Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * linux/mm/memory_hotplug.c
4 *
5 * Copyright (C)
6 */
7
8#include <linux/stddef.h>
9#include <linux/mm.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>
39
40#include <asm/tlbflush.h>
41
42#include "internal.h"
43#include "shuffle.h"
44
45/*
46 * online_page_callback contains pointer to current page onlining function.
47 * Initially it is generic_online_page(). If it is required it could be
48 * changed by calling set_online_page_callback() for callback registration
49 * and restore_online_page_callback() for generic callback restore.
50 */
51
52static void generic_online_page(struct page *page, unsigned int order);
53
54static online_page_callback_t online_page_callback = generic_online_page;
55static DEFINE_MUTEX(online_page_callback_lock);
56
57DEFINE_STATIC_PERCPU_RWSEM(mem_hotplug_lock);
58
59void get_online_mems(void)
60{
61 percpu_down_read(&mem_hotplug_lock);
62}
63
64void put_online_mems(void)
65{
66 percpu_up_read(&mem_hotplug_lock);
67}
68
69bool movable_node_enabled = false;
70
71#ifndef CONFIG_MEMORY_HOTPLUG_DEFAULT_ONLINE
72bool memhp_auto_online;
73#else
74bool memhp_auto_online = true;
75#endif
76EXPORT_SYMBOL_GPL(memhp_auto_online);
77
78static int __init setup_memhp_default_state(char *str)
79{
80 if (!strcmp(str, "online"))
81 memhp_auto_online = true;
82 else if (!strcmp(str, "offline"))
83 memhp_auto_online = false;
84
85 return 1;
86}
87__setup("memhp_default_state=", setup_memhp_default_state);
88
89void mem_hotplug_begin(void)
90{
91 cpus_read_lock();
92 percpu_down_write(&mem_hotplug_lock);
93}
94
95void mem_hotplug_done(void)
96{
97 percpu_up_write(&mem_hotplug_lock);
98 cpus_read_unlock();
99}
100
101u64 max_mem_size = U64_MAX;
102
103/* add this memory to iomem resource */
104static struct resource *register_memory_resource(u64 start, u64 size)
105{
106 struct resource *res;
107 unsigned long flags = IORESOURCE_SYSTEM_RAM | IORESOURCE_BUSY;
108 char *resource_name = "System RAM";
109
110 if (start + size > max_mem_size)
111 return ERR_PTR(-E2BIG);
112
113 /*
114 * Request ownership of the new memory range. This might be
115 * a child of an existing resource that was present but
116 * not marked as busy.
117 */
118 res = __request_region(&iomem_resource, start, size,
119 resource_name, flags);
120
121 if (!res) {
122 pr_debug("Unable to reserve System RAM region: %016llx->%016llx\n",
123 start, start + size);
124 return ERR_PTR(-EEXIST);
125 }
126 return res;
127}
128
129static void release_memory_resource(struct resource *res)
130{
131 if (!res)
132 return;
133 release_resource(res);
134 kfree(res);
135}
136
137#ifdef CONFIG_MEMORY_HOTPLUG_SPARSE
138void get_page_bootmem(unsigned long info, struct page *page,
139 unsigned long type)
140{
141 page->freelist = (void *)type;
142 SetPagePrivate(page);
143 set_page_private(page, info);
144 page_ref_inc(page);
145}
146
147void put_page_bootmem(struct page *page)
148{
149 unsigned long type;
150
151 type = (unsigned long) page->freelist;
152 BUG_ON(type < MEMORY_HOTPLUG_MIN_BOOTMEM_TYPE ||
153 type > MEMORY_HOTPLUG_MAX_BOOTMEM_TYPE);
154
155 if (page_ref_dec_return(page) == 1) {
156 page->freelist = NULL;
157 ClearPagePrivate(page);
158 set_page_private(page, 0);
159 INIT_LIST_HEAD(&page->lru);
160 free_reserved_page(page);
161 }
162}
163
164#ifdef CONFIG_HAVE_BOOTMEM_INFO_NODE
165#ifndef CONFIG_SPARSEMEM_VMEMMAP
166static void register_page_bootmem_info_section(unsigned long start_pfn)
167{
168 unsigned long mapsize, section_nr, i;
169 struct mem_section *ms;
170 struct page *page, *memmap;
171 struct mem_section_usage *usage;
172
173 section_nr = pfn_to_section_nr(start_pfn);
174 ms = __nr_to_section(section_nr);
175
176 /* Get section's memmap address */
177 memmap = sparse_decode_mem_map(ms->section_mem_map, section_nr);
178
179 /*
180 * Get page for the memmap's phys address
181 * XXX: need more consideration for sparse_vmemmap...
182 */
183 page = virt_to_page(memmap);
184 mapsize = sizeof(struct page) * PAGES_PER_SECTION;
185 mapsize = PAGE_ALIGN(mapsize) >> PAGE_SHIFT;
186
187 /* remember memmap's page */
188 for (i = 0; i < mapsize; i++, page++)
189 get_page_bootmem(section_nr, page, SECTION_INFO);
190
191 usage = ms->usage;
192 page = virt_to_page(usage);
193
194 mapsize = PAGE_ALIGN(mem_section_usage_size()) >> PAGE_SHIFT;
195
196 for (i = 0; i < mapsize; i++, page++)
197 get_page_bootmem(section_nr, page, MIX_SECTION_INFO);
198
199}
200#else /* CONFIG_SPARSEMEM_VMEMMAP */
201static void register_page_bootmem_info_section(unsigned long start_pfn)
202{
203 unsigned long mapsize, section_nr, i;
204 struct mem_section *ms;
205 struct page *page, *memmap;
206 struct mem_section_usage *usage;
207
208 section_nr = pfn_to_section_nr(start_pfn);
209 ms = __nr_to_section(section_nr);
210
211 memmap = sparse_decode_mem_map(ms->section_mem_map, section_nr);
212
213 register_page_bootmem_memmap(section_nr, memmap, PAGES_PER_SECTION);
214
215 usage = ms->usage;
216 page = virt_to_page(usage);
217
218 mapsize = PAGE_ALIGN(mem_section_usage_size()) >> PAGE_SHIFT;
219
220 for (i = 0; i < mapsize; i++, page++)
221 get_page_bootmem(section_nr, page, MIX_SECTION_INFO);
222}
223#endif /* !CONFIG_SPARSEMEM_VMEMMAP */
224
225void __init register_page_bootmem_info_node(struct pglist_data *pgdat)
226{
227 unsigned long i, pfn, end_pfn, nr_pages;
228 int node = pgdat->node_id;
229 struct page *page;
230
231 nr_pages = PAGE_ALIGN(sizeof(struct pglist_data)) >> PAGE_SHIFT;
232 page = virt_to_page(pgdat);
233
234 for (i = 0; i < nr_pages; i++, page++)
235 get_page_bootmem(node, page, NODE_INFO);
236
237 pfn = pgdat->node_start_pfn;
238 end_pfn = pgdat_end_pfn(pgdat);
239
240 /* register section info */
241 for (; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
242 /*
243 * Some platforms can assign the same pfn to multiple nodes - on
244 * node0 as well as nodeN. To avoid registering a pfn against
245 * multiple nodes we check that this pfn does not already
246 * reside in some other nodes.
247 */
248 if (pfn_valid(pfn) && (early_pfn_to_nid(pfn) == node))
249 register_page_bootmem_info_section(pfn);
250 }
251}
252#endif /* CONFIG_HAVE_BOOTMEM_INFO_NODE */
253
254static int check_pfn_span(unsigned long pfn, unsigned long nr_pages,
255 const char *reason)
256{
257 /*
258 * Disallow all operations smaller than a sub-section and only
259 * allow operations smaller than a section for
260 * SPARSEMEM_VMEMMAP. Note that check_hotplug_memory_range()
261 * enforces a larger memory_block_size_bytes() granularity for
262 * memory that will be marked online, so this check should only
263 * fire for direct arch_{add,remove}_memory() users outside of
264 * add_memory_resource().
265 */
266 unsigned long min_align;
267
268 if (IS_ENABLED(CONFIG_SPARSEMEM_VMEMMAP))
269 min_align = PAGES_PER_SUBSECTION;
270 else
271 min_align = PAGES_PER_SECTION;
272 if (!IS_ALIGNED(pfn, min_align)
273 || !IS_ALIGNED(nr_pages, min_align)) {
274 WARN(1, "Misaligned __%s_pages start: %#lx end: #%lx\n",
275 reason, pfn, pfn + nr_pages - 1);
276 return -EINVAL;
277 }
278 return 0;
279}
280
281/*
282 * Reasonably generic function for adding memory. It is
283 * expected that archs that support memory hotplug will
284 * call this function after deciding the zone to which to
285 * add the new pages.
286 */
287int __ref __add_pages(int nid, unsigned long pfn, unsigned long nr_pages,
288 struct mhp_restrictions *restrictions)
289{
290 int err;
291 unsigned long nr, start_sec, end_sec;
292 struct vmem_altmap *altmap = restrictions->altmap;
293
294 if (altmap) {
295 /*
296 * Validate altmap is within bounds of the total request
297 */
298 if (altmap->base_pfn != pfn
299 || vmem_altmap_offset(altmap) > nr_pages) {
300 pr_warn_once("memory add fail, invalid altmap\n");
301 return -EINVAL;
302 }
303 altmap->alloc = 0;
304 }
305
306 err = check_pfn_span(pfn, nr_pages, "add");
307 if (err)
308 return err;
309
310 start_sec = pfn_to_section_nr(pfn);
311 end_sec = pfn_to_section_nr(pfn + nr_pages - 1);
312 for (nr = start_sec; nr <= end_sec; nr++) {
313 unsigned long pfns;
314
315 pfns = min(nr_pages, PAGES_PER_SECTION
316 - (pfn & ~PAGE_SECTION_MASK));
317 err = sparse_add_section(nid, pfn, pfns, altmap);
318 if (err)
319 break;
320 pfn += pfns;
321 nr_pages -= pfns;
322 cond_resched();
323 }
324 vmemmap_populate_print_last();
325 return err;
326}
327
328/* find the smallest valid pfn in the range [start_pfn, end_pfn) */
329static unsigned long find_smallest_section_pfn(int nid, struct zone *zone,
330 unsigned long start_pfn,
331 unsigned long end_pfn)
332{
333 for (; start_pfn < end_pfn; start_pfn += PAGES_PER_SUBSECTION) {
334 if (unlikely(!pfn_to_online_page(start_pfn)))
335 continue;
336
337 if (unlikely(pfn_to_nid(start_pfn) != nid))
338 continue;
339
340 if (zone && zone != page_zone(pfn_to_page(start_pfn)))
341 continue;
342
343 return start_pfn;
344 }
345
346 return 0;
347}
348
349/* find the biggest valid pfn in the range [start_pfn, end_pfn). */
350static unsigned long find_biggest_section_pfn(int nid, struct zone *zone,
351 unsigned long start_pfn,
352 unsigned long end_pfn)
353{
354 unsigned long pfn;
355
356 /* pfn is the end pfn of a memory section. */
357 pfn = end_pfn - 1;
358 for (; pfn >= start_pfn; pfn -= PAGES_PER_SUBSECTION) {
359 if (unlikely(!pfn_to_online_page(pfn)))
360 continue;
361
362 if (unlikely(pfn_to_nid(pfn) != nid))
363 continue;
364
365 if (zone && zone != page_zone(pfn_to_page(pfn)))
366 continue;
367
368 return pfn;
369 }
370
371 return 0;
372}
373
374static void shrink_zone_span(struct zone *zone, unsigned long start_pfn,
375 unsigned long end_pfn)
376{
377 unsigned long zone_start_pfn = zone->zone_start_pfn;
378 unsigned long z = zone_end_pfn(zone); /* zone_end_pfn namespace clash */
379 unsigned long zone_end_pfn = z;
380 unsigned long pfn;
381 int nid = zone_to_nid(zone);
382
383 zone_span_writelock(zone);
384 if (zone_start_pfn == start_pfn) {
385 /*
386 * If the section is smallest section in the zone, it need
387 * shrink zone->zone_start_pfn and zone->zone_spanned_pages.
388 * In this case, we find second smallest valid mem_section
389 * for shrinking zone.
390 */
391 pfn = find_smallest_section_pfn(nid, zone, end_pfn,
392 zone_end_pfn);
393 if (pfn) {
394 zone->zone_start_pfn = pfn;
395 zone->spanned_pages = zone_end_pfn - pfn;
396 }
397 } else if (zone_end_pfn == end_pfn) {
398 /*
399 * If the section is biggest section in the zone, it need
400 * shrink zone->spanned_pages.
401 * In this case, we find second biggest valid mem_section for
402 * shrinking zone.
403 */
404 pfn = find_biggest_section_pfn(nid, zone, zone_start_pfn,
405 start_pfn);
406 if (pfn)
407 zone->spanned_pages = pfn - zone_start_pfn + 1;
408 }
409
410 /*
411 * The section is not biggest or smallest mem_section in the zone, it
412 * only creates a hole in the zone. So in this case, we need not
413 * change the zone. But perhaps, the zone has only hole data. Thus
414 * it check the zone has only hole or not.
415 */
416 pfn = zone_start_pfn;
417 for (; pfn < zone_end_pfn; pfn += PAGES_PER_SUBSECTION) {
418 if (unlikely(!pfn_to_online_page(pfn)))
419 continue;
420
421 if (page_zone(pfn_to_page(pfn)) != zone)
422 continue;
423
424 /* Skip range to be removed */
425 if (pfn >= start_pfn && pfn < end_pfn)
426 continue;
427
428 /* If we find valid section, we have nothing to do */
429 zone_span_writeunlock(zone);
430 return;
431 }
432
433 /* The zone has no valid section */
434 zone->zone_start_pfn = 0;
435 zone->spanned_pages = 0;
436 zone_span_writeunlock(zone);
437}
438
439static void update_pgdat_span(struct pglist_data *pgdat)
440{
441 unsigned long node_start_pfn = 0, node_end_pfn = 0;
442 struct zone *zone;
443
444 for (zone = pgdat->node_zones;
445 zone < pgdat->node_zones + MAX_NR_ZONES; zone++) {
446 unsigned long zone_end_pfn = zone->zone_start_pfn +
447 zone->spanned_pages;
448
449 /* No need to lock the zones, they can't change. */
450 if (!zone->spanned_pages)
451 continue;
452 if (!node_end_pfn) {
453 node_start_pfn = zone->zone_start_pfn;
454 node_end_pfn = zone_end_pfn;
455 continue;
456 }
457
458 if (zone_end_pfn > node_end_pfn)
459 node_end_pfn = zone_end_pfn;
460 if (zone->zone_start_pfn < node_start_pfn)
461 node_start_pfn = zone->zone_start_pfn;
462 }
463
464 pgdat->node_start_pfn = node_start_pfn;
465 pgdat->node_spanned_pages = node_end_pfn - node_start_pfn;
466}
467
468static void __remove_zone(struct zone *zone, unsigned long start_pfn,
469 unsigned long nr_pages)
470{
471 struct pglist_data *pgdat = zone->zone_pgdat;
472 unsigned long flags;
473
474#ifdef CONFIG_ZONE_DEVICE
475 /*
476 * Zone shrinking code cannot properly deal with ZONE_DEVICE. So
477 * we will not try to shrink the zones - which is okay as
478 * set_zone_contiguous() cannot deal with ZONE_DEVICE either way.
479 */
480 if (zone_idx(zone) == ZONE_DEVICE)
481 return;
482#endif
483
484 pgdat_resize_lock(zone->zone_pgdat, &flags);
485 shrink_zone_span(zone, start_pfn, start_pfn + nr_pages);
486 update_pgdat_span(pgdat);
487 pgdat_resize_unlock(zone->zone_pgdat, &flags);
488}
489
490static void __remove_section(struct zone *zone, unsigned long pfn,
491 unsigned long nr_pages, unsigned long map_offset,
492 struct vmem_altmap *altmap)
493{
494 struct mem_section *ms = __nr_to_section(pfn_to_section_nr(pfn));
495
496 if (WARN_ON_ONCE(!valid_section(ms)))
497 return;
498
499 __remove_zone(zone, pfn, nr_pages);
500 sparse_remove_section(ms, pfn, nr_pages, map_offset, altmap);
501}
502
503/**
504 * __remove_pages() - remove sections of pages from a zone
505 * @zone: zone from which pages need to be removed
506 * @pfn: starting pageframe (must be aligned to start of a section)
507 * @nr_pages: number of pages to remove (must be multiple of section size)
508 * @altmap: alternative device page map or %NULL if default memmap is used
509 *
510 * Generic helper function to remove section mappings and sysfs entries
511 * for the section of the memory we are removing. Caller needs to make
512 * sure that pages are marked reserved and zones are adjust properly by
513 * calling offline_pages().
