xref: /openbmc/linux/mm/memory_hotplug.c (revision f20c7d91)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  *  linux/mm/memory_hotplug.c
4  *
5  *  Copyright (C)
6  */
7 
8 #include <linux/stddef.h>
9 #include <linux/mm.h>
10 #include <linux/sched/signal.h>
11 #include <linux/swap.h>
12 #include <linux/interrupt.h>
13 #include <linux/pagemap.h>
14 #include <linux/compiler.h>
15 #include <linux/export.h>
16 #include <linux/pagevec.h>
17 #include <linux/writeback.h>
18 #include <linux/slab.h>
19 #include <linux/sysctl.h>
20 #include <linux/cpu.h>
21 #include <linux/memory.h>
22 #include <linux/memremap.h>
23 #include <linux/memory_hotplug.h>
24 #include <linux/highmem.h>
25 #include <linux/vmalloc.h>
26 #include <linux/ioport.h>
27 #include <linux/delay.h>
28 #include <linux/migrate.h>
29 #include <linux/page-isolation.h>
30 #include <linux/pfn.h>
31 #include <linux/suspend.h>
32 #include <linux/mm_inline.h>
33 #include <linux/firmware-map.h>
34 #include <linux/stop_machine.h>
35 #include <linux/hugetlb.h>
36 #include <linux/memblock.h>
37 #include <linux/compaction.h>
38 #include <linux/rmap.h>
39 
40 #include <asm/tlbflush.h>
41 
42 #include "internal.h"
43 #include "shuffle.h"
44 
45 /*
46  * online_page_callback contains pointer to current page onlining function.
47  * Initially it is generic_online_page(). If it is required it could be
48  * changed by calling set_online_page_callback() for callback registration
49  * and restore_online_page_callback() for generic callback restore.
50  */
51 
52 static online_page_callback_t online_page_callback = generic_online_page;
53 static DEFINE_MUTEX(online_page_callback_lock);
54 
55 DEFINE_STATIC_PERCPU_RWSEM(mem_hotplug_lock);
56 
57 void get_online_mems(void)
58 {
59 	percpu_down_read(&mem_hotplug_lock);
60 }
61 
62 void put_online_mems(void)
63 {
64 	percpu_up_read(&mem_hotplug_lock);
65 }
66 
67 bool movable_node_enabled = false;
68 
69 #ifndef CONFIG_MEMORY_HOTPLUG_DEFAULT_ONLINE
70 int memhp_default_online_type = MMOP_OFFLINE;
71 #else
72 int memhp_default_online_type = MMOP_ONLINE;
73 #endif
74 
75 static int __init setup_memhp_default_state(char *str)
76 {
77 	const int online_type = memhp_online_type_from_str(str);
78 
79 	if (online_type >= 0)
80 		memhp_default_online_type = online_type;
81 
82 	return 1;
83 }
84 __setup("memhp_default_state=", setup_memhp_default_state);
85 
86 void mem_hotplug_begin(void)
87 {
88 	cpus_read_lock();
89 	percpu_down_write(&mem_hotplug_lock);
90 }
91 
92 void mem_hotplug_done(void)
93 {
94 	percpu_up_write(&mem_hotplug_lock);
95 	cpus_read_unlock();
96 }
97 
98 u64 max_mem_size = U64_MAX;
99 
100 /* add this memory to iomem resource */
101 static struct resource *register_memory_resource(u64 start, u64 size,
102 						 const char *resource_name)
103 {
104 	struct resource *res;
105 	unsigned long flags =  IORESOURCE_SYSTEM_RAM | IORESOURCE_BUSY;
106 
107 	if (strcmp(resource_name, "System RAM"))
108 		flags |= IORESOURCE_MEM_DRIVER_MANAGED;
109 
110 	/*
111 	 * Make sure value parsed from 'mem=' only restricts memory adding
112 	 * while booting, so that memory hotplug won't be impacted. Please
113 	 * refer to document of 'mem=' in kernel-parameters.txt for more
114 	 * details.
115 	 */
116 	if (start + size > max_mem_size && system_state < SYSTEM_RUNNING)
117 		return ERR_PTR(-E2BIG);
118 
119 	/*
120 	 * Request ownership of the new memory range.  This might be
121 	 * a child of an existing resource that was present but
122 	 * not marked as busy.
123 	 */
124 	res = __request_region(&iomem_resource, start, size,
125 			       resource_name, flags);
126 
127 	if (!res) {
128 		pr_debug("Unable to reserve System RAM region: %016llx->%016llx\n",
129 				start, start + size);
130 		return ERR_PTR(-EEXIST);
131 	}
132 	return res;
133 }
134 
135 static void release_memory_resource(struct resource *res)
136 {
137 	if (!res)
138 		return;
139 	release_resource(res);
140 	kfree(res);
141 }
142 
143 #ifdef CONFIG_MEMORY_HOTPLUG_SPARSE
144 void get_page_bootmem(unsigned long info,  struct page *page,
145 		      unsigned long type)
146 {
147 	page->freelist = (void *)type;
148 	SetPagePrivate(page);
149 	set_page_private(page, info);
150 	page_ref_inc(page);
151 }
152 
153 void put_page_bootmem(struct page *page)
154 {
155 	unsigned long type;
156 
157 	type = (unsigned long) page->freelist;
158 	BUG_ON(type < MEMORY_HOTPLUG_MIN_BOOTMEM_TYPE ||
159 	       type > MEMORY_HOTPLUG_MAX_BOOTMEM_TYPE);
160 
161 	if (page_ref_dec_return(page) == 1) {
162 		page->freelist = NULL;
163 		ClearPagePrivate(page);
164 		set_page_private(page, 0);
165 		INIT_LIST_HEAD(&page->lru);
166 		free_reserved_page(page);
167 	}
168 }
169 
170 #ifdef CONFIG_HAVE_BOOTMEM_INFO_NODE
171 #ifndef CONFIG_SPARSEMEM_VMEMMAP
172 static void register_page_bootmem_info_section(unsigned long start_pfn)
173 {
174 	unsigned long mapsize, section_nr, i;
175 	struct mem_section *ms;
176 	struct page *page, *memmap;
177 	struct mem_section_usage *usage;
178 
179 	section_nr = pfn_to_section_nr(start_pfn);
180 	ms = __nr_to_section(section_nr);
181 
182 	/* Get section's memmap address */
183 	memmap = sparse_decode_mem_map(ms->section_mem_map, section_nr);
184 
185 	/*
186 	 * Get page for the memmap's phys address
187 	 * XXX: need more consideration for sparse_vmemmap...
