Loading...
1/*
2 * linux/mm/memory_hotplug.c
3 *
4 * Copyright (C)
5 */
6
7#include <linux/stddef.h>
8#include <linux/mm.h>
9#include <linux/swap.h>
10#include <linux/interrupt.h>
11#include <linux/pagemap.h>
12#include <linux/bootmem.h>
13#include <linux/compiler.h>
14#include <linux/module.h>
15#include <linux/pagevec.h>
16#include <linux/writeback.h>
17#include <linux/slab.h>
18#include <linux/sysctl.h>
19#include <linux/cpu.h>
20#include <linux/memory.h>
21#include <linux/memory_hotplug.h>
22#include <linux/highmem.h>
23#include <linux/vmalloc.h>
24#include <linux/ioport.h>
25#include <linux/delay.h>
26#include <linux/migrate.h>
27#include <linux/page-isolation.h>
28#include <linux/pfn.h>
29#include <linux/suspend.h>
30#include <linux/mm_inline.h>
31#include <linux/firmware-map.h>
32
33#include <asm/tlbflush.h>
34
35#include "internal.h"
36
37/*
38 * online_page_callback contains pointer to current page onlining function.
39 * Initially it is generic_online_page(). If it is required it could be
40 * changed by calling set_online_page_callback() for callback registration
41 * and restore_online_page_callback() for generic callback restore.
42 */
43
44static void generic_online_page(struct page *page);
45
46static online_page_callback_t online_page_callback = generic_online_page;
47
48DEFINE_MUTEX(mem_hotplug_mutex);
49
50void lock_memory_hotplug(void)
51{
52 mutex_lock(&mem_hotplug_mutex);
53
54 /* for exclusive hibernation if CONFIG_HIBERNATION=y */
55 lock_system_sleep();
56}
57
58void unlock_memory_hotplug(void)
59{
60 unlock_system_sleep();
61 mutex_unlock(&mem_hotplug_mutex);
62}
63
64
65/* add this memory to iomem resource */
66static struct resource *register_memory_resource(u64 start, u64 size)
67{
68 struct resource *res;
69 res = kzalloc(sizeof(struct resource), GFP_KERNEL);
70 BUG_ON(!res);
71
72 res->name = "System RAM";
73 res->start = start;
74 res->end = start + size - 1;
75 res->flags = IORESOURCE_MEM | IORESOURCE_BUSY;
76 if (request_resource(&iomem_resource, res) < 0) {
77 printk("System RAM resource %llx - %llx cannot be added\n",
78 (unsigned long long)res->start, (unsigned long long)res->end);
79 kfree(res);
80 res = NULL;
81 }
82 return res;
83}
84
85static void release_memory_resource(struct resource *res)
86{
87 if (!res)
88 return;
89 release_resource(res);
90 kfree(res);
91 return;
92}
93
94#ifdef CONFIG_MEMORY_HOTPLUG_SPARSE
95#ifndef CONFIG_SPARSEMEM_VMEMMAP
96static void get_page_bootmem(unsigned long info, struct page *page,
97 unsigned long type)
98{
99 page->lru.next = (struct list_head *) type;
100 SetPagePrivate(page);
101 set_page_private(page, info);
102 atomic_inc(&page->_count);
103}
104
105/* reference to __meminit __free_pages_bootmem is valid
106 * so use __ref to tell modpost not to generate a warning */
107void __ref put_page_bootmem(struct page *page)
108{
109 unsigned long type;
110
111 type = (unsigned long) page->lru.next;
112 BUG_ON(type < MEMORY_HOTPLUG_MIN_BOOTMEM_TYPE ||
113 type > MEMORY_HOTPLUG_MAX_BOOTMEM_TYPE);
114
115 if (atomic_dec_return(&page->_count) == 1) {
116 ClearPagePrivate(page);
117 set_page_private(page, 0);
118 INIT_LIST_HEAD(&page->lru);
119 __free_pages_bootmem(page, 0);
120 }
121
122}
123
124static void register_page_bootmem_info_section(unsigned long start_pfn)
125{
126 unsigned long *usemap, mapsize, section_nr, i;
127 struct mem_section *ms;
128 struct page *page, *memmap;
129
130 if (!pfn_valid(start_pfn))
131 return;
132
133 section_nr = pfn_to_section_nr(start_pfn);
134 ms = __nr_to_section(section_nr);
135
136 /* Get section's memmap address */
137 memmap = sparse_decode_mem_map(ms->section_mem_map, section_nr);
138
139 /*
140 * Get page for the memmap's phys address
141 * XXX: need more consideration for sparse_vmemmap...
142 */
143 page = virt_to_page(memmap);
144 mapsize = sizeof(struct page) * PAGES_PER_SECTION;
145 mapsize = PAGE_ALIGN(mapsize) >> PAGE_SHIFT;
146
147 /* remember memmap's page */
148 for (i = 0; i < mapsize; i++, page++)
149 get_page_bootmem(section_nr, page, SECTION_INFO);
150
151 usemap = __nr_to_section(section_nr)->pageblock_flags;
152 page = virt_to_page(usemap);
153
154 mapsize = PAGE_ALIGN(usemap_size()) >> PAGE_SHIFT;
155
156 for (i = 0; i < mapsize; i++, page++)
157 get_page_bootmem(section_nr, page, MIX_SECTION_INFO);
158
159}
160
161void register_page_bootmem_info_node(struct pglist_data *pgdat)
162{
163 unsigned long i, pfn, end_pfn, nr_pages;
164 int node = pgdat->node_id;
165 struct page *page;
166 struct zone *zone;
167
168 nr_pages = PAGE_ALIGN(sizeof(struct pglist_data)) >> PAGE_SHIFT;
169 page = virt_to_page(pgdat);
170
171 for (i = 0; i < nr_pages; i++, page++)
172 get_page_bootmem(node, page, NODE_INFO);
173
174 zone = &pgdat->node_zones[0];
175 for (; zone < pgdat->node_zones + MAX_NR_ZONES - 1; zone++) {
176 if (zone->wait_table) {
177 nr_pages = zone->wait_table_hash_nr_entries
178 * sizeof(wait_queue_head_t);
179 nr_pages = PAGE_ALIGN(nr_pages) >> PAGE_SHIFT;
180 page = virt_to_page(zone->wait_table);
181
182 for (i = 0; i < nr_pages; i++, page++)
183 get_page_bootmem(node, page, NODE_INFO);
184 }
185 }
186
187 pfn = pgdat->node_start_pfn;
188 end_pfn = pfn + pgdat->node_spanned_pages;
189
190 /* register_section info */
191 for (; pfn < end_pfn; pfn += PAGES_PER_SECTION)
192 register_page_bootmem_info_section(pfn);
193
194}
195#endif /* !CONFIG_SPARSEMEM_VMEMMAP */
196
197static void grow_zone_span(struct zone *zone, unsigned long start_pfn,
198 unsigned long end_pfn)
199{
200 unsigned long old_zone_end_pfn;
201
202 zone_span_writelock(zone);
203
204 old_zone_end_pfn = zone->zone_start_pfn + zone->spanned_pages;
205 if (start_pfn < zone->zone_start_pfn)
206 zone->zone_start_pfn = start_pfn;
207
208 zone->spanned_pages = max(old_zone_end_pfn, end_pfn) -
209 zone->zone_start_pfn;
210
211 zone_span_writeunlock(zone);
212}
213
214static void grow_pgdat_span(struct pglist_data *pgdat, unsigned long start_pfn,
215 unsigned long end_pfn)
216{
217 unsigned long old_pgdat_end_pfn =
218 pgdat->node_start_pfn + pgdat->node_spanned_pages;
219
220 if (start_pfn < pgdat->node_start_pfn)
221 pgdat->node_start_pfn = start_pfn;
222
223 pgdat->node_spanned_pages = max(old_pgdat_end_pfn, end_pfn) -
224 pgdat->node_start_pfn;
225}
226
227static int __meminit __add_zone(struct zone *zone, unsigned long phys_start_pfn)
228{
229 struct pglist_data *pgdat = zone->zone_pgdat;
230 int nr_pages = PAGES_PER_SECTION;
231 int nid = pgdat->node_id;
232 int zone_type;
233 unsigned long flags;
234
235 zone_type = zone - pgdat->node_zones;
236 if (!zone->wait_table) {
237 int ret;
238
239 ret = init_currently_empty_zone(zone, phys_start_pfn,
240 nr_pages, MEMMAP_HOTPLUG);
241 if (ret)
242 return ret;
243 }
244 pgdat_resize_lock(zone->zone_pgdat, &flags);
245 grow_zone_span(zone, phys_start_pfn, phys_start_pfn + nr_pages);
246 grow_pgdat_span(zone->zone_pgdat, phys_start_pfn,
247 phys_start_pfn + nr_pages);
248 pgdat_resize_unlock(zone->zone_pgdat, &flags);
249 memmap_init_zone(nr_pages, nid, zone_type,
250 phys_start_pfn, MEMMAP_HOTPLUG);
251 return 0;
252}
253
254static int __meminit __add_section(int nid, struct zone *zone,
255 unsigned long phys_start_pfn)
256{
257 int nr_pages = PAGES_PER_SECTION;
258 int ret;
259
260 if (pfn_valid(phys_start_pfn))
261 return -EEXIST;
262
263 ret = sparse_add_one_section(zone, phys_start_pfn, nr_pages);
264
265 if (ret < 0)
266 return ret;
267
268 ret = __add_zone(zone, phys_start_pfn);
269
270 if (ret < 0)
271 return ret;
272
273 return register_new_memory(nid, __pfn_to_section(phys_start_pfn));
274}
275
276#ifdef CONFIG_SPARSEMEM_VMEMMAP
277static int __remove_section(struct zone *zone, struct mem_section *ms)
278{
279 /*
280 * XXX: Freeing memmap with vmemmap is not implement yet.
281 * This should be removed later.
282 */
283 return -EBUSY;
284}
285#else
286static int __remove_section(struct zone *zone, struct mem_section *ms)
287{
288 unsigned long flags;
289 struct pglist_data *pgdat = zone->zone_pgdat;
290 int ret = -EINVAL;
291
292 if (!valid_section(ms))
293 return ret;
294
295 ret = unregister_memory_section(ms);
296 if (ret)
297 return ret;
298
299 pgdat_resize_lock(pgdat, &flags);
300 sparse_remove_one_section(zone, ms);
301 pgdat_resize_unlock(pgdat, &flags);
302 return 0;
303}
304#endif
305
306/*
307 * Reasonably generic function for adding memory. It is
308 * expected that archs that support memory hotplug will
309 * call this function after deciding the zone to which to
310 * add the new pages.
311 */
312int __ref __add_pages(int nid, struct zone *zone, unsigned long phys_start_pfn,
313 unsigned long nr_pages)
314{
315 unsigned long i;
316 int err = 0;
317 int start_sec, end_sec;
318 /* during initialize mem_map, align hot-added range to section */
319 start_sec = pfn_to_section_nr(phys_start_pfn);
320 end_sec = pfn_to_section_nr(phys_start_pfn + nr_pages - 1);
321
322 for (i = start_sec; i <= end_sec; i++) {
323 err = __add_section(nid, zone, i << PFN_SECTION_SHIFT);
324
325 /*
326 * EEXIST is finally dealt with by ioresource collision
327 * check. see add_memory() => register_memory_resource()
328 * Warning will be printed if there is collision.
329 */
330 if (err && (err != -EEXIST))
331 break;
332 err = 0;
333 }
334
335 return err;
336}
337EXPORT_SYMBOL_GPL(__add_pages);
338
339/**
340 * __remove_pages() - remove sections of pages from a zone
341 * @zone: zone from which pages need to be removed
342 * @phys_start_pfn: starting pageframe (must be aligned to start of a section)
343 * @nr_pages: number of pages to remove (must be multiple of section size)
344 *
345 * Generic helper function to remove section mappings and sysfs entries
346 * for the section of the memory we are removing. Caller needs to make
347 * sure that pages are marked reserved and zones are adjust properly by
348 * calling offline_pages().
349 */
350int __remove_pages(struct zone *zone, unsigned long phys_start_pfn,
351 unsigned long nr_pages)
352{
353 unsigned long i, ret = 0;
354 int sections_to_remove;
355
356 /*
357 * We can only remove entire sections
358 */
359 BUG_ON(phys_start_pfn & ~PAGE_SECTION_MASK);
360 BUG_ON(nr_pages % PAGES_PER_SECTION);
361
362 sections_to_remove = nr_pages / PAGES_PER_SECTION;
363 for (i = 0; i < sections_to_remove; i++) {
364 unsigned long pfn = phys_start_pfn + i*PAGES_PER_SECTION;
365 release_mem_region(pfn << PAGE_SHIFT,
366 PAGES_PER_SECTION << PAGE_SHIFT);
367 ret = __remove_section(zone, __pfn_to_section(pfn));
368 if (ret)
369 break;
370 }
371 return ret;
372}
373EXPORT_SYMBOL_GPL(__remove_pages);
374
375int set_online_page_callback(online_page_callback_t callback)
376{
377 int rc = -EINVAL;
378
379 lock_memory_hotplug();
380
381 if (online_page_callback == generic_online_page) {
382 online_page_callback = callback;
383 rc = 0;
384 }
385
386 unlock_memory_hotplug();
387
388 return rc;
389}
390EXPORT_SYMBOL_GPL(set_online_page_callback);
391
392int restore_online_page_callback(online_page_callback_t callback)
393{
394 int rc = -EINVAL;
395
396 lock_memory_hotplug();
397
398 if (online_page_callback == callback) {
399 online_page_callback = generic_online_page;
400 rc = 0;
401 }
402
403 unlock_memory_hotplug();
404
405 return rc;
406}
407EXPORT_SYMBOL_GPL(restore_online_page_callback);
408
409void __online_page_set_limits(struct page *page)
410{
411 unsigned long pfn = page_to_pfn(page);
412
413 if (pfn >= num_physpages)
414 num_physpages = pfn + 1;
415}
416EXPORT_SYMBOL_GPL(__online_page_set_limits);
417
418void __online_page_increment_counters(struct page *page)
419{
420 totalram_pages++;
421
422#ifdef CONFIG_HIGHMEM
423 if (PageHighMem(page))
424 totalhigh_pages++;
425#endif
426}
427EXPORT_SYMBOL_GPL(__online_page_increment_counters);
428
429void __online_page_free(struct page *page)
430{
431 ClearPageReserved(page);
432 init_page_count(page);
433 __free_page(page);
434}
435EXPORT_SYMBOL_GPL(__online_page_free);
436
437static void generic_online_page(struct page *page)
438{
439 __online_page_set_limits(page);
440 __online_page_increment_counters(page);
441 __online_page_free(page);
442}
443
444static int online_pages_range(unsigned long start_pfn, unsigned long nr_pages,
445 void *arg)
446{
447 unsigned long i;
448 unsigned long onlined_pages = *(unsigned long *)arg;
449 struct page *page;
450 if (PageReserved(pfn_to_page(start_pfn)))
451 for (i = 0; i < nr_pages; i++) {
452 page = pfn_to_page(start_pfn + i);
453 (*online_page_callback)(page);
454 onlined_pages++;
455 }
456 *(unsigned long *)arg = onlined_pages;
457 return 0;
458}
459
460
461int __ref online_pages(unsigned long pfn, unsigned long nr_pages)
462{
463 unsigned long onlined_pages = 0;
464 struct zone *zone;
465 int need_zonelists_rebuild = 0;
466 int nid;
467 int ret;
468 struct memory_notify arg;
469
470 lock_memory_hotplug();
471 arg.start_pfn = pfn;
472 arg.nr_pages = nr_pages;
473 arg.status_change_nid = -1;
474
475 nid = page_to_nid(pfn_to_page(pfn));
476 if (node_present_pages(nid) == 0)
477 arg.status_change_nid = nid;
478
479 ret = memory_notify(MEM_GOING_ONLINE, &arg);
480 ret = notifier_to_errno(ret);
481 if (ret) {
482 memory_notify(MEM_CANCEL_ONLINE, &arg);
483 unlock_memory_hotplug();
484 return ret;
485 }
486 /*
487 * This doesn't need a lock to do pfn_to_page().