514 */
515void __remove_pages(struct zone *zone, unsigned long pfn,
516 unsigned long nr_pages, struct vmem_altmap *altmap)
517{
518 unsigned long map_offset = 0;
519 unsigned long nr, start_sec, end_sec;
520
521 map_offset = vmem_altmap_offset(altmap);
522
523 clear_zone_contiguous(zone);
524
525 if (check_pfn_span(pfn, nr_pages, "remove"))
526 return;
527
528 start_sec = pfn_to_section_nr(pfn);
529 end_sec = pfn_to_section_nr(pfn + nr_pages - 1);
530 for (nr = start_sec; nr <= end_sec; nr++) {
531 unsigned long pfns;
532
533 cond_resched();
534 pfns = min(nr_pages, PAGES_PER_SECTION
535 - (pfn & ~PAGE_SECTION_MASK));
536 __remove_section(zone, pfn, pfns, map_offset, altmap);
537 pfn += pfns;
538 nr_pages -= pfns;
539 map_offset = 0;
540 }
541
542 set_zone_contiguous(zone);
543}
544
545int set_online_page_callback(online_page_callback_t callback)
546{
547 int rc = -EINVAL;
548
549 get_online_mems();
550 mutex_lock(&online_page_callback_lock);
551
552 if (online_page_callback == generic_online_page) {
553 online_page_callback = callback;
554 rc = 0;
555 }
556
557 mutex_unlock(&online_page_callback_lock);
558 put_online_mems();
559
560 return rc;
561}
562EXPORT_SYMBOL_GPL(set_online_page_callback);
563
564int restore_online_page_callback(online_page_callback_t callback)
565{
566 int rc = -EINVAL;
567
568 get_online_mems();
569 mutex_lock(&online_page_callback_lock);
570
571 if (online_page_callback == callback) {
572 online_page_callback = generic_online_page;
573 rc = 0;
574 }
575
576 mutex_unlock(&online_page_callback_lock);
577 put_online_mems();
578
579 return rc;
580}
581EXPORT_SYMBOL_GPL(restore_online_page_callback);
582
583void __online_page_set_limits(struct page *page)
584{
585}
586EXPORT_SYMBOL_GPL(__online_page_set_limits);
587
588void __online_page_increment_counters(struct page *page)
589{
590 adjust_managed_page_count(page, 1);
591}
592EXPORT_SYMBOL_GPL(__online_page_increment_counters);
593
594void __online_page_free(struct page *page)
595{
596 __free_reserved_page(page);
597}
598EXPORT_SYMBOL_GPL(__online_page_free);
599
600static void generic_online_page(struct page *page, unsigned int order)
601{
602 kernel_map_pages(page, 1 << order, 1);
603 __free_pages_core(page, order);
604 totalram_pages_add(1UL << order);
605#ifdef CONFIG_HIGHMEM
606 if (PageHighMem(page))
607 totalhigh_pages_add(1UL << order);
608#endif
609}
610
611static int online_pages_range(unsigned long start_pfn, unsigned long nr_pages,
612 void *arg)
613{
614 const unsigned long end_pfn = start_pfn + nr_pages;
615 unsigned long pfn;
616 int order;
617
618 /*
619 * Online the pages. The callback might decide to keep some pages
620 * PG_reserved (to add them to the buddy later), but we still account
621 * them as being online/belonging to this zone ("present").
622 */
623 for (pfn = start_pfn; pfn < end_pfn; pfn += 1ul << order) {
624 order = min(MAX_ORDER - 1, get_order(PFN_PHYS(end_pfn - pfn)));
625 /* __free_pages_core() wants pfns to be aligned to the order */
626 if (WARN_ON_ONCE(!IS_ALIGNED(pfn, 1ul << order)))
627 order = 0;
628 (*online_page_callback)(pfn_to_page(pfn), order);
629 }
630
631 /* mark all involved sections as online */
632 online_mem_sections(start_pfn, end_pfn);
633
634 *(unsigned long *)arg += nr_pages;
635 return 0;
636}
637
638/* check which state of node_states will be changed when online memory */
639static void node_states_check_changes_online(unsigned long nr_pages,
640 struct zone *zone, struct memory_notify *arg)
641{
642 int nid = zone_to_nid(zone);
643
644 arg->status_change_nid = NUMA_NO_NODE;
645 arg->status_change_nid_normal = NUMA_NO_NODE;
646 arg->status_change_nid_high = NUMA_NO_NODE;
647
648 if (!node_state(nid, N_MEMORY))
649 arg->status_change_nid = nid;
650 if (zone_idx(zone) <= ZONE_NORMAL && !node_state(nid, N_NORMAL_MEMORY))
651 arg->status_change_nid_normal = nid;
652#ifdef CONFIG_HIGHMEM
653 if (zone_idx(zone) <= ZONE_HIGHMEM && !node_state(nid, N_HIGH_MEMORY))
654 arg->status_change_nid_high = nid;
655#endif
656}
657
658static void node_states_set_node(int node, struct memory_notify *arg)
659{
660 if (arg->status_change_nid_normal >= 0)
661 node_set_state(node, N_NORMAL_MEMORY);
662
663 if (arg->status_change_nid_high >= 0)
664 node_set_state(node, N_HIGH_MEMORY);
665
666 if (arg->status_change_nid >= 0)
667 node_set_state(node, N_MEMORY);
668}
669
670static void __meminit resize_zone_range(struct zone *zone, unsigned long start_pfn,
671 unsigned long nr_pages)
672{
673 unsigned long old_end_pfn = zone_end_pfn(zone);
674
675 if (zone_is_empty(zone) || start_pfn < zone->zone_start_pfn)
676 zone->zone_start_pfn = start_pfn;
677
678 zone->spanned_pages = max(start_pfn + nr_pages, old_end_pfn) - zone->zone_start_pfn;
679}
680
681static void __meminit resize_pgdat_range(struct pglist_data *pgdat, unsigned long start_pfn,
682 unsigned long nr_pages)
683{
684 unsigned long old_end_pfn = pgdat_end_pfn(pgdat);
685
686 if (!pgdat->node_spanned_pages || start_pfn < pgdat->node_start_pfn)
687 pgdat->node_start_pfn = start_pfn;
688
689 pgdat->node_spanned_pages = max(start_pfn + nr_pages, old_end_pfn) - pgdat->node_start_pfn;
690
691}
692/*
693 * Associate the pfn range with the given zone, initializing the memmaps
694 * and resizing the pgdat/zone data to span the added pages. After this
695 * call, all affected pages are PG_reserved.
696 */
697void __ref move_pfn_range_to_zone(struct zone *zone, unsigned long start_pfn,
698 unsigned long nr_pages, struct vmem_altmap *altmap)
699{
700 struct pglist_data *pgdat = zone->zone_pgdat;
701 int nid = pgdat->node_id;
702 unsigned long flags;
703
704 clear_zone_contiguous(zone);
705
706 /* TODO Huh pgdat is irqsave while zone is not. It used to be like that before */
707 pgdat_resize_lock(pgdat, &flags);
708 zone_span_writelock(zone);
709 if (zone_is_empty(zone))
710 init_currently_empty_zone(zone, start_pfn, nr_pages);
711 resize_zone_range(zone, start_pfn, nr_pages);
712 zone_span_writeunlock(zone);
713 resize_pgdat_range(pgdat, start_pfn, nr_pages);
714 pgdat_resize_unlock(pgdat, &flags);
715
716 /*
717 * TODO now we have a visible range of pages which are not associated
718 * with their zone properly. Not nice but set_pfnblock_flags_mask
719 * expects the zone spans the pfn range. All the pages in the range
720 * are reserved so nobody should be touching them so we should be safe
721 */
722 memmap_init_zone(nr_pages, nid, zone_idx(zone), start_pfn,
723 MEMMAP_HOTPLUG, altmap);
724
725 set_zone_contiguous(zone);
726}
727
728/*
729 * Returns a default kernel memory zone for the given pfn range.
730 * If no kernel zone covers this pfn range it will automatically go
731 * to the ZONE_NORMAL.
732 */
733static struct zone *default_kernel_zone_for_pfn(int nid, unsigned long start_pfn,
734 unsigned long nr_pages)
735{
736 struct pglist_data *pgdat = NODE_DATA(nid);
737 int zid;
738
739 for (zid = 0; zid <= ZONE_NORMAL; zid++) {
740 struct zone *zone = &pgdat->node_zones[zid];
741
742 if (zone_intersects(zone, start_pfn, nr_pages))
743 return zone;
744 }
745
746 return &pgdat->node_zones[ZONE_NORMAL];
747}
748
749static inline struct zone *default_zone_for_pfn(int nid, unsigned long start_pfn,
750 unsigned long nr_pages)
751{
752 struct zone *kernel_zone = default_kernel_zone_for_pfn(nid, start_pfn,
753 nr_pages);
754 struct zone *movable_zone = &NODE_DATA(nid)->node_zones[ZONE_MOVABLE];
755 bool in_kernel = zone_intersects(kernel_zone, start_pfn, nr_pages);
756 bool in_movable = zone_intersects(movable_zone, start_pfn, nr_pages);
757
758 /*
759 * We inherit the existing zone in a simple case where zones do not
760 * overlap in the given range
761 */
762 if (in_kernel ^ in_movable)
763 return (in_kernel) ? kernel_zone : movable_zone;
764
765 /*
766 * If the range doesn't belong to any zone or two zones overlap in the
767 * given range then we use movable zone only if movable_node is
768 * enabled because we always online to a kernel zone by default.
769 */
770 return movable_node_enabled ? movable_zone : kernel_zone;
771}
772
773struct zone * zone_for_pfn_range(int online_type, int nid, unsigned start_pfn,
774 unsigned long nr_pages)
775{
776 if (online_type == MMOP_ONLINE_KERNEL)
777 return default_kernel_zone_for_pfn(nid, start_pfn, nr_pages);
778
779 if (online_type == MMOP_ONLINE_MOVABLE)
780 return &NODE_DATA(nid)->node_zones[ZONE_MOVABLE];
781
782 return default_zone_for_pfn(nid, start_pfn, nr_pages);
783}
784
785int __ref online_pages(unsigned long pfn, unsigned long nr_pages, int online_type)
786{
787 unsigned long flags;
788 unsigned long onlined_pages = 0;
789 struct zone *zone;
790 int need_zonelists_rebuild = 0;
791 int nid;
792 int ret;
793 struct memory_notify arg;
794 struct memory_block *mem;
795
796 mem_hotplug_begin();
797
798 /*
799 * We can't use pfn_to_nid() because nid might be stored in struct page
800 * which is not yet initialized. Instead, we find nid from memory block.
801 */
802 mem = find_memory_block(__pfn_to_section(pfn));
803 nid = mem->nid;
804 put_device(&mem->dev);
805
806 /* associate pfn range with the zone */
807 zone = zone_for_pfn_range(online_type, nid, pfn, nr_pages);
808 move_pfn_range_to_zone(zone, pfn, nr_pages, NULL);
809
810 arg.start_pfn = pfn;
811 arg.nr_pages = nr_pages;
812 node_states_check_changes_online(nr_pages, zone, &arg);
813
814 ret = memory_notify(MEM_GOING_ONLINE, &arg);
815 ret = notifier_to_errno(ret);
816 if (ret)
817 goto failed_addition;
818
819 /*
820 * If this zone is not populated, then it is not in zonelist.
821 * This means the page allocator ignores this zone.
822 * So, zonelist must be updated after online.
823 */
824 if (!populated_zone(zone)) {
825 need_zonelists_rebuild = 1;
826 setup_zone_pageset(zone);
827 }
828
829 ret = walk_system_ram_range(pfn, nr_pages, &onlined_pages,
830 online_pages_range);
831 if (ret) {
832 /* not a single memory resource was applicable */
833 if (need_zonelists_rebuild)
834 zone_pcp_reset(zone);
835 goto failed_addition;
836 }
837
838 zone->present_pages += onlined_pages;
839
840 pgdat_resize_lock(zone->zone_pgdat, &flags);
841 zone->zone_pgdat->node_present_pages += onlined_pages;
842 pgdat_resize_unlock(zone->zone_pgdat, &flags);
843
844 shuffle_zone(zone);
845
846 node_states_set_node(nid, &arg);
847 if (need_zonelists_rebuild)
848 build_all_zonelists(NULL);
849 else
850 zone_pcp_update(zone);
851
852 init_per_zone_wmark_min();
853
854 kswapd_run(nid);
855 kcompactd_run(nid);
856
857 vm_total_pages = nr_free_pagecache_pages();
858
859 writeback_set_ratelimit();
860
861 memory_notify(MEM_ONLINE, &arg);
862 mem_hotplug_done();
863 return 0;
864
865failed_addition:
866 pr_debug("online_pages [mem %#010llx-%#010llx] failed\n",
867 (unsigned long long) pfn << PAGE_SHIFT,
868 (((unsigned long long) pfn + nr_pages) << PAGE_SHIFT) - 1);
869 memory_notify(MEM_CANCEL_ONLINE, &arg);
870 mem_hotplug_done();
871 return ret;
872}
873#endif /* CONFIG_MEMORY_HOTPLUG_SPARSE */
874
875static void reset_node_present_pages(pg_data_t *pgdat)
876{
877 struct zone *z;
878
879 for (z = pgdat->node_zones; z < pgdat->node_zones + MAX_NR_ZONES; z++)
880 z->present_pages = 0;
881
882 pgdat->node_present_pages = 0;
883}
884
885/* we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG */
886static pg_data_t __ref *hotadd_new_pgdat(int nid, u64 start)
887{
888 struct pglist_data *pgdat;
889 unsigned long start_pfn = PFN_DOWN(start);
890
891 pgdat = NODE_DATA(nid);
892 if (!pgdat) {
893 pgdat = arch_alloc_nodedata(nid);
894 if (!pgdat)
895 return NULL;
896
897 pgdat->per_cpu_nodestats =
898 alloc_percpu(struct per_cpu_nodestat);
899 arch_refresh_nodedata(nid, pgdat);
900 } else {
901 int cpu;
902 /*
903 * Reset the nr_zones, order and classzone_idx before reuse.
904 * Note that kswapd will init kswapd_classzone_idx properly
905 * when it starts in the near future.
906 */
907 pgdat->nr_zones = 0;
908 pgdat->kswapd_order = 0;
909 pgdat->kswapd_classzone_idx = 0;
910 for_each_online_cpu(cpu) {
911 struct per_cpu_nodestat *p;
912
913 p = per_cpu_ptr(pgdat->per_cpu_nodestats, cpu);
914 memset(p, 0, sizeof(*p));
915 }
916 }
917
918 /* we can use NODE_DATA(nid) from here */
919
920 pgdat->node_id = nid;
921 pgdat->node_start_pfn = start_pfn;
922
923 /* init node's zones as empty zones, we don't have any present pages.*/
924 free_area_init_core_hotplug(nid);
925
926 /*
927 * The node we allocated has no zone fallback lists. For avoiding
928 * to access not-initialized zonelist, build here.
929 */
930 build_all_zonelists(pgdat);
931
932 /*
933 * When memory is hot-added, all the memory is in offline state. So
934 * clear all zones' present_pages because they will be updated in
935 * online_pages() and offline_pages().
936 */
937 reset_node_managed_pages(pgdat);
938 reset_node_present_pages(pgdat);
939
940 return pgdat;
941}
942
943static void rollback_node_hotadd(int nid)
944{
945 pg_data_t *pgdat = NODE_DATA(nid);
946
947 arch_refresh_nodedata(nid, NULL);
948 free_percpu(pgdat->per_cpu_nodestats);
949 arch_free_nodedata(pgdat);
950}
951
952
953/**
954 * try_online_node - online a node if offlined
955 * @nid: the node ID
956 * @start: start addr of the node
957 * @set_node_online: Whether we want to online the node
958 * called by cpu_up() to online a node without onlined memory.
959 *
960 * Returns:
961 * 1 -> a new node has been allocated
962 * 0 -> the node is already online
963 * -ENOMEM -> the node could not be allocated
964 */
965static int __try_online_node(int nid, u64 start, bool set_node_online)
966{
967 pg_data_t *pgdat;
968 int ret = 1;
969
970 if (node_online(nid))
971 return 0;
972
973 pgdat = hotadd_new_pgdat(nid, start);
974 if (!pgdat) {
975 pr_err("Cannot online node %d due to NULL pgdat\n", nid);
976 ret = -ENOMEM;
977 goto out;
978 }
979
980 if (set_node_online) {
981 node_set_online(nid);
982 ret = register_one_node(nid);
983 BUG_ON(ret);
984 }
985out:
986 return ret;
987}
988
989/*
990 * Users of this function always want to online/register the node
991 */
992int try_online_node(int nid)
993{
994 int ret;
995
996 mem_hotplug_begin();
997 ret = __try_online_node(nid, 0, true);
998 mem_hotplug_done();
999 return ret;
1000}
1001
1002static int check_hotplug_memory_range(u64 start, u64 size)
1003{
1004 /* memory range must be block size aligned */
1005 if (!size || !IS_ALIGNED(start, memory_block_size_bytes()) ||
1006 !IS_ALIGNED(size, memory_block_size_bytes())) {
1007 pr_err("Block size [%#lx] unaligned hotplug range: start %#llx, size %#llx",
1008 memory_block_size_bytes(), start, size);
1009 return -EINVAL;
1010 }
1011
1012 return 0;
1013}
1014
1015static int online_memory_block(struct memory_block *mem, void *arg)
1016{
1017 return device_online(&mem->dev);
1018}
1019
1020/*
1021 * NOTE: The caller must call lock_device_hotplug() to serialize hotplug
1022 * and online/offline operations (triggered e.g. by sysfs).
