single driver.
887 * We rely on "present pages" instead of "managed pages", as the latter is
888 * highly unreliable and dynamic in virtualized environments, and does not
889 * consider boot time allocations. For example, memory ballooning adjusts the
890 * managed pages when inflating/deflating the balloon, and balloon compaction
891 * can even migrate inflated pages between zones.
893 * Using "present pages" is better but some things to keep in mind are:
895 * a) Some memblock allocations, such as for the crashkernel area, are
896 * effectively unused by the kernel, yet they account to "present pages".
897 * Fortunately, these allocations are comparatively small in relevant setups
898 * (e.g., fraction of system memory).
899 * b) Some hotplugged memory blocks in virtualized environments, esecially
900 * hotplugged by virtio-mem, look like they are completely present, however,
901 * only parts of the memory block are actually currently usable.
902 * "present pages" is an upper limit that can get reached at runtime. As
903 * we base our calculations on KERNEL_EARLY, this is not an issue.
905 static struct zone
*auto_movable_zone_for_pfn(int nid
,
906 struct memory_group
*group
,
908 unsigned long nr_pages
)
910 unsigned long online_pages
= 0, max_pages
, end_pfn
;
913 if (!auto_movable_ratio
)
916 if (group
&& !group
->is_dynamic
) {
917 max_pages
= group
->s
.max_pages
;
918 online_pages
= group
->present_movable_pages
;
920 /* If anything is !MOVABLE online the rest !MOVABLE. */
921 if (group
->present_kernel_pages
)
923 } else if (!group
|| group
->d
.unit_pages
== nr_pages
) {
924 max_pages
= nr_pages
;
926 max_pages
= group
->d
.unit_pages
;
928 * Take a look at all online sections in the current unit.
929 * We can safely assume that all pages within a section belong
930 * to the same zone, because dynamic memory groups only deal
931 * with hotplugged memory.
933 pfn
= ALIGN_DOWN(pfn
, group
->d
.unit_pages
);
934 end_pfn
= pfn
+ group
->d
.unit_pages
;
935 for (; pfn
< end_pfn
; pfn
+= PAGES_PER_SECTION
) {
936 page
= pfn_to_online_page(pfn
);
939 /* If anything is !MOVABLE online the rest !MOVABLE. */
940 if (page_zonenum(page
) != ZONE_MOVABLE
)
942 online_pages
+= PAGES_PER_SECTION
;
947 * Online MOVABLE if we could *currently* online all remaining parts
948 * MOVABLE. We expect to (add+) online them immediately next, so if
949 * nobody interferes, all will be MOVABLE if possible.
951 nr_pages
= max_pages
- online_pages
;
952 if (!auto_movable_can_online_movable(NUMA_NO_NODE
, group
, nr_pages
))
956 if (auto_movable_numa_aware
&&
957 !auto_movable_can_online_movable(nid
, group
, nr_pages
))
959 #endif /* CONFIG_NUMA */
961 return &NODE_DATA(nid
)->node_zones
[ZONE_MOVABLE
];
963 return default_kernel_zone_for_pfn(nid
, pfn
, nr_pages
);
966 static inline struct zone
*default_zone_for_pfn(int nid
, unsigned long start_pfn
,
967 unsigned long nr_pages
)
969 struct zone
*kernel_zone
= default_kernel_zone_for_pfn(nid
, start_pfn
,
971 struct zone
*movable_zone
= &NODE_DATA(nid
)->node_zones
[ZONE_MOVABLE
];
972 bool in_kernel
= zone_intersects(kernel_zone
, start_pfn
, nr_pages
);
973 bool in_movable
= zone_intersects(movable_zone
, start_pfn
, nr_pages
);
976 * We inherit the existing zone in a simple case where zones do not
977 * overlap in the given range
979 if (in_kernel
^ in_movable
)
980 return (in_kernel
) ? kernel_zone
: movable_zone
;
983 * If the range doesn't belong to any zone or two zones overlap in the
984 * given range then we use movable zone only if movable_node is
985 * enabled because we always online to a kernel zone by default.
987 return movable_node_enabled
? movable_zone
: kernel_zone
;
990 struct zone
*zone_for_pfn_range(int online_type
, int nid
,
991 struct memory_group
*group
, unsigned long start_pfn
,
992 unsigned long nr_pages
)
994 if (online_type
== MMOP_ONLINE_KERNEL
)
995 return default_kernel_zone_for_pfn(nid
, start_pfn
, nr_pages
);
997 if (online_type
== MMOP_ONLINE_MOVABLE
)
998 return &NODE_DATA(nid
)->node_zones
[ZONE_MOVABLE
];
1000 if (online_policy
== ONLINE_POLICY_AUTO_MOVABLE
)
1001 return auto_movable_zone_for_pfn(nid
, group
, start_pfn
, nr_pages
);
1003 return default_zone_for_pfn(nid
, start_pfn
, nr_pages
);
1007 * This function should only be called by memory_block_{online,offline},
1008 * and {online,offline}_pages.
1010 void adjust_present_page_count(struct page
*page
, struct memory_group
*group
,
1013 struct zone
*zone
= page_zone(page
);
1014 const bool movable
= zone_idx(zone
) == ZONE_MOVABLE
;
1017 * We only support onlining/offlining/adding/removing of complete
1018 * memory blocks; therefore, either all is either early or hotplugged.