488 * The section can't be removed here because of the
489 * memory_block->state_mutex.
490 */
491 zone = page_zone(pfn_to_page(pfn));
492 /*
493 * If this zone is not populated, then it is not in zonelist.
494 * This means the page allocator ignores this zone.
495 * So, zonelist must be updated after online.
496 */
497 mutex_lock(&zonelists_mutex);
498 if (!populated_zone(zone))
499 need_zonelists_rebuild = 1;
500
501 ret = walk_system_ram_range(pfn, nr_pages, &onlined_pages,
502 online_pages_range);
503 if (ret) {
504 mutex_unlock(&zonelists_mutex);
505 printk(KERN_DEBUG "online_pages %lx at %lx failed\n",
506 nr_pages, pfn);
507 memory_notify(MEM_CANCEL_ONLINE, &arg);
508 unlock_memory_hotplug();
509 return ret;
510 }
511
512 zone->present_pages += onlined_pages;
513 zone->zone_pgdat->node_present_pages += onlined_pages;
514 if (need_zonelists_rebuild)
515 build_all_zonelists(zone);
516 else
517 zone_pcp_update(zone);
518
519 mutex_unlock(&zonelists_mutex);
520
521 init_per_zone_wmark_min();
522
523 if (onlined_pages) {
524 kswapd_run(zone_to_nid(zone));
525 node_set_state(zone_to_nid(zone), N_HIGH_MEMORY);
526 }
527
528 vm_total_pages = nr_free_pagecache_pages();
529
530 writeback_set_ratelimit();
531
532 if (onlined_pages)
533 memory_notify(MEM_ONLINE, &arg);
534 unlock_memory_hotplug();
535
536 return 0;
537}
538#endif /* CONFIG_MEMORY_HOTPLUG_SPARSE */
539
540/* we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG */
541static pg_data_t __ref *hotadd_new_pgdat(int nid, u64 start)
542{
543 struct pglist_data *pgdat;
544 unsigned long zones_size[MAX_NR_ZONES] = {0};
545 unsigned long zholes_size[MAX_NR_ZONES] = {0};
546 unsigned long start_pfn = start >> PAGE_SHIFT;
547
548 pgdat = arch_alloc_nodedata(nid);
549 if (!pgdat)
550 return NULL;
551
552 arch_refresh_nodedata(nid, pgdat);
553
554 /* we can use NODE_DATA(nid) from here */
555
556 /* init node's zones as empty zones, we don't have any present pages.*/
557 free_area_init_node(nid, zones_size, start_pfn, zholes_size);
558
559 /*
560 * The node we allocated has no zone fallback lists. For avoiding
561 * to access not-initialized zonelist, build here.
562 */
563 mutex_lock(&zonelists_mutex);
564 build_all_zonelists(NULL);
565 mutex_unlock(&zonelists_mutex);
566
567 return pgdat;
568}
569
570static void rollback_node_hotadd(int nid, pg_data_t *pgdat)
571{
572 arch_refresh_nodedata(nid, NULL);
573 arch_free_nodedata(pgdat);
574 return;
575}
576
577
578/*
579 * called by cpu_up() to online a node without onlined memory.
580 */
581int mem_online_node(int nid)
582{
583 pg_data_t *pgdat;
584 int ret;
585
586 lock_memory_hotplug();
587 pgdat = hotadd_new_pgdat(nid, 0);
588 if (!pgdat) {
589 ret = -ENOMEM;
590 goto out;
591 }
592 node_set_online(nid);
593 ret = register_one_node(nid);
594 BUG_ON(ret);
595
596out:
597 unlock_memory_hotplug();
598 return ret;
599}
600
601/* we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG */
602int __ref add_memory(int nid, u64 start, u64 size)
603{
604 pg_data_t *pgdat = NULL;
605 int new_pgdat = 0;
606 struct resource *res;
607 int ret;
608
609 lock_memory_hotplug();
610
611 res = register_memory_resource(start, size);
612 ret = -EEXIST;
613 if (!res)
614 goto out;
615
616 if (!node_online(nid)) {
617 pgdat = hotadd_new_pgdat(nid, start);
618 ret = -ENOMEM;
619 if (!pgdat)
620 goto out;
621 new_pgdat = 1;
622 }
623
624 /* call arch's memory hotadd */
625 ret = arch_add_memory(nid, start, size);
626
627 if (ret < 0)
628 goto error;
629
630 /* we online node here. we can't roll back from here. */
631 node_set_online(nid);
632
633 if (new_pgdat) {
634 ret = register_one_node(nid);
635 /*
636 * If sysfs file of new node can't create, cpu on the node
637 * can't be hot-added. There is no rollback way now.
638 * So, check by BUG_ON() to catch it reluctantly..
639 */
640 BUG_ON(ret);
641 }
642
643 /* create new memmap entry */
644 firmware_map_add_hotplug(start, start + size, "System RAM");
645
646 goto out;
647
648error:
649 /* rollback pgdat allocation and others */
650 if (new_pgdat)
651 rollback_node_hotadd(nid, pgdat);
652 if (res)
653 release_memory_resource(res);
654
655out:
656 unlock_memory_hotplug();
657 return ret;
658}
659EXPORT_SYMBOL_GPL(add_memory);
660
661#ifdef CONFIG_MEMORY_HOTREMOVE
662/*
663 * A free page on the buddy free lists (not the per-cpu lists) has PageBuddy
664 * set and the size of the free page is given by page_order(). Using this,
665 * the function determines if the pageblock contains only free pages.
666 * Due to buddy contraints, a free page at least the size of a pageblock will
667 * be located at the start of the pageblock
668 */
669static inline int pageblock_free(struct page *page)
670{
671 return PageBuddy(page) && page_order(page) >= pageblock_order;
672}
673
674/* Return the start of the next active pageblock after a given page */
675static struct page *next_active_pageblock(struct page *page)
676{
677 /* Ensure the starting page is pageblock-aligned */
678 BUG_ON(page_to_pfn(page) & (pageblock_nr_pages - 1));
679
680 /* If the entire pageblock is free, move to the end of free page */
681 if (pageblock_free(page)) {
682 int order;
683 /* be careful. we don't have locks, page_order can be changed.*/
684 order = page_order(page);
685 if ((order < MAX_ORDER) && (order >= pageblock_order))
686 return page + (1 << order);
687 }
688
689 return page + pageblock_nr_pages;
690}
691
692/* Checks if this range of memory is likely to be hot-removable. */
693int is_mem_section_removable(unsigned long start_pfn, unsigned long nr_pages)
694{
695 struct page *page = pfn_to_page(start_pfn);
696 struct page *end_page = page + nr_pages;
697
698 /* Check the starting page of each pageblock within the range */
699 for (; page < end_page; page = next_active_pageblock(page)) {
700 if (!is_pageblock_removable_nolock(page))
701 return 0;
702 cond_resched();
703 }
704
705 /* All pageblocks in the memory block are likely to be hot-removable */
706 return 1;
707}
708
709/*
710 * Confirm all pages in a range [start, end) is belongs to the same zone.
711 */
712static int test_pages_in_a_zone(unsigned long start_pfn, unsigned long end_pfn)
713{
714 unsigned long pfn;
715 struct zone *zone = NULL;
716 struct page *page;
717 int i;
718 for (pfn = start_pfn;
719 pfn < end_pfn;
720 pfn += MAX_ORDER_NR_PAGES) {
721 i = 0;
722 /* This is just a CONFIG_HOLES_IN_ZONE check.*/
723 while ((i < MAX_ORDER_NR_PAGES) && !pfn_valid_within(pfn + i))
724 i++;
725 if (i == MAX_ORDER_NR_PAGES)
726 continue;
727 page = pfn_to_page(pfn + i);
728 if (zone && page_zone(page) != zone)
729 return 0;
730 zone = page_zone(page);
731 }
732 return 1;
733}
734
735/*
736 * Scanning pfn is much easier than scanning lru list.
737 * Scan pfn from start to end and Find LRU page.
738 */
739static unsigned long scan_lru_pages(unsigned long start, unsigned long end)
740{
741 unsigned long pfn;
742 struct page *page;
743 for (pfn = start; pfn < end; pfn++) {
744 if (pfn_valid(pfn)) {
745 page = pfn_to_page(pfn);
746 if (PageLRU(page))
747 return pfn;
748 }
749 }
750 return 0;
751}
752
753static struct page *
754hotremove_migrate_alloc(struct page *page, unsigned long private, int **x)
755{
756 /* This should be improooooved!! */
757 return alloc_page(GFP_HIGHUSER_MOVABLE);
758}
759
760#define NR_OFFLINE_AT_ONCE_PAGES (256)
761static int
762do_migrate_range(unsigned long start_pfn, unsigned long end_pfn)
763{
764 unsigned long pfn;
765 struct page *page;
766 int move_pages = NR_OFFLINE_AT_ONCE_PAGES;
767 int not_managed = 0;
768 int ret = 0;
769 LIST_HEAD(source);
770
771 for (pfn = start_pfn; pfn < end_pfn && move_pages > 0; pfn++) {
772 if (!pfn_valid(pfn))
773 continue;
774 page = pfn_to_page(pfn);
775 if (!get_page_unless_zero(page))
776 continue;
777 /*
778 * We can skip free pages. And we can only deal with pages on
779 * LRU.
780 */
781 ret = isolate_lru_page(page);
782 if (!ret) { /* Success */
783 put_page(page);
784 list_add_tail(&page->lru, &source);
785 move_pages--;
786 inc_zone_page_state(page, NR_ISOLATED_ANON +
787 page_is_file_cache(page));
788
789 } else {
790#ifdef CONFIG_DEBUG_VM
791 printk(KERN_ALERT "removing pfn %lx from LRU failed\n",
792 pfn);
793 dump_page(page);
794#endif
795 put_page(page);
796 /* Because we don't have big zone->lock. we should
797 check this again here. */
798 if (page_count(page)) {
799 not_managed++;
800 ret = -EBUSY;
801 break;
802 }
803 }
804 }
805 if (!list_empty(&source)) {
806 if (not_managed) {
807 putback_lru_pages(&source);
808 goto out;
809 }
810 /* this function returns # of failed pages */
811 ret = migrate_pages(&source, hotremove_migrate_alloc, 0,
812 true, true);
813 if (ret)
814 putback_lru_pages(&source);
815 }
816out:
817 return ret;
818}
819
820/*
821 * remove from free_area[] and mark all as Reserved.
822 */
823static int
824offline_isolated_pages_cb(unsigned long start, unsigned long nr_pages,
825 void *data)
826{
827 __offline_isolated_pages(start, start + nr_pages);
828 return 0;
829}
830
831static void
832offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
833{
834 walk_system_ram_range(start_pfn, end_pfn - start_pfn, NULL,
835 offline_isolated_pages_cb);
836}
837
838/*
839 * Check all pages in range, recoreded as memory resource, are isolated.
840 */
841static int
842check_pages_isolated_cb(unsigned long start_pfn, unsigned long nr_pages,
843 void *data)
844{
845 int ret;
846 long offlined = *(long *)data;
847 ret = test_pages_isolated(start_pfn, start_pfn + nr_pages);
848 offlined = nr_pages;
849 if (!ret)
850 *(long *)data += offlined;
851 return ret;
852}
853
854static long
855check_pages_isolated(unsigned long start_pfn, unsigned long end_pfn)
856{
857 long offlined = 0;
858 int ret;
859
860 ret = walk_system_ram_range(start_pfn, end_pfn - start_pfn, &offlined,
861 check_pages_isolated_cb);
862 if (ret < 0)
863 offlined = (long)ret;
864 return offlined;
865}
866
867static int __ref offline_pages(unsigned long start_pfn,
868 unsigned long end_pfn, unsigned long timeout)
869{
870 unsigned long pfn, nr_pages, expire;
871 long offlined_pages;
872 int ret, drain, retry_max, node;
873 struct zone *zone;
874 struct memory_notify arg;
875
876 BUG_ON(start_pfn >= end_pfn);
877 /* at least, alignment against pageblock is necessary */
878 if (!IS_ALIGNED(start_pfn, pageblock_nr_pages))
879 return -EINVAL;
880 if (!IS_ALIGNED(end_pfn, pageblock_nr_pages))
881 return -EINVAL;
882 /* This makes hotplug much easier...and readable.
883 we assume this for now. .*/
884 if (!test_pages_in_a_zone(start_pfn, end_pfn))
885 return -EINVAL;
886
887 lock_memory_hotplug();
888
889 zone = page_zone(pfn_to_page(start_pfn));
890 node = zone_to_nid(zone);
891 nr_pages = end_pfn - start_pfn;
892
893 /* set above range as isolated */
894 ret = start_isolate_page_range(start_pfn, end_pfn);
895 if (ret)
896 goto out;
897
898 arg.start_pfn = start_pfn;
899 arg.nr_pages = nr_pages;
900 arg.status_change_nid = -1;
901 if (nr_pages >= node_present_pages(node))
902 arg.status_change_nid = node;
903
904 ret = memory_notify(MEM_GOING_OFFLINE, &arg);
905 ret = notifier_to_errno(ret);
906 if (ret)
907 goto failed_removal;
908
909 pfn = start_pfn;
910 expire = jiffies + timeout;
911 drain = 0;
912 retry_max = 5;
913repeat:
914 /* start memory hot removal */
915 ret = -EAGAIN;
916 if (time_after(jiffies, expire))
917 goto failed_removal;
918 ret = -EINTR;
919 if (signal_pending(current))
920 goto failed_removal;
921 ret = 0;
922 if (drain) {
923 lru_add_drain_all();
924 cond_resched();
925 drain_all_pages();
926 }
927
928 pfn = scan_lru_pages(start_pfn, end_pfn);
929 if (pfn) { /* We have page on LRU */
930 ret = do_migrate_range(pfn, end_pfn);
931 if (!ret) {
932 drain = 1;
933 goto repeat;
934 } else {
935 if (ret < 0)
936 if (--retry_max == 0)
937 goto failed_removal;
938 yield();
939 drain = 1;
940 goto repeat;
941 }
942 }
943 /* drain all zone's lru pagevec, this is asyncronous... */
944 lru_add_drain_all();
945 yield();
946 /* drain pcp pages , this is synchrouns. */
947 drain_all_pages();
948 /* check again */
949 offlined_pages = check_pages_isolated(start_pfn, end_pfn);
950 if (offlined_pages < 0) {
951 ret = -EBUSY;
952 goto failed_removal;
953 }
954 printk(KERN_INFO "Offlined Pages %ld\n", offlined_pages);
955 /* Ok, all of our target is islaoted.