1023 *
1024 * we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG
1025 */
1026int __ref add_memory_resource(int nid, struct resource *res)
1027{
1028 struct mhp_restrictions restrictions = {};
1029 u64 start, size;
1030 bool new_node = false;
1031 int ret;
1032
1033 start = res->start;
1034 size = resource_size(res);
1035
1036 ret = check_hotplug_memory_range(start, size);
1037 if (ret)
1038 return ret;
1039
1040 mem_hotplug_begin();
1041
1042 /*
1043 * Add new range to memblock so that when hotadd_new_pgdat() is called
1044 * to allocate new pgdat, get_pfn_range_for_nid() will be able to find
1045 * this new range and calculate total pages correctly. The range will
1046 * be removed at hot-remove time.
1047 */
1048 memblock_add_node(start, size, nid);
1049
1050 ret = __try_online_node(nid, start, false);
1051 if (ret < 0)
1052 goto error;
1053 new_node = ret;
1054
1055 /* call arch's memory hotadd */
1056 ret = arch_add_memory(nid, start, size, &restrictions);
1057 if (ret < 0)
1058 goto error;
1059
1060 /* create memory block devices after memory was added */
1061 ret = create_memory_block_devices(start, size);
1062 if (ret) {
1063 arch_remove_memory(nid, start, size, NULL);
1064 goto error;
1065 }
1066
1067 if (new_node) {
1068 /* If sysfs file of new node can't be created, cpu on the node
1069 * can't be hot-added. There is no rollback way now.
1070 * So, check by BUG_ON() to catch it reluctantly..
1071 * We online node here. We can't roll back from here.
1072 */
1073 node_set_online(nid);
1074 ret = __register_one_node(nid);
1075 BUG_ON(ret);
1076 }
1077
1078 /* link memory sections under this node.*/
1079 ret = link_mem_sections(nid, PFN_DOWN(start), PFN_UP(start + size - 1));
1080 BUG_ON(ret);
1081
1082 /* create new memmap entry */
1083 firmware_map_add_hotplug(start, start + size, "System RAM");
1084
1085 /* device_online() will take the lock when calling online_pages() */
1086 mem_hotplug_done();
1087
1088 /* online pages if requested */
1089 if (memhp_auto_online)
1090 walk_memory_blocks(start, size, NULL, online_memory_block);
1091
1092 return ret;
1093error:
1094 /* rollback pgdat allocation and others */
1095 if (new_node)
1096 rollback_node_hotadd(nid);
1097 memblock_remove(start, size);
1098 mem_hotplug_done();
1099 return ret;
1100}
1101
1102/* requires device_hotplug_lock, see add_memory_resource() */
1103int __ref __add_memory(int nid, u64 start, u64 size)
1104{
1105 struct resource *res;
1106 int ret;
1107
1108 res = register_memory_resource(start, size);
1109 if (IS_ERR(res))
1110 return PTR_ERR(res);
1111
1112 ret = add_memory_resource(nid, res);
1113 if (ret < 0)
1114 release_memory_resource(res);
1115 return ret;
1116}
1117
1118int add_memory(int nid, u64 start, u64 size)
1119{
1120 int rc;
1121
1122 lock_device_hotplug();
1123 rc = __add_memory(nid, start, size);
1124 unlock_device_hotplug();
1125
1126 return rc;
1127}
1128EXPORT_SYMBOL_GPL(add_memory);
1129
1130#ifdef CONFIG_MEMORY_HOTREMOVE
1131/*
1132 * A free page on the buddy free lists (not the per-cpu lists) has PageBuddy
1133 * set and the size of the free page is given by page_order(). Using this,
1134 * the function determines if the pageblock contains only free pages.
1135 * Due to buddy contraints, a free page at least the size of a pageblock will
1136 * be located at the start of the pageblock
1137 */
1138static inline int pageblock_free(struct page *page)
1139{
1140 return PageBuddy(page) && page_order(page) >= pageblock_order;
1141}
1142
1143/* Return the pfn of the start of the next active pageblock after a given pfn */
1144static unsigned long next_active_pageblock(unsigned long pfn)
1145{
1146 struct page *page = pfn_to_page(pfn);
1147
1148 /* Ensure the starting page is pageblock-aligned */
1149 BUG_ON(pfn & (pageblock_nr_pages - 1));
1150
1151 /* If the entire pageblock is free, move to the end of free page */
1152 if (pageblock_free(page)) {
1153 int order;
1154 /* be careful. we don't have locks, page_order can be changed.*/
1155 order = page_order(page);
1156 if ((order < MAX_ORDER) && (order >= pageblock_order))
1157 return pfn + (1 << order);
1158 }
1159
1160 return pfn + pageblock_nr_pages;
1161}
1162
1163static bool is_pageblock_removable_nolock(unsigned long pfn)
1164{
1165 struct page *page = pfn_to_page(pfn);
1166 struct zone *zone;
1167
1168 /*
1169 * We have to be careful here because we are iterating over memory
1170 * sections which are not zone aware so we might end up outside of
1171 * the zone but still within the section.
1172 * We have to take care about the node as well. If the node is offline
1173 * its NODE_DATA will be NULL - see page_zone.
1174 */
1175 if (!node_online(page_to_nid(page)))
1176 return false;
1177
1178 zone = page_zone(page);
1179 pfn = page_to_pfn(page);
1180 if (!zone_spans_pfn(zone, pfn))
1181 return false;
1182
1183 return !has_unmovable_pages(zone, page, 0, MIGRATE_MOVABLE, SKIP_HWPOISON);
1184}
1185
1186/* Checks if this range of memory is likely to be hot-removable. */
1187bool is_mem_section_removable(unsigned long start_pfn, unsigned long nr_pages)
1188{
1189 unsigned long end_pfn, pfn;
1190
1191 end_pfn = min(start_pfn + nr_pages,
1192 zone_end_pfn(page_zone(pfn_to_page(start_pfn))));
1193
1194 /* Check the starting page of each pageblock within the range */
1195 for (pfn = start_pfn; pfn < end_pfn; pfn = next_active_pageblock(pfn)) {
1196 if (!is_pageblock_removable_nolock(pfn))
1197 return false;
1198 cond_resched();
1199 }
1200
1201 /* All pageblocks in the memory block are likely to be hot-removable */
1202 return true;
1203}
1204
1205/*
1206 * Confirm all pages in a range [start, end) belong to the same zone.
1207 * When true, return its valid [start, end).
1208 */
1209int test_pages_in_a_zone(unsigned long start_pfn, unsigned long end_pfn,
1210 unsigned long *valid_start, unsigned long *valid_end)
1211{
1212 unsigned long pfn, sec_end_pfn;
1213 unsigned long start, end;
1214 struct zone *zone = NULL;
1215 struct page *page;
1216 int i;
1217 for (pfn = start_pfn, sec_end_pfn = SECTION_ALIGN_UP(start_pfn + 1);
1218 pfn < end_pfn;
1219 pfn = sec_end_pfn, sec_end_pfn += PAGES_PER_SECTION) {
1220 /* Make sure the memory section is present first */
1221 if (!present_section_nr(pfn_to_section_nr(pfn)))
1222 continue;
1223 for (; pfn < sec_end_pfn && pfn < end_pfn;
1224 pfn += MAX_ORDER_NR_PAGES) {
1225 i = 0;
1226 /* This is just a CONFIG_HOLES_IN_ZONE check.*/
1227 while ((i < MAX_ORDER_NR_PAGES) &&
1228 !pfn_valid_within(pfn + i))
1229 i++;
1230 if (i == MAX_ORDER_NR_PAGES || pfn + i >= end_pfn)
1231 continue;
1232 /* Check if we got outside of the zone */
1233 if (zone && !zone_spans_pfn(zone, pfn + i))
1234 return 0;
1235 page = pfn_to_page(pfn + i);
1236 if (zone && page_zone(page) != zone)
1237 return 0;
1238 if (!zone)
1239 start = pfn + i;
1240 zone = page_zone(page);
1241 end = pfn + MAX_ORDER_NR_PAGES;
1242 }
1243 }
1244
1245 if (zone) {
1246 *valid_start = start;
1247 *valid_end = min(end, end_pfn);
1248 return 1;
1249 } else {
1250 return 0;
1251 }
1252}
1253
1254/*
1255 * Scan pfn range [start,end) to find movable/migratable pages (LRU pages,
1256 * non-lru movable pages and hugepages). We scan pfn because it's much
1257 * easier than scanning over linked list. This function returns the pfn
1258 * of the first found movable page if it's found, otherwise 0.
1259 */
1260static unsigned long scan_movable_pages(unsigned long start, unsigned long end)
1261{
1262 unsigned long pfn;
1263
1264 for (pfn = start; pfn < end; pfn++) {
1265 struct page *page, *head;
1266 unsigned long skip;
1267
1268 if (!pfn_valid(pfn))
1269 continue;
1270 page = pfn_to_page(pfn);
1271 if (PageLRU(page))
1272 return pfn;
1273 if (__PageMovable(page))
1274 return pfn;
1275
1276 if (!PageHuge(page))
1277 continue;
1278 head = compound_head(page);
1279 if (page_huge_active(head))
1280 return pfn;
1281 skip = compound_nr(head) - (page - head);
1282 pfn += skip - 1;
1283 }
1284 return 0;
1285}
1286
1287static struct page *new_node_page(struct page *page, unsigned long private)
1288{
1289 int nid = page_to_nid(page);
1290 nodemask_t nmask = node_states[N_MEMORY];
1291
1292 /*
1293 * try to allocate from a different node but reuse this node if there
1294 * are no other online nodes to be used (e.g. we are offlining a part
1295 * of the only existing node)
1296 */
1297 node_clear(nid, nmask);
1298 if (nodes_empty(nmask))
1299 node_set(nid, nmask);
1300
1301 return new_page_nodemask(page, nid, &nmask);
1302}
1303
1304static int
1305do_migrate_range(unsigned long start_pfn, unsigned long end_pfn)
1306{
1307 unsigned long pfn;
1308 struct page *page;
1309 int ret = 0;
1310 LIST_HEAD(source);
1311
1312 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
1313 if (!pfn_valid(pfn))
1314 continue;
1315 page = pfn_to_page(pfn);
1316
1317 if (PageHuge(page)) {
1318 struct page *head = compound_head(page);
1319 pfn = page_to_pfn(head) + compound_nr(head) - 1;
1320 isolate_huge_page(head, &source);
1321 continue;
1322 } else if (PageTransHuge(page))
1323 pfn = page_to_pfn(compound_head(page))
1324 + hpage_nr_pages(page) - 1;
1325
1326 /*
1327 * HWPoison pages have elevated reference counts so the migration would
1328 * fail on them. It also doesn't make any sense to migrate them in the
1329 * first place. Still try to unmap such a page in case it is still mapped
1330 * (e.g. current hwpoison implementation doesn't unmap KSM pages but keep
1331 * the unmap as the catch all safety net).
1332 */
1333 if (PageHWPoison(page)) {
1334 if (WARN_ON(PageLRU(page)))
1335 isolate_lru_page(page);
1336 if (page_mapped(page))
1337 try_to_unmap(page, TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS);
1338 continue;
1339 }
1340
1341 if (!get_page_unless_zero(page))
1342 continue;
1343 /*
1344 * We can skip free pages. And we can deal with pages on
1345 * LRU and non-lru movable pages.
1346 */
1347 if (PageLRU(page))
1348 ret = isolate_lru_page(page);
1349 else
1350 ret = isolate_movable_page(page, ISOLATE_UNEVICTABLE);
1351 if (!ret) { /* Success */
1352 list_add_tail(&page->lru, &source);
1353 if (!__PageMovable(page))
1354 inc_node_page_state(page, NR_ISOLATED_ANON +
1355 page_is_file_cache(page));
1356
1357 } else {
1358 pr_warn("failed to isolate pfn %lx\n", pfn);
1359 dump_page(page, "isolation failed");
1360 }
1361 put_page(page);
1362 }
1363 if (!list_empty(&source)) {
1364 /* Allocate a new page from the nearest neighbor node */
1365 ret = migrate_pages(&source, new_node_page, NULL, 0,
1366 MIGRATE_SYNC, MR_MEMORY_HOTPLUG);
1367 if (ret) {
1368 list_for_each_entry(page, &source, lru) {
1369 pr_warn("migrating pfn %lx failed ret:%d ",
1370 page_to_pfn(page), ret);
1371 dump_page(page, "migration failure");
1372 }
1373 putback_movable_pages(&source);
1374 }
1375 }
1376
1377 return ret;
1378}
1379
1380/*
1381 * remove from free_area[] and mark all as Reserved.
1382 */
1383static int
1384offline_isolated_pages_cb(unsigned long start, unsigned long nr_pages,
1385 void *data)
1386{
1387 unsigned long *offlined_pages = (unsigned long *)data;
1388
1389 *offlined_pages += __offline_isolated_pages(start, start + nr_pages);
1390 return 0;
1391}
1392
1393/*
1394 * Check all pages in range, recoreded as memory resource, are isolated.
1395 */
1396static int
1397check_pages_isolated_cb(unsigned long start_pfn, unsigned long nr_pages,
1398 void *data)
1399{
1400 return test_pages_isolated(start_pfn, start_pfn + nr_pages, true);
1401}
1402
1403static int __init cmdline_parse_movable_node(char *p)
1404{
1405#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
1406 movable_node_enabled = true;
1407#else
1408 pr_warn("movable_node parameter depends on CONFIG_HAVE_MEMBLOCK_NODE_MAP to work properly\n");
1409#endif
1410 return 0;
1411}
1412early_param("movable_node", cmdline_parse_movable_node);
1413
1414/* check which state of node_states will be changed when offline memory */
1415static void node_states_check_changes_offline(unsigned long nr_pages,
1416 struct zone *zone, struct memory_notify *arg)
1417{
1418 struct pglist_data *pgdat = zone->zone_pgdat;
1419 unsigned long present_pages = 0;
1420 enum zone_type zt;
1421
1422 arg->status_change_nid = NUMA_NO_NODE;
1423 arg->status_change_nid_normal = NUMA_NO_NODE;
1424 arg->status_change_nid_high = NUMA_NO_NODE;
1425
1426 /*
1427 * Check whether node_states[N_NORMAL_MEMORY] will be changed.
1428 * If the memory to be offline is within the range
1429 * [0..ZONE_NORMAL], and it is the last present memory there,
1430 * the zones in that range will become empty after the offlining,
1431 * thus we can determine that we need to clear the node from
1432 * node_states[N_NORMAL_MEMORY].
1433 */
1434 for (zt = 0; zt <= ZONE_NORMAL; zt++)
1435 present_pages += pgdat->node_zones[zt].present_pages;
1436 if (zone_idx(zone) <= ZONE_NORMAL && nr_pages >= present_pages)
1437 arg->status_change_nid_normal = zone_to_nid(zone);
1438
1439#ifdef CONFIG_HIGHMEM
1440 /*
1441 * node_states[N_HIGH_MEMORY] contains nodes which
1442 * have normal memory or high memory.
1443 * Here we add the present_pages belonging to ZONE_HIGHMEM.
1444 * If the zone is within the range of [0..ZONE_HIGHMEM), and
1445 * we determine that the zones in that range become empty,
1446 * we need to clear the node for N_HIGH_MEMORY.
1447 */
1448 present_pages += pgdat->node_zones[ZONE_HIGHMEM].present_pages;
1449 if (zone_idx(zone) <= ZONE_HIGHMEM && nr_pages >= present_pages)
1450 arg->status_change_nid_high = zone_to_nid(zone);
1451#endif
1452
1453 /*
1454 * We have accounted the pages from [0..ZONE_NORMAL), and
1455 * in case of CONFIG_HIGHMEM the pages from ZONE_HIGHMEM
1456 * as well.
1457 * Here we count the possible pages from ZONE_MOVABLE.
1458 * If after having accounted all the pages, we see that the nr_pages
1459 * to be offlined is over or equal to the accounted pages,
1460 * we know that the node will become empty, and so, we can clear
1461 * it for N_MEMORY as well.