1020 if (early_section(__pfn_to_section(page_to_pfn(page
))))
1021 zone
->present_early_pages
+= nr_pages
;
1022 zone
->present_pages
+= nr_pages
;
1023 zone
->zone_pgdat
->node_present_pages
+= nr_pages
;
1025 if (group
&& movable
)
1026 group
->present_movable_pages
+= nr_pages
;
1027 else if (group
&& !movable
)
1028 group
->present_kernel_pages
+= nr_pages
;
1031 int mhp_init_memmap_on_memory(unsigned long pfn
, unsigned long nr_pages
,
1034 unsigned long end_pfn
= pfn
+ nr_pages
;
1037 ret
= kasan_add_zero_shadow(__va(PFN_PHYS(pfn
)), PFN_PHYS(nr_pages
));
1041 move_pfn_range_to_zone(zone
, pfn
, nr_pages
, NULL
, MIGRATE_UNMOVABLE
);
1044 * It might be that the vmemmap_pages fully span sections. If that is
1045 * the case, mark those sections online here as otherwise they will be
1048 if (nr_pages
>= PAGES_PER_SECTION
)
1049 online_mem_sections(pfn
, ALIGN_DOWN(end_pfn
, PAGES_PER_SECTION
));
1054 void mhp_deinit_memmap_on_memory(unsigned long pfn
, unsigned long nr_pages
)
1056 unsigned long end_pfn
= pfn
+ nr_pages
;
1059 * It might be that the vmemmap_pages fully span sections. If that is
1060 * the case, mark those sections offline here as otherwise they will be
1063 if (nr_pages
>= PAGES_PER_SECTION
)
1064 offline_mem_sections(pfn
, ALIGN_DOWN(end_pfn
, PAGES_PER_SECTION
));
1067 * The pages associated with this vmemmap have been offlined, so
1068 * we can reset its state here.
1070 remove_pfn_range_from_zone(page_zone(pfn_to_page(pfn
)), pfn
, nr_pages
);
1071 kasan_remove_zero_shadow(__va(PFN_PHYS(pfn
)), PFN_PHYS(nr_pages
));
1074 int __ref
online_pages(unsigned long pfn
, unsigned long nr_pages
,
1075 struct zone
*zone
, struct memory_group
*group
)
1077 unsigned long flags
;
1078 int need_zonelists_rebuild
= 0;
1079 const int nid
= zone_to_nid(zone
);
1081 struct memory_notify arg
;
1084 * {on,off}lining is constrained to full memory sections (or more
1085 * precisely to memory blocks from the user space POV).
1086 * memmap_on_memory is an exception because it reserves initial part
1087 * of the physical memory space for vmemmaps. That space is pageblock
1090 if (WARN_ON_ONCE(!nr_pages
||
1091 !IS_ALIGNED(pfn
, pageblock_nr_pages
) ||
1092 !IS_ALIGNED(pfn
+ nr_pages
, PAGES_PER_SECTION
)))
1095 mem_hotplug_begin();
1097 /* associate pfn range with the zone */
1098 move_pfn_range_to_zone(zone
, pfn
, nr_pages
, NULL
, MIGRATE_ISOLATE
);
1100 arg
.start_pfn
= pfn
;
1101 arg
.nr_pages
= nr_pages
;
1102 node_states_check_changes_online(nr_pages
, zone
, &arg
);
1104 ret
= memory_notify(MEM_GOING_ONLINE
, &arg
);
1105 ret
= notifier_to_errno(ret
);
1107 goto failed_addition
;
1110 * Fixup the number of isolated pageblocks before marking the sections
1111 * onlining, such that undo_isolate_page_range() works correctly.
1113 spin_lock_irqsave(&zone
->lock
, flags
);
1114 zone
->nr_isolate_pageblock
+= nr_pages
/ pageblock_nr_pages
;
1115 spin_unlock_irqrestore(&zone
->lock
, flags
);
1118 * If this zone is not populated, then it is not in zonelist.
1119 * This means the page allocator ignores this zone.
1120 * So, zonelist must be updated after online.
1122 if (!populated_zone(zone
)) {
1123 need_zonelists_rebuild
= 1;
1124 setup_zone_pageset(zone
);
1127 online_pages_range(pfn
, nr_pages
);
1128 adjust_present_page_count(pfn_to_page(pfn
), group
, nr_pages
);
1130 node_states_set_node(nid
, &arg
);
1131 if (need_zonelists_rebuild
)
1132 build_all_zonelists(NULL
);
1134 /* Basic onlining is complete, allow allocation of onlined pages. */
1135 undo_isolate_page_range(pfn
, pfn
+ nr_pages
, MIGRATE_MOVABLE
);
1138 * Freshly onlined pages aren't shuffled (e.g., all pages are placed to
1139 * the tail of the freelist when undoing isolation). Shuffle the whole
1140 * zone to make sure the just onlined pages are properly distributed
1141 * across the whole freelist - to create an initial shuffle.
1145 /* reinitialise watermarks and update pcp limits */
1146 init_per_zone_wmark_min();
1151 writeback_set_ratelimit();
1153 memory_notify(MEM_ONLINE
, &arg
);
1158 pr_debug("online_pages [mem %#010llx-%#010llx] failed\n",
1159 (unsigned long long) pfn
<< PAGE_SHIFT
,
1160 (((unsigned long long) pfn
+ nr_pages
) << PAGE_SHIFT
) - 1);
1161 memory_notify(MEM_CANCEL_ONLINE
, &arg
);
1162 remove_pfn_range_from_zone(zone
, pfn
, nr_pages
);
1166 #endif /* CONFIG_MEMORY_HOTPLUG_SPARSE */
1168 static void reset_node_present_pages(pg_data_t
*pgdat
)
1172 for (z
= pgdat
->node_zones
; z
< pgdat
->node_zones
+ MAX_NR_ZONES
; z
++)
1173 z
->present_pages
= 0;
1175 pgdat
->node_present_pages
= 0;
1178 /* we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG */
1179 static pg_data_t __ref
*hotadd_new_pgdat(int nid
)
1181 struct pglist_data
*pgdat
;
1183 pgdat
= NODE_DATA(nid
);
1185 pgdat
= arch_alloc_nodedata(nid
);
1189 pgdat
->per_cpu_nodestats
=
1190 alloc_percpu(struct per_cpu_nodestat
);
1191 arch_refresh_nodedata(nid
, pgdat
);
1195 * Reset the nr_zones, order and highest_zoneidx before reuse.
1196 * Note that kswapd will init kswapd_highest_zoneidx properly
1197 * when it starts in the near future.