956 We cannot do rollback at this point. */
957 offline_isolated_pages(start_pfn, end_pfn);
958 /* reset pagetype flags and makes migrate type to be MOVABLE */
959 undo_isolate_page_range(start_pfn, end_pfn);
960 /* removal success */
961 zone->present_pages -= offlined_pages;
962 zone->zone_pgdat->node_present_pages -= offlined_pages;
963 totalram_pages -= offlined_pages;
964
965 init_per_zone_wmark_min();
966
967 if (!node_present_pages(node)) {
968 node_clear_state(node, N_HIGH_MEMORY);
969 kswapd_stop(node);
970 }
971
972 vm_total_pages = nr_free_pagecache_pages();
973 writeback_set_ratelimit();
974
975 memory_notify(MEM_OFFLINE, &arg);
976 unlock_memory_hotplug();
977 return 0;
978
979failed_removal:
980 printk(KERN_INFO "memory offlining %lx to %lx failed\n",
981 start_pfn, end_pfn);
982 memory_notify(MEM_CANCEL_OFFLINE, &arg);
983 /* pushback to free area */
984 undo_isolate_page_range(start_pfn, end_pfn);
985
986out:
987 unlock_memory_hotplug();
988 return ret;
989}
990
991int remove_memory(u64 start, u64 size)
992{
993 unsigned long start_pfn, end_pfn;
994
995 start_pfn = PFN_DOWN(start);
996 end_pfn = start_pfn + PFN_DOWN(size);
997 return offline_pages(start_pfn, end_pfn, 120 * HZ);
998}
999#else
1000int remove_memory(u64 start, u64 size)
1001{
1002 return -EINVAL;
1003}
1004#endif /* CONFIG_MEMORY_HOTREMOVE */
1005EXPORT_SYMBOL_GPL(remove_memory);
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/writeback.h>
17#include <linux/slab.h>
18#include <linux/sysctl.h>
19#include <linux/cpu.h>
20#include <linux/memory.h>
21#include <linux/memremap.h>
22#include <linux/memory_hotplug.h>
23#include <linux/vmalloc.h>
24#include <linux/ioport.h>
25#include <linux/delay.h>
26#include <linux/migrate.h>
27#include <linux/page-isolation.h>
28#include <linux/pfn.h>
29#include <linux/suspend.h>
30#include <linux/mm_inline.h>
31#include <linux/firmware-map.h>
32#include <linux/stop_machine.h>
33#include <linux/hugetlb.h>
34#include <linux/memblock.h>
35#include <linux/compaction.h>
36#include <linux/rmap.h>
37#include <linux/module.h>
38
39#include <asm/tlbflush.h>
40
41#include "internal.h"
42#include "shuffle.h"
43
44enum {
45 MEMMAP_ON_MEMORY_DISABLE = 0,
46 MEMMAP_ON_MEMORY_ENABLE,
47 MEMMAP_ON_MEMORY_FORCE,
48};
49
50static int memmap_mode __read_mostly = MEMMAP_ON_MEMORY_DISABLE;
51
52static inline unsigned long memory_block_memmap_size(void)
53{
54 return PHYS_PFN(memory_block_size_bytes()) * sizeof(struct page);
55}
56
57static inline unsigned long memory_block_memmap_on_memory_pages(void)
58{
59 unsigned long nr_pages = PFN_UP(memory_block_memmap_size());
60
61 /*
62 * In "forced" memmap_on_memory mode, we add extra pages to align the
63 * vmemmap size to cover full pageblocks. That way, we can add memory
64 * even if the vmemmap size is not properly aligned, however, we might waste
65 * memory.
66 */
67 if (memmap_mode == MEMMAP_ON_MEMORY_FORCE)
68 return pageblock_align(nr_pages);
69 return nr_pages;
70}
71
72#ifdef CONFIG_MHP_MEMMAP_ON_MEMORY
73/*
74 * memory_hotplug.memmap_on_memory parameter
75 */
76static int set_memmap_mode(const char *val, const struct kernel_param *kp)
77{
78 int ret, mode;
79 bool enabled;
80
81 if (sysfs_streq(val, "force") || sysfs_streq(val, "FORCE")) {
82 mode = MEMMAP_ON_MEMORY_FORCE;
83 } else {
84 ret = kstrtobool(val, &enabled);
85 if (ret < 0)
86 return ret;
87 if (enabled)
88 mode = MEMMAP_ON_MEMORY_ENABLE;
89 else
90 mode = MEMMAP_ON_MEMORY_DISABLE;
91 }
92 *((int *)kp->arg) = mode;
93 if (mode == MEMMAP_ON_MEMORY_FORCE) {
94 unsigned long memmap_pages = memory_block_memmap_on_memory_pages();
95
96 pr_info_once("Memory hotplug will waste %ld pages in each memory block\n",
97 memmap_pages - PFN_UP(memory_block_memmap_size()));
98 }
99 return 0;
100}
101
102static int get_memmap_mode(char *buffer, const struct kernel_param *kp)
103{
104 int mode = *((int *)kp->arg);
105
106 if (mode == MEMMAP_ON_MEMORY_FORCE)
107 return sprintf(buffer, "force\n");
108 return sprintf(buffer, "%c\n", mode ? 'Y' : 'N');
109}
110
111static const struct kernel_param_ops memmap_mode_ops = {
112 .set = set_memmap_mode,
113 .get = get_memmap_mode,
114};
115module_param_cb(memmap_on_memory, &memmap_mode_ops, &memmap_mode, 0444);
116MODULE_PARM_DESC(memmap_on_memory, "Enable memmap on memory for memory hotplug\n"
117 "With value \"force\" it could result in memory wastage due "
118 "to memmap size limitations (Y/N/force)");
119
120static inline bool mhp_memmap_on_memory(void)
121{
122 return memmap_mode != MEMMAP_ON_MEMORY_DISABLE;
123}
124#else
125static inline bool mhp_memmap_on_memory(void)
126{
127 return false;
128}
129#endif
130
131enum {
132 ONLINE_POLICY_CONTIG_ZONES = 0,
133 ONLINE_POLICY_AUTO_MOVABLE,
134};
135
136static const char * const online_policy_to_str[] = {
137 [ONLINE_POLICY_CONTIG_ZONES] = "contig-zones",
138 [ONLINE_POLICY_AUTO_MOVABLE] = "auto-movable",
139};
140
141static int set_online_policy(const char *val, const struct kernel_param *kp)
142{
143 int ret = sysfs_match_string(online_policy_to_str, val);
144
145 if (ret < 0)
146 return ret;
147 *((int *)kp->arg) = ret;
148 return 0;
149}
150
151static int get_online_policy(char *buffer, const struct kernel_param *kp)
152{
153 return sprintf(buffer, "%s\n", online_policy_to_str[*((int *)kp->arg)]);
154}
155
156/*
157 * memory_hotplug.online_policy: configure online behavior when onlining without
158 * specifying a zone (MMOP_ONLINE)
159 *
160 * "contig-zones": keep zone contiguous
161 * "auto-movable": online memory to ZONE_MOVABLE if the configuration
162 * (auto_movable_ratio, auto_movable_numa_aware) allows for it
163 */
164static int online_policy __read_mostly = ONLINE_POLICY_CONTIG_ZONES;
165static const struct kernel_param_ops online_policy_ops = {
166 .set = set_online_policy,
167 .get = get_online_policy,
168};
169module_param_cb(online_policy, &online_policy_ops, &online_policy, 0644);
170MODULE_PARM_DESC(online_policy,
171 "Set the online policy (\"contig-zones\", \"auto-movable\") "
172 "Default: \"contig-zones\"");
173
174/*
175 * memory_hotplug.auto_movable_ratio: specify maximum MOVABLE:KERNEL ratio
176 *
177 * The ratio represent an upper limit and the kernel might decide to not
178 * online some memory to ZONE_MOVABLE -- e.g., because hotplugged KERNEL memory
179 * doesn't allow for more MOVABLE memory.
180 */
181static unsigned int auto_movable_ratio __read_mostly = 301;
182module_param(auto_movable_ratio, uint, 0644);
183MODULE_PARM_DESC(auto_movable_ratio,
184 "Set the maximum ratio of MOVABLE:KERNEL memory in the system "
185 "in percent for \"auto-movable\" online policy. Default: 301");
186
187/*
188 * memory_hotplug.auto_movable_numa_aware: consider numa node stats
189 */
190#ifdef CONFIG_NUMA
191static bool auto_movable_numa_aware __read_mostly = true;
192module_param(auto_movable_numa_aware, bool, 0644);
193MODULE_PARM_DESC(auto_movable_numa_aware,
194 "Consider numa node stats in addition to global stats in "
195 "\"auto-movable\" online policy. Default: true");
196#endif /* CONFIG_NUMA */
197
198/*
199 * online_page_callback contains pointer to current page onlining function.
200 * Initially it is generic_online_page(). If it is required it could be
201 * changed by calling set_online_page_callback() for callback registration
202 * and restore_online_page_callback() for generic callback restore.
203 */
204
205static online_page_callback_t online_page_callback = generic_online_page;
206static DEFINE_MUTEX(online_page_callback_lock);
207
208DEFINE_STATIC_PERCPU_RWSEM(mem_hotplug_lock);
209
210void get_online_mems(void)
211{
212 percpu_down_read(&mem_hotplug_lock);
213}
214
215void put_online_mems(void)
216{
217 percpu_up_read(&mem_hotplug_lock);
218}
219
220bool movable_node_enabled = false;
221
222#ifndef CONFIG_MEMORY_HOTPLUG_DEFAULT_ONLINE
223int mhp_default_online_type = MMOP_OFFLINE;
224#else
225int mhp_default_online_type = MMOP_ONLINE;
226#endif
227
228static int __init setup_memhp_default_state(char *str)
229{
230 const int online_type = mhp_online_type_from_str(str);
231
232 if (online_type >= 0)
233 mhp_default_online_type = online_type;
234
235 return 1;
236}
237__setup("memhp_default_state=", setup_memhp_default_state);
238
239void mem_hotplug_begin(void)
240{
241 cpus_read_lock();
242 percpu_down_write(&mem_hotplug_lock);
243}
244
245void mem_hotplug_done(void)
246{
247 percpu_up_write(&mem_hotplug_lock);
248 cpus_read_unlock();
249}
250
251u64 max_mem_size = U64_MAX;
252
253/* add this memory to iomem resource */
254static struct resource *register_memory_resource(u64 start, u64 size,
255 const char *resource_name)
256{
257 struct resource *res;
258 unsigned long flags = IORESOURCE_SYSTEM_RAM | IORESOURCE_BUSY;
259
260 if (strcmp(resource_name, "System RAM"))
261 flags |= IORESOURCE_SYSRAM_DRIVER_MANAGED;
262
263 if (!mhp_range_allowed(start, size, true))
264 return ERR_PTR(-E2BIG);
265
266 /*
267 * Make sure value parsed from 'mem=' only restricts memory adding
268 * while booting, so that memory hotplug won't be impacted. Please
269 * refer to document of 'mem=' in kernel-parameters.txt for more
270 * details.
271 */
272 if (start + size > max_mem_size && system_state < SYSTEM_RUNNING)
273 return ERR_PTR(-E2BIG);
274
275 /*
276 * Request ownership of the new memory range. This might be
277 * a child of an existing resource that was present but
278 * not marked as busy.
279 */
280 res = __request_region(&iomem_resource, start, size,
281 resource_name, flags);
282
283 if (!res) {
284 pr_debug("Unable to reserve System RAM region: %016llx->%016llx\n",
285 start, start + size);
286 return ERR_PTR(-EEXIST);
287 }
288 return res;
289}
290
291static void release_memory_resource(struct resource *res)
292{
293 if (!res)
294 return;
295 release_resource(res);
296 kfree(res);
297}
298
299static int check_pfn_span(unsigned long pfn, unsigned long nr_pages)
300{
301 /*
302 * Disallow all operations smaller than a sub-section and only
303 * allow operations smaller than a section for
304 * SPARSEMEM_VMEMMAP. Note that check_hotplug_memory_range()
305 * enforces a larger memory_block_size_bytes() granularity for
306 * memory that will be marked online, so this check should only
307 * fire for direct arch_{add,remove}_memory() users outside of
308 * add_memory_resource().
309 */
310 unsigned long min_align;
311
312 if (IS_ENABLED(CONFIG_SPARSEMEM_VMEMMAP))
313 min_align = PAGES_PER_SUBSECTION;
314 else
315 min_align = PAGES_PER_SECTION;
316 if (!IS_ALIGNED(pfn | nr_pages, min_align))
317 return -EINVAL;
318 return 0;
319}
320
321/*
322 * Return page for the valid pfn only if the page is online. All pfn
323 * walkers which rely on the fully initialized page->flags and others
324 * should use this rather than pfn_valid && pfn_to_page
325 */
326struct page *pfn_to_online_page(unsigned long pfn)
327{
328 unsigned long nr = pfn_to_section_nr(pfn);
329 struct dev_pagemap *pgmap;
330 struct mem_section *ms;
331
332 if (nr >= NR_MEM_SECTIONS)
333 return NULL;
334
335 ms = __nr_to_section(nr);
336 if (!online_section(ms))
337 return NULL;
338
339 /*
340 * Save some code text when online_section() +
341 * pfn_section_valid() are sufficient.
342 */
343 if (IS_ENABLED(CONFIG_HAVE_ARCH_PFN_VALID) && !pfn_valid(pfn))
344 return NULL;
345
346 if (!pfn_section_valid(ms, pfn))
347 return NULL;
348
349 if (!online_device_section(ms))
350 return pfn_to_page(pfn);
351
352 /*
353 * Slowpath: when ZONE_DEVICE collides with
354 * ZONE_{NORMAL,MOVABLE} within the same section some pfns in
355 * the section may be 'offline' but 'valid'. Only
356 * get_dev_pagemap() can determine sub-section online status.