188 	 */
189 	page = virt_to_page(memmap);
190 	mapsize = sizeof(struct page) * PAGES_PER_SECTION;
191 	mapsize = PAGE_ALIGN(mapsize) >> PAGE_SHIFT;
192 
193 	/* remember memmap's page */
194 	for (i = 0; i < mapsize; i++, page++)
195 		get_page_bootmem(section_nr, page, SECTION_INFO);
196 
197 	usage = ms->usage;
198 	page = virt_to_page(usage);
199 
200 	mapsize = PAGE_ALIGN(mem_section_usage_size()) >> PAGE_SHIFT;
201 
202 	for (i = 0; i < mapsize; i++, page++)
203 		get_page_bootmem(section_nr, page, MIX_SECTION_INFO);
204 
205 }
206 #else /* CONFIG_SPARSEMEM_VMEMMAP */
207 static void register_page_bootmem_info_section(unsigned long start_pfn)
208 {
209 	unsigned long mapsize, section_nr, i;
210 	struct mem_section *ms;
211 	struct page *page, *memmap;
212 	struct mem_section_usage *usage;
213 
214 	section_nr = pfn_to_section_nr(start_pfn);
215 	ms = __nr_to_section(section_nr);
216 
217 	memmap = sparse_decode_mem_map(ms->section_mem_map, section_nr);
218 
219 	register_page_bootmem_memmap(section_nr, memmap, PAGES_PER_SECTION);
220 
221 	usage = ms->usage;
222 	page = virt_to_page(usage);
223 
224 	mapsize = PAGE_ALIGN(mem_section_usage_size()) >> PAGE_SHIFT;
225 
226 	for (i = 0; i < mapsize; i++, page++)
227 		get_page_bootmem(section_nr, page, MIX_SECTION_INFO);
228 }
229 #endif /* !CONFIG_SPARSEMEM_VMEMMAP */
230 
231 void __init register_page_bootmem_info_node(struct pglist_data *pgdat)
232 {
233 	unsigned long i, pfn, end_pfn, nr_pages;
234 	int node = pgdat->node_id;
235 	struct page *page;
236 
237 	nr_pages = PAGE_ALIGN(sizeof(struct pglist_data)) >> PAGE_SHIFT;
238 	page = virt_to_page(pgdat);
239 
240 	for (i = 0; i < nr_pages; i++, page++)
241 		get_page_bootmem(node, page, NODE_INFO);
242 
243 	pfn = pgdat->node_start_pfn;
244 	end_pfn = pgdat_end_pfn(pgdat);
245 
246 	/* register section info */
247 	for (; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
248 		/*
249 		 * Some platforms can assign the same pfn to multiple nodes - on
250 		 * node0 as well as nodeN.  To avoid registering a pfn against
251 		 * multiple nodes we check that this pfn does not already
252 		 * reside in some other nodes.
253 		 */
254 		if (pfn_valid(pfn) && (early_pfn_to_nid(pfn) == node))
255 			register_page_bootmem_info_section(pfn);
256 	}
257 }
258 #endif /* CONFIG_HAVE_BOOTMEM_INFO_NODE */
259 
260 static int check_pfn_span(unsigned long pfn, unsigned long nr_pages,
261 		const char *reason)
262 {
263 	/*
264 	 * Disallow all operations smaller than a sub-section and only
265 	 * allow operations smaller than a section for
266 	 * SPARSEMEM_VMEMMAP. Note that check_hotplug_memory_range()
267 	 * enforces a larger memory_block_size_bytes() granularity for
268 	 * memory that will be marked online, so this check should only
269 	 * fire for direct arch_{add,remove}_memory() users outside of
270 	 * add_memory_resource().
271 	 */
272 	unsigned long min_align;
273 
274 	if (IS_ENABLED(CONFIG_SPARSEMEM_VMEMMAP))
275 		min_align = PAGES_PER_SUBSECTION;
276 	else
277 		min_align = PAGES_PER_SECTION;
278 	if (!IS_ALIGNED(pfn, min_align)
279 			|| !IS_ALIGNED(nr_pages, min_align)) {
280 		WARN(1, "Misaligned __%s_pages start: %#lx end: #%lx\n",
281 				reason, pfn, pfn + nr_pages - 1);
282 		return -EINVAL;
283 	}
284 	return 0;
285 }
286 
287 static int check_hotplug_memory_addressable(unsigned long pfn,
288 					    unsigned long nr_pages)
289 {
290 	const u64 max_addr = PFN_PHYS(pfn + nr_pages) - 1;
291 
292 	if (max_addr >> MAX_PHYSMEM_BITS) {
293 		const u64 max_allowed = (1ull << (MAX_PHYSMEM_BITS + 1)) - 1;
294 		WARN(1,
295 		     "Hotplugged memory exceeds maximum addressable address, range=%#llx-%#llx, maximum=%#llx\n",
296 		     (u64)PFN_PHYS(pfn), max_addr, max_allowed);
297 		return -E2BIG;
298 	}
299 
300 	return 0;
301 }
302 
303 /*
304  * Reasonably generic function for adding memory.  It is
305  * expected that archs that support memory hotplug will
306  * call this function after deciding the zone to which to
307  * add the new pages.
308  */
309 int __ref __add_pages(int nid, unsigned long pfn, unsigned long nr_pages,
310 		struct mhp_params *params)
311 {
312 	const unsigned long end_pfn = pfn + nr_pages;
313 	unsigned long cur_nr_pages;
314 	int err;
315 	struct vmem_altmap *altmap = params->altmap;
316 
317 	if (WARN_ON_ONCE(!params->pgprot.pgprot))
318 		return -EINVAL;
319 
320 	err = check_hotplug_memory_addressable(pfn, nr_pages);
321 	if (err)
322 		return err;
323 
324 	if (altmap) {
325 		/*
326 		 * Validate altmap is within bounds of the total request
327 		 */
328 		if (altmap->base_pfn != pfn
329 				|| vmem_altmap_offset(altmap) > nr_pages) {
330 			pr_warn_once("memory add fail, invalid altmap\n");
331 			return -EINVAL;
332 		}
333 		altmap->alloc = 0;
334 	}
335 
336 	err = check_pfn_span(pfn, nr_pages, "add");
337 	if (err)
338 		return err;
339 
340 	for (; pfn < end_pfn; pfn += cur_nr_pages) {
341 		/* Select all remaining pages up to the next section boundary */
342 		cur_nr_pages = min(end_pfn - pfn,
343 				   SECTION_ALIGN_UP(pfn + 1) - pfn);
344 		err = sparse_add_section(nid, pfn, cur_nr_pages, altmap);
345 		if (err)
346 			break;
347 		cond_resched();
348 	}
349 	vmemmap_populate_print_last();
350 	return err;
351 }
352 
353 /* find the smallest valid pfn in the range [start_pfn, end_pfn) */
354 static unsigned long find_smallest_section_pfn(int nid, struct zone *zone,
355 				     unsigned long start_pfn,
356 				     unsigned long end_pfn)
357 {
358 	for (; start_pfn < end_pfn; start_pfn += PAGES_PER_SUBSECTION) {
359 		if (unlikely(!pfn_to_online_page(start_pfn)))
360 			continue;
361 
362 		if (unlikely(pfn_to_nid(start_pfn) != nid))
363 			continue;
364 
365 		if (zone != page_zone(pfn_to_page(start_pfn)))
366 			continue;
367 
368 		return start_pfn;
369 	}
370 
371 	return 0;
372 }
373 
374 /* find the biggest valid pfn in the range [start_pfn, end_pfn). */
375 static unsigned long find_biggest_section_pfn(int nid, struct zone *zone,
376 				    unsigned long start_pfn,
377 				    unsigned long end_pfn)
378 {
379 	unsigned long pfn;
380 
381 	/* pfn is the end pfn of a memory section. */
382 	pfn = end_pfn - 1;
383 	for (; pfn >= start_pfn; pfn -= PAGES_PER_SUBSECTION) {
384 		if (unlikely(!pfn_to_online_page(pfn)))
385 			continue;
386 
387 		if (unlikely(pfn_to_nid(pfn) != nid))
388 			continue;
389 
390 		if (zone != page_zone(pfn_to_page(pfn)))
391 			continue;
392 
393 		return pfn;
394 	}
395 
396 	return 0;
397 }
398 
399 static void shrink_zone_span(struct zone *zone, unsigned long start_pfn,
400 			     unsigned long end_pfn)
401 {
402 	unsigned long pfn;
403 	int nid = zone_to_nid(zone);
404 
405 	zone_span_writelock(zone);
406 	if (zone->zone_start_pfn == start_pfn) {
407 		/*
408 		 * If the section is smallest section in the zone, it need
409 		 * shrink zone->zone_start_pfn and zone->zone_spanned_pages.