1199 pgdat
->nr_zones
= 0;
1200 pgdat
->kswapd_order
= 0;
1201 pgdat
->kswapd_highest_zoneidx
= 0;
1202 for_each_online_cpu(cpu
) {
1203 struct per_cpu_nodestat
*p
;
1205 p
= per_cpu_ptr(pgdat
->per_cpu_nodestats
, cpu
);
1206 memset(p
, 0, sizeof(*p
));
1210 /* we can use NODE_DATA(nid) from here */
1211 pgdat
->node_id
= nid
;
1212 pgdat
->node_start_pfn
= 0;
1214 /* init node's zones as empty zones, we don't have any present pages.*/
1215 free_area_init_core_hotplug(nid
);
1218 * The node we allocated has no zone fallback lists. For avoiding
1219 * to access not-initialized zonelist, build here.
1221 build_all_zonelists(pgdat
);
1224 * When memory is hot-added, all the memory is in offline state. So
1225 * clear all zones' present_pages because they will be updated in
1226 * online_pages() and offline_pages().
1228 reset_node_managed_pages(pgdat
);
1229 reset_node_present_pages(pgdat
);
1234 static void rollback_node_hotadd(int nid
)
1236 pg_data_t
*pgdat
= NODE_DATA(nid
);
1238 arch_refresh_nodedata(nid
, NULL
);
1239 free_percpu(pgdat
->per_cpu_nodestats
);
1240 arch_free_nodedata(pgdat
);
1245 * __try_online_node - online a node if offlined
1247 * @set_node_online: Whether we want to online the node
1248 * called by cpu_up() to online a node without onlined memory.
1251 * 1 -> a new node has been allocated
1252 * 0 -> the node is already online
1253 * -ENOMEM -> the node could not be allocated
1255 static int __try_online_node(int nid
, bool set_node_online
)
1260 if (node_online(nid
))
1263 pgdat
= hotadd_new_pgdat(nid
);
1265 pr_err("Cannot online node %d due to NULL pgdat\n", nid
);
1270 if (set_node_online
) {
1271 node_set_online(nid
);
1272 ret
= register_one_node(nid
);
1280 * Users of this function always want to online/register the node
1282 int try_online_node(int nid
)
1286 mem_hotplug_begin();
1287 ret
= __try_online_node(nid
, true);
1292 static int check_hotplug_memory_range(u64 start
, u64 size
)
1294 /* memory range must be block size aligned */
1295 if (!size
|| !IS_ALIGNED(start
, memory_block_size_bytes()) ||
1296 !IS_ALIGNED(size
, memory_block_size_bytes())) {
1297 pr_err("Block size [%#lx] unaligned hotplug range: start %#llx, size %#llx",
1298 memory_block_size_bytes(), start
, size
);
1305 static int online_memory_block(struct memory_block
*mem
, void *arg
)
1307 mem
->online_type
= mhp_default_online_type
;
1308 return device_online(&mem
->dev
);
1311 bool mhp_supports_memmap_on_memory(unsigned long size
)
1313 unsigned long nr_vmemmap_pages
= size
/ PAGE_SIZE
;
1314 unsigned long vmemmap_size
= nr_vmemmap_pages
* sizeof(struct page
);
1315 unsigned long remaining_size
= size
- vmemmap_size
;
1318 * Besides having arch support and the feature enabled at runtime, we
1319 * need a few more assumptions to hold true:
1321 * a) We span a single memory block: memory onlining/offlinin;g happens
1322 * in memory block granularity. We don't want the vmemmap of online
1323 * memory blocks to reside on offline memory blocks. In the future,
1324 * we might want to support variable-sized memory blocks to make the
1325 * feature more versatile.
1327 * b) The vmemmap pages span complete PMDs: We don't want vmemmap code
1328 * to populate memory from the altmap for unrelated parts (i.e.,
1329 * other memory blocks)
1331 * c) The vmemmap pages (and thereby the pages that will be exposed to
1332 * the buddy) have to cover full pageblocks: memory onlining/offlining
1333 * code requires applicable ranges to be page-aligned, for example, to
1334 * set the migratetypes properly.
1336 * TODO: Although we have a check here to make sure that vmemmap pages
1337 * fully populate a PMD, it is not the right place to check for
1338 * this. A much better solution involves improving vmemmap code
1339 * to fallback to base pages when trying to populate vmemmap using
1340 * altmap as an alternative source of memory, and we do not exactly
1341 * populate a single PMD.
1343 return memmap_on_memory
&&
1344 !hugetlb_free_vmemmap_enabled
&&
1345 IS_ENABLED(CONFIG_MHP_MEMMAP_ON_MEMORY
) &&
1346 size
== memory_block_size_bytes() &&
1347 IS_ALIGNED(vmemmap_size
, PMD_SIZE
) &&
1348 IS_ALIGNED(remaining_size
, (pageblock_nr_pages
<< PAGE_SHIFT
));
1352 * NOTE: The caller must call lock_device_hotplug() to serialize hotplug
1353 * and online/offline operations (triggered e.g. by sysfs).
1355 * we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG
1357 int __ref
add_memory_resource(int nid
, struct resource
*res
, mhp_t mhp_flags
)
1359 struct mhp_params params
= { .pgprot
= pgprot_mhp(PAGE_KERNEL
) };
1360 struct vmem_altmap mhp_altmap
= {};
1361 struct memory_group
*group
= NULL
;
1363 bool new_node
= false;
1367 size
= resource_size(res
);
1369 ret
= check_hotplug_memory_range(start
, size
);
1373 if (mhp_flags
& MHP_NID_IS_MGID
) {
1374 group
= memory_group_find_by_id(nid
);
1380 if (!node_possible(nid
)) {
1381 WARN(1, "node %d was absent from the node_possible_map\n", nid
);
1385 mem_hotplug_begin();
1387 if (IS_ENABLED(CONFIG_ARCH_KEEP_MEMBLOCK
))
1388 memblock_add_node(start
, size
, nid
);
1390 ret
= __try_online_node(nid
, false);
1396 * Self hosted memmap array
1398 if (mhp_flags
& MHP_MEMMAP_ON_MEMORY
) {
1399 if (!mhp_supports_memmap_on_memory(size
)) {
1403 mhp_altmap
.free
= PHYS_PFN(size
);
1404 mhp_altmap
.base_pfn
= PHYS_PFN(start
);
1405 params
.altmap
= &mhp_altmap
;
1408 /* call arch's memory hotadd */
1409 ret
= arch_add_memory(nid
, start
, size
, ¶ms
);
1413 /* create memory block devices after memory was added */
1414 ret
= create_memory_block_devices(start
, size
, mhp_altmap
.alloc
,
1417 arch_remove_memory(start
, size
, NULL
);
1422 /* If sysfs file of new node can't be created, cpu on the node
1423 * can't be hot-added. There is no rollback way now.