357 */
358 pgmap = get_dev_pagemap(pfn, NULL);
359 put_dev_pagemap(pgmap);
360
361 /* The presence of a pgmap indicates ZONE_DEVICE offline pfn */
362 if (pgmap)
363 return NULL;
364
365 return pfn_to_page(pfn);
366}
367EXPORT_SYMBOL_GPL(pfn_to_online_page);
368
369int __add_pages(int nid, unsigned long pfn, unsigned long nr_pages,
370 struct mhp_params *params)
371{
372 const unsigned long end_pfn = pfn + nr_pages;
373 unsigned long cur_nr_pages;
374 int err;
375 struct vmem_altmap *altmap = params->altmap;
376
377 if (WARN_ON_ONCE(!pgprot_val(params->pgprot)))
378 return -EINVAL;
379
380 VM_BUG_ON(!mhp_range_allowed(PFN_PHYS(pfn), nr_pages * PAGE_SIZE, false));
381
382 if (altmap) {
383 /*
384 * Validate altmap is within bounds of the total request
385 */
386 if (altmap->base_pfn != pfn
387 || vmem_altmap_offset(altmap) > nr_pages) {
388 pr_warn_once("memory add fail, invalid altmap\n");
389 return -EINVAL;
390 }
391 altmap->alloc = 0;
392 }
393
394 if (check_pfn_span(pfn, nr_pages)) {
395 WARN(1, "Misaligned %s start: %#lx end: %#lx\n", __func__, pfn, pfn + nr_pages - 1);
396 return -EINVAL;
397 }
398
399 for (; pfn < end_pfn; pfn += cur_nr_pages) {
400 /* Select all remaining pages up to the next section boundary */
401 cur_nr_pages = min(end_pfn - pfn,
402 SECTION_ALIGN_UP(pfn + 1) - pfn);
403 err = sparse_add_section(nid, pfn, cur_nr_pages, altmap,
404 params->pgmap);
405 if (err)
406 break;
407 cond_resched();
408 }
409 vmemmap_populate_print_last();
410 return err;
411}
412
413/* find the smallest valid pfn in the range [start_pfn, end_pfn) */
414static unsigned long find_smallest_section_pfn(int nid, struct zone *zone,
415 unsigned long start_pfn,
416 unsigned long end_pfn)
417{
418 for (; start_pfn < end_pfn; start_pfn += PAGES_PER_SUBSECTION) {
419 if (unlikely(!pfn_to_online_page(start_pfn)))
420 continue;
421
422 if (unlikely(pfn_to_nid(start_pfn) != nid))
423 continue;
424
425 if (zone != page_zone(pfn_to_page(start_pfn)))
426 continue;
427
428 return start_pfn;
429 }
430
431 return 0;
432}
433
434/* find the biggest valid pfn in the range [start_pfn, end_pfn). */
435static unsigned long find_biggest_section_pfn(int nid, struct zone *zone,
436 unsigned long start_pfn,
437 unsigned long end_pfn)
438{
439 unsigned long pfn;
440
441 /* pfn is the end pfn of a memory section. */
442 pfn = end_pfn - 1;
443 for (; pfn >= start_pfn; pfn -= PAGES_PER_SUBSECTION) {
444 if (unlikely(!pfn_to_online_page(pfn)))
445 continue;
446
447 if (unlikely(pfn_to_nid(pfn) != nid))
448 continue;
449
450 if (zone != page_zone(pfn_to_page(pfn)))
451 continue;
452
453 return pfn;
454 }
455
456 return 0;
457}
458
459static void shrink_zone_span(struct zone *zone, unsigned long start_pfn,
460 unsigned long end_pfn)
461{
462 unsigned long pfn;
463 int nid = zone_to_nid(zone);
464
465 if (zone->zone_start_pfn == start_pfn) {
466 /*
467 * If the section is smallest section in the zone, it need
468 * shrink zone->zone_start_pfn and zone->zone_spanned_pages.
469 * In this case, we find second smallest valid mem_section
470 * for shrinking zone.
471 */
472 pfn = find_smallest_section_pfn(nid, zone, end_pfn,
473 zone_end_pfn(zone));
474 if (pfn) {
475 zone->spanned_pages = zone_end_pfn(zone) - pfn;
476 zone->zone_start_pfn = pfn;
477 } else {
478 zone->zone_start_pfn = 0;
479 zone->spanned_pages = 0;
480 }
481 } else if (zone_end_pfn(zone) == end_pfn) {
482 /*
483 * If the section is biggest section in the zone, it need
484 * shrink zone->spanned_pages.
485 * In this case, we find second biggest valid mem_section for
486 * shrinking zone.
487 */
488 pfn = find_biggest_section_pfn(nid, zone, zone->zone_start_pfn,
489 start_pfn);
490 if (pfn)
491 zone->spanned_pages = pfn - zone->zone_start_pfn + 1;
492 else {
493 zone->zone_start_pfn = 0;
494 zone->spanned_pages = 0;
495 }
496 }
497}
498
499static void update_pgdat_span(struct pglist_data *pgdat)
500{
501 unsigned long node_start_pfn = 0, node_end_pfn = 0;
502 struct zone *zone;
503
504 for (zone = pgdat->node_zones;
505 zone < pgdat->node_zones + MAX_NR_ZONES; zone++) {
506 unsigned long end_pfn = zone_end_pfn(zone);
507
508 /* No need to lock the zones, they can't change. */
509 if (!zone->spanned_pages)
510 continue;
511 if (!node_end_pfn) {
512 node_start_pfn = zone->zone_start_pfn;
513 node_end_pfn = end_pfn;
514 continue;
515 }
516
517 if (end_pfn > node_end_pfn)
518 node_end_pfn = end_pfn;
519 if (zone->zone_start_pfn < node_start_pfn)
520 node_start_pfn = zone->zone_start_pfn;
521 }
522
523 pgdat->node_start_pfn = node_start_pfn;
524 pgdat->node_spanned_pages = node_end_pfn - node_start_pfn;
525}
526
527void remove_pfn_range_from_zone(struct zone *zone,
528 unsigned long start_pfn,
529 unsigned long nr_pages)
530{
531 const unsigned long end_pfn = start_pfn + nr_pages;
532 struct pglist_data *pgdat = zone->zone_pgdat;
533 unsigned long pfn, cur_nr_pages;
534
535 /* Poison struct pages because they are now uninitialized again. */
536 for (pfn = start_pfn; pfn < end_pfn; pfn += cur_nr_pages) {
537 cond_resched();
538
539 /* Select all remaining pages up to the next section boundary */
540 cur_nr_pages =
541 min(end_pfn - pfn, SECTION_ALIGN_UP(pfn + 1) - pfn);
542 page_init_poison(pfn_to_page(pfn),
543 sizeof(struct page) * cur_nr_pages);
544 }
545
546 /*
547 * Zone shrinking code cannot properly deal with ZONE_DEVICE. So
548 * we will not try to shrink the zones - which is okay as
549 * set_zone_contiguous() cannot deal with ZONE_DEVICE either way.
550 */
551 if (zone_is_zone_device(zone))
552 return;
553
554 clear_zone_contiguous(zone);
555
556 shrink_zone_span(zone, start_pfn, start_pfn + nr_pages);
557 update_pgdat_span(pgdat);
558
559 set_zone_contiguous(zone);
560}
561
562/**
563 * __remove_pages() - remove sections of pages
564 * @pfn: starting pageframe (must be aligned to start of a section)
565 * @nr_pages: number of pages to remove (must be multiple of section size)
566 * @altmap: alternative device page map or %NULL if default memmap is used
567 *
568 * Generic helper function to remove section mappings and sysfs entries
569 * for the section of the memory we are removing. Caller needs to make
570 * sure that pages are marked reserved and zones are adjust properly by
571 * calling offline_pages().
572 */
573void __remove_pages(unsigned long pfn, unsigned long nr_pages,
574 struct vmem_altmap *altmap)
575{
576 const unsigned long end_pfn = pfn + nr_pages;
577 unsigned long cur_nr_pages;
578
579 if (check_pfn_span(pfn, nr_pages)) {
580 WARN(1, "Misaligned %s start: %#lx end: %#lx\n", __func__, pfn, pfn + nr_pages - 1);
581 return;
582 }
583
584 for (; pfn < end_pfn; pfn += cur_nr_pages) {
585 cond_resched();
586 /* Select all remaining pages up to the next section boundary */
587 cur_nr_pages = min(end_pfn - pfn,
588 SECTION_ALIGN_UP(pfn + 1) - pfn);
589 sparse_remove_section(pfn, cur_nr_pages, altmap);
590 }
591}
592
593int set_online_page_callback(online_page_callback_t callback)
594{
595 int rc = -EINVAL;
596
597 get_online_mems();
598 mutex_lock(&online_page_callback_lock);
599
600 if (online_page_callback == generic_online_page) {
601 online_page_callback = callback;
602 rc = 0;
603 }
604
605 mutex_unlock(&online_page_callback_lock);
606 put_online_mems();
607
608 return rc;
609}
610EXPORT_SYMBOL_GPL(set_online_page_callback);
611
612int restore_online_page_callback(online_page_callback_t callback)
613{
614 int rc = -EINVAL;
615
616 get_online_mems();
617 mutex_lock(&online_page_callback_lock);
618
619 if (online_page_callback == callback) {
620 online_page_callback = generic_online_page;
621 rc = 0;
622 }
623
624 mutex_unlock(&online_page_callback_lock);
625 put_online_mems();
626
627 return rc;
628}
629EXPORT_SYMBOL_GPL(restore_online_page_callback);
630
631/* we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG */
632void generic_online_page(struct page *page, unsigned int order)
633{
634 __free_pages_core(page, order, MEMINIT_HOTPLUG);
635}
636EXPORT_SYMBOL_GPL(generic_online_page);
637
638static void online_pages_range(unsigned long start_pfn, unsigned long nr_pages)
639{
640 const unsigned long end_pfn = start_pfn + nr_pages;
641 unsigned long pfn;
642
643 /*
644 * Online the pages in MAX_PAGE_ORDER aligned chunks. The callback might
645 * decide to not expose all pages to the buddy (e.g., expose them
646 * later). We account all pages as being online and belonging to this
647 * zone ("present").
648 * When using memmap_on_memory, the range might not be aligned to
649 * MAX_ORDER_NR_PAGES - 1, but pageblock aligned. __ffs() will detect
650 * this and the first chunk to online will be pageblock_nr_pages.
651 */
652 for (pfn = start_pfn; pfn < end_pfn;) {
653 int order;
654
655 /*
656 * Free to online pages in the largest chunks alignment allows.
657 *
658 * __ffs() behaviour is undefined for 0. start == 0 is
659 * MAX_PAGE_ORDER-aligned, Set order to MAX_PAGE_ORDER for
660 * the case.
661 */
662 if (pfn)
663 order = min_t(int, MAX_PAGE_ORDER, __ffs(pfn));
664 else
665 order = MAX_PAGE_ORDER;
666
667 (*online_page_callback)(pfn_to_page(pfn), order);
668 pfn += (1UL << order);
669 }
670
671 /* mark all involved sections as online */
672 online_mem_sections(start_pfn, end_pfn);
673}
674
675/* check which state of node_states will be changed when online memory */
676static void node_states_check_changes_online(unsigned long nr_pages,
677 struct zone *zone, struct memory_notify *arg)
678{
679 int nid = zone_to_nid(zone);
680
681 arg->status_change_nid = NUMA_NO_NODE;
682 arg->status_change_nid_normal = NUMA_NO_NODE;
683
684 if (!node_state(nid, N_MEMORY))
685 arg->status_change_nid = nid;
686 if (zone_idx(zone) <= ZONE_NORMAL && !node_state(nid, N_NORMAL_MEMORY))
687 arg->status_change_nid_normal = nid;
688}
689
690static void node_states_set_node(int node, struct memory_notify *arg)
691{
692 if (arg->status_change_nid_normal >= 0)
693 node_set_state(node, N_NORMAL_MEMORY);
694
695 if (arg->status_change_nid >= 0)
696 node_set_state(node, N_MEMORY);
697}
698
699static void __meminit resize_zone_range(struct zone *zone, unsigned long start_pfn,
700 unsigned long nr_pages)
701{
702 unsigned long old_end_pfn = zone_end_pfn(zone);
703
704 if (zone_is_empty(zone) || start_pfn < zone->zone_start_pfn)
705 zone->zone_start_pfn = start_pfn;
706
707 zone->spanned_pages = max(start_pfn + nr_pages, old_end_pfn) - zone->zone_start_pfn;
708}
709
710static void __meminit resize_pgdat_range(struct pglist_data *pgdat, unsigned long start_pfn,
711 unsigned long nr_pages)
712{
713 unsigned long old_end_pfn = pgdat_end_pfn(pgdat);
714
715 if (!pgdat->node_spanned_pages || start_pfn < pgdat->node_start_pfn)
716 pgdat->node_start_pfn = start_pfn;
717
718 pgdat->node_spanned_pages = max(start_pfn + nr_pages, old_end_pfn) - pgdat->node_start_pfn;
719
720}
721
722#ifdef CONFIG_ZONE_DEVICE
723static void section_taint_zone_device(unsigned long pfn)
724{
725 struct mem_section *ms = __pfn_to_section(pfn);
726
727 ms->section_mem_map |= SECTION_TAINT_ZONE_DEVICE;
728}
729#else
730static inline void section_taint_zone_device(unsigned long pfn)
731{
732}
733#endif
734
735/*
736 * Associate the pfn range with the given zone, initializing the memmaps
737 * and resizing the pgdat/zone data to span the added pages. After this
738 * call, all affected pages are PageOffline().
739 *
740 * All aligned pageblocks are initialized to the specified migratetype
741 * (usually MIGRATE_MOVABLE). Besides setting the migratetype, no related
742 * zone stats (e.g., nr_isolate_pageblock) are touched.
743 */
744void move_pfn_range_to_zone(struct zone *zone, unsigned long start_pfn,
745 unsigned long nr_pages,
746 struct vmem_altmap *altmap, int migratetype)
747{
748 struct pglist_data *pgdat = zone->zone_pgdat;
749 int nid = pgdat->node_id;
750
751 clear_zone_contiguous(zone);
752
753 if (zone_is_empty(zone))
754 init_currently_empty_zone(zone, start_pfn, nr_pages);
755 resize_zone_range(zone, start_pfn, nr_pages);
756 resize_pgdat_range(pgdat, start_pfn, nr_pages);
757
758 /*
759 * Subsection population requires care in pfn_to_online_page().
760 * Set the taint to enable the slow path detection of
761 * ZONE_DEVICE pages in an otherwise ZONE_{NORMAL,MOVABLE}
762 * section.
763 */
764 if (zone_is_zone_device(zone)) {
765 if (!IS_ALIGNED(start_pfn, PAGES_PER_SECTION))
766 section_taint_zone_device(start_pfn);
767 if (!IS_ALIGNED(start_pfn + nr_pages, PAGES_PER_SECTION))
768 section_taint_zone_device(start_pfn + nr_pages);
769 }
770
771 /*
772 * TODO now we have a visible range of pages which are not associated
773 * with their zone properly. Not nice but set_pfnblock_flags_mask
774 * expects the zone spans the pfn range. All the pages in the range
775 * are reserved so nobody should be touching them so we should be safe
776 */
777 memmap_init_range(nr_pages, nid, zone_idx(zone), start_pfn, 0,
778 MEMINIT_HOTPLUG, altmap, migratetype);
779
780 set_zone_contiguous(zone);
781}
782
783struct auto_movable_stats {
784 unsigned long kernel_early_pages;
785 unsigned long movable_pages;
786};
787
788static void auto_movable_stats_account_zone(struct auto_movable_stats *stats,
789 struct zone *zone)
790{
791 if (zone_idx(zone) == ZONE_MOVABLE) {
792 stats->movable_pages += zone->present_pages;
793 } else {
794 stats->kernel_early_pages += zone->present_early_pages;
795#ifdef CONFIG_CMA
796 /*
797 * CMA pages (never on hotplugged memory) behave like
798 * ZONE_MOVABLE.