410 		 * In this case, we find second smallest valid mem_section
411 		 * for shrinking zone.
412 		 */
413 		pfn = find_smallest_section_pfn(nid, zone, end_pfn,
414 						zone_end_pfn(zone));
415 		if (pfn) {
416 			zone->spanned_pages = zone_end_pfn(zone) - pfn;
417 			zone->zone_start_pfn = pfn;
418 		} else {
419 			zone->zone_start_pfn = 0;
420 			zone->spanned_pages = 0;
421 		}
422 	} else if (zone_end_pfn(zone) == end_pfn) {
423 		/*
424 		 * If the section is biggest section in the zone, it need
425 		 * shrink zone->spanned_pages.
426 		 * In this case, we find second biggest valid mem_section for
427 		 * shrinking zone.
428 		 */
429 		pfn = find_biggest_section_pfn(nid, zone, zone->zone_start_pfn,
430 					       start_pfn);
431 		if (pfn)
432 			zone->spanned_pages = pfn - zone->zone_start_pfn + 1;
433 		else {
434 			zone->zone_start_pfn = 0;
435 			zone->spanned_pages = 0;
436 		}
437 	}
438 	zone_span_writeunlock(zone);
439 }
440 
441 static void update_pgdat_span(struct pglist_data *pgdat)
442 {
443 	unsigned long node_start_pfn = 0, node_end_pfn = 0;
444 	struct zone *zone;
445 
446 	for (zone = pgdat->node_zones;
447 	     zone < pgdat->node_zones + MAX_NR_ZONES; zone++) {
448 		unsigned long zone_end_pfn = zone->zone_start_pfn +
449 					     zone->spanned_pages;
450 
451 		/* No need to lock the zones, they can't change. */
452 		if (!zone->spanned_pages)
453 			continue;
454 		if (!node_end_pfn) {
455 			node_start_pfn = zone->zone_start_pfn;
456 			node_end_pfn = zone_end_pfn;
457 			continue;
458 		}
459 
460 		if (zone_end_pfn > node_end_pfn)
461 			node_end_pfn = zone_end_pfn;
462 		if (zone->zone_start_pfn < node_start_pfn)
463 			node_start_pfn = zone->zone_start_pfn;
464 	}
465 
466 	pgdat->node_start_pfn = node_start_pfn;
467 	pgdat->node_spanned_pages = node_end_pfn - node_start_pfn;
468 }
469 
470 void __ref remove_pfn_range_from_zone(struct zone *zone,
471 				      unsigned long start_pfn,
472 				      unsigned long nr_pages)
473 {
474 	struct pglist_data *pgdat = zone->zone_pgdat;
475 	unsigned long flags;
476 
477 	/* Poison struct pages because they are now uninitialized again. */
478 	page_init_poison(pfn_to_page(start_pfn), sizeof(struct page) * nr_pages);
479 
480 #ifdef CONFIG_ZONE_DEVICE
481 	/*
482 	 * Zone shrinking code cannot properly deal with ZONE_DEVICE. So
483 	 * we will not try to shrink the zones - which is okay as
484 	 * set_zone_contiguous() cannot deal with ZONE_DEVICE either way.
485 	 */
486 	if (zone_idx(zone) == ZONE_DEVICE)
487 		return;
488 #endif
489 
490 	clear_zone_contiguous(zone);
491 
492 	pgdat_resize_lock(zone->zone_pgdat, &flags);
493 	shrink_zone_span(zone, start_pfn, start_pfn + nr_pages);
494 	update_pgdat_span(pgdat);
495 	pgdat_resize_unlock(zone->zone_pgdat, &flags);
496 
497 	set_zone_contiguous(zone);
498 }
499 
500 static void __remove_section(unsigned long pfn, unsigned long nr_pages,
501 			     unsigned long map_offset,
502 			     struct vmem_altmap *altmap)
503 {
504 	struct mem_section *ms = __pfn_to_section(pfn);
505 
506 	if (WARN_ON_ONCE(!valid_section(ms)))
507 		return;
508 
509 	sparse_remove_section(ms, pfn, nr_pages, map_offset, altmap);
510 }
511 
512 /**
513  * __remove_pages() - remove sections of pages
514  * @pfn: starting pageframe (must be aligned to start of a section)
515  * @nr_pages: number of pages to remove (must be multiple of section size)
516  * @altmap: alternative device page map or %NULL if default memmap is used
517  *
518  * Generic helper function to remove section mappings and sysfs entries
519  * for the section of the memory we are removing. Caller needs to make
520  * sure that pages are marked reserved and zones are adjust properly by
521  * calling offline_pages().