1424 * So, check by BUG_ON() to catch it reluctantly..
1425 * We online node here. We can't roll back from here.
1427 node_set_online(nid
);
1428 ret
= __register_one_node(nid
);
1432 /* link memory sections under this node.*/
1433 link_mem_sections(nid
, PFN_DOWN(start
), PFN_UP(start
+ size
- 1),
1436 /* create new memmap entry */
1437 if (!strcmp(res
->name
, "System RAM"))
1438 firmware_map_add_hotplug(start
, start
+ size
, "System RAM");
1440 /* device_online() will take the lock when calling online_pages() */
1444 * In case we're allowed to merge the resource, flag it and trigger
1445 * merging now that adding succeeded.
1447 if (mhp_flags
& MHP_MERGE_RESOURCE
)
1448 merge_system_ram_resource(res
);
1450 /* online pages if requested */
1451 if (mhp_default_online_type
!= MMOP_OFFLINE
)
1452 walk_memory_blocks(start
, size
, NULL
, online_memory_block
);
1456 /* rollback pgdat allocation and others */
1458 rollback_node_hotadd(nid
);
1459 if (IS_ENABLED(CONFIG_ARCH_KEEP_MEMBLOCK
))
1460 memblock_remove(start
, size
);
1465 /* requires device_hotplug_lock, see add_memory_resource() */
1466 int __ref
__add_memory(int nid
, u64 start
, u64 size
, mhp_t mhp_flags
)
1468 struct resource
*res
;
1471 res
= register_memory_resource(start
, size
, "System RAM");
1473 return PTR_ERR(res
);
1475 ret
= add_memory_resource(nid
, res
, mhp_flags
);
1477 release_memory_resource(res
);
1481 int add_memory(int nid
, u64 start
, u64 size
, mhp_t mhp_flags
)
1485 lock_device_hotplug();
1486 rc
= __add_memory(nid
, start
, size
, mhp_flags
);
1487 unlock_device_hotplug();
1491 EXPORT_SYMBOL_GPL(add_memory
);
1494 * Add special, driver-managed memory to the system as system RAM. Such
1495 * memory is not exposed via the raw firmware-provided memmap as system
1496 * RAM, instead, it is detected and added by a driver - during cold boot,
1497 * after a reboot, and after kexec.
1499 * Reasons why this memory should not be used for the initial memmap of a
1500 * kexec kernel or for placing kexec images:
1501 * - The booting kernel is in charge of determining how this memory will be
1502 * used (e.g., use persistent memory as system RAM)
1503 * - Coordination with a hypervisor is required before this memory
1504 * can be used (e.g., inaccessible parts).
1506 * For this memory, no entries in /sys/firmware/memmap ("raw firmware-provided
1507 * memory map") are created. Also, the created memory resource is flagged
1508 * with IORESOURCE_SYSRAM_DRIVER_MANAGED, so in-kernel users can special-case
1509 * this memory as well (esp., not place kexec images onto it).
1511 * The resource_name (visible via /proc/iomem) has to have the format
1512 * "System RAM ($DRIVER)".
1514 int add_memory_driver_managed(int nid
, u64 start
, u64 size
,
1515 const char *resource_name
, mhp_t mhp_flags
)
1517 struct resource
*res
;
1520 if (!resource_name
||
1521 strstr(resource_name
, "System RAM (") != resource_name
||
1522 resource_name
[strlen(resource_name
) - 1] != ')')
1525 lock_device_hotplug();
1527 res
= register_memory_resource(start
, size
, resource_name
);
1533 rc
= add_memory_resource(nid
, res
, mhp_flags
);
1535 release_memory_resource(res
);
1538 unlock_device_hotplug();
1541 EXPORT_SYMBOL_GPL(add_memory_driver_managed
);
1544 * Platforms should define arch_get_mappable_range() that provides
1545 * maximum possible addressable physical memory range for which the
1546 * linear mapping could be created. The platform returned address
1547 * range must adhere to these following semantics.
1549 * - range.start <= range.end
1550 * - Range includes both end points [range.start..range.end]
1552 * There is also a fallback definition provided here, allowing the
1553 * entire possible physical address range in case any platform does
1554 * not define arch_get_mappable_range().
1556 struct range __weak
arch_get_mappable_range(void)
1558 struct range mhp_range
= {
1565 struct range
mhp_get_pluggable_range(bool need_mapping
)
1567 const u64 max_phys
= (1ULL << MAX_PHYSMEM_BITS
) - 1;
1568 struct range mhp_range
;
1571 mhp_range
= arch_get_mappable_range();
1572 if (mhp_range
.start
> max_phys
) {
1573 mhp_range
.start
= 0;
1576 mhp_range
.end
= min_t(u64
, mhp_range
.end
, max_phys
);
1578 mhp_range
.start
= 0;
1579 mhp_range
.end
= max_phys
;
1583 EXPORT_SYMBOL_GPL(mhp_get_pluggable_range
);
1585 bool mhp_range_allowed(u64 start
, u64 size
, bool need_mapping
)
1587 struct range mhp_range
= mhp_get_pluggable_range(need_mapping
);
1588 u64 end
= start
+ size
;
1590 if (start
< end
&& start
>= mhp_range
.start
&& (end
- 1) <= mhp_range
.end
)
1593 pr_warn("Hotplug memory [%#llx-%#llx] exceeds maximum addressable range [%#llx-%#llx]\n",
1594 start
, end
, mhp_range
.start
, mhp_range
.end
);
1598 #ifdef CONFIG_MEMORY_HOTREMOVE
1600 * Confirm all pages in a range [start, end) belong to the same zone (skipping
1601 * memory holes). When true, return the zone.