799 */
800 stats->movable_pages += zone->cma_pages;
801 stats->kernel_early_pages -= zone->cma_pages;
802#endif /* CONFIG_CMA */
803 }
804}
805struct auto_movable_group_stats {
806 unsigned long movable_pages;
807 unsigned long req_kernel_early_pages;
808};
809
810static int auto_movable_stats_account_group(struct memory_group *group,
811 void *arg)
812{
813 const int ratio = READ_ONCE(auto_movable_ratio);
814 struct auto_movable_group_stats *stats = arg;
815 long pages;
816
817 /*
818 * We don't support modifying the config while the auto-movable online
819 * policy is already enabled. Just avoid the division by zero below.
820 */
821 if (!ratio)
822 return 0;
823
824 /*
825 * Calculate how many early kernel pages this group requires to
826 * satisfy the configured zone ratio.
827 */
828 pages = group->present_movable_pages * 100 / ratio;
829 pages -= group->present_kernel_pages;
830
831 if (pages > 0)
832 stats->req_kernel_early_pages += pages;
833 stats->movable_pages += group->present_movable_pages;
834 return 0;
835}
836
837static bool auto_movable_can_online_movable(int nid, struct memory_group *group,
838 unsigned long nr_pages)
839{
840 unsigned long kernel_early_pages, movable_pages;
841 struct auto_movable_group_stats group_stats = {};
842 struct auto_movable_stats stats = {};
843 struct zone *zone;
844 int i;
845
846 /* Walk all relevant zones and collect MOVABLE vs. KERNEL stats. */
847 if (nid == NUMA_NO_NODE) {
848 /* TODO: cache values */
849 for_each_populated_zone(zone)
850 auto_movable_stats_account_zone(&stats, zone);
851 } else {
852 for (i = 0; i < MAX_NR_ZONES; i++) {
853 pg_data_t *pgdat = NODE_DATA(nid);
854
855 zone = pgdat->node_zones + i;
856 if (populated_zone(zone))
857 auto_movable_stats_account_zone(&stats, zone);
858 }
859 }
860
861 kernel_early_pages = stats.kernel_early_pages;
862 movable_pages = stats.movable_pages;
863
864 /*
865 * Kernel memory inside dynamic memory group allows for more MOVABLE
866 * memory within the same group. Remove the effect of all but the
867 * current group from the stats.
868 */
869 walk_dynamic_memory_groups(nid, auto_movable_stats_account_group,
870 group, &group_stats);
871 if (kernel_early_pages <= group_stats.req_kernel_early_pages)
872 return false;
873 kernel_early_pages -= group_stats.req_kernel_early_pages;
874 movable_pages -= group_stats.movable_pages;
875
876 if (group && group->is_dynamic)
877 kernel_early_pages += group->present_kernel_pages;
878
879 /*
880 * Test if we could online the given number of pages to ZONE_MOVABLE
881 * and still stay in the configured ratio.
882 */
883 movable_pages += nr_pages;
884 return movable_pages <= (auto_movable_ratio * kernel_early_pages) / 100;
885}
886
887/*
888 * Returns a default kernel memory zone for the given pfn range.
889 * If no kernel zone covers this pfn range it will automatically go
890 * to the ZONE_NORMAL.
891 */
892static struct zone *default_kernel_zone_for_pfn(int nid, unsigned long start_pfn,
893 unsigned long nr_pages)
894{
895 struct pglist_data *pgdat = NODE_DATA(nid);
896 int zid;
897
898 for (zid = 0; zid < ZONE_NORMAL; zid++) {
899 struct zone *zone = &pgdat->node_zones[zid];
900
901 if (zone_intersects(zone, start_pfn, nr_pages))
902 return zone;
903 }
904
905 return &pgdat->node_zones[ZONE_NORMAL];
906}
907
908/*
909 * Determine to which zone to online memory dynamically based on user
910 * configuration and system stats. We care about the following ratio:
911 *
912 * MOVABLE : KERNEL
913 *
914 * Whereby MOVABLE is memory in ZONE_MOVABLE and KERNEL is memory in
915 * one of the kernel zones. CMA pages inside one of the kernel zones really
916 * behaves like ZONE_MOVABLE, so we treat them accordingly.
917 *
918 * We don't allow for hotplugged memory in a KERNEL zone to increase the
919 * amount of MOVABLE memory we can have, so we end up with:
920 *
921 * MOVABLE : KERNEL_EARLY
922 *
923 * Whereby KERNEL_EARLY is memory in one of the kernel zones, available sinze
924 * boot. We base our calculation on KERNEL_EARLY internally, because:
925 *
926 * a) Hotplugged memory in one of the kernel zones can sometimes still get
927 * hotunplugged, especially when hot(un)plugging individual memory blocks.
928 * There is no coordination across memory devices, therefore "automatic"
929 * hotunplugging, as implemented in hypervisors, could result in zone
930 * imbalances.
931 * b) Early/boot memory in one of the kernel zones can usually not get
932 * hotunplugged again (e.g., no firmware interface to unplug, fragmented
933 * with unmovable allocations). While there are corner cases where it might
934 * still work, it is barely relevant in practice.
935 *
936 * Exceptions are dynamic memory groups, which allow for more MOVABLE
937 * memory within the same memory group -- because in that case, there is
938 * coordination within the single memory device managed by a single driver.
939 *
940 * We rely on "present pages" instead of "managed pages", as the latter is
941 * highly unreliable and dynamic in virtualized environments, and does not
942 * consider boot time allocations. For example, memory ballooning adjusts the
943 * managed pages when inflating/deflating the balloon, and balloon compaction
944 * can even migrate inflated pages between zones.
945 *
946 * Using "present pages" is better but some things to keep in mind are:
947 *
948 * a) Some memblock allocations, such as for the crashkernel area, are
949 * effectively unused by the kernel, yet they account to "present pages".
950 * Fortunately, these allocations are comparatively small in relevant setups
951 * (e.g., fraction of system memory).
952 * b) Some hotplugged memory blocks in virtualized environments, esecially
953 * hotplugged by virtio-mem, look like they are completely present, however,
954 * only parts of the memory block are actually currently usable.
955 * "present pages" is an upper limit that can get reached at runtime. As
956 * we base our calculations on KERNEL_EARLY, this is not an issue.
957 */
958static struct zone *auto_movable_zone_for_pfn(int nid,
959 struct memory_group *group,
960 unsigned long pfn,
961 unsigned long nr_pages)
962{
963 unsigned long online_pages = 0, max_pages, end_pfn;
964 struct page *page;
965
966 if (!auto_movable_ratio)
967 goto kernel_zone;
968
969 if (group && !group->is_dynamic) {
970 max_pages = group->s.max_pages;
971 online_pages = group->present_movable_pages;
972
973 /* If anything is !MOVABLE online the rest !MOVABLE. */
974 if (group->present_kernel_pages)
975 goto kernel_zone;
976 } else if (!group || group->d.unit_pages == nr_pages) {
977 max_pages = nr_pages;
978 } else {
979 max_pages = group->d.unit_pages;
980 /*
981 * Take a look at all online sections in the current unit.
982 * We can safely assume that all pages within a section belong
983 * to the same zone, because dynamic memory groups only deal
984 * with hotplugged memory.
985 */
986 pfn = ALIGN_DOWN(pfn, group->d.unit_pages);
987 end_pfn = pfn + group->d.unit_pages;
988 for (; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
989 page = pfn_to_online_page(pfn);
990 if (!page)
991 continue;
992 /* If anything is !MOVABLE online the rest !MOVABLE. */
993 if (!is_zone_movable_page(page))
994 goto kernel_zone;
995 online_pages += PAGES_PER_SECTION;
996 }
997 }
998
999 /*
1000 * Online MOVABLE if we could *currently* online all remaining parts
1001 * MOVABLE. We expect to (add+) online them immediately next, so if
1002 * nobody interferes, all will be MOVABLE if possible.
1003 */
1004 nr_pages = max_pages - online_pages;
1005 if (!auto_movable_can_online_movable(NUMA_NO_NODE, group, nr_pages))
1006 goto kernel_zone;
1007
1008#ifdef CONFIG_NUMA
1009 if (auto_movable_numa_aware &&
1010 !auto_movable_can_online_movable(nid, group, nr_pages))
1011 goto kernel_zone;
1012#endif /* CONFIG_NUMA */
1013
1014 return &NODE_DATA(nid)->node_zones[ZONE_MOVABLE];
1015kernel_zone:
1016 return default_kernel_zone_for_pfn(nid, pfn, nr_pages);
1017}
1018
1019static inline struct zone *default_zone_for_pfn(int nid, unsigned long start_pfn,
1020 unsigned long nr_pages)
1021{
1022 struct zone *kernel_zone = default_kernel_zone_for_pfn(nid, start_pfn,
1023 nr_pages);
1024 struct zone *movable_zone = &NODE_DATA(nid)->node_zones[ZONE_MOVABLE];
1025 bool in_kernel = zone_intersects(kernel_zone, start_pfn, nr_pages);
1026 bool in_movable = zone_intersects(movable_zone, start_pfn, nr_pages);
1027
1028 /*
1029 * We inherit the existing zone in a simple case where zones do not
1030 * overlap in the given range
1031 */
1032 if (in_kernel ^ in_movable)
1033 return (in_kernel) ? kernel_zone : movable_zone;
1034
1035 /*
1036 * If the range doesn't belong to any zone or two zones overlap in the
1037 * given range then we use movable zone only if movable_node is
1038 * enabled because we always online to a kernel zone by default.
1039 */
1040 return movable_node_enabled ? movable_zone : kernel_zone;
1041}
1042
1043struct zone *zone_for_pfn_range(int online_type, int nid,
1044 struct memory_group *group, unsigned long start_pfn,
1045 unsigned long nr_pages)
1046{
1047 if (online_type == MMOP_ONLINE_KERNEL)
1048 return default_kernel_zone_for_pfn(nid, start_pfn, nr_pages);
1049
1050 if (online_type == MMOP_ONLINE_MOVABLE)
1051 return &NODE_DATA(nid)->node_zones[ZONE_MOVABLE];
1052
1053 if (online_policy == ONLINE_POLICY_AUTO_MOVABLE)
1054 return auto_movable_zone_for_pfn(nid, group, start_pfn, nr_pages);
1055
1056 return default_zone_for_pfn(nid, start_pfn, nr_pages);
1057}
1058
1059/*
1060 * This function should only be called by memory_block_{online,offline},
1061 * and {online,offline}_pages.
1062 */
1063void adjust_present_page_count(struct page *page, struct memory_group *group,
1064 long nr_pages)
1065{
1066 struct zone *zone = page_zone(page);
1067 const bool movable = zone_idx(zone) == ZONE_MOVABLE;
1068
1069 /*
1070 * We only support onlining/offlining/adding/removing of complete
1071 * memory blocks; therefore, either all is either early or hotplugged.
1072 */
1073 if (early_section(__pfn_to_section(page_to_pfn(page))))
1074 zone->present_early_pages += nr_pages;
1075 zone->present_pages += nr_pages;
1076 zone->zone_pgdat->node_present_pages += nr_pages;
1077
1078 if (group && movable)
1079 group->present_movable_pages += nr_pages;
1080 else if (group && !movable)
1081 group->present_kernel_pages += nr_pages;
1082}
1083
1084int mhp_init_memmap_on_memory(unsigned long pfn, unsigned long nr_pages,
1085 struct zone *zone, bool mhp_off_inaccessible)
1086{
1087 unsigned long end_pfn = pfn + nr_pages;
1088 int ret, i;
1089
1090 ret = kasan_add_zero_shadow(__va(PFN_PHYS(pfn)), PFN_PHYS(nr_pages));
1091 if (ret)
1092 return ret;
1093
1094 /*
1095 * Memory block is accessible at this stage and hence poison the struct
1096 * pages now. If the memory block is accessible during memory hotplug
1097 * addition phase, then page poisining is already performed in
1098 * sparse_add_section().
1099 */
1100 if (mhp_off_inaccessible)
1101 page_init_poison(pfn_to_page(pfn), sizeof(struct page) * nr_pages);
1102
1103 move_pfn_range_to_zone(zone, pfn, nr_pages, NULL, MIGRATE_UNMOVABLE);
1104
1105 for (i = 0; i < nr_pages; i++) {
1106 struct page *page = pfn_to_page(pfn + i);
1107
1108 __ClearPageOffline(page);
1109 SetPageVmemmapSelfHosted(page);
1110 }
1111
1112 /*
1113 * It might be that the vmemmap_pages fully span sections. If that is
1114 * the case, mark those sections online here as otherwise they will be
1115 * left offline.
1116 */
1117 if (nr_pages >= PAGES_PER_SECTION)
1118 online_mem_sections(pfn, ALIGN_DOWN(end_pfn, PAGES_PER_SECTION));
1119
1120 return ret;
1121}
1122
1123void mhp_deinit_memmap_on_memory(unsigned long pfn, unsigned long nr_pages)
1124{
1125 unsigned long end_pfn = pfn + nr_pages;
1126
1127 /*
1128 * It might be that the vmemmap_pages fully span sections. If that is
1129 * the case, mark those sections offline here as otherwise they will be
1130 * left online.
1131 */
1132 if (nr_pages >= PAGES_PER_SECTION)
1133 offline_mem_sections(pfn, ALIGN_DOWN(end_pfn, PAGES_PER_SECTION));
1134
1135 /*
1136 * The pages associated with this vmemmap have been offlined, so
1137 * we can reset its state here.
1138 */
1139 remove_pfn_range_from_zone(page_zone(pfn_to_page(pfn)), pfn, nr_pages);
1140 kasan_remove_zero_shadow(__va(PFN_PHYS(pfn)), PFN_PHYS(nr_pages));
1141}
1142
1143/*
1144 * Must be called with mem_hotplug_lock in write mode.
1145 */
1146int online_pages(unsigned long pfn, unsigned long nr_pages,
1147 struct zone *zone, struct memory_group *group)
1148{
1149 unsigned long flags;
1150 int need_zonelists_rebuild = 0;
1151 const int nid = zone_to_nid(zone);
1152 int ret;
1153 struct memory_notify arg;
1154
1155 /*
1156 * {on,off}lining is constrained to full memory sections (or more
1157 * precisely to memory blocks from the user space POV).
1158 * memmap_on_memory is an exception because it reserves initial part
1159 * of the physical memory space for vmemmaps. That space is pageblock
1160 * aligned.
1161 */
1162 if (WARN_ON_ONCE(!nr_pages || !pageblock_aligned(pfn) ||
1163 !IS_ALIGNED(pfn + nr_pages, PAGES_PER_SECTION)))
1164 return -EINVAL;
1165
1166
1167 /* associate pfn range with the zone */
1168 move_pfn_range_to_zone(zone, pfn, nr_pages, NULL, MIGRATE_ISOLATE);
1169
1170 arg.start_pfn = pfn;
1171 arg.nr_pages = nr_pages;
1172 node_states_check_changes_online(nr_pages, zone, &arg);
1173
1174 ret = memory_notify(MEM_GOING_ONLINE, &arg);
1175 ret = notifier_to_errno(ret);
1176 if (ret)
1177 goto failed_addition;
1178
1179 /*
1180 * Fixup the number of isolated pageblocks before marking the sections
1181 * onlining, such that undo_isolate_page_range() works correctly.