522  */
523 void __remove_pages(unsigned long pfn, unsigned long nr_pages,
524 		    struct vmem_altmap *altmap)
525 {
526 	const unsigned long end_pfn = pfn + nr_pages;
527 	unsigned long cur_nr_pages;
528 	unsigned long map_offset = 0;
529 
530 	map_offset = vmem_altmap_offset(altmap);
531 
532 	if (check_pfn_span(pfn, nr_pages, "remove"))
533 		return;
534 
535 	for (; pfn < end_pfn; pfn += cur_nr_pages) {
536 		cond_resched();
537 		/* Select all remaining pages up to the next section boundary */
538 		cur_nr_pages = min(end_pfn - pfn,
539 				   SECTION_ALIGN_UP(pfn + 1) - pfn);
540 		__remove_section(pfn, cur_nr_pages, map_offset, altmap);
541 		map_offset = 0;
542 	}
543 }
544 
545 int set_online_page_callback(online_page_callback_t callback)
546 {
547 	int rc = -EINVAL;
548 
549 	get_online_mems();
550 	mutex_lock(&online_page_callback_lock);
551 
552 	if (online_page_callback == generic_online_page) {
553 		online_page_callback = callback;
554 		rc = 0;
555 	}
556 
557 	mutex_unlock(&online_page_callback_lock);
558 	put_online_mems();
559 
560 	return rc;
561 }
562 EXPORT_SYMBOL_GPL(set_online_page_callback);
563 
564 int restore_online_page_callback(online_page_callback_t callback)
565 {
566 	int rc = -EINVAL;
567 
568 	get_online_mems();
569 	mutex_lock(&online_page_callback_lock);
570 
571 	if (online_page_callback == callback) {
572 		online_page_callback = generic_online_page;
573 		rc = 0;
574 	}
575 
576 	mutex_unlock(&online_page_callback_lock);
577 	put_online_mems();
578 
579 	return rc;
580 }
581 EXPORT_SYMBOL_GPL(restore_online_page_callback);
582 
583 void generic_online_page(struct page *page, unsigned int order)
584 {
585 	/*
586 	 * Freeing the page with debug_pagealloc enabled will try to unmap it,
587 	 * so we should map it first. This is better than introducing a special
588 	 * case in page freeing fast path.
589 	 */
590 	if (debug_pagealloc_enabled_static())
591 		kernel_map_pages(page, 1 << order, 1);
592 	__free_pages_core(page, order);
593 	totalram_pages_add(1UL << order);
594 #ifdef CONFIG_HIGHMEM
595 	if (PageHighMem(page))
596 		totalhigh_pages_add(1UL << order);
597 #endif
598 }
599 EXPORT_SYMBOL_GPL(generic_online_page);
600 
601 static int online_pages_range(unsigned long start_pfn, unsigned long nr_pages,
602 			void *arg)
603 {
604 	const unsigned long end_pfn = start_pfn + nr_pages;
605 	unsigned long pfn;
606 	int order;
607 
608 	/*
609 	 * Online the pages. The callback might decide to keep some pages
610 	 * PG_reserved (to add them to the buddy later), but we still account
611 	 * them as being online/belonging to this zone ("present").
612 	 */
613 	for (pfn = start_pfn; pfn < end_pfn; pfn += 1ul << order) {
614 		order = min(MAX_ORDER - 1, get_order(PFN_PHYS(end_pfn - pfn)));
615 		/* __free_pages_core() wants pfns to be aligned to the order */
616 		if (WARN_ON_ONCE(!IS_ALIGNED(pfn, 1ul << order)))
617 			order = 0;
618 		(*online_page_callback)(pfn_to_page(pfn), order);
619 	}
620 
621 	/* mark all involved sections as online */
622 	online_mem_sections(start_pfn, end_pfn);
623 
624 	*(unsigned long *)arg += nr_pages;
625 	return 0;
626 }
627 
628 /* check which state of node_states will be changed when online memory */
629 static void node_states_check_changes_online(unsigned long nr_pages,
630 	struct zone *zone, struct memory_notify *arg)
631 {
632 	int nid = zone_to_nid(zone);
633 
634 	arg->status_change_nid = NUMA_NO_NODE;
635 	arg->status_change_nid_normal = NUMA_NO_NODE;
636 	arg->status_change_nid_high = NUMA_NO_NODE;
637 
638 	if (!node_state(nid, N_MEMORY))
639 		arg->status_change_nid = nid;
640 	if (zone_idx(zone) <= ZONE_NORMAL && !node_state(nid, N_NORMAL_MEMORY))
641 		arg->status_change_nid_normal = nid;
642 #ifdef CONFIG_HIGHMEM
643 	if (zone_idx(zone) <= ZONE_HIGHMEM && !node_state(nid, N_HIGH_MEMORY))
644 		arg->status_change_nid_high = nid;
645 #endif
646 }
647 
648 static void node_states_set_node(int node, struct memory_notify *arg)
649 {
650 	if (arg->status_change_nid_normal >= 0)
651 		node_set_state(node, N_NORMAL_MEMORY);
652 
653 	if (arg->status_change_nid_high >= 0)
654 		node_set_state(node, N_HIGH_MEMORY);
655 
656 	if (arg->status_change_nid >= 0)
657 		node_set_state(node, N_MEMORY);
658 }
659 
660 static void __meminit resize_zone_range(struct zone *zone, unsigned long start_pfn,
661 		unsigned long nr_pages)
662 {
663 	unsigned long old_end_pfn = zone_end_pfn(zone);
664 
665 	if (zone_is_empty(zone) || start_pfn < zone->zone_start_pfn)
666 		zone->zone_start_pfn = start_pfn;
667 
668 	zone->spanned_pages = max(start_pfn + nr_pages, old_end_pfn) - zone->zone_start_pfn;
669 }
670 
671 static void __meminit resize_pgdat_range(struct pglist_data *pgdat, unsigned long start_pfn,
672                                      unsigned long nr_pages)
673 {
674 	unsigned long old_end_pfn = pgdat_end_pfn(pgdat);
675 
676 	if (!pgdat->node_spanned_pages || start_pfn < pgdat->node_start_pfn)
677 		pgdat->node_start_pfn = start_pfn;
678 
679 	pgdat->node_spanned_pages = max(start_pfn + nr_pages, old_end_pfn) - pgdat->node_start_pfn;
680 
681 }
682 /*
683  * Associate the pfn range with the given zone, initializing the memmaps
684  * and resizing the pgdat/zone data to span the added pages. After this
685  * call, all affected pages are PG_reserved.
686  */
687 void __ref move_pfn_range_to_zone(struct zone *zone, unsigned long start_pfn,
688 		unsigned long nr_pages, struct vmem_altmap *altmap)
689 {
690 	struct pglist_data *pgdat = zone->zone_pgdat;
691 	int nid = pgdat->node_id;
692 	unsigned long flags;
693 
694 	clear_zone_contiguous(zone);
695 
696 	/* TODO Huh pgdat is irqsave while zone is not. It used to be like that before */
697 	pgdat_resize_lock(pgdat, &flags);
698 	zone_span_writelock(zone);
699 	if (zone_is_empty(zone))
700 		init_currently_empty_zone(zone, start_pfn, nr_pages);
701 	resize_zone_range(zone, start_pfn, nr_pages);
702 	zone_span_writeunlock(zone);
703 	resize_pgdat_range(pgdat, start_pfn, nr_pages);
704 	pgdat_resize_unlock(pgdat, &flags);
705 
706 	/*
707 	 * TODO now we have a visible range of pages which are not associated
708 	 * with their zone properly. Not nice but set_pfnblock_flags_mask
709 	 * expects the zone spans the pfn range. All the pages in the range
710 	 * are reserved so nobody should be touching them so we should be safe
711 	 */
712 	memmap_init_zone(nr_pages, nid, zone_idx(zone), start_pfn,
713 			MEMMAP_HOTPLUG, altmap);
714 
715 	set_zone_contiguous(zone);
716 }
717 
718 /*
719  * Returns a default kernel memory zone for the given pfn range.