1603 struct zone
*test_pages_in_a_zone(unsigned long start_pfn
,
1604 unsigned long end_pfn
)
1606 unsigned long pfn
, sec_end_pfn
;
1607 struct zone
*zone
= NULL
;
1610 for (pfn
= start_pfn
, sec_end_pfn
= SECTION_ALIGN_UP(start_pfn
+ 1);
1612 pfn
= sec_end_pfn
, sec_end_pfn
+= PAGES_PER_SECTION
) {
1613 /* Make sure the memory section is present first */
1614 if (!present_section_nr(pfn_to_section_nr(pfn
)))
1616 for (; pfn
< sec_end_pfn
&& pfn
< end_pfn
;
1617 pfn
+= MAX_ORDER_NR_PAGES
) {
1618 /* Check if we got outside of the zone */
1619 if (zone
&& !zone_spans_pfn(zone
, pfn
))
1621 page
= pfn_to_page(pfn
);
1622 if (zone
&& page_zone(page
) != zone
)
1624 zone
= page_zone(page
);
1632 * Scan pfn range [start,end) to find movable/migratable pages (LRU pages,
1633 * non-lru movable pages and hugepages). Will skip over most unmovable
1634 * pages (esp., pages that can be skipped when offlining), but bail out on
1635 * definitely unmovable pages.
1638 * 0 in case a movable page is found and movable_pfn was updated.
1639 * -ENOENT in case no movable page was found.
1640 * -EBUSY in case a definitely unmovable page was found.
1642 static int scan_movable_pages(unsigned long start
, unsigned long end
,
1643 unsigned long *movable_pfn
)
1647 for (pfn
= start
; pfn
< end
; pfn
++) {
1648 struct page
*page
, *head
;
1651 if (!pfn_valid(pfn
))
1653 page
= pfn_to_page(pfn
);
1656 if (__PageMovable(page
))
1660 * PageOffline() pages that are not marked __PageMovable() and
1661 * have a reference count > 0 (after MEM_GOING_OFFLINE) are
1662 * definitely unmovable. If their reference count would be 0,
1663 * they could at least be skipped when offlining memory.
1665 if (PageOffline(page
) && page_count(page
))
1668 if (!PageHuge(page
))
1670 head
= compound_head(page
);
1672 * This test is racy as we hold no reference or lock. The
1673 * hugetlb page could have been free'ed and head is no longer
1674 * a hugetlb page before the following check. In such unlikely
1675 * cases false positives and negatives are possible. Calling
1676 * code must deal with these scenarios.
1678 if (HPageMigratable(head
))
1680 skip
= compound_nr(head
) - (page
- head
);
1690 do_migrate_range(unsigned long start_pfn
, unsigned long end_pfn
)
1693 struct page
*page
, *head
;
1696 static DEFINE_RATELIMIT_STATE(migrate_rs
, DEFAULT_RATELIMIT_INTERVAL
,
1697 DEFAULT_RATELIMIT_BURST
);
1699 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
1700 if (!pfn_valid(pfn
))
1702 page
= pfn_to_page(pfn
);
1703 head
= compound_head(page
);
1705 if (PageHuge(page
)) {
1706 pfn
= page_to_pfn(head
) + compound_nr(head
) - 1;
1707 isolate_huge_page(head
, &source
);
1709 } else if (PageTransHuge(page
))
1710 pfn
= page_to_pfn(head
) + thp_nr_pages(page
) - 1;
1713 * HWPoison pages have elevated reference counts so the migration would
1714 * fail on them. It also doesn't make any sense to migrate them in the
1715 * first place. Still try to unmap such a page in case it is still mapped
1716 * (e.g. current hwpoison implementation doesn't unmap KSM pages but keep
1717 * the unmap as the catch all safety net).
1719 if (PageHWPoison(page
)) {
1720 if (WARN_ON(PageLRU(page
)))
1721 isolate_lru_page(page
);
1722 if (page_mapped(page
))
1723 try_to_unmap(page
, TTU_IGNORE_MLOCK
);
1727 if (!get_page_unless_zero(page
))
1730 * We can skip free pages. And we can deal with pages on
1731 * LRU and non-lru movable pages.
1734 ret
= isolate_lru_page(page
);
1736 ret
= isolate_movable_page(page
, ISOLATE_UNEVICTABLE
);
1737 if (!ret
) { /* Success */
1738 list_add_tail(&page
->lru
, &source
);
1739 if (!__PageMovable(page
))
1740 inc_node_page_state(page
, NR_ISOLATED_ANON
+
1741 page_is_file_lru(page
));
1744 if (__ratelimit(&migrate_rs
)) {
1745 pr_warn("failed to isolate pfn %lx\n", pfn
);
1746 dump_page(page
, "isolation failed");
1751 if (!list_empty(&source
)) {
1752 nodemask_t nmask
= node_states
[N_MEMORY
];
1753 struct migration_target_control mtc
= {
1755 .gfp_mask
= GFP_USER
| __GFP_MOVABLE
| __GFP_RETRY_MAYFAIL
,
1759 * We have checked that migration range is on a single zone so
1760 * we can use the nid of the first page to all the others.