1182 */
1183 spin_lock_irqsave(&zone->lock, flags);
1184 zone->nr_isolate_pageblock += nr_pages / pageblock_nr_pages;
1185 spin_unlock_irqrestore(&zone->lock, flags);
1186
1187 /*
1188 * If this zone is not populated, then it is not in zonelist.
1189 * This means the page allocator ignores this zone.
1190 * So, zonelist must be updated after online.
1191 */
1192 if (!populated_zone(zone)) {
1193 need_zonelists_rebuild = 1;
1194 setup_zone_pageset(zone);
1195 }
1196
1197 online_pages_range(pfn, nr_pages);
1198 adjust_present_page_count(pfn_to_page(pfn), group, nr_pages);
1199
1200 node_states_set_node(nid, &arg);
1201 if (need_zonelists_rebuild)
1202 build_all_zonelists(NULL);
1203
1204 /* Basic onlining is complete, allow allocation of onlined pages. */
1205 undo_isolate_page_range(pfn, pfn + nr_pages, MIGRATE_MOVABLE);
1206
1207 /*
1208 * Freshly onlined pages aren't shuffled (e.g., all pages are placed to
1209 * the tail of the freelist when undoing isolation). Shuffle the whole
1210 * zone to make sure the just onlined pages are properly distributed
1211 * across the whole freelist - to create an initial shuffle.
1212 */
1213 shuffle_zone(zone);
1214
1215 /* reinitialise watermarks and update pcp limits */
1216 init_per_zone_wmark_min();
1217
1218 kswapd_run(nid);
1219 kcompactd_run(nid);
1220
1221 writeback_set_ratelimit();
1222
1223 memory_notify(MEM_ONLINE, &arg);
1224 return 0;
1225
1226failed_addition:
1227 pr_debug("online_pages [mem %#010llx-%#010llx] failed\n",
1228 (unsigned long long) pfn << PAGE_SHIFT,
1229 (((unsigned long long) pfn + nr_pages) << PAGE_SHIFT) - 1);
1230 memory_notify(MEM_CANCEL_ONLINE, &arg);
1231 remove_pfn_range_from_zone(zone, pfn, nr_pages);
1232 return ret;
1233}
1234
1235/* we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG */
1236static pg_data_t *hotadd_init_pgdat(int nid)
1237{
1238 struct pglist_data *pgdat;
1239
1240 /*
1241 * NODE_DATA is preallocated (free_area_init) but its internal
1242 * state is not allocated completely. Add missing pieces.
1243 * Completely offline nodes stay around and they just need
1244 * reintialization.
1245 */
1246 pgdat = NODE_DATA(nid);
1247
1248 /* init node's zones as empty zones, we don't have any present pages.*/
1249 free_area_init_core_hotplug(pgdat);
1250
1251 /*
1252 * The node we allocated has no zone fallback lists. For avoiding
1253 * to access not-initialized zonelist, build here.
1254 */
1255 build_all_zonelists(pgdat);
1256
1257 return pgdat;
1258}
1259
1260/*
1261 * __try_online_node - online a node if offlined
1262 * @nid: the node ID
1263 * @set_node_online: Whether we want to online the node
1264 * called by cpu_up() to online a node without onlined memory.
1265 *
1266 * Returns:
1267 * 1 -> a new node has been allocated
1268 * 0 -> the node is already online
1269 * -ENOMEM -> the node could not be allocated
1270 */
1271static int __try_online_node(int nid, bool set_node_online)
1272{
1273 pg_data_t *pgdat;
1274 int ret = 1;
1275
1276 if (node_online(nid))
1277 return 0;
1278
1279 pgdat = hotadd_init_pgdat(nid);
1280 if (!pgdat) {
1281 pr_err("Cannot online node %d due to NULL pgdat\n", nid);
1282 ret = -ENOMEM;
1283 goto out;
1284 }
1285
1286 if (set_node_online) {
1287 node_set_online(nid);
1288 ret = register_one_node(nid);
1289 BUG_ON(ret);
1290 }
1291out:
1292 return ret;
1293}
1294
1295/*
1296 * Users of this function always want to online/register the node
1297 */
1298int try_online_node(int nid)
1299{
1300 int ret;
1301
1302 mem_hotplug_begin();
1303 ret = __try_online_node(nid, true);
1304 mem_hotplug_done();
1305 return ret;
1306}
1307
1308static int check_hotplug_memory_range(u64 start, u64 size)
1309{
1310 /* memory range must be block size aligned */
1311 if (!size || !IS_ALIGNED(start, memory_block_size_bytes()) ||
1312 !IS_ALIGNED(size, memory_block_size_bytes())) {
1313 pr_err("Block size [%#lx] unaligned hotplug range: start %#llx, size %#llx",
1314 memory_block_size_bytes(), start, size);
1315 return -EINVAL;
1316 }
1317
1318 return 0;
1319}
1320
1321static int online_memory_block(struct memory_block *mem, void *arg)
1322{
1323 mem->online_type = mhp_default_online_type;
1324 return device_online(&mem->dev);
1325}
1326
1327#ifndef arch_supports_memmap_on_memory
1328static inline bool arch_supports_memmap_on_memory(unsigned long vmemmap_size)
1329{
1330 /*
1331 * As default, we want the vmemmap to span a complete PMD such that we
1332 * can map the vmemmap using a single PMD if supported by the
1333 * architecture.
1334 */
1335 return IS_ALIGNED(vmemmap_size, PMD_SIZE);
1336}
1337#endif
1338
1339bool mhp_supports_memmap_on_memory(void)
1340{
1341 unsigned long vmemmap_size = memory_block_memmap_size();
1342 unsigned long memmap_pages = memory_block_memmap_on_memory_pages();
1343
1344 /*
1345 * Besides having arch support and the feature enabled at runtime, we
1346 * need a few more assumptions to hold true:
1347 *
1348 * a) The vmemmap pages span complete PMDs: We don't want vmemmap code
1349 * to populate memory from the altmap for unrelated parts (i.e.,
1350 * other memory blocks)
1351 *
1352 * b) The vmemmap pages (and thereby the pages that will be exposed to
1353 * the buddy) have to cover full pageblocks: memory onlining/offlining
1354 * code requires applicable ranges to be page-aligned, for example, to
1355 * set the migratetypes properly.
1356 *
1357 * TODO: Although we have a check here to make sure that vmemmap pages
1358 * fully populate a PMD, it is not the right place to check for
1359 * this. A much better solution involves improving vmemmap code
1360 * to fallback to base pages when trying to populate vmemmap using
1361 * altmap as an alternative source of memory, and we do not exactly
1362 * populate a single PMD.
1363 */
1364 if (!mhp_memmap_on_memory())
1365 return false;
1366
1367 /*
1368 * Make sure the vmemmap allocation is fully contained
1369 * so that we always allocate vmemmap memory from altmap area.
1370 */
1371 if (!IS_ALIGNED(vmemmap_size, PAGE_SIZE))
1372 return false;
1373
1374 /*
1375 * start pfn should be pageblock_nr_pages aligned for correctly
1376 * setting migrate types
1377 */
1378 if (!pageblock_aligned(memmap_pages))
1379 return false;
1380
1381 if (memmap_pages == PHYS_PFN(memory_block_size_bytes()))
1382 /* No effective hotplugged memory doesn't make sense. */
1383 return false;
1384
1385 return arch_supports_memmap_on_memory(vmemmap_size);
1386}
1387EXPORT_SYMBOL_GPL(mhp_supports_memmap_on_memory);
1388
1389static void remove_memory_blocks_and_altmaps(u64 start, u64 size)
1390{
1391 unsigned long memblock_size = memory_block_size_bytes();
1392 u64 cur_start;
1393
1394 /*
1395 * For memmap_on_memory, the altmaps were added on a per-memblock
1396 * basis; we have to process each individual memory block.
1397 */
1398 for (cur_start = start; cur_start < start + size;
1399 cur_start += memblock_size) {
1400 struct vmem_altmap *altmap = NULL;
1401 struct memory_block *mem;
1402
1403 mem = find_memory_block(pfn_to_section_nr(PFN_DOWN(cur_start)));
1404 if (WARN_ON_ONCE(!mem))
1405 continue;
1406
1407 altmap = mem->altmap;
1408 mem->altmap = NULL;
1409
1410 remove_memory_block_devices(cur_start, memblock_size);
1411
1412 arch_remove_memory(cur_start, memblock_size, altmap);
1413
1414 /* Verify that all vmemmap pages have actually been freed. */
1415 WARN(altmap->alloc, "Altmap not fully unmapped");
1416 kfree(altmap);
1417 }
1418}
1419
1420static int create_altmaps_and_memory_blocks(int nid, struct memory_group *group,
1421 u64 start, u64 size, mhp_t mhp_flags)
1422{
1423 unsigned long memblock_size = memory_block_size_bytes();
1424 u64 cur_start;
1425 int ret;
1426
1427 for (cur_start = start; cur_start < start + size;
1428 cur_start += memblock_size) {
1429 struct mhp_params params = { .pgprot =
1430 pgprot_mhp(PAGE_KERNEL) };
1431 struct vmem_altmap mhp_altmap = {
1432 .base_pfn = PHYS_PFN(cur_start),
1433 .end_pfn = PHYS_PFN(cur_start + memblock_size - 1),
1434 };
1435
1436 mhp_altmap.free = memory_block_memmap_on_memory_pages();
1437 if (mhp_flags & MHP_OFFLINE_INACCESSIBLE)
1438 mhp_altmap.inaccessible = true;
1439 params.altmap = kmemdup(&mhp_altmap, sizeof(struct vmem_altmap),
1440 GFP_KERNEL);
1441 if (!params.altmap) {
1442 ret = -ENOMEM;
1443 goto out;
1444 }
1445
1446 /* call arch's memory hotadd */
1447 ret = arch_add_memory(nid, cur_start, memblock_size, ¶ms);
1448 if (ret < 0) {
1449 kfree(params.altmap);
1450 goto out;
1451 }
1452
1453 /* create memory block devices after memory was added */
1454 ret = create_memory_block_devices(cur_start, memblock_size,
1455 params.altmap, group);
1456 if (ret) {
1457 arch_remove_memory(cur_start, memblock_size, NULL);
1458 kfree(params.altmap);
1459 goto out;
1460 }
1461 }
1462
1463 return 0;
1464out:
1465 if (ret && cur_start != start)
1466 remove_memory_blocks_and_altmaps(start, cur_start - start);
1467 return ret;
1468}
1469
1470/*
1471 * NOTE: The caller must call lock_device_hotplug() to serialize hotplug
1472 * and online/offline operations (triggered e.g. by sysfs).
1473 *
1474 * we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG
1475 */
1476int add_memory_resource(int nid, struct resource *res, mhp_t mhp_flags)
1477{
1478 struct mhp_params params = { .pgprot = pgprot_mhp(PAGE_KERNEL) };
1479 enum memblock_flags memblock_flags = MEMBLOCK_NONE;
1480 struct memory_group *group = NULL;
1481 u64 start, size;
1482 bool new_node = false;
1483 int ret;
1484
1485 start = res->start;
1486 size = resource_size(res);
1487
1488 ret = check_hotplug_memory_range(start, size);
1489 if (ret)
1490 return ret;
1491
1492 if (mhp_flags & MHP_NID_IS_MGID) {
1493 group = memory_group_find_by_id(nid);
1494 if (!group)
1495 return -EINVAL;
1496 nid = group->nid;
1497 }
1498
1499 if (!node_possible(nid)) {
1500 WARN(1, "node %d was absent from the node_possible_map\n", nid);
1501 return -EINVAL;
1502 }
1503
1504 mem_hotplug_begin();
1505
1506 if (IS_ENABLED(CONFIG_ARCH_KEEP_MEMBLOCK)) {
1507 if (res->flags & IORESOURCE_SYSRAM_DRIVER_MANAGED)
1508 memblock_flags = MEMBLOCK_DRIVER_MANAGED;
1509 ret = memblock_add_node(start, size, nid, memblock_flags);
1510 if (ret)
1511 goto error_mem_hotplug_end;
1512 }
1513
1514 ret = __try_online_node(nid, false);
1515 if (ret < 0)
1516 goto error;
1517 new_node = ret;
1518
1519 /*
1520 * Self hosted memmap array
1521 */
1522 if ((mhp_flags & MHP_MEMMAP_ON_MEMORY) &&
1523 mhp_supports_memmap_on_memory()) {
1524 ret = create_altmaps_and_memory_blocks(nid, group, start, size, mhp_flags);
1525 if (ret)
1526 goto error;
1527 } else {
1528 ret = arch_add_memory(nid, start, size, ¶ms);
1529 if (ret < 0)
1530 goto error;
1531
1532 /* create memory block devices after memory was added */
1533 ret = create_memory_block_devices(start, size, NULL, group);
1534 if (ret) {
1535 arch_remove_memory(start, size, params.altmap);
1536 goto error;
1537 }
1538 }
1539
1540 if (new_node) {
1541 /* If sysfs file of new node can't be created, cpu on the node
1542 * can't be hot-added. There is no rollback way now.
1543 * So, check by BUG_ON() to catch it reluctantly..
1544 * We online node here. We can't roll back from here.
1545 */
1546 node_set_online(nid);
1547 ret = __register_one_node(nid);
1548 BUG_ON(ret);
1549 }
1550
1551 register_memory_blocks_under_node(nid, PFN_DOWN(start),
1552 PFN_UP(start + size - 1),
1553 MEMINIT_HOTPLUG);
1554
1555 /* create new memmap entry */
1556 if (!strcmp(res->name, "System RAM"))
1557 firmware_map_add_hotplug(start, start + size, "System RAM");
1558
1559 /* device_online() will take the lock when calling online_pages() */
1560 mem_hotplug_done();
1561
1562 /*
1563 * In case we're allowed to merge the resource, flag it and trigger
1564 * merging now that adding succeeded.
1565 */
1566 if (mhp_flags & MHP_MERGE_RESOURCE)
1567 merge_system_ram_resource(res);
1568
1569 /* online pages if requested */
1570 if (mhp_default_online_type != MMOP_OFFLINE)
1571 walk_memory_blocks(start, size, NULL, online_memory_block);
1572
1573 return ret;
1574error:
1575 if (IS_ENABLED(CONFIG_ARCH_KEEP_MEMBLOCK))
1576 memblock_remove(start, size);
1577error_mem_hotplug_end:
1578 mem_hotplug_done();
1579 return ret;
1580}
1581
1582/* requires device_hotplug_lock, see add_memory_resource() */
1583int __add_memory(int nid, u64 start, u64 size, mhp_t mhp_flags)
1584{
1585 struct resource *res;
1586 int ret;
1587
1588 res = register_memory_resource(start, size, "System RAM");
1589 if (IS_ERR(res))
1590 return PTR_ERR(res);
1591
1592 ret = add_memory_resource(nid, res, mhp_flags);
1593 if (ret < 0)
1594 release_memory_resource(res);
1595 return ret;
1596}
1597
1598int add_memory(int nid, u64 start, u64 size, mhp_t mhp_flags)
1599{
1600 int rc;
1601
1602 lock_device_hotplug();
1603 rc = __add_memory(nid, start, size, mhp_flags);
1604 unlock_device_hotplug();
1605
1606 return rc;
1607}
1608EXPORT_SYMBOL_GPL(add_memory);
1609
1610/*
1611 * Add special, driver-managed memory to the system as system RAM. Such
1612 * memory is not exposed via the raw firmware-provided memmap as system
1613 * RAM, instead, it is detected and added by a driver - during cold boot,
1614 * after a reboot, and after kexec.