1462 */
1463 present_pages += pgdat->node_zones[ZONE_MOVABLE].present_pages;
1464
1465 if (nr_pages >= present_pages)
1466 arg->status_change_nid = zone_to_nid(zone);
1467}
1468
1469static void node_states_clear_node(int node, struct memory_notify *arg)
1470{
1471 if (arg->status_change_nid_normal >= 0)
1472 node_clear_state(node, N_NORMAL_MEMORY);
1473
1474 if (arg->status_change_nid_high >= 0)
1475 node_clear_state(node, N_HIGH_MEMORY);
1476
1477 if (arg->status_change_nid >= 0)
1478 node_clear_state(node, N_MEMORY);
1479}
1480
1481static int __ref __offline_pages(unsigned long start_pfn,
1482 unsigned long end_pfn)
1483{
1484 unsigned long pfn, nr_pages;
1485 unsigned long offlined_pages = 0;
1486 int ret, node, nr_isolate_pageblock;
1487 unsigned long flags;
1488 unsigned long valid_start, valid_end;
1489 struct zone *zone;
1490 struct memory_notify arg;
1491 char *reason;
1492
1493 mem_hotplug_begin();
1494
1495 /* This makes hotplug much easier...and readable.
1496 we assume this for now. .*/
1497 if (!test_pages_in_a_zone(start_pfn, end_pfn, &valid_start,
1498 &valid_end)) {
1499 ret = -EINVAL;
1500 reason = "multizone range";
1501 goto failed_removal;
1502 }
1503
1504 zone = page_zone(pfn_to_page(valid_start));
1505 node = zone_to_nid(zone);
1506 nr_pages = end_pfn - start_pfn;
1507
1508 /* set above range as isolated */
1509 ret = start_isolate_page_range(start_pfn, end_pfn,
1510 MIGRATE_MOVABLE,
1511 SKIP_HWPOISON | REPORT_FAILURE);
1512 if (ret < 0) {
1513 reason = "failure to isolate range";
1514 goto failed_removal;
1515 }
1516 nr_isolate_pageblock = ret;
1517
1518 arg.start_pfn = start_pfn;
1519 arg.nr_pages = nr_pages;
1520 node_states_check_changes_offline(nr_pages, zone, &arg);
1521
1522 ret = memory_notify(MEM_GOING_OFFLINE, &arg);
1523 ret = notifier_to_errno(ret);
1524 if (ret) {
1525 reason = "notifier failure";
1526 goto failed_removal_isolated;
1527 }
1528
1529 do {
1530 for (pfn = start_pfn; pfn;) {
1531 if (signal_pending(current)) {
1532 ret = -EINTR;
1533 reason = "signal backoff";
1534 goto failed_removal_isolated;
1535 }
1536
1537 cond_resched();
1538 lru_add_drain_all();
1539
1540 pfn = scan_movable_pages(pfn, end_pfn);
1541 if (pfn) {
1542 /*
1543 * TODO: fatal migration failures should bail
1544 * out
1545 */
1546 do_migrate_range(pfn, end_pfn);
1547 }
1548 }
1549
1550 /*
1551 * Dissolve free hugepages in the memory block before doing
1552 * offlining actually in order to make hugetlbfs's object
1553 * counting consistent.
1554 */
1555 ret = dissolve_free_huge_pages(start_pfn, end_pfn);
1556 if (ret) {
1557 reason = "failure to dissolve huge pages";
1558 goto failed_removal_isolated;
1559 }
1560 /* check again */
1561 ret = walk_system_ram_range(start_pfn, end_pfn - start_pfn,
1562 NULL, check_pages_isolated_cb);
1563 } while (ret);
1564
1565 /* Ok, all of our target is isolated.
1566 We cannot do rollback at this point. */
1567 walk_system_ram_range(start_pfn, end_pfn - start_pfn,
1568 &offlined_pages, offline_isolated_pages_cb);
1569 pr_info("Offlined Pages %ld\n", offlined_pages);
1570 /*
1571 * Onlining will reset pagetype flags and makes migrate type
1572 * MOVABLE, so just need to decrease the number of isolated
1573 * pageblocks zone counter here.
1574 */
1575 spin_lock_irqsave(&zone->lock, flags);
1576 zone->nr_isolate_pageblock -= nr_isolate_pageblock;
1577 spin_unlock_irqrestore(&zone->lock, flags);
1578
1579 /* removal success */
1580 adjust_managed_page_count(pfn_to_page(start_pfn), -offlined_pages);
1581 zone->present_pages -= offlined_pages;
1582
1583 pgdat_resize_lock(zone->zone_pgdat, &flags);
1584 zone->zone_pgdat->node_present_pages -= offlined_pages;
1585 pgdat_resize_unlock(zone->zone_pgdat, &flags);
1586
1587 init_per_zone_wmark_min();
1588
1589 if (!populated_zone(zone)) {
1590 zone_pcp_reset(zone);
1591 build_all_zonelists(NULL);
1592 } else
1593 zone_pcp_update(zone);
1594
1595 node_states_clear_node(node, &arg);
1596 if (arg.status_change_nid >= 0) {
1597 kswapd_stop(node);
1598 kcompactd_stop(node);
1599 }
1600
1601 vm_total_pages = nr_free_pagecache_pages();
1602 writeback_set_ratelimit();
1603
1604 memory_notify(MEM_OFFLINE, &arg);
1605 mem_hotplug_done();
1606 return 0;
1607
1608failed_removal_isolated:
1609 undo_isolate_page_range(start_pfn, end_pfn, MIGRATE_MOVABLE);
1610 memory_notify(MEM_CANCEL_OFFLINE, &arg);
1611failed_removal:
1612 pr_debug("memory offlining [mem %#010llx-%#010llx] failed due to %s\n",
1613 (unsigned long long) start_pfn << PAGE_SHIFT,
1614 ((unsigned long long) end_pfn << PAGE_SHIFT) - 1,
1615 reason);
1616 /* pushback to free area */
1617 mem_hotplug_done();
1618 return ret;
1619}
1620
1621int offline_pages(unsigned long start_pfn, unsigned long nr_pages)
1622{
1623 return __offline_pages(start_pfn, start_pfn + nr_pages);
1624}
1625
1626static int check_memblock_offlined_cb(struct memory_block *mem, void *arg)
1627{
1628 int ret = !is_memblock_offlined(mem);
1629
1630 if (unlikely(ret)) {
1631 phys_addr_t beginpa, endpa;
1632
1633 beginpa = PFN_PHYS(section_nr_to_pfn(mem->start_section_nr));
1634 endpa = beginpa + memory_block_size_bytes() - 1;
1635 pr_warn("removing memory fails, because memory [%pa-%pa] is onlined\n",
1636 &beginpa, &endpa);
1637
1638 return -EBUSY;
1639 }
1640 return 0;
1641}
1642
1643static int check_cpu_on_node(pg_data_t *pgdat)
1644{
1645 int cpu;
1646
1647 for_each_present_cpu(cpu) {
1648 if (cpu_to_node(cpu) == pgdat->node_id)
1649 /*
1650 * the cpu on this node isn't removed, and we can't
1651 * offline this node.
1652 */
1653 return -EBUSY;
1654 }
1655
1656 return 0;
1657}
1658
1659static int check_no_memblock_for_node_cb(struct memory_block *mem, void *arg)
1660{
1661 int nid = *(int *)arg;
1662
1663 /*
1664 * If a memory block belongs to multiple nodes, the stored nid is not
1665 * reliable. However, such blocks are always online (e.g., cannot get
1666 * offlined) and, therefore, are still spanned by the node.
1667 */
1668 return mem->nid == nid ? -EEXIST : 0;
1669}
1670
1671/**
1672 * try_offline_node
1673 * @nid: the node ID
1674 *
1675 * Offline a node if all memory sections and cpus of the node are removed.
1676 *
1677 * NOTE: The caller must call lock_device_hotplug() to serialize hotplug
1678 * and online/offline operations before this call.
1679 */
1680void try_offline_node(int nid)
1681{
1682 pg_data_t *pgdat = NODE_DATA(nid);
1683 int rc;
1684
1685 /*
1686 * If the node still spans pages (especially ZONE_DEVICE), don't
1687 * offline it. A node spans memory after move_pfn_range_to_zone(),
1688 * e.g., after the memory block was onlined.
1689 */
1690 if (pgdat->node_spanned_pages)
1691 return;
1692
1693 /*
1694 * Especially offline memory blocks might not be spanned by the
1695 * node. They will get spanned by the node once they get onlined.
1696 * However, they link to the node in sysfs and can get onlined later.
1697 */
1698 rc = for_each_memory_block(&nid, check_no_memblock_for_node_cb);
1699 if (rc)
1700 return;
1701
1702 if (check_cpu_on_node(pgdat))
1703 return;
1704
1705 /*
1706 * all memory/cpu of this node are removed, we can offline this
1707 * node now.
1708 */
1709 node_set_offline(nid);
1710 unregister_one_node(nid);
1711}
1712EXPORT_SYMBOL(try_offline_node);
1713
1714static void __release_memory_resource(resource_size_t start,
1715 resource_size_t size)
1716{
1717 int ret;
1718
1719 /*
1720 * When removing memory in the same granularity as it was added,
1721 * this function never fails. It might only fail if resources
1722 * have to be adjusted or split. We'll ignore the error, as
1723 * removing of memory cannot fail.
1724 */
1725 ret = release_mem_region_adjustable(&iomem_resource, start, size);
1726 if (ret) {
1727 resource_size_t endres = start + size - 1;
1728
1729 pr_warn("Unable to release resource <%pa-%pa> (%d)\n",
1730 &start, &endres, ret);
1731 }
1732}
1733
1734static int __ref try_remove_memory(int nid, u64 start, u64 size)
1735{
1736 int rc = 0;
1737
1738 BUG_ON(check_hotplug_memory_range(start, size));
1739
1740 mem_hotplug_begin();
1741
1742 /*
1743 * All memory blocks must be offlined before removing memory. Check
1744 * whether all memory blocks in question are offline and return error
1745 * if this is not the case.
1746 */
1747 rc = walk_memory_blocks(start, size, NULL, check_memblock_offlined_cb);
1748 if (rc)
1749 goto done;
1750
1751 /* remove memmap entry */
1752 firmware_map_remove(start, start + size, "System RAM");
1753 memblock_free(start, size);
1754 memblock_remove(start, size);
1755
1756 /* remove memory block devices before removing memory */
1757 remove_memory_block_devices(start, size);
1758
1759 arch_remove_memory(nid, start, size, NULL);
1760 __release_memory_resource(start, size);
1761
1762 try_offline_node(nid);
1763
1764done:
1765 mem_hotplug_done();
1766 return rc;
1767}
1768
1769/**
1770 * remove_memory
1771 * @nid: the node ID
1772 * @start: physical address of the region to remove
1773 * @size: size of the region to remove
1774 *
1775 * NOTE: The caller must call lock_device_hotplug() to serialize hotplug
1776 * and online/offline operations before this call, as required by
1777 * try_offline_node().
1778 */
1779void __remove_memory(int nid, u64 start, u64 size)
1780{
1781
1782 /*
1783 * trigger BUG() if some memory is not offlined prior to calling this
1784 * function
1785 */
1786 if (try_remove_memory(nid, start, size))
1787 BUG();
1788}
1789
1790/*
1791 * Remove memory if every memory block is offline, otherwise return -EBUSY is
1792 * some memory is not offline
1793 */
1794int remove_memory(int nid, u64 start, u64 size)
1795{
1796 int rc;
1797
1798 lock_device_hotplug();
1799 rc = try_remove_memory(nid, start, size);
1800 unlock_device_hotplug();
1801
1802 return rc;
1803}
1804EXPORT_SYMBOL_GPL(remove_memory);
1805#endif /* CONFIG_MEMORY_HOTREMOVE */
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * linux/mm/memory_hotplug.c
4 *
5 * Copyright (C)
6 */
7
8#include <linux/stddef.h>
9#include <linux/mm.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>
39
40#include <asm/tlbflush.h>
41
42#include "internal.h"
43#include "shuffle.h"
44
45
46/*
47 * memory_hotplug.memmap_on_memory parameter
48 */
49static bool memmap_on_memory __ro_after_init;
50#ifdef CONFIG_MHP_MEMMAP_ON_MEMORY
51module_param(memmap_on_memory, bool, 0444);
52MODULE_PARM_DESC(memmap_on_memory, "Enable memmap on memory for memory hotplug");
53#endif
54
55/*
56 * online_page_callback contains pointer to current page onlining function.
57 * Initially it is generic_online_page(). If it is required it could be
58 * changed by calling set_online_page_callback() for callback registration
59 * and restore_online_page_callback() for generic callback restore.
60 */
61
62static online_page_callback_t online_page_callback = generic_online_page;
63static DEFINE_MUTEX(online_page_callback_lock);
64
65DEFINE_STATIC_PERCPU_RWSEM(mem_hotplug_lock);
66
67void get_online_mems(void)
68{
69 percpu_down_read(&mem_hotplug_lock);
70}
71
72void put_online_mems(void)
73{
74 percpu_up_read(&mem_hotplug_lock);
75}
76
77bool movable_node_enabled = false;
78
79#ifndef CONFIG_MEMORY_HOTPLUG_DEFAULT_ONLINE
80int mhp_default_online_type = MMOP_OFFLINE;
81#else
82int mhp_default_online_type = MMOP_ONLINE;
83#endif
84
85static int __init setup_memhp_default_state(char *str)
86{
87 const int online_type = mhp_online_type_from_str(str);
88
89 if (online_type >= 0)
90 mhp_default_online_type = online_type;
91
92 return 1;
93}
94__setup("memhp_default_state=", setup_memhp_default_state);
95
96void mem_hotplug_begin(void)
97{
98 cpus_read_lock();
99 percpu_down_write(&mem_hotplug_lock);
100}
101
102void mem_hotplug_done(void)
103{
104 percpu_up_write(&mem_hotplug_lock);
105 cpus_read_unlock();
106}
107
108u64 max_mem_size = U64_MAX;
109
110/* add this memory to iomem resource */
111static struct resource *register_memory_resource(u64 start, u64 size,
112 const char *resource_name)
113{
114 struct resource *res;
115 unsigned long flags = IORESOURCE_SYSTEM_RAM | IORESOURCE_BUSY;
116
117 if (strcmp(resource_name, "System RAM"))
118 flags |= IORESOURCE_SYSRAM_DRIVER_MANAGED;
119
120 if (!mhp_range_allowed(start, size, true))
121 return ERR_PTR(-E2BIG);
122
123 /*
124 * Make sure value parsed from 'mem=' only restricts memory adding
125 * while booting, so that memory hotplug won't be impacted. Please
126 * refer to document of 'mem=' in kernel-parameters.txt for more
127 * details.
128 */
129 if (start + size > max_mem_size && system_state < SYSTEM_RUNNING)
130 return ERR_PTR(-E2BIG);
131
132 /*
133 * Request ownership of the new memory range. This might be
134 * a child of an existing resource that was present but
135 * not marked as busy.
136 */
137 res = __request_region(&iomem_resource, start, size,
138 resource_name, flags);
139
140 if (!res) {
141 pr_debug("Unable to reserve System RAM region: %016llx->%016llx\n",
142 start, start + size);
143 return ERR_PTR(-EEXIST);
144 }
145 return res;
146}
147
148static void release_memory_resource(struct resource *res)
149{
150 if (!res)
151 return;
152 release_resource(res);
153 kfree(res);
154}
155
156#ifdef CONFIG_MEMORY_HOTPLUG_SPARSE
157static int check_pfn_span(unsigned long pfn, unsigned long nr_pages,
158 const char *reason)
159{
160 /*
161 * Disallow all operations smaller than a sub-section and only
162 * allow operations smaller than a section for
163 * SPARSEMEM_VMEMMAP. Note that check_hotplug_memory_range()
164 * enforces a larger memory_block_size_bytes() granularity for
165 * memory that will be marked online, so this check should only
166 * fire for direct arch_{add,remove}_memory() users outside of
167 * add_memory_resource().
168 */
169 unsigned long min_align;
170
171 if (IS_ENABLED(CONFIG_SPARSEMEM_VMEMMAP))
172 min_align = PAGES_PER_SUBSECTION;
173 else
174 min_align = PAGES_PER_SECTION;
175 if (!IS_ALIGNED(pfn, min_align)
176 || !IS_ALIGNED(nr_pages, min_align)) {
177 WARN(1, "Misaligned __%s_pages start: %#lx end: #%lx\n",
178 reason, pfn, pfn + nr_pages - 1);
179 return -EINVAL;
180 }
181 return 0;
182}
183
184/*
185 * Return page for the valid pfn only if the page is online. All pfn
186 * walkers which rely on the fully initialized page->flags and others
187 * should use this rather than pfn_valid && pfn_to_page
188 */
189struct page *pfn_to_online_page(unsigned long pfn)
190{
191 unsigned long nr = pfn_to_section_nr(pfn);
192 struct dev_pagemap *pgmap;
193 struct mem_section *ms;
194
195 if (nr >= NR_MEM_SECTIONS)
196 return NULL;
197
198 ms = __nr_to_section(nr);
199 if (!online_section(ms))
200 return NULL;
201
202 /*
203 * Save some code text when online_section() +
204 * pfn_section_valid() are sufficient.