1762 mtc
.nid
= page_to_nid(list_first_entry(&source
, struct page
, lru
));
1765 * try to allocate from a different node but reuse this node
1766 * if there are no other online nodes to be used (e.g. we are
1767 * offlining a part of the only existing node)
1769 node_clear(mtc
.nid
, nmask
);
1770 if (nodes_empty(nmask
))
1771 node_set(mtc
.nid
, nmask
);
1772 ret
= migrate_pages(&source
, alloc_migration_target
, NULL
,
1773 (unsigned long)&mtc
, MIGRATE_SYNC
, MR_MEMORY_HOTPLUG
, NULL
);
1775 list_for_each_entry(page
, &source
, lru
) {
1776 if (__ratelimit(&migrate_rs
)) {
1777 pr_warn("migrating pfn %lx failed ret:%d\n",
1778 page_to_pfn(page
), ret
);
1779 dump_page(page
, "migration failure");
1782 putback_movable_pages(&source
);
1789 static int __init
cmdline_parse_movable_node(char *p
)
1791 movable_node_enabled
= true;
1794 early_param("movable_node", cmdline_parse_movable_node
);
1796 /* check which state of node_states will be changed when offline memory */
1797 static void node_states_check_changes_offline(unsigned long nr_pages
,
1798 struct zone
*zone
, struct memory_notify
*arg
)
1800 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
1801 unsigned long present_pages
= 0;
1804 arg
->status_change_nid
= NUMA_NO_NODE
;
1805 arg
->status_change_nid_normal
= NUMA_NO_NODE
;
1806 arg
->status_change_nid_high
= NUMA_NO_NODE
;
1809 * Check whether node_states[N_NORMAL_MEMORY] will be changed.
1810 * If the memory to be offline is within the range
1811 * [0..ZONE_NORMAL], and it is the last present memory there,
1812 * the zones in that range will become empty after the offlining,
1813 * thus we can determine that we need to clear the node from
1814 * node_states[N_NORMAL_MEMORY].
1816 for (zt
= 0; zt
<= ZONE_NORMAL
; zt
++)
1817 present_pages
+= pgdat
->node_zones
[zt
].present_pages
;
1818 if (zone_idx(zone
) <= ZONE_NORMAL
&& nr_pages
>= present_pages
)
1819 arg
->status_change_nid_normal
= zone_to_nid(zone
);
1821 #ifdef CONFIG_HIGHMEM
1823 * node_states[N_HIGH_MEMORY] contains nodes which
1824 * have normal memory or high memory.
1825 * Here we add the present_pages belonging to ZONE_HIGHMEM.
1826 * If the zone is within the range of [0..ZONE_HIGHMEM), and
1827 * we determine that the zones in that range become empty,
1828 * we need to clear the node for N_HIGH_MEMORY.
1830 present_pages
+= pgdat
->node_zones
[ZONE_HIGHMEM
].present_pages
;
1831 if (zone_idx(zone
) <= ZONE_HIGHMEM
&& nr_pages
>= present_pages
)
1832 arg
->status_change_nid_high
= zone_to_nid(zone
);
1836 * We have accounted the pages from [0..ZONE_NORMAL), and
1837 * in case of CONFIG_HIGHMEM the pages from ZONE_HIGHMEM
1839 * Here we count the possible pages from ZONE_MOVABLE.
1840 * If after having accounted all the pages, we see that the nr_pages
1841 * to be offlined is over or equal to the accounted pages,
1842 * we know that the node will become empty, and so, we can clear
1843 * it for N_MEMORY as well.
1845 present_pages
+= pgdat
->node_zones
[ZONE_MOVABLE
].present_pages
;
1847 if (nr_pages
>= present_pages
)
1848 arg
->status_change_nid
= zone_to_nid(zone
);
1851 static void node_states_clear_node(int node
, struct memory_notify
*arg
)
1853 if (arg
->status_change_nid_normal
>= 0)
1854 node_clear_state(node
, N_NORMAL_MEMORY
);
1856 if (arg
->status_change_nid_high
>= 0)
1857 node_clear_state(node
, N_HIGH_MEMORY
);
1859 if (arg
->status_change_nid
>= 0)
1860 node_clear_state(node
, N_MEMORY
);
1863 static int count_system_ram_pages_cb(unsigned long start_pfn
,
1864 unsigned long nr_pages
, void *data
)
1866 unsigned long *nr_system_ram_pages
= data
;
1868 *nr_system_ram_pages
+= nr_pages
;
1872 int __ref
offline_pages(unsigned long start_pfn
, unsigned long nr_pages
,
1873 struct memory_group
*group
)
1875 const unsigned long end_pfn
= start_pfn
+ nr_pages
;
1876 unsigned long pfn
, system_ram_pages
= 0;
1877 unsigned long flags
;
1879 struct memory_notify arg
;
1884 * {on,off}lining is constrained to full memory sections (or more
1885 * precisely to memory blocks from the user space POV).
1886 * memmap_on_memory is an exception because it reserves initial part
1887 * of the physical memory space for vmemmaps. That space is pageblock
1890 if (WARN_ON_ONCE(!nr_pages
||
1891 !IS_ALIGNED(start_pfn
, pageblock_nr_pages
) ||
1892 !IS_ALIGNED(start_pfn
+ nr_pages
, PAGES_PER_SECTION
)))
1895 mem_hotplug_begin();
1898 * Don't allow to offline memory blocks that contain holes.
1899 * Consequently, memory blocks with holes can never get onlined
1900 * via the hotplug path - online_pages() - as hotplugged memory has
1901 * no holes. This way, we e.g., don't have to worry about marking
1902 * memory holes PG_reserved, don't need pfn_valid() checks, and can
1903 * avoid using walk_system_ram_range() later.
1905 walk_system_ram_range(start_pfn
, nr_pages
, &system_ram_pages
,
1906 count_system_ram_pages_cb
);
1907 if (system_ram_pages
!= nr_pages
) {
1909 reason
= "memory holes";
1910 goto failed_removal
;
1913 /* This makes hotplug much easier...and readable.
1914 we assume this for now. .*/
1915 zone
= test_pages_in_a_zone(start_pfn
, end_pfn
);
1918 reason
= "multizone range";
1919 goto failed_removal
;
1921 node
= zone_to_nid(zone
);
1924 * Disable pcplists so that page isolation cannot race with freeing
1925 * in a way that pages from isolated pageblock are left on pcplists.