1615 *
1616 * Reasons why this memory should not be used for the initial memmap of a
1617 * kexec kernel or for placing kexec images:
1618 * - The booting kernel is in charge of determining how this memory will be
1619 * used (e.g., use persistent memory as system RAM)
1620 * - Coordination with a hypervisor is required before this memory
1621 * can be used (e.g., inaccessible parts).
1622 *
1623 * For this memory, no entries in /sys/firmware/memmap ("raw firmware-provided
1624 * memory map") are created. Also, the created memory resource is flagged
1625 * with IORESOURCE_SYSRAM_DRIVER_MANAGED, so in-kernel users can special-case
1626 * this memory as well (esp., not place kexec images onto it).
1627 *
1628 * The resource_name (visible via /proc/iomem) has to have the format
1629 * "System RAM ($DRIVER)".
1630 */
1631int add_memory_driver_managed(int nid, u64 start, u64 size,
1632 const char *resource_name, mhp_t mhp_flags)
1633{
1634 struct resource *res;
1635 int rc;
1636
1637 if (!resource_name ||
1638 strstr(resource_name, "System RAM (") != resource_name ||
1639 resource_name[strlen(resource_name) - 1] != ')')
1640 return -EINVAL;
1641
1642 lock_device_hotplug();
1643
1644 res = register_memory_resource(start, size, resource_name);
1645 if (IS_ERR(res)) {
1646 rc = PTR_ERR(res);
1647 goto out_unlock;
1648 }
1649
1650 rc = add_memory_resource(nid, res, mhp_flags);
1651 if (rc < 0)
1652 release_memory_resource(res);
1653
1654out_unlock:
1655 unlock_device_hotplug();
1656 return rc;
1657}
1658EXPORT_SYMBOL_GPL(add_memory_driver_managed);
1659
1660/*
1661 * Platforms should define arch_get_mappable_range() that provides
1662 * maximum possible addressable physical memory range for which the
1663 * linear mapping could be created. The platform returned address
1664 * range must adhere to these following semantics.
1665 *
1666 * - range.start <= range.end
1667 * - Range includes both end points [range.start..range.end]
1668 *
1669 * There is also a fallback definition provided here, allowing the
1670 * entire possible physical address range in case any platform does
1671 * not define arch_get_mappable_range().
1672 */
1673struct range __weak arch_get_mappable_range(void)
1674{
1675 struct range mhp_range = {
1676 .start = 0UL,
1677 .end = -1ULL,
1678 };
1679 return mhp_range;
1680}
1681
1682struct range mhp_get_pluggable_range(bool need_mapping)
1683{
1684 const u64 max_phys = DIRECT_MAP_PHYSMEM_END;
1685 struct range mhp_range;
1686
1687 if (need_mapping) {
1688 mhp_range = arch_get_mappable_range();
1689 if (mhp_range.start > max_phys) {
1690 mhp_range.start = 0;
1691 mhp_range.end = 0;
1692 }
1693 mhp_range.end = min_t(u64, mhp_range.end, max_phys);
1694 } else {
1695 mhp_range.start = 0;
1696 mhp_range.end = max_phys;
1697 }
1698 return mhp_range;
1699}
1700EXPORT_SYMBOL_GPL(mhp_get_pluggable_range);
1701
1702bool mhp_range_allowed(u64 start, u64 size, bool need_mapping)
1703{
1704 struct range mhp_range = mhp_get_pluggable_range(need_mapping);
1705 u64 end = start + size;
1706
1707 if (start < end && start >= mhp_range.start && (end - 1) <= mhp_range.end)
1708 return true;
1709
1710 pr_warn("Hotplug memory [%#llx-%#llx] exceeds maximum addressable range [%#llx-%#llx]\n",
1711 start, end, mhp_range.start, mhp_range.end);
1712 return false;
1713}
1714
1715#ifdef CONFIG_MEMORY_HOTREMOVE
1716/*
1717 * Scan pfn range [start,end) to find movable/migratable pages (LRU pages,
1718 * non-lru movable pages and hugepages). Will skip over most unmovable
1719 * pages (esp., pages that can be skipped when offlining), but bail out on
1720 * definitely unmovable pages.
1721 *
1722 * Returns:
1723 * 0 in case a movable page is found and movable_pfn was updated.
1724 * -ENOENT in case no movable page was found.
1725 * -EBUSY in case a definitely unmovable page was found.
1726 */
1727static int scan_movable_pages(unsigned long start, unsigned long end,
1728 unsigned long *movable_pfn)
1729{
1730 unsigned long pfn;
1731
1732 for (pfn = start; pfn < end; pfn++) {
1733 struct page *page;
1734 struct folio *folio;
1735
1736 if (!pfn_valid(pfn))
1737 continue;
1738 page = pfn_to_page(pfn);
1739 if (PageLRU(page))
1740 goto found;
1741 if (__PageMovable(page))
1742 goto found;
1743
1744 /*
1745 * PageOffline() pages that are not marked __PageMovable() and
1746 * have a reference count > 0 (after MEM_GOING_OFFLINE) are
1747 * definitely unmovable. If their reference count would be 0,
1748 * they could at least be skipped when offlining memory.
1749 */
1750 if (PageOffline(page) && page_count(page))
1751 return -EBUSY;
1752
1753 if (!PageHuge(page))
1754 continue;
1755 folio = page_folio(page);
1756 /*
1757 * This test is racy as we hold no reference or lock. The
1758 * hugetlb page could have been free'ed and head is no longer
1759 * a hugetlb page before the following check. In such unlikely
1760 * cases false positives and negatives are possible. Calling
1761 * code must deal with these scenarios.
1762 */
1763 if (folio_test_hugetlb_migratable(folio))
1764 goto found;
1765 pfn |= folio_nr_pages(folio) - 1;
1766 }
1767 return -ENOENT;
1768found:
1769 *movable_pfn = pfn;
1770 return 0;
1771}
1772
1773static void do_migrate_range(unsigned long start_pfn, unsigned long end_pfn)
1774{
1775 struct folio *folio;
1776 unsigned long pfn;
1777 LIST_HEAD(source);
1778 static DEFINE_RATELIMIT_STATE(migrate_rs, DEFAULT_RATELIMIT_INTERVAL,
1779 DEFAULT_RATELIMIT_BURST);
1780
1781 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
1782 struct page *page;
1783
1784 if (!pfn_valid(pfn))
1785 continue;
1786 page = pfn_to_page(pfn);
1787 folio = page_folio(page);
1788
1789 /*
1790 * No reference or lock is held on the folio, so it might
1791 * be modified concurrently (e.g. split). As such,
1792 * folio_nr_pages() may read garbage. This is fine as the outer
1793 * loop will revisit the split folio later.
1794 */
1795 if (folio_test_large(folio))
1796 pfn = folio_pfn(folio) + folio_nr_pages(folio) - 1;
1797
1798 if (!folio_try_get(folio))
1799 continue;
1800
1801 if (unlikely(page_folio(page) != folio))
1802 goto put_folio;
1803
1804 if (folio_test_hwpoison(folio) ||
1805 (folio_test_large(folio) && folio_test_has_hwpoisoned(folio))) {
1806 if (WARN_ON(folio_test_lru(folio)))
1807 folio_isolate_lru(folio);
1808 if (folio_mapped(folio)) {
1809 folio_lock(folio);
1810 unmap_poisoned_folio(folio, pfn, false);
1811 folio_unlock(folio);
1812 }
1813
1814 goto put_folio;
1815 }
1816
1817 if (!isolate_folio_to_list(folio, &source)) {
1818 if (__ratelimit(&migrate_rs)) {
1819 pr_warn("failed to isolate pfn %lx\n",
1820 page_to_pfn(page));
1821 dump_page(page, "isolation failed");
1822 }
1823 }
1824put_folio:
1825 folio_put(folio);
1826 }
1827 if (!list_empty(&source)) {
1828 nodemask_t nmask = node_states[N_MEMORY];
1829 struct migration_target_control mtc = {
1830 .nmask = &nmask,
1831 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
1832 .reason = MR_MEMORY_HOTPLUG,
1833 };
1834 int ret;
1835
1836 /*
1837 * We have checked that migration range is on a single zone so
1838 * we can use the nid of the first page to all the others.
1839 */
1840 mtc.nid = folio_nid(list_first_entry(&source, struct folio, lru));
1841
1842 /*
1843 * try to allocate from a different node but reuse this node
1844 * if there are no other online nodes to be used (e.g. we are
1845 * offlining a part of the only existing node)
1846 */
1847 node_clear(mtc.nid, nmask);
1848 if (nodes_empty(nmask))
1849 node_set(mtc.nid, nmask);
1850 ret = migrate_pages(&source, alloc_migration_target, NULL,
1851 (unsigned long)&mtc, MIGRATE_SYNC, MR_MEMORY_HOTPLUG, NULL);
1852 if (ret) {
1853 list_for_each_entry(folio, &source, lru) {
1854 if (__ratelimit(&migrate_rs)) {
1855 pr_warn("migrating pfn %lx failed ret:%d\n",
1856 folio_pfn(folio), ret);
1857 dump_page(&folio->page,
1858 "migration failure");
1859 }
1860 }
1861 putback_movable_pages(&source);
1862 }
1863 }
1864}
1865
1866static int __init cmdline_parse_movable_node(char *p)
1867{
1868 movable_node_enabled = true;
1869 return 0;
1870}
1871early_param("movable_node", cmdline_parse_movable_node);
1872
1873/* check which state of node_states will be changed when offline memory */
1874static void node_states_check_changes_offline(unsigned long nr_pages,
1875 struct zone *zone, struct memory_notify *arg)
1876{
1877 struct pglist_data *pgdat = zone->zone_pgdat;
1878 unsigned long present_pages = 0;
1879 enum zone_type zt;
1880
1881 arg->status_change_nid = NUMA_NO_NODE;
1882 arg->status_change_nid_normal = NUMA_NO_NODE;
1883
1884 /*
1885 * Check whether node_states[N_NORMAL_MEMORY] will be changed.
1886 * If the memory to be offline is within the range
1887 * [0..ZONE_NORMAL], and it is the last present memory there,
1888 * the zones in that range will become empty after the offlining,
1889 * thus we can determine that we need to clear the node from
1890 * node_states[N_NORMAL_MEMORY].
1891 */
1892 for (zt = 0; zt <= ZONE_NORMAL; zt++)
1893 present_pages += pgdat->node_zones[zt].present_pages;
1894 if (zone_idx(zone) <= ZONE_NORMAL && nr_pages >= present_pages)
1895 arg->status_change_nid_normal = zone_to_nid(zone);
1896
1897 /*
1898 * We have accounted the pages from [0..ZONE_NORMAL); ZONE_HIGHMEM
1899 * does not apply as we don't support 32bit.
1900 * Here we count the possible pages from ZONE_MOVABLE.
1901 * If after having accounted all the pages, we see that the nr_pages
1902 * to be offlined is over or equal to the accounted pages,
1903 * we know that the node will become empty, and so, we can clear
1904 * it for N_MEMORY as well.
1905 */
1906 present_pages += pgdat->node_zones[ZONE_MOVABLE].present_pages;
1907
1908 if (nr_pages >= present_pages)
1909 arg->status_change_nid = zone_to_nid(zone);
1910}
1911
1912static void node_states_clear_node(int node, struct memory_notify *arg)
1913{
1914 if (arg->status_change_nid_normal >= 0)
1915 node_clear_state(node, N_NORMAL_MEMORY);
1916
1917 if (arg->status_change_nid >= 0)
1918 node_clear_state(node, N_MEMORY);
1919}
1920
1921static int count_system_ram_pages_cb(unsigned long start_pfn,
1922 unsigned long nr_pages, void *data)
1923{
1924 unsigned long *nr_system_ram_pages = data;
1925
1926 *nr_system_ram_pages += nr_pages;
1927 return 0;
1928}
1929
1930/*
1931 * Must be called with mem_hotplug_lock in write mode.
1932 */
1933int offline_pages(unsigned long start_pfn, unsigned long nr_pages,
1934 struct zone *zone, struct memory_group *group)
1935{
1936 const unsigned long end_pfn = start_pfn + nr_pages;
1937 unsigned long pfn, managed_pages, system_ram_pages = 0;
1938 const int node = zone_to_nid(zone);
1939 unsigned long flags;
1940 struct memory_notify arg;
1941 char *reason;
1942 int ret;
1943
1944 /*
1945 * {on,off}lining is constrained to full memory sections (or more
1946 * precisely to memory blocks from the user space POV).
1947 * memmap_on_memory is an exception because it reserves initial part
1948 * of the physical memory space for vmemmaps. That space is pageblock
1949 * aligned.
1950 */
1951 if (WARN_ON_ONCE(!nr_pages || !pageblock_aligned(start_pfn) ||
1952 !IS_ALIGNED(start_pfn + nr_pages, PAGES_PER_SECTION)))
1953 return -EINVAL;
1954
1955 /*
1956 * Don't allow to offline memory blocks that contain holes.
1957 * Consequently, memory blocks with holes can never get onlined
1958 * via the hotplug path - online_pages() - as hotplugged memory has
1959 * no holes. This way, we don't have to worry about memory holes,
1960 * don't need pfn_valid() checks, and can avoid using
1961 * walk_system_ram_range() later.
1962 */
1963 walk_system_ram_range(start_pfn, nr_pages, &system_ram_pages,
1964 count_system_ram_pages_cb);
1965 if (system_ram_pages != nr_pages) {
1966 ret = -EINVAL;
1967 reason = "memory holes";
1968 goto failed_removal;
1969 }
1970
1971 /*
1972 * We only support offlining of memory blocks managed by a single zone,
1973 * checked by calling code. This is just a sanity check that we might
1974 * want to remove in the future.
1975 */
1976 if (WARN_ON_ONCE(page_zone(pfn_to_page(start_pfn)) != zone ||
1977 page_zone(pfn_to_page(end_pfn - 1)) != zone)) {
1978 ret = -EINVAL;
1979 reason = "multizone range";
1980 goto failed_removal;
1981 }
1982
1983 /*
1984 * Disable pcplists so that page isolation cannot race with freeing
1985 * in a way that pages from isolated pageblock are left on pcplists.