720  * If no kernel zone covers this pfn range it will automatically go
721  * to the ZONE_NORMAL.
722  */
723 static struct zone *default_kernel_zone_for_pfn(int nid, unsigned long start_pfn,
724 		unsigned long nr_pages)
725 {
726 	struct pglist_data *pgdat = NODE_DATA(nid);
727 	int zid;
728 
729 	for (zid = 0; zid <= ZONE_NORMAL; zid++) {
730 		struct zone *zone = &pgdat->node_zones[zid];
731 
732 		if (zone_intersects(zone, start_pfn, nr_pages))
733 			return zone;
734 	}
735 
736 	return &pgdat->node_zones[ZONE_NORMAL];
737 }
738 
739 static inline struct zone *default_zone_for_pfn(int nid, unsigned long start_pfn,
740 		unsigned long nr_pages)
741 {
742 	struct zone *kernel_zone = default_kernel_zone_for_pfn(nid, start_pfn,
743 			nr_pages);
744 	struct zone *movable_zone = &NODE_DATA(nid)->node_zones[ZONE_MOVABLE];
745 	bool in_kernel = zone_intersects(kernel_zone, start_pfn, nr_pages);
746 	bool in_movable = zone_intersects(movable_zone, start_pfn, nr_pages);
747 
748 	/*
749 	 * We inherit the existing zone in a simple case where zones do not
750 	 * overlap in the given range
751 	 */
752 	if (in_kernel ^ in_movable)
753 		return (in_kernel) ? kernel_zone : movable_zone;
754 
755 	/*
756 	 * If the range doesn't belong to any zone or two zones overlap in the
757 	 * given range then we use movable zone only if movable_node is
758 	 * enabled because we always online to a kernel zone by default.
759 	 */
760 	return movable_node_enabled ? movable_zone : kernel_zone;
761 }
762 
763 struct zone * zone_for_pfn_range(int online_type, int nid, unsigned start_pfn,
764 		unsigned long nr_pages)
765 {
766 	if (online_type == MMOP_ONLINE_KERNEL)
767 		return default_kernel_zone_for_pfn(nid, start_pfn, nr_pages);
768 
769 	if (online_type == MMOP_ONLINE_MOVABLE)
770 		return &NODE_DATA(nid)->node_zones[ZONE_MOVABLE];
771 
772 	return default_zone_for_pfn(nid, start_pfn, nr_pages);
773 }
774 
775 int __ref online_pages(unsigned long pfn, unsigned long nr_pages,
776 		       int online_type, int nid)
777 {
778 	unsigned long flags;
779 	unsigned long onlined_pages = 0;
780 	struct zone *zone;
781 	int need_zonelists_rebuild = 0;
782 	int ret;
783 	struct memory_notify arg;
784 
785 	mem_hotplug_begin();
786 
787 	/* associate pfn range with the zone */
788 	zone = zone_for_pfn_range(online_type, nid, pfn, nr_pages);
789 	move_pfn_range_to_zone(zone, pfn, nr_pages, NULL);
790 
791 	arg.start_pfn = pfn;
792 	arg.nr_pages = nr_pages;
793 	node_states_check_changes_online(nr_pages, zone, &arg);
794 
795 	ret = memory_notify(MEM_GOING_ONLINE, &arg);
796 	ret = notifier_to_errno(ret);
797 	if (ret)
798 		goto failed_addition;
799 
800 	/*
801 	 * If this zone is not populated, then it is not in zonelist.
802 	 * This means the page allocator ignores this zone.
803 	 * So, zonelist must be updated after online.
804 	 */
805 	if (!populated_zone(zone)) {
806 		need_zonelists_rebuild = 1;
807 		setup_zone_pageset(zone);
808 	}
809 
810 	ret = walk_system_ram_range(pfn, nr_pages, &onlined_pages,
811 		online_pages_range);
812 	if (ret) {
813 		/* not a single memory resource was applicable */
814 		if (need_zonelists_rebuild)
815 			zone_pcp_reset(zone);
816 		goto failed_addition;
817 	}
818 
819 	zone->present_pages += onlined_pages;
820 
821 	pgdat_resize_lock(zone->zone_pgdat, &flags);
822 	zone->zone_pgdat->node_present_pages += onlined_pages;
823 	pgdat_resize_unlock(zone->zone_pgdat, &flags);
824 
825 	shuffle_zone(zone);
826 
827 	node_states_set_node(nid, &arg);
828 	if (need_zonelists_rebuild)
829 		build_all_zonelists(NULL);
830 	else
831 		zone_pcp_update(zone);
832 
833 	init_per_zone_wmark_min();
834 
835 	kswapd_run(nid);
836 	kcompactd_run(nid);
837 
838 	vm_total_pages = nr_free_pagecache_pages();
839 
840 	writeback_set_ratelimit();
841 
842 	memory_notify(MEM_ONLINE, &arg);
843 	mem_hotplug_done();
844 	return 0;
845 
846 failed_addition:
847 	pr_debug("online_pages [mem %#010llx-%#010llx] failed\n",
848 		 (unsigned long long) pfn << PAGE_SHIFT,
849 		 (((unsigned long long) pfn + nr_pages) << PAGE_SHIFT) - 1);
850 	memory_notify(MEM_CANCEL_ONLINE, &arg);
851 	remove_pfn_range_from_zone(zone, pfn, nr_pages);
852 	mem_hotplug_done();
853 	return ret;
854 }
855 #endif /* CONFIG_MEMORY_HOTPLUG_SPARSE */
856 
857 static void reset_node_present_pages(pg_data_t *pgdat)
858 {
859 	struct zone *z;
860 
861 	for (z = pgdat->node_zones; z < pgdat->node_zones + MAX_NR_ZONES; z++)
862 		z->present_pages = 0;
863 
864 	pgdat->node_present_pages = 0;
865 }
866 
867 /* we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG */
868 static pg_data_t __ref *hotadd_new_pgdat(int nid)
869 {
870 	struct pglist_data *pgdat;
871 
872 	pgdat = NODE_DATA(nid);
873 	if (!pgdat) {
874 		pgdat = arch_alloc_nodedata(nid);
875 		if (!pgdat)
876 			return NULL;
877 
878 		pgdat->per_cpu_nodestats =
879 			alloc_percpu(struct per_cpu_nodestat);
880 		arch_refresh_nodedata(nid, pgdat);
881 	} else {
882 		int cpu;
883 		/*
884 		 * Reset the nr_zones, order and highest_zoneidx before reuse.