1927 zone_pcp_disable(zone
);
1928 lru_cache_disable();
1930 /* set above range as isolated */
1931 ret
= start_isolate_page_range(start_pfn
, end_pfn
,
1933 MEMORY_OFFLINE
| REPORT_FAILURE
);
1935 reason
= "failure to isolate range";
1936 goto failed_removal_pcplists_disabled
;
1939 arg
.start_pfn
= start_pfn
;
1940 arg
.nr_pages
= nr_pages
;
1941 node_states_check_changes_offline(nr_pages
, zone
, &arg
);
1943 ret
= memory_notify(MEM_GOING_OFFLINE
, &arg
);
1944 ret
= notifier_to_errno(ret
);
1946 reason
= "notifier failure";
1947 goto failed_removal_isolated
;
1953 if (signal_pending(current
)) {
1955 reason
= "signal backoff";
1956 goto failed_removal_isolated
;
1961 ret
= scan_movable_pages(pfn
, end_pfn
, &pfn
);
1964 * TODO: fatal migration failures should bail
1967 do_migrate_range(pfn
, end_pfn
);
1971 if (ret
!= -ENOENT
) {
1972 reason
= "unmovable page";
1973 goto failed_removal_isolated
;
1977 * Dissolve free hugepages in the memory block before doing
1978 * offlining actually in order to make hugetlbfs's object
1979 * counting consistent.
1981 ret
= dissolve_free_huge_pages(start_pfn
, end_pfn
);
1983 reason
= "failure to dissolve huge pages";
1984 goto failed_removal_isolated
;
1987 ret
= test_pages_isolated(start_pfn
, end_pfn
, MEMORY_OFFLINE
);
1991 /* Mark all sections offline and remove free pages from the buddy. */
1992 __offline_isolated_pages(start_pfn
, end_pfn
);
1993 pr_debug("Offlined Pages %ld\n", nr_pages
);
1996 * The memory sections are marked offline, and the pageblock flags
1997 * effectively stale; nobody should be touching them. Fixup the number
1998 * of isolated pageblocks, memory onlining will properly revert this.
2000 spin_lock_irqsave(&zone
->lock
, flags
);
2001 zone
->nr_isolate_pageblock
-= nr_pages
/ pageblock_nr_pages
;
2002 spin_unlock_irqrestore(&zone
->lock
, flags
);
2005 zone_pcp_enable(zone
);
2007 /* removal success */
2008 adjust_managed_page_count(pfn_to_page(start_pfn
), -nr_pages
);
2009 adjust_present_page_count(pfn_to_page(start_pfn
), group
, -nr_pages
);
2011 /* reinitialise watermarks and update pcp limits */
2012 init_per_zone_wmark_min();
2014 if (!populated_zone(zone
)) {
2015 zone_pcp_reset(zone
);
2016 build_all_zonelists(NULL
);
2019 node_states_clear_node(node
, &arg
);
2020 if (arg
.status_change_nid
>= 0) {
2022 kcompactd_stop(node
);
2025 writeback_set_ratelimit();
2027 memory_notify(MEM_OFFLINE
, &arg
);
2028 remove_pfn_range_from_zone(zone
, start_pfn
, nr_pages
);
2032 failed_removal_isolated
:
2033 undo_isolate_page_range(start_pfn
, end_pfn
, MIGRATE_MOVABLE
);
2034 memory_notify(MEM_CANCEL_OFFLINE
, &arg
);
2035 failed_removal_pcplists_disabled
:
2037 zone_pcp_enable(zone
);
2039 pr_debug("memory offlining [mem %#010llx-%#010llx] failed due to %s\n",
2040 (unsigned long long) start_pfn
<< PAGE_SHIFT
,
2041 ((unsigned long long) end_pfn
<< PAGE_SHIFT
) - 1,
2043 /* pushback to free area */
2048 static int check_memblock_offlined_cb(struct memory_block
*mem
, void *arg
)
2050 int ret
= !is_memblock_offlined(mem
);
2054 if (unlikely(ret
)) {
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 get_nr_vmemmap_pages_cb(struct memory_block
*mem
, void *arg
)
2070 * If not set, continue with the next block.
2072 return mem
->nr_vmemmap_pages
;
2075 static int check_cpu_on_node(pg_data_t
*pgdat
)
2079 for_each_present_cpu(cpu
) {
2080 if (cpu_to_node(cpu
) == pgdat
->node_id
)
2082 * the cpu on this node isn't removed, and we can't
2083 * offline this node.
2091 static int check_no_memblock_for_node_cb(struct memory_block
*mem
, void *arg
)
2093 int nid
= *(int *)arg
;
2096 * If a memory block belongs to multiple nodes, the stored nid is not
2097 * reliable. However, such blocks are always online (e.g., cannot get
2098 * offlined) and, therefore, are still spanned by the node.
2100 return mem
->nid
== nid
? -EEXIST
: 0;
2107 * Offline a node if all memory sections and cpus of the node are removed.
2109 * NOTE: The caller must call lock_device_hotplug() to serialize hotplug
2110 * and online/offline operations before this call.
2112 void try_offline_node(int nid
)
2114 pg_data_t
*pgdat
= NODE_DATA(nid
);
2118 * If the node still spans pages (especially ZONE_DEVICE), don't
2119 * offline it. A node spans memory after move_pfn_range_to_zone(),
2120 * e.g., after the memory block was onlined.
2122 if (pgdat
->node_spanned_pages
)
2126 * Especially offline memory blocks might not be spanned by the
2127 * node. They will get spanned by the node once they get onlined.
2128 * However, they link to the node in sysfs and can get onlined later.
2130 rc
= for_each_memory_block(&nid
, check_no_memblock_for_node_cb
);
2134 if (check_cpu_on_node(pgdat
))
2138 * all memory/cpu of this node are removed, we can offline this
2141 node_set_offline(nid
);
2142 unregister_one_node(nid
);
2144 EXPORT_SYMBOL(try_offline_node
);
2146 static int __ref
try_remove_memory(u64 start
, u64 size
)
2148 struct vmem_altmap mhp_altmap
= {};
2149 struct vmem_altmap
*altmap
= NULL
;
2150 unsigned long nr_vmemmap_pages
;
2151 int rc
= 0, nid
= NUMA_NO_NODE
;
2153 BUG_ON(check_hotplug_memory_range(start
, size
));
2156 * All memory blocks must be offlined before removing memory. Check
2157 * whether all memory blocks in question are offline and return error
2158 * if this is not the case.