205 */
206 if (IS_ENABLED(CONFIG_HAVE_ARCH_PFN_VALID) && !pfn_valid(pfn))
207 return NULL;
208
209 if (!pfn_section_valid(ms, pfn))
210 return NULL;
211
212 if (!online_device_section(ms))
213 return pfn_to_page(pfn);
214
215 /*
216 * Slowpath: when ZONE_DEVICE collides with
217 * ZONE_{NORMAL,MOVABLE} within the same section some pfns in
218 * the section may be 'offline' but 'valid'. Only
219 * get_dev_pagemap() can determine sub-section online status.
220 */
221 pgmap = get_dev_pagemap(pfn, NULL);
222 put_dev_pagemap(pgmap);
223
224 /* The presence of a pgmap indicates ZONE_DEVICE offline pfn */
225 if (pgmap)
226 return NULL;
227
228 return pfn_to_page(pfn);
229}
230EXPORT_SYMBOL_GPL(pfn_to_online_page);
231
232/*
233 * Reasonably generic function for adding memory. It is
234 * expected that archs that support memory hotplug will
235 * call this function after deciding the zone to which to
236 * add the new pages.
237 */
238int __ref __add_pages(int nid, unsigned long pfn, unsigned long nr_pages,
239 struct mhp_params *params)
240{
241 const unsigned long end_pfn = pfn + nr_pages;
242 unsigned long cur_nr_pages;
243 int err;
244 struct vmem_altmap *altmap = params->altmap;
245
246 if (WARN_ON_ONCE(!params->pgprot.pgprot))
247 return -EINVAL;
248
249 VM_BUG_ON(!mhp_range_allowed(PFN_PHYS(pfn), nr_pages * PAGE_SIZE, false));
250
251 if (altmap) {
252 /*
253 * Validate altmap is within bounds of the total request
254 */
255 if (altmap->base_pfn != pfn
256 || vmem_altmap_offset(altmap) > nr_pages) {
257 pr_warn_once("memory add fail, invalid altmap\n");
258 return -EINVAL;
259 }
260 altmap->alloc = 0;
261 }
262
263 err = check_pfn_span(pfn, nr_pages, "add");
264 if (err)
265 return err;
266
267 for (; pfn < end_pfn; pfn += cur_nr_pages) {
268 /* Select all remaining pages up to the next section boundary */
269 cur_nr_pages = min(end_pfn - pfn,
270 SECTION_ALIGN_UP(pfn + 1) - pfn);
271 err = sparse_add_section(nid, pfn, cur_nr_pages, altmap);
272 if (err)
273 break;
274 cond_resched();
275 }
276 vmemmap_populate_print_last();
277 return err;
278}
279
280/* find the smallest valid pfn in the range [start_pfn, end_pfn) */
281static unsigned long find_smallest_section_pfn(int nid, struct zone *zone,
282 unsigned long start_pfn,
283 unsigned long end_pfn)
284{
285 for (; start_pfn < end_pfn; start_pfn += PAGES_PER_SUBSECTION) {
286 if (unlikely(!pfn_to_online_page(start_pfn)))
287 continue;
288
289 if (unlikely(pfn_to_nid(start_pfn) != nid))
290 continue;
291
292 if (zone != page_zone(pfn_to_page(start_pfn)))
293 continue;
294
295 return start_pfn;
296 }
297
298 return 0;
299}
300
301/* find the biggest valid pfn in the range [start_pfn, end_pfn). */
302static unsigned long find_biggest_section_pfn(int nid, struct zone *zone,
303 unsigned long start_pfn,
304 unsigned long end_pfn)
305{
306 unsigned long pfn;
307
308 /* pfn is the end pfn of a memory section. */
309 pfn = end_pfn - 1;
310 for (; pfn >= start_pfn; pfn -= PAGES_PER_SUBSECTION) {
311 if (unlikely(!pfn_to_online_page(pfn)))
312 continue;
313
314 if (unlikely(pfn_to_nid(pfn) != nid))
315 continue;
316
317 if (zone != page_zone(pfn_to_page(pfn)))
318 continue;
319
320 return pfn;
321 }
322
323 return 0;
324}
325
326static void shrink_zone_span(struct zone *zone, unsigned long start_pfn,
327 unsigned long end_pfn)
328{
329 unsigned long pfn;
330 int nid = zone_to_nid(zone);
331
332 if (zone->zone_start_pfn == start_pfn) {
333 /*
334 * If the section is smallest section in the zone, it need
335 * shrink zone->zone_start_pfn and zone->zone_spanned_pages.
336 * In this case, we find second smallest valid mem_section
337 * for shrinking zone.
338 */
339 pfn = find_smallest_section_pfn(nid, zone, end_pfn,
340 zone_end_pfn(zone));
341 if (pfn) {
342 zone->spanned_pages = zone_end_pfn(zone) - pfn;
343 zone->zone_start_pfn = pfn;
344 } else {
345 zone->zone_start_pfn = 0;
346 zone->spanned_pages = 0;
347 }
348 } else if (zone_end_pfn(zone) == end_pfn) {
349 /*
350 * If the section is biggest section in the zone, it need
351 * shrink zone->spanned_pages.
352 * In this case, we find second biggest valid mem_section for
353 * shrinking zone.
354 */
355 pfn = find_biggest_section_pfn(nid, zone, zone->zone_start_pfn,
356 start_pfn);
357 if (pfn)
358 zone->spanned_pages = pfn - zone->zone_start_pfn + 1;
359 else {
360 zone->zone_start_pfn = 0;
361 zone->spanned_pages = 0;
362 }
363 }
364}
365
366static void update_pgdat_span(struct pglist_data *pgdat)
367{
368 unsigned long node_start_pfn = 0, node_end_pfn = 0;
369 struct zone *zone;
370
371 for (zone = pgdat->node_zones;
372 zone < pgdat->node_zones + MAX_NR_ZONES; zone++) {
373 unsigned long end_pfn = zone_end_pfn(zone);
374
375 /* No need to lock the zones, they can't change. */
376 if (!zone->spanned_pages)
377 continue;
378 if (!node_end_pfn) {
379 node_start_pfn = zone->zone_start_pfn;
380 node_end_pfn = end_pfn;
381 continue;
382 }
383
384 if (end_pfn > node_end_pfn)
385 node_end_pfn = end_pfn;
386 if (zone->zone_start_pfn < node_start_pfn)
387 node_start_pfn = zone->zone_start_pfn;
388 }
389
390 pgdat->node_start_pfn = node_start_pfn;
391 pgdat->node_spanned_pages = node_end_pfn - node_start_pfn;
392}
393
394void __ref remove_pfn_range_from_zone(struct zone *zone,
395 unsigned long start_pfn,
396 unsigned long nr_pages)
397{
398 const unsigned long end_pfn = start_pfn + nr_pages;
399 struct pglist_data *pgdat = zone->zone_pgdat;
400 unsigned long pfn, cur_nr_pages;
401
402 /* Poison struct pages because they are now uninitialized again. */
403 for (pfn = start_pfn; pfn < end_pfn; pfn += cur_nr_pages) {
404 cond_resched();
405
406 /* Select all remaining pages up to the next section boundary */
407 cur_nr_pages =
408 min(end_pfn - pfn, SECTION_ALIGN_UP(pfn + 1) - pfn);
409 page_init_poison(pfn_to_page(pfn),
410 sizeof(struct page) * cur_nr_pages);
411 }
412
413#ifdef CONFIG_ZONE_DEVICE
414 /*
415 * Zone shrinking code cannot properly deal with ZONE_DEVICE. So
416 * we will not try to shrink the zones - which is okay as
417 * set_zone_contiguous() cannot deal with ZONE_DEVICE either way.
418 */
419 if (zone_idx(zone) == ZONE_DEVICE)
420 return;
421#endif
422
423 clear_zone_contiguous(zone);
424
425 shrink_zone_span(zone, start_pfn, start_pfn + nr_pages);
426 update_pgdat_span(pgdat);
427
428 set_zone_contiguous(zone);
429}
430
431static void __remove_section(unsigned long pfn, unsigned long nr_pages,
432 unsigned long map_offset,
433 struct vmem_altmap *altmap)
434{
435 struct mem_section *ms = __pfn_to_section(pfn);
436
437 if (WARN_ON_ONCE(!valid_section(ms)))
438 return;
439
440 sparse_remove_section(ms, pfn, nr_pages, map_offset, altmap);
441}
442
443/**
444 * __remove_pages() - remove sections of pages
445 * @pfn: starting pageframe (must be aligned to start of a section)
446 * @nr_pages: number of pages to remove (must be multiple of section size)
447 * @altmap: alternative device page map or %NULL if default memmap is used
448 *
449 * Generic helper function to remove section mappings and sysfs entries
450 * for the section of the memory we are removing. Caller needs to make
451 * sure that pages are marked reserved and zones are adjust properly by
452 * calling offline_pages().
453 */
454void __remove_pages(unsigned long pfn, unsigned long nr_pages,
455 struct vmem_altmap *altmap)
456{
457 const unsigned long end_pfn = pfn + nr_pages;
458 unsigned long cur_nr_pages;
459 unsigned long map_offset = 0;
460
461 map_offset = vmem_altmap_offset(altmap);
462
463 if (check_pfn_span(pfn, nr_pages, "remove"))
464 return;
465
466 for (; pfn < end_pfn; pfn += cur_nr_pages) {
467 cond_resched();
468 /* Select all remaining pages up to the next section boundary */
469 cur_nr_pages = min(end_pfn - pfn,
470 SECTION_ALIGN_UP(pfn + 1) - pfn);
471 __remove_section(pfn, cur_nr_pages, map_offset, altmap);
472 map_offset = 0;
473 }
474}
475
476int set_online_page_callback(online_page_callback_t callback)
477{
478 int rc = -EINVAL;
479
480 get_online_mems();
481 mutex_lock(&online_page_callback_lock);
482
483 if (online_page_callback == generic_online_page) {
484 online_page_callback = callback;
485 rc = 0;
486 }
487
488 mutex_unlock(&online_page_callback_lock);
489 put_online_mems();
490
491 return rc;
492}
493EXPORT_SYMBOL_GPL(set_online_page_callback);
494
495int restore_online_page_callback(online_page_callback_t callback)
496{
497 int rc = -EINVAL;
498
499 get_online_mems();
500 mutex_lock(&online_page_callback_lock);
501
502 if (online_page_callback == callback) {
503 online_page_callback = generic_online_page;
504 rc = 0;
505 }
506
507 mutex_unlock(&online_page_callback_lock);
508 put_online_mems();
509
510 return rc;
511}
512EXPORT_SYMBOL_GPL(restore_online_page_callback);
513
514void generic_online_page(struct page *page, unsigned int order)
515{
516 /*
517 * Freeing the page with debug_pagealloc enabled will try to unmap it,
518 * so we should map it first. This is better than introducing a special
519 * case in page freeing fast path.
520 */
521 debug_pagealloc_map_pages(page, 1 << order);
522 __free_pages_core(page, order);
523 totalram_pages_add(1UL << order);
524#ifdef CONFIG_HIGHMEM
525 if (PageHighMem(page))
526 totalhigh_pages_add(1UL << order);
527#endif
528}
529EXPORT_SYMBOL_GPL(generic_online_page);
530
531static void online_pages_range(unsigned long start_pfn, unsigned long nr_pages)
532{
533 const unsigned long end_pfn = start_pfn + nr_pages;
534 unsigned long pfn;
535
536 /*
537 * Online the pages in MAX_ORDER - 1 aligned chunks. The callback might
538 * decide to not expose all pages to the buddy (e.g., expose them
539 * later). We account all pages as being online and belonging to this
540 * zone ("present").
541 * When using memmap_on_memory, the range might not be aligned to
542 * MAX_ORDER_NR_PAGES - 1, but pageblock aligned. __ffs() will detect
543 * this and the first chunk to online will be pageblock_nr_pages.
544 */
545 for (pfn = start_pfn; pfn < end_pfn;) {
546 int order = min(MAX_ORDER - 1UL, __ffs(pfn));
547
548 (*online_page_callback)(pfn_to_page(pfn), order);
549 pfn += (1UL << order);
550 }
551
552 /* mark all involved sections as online */
553 online_mem_sections(start_pfn, end_pfn);
554}
555
556/* check which state of node_states will be changed when online memory */
557static void node_states_check_changes_online(unsigned long nr_pages,
558 struct zone *zone, struct memory_notify *arg)
559{
560 int nid = zone_to_nid(zone);
561
562 arg->status_change_nid = NUMA_NO_NODE;
563 arg->status_change_nid_normal = NUMA_NO_NODE;
564 arg->status_change_nid_high = NUMA_NO_NODE;
565
566 if (!node_state(nid, N_MEMORY))
567 arg->status_change_nid = nid;
568 if (zone_idx(zone) <= ZONE_NORMAL && !node_state(nid, N_NORMAL_MEMORY))
569 arg->status_change_nid_normal = nid;
570#ifdef CONFIG_HIGHMEM
571 if (zone_idx(zone) <= ZONE_HIGHMEM && !node_state(nid, N_HIGH_MEMORY))
572 arg->status_change_nid_high = nid;
573#endif
574}
575
576static void node_states_set_node(int node, struct memory_notify *arg)
577{
578 if (arg->status_change_nid_normal >= 0)
579 node_set_state(node, N_NORMAL_MEMORY);
580
581 if (arg->status_change_nid_high >= 0)
582 node_set_state(node, N_HIGH_MEMORY);
583
584 if (arg->status_change_nid >= 0)
585 node_set_state(node, N_MEMORY);
586}
587
588static void __meminit resize_zone_range(struct zone *zone, unsigned long start_pfn,
589 unsigned long nr_pages)
590{
591 unsigned long old_end_pfn = zone_end_pfn(zone);
592
593 if (zone_is_empty(zone) || start_pfn < zone->zone_start_pfn)
594 zone->zone_start_pfn = start_pfn;
595
596 zone->spanned_pages = max(start_pfn + nr_pages, old_end_pfn) - zone->zone_start_pfn;
597}
598
599static void __meminit resize_pgdat_range(struct pglist_data *pgdat, unsigned long start_pfn,
600 unsigned long nr_pages)
601{
602 unsigned long old_end_pfn = pgdat_end_pfn(pgdat);
603
604 if (!pgdat->node_spanned_pages || start_pfn < pgdat->node_start_pfn)
605 pgdat->node_start_pfn = start_pfn;
606
607 pgdat->node_spanned_pages = max(start_pfn + nr_pages, old_end_pfn) - pgdat->node_start_pfn;
608
609}
610
611static void section_taint_zone_device(unsigned long pfn)
612{
613 struct mem_section *ms = __pfn_to_section(pfn);
614
615 ms->section_mem_map |= SECTION_TAINT_ZONE_DEVICE;
616}
617
618/*
619 * Associate the pfn range with the given zone, initializing the memmaps
620 * and resizing the pgdat/zone data to span the added pages. After this
621 * call, all affected pages are PG_reserved.
622 *
623 * All aligned pageblocks are initialized to the specified migratetype
624 * (usually MIGRATE_MOVABLE). Besides setting the migratetype, no related
625 * zone stats (e.g., nr_isolate_pageblock) are touched.
626 */
627void __ref move_pfn_range_to_zone(struct zone *zone, unsigned long start_pfn,
628 unsigned long nr_pages,
629 struct vmem_altmap *altmap, int migratetype)
630{
631 struct pglist_data *pgdat = zone->zone_pgdat;
632 int nid = pgdat->node_id;
633
634 clear_zone_contiguous(zone);
635
636 if (zone_is_empty(zone))
637 init_currently_empty_zone(zone, start_pfn, nr_pages);
638 resize_zone_range(zone, start_pfn, nr_pages);
639 resize_pgdat_range(pgdat, start_pfn, nr_pages);
640
641 /*
642 * Subsection population requires care in pfn_to_online_page().
643 * Set the taint to enable the slow path detection of
644 * ZONE_DEVICE pages in an otherwise ZONE_{NORMAL,MOVABLE}
645 * section.
646 */
647 if (zone_is_zone_device(zone)) {
648 if (!IS_ALIGNED(start_pfn, PAGES_PER_SECTION))
649 section_taint_zone_device(start_pfn);
650 if (!IS_ALIGNED(start_pfn + nr_pages, PAGES_PER_SECTION))
651 section_taint_zone_device(start_pfn + nr_pages);
652 }
653
654 /*
655 * TODO now we have a visible range of pages which are not associated
656 * with their zone properly. Not nice but set_pfnblock_flags_mask
657 * expects the zone spans the pfn range. All the pages in the range
658 * are reserved so nobody should be touching them so we should be safe
659 */
660 memmap_init_range(nr_pages, nid, zone_idx(zone), start_pfn, 0,
661 MEMINIT_HOTPLUG, altmap, migratetype);
662
663 set_zone_contiguous(zone);
664}
665
666/*
667 * Returns a default kernel memory zone for the given pfn range.