885 		 * Note that kswapd will init kswapd_highest_zoneidx properly
886 		 * when it starts in the near future.
887 		 */
888 		pgdat->nr_zones = 0;
889 		pgdat->kswapd_order = 0;
890 		pgdat->kswapd_highest_zoneidx = 0;
891 		for_each_online_cpu(cpu) {
892 			struct per_cpu_nodestat *p;
893 
894 			p = per_cpu_ptr(pgdat->per_cpu_nodestats, cpu);
895 			memset(p, 0, sizeof(*p));
896 		}
897 	}
898 
899 	/* we can use NODE_DATA(nid) from here */
900 	pgdat->node_id = nid;
901 	pgdat->node_start_pfn = 0;
902 
903 	/* init node's zones as empty zones, we don't have any present pages.*/
904 	free_area_init_core_hotplug(nid);
905 
906 	/*
907 	 * The node we allocated has no zone fallback lists. For avoiding
908 	 * to access not-initialized zonelist, build here.
909 	 */
910 	build_all_zonelists(pgdat);
911 
912 	/*
913 	 * When memory is hot-added, all the memory is in offline state. So
914 	 * clear all zones' present_pages because they will be updated in
915 	 * online_pages() and offline_pages().
916 	 */
917 	reset_node_managed_pages(pgdat);
918 	reset_node_present_pages(pgdat);
919 
920 	return pgdat;
921 }
922 
923 static void rollback_node_hotadd(int nid)
924 {
925 	pg_data_t *pgdat = NODE_DATA(nid);
926 
927 	arch_refresh_nodedata(nid, NULL);
928 	free_percpu(pgdat->per_cpu_nodestats);
929 	arch_free_nodedata(pgdat);
930 }
931 
932 
933 /**
934  * try_online_node - online a node if offlined
935  * @nid: the node ID
936  * @set_node_online: Whether we want to online the node
937  * called by cpu_up() to online a node without onlined memory.
938  *
939  * Returns:
940  * 1 -> a new node has been allocated
941  * 0 -> the node is already online
942  * -ENOMEM -> the node could not be allocated
943  */
944 static int __try_online_node(int nid, bool set_node_online)
945 {
946 	pg_data_t *pgdat;
947 	int ret = 1;
948 
949 	if (node_online(nid))
950 		return 0;
951 
952 	pgdat = hotadd_new_pgdat(nid);
953 	if (!pgdat) {
954 		pr_err("Cannot online node %d due to NULL pgdat\n", nid);
955 		ret = -ENOMEM;
956 		goto out;
957 	}
958 
959 	if (set_node_online) {
960 		node_set_online(nid);
961 		ret = register_one_node(nid);
962 		BUG_ON(ret);
963 	}
964 out:
965 	return ret;
966 }
967 
968 /*
969  * Users of this function always want to online/register the node
970  */
971 int try_online_node(int nid)
972 {
973 	int ret;
974 
975 	mem_hotplug_begin();
976 	ret =  __try_online_node(nid, true);
977 	mem_hotplug_done();
978 	return ret;
979 }
980 
981 static int check_hotplug_memory_range(u64 start, u64 size)
982 {
983 	/* memory range must be block size aligned */
984 	if (!size || !IS_ALIGNED(start, memory_block_size_bytes()) ||
985 	    !IS_ALIGNED(size, memory_block_size_bytes())) {
986 		pr_err("Block size [%#lx] unaligned hotplug range: start %#llx, size %#llx",
987 		       memory_block_size_bytes(), start, size);
988 		return -EINVAL;
989 	}
990 
991 	return 0;
992 }
993 
994 static int online_memory_block(struct memory_block *mem, void *arg)
995 {
996 	mem->online_type = memhp_default_online_type;
997 	return device_online(&mem->dev);
998 }
999 
1000 /*
1001  * NOTE: The caller must call lock_device_hotplug() to serialize hotplug
1002  * and online/offline operations (triggered e.g. by sysfs).
1003  *
1004  * we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG
1005  */
1006 int __ref add_memory_resource(int nid, struct resource *res)
1007 {
1008 	struct mhp_params params = { .pgprot = PAGE_KERNEL };
1009 	u64 start, size;
1010 	bool new_node = false;
1011 	int ret;
1012 
1013 	start = res->start;
1014 	size = resource_size(res);
1015 
1016 	ret = check_hotplug_memory_range(start, size);
1017 	if (ret)
1018 		return ret;
1019 
1020 	if (!node_possible(nid)) {
1021 		WARN(1, "node %d was absent from the node_possible_map\n", nid);
1022 		return -EINVAL;
1023 	}
1024 
1025 	mem_hotplug_begin();
1026 
1027 	if (IS_ENABLED(CONFIG_ARCH_KEEP_MEMBLOCK))
1028 		memblock_add_node(start, size, nid);
1029 
1030 	ret = __try_online_node(nid, false);
1031 	if (ret < 0)
1032 		goto error;
1033 	new_node = ret;
1034 
1035 	/* call arch's memory hotadd */
1036 	ret = arch_add_memory(nid, start, size, &params);
1037 	if (ret < 0)
1038 		goto error;
1039 
1040 	/* create memory block devices after memory was added */
1041 	ret = create_memory_block_devices(start, size);
1042 	if (ret) {
1043 		arch_remove_memory(nid, start, size, NULL);
1044 		goto error;
1045 	}
1046 
1047 	if (new_node) {
1048 		/* If sysfs file of new node can't be created, cpu on the node
1049 		 * can't be hot-added. There is no rollback way now.
1050 		 * So, check by BUG_ON() to catch it reluctantly..
1051 		 * We online node here. We can't roll back from here.
1052 		 */
1053 		node_set_online(nid);
1054 		ret = __register_one_node(nid);
1055 		BUG_ON(ret);
1056 	}
1057 
1058 	/* link memory sections under this node.*/
1059 	ret = link_mem_sections(nid, PFN_DOWN(start), PFN_UP(start + size - 1));
1060 	BUG_ON(ret);
1061 
1062 	/* create new memmap entry */
1063 	if (!strcmp(res->name, "System RAM"))
1064 		firmware_map_add_hotplug(start, start + size, "System RAM");
1065 
1066 	/* device_online() will take the lock when calling online_pages() */
1067 	mem_hotplug_done();
1068 
1069 	/* online pages if requested */
1070 	if (memhp_default_online_type != MMOP_OFFLINE)
1071 		walk_memory_blocks(start, size, NULL, online_memory_block);
1072 
1073 	return ret;
1074 error:
1075 	/* rollback pgdat allocation and others */
1076 	if (new_node)
1077 		rollback_node_hotadd(nid);
1078 	if (IS_ENABLED(CONFIG_ARCH_KEEP_MEMBLOCK))
1079 		memblock_remove(start, size);
1080 	mem_hotplug_done();
1081 	return ret;
1082 }
1083 
1084 /* requires device_hotplug_lock, see add_memory_resource() */
1085 int __ref __add_memory(int nid, u64 start, u64 size)
1086 {
1087 	struct resource *res;
1088 	int ret;
1089 
1090 	res = register_memory_resource(start, size, "System RAM");
1091 	if (IS_ERR(res))
1092 		return PTR_ERR(res);
1093 
1094 	ret = add_memory_resource(nid, res);
1095 	if (ret < 0)
1096 		release_memory_resource(res);
1097 	return ret;
1098 }
1099 
1100 int add_memory(int nid, u64 start, u64 size)
1101 {
1102 	int rc;
1103 
1104 	lock_device_hotplug();
1105 	rc = __add_memory(nid, start, size);
1106 	unlock_device_hotplug();
1107 
1108 	return rc;
1109 }
1110 EXPORT_SYMBOL_GPL(add_memory);
1111 
1112 /*
1113  * Add special, driver-managed memory to the system as system RAM. Such
1114  * memory is not exposed via the raw firmware-provided memmap as system
1115  * RAM, instead, it is detected and added by a driver - during cold boot,
1116  * after a reboot, and after kexec.