2160 * While at it, determine the nid. Note that if we'd have mixed nodes,
2161 * we'd only try to offline the last determined one -- which is good
2162 * enough for the cases we care about.
2164 rc
= walk_memory_blocks(start
, size
, &nid
, check_memblock_offlined_cb
);
2169 * We only support removing memory added with MHP_MEMMAP_ON_MEMORY in
2170 * the same granularity it was added - a single memory block.
2172 if (memmap_on_memory
) {
2173 nr_vmemmap_pages
= walk_memory_blocks(start
, size
, NULL
,
2174 get_nr_vmemmap_pages_cb
);
2175 if (nr_vmemmap_pages
) {
2176 if (size
!= memory_block_size_bytes()) {
2177 pr_warn("Refuse to remove %#llx - %#llx,"
2178 "wrong granularity\n",
2179 start
, start
+ size
);
2184 * Let remove_pmd_table->free_hugepage_table do the
2185 * right thing if we used vmem_altmap when hot-adding
2188 mhp_altmap
.alloc
= nr_vmemmap_pages
;
2189 altmap
= &mhp_altmap
;
2193 /* remove memmap entry */
2194 firmware_map_remove(start
, start
+ size
, "System RAM");
2197 * Memory block device removal under the device_hotplug_lock is
2198 * a barrier against racing online attempts.
2200 remove_memory_block_devices(start
, size
);
2202 mem_hotplug_begin();
2204 arch_remove_memory(start
, size
, altmap
);
2206 if (IS_ENABLED(CONFIG_ARCH_KEEP_MEMBLOCK
)) {
2207 memblock_free(start
, size
);
2208 memblock_remove(start
, size
);
2211 release_mem_region_adjustable(start
, size
);
2213 if (nid
!= NUMA_NO_NODE
)
2214 try_offline_node(nid
);
2221 * __remove_memory - Remove memory if every memory block is offline
2222 * @start: physical address of the region to remove
2223 * @size: size of the region to remove
2225 * NOTE: The caller must call lock_device_hotplug() to serialize hotplug
2226 * and online/offline operations before this call, as required by
2227 * try_offline_node().
2229 void __remove_memory(u64 start
, u64 size
)
2233 * trigger BUG() if some memory is not offlined prior to calling this
2236 if (try_remove_memory(start
, size
))
2241 * Remove memory if every memory block is offline, otherwise return -EBUSY is
2242 * some memory is not offline
2244 int remove_memory(u64 start
, u64 size
)
2248 lock_device_hotplug();
2249 rc
= try_remove_memory(start
, size
);
2250 unlock_device_hotplug();
2254 EXPORT_SYMBOL_GPL(remove_memory
);
2256 static int try_offline_memory_block(struct memory_block
*mem
, void *arg
)
2258 uint8_t online_type
= MMOP_ONLINE_KERNEL
;
2259 uint8_t **online_types
= arg
;
2264 * Sense the online_type via the zone of the memory block. Offlining
2265 * with multiple zones within one memory block will be rejected
2266 * by offlining code ... so we don't care about that.
2268 page
= pfn_to_online_page(section_nr_to_pfn(mem
->start_section_nr
));
2269 if (page
&& zone_idx(page_zone(page
)) == ZONE_MOVABLE
)
2270 online_type
= MMOP_ONLINE_MOVABLE
;
2272 rc
= device_offline(&mem
->dev
);
2274 * Default is MMOP_OFFLINE - change it only if offlining succeeded,
2275 * so try_reonline_memory_block() can do the right thing.
2278 **online_types
= online_type
;
2281 /* Ignore if already offline. */
2282 return rc
< 0 ? rc
: 0;
2285 static int try_reonline_memory_block(struct memory_block
*mem
, void *arg
)
2287 uint8_t **online_types
= arg
;
2290 if (**online_types
!= MMOP_OFFLINE
) {
2291 mem
->online_type
= **online_types
;
2292 rc
= device_online(&mem
->dev
);
2294 pr_warn("%s: Failed to re-online memory: %d",
2298 /* Continue processing all remaining memory blocks. */
2304 * Try to offline and remove memory. Might take a long time to finish in case
2305 * memory is still in use. Primarily useful for memory devices that logically
2306 * unplugged all memory (so it's no longer in use) and want to offline + remove
2309 int offline_and_remove_memory(u64 start
, u64 size
)
2311 const unsigned long mb_count
= size
/ memory_block_size_bytes();
2312 uint8_t *online_types
, *tmp
;
2315 if (!IS_ALIGNED(start
, memory_block_size_bytes()) ||
2316 !IS_ALIGNED(size
, memory_block_size_bytes()) || !size
)
2320 * We'll remember the old online type of each memory block, so we can
2321 * try to revert whatever we did when offlining one memory block fails
2322 * after offlining some others succeeded.
2324 online_types
= kmalloc_array(mb_count
, sizeof(*online_types
),
2329 * Initialize all states to MMOP_OFFLINE, so when we abort processing in
2330 * try_offline_memory_block(), we'll skip all unprocessed blocks in
2331 * try_reonline_memory_block().
2333 memset(online_types
, MMOP_OFFLINE
, mb_count
);
2335 lock_device_hotplug();
2338 rc
= walk_memory_blocks(start
, size
, &tmp
, try_offline_memory_block
);
2341 * In case we succeeded to offline all memory, remove it.
2342 * This cannot fail as it cannot get onlined in the meantime.
2345 rc
= try_remove_memory(start
, size
);
2347 pr_err("%s: Failed to remove memory: %d", __func__
, rc
);
2351 * Rollback what we did. While memory onlining might theoretically fail
2352 * (nacked by a notifier), it barely ever happens.
2356 walk_memory_blocks(start
, size
, &tmp
,
2357 try_reonline_memory_block
);
2359 unlock_device_hotplug();
2361 kfree(online_types
);
2364 EXPORT_SYMBOL_GPL(offline_and_remove_memory
);
2365 #endif /* CONFIG_MEMORY_HOTREMOVE */