668 * If no kernel zone covers this pfn range it will automatically go
669 * to the ZONE_NORMAL.
670 */
671static struct zone *default_kernel_zone_for_pfn(int nid, unsigned long start_pfn,
672 unsigned long nr_pages)
673{
674 struct pglist_data *pgdat = NODE_DATA(nid);
675 int zid;
676
677 for (zid = 0; zid <= ZONE_NORMAL; zid++) {
678 struct zone *zone = &pgdat->node_zones[zid];
679
680 if (zone_intersects(zone, start_pfn, nr_pages))
681 return zone;
682 }
683
684 return &pgdat->node_zones[ZONE_NORMAL];
685}
686
687static inline struct zone *default_zone_for_pfn(int nid, unsigned long start_pfn,
688 unsigned long nr_pages)
689{
690 struct zone *kernel_zone = default_kernel_zone_for_pfn(nid, start_pfn,
691 nr_pages);
692 struct zone *movable_zone = &NODE_DATA(nid)->node_zones[ZONE_MOVABLE];
693 bool in_kernel = zone_intersects(kernel_zone, start_pfn, nr_pages);
694 bool in_movable = zone_intersects(movable_zone, start_pfn, nr_pages);
695
696 /*
697 * We inherit the existing zone in a simple case where zones do not
698 * overlap in the given range
699 */
700 if (in_kernel ^ in_movable)
701 return (in_kernel) ? kernel_zone : movable_zone;
702
703 /*
704 * If the range doesn't belong to any zone or two zones overlap in the
705 * given range then we use movable zone only if movable_node is
706 * enabled because we always online to a kernel zone by default.
707 */
708 return movable_node_enabled ? movable_zone : kernel_zone;
709}
710
711struct zone *zone_for_pfn_range(int online_type, int nid,
712 unsigned long start_pfn, unsigned long nr_pages)
713{
714 if (online_type == MMOP_ONLINE_KERNEL)
715 return default_kernel_zone_for_pfn(nid, start_pfn, nr_pages);
716
717 if (online_type == MMOP_ONLINE_MOVABLE)
718 return &NODE_DATA(nid)->node_zones[ZONE_MOVABLE];
719
720 return default_zone_for_pfn(nid, start_pfn, nr_pages);
721}
722
723/*
724 * This function should only be called by memory_block_{online,offline},
725 * and {online,offline}_pages.
726 */
727void adjust_present_page_count(struct zone *zone, long nr_pages)
728{
729 zone->present_pages += nr_pages;
730 zone->zone_pgdat->node_present_pages += nr_pages;
731}
732
733int mhp_init_memmap_on_memory(unsigned long pfn, unsigned long nr_pages,
734 struct zone *zone)
735{
736 unsigned long end_pfn = pfn + nr_pages;
737 int ret;
738
739 ret = kasan_add_zero_shadow(__va(PFN_PHYS(pfn)), PFN_PHYS(nr_pages));
740 if (ret)
741 return ret;
742
743 move_pfn_range_to_zone(zone, pfn, nr_pages, NULL, MIGRATE_UNMOVABLE);
744
745 /*
746 * It might be that the vmemmap_pages fully span sections. If that is
747 * the case, mark those sections online here as otherwise they will be
748 * left offline.
749 */
750 if (nr_pages >= PAGES_PER_SECTION)
751 online_mem_sections(pfn, ALIGN_DOWN(end_pfn, PAGES_PER_SECTION));
752
753 return ret;
754}
755
756void mhp_deinit_memmap_on_memory(unsigned long pfn, unsigned long nr_pages)
757{
758 unsigned long end_pfn = pfn + nr_pages;
759
760 /*
761 * It might be that the vmemmap_pages fully span sections. If that is
762 * the case, mark those sections offline here as otherwise they will be
763 * left online.
764 */
765 if (nr_pages >= PAGES_PER_SECTION)
766 offline_mem_sections(pfn, ALIGN_DOWN(end_pfn, PAGES_PER_SECTION));
767
768 /*
769 * The pages associated with this vmemmap have been offlined, so
770 * we can reset its state here.
771 */
772 remove_pfn_range_from_zone(page_zone(pfn_to_page(pfn)), pfn, nr_pages);
773 kasan_remove_zero_shadow(__va(PFN_PHYS(pfn)), PFN_PHYS(nr_pages));
774}
775
776int __ref online_pages(unsigned long pfn, unsigned long nr_pages, struct zone *zone)
777{
778 unsigned long flags;
779 int need_zonelists_rebuild = 0;
780 const int nid = zone_to_nid(zone);
781 int ret;
782 struct memory_notify arg;
783
784 /*
785 * {on,off}lining is constrained to full memory sections (or more
786 * precisely to memory blocks from the user space POV).
787 * memmap_on_memory is an exception because it reserves initial part
788 * of the physical memory space for vmemmaps. That space is pageblock
789 * aligned.
790 */
791 if (WARN_ON_ONCE(!nr_pages ||
792 !IS_ALIGNED(pfn, pageblock_nr_pages) ||
793 !IS_ALIGNED(pfn + nr_pages, PAGES_PER_SECTION)))
794 return -EINVAL;
795
796 mem_hotplug_begin();
797
798 /* associate pfn range with the zone */
799 move_pfn_range_to_zone(zone, pfn, nr_pages, NULL, MIGRATE_ISOLATE);
800
801 arg.start_pfn = pfn;
802 arg.nr_pages = nr_pages;
803 node_states_check_changes_online(nr_pages, zone, &arg);
804
805 ret = memory_notify(MEM_GOING_ONLINE, &arg);
806 ret = notifier_to_errno(ret);
807 if (ret)
808 goto failed_addition;
809
810 /*
811 * Fixup the number of isolated pageblocks before marking the sections
812 * onlining, such that undo_isolate_page_range() works correctly.
813 */
814 spin_lock_irqsave(&zone->lock, flags);
815 zone->nr_isolate_pageblock += nr_pages / pageblock_nr_pages;
816 spin_unlock_irqrestore(&zone->lock, flags);
817
818 /*
819 * If this zone is not populated, then it is not in zonelist.
820 * This means the page allocator ignores this zone.
821 * So, zonelist must be updated after online.
822 */
823 if (!populated_zone(zone)) {
824 need_zonelists_rebuild = 1;
825 setup_zone_pageset(zone);
826 }
827
828 online_pages_range(pfn, nr_pages);
829 adjust_present_page_count(zone, nr_pages);
830
831 node_states_set_node(nid, &arg);
832 if (need_zonelists_rebuild)
833 build_all_zonelists(NULL);
834
835 /* Basic onlining is complete, allow allocation of onlined pages. */
836 undo_isolate_page_range(pfn, pfn + nr_pages, MIGRATE_MOVABLE);
837
838 /*
839 * Freshly onlined pages aren't shuffled (e.g., all pages are placed to
840 * the tail of the freelist when undoing isolation). Shuffle the whole
841 * zone to make sure the just onlined pages are properly distributed
842 * across the whole freelist - to create an initial shuffle.
843 */
844 shuffle_zone(zone);
845
846 /* reinitialise watermarks and update pcp limits */
847 init_per_zone_wmark_min();
848
849 kswapd_run(nid);
850 kcompactd_run(nid);
851
852 writeback_set_ratelimit();
853
854 memory_notify(MEM_ONLINE, &arg);
855 mem_hotplug_done();
856 return 0;
857
858failed_addition:
859 pr_debug("online_pages [mem %#010llx-%#010llx] failed\n",
860 (unsigned long long) pfn << PAGE_SHIFT,
861 (((unsigned long long) pfn + nr_pages) << PAGE_SHIFT) - 1);
862 memory_notify(MEM_CANCEL_ONLINE, &arg);
863 remove_pfn_range_from_zone(zone, pfn, nr_pages);
864 mem_hotplug_done();
865 return ret;
866}
867#endif /* CONFIG_MEMORY_HOTPLUG_SPARSE */
868
869static void reset_node_present_pages(pg_data_t *pgdat)
870{
871 struct zone *z;
872
873 for (z = pgdat->node_zones; z < pgdat->node_zones + MAX_NR_ZONES; z++)
874 z->present_pages = 0;
875
876 pgdat->node_present_pages = 0;
877}
878
879/* we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG */
880static pg_data_t __ref *hotadd_new_pgdat(int nid)
881{
882 struct pglist_data *pgdat;
883
884 pgdat = NODE_DATA(nid);
885 if (!pgdat) {
886 pgdat = arch_alloc_nodedata(nid);
887 if (!pgdat)
888 return NULL;
889
890 pgdat->per_cpu_nodestats =
891 alloc_percpu(struct per_cpu_nodestat);
892 arch_refresh_nodedata(nid, pgdat);
893 } else {
894 int cpu;
895 /*
896 * Reset the nr_zones, order and highest_zoneidx before reuse.
897 * Note that kswapd will init kswapd_highest_zoneidx properly
898 * when it starts in the near future.
899 */
900 pgdat->nr_zones = 0;
901 pgdat->kswapd_order = 0;
902 pgdat->kswapd_highest_zoneidx = 0;
903 for_each_online_cpu(cpu) {
904 struct per_cpu_nodestat *p;
905
906 p = per_cpu_ptr(pgdat->per_cpu_nodestats, cpu);
907 memset(p, 0, sizeof(*p));
908 }
909 }
910
911 /* we can use NODE_DATA(nid) from here */
912 pgdat->node_id = nid;
913 pgdat->node_start_pfn = 0;
914
915 /* init node's zones as empty zones, we don't have any present pages.*/
916 free_area_init_core_hotplug(nid);
917
918 /*
919 * The node we allocated has no zone fallback lists. For avoiding
920 * to access not-initialized zonelist, build here.
921 */
922 build_all_zonelists(pgdat);
923
924 /*
925 * When memory is hot-added, all the memory is in offline state. So
926 * clear all zones' present_pages because they will be updated in
927 * online_pages() and offline_pages().
928 */
929 reset_node_managed_pages(pgdat);
930 reset_node_present_pages(pgdat);
931
932 return pgdat;
933}
934
935static void rollback_node_hotadd(int nid)
936{
937 pg_data_t *pgdat = NODE_DATA(nid);
938
939 arch_refresh_nodedata(nid, NULL);
940 free_percpu(pgdat->per_cpu_nodestats);
941 arch_free_nodedata(pgdat);
942}
943
944
945/*
946 * __try_online_node - online a node if offlined
947 * @nid: the node ID
948 * @set_node_online: Whether we want to online the node
949 * called by cpu_up() to online a node without onlined memory.
950 *
951 * Returns:
952 * 1 -> a new node has been allocated
953 * 0 -> the node is already online
954 * -ENOMEM -> the node could not be allocated
955 */
956static int __try_online_node(int nid, bool set_node_online)
957{
958 pg_data_t *pgdat;
959 int ret = 1;
960
961 if (node_online(nid))
962 return 0;
963
964 pgdat = hotadd_new_pgdat(nid);
965 if (!pgdat) {
966 pr_err("Cannot online node %d due to NULL pgdat\n", nid);
967 ret = -ENOMEM;
968 goto out;
969 }
970
971 if (set_node_online) {
972 node_set_online(nid);
973 ret = register_one_node(nid);
974 BUG_ON(ret);
975 }
976out:
977 return ret;
978}
979
980/*
981 * Users of this function always want to online/register the node
982 */
983int try_online_node(int nid)
984{
985 int ret;
986
987 mem_hotplug_begin();
988 ret = __try_online_node(nid, true);
989 mem_hotplug_done();
990 return ret;
991}
992
993static int check_hotplug_memory_range(u64 start, u64 size)
994{
995 /* memory range must be block size aligned */
996 if (!size || !IS_ALIGNED(start, memory_block_size_bytes()) ||
997 !IS_ALIGNED(size, memory_block_size_bytes())) {
998 pr_err("Block size [%#lx] unaligned hotplug range: start %#llx, size %#llx",
999 memory_block_size_bytes(), start, size);
1000 return -EINVAL;
1001 }
1002
1003 return 0;
1004}
1005
1006static int online_memory_block(struct memory_block *mem, void *arg)
1007{
1008 mem->online_type = mhp_default_online_type;
1009 return device_online(&mem->dev);
1010}
1011
1012bool mhp_supports_memmap_on_memory(unsigned long size)
1013{
1014 unsigned long nr_vmemmap_pages = size / PAGE_SIZE;
1015 unsigned long vmemmap_size = nr_vmemmap_pages * sizeof(struct page);
1016 unsigned long remaining_size = size - vmemmap_size;
1017
1018 /*
1019 * Besides having arch support and the feature enabled at runtime, we
1020 * need a few more assumptions to hold true:
1021 *
1022 * a) We span a single memory block: memory onlining/offlinin;g happens
1023 * in memory block granularity. We don't want the vmemmap of online
1024 * memory blocks to reside on offline memory blocks. In the future,
1025 * we might want to support variable-sized memory blocks to make the
1026 * feature more versatile.
1027 *
1028 * b) The vmemmap pages span complete PMDs: We don't want vmemmap code
1029 * to populate memory from the altmap for unrelated parts (i.e.,
1030 * other memory blocks)
1031 *
1032 * c) The vmemmap pages (and thereby the pages that will be exposed to
1033 * the buddy) have to cover full pageblocks: memory onlining/offlining
1034 * code requires applicable ranges to be page-aligned, for example, to
1035 * set the migratetypes properly.
1036 *
1037 * TODO: Although we have a check here to make sure that vmemmap pages
1038 * fully populate a PMD, it is not the right place to check for
1039 * this. A much better solution involves improving vmemmap code
1040 * to fallback to base pages when trying to populate vmemmap using
1041 * altmap as an alternative source of memory, and we do not exactly
1042 * populate a single PMD.
1043 */
1044 return memmap_on_memory &&
1045 !hugetlb_free_vmemmap_enabled &&
1046 IS_ENABLED(CONFIG_MHP_MEMMAP_ON_MEMORY) &&
1047 size == memory_block_size_bytes() &&
1048 IS_ALIGNED(vmemmap_size, PMD_SIZE) &&
1049 IS_ALIGNED(remaining_size, (pageblock_nr_pages << PAGE_SHIFT));
1050}
1051
1052/*
1053 * NOTE: The caller must call lock_device_hotplug() to serialize hotplug
1054 * and online/offline operations (triggered e.g. by sysfs).
1055 *
1056 * we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG
1057 */
1058int __ref add_memory_resource(int nid, struct resource *res, mhp_t mhp_flags)
1059{
1060 struct mhp_params params = { .pgprot = pgprot_mhp(PAGE_KERNEL) };
1061 struct vmem_altmap mhp_altmap = {};
1062 u64 start, size;
1063 bool new_node = false;
1064 int ret;
1065
1066 start = res->start;
1067 size = resource_size(res);
1068
1069 ret = check_hotplug_memory_range(start, size);
1070 if (ret)
1071 return ret;
1072
1073 if (!node_possible(nid)) {
1074 WARN(1, "node %d was absent from the node_possible_map\n", nid);
1075 return -EINVAL;
1076 }
1077
1078 mem_hotplug_begin();
1079
1080 if (IS_ENABLED(CONFIG_ARCH_KEEP_MEMBLOCK))
1081 memblock_add_node(start, size, nid);
1082
1083 ret = __try_online_node(nid, false);
1084 if (ret < 0)
1085 goto error;
1086 new_node = ret;
1087
1088 /*
1089 * Self hosted memmap array
1090 */
1091 if (mhp_flags & MHP_MEMMAP_ON_MEMORY) {
1092 if (!mhp_supports_memmap_on_memory(size)) {
1093 ret = -EINVAL;
1094 goto error;
1095 }
1096 mhp_altmap.free = PHYS_PFN(size);
1097 mhp_altmap.base_pfn = PHYS_PFN(start);
1098 params.altmap = &mhp_altmap;
1099 }
1100
1101 /* call arch's memory hotadd */
1102 ret = arch_add_memory(nid, start, size, ¶ms);
1103 if (ret < 0)
1104 goto error;
1105
1106 /* create memory block devices after memory was added */
1107 ret = create_memory_block_devices(start, size, mhp_altmap.alloc);
1108 if (ret) {
1109 arch_remove_memory(nid, start, size, NULL);
1110 goto error;
1111 }
1112
1113 if (new_node) {
1114 /* If sysfs file of new node can't be created, cpu on the node
1115 * can't be hot-added. There is no rollback way now.
1116 * So, check by BUG_ON() to catch it reluctantly..
1117 * We online node here. We can't roll back from here.
1118 */
1119 node_set_online(nid);
1120 ret = __register_one_node(nid);
1121 BUG_ON(ret);
1122 }
1123
1124 /* link memory sections under this node.*/
1125 link_mem_sections(nid, PFN_DOWN(start), PFN_UP(start + size - 1),
1126 MEMINIT_HOTPLUG);
1127
1128 /* create new memmap entry */
1129 if (!strcmp(res->name, "System RAM"))
1130 firmware_map_add_hotplug(start, start + size, "System RAM");
1131
1132 /* device_online() will take the lock when calling online_pages() */
1133 mem_hotplug_done();
1134
1135 /*
1136 * In case we're allowed to merge the resource, flag it and trigger
1137 * merging now that adding succeeded.