1117  *
1118  * Reasons why this memory should not be used for the initial memmap of a
1119  * kexec kernel or for placing kexec images:
1120  * - The booting kernel is in charge of determining how this memory will be
1121  *   used (e.g., use persistent memory as system RAM)
1122  * - Coordination with a hypervisor is required before this memory
1123  *   can be used (e.g., inaccessible parts).
1124  *
1125  * For this memory, no entries in /sys/firmware/memmap ("raw firmware-provided
1126  * memory map") are created. Also, the created memory resource is flagged
1127  * with IORESOURCE_MEM_DRIVER_MANAGED, so in-kernel users can special-case
1128  * this memory as well (esp., not place kexec images onto it).
1129  *
1130  * The resource_name (visible via /proc/iomem) has to have the format
1131  * "System RAM ($DRIVER)".
1132  */
1133 int add_memory_driver_managed(int nid, u64 start, u64 size,
1134 			      const char *resource_name)
1135 {
1136 	struct resource *res;
1137 	int rc;
1138 
1139 	if (!resource_name ||
1140 	    strstr(resource_name, "System RAM (") != resource_name ||
1141 	    resource_name[strlen(resource_name) - 1] != ')')
1142 		return -EINVAL;
1143 
1144 	lock_device_hotplug();
1145 
1146 	res = register_memory_resource(start, size, resource_name);
1147 	if (IS_ERR(res)) {
1148 		rc = PTR_ERR(res);
1149 		goto out_unlock;
1150 	}
1151 
1152 	rc = add_memory_resource(nid, res);
1153 	if (rc < 0)
1154 		release_memory_resource(res);
1155 
1156 out_unlock:
1157 	unlock_device_hotplug();
1158 	return rc;
1159 }
1160 EXPORT_SYMBOL_GPL(add_memory_driver_managed);
1161 
1162 #ifdef CONFIG_MEMORY_HOTREMOVE
1163 /*
1164  * Confirm all pages in a range [start, end) belong to the same zone (skipping
1165  * memory holes). When true, return the zone.
1166  */
1167 struct zone *test_pages_in_a_zone(unsigned long start_pfn,
1168 				  unsigned long end_pfn)
1169 {
1170 	unsigned long pfn, sec_end_pfn;
1171 	struct zone *zone = NULL;
1172 	struct page *page;
1173 	int i;
1174 	for (pfn = start_pfn, sec_end_pfn = SECTION_ALIGN_UP(start_pfn + 1);
1175 	     pfn < end_pfn;
1176 	     pfn = sec_end_pfn, sec_end_pfn += PAGES_PER_SECTION) {
1177 		/* Make sure the memory section is present first */
1178 		if (!present_section_nr(pfn_to_section_nr(pfn)))
1179 			continue;
1180 		for (; pfn < sec_end_pfn && pfn < end_pfn;
1181 		     pfn += MAX_ORDER_NR_PAGES) {
1182 			i = 0;
1183 			/* This is just a CONFIG_HOLES_IN_ZONE check.*/
1184 			while ((i < MAX_ORDER_NR_PAGES) &&
1185 				!pfn_valid_within(pfn + i))
1186 				i++;
1187 			if (i == MAX_ORDER_NR_PAGES || pfn + i >= end_pfn)
1188 				continue;
1189 			/* Check if we got outside of the zone */
1190 			if (zone && !zone_spans_pfn(zone, pfn + i))
1191 				return NULL;
1192 			page = pfn_to_page(pfn + i);
1193 			if (zone && page_zone(page) != zone)
1194 				return NULL;
1195 			zone = page_zone(page);
1196 		}
1197 	}
1198 
1199 	return zone;
1200 }
1201 
1202 /*
1203  * Scan pfn range [start,end) to find movable/migratable pages (LRU pages,
1204  * non-lru movable pages and hugepages). Will skip over most unmovable
1205  * pages (esp., pages that can be skipped when offlining), but bail out on
1206  * definitely unmovable pages.
1207  *
1208  * Returns:
1209  *	0 in case a movable page is found and movable_pfn was updated.
1210  *	-ENOENT in case no movable page was found.
1211  *	-EBUSY in case a definitely unmovable page was found.
1212  */
1213 static int scan_movable_pages(unsigned long start, unsigned long end,
1214 			      unsigned long *movable_pfn)
1215 {
1216 	unsigned long pfn;
1217 
1218 	for (pfn = start; pfn < end; pfn++) {
1219 		struct page *page, *head;
1220 		unsigned long skip;
1221 
1222 		if (!pfn_valid(pfn))
1223 			continue;
1224 		page = pfn_to_page(pfn);
1225 		if (PageLRU(page))
1226 			goto found;
1227 		if (__PageMovable(page))
1228 			goto found;
1229 
1230 		/*
1231 		 * PageOffline() pages that are not marked __PageMovable() and
1232 		 * have a reference count > 0 (after MEM_GOING_OFFLINE) are
1233 		 * definitely unmovable. If their reference count would be 0,
1234 		 * they could at least be skipped when offlining memory.