1986 */
1987 zone_pcp_disable(zone);
1988 lru_cache_disable();
1989
1990 /* set above range as isolated */
1991 ret = start_isolate_page_range(start_pfn, end_pfn,
1992 MIGRATE_MOVABLE,
1993 MEMORY_OFFLINE | REPORT_FAILURE,
1994 GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL);
1995 if (ret) {
1996 reason = "failure to isolate range";
1997 goto failed_removal_pcplists_disabled;
1998 }
1999
2000 arg.start_pfn = start_pfn;
2001 arg.nr_pages = nr_pages;
2002 node_states_check_changes_offline(nr_pages, zone, &arg);
2003
2004 ret = memory_notify(MEM_GOING_OFFLINE, &arg);
2005 ret = notifier_to_errno(ret);
2006 if (ret) {
2007 reason = "notifier failure";
2008 goto failed_removal_isolated;
2009 }
2010
2011 do {
2012 pfn = start_pfn;
2013 do {
2014 /*
2015 * Historically we always checked for any signal and
2016 * can't limit it to fatal signals without eventually
2017 * breaking user space.
2018 */
2019 if (signal_pending(current)) {
2020 ret = -EINTR;
2021 reason = "signal backoff";
2022 goto failed_removal_isolated;
2023 }
2024
2025 cond_resched();
2026
2027 ret = scan_movable_pages(pfn, end_pfn, &pfn);
2028 if (!ret) {
2029 /*
2030 * TODO: fatal migration failures should bail
2031 * out
2032 */
2033 do_migrate_range(pfn, end_pfn);
2034 }
2035 } while (!ret);
2036
2037 if (ret != -ENOENT) {
2038 reason = "unmovable page";
2039 goto failed_removal_isolated;
2040 }
2041
2042 /*
2043 * Dissolve free hugetlb folios in the memory block before doing
2044 * offlining actually in order to make hugetlbfs's object
2045 * counting consistent.
2046 */
2047 ret = dissolve_free_hugetlb_folios(start_pfn, end_pfn);
2048 if (ret) {
2049 reason = "failure to dissolve huge pages";
2050 goto failed_removal_isolated;
2051 }
2052
2053 ret = test_pages_isolated(start_pfn, end_pfn, MEMORY_OFFLINE);
2054
2055 } while (ret);
2056
2057 /* Mark all sections offline and remove free pages from the buddy. */
2058 managed_pages = __offline_isolated_pages(start_pfn, end_pfn);
2059 pr_debug("Offlined Pages %ld\n", nr_pages);
2060
2061 /*
2062 * The memory sections are marked offline, and the pageblock flags
2063 * effectively stale; nobody should be touching them. Fixup the number
2064 * of isolated pageblocks, memory onlining will properly revert this.
2065 */
2066 spin_lock_irqsave(&zone->lock, flags);
2067 zone->nr_isolate_pageblock -= nr_pages / pageblock_nr_pages;
2068 spin_unlock_irqrestore(&zone->lock, flags);
2069
2070 lru_cache_enable();
2071 zone_pcp_enable(zone);
2072
2073 /* removal success */
2074 adjust_managed_page_count(pfn_to_page(start_pfn), -managed_pages);
2075 adjust_present_page_count(pfn_to_page(start_pfn), group, -nr_pages);
2076
2077 /* reinitialise watermarks and update pcp limits */
2078 init_per_zone_wmark_min();
2079
2080 /*
2081 * Make sure to mark the node as memory-less before rebuilding the zone
2082 * list. Otherwise this node would still appear in the fallback lists.
2083 */
2084 node_states_clear_node(node, &arg);
2085 if (!populated_zone(zone)) {
2086 zone_pcp_reset(zone);
2087 build_all_zonelists(NULL);
2088 }
2089
2090 if (arg.status_change_nid >= 0) {
2091 kcompactd_stop(node);
2092 kswapd_stop(node);
2093 }
2094
2095 writeback_set_ratelimit();
2096
2097 memory_notify(MEM_OFFLINE, &arg);
2098 remove_pfn_range_from_zone(zone, start_pfn, nr_pages);
2099 return 0;
2100
2101failed_removal_isolated:
2102 /* pushback to free area */
2103 undo_isolate_page_range(start_pfn, end_pfn, MIGRATE_MOVABLE);
2104 memory_notify(MEM_CANCEL_OFFLINE, &arg);
2105failed_removal_pcplists_disabled:
2106 lru_cache_enable();
2107 zone_pcp_enable(zone);
2108failed_removal:
2109 pr_debug("memory offlining [mem %#010llx-%#010llx] failed due to %s\n",
2110 (unsigned long long) start_pfn << PAGE_SHIFT,
2111 ((unsigned long long) end_pfn << PAGE_SHIFT) - 1,
2112 reason);
2113 return ret;
2114}
2115
2116static int check_memblock_offlined_cb(struct memory_block *mem, void *arg)
2117{
2118 int *nid = arg;
2119
2120 *nid = mem->nid;
2121 if (unlikely(mem->state != MEM_OFFLINE)) {
2122 phys_addr_t beginpa, endpa;
2123
2124 beginpa = PFN_PHYS(section_nr_to_pfn(mem->start_section_nr));
2125 endpa = beginpa + memory_block_size_bytes() - 1;
2126 pr_warn("removing memory fails, because memory [%pa-%pa] is onlined\n",
2127 &beginpa, &endpa);
2128
2129 return -EBUSY;
2130 }
2131 return 0;
2132}
2133
2134static int count_memory_range_altmaps_cb(struct memory_block *mem, void *arg)
2135{
2136 u64 *num_altmaps = (u64 *)arg;
2137
2138 if (mem->altmap)
2139 *num_altmaps += 1;
2140
2141 return 0;
2142}
2143
2144static int check_cpu_on_node(int nid)
2145{
2146 int cpu;
2147
2148 for_each_present_cpu(cpu) {
2149 if (cpu_to_node(cpu) == nid)
2150 /*
2151 * the cpu on this node isn't removed, and we can't
2152 * offline this node.
2153 */
2154 return -EBUSY;
2155 }
2156
2157 return 0;
2158}
2159
2160static int check_no_memblock_for_node_cb(struct memory_block *mem, void *arg)
2161{
2162 int nid = *(int *)arg;
2163
2164 /*
2165 * If a memory block belongs to multiple nodes, the stored nid is not
2166 * reliable. However, such blocks are always online (e.g., cannot get
2167 * offlined) and, therefore, are still spanned by the node.
2168 */
2169 return mem->nid == nid ? -EEXIST : 0;
2170}
2171
2172/**
2173 * try_offline_node
2174 * @nid: the node ID
2175 *
2176 * Offline a node if all memory sections and cpus of the node are removed.
2177 *
2178 * NOTE: The caller must call lock_device_hotplug() to serialize hotplug
2179 * and online/offline operations before this call.
2180 */
2181void try_offline_node(int nid)
2182{
2183 int rc;
2184
2185 /*
2186 * If the node still spans pages (especially ZONE_DEVICE), don't
2187 * offline it. A node spans memory after move_pfn_range_to_zone(),
2188 * e.g., after the memory block was onlined.
2189 */
2190 if (node_spanned_pages(nid))
2191 return;
2192
2193 /*
2194 * Especially offline memory blocks might not be spanned by the
2195 * node. They will get spanned by the node once they get onlined.
2196 * However, they link to the node in sysfs and can get onlined later.
2197 */
2198 rc = for_each_memory_block(&nid, check_no_memblock_for_node_cb);
2199 if (rc)
2200 return;
2201
2202 if (check_cpu_on_node(nid))
2203 return;
2204
2205 /*
2206 * all memory/cpu of this node are removed, we can offline this
2207 * node now.
2208 */
2209 node_set_offline(nid);
2210 unregister_one_node(nid);
2211}
2212EXPORT_SYMBOL(try_offline_node);
2213
2214static int memory_blocks_have_altmaps(u64 start, u64 size)
2215{
2216 u64 num_memblocks = size / memory_block_size_bytes();
2217 u64 num_altmaps = 0;
2218
2219 if (!mhp_memmap_on_memory())
2220 return 0;
2221
2222 walk_memory_blocks(start, size, &num_altmaps,
2223 count_memory_range_altmaps_cb);
2224
2225 if (num_altmaps == 0)
2226 return 0;
2227
2228 if (WARN_ON_ONCE(num_memblocks != num_altmaps))
2229 return -EINVAL;
2230
2231 return 1;
2232}
2233
2234static int try_remove_memory(u64 start, u64 size)
2235{
2236 int rc, nid = NUMA_NO_NODE;
2237
2238 BUG_ON(check_hotplug_memory_range(start, size));
2239
2240 /*
2241 * All memory blocks must be offlined before removing memory. Check
2242 * whether all memory blocks in question are offline and return error
2243 * if this is not the case.
2244 *
2245 * While at it, determine the nid. Note that if we'd have mixed nodes,
2246 * we'd only try to offline the last determined one -- which is good
2247 * enough for the cases we care about.
2248 */
2249 rc = walk_memory_blocks(start, size, &nid, check_memblock_offlined_cb);
2250 if (rc)
2251 return rc;
2252
2253 /* remove memmap entry */
2254 firmware_map_remove(start, start + size, "System RAM");
2255
2256 mem_hotplug_begin();
2257
2258 rc = memory_blocks_have_altmaps(start, size);
2259 if (rc < 0) {
2260 mem_hotplug_done();
2261 return rc;
2262 } else if (!rc) {
2263 /*
2264 * Memory block device removal under the device_hotplug_lock is
2265 * a barrier against racing online attempts.
2266 * No altmaps present, do the removal directly
2267 */
2268 remove_memory_block_devices(start, size);
2269 arch_remove_memory(start, size, NULL);
2270 } else {
2271 /* all memblocks in the range have altmaps */
2272 remove_memory_blocks_and_altmaps(start, size);
2273 }
2274
2275 if (IS_ENABLED(CONFIG_ARCH_KEEP_MEMBLOCK))
2276 memblock_remove(start, size);
2277
2278 release_mem_region_adjustable(start, size);
2279
2280 if (nid != NUMA_NO_NODE)
2281 try_offline_node(nid);
2282
2283 mem_hotplug_done();
2284 return 0;
2285}
2286
2287/**
2288 * __remove_memory - Remove memory if every memory block is offline
2289 * @start: physical address of the region to remove
2290 * @size: size of the region to remove
2291 *
2292 * NOTE: The caller must call lock_device_hotplug() to serialize hotplug
2293 * and online/offline operations before this call, as required by
2294 * try_offline_node().
2295 */
2296void __remove_memory(u64 start, u64 size)
2297{
2298
2299 /*
2300 * trigger BUG() if some memory is not offlined prior to calling this
2301 * function
2302 */
2303 if (try_remove_memory(start, size))
2304 BUG();
2305}
2306
2307/*
2308 * Remove memory if every memory block is offline, otherwise return -EBUSY is
2309 * some memory is not offline
2310 */
2311int remove_memory(u64 start, u64 size)
2312{
2313 int rc;
2314
2315 lock_device_hotplug();
2316 rc = try_remove_memory(start, size);
2317 unlock_device_hotplug();
2318
2319 return rc;
2320}
2321EXPORT_SYMBOL_GPL(remove_memory);
2322
2323static int try_offline_memory_block(struct memory_block *mem, void *arg)
2324{
2325 uint8_t online_type = MMOP_ONLINE_KERNEL;
2326 uint8_t **online_types = arg;
2327 struct page *page;
2328 int rc;
2329
2330 /*
2331 * Sense the online_type via the zone of the memory block. Offlining
2332 * with multiple zones within one memory block will be rejected
2333 * by offlining code ... so we don't care about that.
2334 */
2335 page = pfn_to_online_page(section_nr_to_pfn(mem->start_section_nr));
2336 if (page && zone_idx(page_zone(page)) == ZONE_MOVABLE)
2337 online_type = MMOP_ONLINE_MOVABLE;
2338
2339 rc = device_offline(&mem->dev);
2340 /*
2341 * Default is MMOP_OFFLINE - change it only if offlining succeeded,
2342 * so try_reonline_memory_block() can do the right thing.
2343 */
2344 if (!rc)
2345 **online_types = online_type;
2346
2347 (*online_types)++;
2348 /* Ignore if already offline. */
2349 return rc < 0 ? rc : 0;
2350}
2351
2352static int try_reonline_memory_block(struct memory_block *mem, void *arg)
2353{
2354 uint8_t **online_types = arg;
2355 int rc;
2356
2357 if (**online_types != MMOP_OFFLINE) {
2358 mem->online_type = **online_types;
2359 rc = device_online(&mem->dev);
2360 if (rc < 0)
2361 pr_warn("%s: Failed to re-online memory: %d",
2362 __func__, rc);
2363 }
2364
2365 /* Continue processing all remaining memory blocks. */
2366 (*online_types)++;
2367 return 0;
2368}
2369
2370/*
2371 * Try to offline and remove memory. Might take a long time to finish in case
2372 * memory is still in use. Primarily useful for memory devices that logically
2373 * unplugged all memory (so it's no longer in use) and want to offline + remove
2374 * that memory.
2375 */
2376int offline_and_remove_memory(u64 start, u64 size)
2377{
2378 const unsigned long mb_count = size / memory_block_size_bytes();
2379 uint8_t *online_types, *tmp;
2380 int rc;
2381
2382 if (!IS_ALIGNED(start, memory_block_size_bytes()) ||
2383 !IS_ALIGNED(size, memory_block_size_bytes()) || !size)
2384 return -EINVAL;
2385
2386 /*
2387 * We'll remember the old online type of each memory block, so we can
2388 * try to revert whatever we did when offlining one memory block fails
2389 * after offlining some others succeeded.
2390 */
2391 online_types = kmalloc_array(mb_count, sizeof(*online_types),
2392 GFP_KERNEL);
2393 if (!online_types)
2394 return -ENOMEM;
2395 /*
2396 * Initialize all states to MMOP_OFFLINE, so when we abort processing in
2397 * try_offline_memory_block(), we'll skip all unprocessed blocks in
2398 * try_reonline_memory_block().
2399 */
2400 memset(online_types, MMOP_OFFLINE, mb_count);
2401
2402 lock_device_hotplug();
2403
2404 tmp = online_types;
2405 rc = walk_memory_blocks(start, size, &tmp, try_offline_memory_block);
2406
2407 /*
2408 * In case we succeeded to offline all memory, remove it.
2409 * This cannot fail as it cannot get onlined in the meantime.
2410 */
2411 if (!rc) {
2412 rc = try_remove_memory(start, size);
2413 if (rc)
2414 pr_err("%s: Failed to remove memory: %d", __func__, rc);
2415 }
2416
2417 /*
2418 * Rollback what we did. While memory onlining might theoretically fail
2419 * (nacked by a notifier), it barely ever happens.
2420 */
2421 if (rc) {
2422 tmp = online_types;
2423 walk_memory_blocks(start, size, &tmp,
2424 try_reonline_memory_block);
2425 }
2426 unlock_device_hotplug();
2427
2428 kfree(online_types);
2429 return rc;
2430}
2431EXPORT_SYMBOL_GPL(offline_and_remove_memory);
2432#endif /* CONFIG_MEMORY_HOTREMOVE */