1138 */
1139 if (mhp_flags & MHP_MERGE_RESOURCE)
1140 merge_system_ram_resource(res);
1141
1142 /* online pages if requested */
1143 if (mhp_default_online_type != MMOP_OFFLINE)
1144 walk_memory_blocks(start, size, NULL, online_memory_block);
1145
1146 return ret;
1147error:
1148 /* rollback pgdat allocation and others */
1149 if (new_node)
1150 rollback_node_hotadd(nid);
1151 if (IS_ENABLED(CONFIG_ARCH_KEEP_MEMBLOCK))
1152 memblock_remove(start, size);
1153 mem_hotplug_done();
1154 return ret;
1155}
1156
1157/* requires device_hotplug_lock, see add_memory_resource() */
1158int __ref __add_memory(int nid, u64 start, u64 size, mhp_t mhp_flags)
1159{
1160 struct resource *res;
1161 int ret;
1162
1163 res = register_memory_resource(start, size, "System RAM");
1164 if (IS_ERR(res))
1165 return PTR_ERR(res);
1166
1167 ret = add_memory_resource(nid, res, mhp_flags);
1168 if (ret < 0)
1169 release_memory_resource(res);
1170 return ret;
1171}
1172
1173int add_memory(int nid, u64 start, u64 size, mhp_t mhp_flags)
1174{
1175 int rc;
1176
1177 lock_device_hotplug();
1178 rc = __add_memory(nid, start, size, mhp_flags);
1179 unlock_device_hotplug();
1180
1181 return rc;
1182}
1183EXPORT_SYMBOL_GPL(add_memory);
1184
1185/*
1186 * Add special, driver-managed memory to the system as system RAM. Such
1187 * memory is not exposed via the raw firmware-provided memmap as system
1188 * RAM, instead, it is detected and added by a driver - during cold boot,
1189 * after a reboot, and after kexec.
1190 *
1191 * Reasons why this memory should not be used for the initial memmap of a
1192 * kexec kernel or for placing kexec images:
1193 * - The booting kernel is in charge of determining how this memory will be
1194 * used (e.g., use persistent memory as system RAM)
1195 * - Coordination with a hypervisor is required before this memory
1196 * can be used (e.g., inaccessible parts).
1197 *
1198 * For this memory, no entries in /sys/firmware/memmap ("raw firmware-provided
1199 * memory map") are created. Also, the created memory resource is flagged
1200 * with IORESOURCE_SYSRAM_DRIVER_MANAGED, so in-kernel users can special-case
1201 * this memory as well (esp., not place kexec images onto it).
1202 *
1203 * The resource_name (visible via /proc/iomem) has to have the format
1204 * "System RAM ($DRIVER)".
1205 */
1206int add_memory_driver_managed(int nid, u64 start, u64 size,
1207 const char *resource_name, mhp_t mhp_flags)
1208{
1209 struct resource *res;
1210 int rc;
1211
1212 if (!resource_name ||
1213 strstr(resource_name, "System RAM (") != resource_name ||
1214 resource_name[strlen(resource_name) - 1] != ')')
1215 return -EINVAL;
1216
1217 lock_device_hotplug();
1218
1219 res = register_memory_resource(start, size, resource_name);
1220 if (IS_ERR(res)) {
1221 rc = PTR_ERR(res);
1222 goto out_unlock;
1223 }
1224
1225 rc = add_memory_resource(nid, res, mhp_flags);
1226 if (rc < 0)
1227 release_memory_resource(res);
1228
1229out_unlock:
1230 unlock_device_hotplug();
1231 return rc;
1232}
1233EXPORT_SYMBOL_GPL(add_memory_driver_managed);
1234
1235/*
1236 * Platforms should define arch_get_mappable_range() that provides
1237 * maximum possible addressable physical memory range for which the
1238 * linear mapping could be created. The platform returned address
1239 * range must adhere to these following semantics.
1240 *
1241 * - range.start <= range.end
1242 * - Range includes both end points [range.start..range.end]
1243 *
1244 * There is also a fallback definition provided here, allowing the
1245 * entire possible physical address range in case any platform does
1246 * not define arch_get_mappable_range().
1247 */
1248struct range __weak arch_get_mappable_range(void)
1249{
1250 struct range mhp_range = {
1251 .start = 0UL,
1252 .end = -1ULL,
1253 };
1254 return mhp_range;
1255}
1256
1257struct range mhp_get_pluggable_range(bool need_mapping)
1258{
1259 const u64 max_phys = (1ULL << MAX_PHYSMEM_BITS) - 1;
1260 struct range mhp_range;
1261
1262 if (need_mapping) {
1263 mhp_range = arch_get_mappable_range();
1264 if (mhp_range.start > max_phys) {
1265 mhp_range.start = 0;
1266 mhp_range.end = 0;
1267 }
1268 mhp_range.end = min_t(u64, mhp_range.end, max_phys);
1269 } else {
1270 mhp_range.start = 0;
1271 mhp_range.end = max_phys;
1272 }
1273 return mhp_range;
1274}
1275EXPORT_SYMBOL_GPL(mhp_get_pluggable_range);
1276
1277bool mhp_range_allowed(u64 start, u64 size, bool need_mapping)
1278{
1279 struct range mhp_range = mhp_get_pluggable_range(need_mapping);
1280 u64 end = start + size;
1281
1282 if (start < end && start >= mhp_range.start && (end - 1) <= mhp_range.end)
1283 return true;
1284
1285 pr_warn("Hotplug memory [%#llx-%#llx] exceeds maximum addressable range [%#llx-%#llx]\n",
1286 start, end, mhp_range.start, mhp_range.end);
1287 return false;
1288}
1289
1290#ifdef CONFIG_MEMORY_HOTREMOVE
1291/*
1292 * Confirm all pages in a range [start, end) belong to the same zone (skipping
1293 * memory holes). When true, return the zone.
1294 */
1295struct zone *test_pages_in_a_zone(unsigned long start_pfn,
1296 unsigned long end_pfn)
1297{
1298 unsigned long pfn, sec_end_pfn;
1299 struct zone *zone = NULL;
1300 struct page *page;
1301 int i;
1302 for (pfn = start_pfn, sec_end_pfn = SECTION_ALIGN_UP(start_pfn + 1);
1303 pfn < end_pfn;
1304 pfn = sec_end_pfn, sec_end_pfn += PAGES_PER_SECTION) {
1305 /* Make sure the memory section is present first */
1306 if (!present_section_nr(pfn_to_section_nr(pfn)))
1307 continue;
1308 for (; pfn < sec_end_pfn && pfn < end_pfn;
1309 pfn += MAX_ORDER_NR_PAGES) {
1310 i = 0;
1311 /* This is just a CONFIG_HOLES_IN_ZONE check.*/
1312 while ((i < MAX_ORDER_NR_PAGES) &&
1313 !pfn_valid_within(pfn + i))
1314 i++;
1315 if (i == MAX_ORDER_NR_PAGES || pfn + i >= end_pfn)
1316 continue;
1317 /* Check if we got outside of the zone */
1318 if (zone && !zone_spans_pfn(zone, pfn + i))
1319 return NULL;
1320 page = pfn_to_page(pfn + i);
1321 if (zone && page_zone(page) != zone)
1322 return NULL;
1323 zone = page_zone(page);
1324 }
1325 }
1326
1327 return zone;
1328}
1329
1330/*
1331 * Scan pfn range [start,end) to find movable/migratable pages (LRU pages,
1332 * non-lru movable pages and hugepages). Will skip over most unmovable
1333 * pages (esp., pages that can be skipped when offlining), but bail out on
1334 * definitely unmovable pages.
1335 *
1336 * Returns:
1337 * 0 in case a movable page is found and movable_pfn was updated.
1338 * -ENOENT in case no movable page was found.
1339 * -EBUSY in case a definitely unmovable page was found.
1340 */
1341static int scan_movable_pages(unsigned long start, unsigned long end,
1342 unsigned long *movable_pfn)
1343{
1344 unsigned long pfn;
1345
1346 for (pfn = start; pfn < end; pfn++) {
1347 struct page *page, *head;
1348 unsigned long skip;
1349
1350 if (!pfn_valid(pfn))
1351 continue;
1352 page = pfn_to_page(pfn);
1353 if (PageLRU(page))
1354 goto found;
1355 if (__PageMovable(page))
1356 goto found;
1357
1358 /*
1359 * PageOffline() pages that are not marked __PageMovable() and
1360 * have a reference count > 0 (after MEM_GOING_OFFLINE) are
1361 * definitely unmovable. If their reference count would be 0,
1362 * they could at least be skipped when offlining memory.
1363 */
1364 if (PageOffline(page) && page_count(page))
1365 return -EBUSY;
1366
1367 if (!PageHuge(page))
1368 continue;
1369 head = compound_head(page);
1370 /*
1371 * This test is racy as we hold no reference or lock. The
1372 * hugetlb page could have been free'ed and head is no longer
1373 * a hugetlb page before the following check. In such unlikely
1374 * cases false positives and negatives are possible. Calling
1375 * code must deal with these scenarios.
1376 */
1377 if (HPageMigratable(head))
1378 goto found;
1379 skip = compound_nr(head) - (page - head);
1380 pfn += skip - 1;
1381 }
1382 return -ENOENT;
1383found:
1384 *movable_pfn = pfn;
1385 return 0;
1386}
1387
1388static int
1389do_migrate_range(unsigned long start_pfn, unsigned long end_pfn)
1390{
1391 unsigned long pfn;
1392 struct page *page, *head;
1393 int ret = 0;
1394 LIST_HEAD(source);
1395 static DEFINE_RATELIMIT_STATE(migrate_rs, DEFAULT_RATELIMIT_INTERVAL,
1396 DEFAULT_RATELIMIT_BURST);
1397
1398 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
1399 if (!pfn_valid(pfn))
1400 continue;
1401 page = pfn_to_page(pfn);
1402 head = compound_head(page);
1403
1404 if (PageHuge(page)) {
1405 pfn = page_to_pfn(head) + compound_nr(head) - 1;
1406 isolate_huge_page(head, &source);
1407 continue;
1408 } else if (PageTransHuge(page))
1409 pfn = page_to_pfn(head) + thp_nr_pages(page) - 1;
1410
1411 /*
1412 * HWPoison pages have elevated reference counts so the migration would
1413 * fail on them. It also doesn't make any sense to migrate them in the
1414 * first place. Still try to unmap such a page in case it is still mapped
1415 * (e.g. current hwpoison implementation doesn't unmap KSM pages but keep
1416 * the unmap as the catch all safety net).
1417 */
1418 if (PageHWPoison(page)) {
1419 if (WARN_ON(PageLRU(page)))
1420 isolate_lru_page(page);
1421 if (page_mapped(page))
1422 try_to_unmap(page, TTU_IGNORE_MLOCK);
1423 continue;
1424 }
1425
1426 if (!get_page_unless_zero(page))
1427 continue;
1428 /*
1429 * We can skip free pages. And we can deal with pages on
1430 * LRU and non-lru movable pages.
1431 */
1432 if (PageLRU(page))
1433 ret = isolate_lru_page(page);
1434 else
1435 ret = isolate_movable_page(page, ISOLATE_UNEVICTABLE);
1436 if (!ret) { /* Success */
1437 list_add_tail(&page->lru, &source);
1438 if (!__PageMovable(page))
1439 inc_node_page_state(page, NR_ISOLATED_ANON +
1440 page_is_file_lru(page));
1441
1442 } else {
1443 if (__ratelimit(&migrate_rs)) {
1444 pr_warn("failed to isolate pfn %lx\n", pfn);
1445 dump_page(page, "isolation failed");
1446 }
1447 }
1448 put_page(page);
1449 }
1450 if (!list_empty(&source)) {
1451 nodemask_t nmask = node_states[N_MEMORY];
1452 struct migration_target_control mtc = {
1453 .nmask = &nmask,
1454 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
1455 };
1456
1457 /*
1458 * We have checked that migration range is on a single zone so
1459 * we can use the nid of the first page to all the others.
1460 */
1461 mtc.nid = page_to_nid(list_first_entry(&source, struct page, lru));
1462
1463 /*
1464 * try to allocate from a different node but reuse this node
1465 * if there are no other online nodes to be used (e.g. we are
1466 * offlining a part of the only existing node)
1467 */
1468 node_clear(mtc.nid, nmask);
1469 if (nodes_empty(nmask))
1470 node_set(mtc.nid, nmask);
1471 ret = migrate_pages(&source, alloc_migration_target, NULL,
1472 (unsigned long)&mtc, MIGRATE_SYNC, MR_MEMORY_HOTPLUG);
1473 if (ret) {
1474 list_for_each_entry(page, &source, lru) {
1475 if (__ratelimit(&migrate_rs)) {
1476 pr_warn("migrating pfn %lx failed ret:%d\n",
1477 page_to_pfn(page), ret);
1478 dump_page(page, "migration failure");
1479 }
1480 }
1481 putback_movable_pages(&source);
1482 }
1483 }
1484
1485 return ret;
1486}
1487
1488static int __init cmdline_parse_movable_node(char *p)
1489{
1490 movable_node_enabled = true;
1491 return 0;
1492}
1493early_param("movable_node", cmdline_parse_movable_node);
1494
1495/* check which state of node_states will be changed when offline memory */
1496static void node_states_check_changes_offline(unsigned long nr_pages,
1497 struct zone *zone, struct memory_notify *arg)
1498{
1499 struct pglist_data *pgdat = zone->zone_pgdat;
1500 unsigned long present_pages = 0;
1501 enum zone_type zt;
1502
1503 arg->status_change_nid = NUMA_NO_NODE;
1504 arg->status_change_nid_normal = NUMA_NO_NODE;
1505 arg->status_change_nid_high = NUMA_NO_NODE;
1506
1507 /*
1508 * Check whether node_states[N_NORMAL_MEMORY] will be changed.
1509 * If the memory to be offline is within the range
1510 * [0..ZONE_NORMAL], and it is the last present memory there,
1511 * the zones in that range will become empty after the offlining,
1512 * thus we can determine that we need to clear the node from
1513 * node_states[N_NORMAL_MEMORY].
1514 */
1515 for (zt = 0; zt <= ZONE_NORMAL; zt++)
1516 present_pages += pgdat->node_zones[zt].present_pages;
1517 if (zone_idx(zone) <= ZONE_NORMAL && nr_pages >= present_pages)
1518 arg->status_change_nid_normal = zone_to_nid(zone);
1519
1520#ifdef CONFIG_HIGHMEM
1521 /*
1522 * node_states[N_HIGH_MEMORY] contains nodes which
1523 * have normal memory or high memory.
1524 * Here we add the present_pages belonging to ZONE_HIGHMEM.
1525 * If the zone is within the range of [0..ZONE_HIGHMEM), and
1526 * we determine that the zones in that range become empty,
1527 * we need to clear the node for N_HIGH_MEMORY.
1528 */
1529 present_pages += pgdat->node_zones[ZONE_HIGHMEM].present_pages;
1530 if (zone_idx(zone) <= ZONE_HIGHMEM && nr_pages >= present_pages)
1531 arg->status_change_nid_high = zone_to_nid(zone);
1532#endif
1533
1534 /*
1535 * We have accounted the pages from [0..ZONE_NORMAL), and
1536 * in case of CONFIG_HIGHMEM the pages from ZONE_HIGHMEM
1537 * as well.
1538 * Here we count the possible pages from ZONE_MOVABLE.
1539 * If after having accounted all the pages, we see that the nr_pages
1540 * to be offlined is over or equal to the accounted pages,
1541 * we know that the node will become empty, and so, we can clear
1542 * it for N_MEMORY as well.
1543 */
1544 present_pages += pgdat->node_zones[ZONE_MOVABLE].present_pages;
1545
1546 if (nr_pages >= present_pages)
1547 arg->status_change_nid = zone_to_nid(zone);
1548}
1549
1550static void node_states_clear_node(int node, struct memory_notify *arg)
1551{
1552 if (arg->status_change_nid_normal >= 0)
1553 node_clear_state(node, N_NORMAL_MEMORY);
1554
1555 if (arg->status_change_nid_high >= 0)
1556 node_clear_state(node, N_HIGH_MEMORY);
1557
1558 if (arg->status_change_nid >= 0)
1559 node_clear_state(node, N_MEMORY);
1560}
1561
1562static int count_system_ram_pages_cb(unsigned long start_pfn,
1563 unsigned long nr_pages, void *data)
1564{
1565 unsigned long *nr_system_ram_pages = data;
1566
1567 *nr_system_ram_pages += nr_pages;
1568 return 0;
1569}
1570
1571int __ref offline_pages(unsigned long start_pfn, unsigned long nr_pages)
1572{
1573 const unsigned long end_pfn = start_pfn + nr_pages;
1574 unsigned long pfn, system_ram_pages = 0;
1575 unsigned long flags;
1576 struct zone *zone;
1577 struct memory_notify arg;
1578 int ret, node;
1579 char *reason;
1580
1581 /*
1582 * {on,off}lining is constrained to full memory sections (or more
1583 * precisely to memory blocks from the user space POV).
1584 * memmap_on_memory is an exception because it reserves initial part
1585 * of the physical memory space for vmemmaps. That space is pageblock
1586 * aligned.
1587 */
1588 if (WARN_ON_ONCE(!nr_pages ||
1589 !IS_ALIGNED(start_pfn, pageblock_nr_pages) ||
1590 !IS_ALIGNED(start_pfn + nr_pages, PAGES_PER_SECTION)))
1591 return -EINVAL;
1592
1593 mem_hotplug_begin();
1594
1595 /*
1596 * Don't allow to offline memory blocks that contain holes.
1597 * Consequently, memory blocks with holes can never get onlined
1598 * via the hotplug path - online_pages() - as hotplugged memory has
1599 * no holes. This way, we e.g., don't have to worry about marking
1600 * memory holes PG_reserved, don't need pfn_valid() checks, and can
1601 * avoid using walk_system_ram_range() later.
1602 */
1603 walk_system_ram_range(start_pfn, nr_pages, &system_ram_pages,
1604 count_system_ram_pages_cb);
1605 if (system_ram_pages != nr_pages) {
1606 ret = -EINVAL;
1607 reason = "memory holes";
1608 goto failed_removal;
1609 }
1610
1611 /* This makes hotplug much easier...and readable.
1612 we assume this for now. .*/
1613 zone = test_pages_in_a_zone(start_pfn, end_pfn);
1614 if (!zone) {
1615 ret = -EINVAL;
1616 reason = "multizone range";
1617 goto failed_removal;
1618 }
1619 node = zone_to_nid(zone);
1620
1621 /*
1622 * Disable pcplists so that page isolation cannot race with freeing
1623 * in a way that pages from isolated pageblock are left on pcplists.
1624 */
1625 zone_pcp_disable(zone);
1626 lru_cache_disable();
1627
1628 /* set above range as isolated */
1629 ret = start_isolate_page_range(start_pfn, end_pfn,
1630 MIGRATE_MOVABLE,
1631 MEMORY_OFFLINE | REPORT_FAILURE);
1632 if (ret) {
1633 reason = "failure to isolate range";
1634 goto failed_removal_pcplists_disabled;
1635 }
1636
1637 arg.start_pfn = start_pfn;
1638 arg.nr_pages = nr_pages;
1639 node_states_check_changes_offline(nr_pages, zone, &arg);
1640
1641 ret = memory_notify(MEM_GOING_OFFLINE, &arg);
1642 ret = notifier_to_errno(ret);
1643 if (ret) {
1644 reason = "notifier failure";
1645 goto failed_removal_isolated;
1646 }
1647
1648 do {
1649 pfn = start_pfn;
1650 do {
1651 if (signal_pending(current)) {
1652 ret = -EINTR;
1653 reason = "signal backoff";
1654 goto failed_removal_isolated;
1655 }
1656
1657 cond_resched();
1658
1659 ret = scan_movable_pages(pfn, end_pfn, &pfn);
1660 if (!ret) {
1661 /*
1662 * TODO: fatal migration failures should bail
1663 * out
1664 */
1665 do_migrate_range(pfn, end_pfn);
1666 }
1667 } while (!ret);
1668
1669 if (ret != -ENOENT) {
1670 reason = "unmovable page";
1671 goto failed_removal_isolated;
1672 }
1673
1674 /*
1675 * Dissolve free hugepages in the memory block before doing
1676 * offlining actually in order to make hugetlbfs's object
1677 * counting consistent.
1678 */
1679 ret = dissolve_free_huge_pages(start_pfn, end_pfn);
1680 if (ret) {
1681 reason = "failure to dissolve huge pages";
1682 goto failed_removal_isolated;
1683 }
1684
1685 ret = test_pages_isolated(start_pfn, end_pfn, MEMORY_OFFLINE);
1686
1687 } while (ret);
1688
1689 /* Mark all sections offline and remove free pages from the buddy. */
1690 __offline_isolated_pages(start_pfn, end_pfn);
1691 pr_debug("Offlined Pages %ld\n", nr_pages);
1692
1693 /*
1694 * The memory sections are marked offline, and the pageblock flags
1695 * effectively stale; nobody should be touching them. Fixup the number
1696 * of isolated pageblocks, memory onlining will properly revert this.
1697 */
1698 spin_lock_irqsave(&zone->lock, flags);
1699 zone->nr_isolate_pageblock -= nr_pages / pageblock_nr_pages;
1700 spin_unlock_irqrestore(&zone->lock, flags);
1701
1702 lru_cache_enable();
1703 zone_pcp_enable(zone);
1704
1705 /* removal success */
1706 adjust_managed_page_count(pfn_to_page(start_pfn), -nr_pages);
1707 adjust_present_page_count(zone, -nr_pages);
1708
1709 /* reinitialise watermarks and update pcp limits */
1710 init_per_zone_wmark_min();
1711
1712 if (!populated_zone(zone)) {
1713 zone_pcp_reset(zone);
1714 build_all_zonelists(NULL);
1715 }
1716
1717 node_states_clear_node(node, &arg);
1718 if (arg.status_change_nid >= 0) {
1719 kswapd_stop(node);
1720 kcompactd_stop(node);
1721 }
1722
1723 writeback_set_ratelimit();
1724
1725 memory_notify(MEM_OFFLINE, &arg);
1726 remove_pfn_range_from_zone(zone, start_pfn, nr_pages);
1727 mem_hotplug_done();
1728 return 0;
1729
1730failed_removal_isolated:
1731 undo_isolate_page_range(start_pfn, end_pfn, MIGRATE_MOVABLE);
1732 memory_notify(MEM_CANCEL_OFFLINE, &arg);
1733failed_removal_pcplists_disabled:
1734 lru_cache_enable();
1735 zone_pcp_enable(zone);
1736failed_removal:
1737 pr_debug("memory offlining [mem %#010llx-%#010llx] failed due to %s\n",
1738 (unsigned long long) start_pfn << PAGE_SHIFT,
1739 ((unsigned long long) end_pfn << PAGE_SHIFT) - 1,
1740 reason);
1741 /* pushback to free area */
1742 mem_hotplug_done();
1743 return ret;
1744}
1745
1746static int check_memblock_offlined_cb(struct memory_block *mem, void *arg)
1747{
1748 int ret = !is_memblock_offlined(mem);
1749
1750 if (unlikely(ret)) {
1751 phys_addr_t beginpa, endpa;
1752
1753 beginpa = PFN_PHYS(section_nr_to_pfn(mem->start_section_nr));
1754 endpa = beginpa + memory_block_size_bytes() - 1;
1755 pr_warn("removing memory fails, because memory [%pa-%pa] is onlined\n",
1756 &beginpa, &endpa);
1757
1758 return -EBUSY;
1759 }
1760 return 0;
1761}
1762
1763static int get_nr_vmemmap_pages_cb(struct memory_block *mem, void *arg)
1764{
1765 /*
1766 * If not set, continue with the next block.
1767 */
1768 return mem->nr_vmemmap_pages;
1769}
1770
1771static int check_cpu_on_node(pg_data_t *pgdat)
1772{
1773 int cpu;
1774
1775 for_each_present_cpu(cpu) {
1776 if (cpu_to_node(cpu) == pgdat->node_id)
1777 /*
1778 * the cpu on this node isn't removed, and we can't
1779 * offline this node.
1780 */
1781 return -EBUSY;
1782 }
1783
1784 return 0;
1785}
1786
1787static int check_no_memblock_for_node_cb(struct memory_block *mem, void *arg)
1788{
1789 int nid = *(int *)arg;
1790
1791 /*
1792 * If a memory block belongs to multiple nodes, the stored nid is not
1793 * reliable. However, such blocks are always online (e.g., cannot get
1794 * offlined) and, therefore, are still spanned by the node.
1795 */
1796 return mem->nid == nid ? -EEXIST : 0;
1797}
1798
1799/**
1800 * try_offline_node
1801 * @nid: the node ID
1802 *
1803 * Offline a node if all memory sections and cpus of the node are removed.
1804 *
1805 * NOTE: The caller must call lock_device_hotplug() to serialize hotplug
1806 * and online/offline operations before this call.
1807 */
1808void try_offline_node(int nid)
1809{
1810 pg_data_t *pgdat = NODE_DATA(nid);
1811 int rc;
1812
1813 /*
1814 * If the node still spans pages (especially ZONE_DEVICE), don't
1815 * offline it. A node spans memory after move_pfn_range_to_zone(),
1816 * e.g., after the memory block was onlined.
1817 */
1818 if (pgdat->node_spanned_pages)
1819 return;
1820
1821 /*
1822 * Especially offline memory blocks might not be spanned by the
1823 * node. They will get spanned by the node once they get onlined.
1824 * However, they link to the node in sysfs and can get onlined later.
1825 */
1826 rc = for_each_memory_block(&nid, check_no_memblock_for_node_cb);
1827 if (rc)
1828 return;
1829
1830 if (check_cpu_on_node(pgdat))
1831 return;
1832
1833 /*
1834 * all memory/cpu of this node are removed, we can offline this
1835 * node now.
1836 */
1837 node_set_offline(nid);
1838 unregister_one_node(nid);
1839}
1840EXPORT_SYMBOL(try_offline_node);
1841
1842static int __ref try_remove_memory(int nid, u64 start, u64 size)
1843{
1844 int rc = 0;
1845 struct vmem_altmap mhp_altmap = {};
1846 struct vmem_altmap *altmap = NULL;
1847 unsigned long nr_vmemmap_pages;
1848
1849 BUG_ON(check_hotplug_memory_range(start, size));
1850
1851 /*
1852 * All memory blocks must be offlined before removing memory. Check
1853 * whether all memory blocks in question are offline and return error
1854 * if this is not the case.
1855 */
1856 rc = walk_memory_blocks(start, size, NULL, check_memblock_offlined_cb);
1857 if (rc)
1858 return rc;
1859
1860 /*
1861 * We only support removing memory added with MHP_MEMMAP_ON_MEMORY in
1862 * the same granularity it was added - a single memory block.
1863 */
1864 if (memmap_on_memory) {
1865 nr_vmemmap_pages = walk_memory_blocks(start, size, NULL,
1866 get_nr_vmemmap_pages_cb);
1867 if (nr_vmemmap_pages) {
1868 if (size != memory_block_size_bytes()) {
1869 pr_warn("Refuse to remove %#llx - %#llx,"
1870 "wrong granularity\n",
1871 start, start + size);
1872 return -EINVAL;
1873 }
1874
1875 /*
1876 * Let remove_pmd_table->free_hugepage_table do the
1877 * right thing if we used vmem_altmap when hot-adding
1878 * the range.
1879 */
1880 mhp_altmap.alloc = nr_vmemmap_pages;
1881 altmap = &mhp_altmap;
1882 }
1883 }
1884
1885 /* remove memmap entry */
1886 firmware_map_remove(start, start + size, "System RAM");
1887
1888 /*
1889 * Memory block device removal under the device_hotplug_lock is
1890 * a barrier against racing online attempts.
1891 */
1892 remove_memory_block_devices(start, size);
1893
1894 mem_hotplug_begin();
1895
1896 arch_remove_memory(nid, start, size, altmap);
1897
1898 if (IS_ENABLED(CONFIG_ARCH_KEEP_MEMBLOCK)) {
1899 memblock_free(start, size);
1900 memblock_remove(start, size);
1901 }
1902
1903 release_mem_region_adjustable(start, size);
1904
1905 try_offline_node(nid);
1906
1907 mem_hotplug_done();
1908 return 0;
1909}
1910
1911/**
1912 * __remove_memory - Remove memory if every memory block is offline
1913 * @nid: the node ID
1914 * @start: physical address of the region to remove
1915 * @size: size of the region to remove
1916 *
1917 * NOTE: The caller must call lock_device_hotplug() to serialize hotplug
1918 * and online/offline operations before this call, as required by
1919 * try_offline_node().
1920 */
1921void __remove_memory(int nid, u64 start, u64 size)
1922{
1923
1924 /*
1925 * trigger BUG() if some memory is not offlined prior to calling this
1926 * function
1927 */
1928 if (try_remove_memory(nid, start, size))
1929 BUG();
1930}
1931
1932/*
1933 * Remove memory if every memory block is offline, otherwise return -EBUSY is
1934 * some memory is not offline
1935 */
1936int remove_memory(int nid, u64 start, u64 size)
1937{
1938 int rc;
1939
1940 lock_device_hotplug();
1941 rc = try_remove_memory(nid, start, size);
1942 unlock_device_hotplug();
1943
1944 return rc;
1945}
1946EXPORT_SYMBOL_GPL(remove_memory);
1947
1948static int try_offline_memory_block(struct memory_block *mem, void *arg)
1949{
1950 uint8_t online_type = MMOP_ONLINE_KERNEL;
1951 uint8_t **online_types = arg;
1952 struct page *page;
1953 int rc;
1954
1955 /*
1956 * Sense the online_type via the zone of the memory block. Offlining
1957 * with multiple zones within one memory block will be rejected
1958 * by offlining code ... so we don't care about that.
1959 */
1960 page = pfn_to_online_page(section_nr_to_pfn(mem->start_section_nr));
1961 if (page && zone_idx(page_zone(page)) == ZONE_MOVABLE)
1962 online_type = MMOP_ONLINE_MOVABLE;
1963
1964 rc = device_offline(&mem->dev);
1965 /*
1966 * Default is MMOP_OFFLINE - change it only if offlining succeeded,
1967 * so try_reonline_memory_block() can do the right thing.
1968 */
1969 if (!rc)
1970 **online_types = online_type;
1971
1972 (*online_types)++;
1973 /* Ignore if already offline. */
1974 return rc < 0 ? rc : 0;
1975}
1976
1977static int try_reonline_memory_block(struct memory_block *mem, void *arg)
1978{
1979 uint8_t **online_types = arg;
1980 int rc;
1981
1982 if (**online_types != MMOP_OFFLINE) {
1983 mem->online_type = **online_types;
1984 rc = device_online(&mem->dev);
1985 if (rc < 0)
1986 pr_warn("%s: Failed to re-online memory: %d",
1987 __func__, rc);
1988 }
1989
1990 /* Continue processing all remaining memory blocks. */
1991 (*online_types)++;
1992 return 0;
1993}
1994
1995/*
1996 * Try to offline and remove memory. Might take a long time to finish in case
1997 * memory is still in use. Primarily useful for memory devices that logically
1998 * unplugged all memory (so it's no longer in use) and want to offline + remove
1999 * that memory.
2000 */
2001int offline_and_remove_memory(int nid, u64 start, u64 size)
2002{
2003 const unsigned long mb_count = size / memory_block_size_bytes();
2004 uint8_t *online_types, *tmp;
2005 int rc;
2006
2007 if (!IS_ALIGNED(start, memory_block_size_bytes()) ||
2008 !IS_ALIGNED(size, memory_block_size_bytes()) || !size)
2009 return -EINVAL;
2010
2011 /*
2012 * We'll remember the old online type of each memory block, so we can
2013 * try to revert whatever we did when offlining one memory block fails
2014 * after offlining some others succeeded.
2015 */
2016 online_types = kmalloc_array(mb_count, sizeof(*online_types),
2017 GFP_KERNEL);
2018 if (!online_types)
2019 return -ENOMEM;
2020 /*
2021 * Initialize all states to MMOP_OFFLINE, so when we abort processing in
2022 * try_offline_memory_block(), we'll skip all unprocessed blocks in
2023 * try_reonline_memory_block().
2024 */
2025 memset(online_types, MMOP_OFFLINE, mb_count);
2026
2027 lock_device_hotplug();
2028
2029 tmp = online_types;
2030 rc = walk_memory_blocks(start, size, &tmp, try_offline_memory_block);
2031
2032 /*
2033 * In case we succeeded to offline all memory, remove it.
2034 * This cannot fail as it cannot get onlined in the meantime.
2035 */
2036 if (!rc) {
2037 rc = try_remove_memory(nid, start, size);
2038 if (rc)
2039 pr_err("%s: Failed to remove memory: %d", __func__, rc);
2040 }
2041
2042 /*
2043 * Rollback what we did. While memory onlining might theoretically fail
2044 * (nacked by a notifier), it barely ever happens.
2045 */
2046 if (rc) {
2047 tmp = online_types;
2048 walk_memory_blocks(start, size, &tmp,
2049 try_reonline_memory_block);
2050 }
2051 unlock_device_hotplug();
2052
2053 kfree(online_types);
2054 return rc;
2055}
2056EXPORT_SYMBOL_GPL(offline_and_remove_memory);
2057#endif /* CONFIG_MEMORY_HOTREMOVE */