1235 		 */
1236 		if (PageOffline(page) && page_count(page))
1237 			return -EBUSY;
1238 
1239 		if (!PageHuge(page))
1240 			continue;
1241 		head = compound_head(page);
1242 		if (page_huge_active(head))
1243 			goto found;
1244 		skip = compound_nr(head) - (page - head);
1245 		pfn += skip - 1;
1246 	}
1247 	return -ENOENT;
1248 found:
1249 	*movable_pfn = pfn;
1250 	return 0;
1251 }
1252 
1253 static struct page *new_node_page(struct page *page, unsigned long private)
1254 {
1255 	int nid = page_to_nid(page);
1256 	nodemask_t nmask = node_states[N_MEMORY];
1257 
1258 	/*
1259 	 * try to allocate from a different node but reuse this node if there
1260 	 * are no other online nodes to be used (e.g. we are offlining a part
1261 	 * of the only existing node)
1262 	 */
1263 	node_clear(nid, nmask);
1264 	if (nodes_empty(nmask))
1265 		node_set(nid, nmask);
1266 
1267 	return new_page_nodemask(page, nid, &nmask);
1268 }
1269 
1270 static int
1271 do_migrate_range(unsigned long start_pfn, unsigned long end_pfn)
1272 {
1273 	unsigned long pfn;
1274 	struct page *page;
1275 	int ret = 0;
1276 	LIST_HEAD(source);
1277 
1278 	for (pfn = start_pfn; pfn < end_pfn; pfn++) {
1279 		if (!pfn_valid(pfn))
1280 			continue;
1281 		page = pfn_to_page(pfn);
1282 
1283 		if (PageHuge(page)) {
1284 			struct page *head = compound_head(page);
1285 			pfn = page_to_pfn(head) + compound_nr(head) - 1;
1286 			isolate_huge_page(head, &source);
1287 			continue;
1288 		} else if (PageTransHuge(page))
1289 			pfn = page_to_pfn(compound_head(page))
1290 				+ hpage_nr_pages(page) - 1;
1291 
1292 		/*
1293 		 * HWPoison pages have elevated reference counts so the migration would
1294 		 * fail on them. It also doesn't make any sense to migrate them in the
1295 		 * first place. Still try to unmap such a page in case it is still mapped
1296 		 * (e.g. current hwpoison implementation doesn't unmap KSM pages but keep
1297 		 * the unmap as the catch all safety net).
1298 		 */
1299 		if (PageHWPoison(page)) {
1300 			if (WARN_ON(PageLRU(page)))
1301 				isolate_lru_page(page);
1302 			if (page_mapped(page))
1303 				try_to_unmap(page, TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS);
1304 			continue;
1305 		}
1306 
1307 		if (!get_page_unless_zero(page))
1308 			continue;
1309 		/*
1310 		 * We can skip free pages. And we can deal with pages on
1311 		 * LRU and non-lru movable pages.
1312 		 */
1313 		if (PageLRU(page))
1314 			ret = isolate_lru_page(page);
1315 		else
1316 			ret = isolate_movable_page(page, ISOLATE_UNEVICTABLE);
1317 		if (!ret) { /* Success */
1318 			list_add_tail(&page->lru, &source);
1319 			if (!__PageMovable(page))
1320 				inc_node_page_state(page, NR_ISOLATED_ANON +
1321 						    page_is_file_lru(page));
1322 
1323 		} else {
1324 			pr_warn("failed to isolate pfn %lx\n", pfn);
1325 			dump_page(page, "isolation failed");
1326 		}
1327 		put_page(page);
1328 	}
1329 	if (!list_empty(&source)) {
1330 		/* Allocate a new page from the nearest neighbor node */
1331 		ret = migrate_pages(&source, new_node_page, NULL, 0,
1332 					MIGRATE_SYNC, MR_MEMORY_HOTPLUG);
1333 		if (ret) {
1334 			list_for_each_entry(page, &source, lru) {
1335 				pr_warn("migrating pfn %lx failed ret:%d ",
1336 				       page_to_pfn(page), ret);
1337 				dump_page(page, "migration failure");
1338 			}
1339 			putback_movable_pages(&source);
1340 		}
1341 	}
1342 
1343 	return ret;
1344 }
1345 
1346 /* Mark all sections offline and remove all free pages from the buddy. */
1347 static int
1348 offline_isolated_pages_cb(unsigned long start, unsigned long nr_pages,
1349 			void *data)
1350 {
1351 	unsigned long *offlined_pages = (unsigned long *)data;
1352 
1353 	*offlined_pages += __offline_isolated_pages(start, start + nr_pages);
1354 	return 0;
1355 }
1356 
1357 /*
1358  * Check all pages in range, recorded as memory resource, are isolated.
1359  */
1360 static int
1361 check_pages_isolated_cb(unsigned long start_pfn, unsigned long nr_pages,
1362 			void *data)
1363 {
1364 	return test_pages_isolated(start_pfn, start_pfn + nr_pages,
1365 				   MEMORY_OFFLINE);
1366 }
1367 
1368 static int __init cmdline_parse_movable_node(char *p)
1369 {
1370 	movable_node_enabled = true;
1371 	return 0;
1372 }
1373 early_param("movable_node", cmdline_parse_movable_node);
1374 
1375 /* check which state of node_states will be changed when offline memory */
1376 static void node_states_check_changes_offline(unsigned long nr_pages,
1377 		struct zone *zone, struct memory_notify *arg)
1378 {
1379 	struct pglist_data *pgdat = zone->zone_pgdat;
1380 	unsigned long present_pages = 0;
1381 	enum zone_type zt;
1382 
1383 	arg->status_change_nid = NUMA_NO_NODE;
1384 	arg->status_change_nid_normal = NUMA_NO_NODE;
1385 	arg->status_change_nid_high = NUMA_NO_NODE;
1386 
1387 	/*
1388 	 * Check whether node_states[N_NORMAL_MEMORY] will be changed.
1389 	 * If the memory to be offline is within the range
1390 	 * [0..ZONE_NORMAL], and it is the last present memory there,
1391 	 * the zones in that range will become empty after the offlining,
1392 	 * thus we can determine that we need to clear the node from
1393 	 * node_states[N_NORMAL_MEMORY].
1394 	 */
1395 	for (zt = 0; zt <= ZONE_NORMAL; zt++)
1396 		present_pages += pgdat->node_zones[zt].present_pages;
1397 	if (zone_idx(zone) <= ZONE_NORMAL && nr_pages >= present_pages)
1398 		arg->status_change_nid_normal = zone_to_nid(zone);
1399 
1400 #ifdef CONFIG_HIGHMEM
1401 	/*
1402 	 * node_states[N_HIGH_MEMORY] contains nodes which
1403 	 * have normal memory or high memory.
1404 	 * Here we add the present_pages belonging to ZONE_HIGHMEM.
1405 	 * If the zone is within the range of [0..ZONE_HIGHMEM), and
1406 	 * we determine that the zones in that range become empty,
1407 	 * we need to clear the node for N_HIGH_MEMORY.
1408 	 */
1409 	present_pages += pgdat->node_zones[ZONE_HIGHMEM].present_pages;
1410 	if (zone_idx(zone) <= ZONE_HIGHMEM && nr_pages >= present_pages)
1411 		arg->status_change_nid_high = zone_to_nid(zone);
1412 #endif
1413 
1414 	/*
1415 	 * We have accounted the pages from [0..ZONE_NORMAL), and
1416 	 * in case of CONFIG_HIGHMEM the pages from ZONE_HIGHMEM
1417 	 * as well.
1418 	 * Here we count the possible pages from ZONE_MOVABLE.
1419 	 * If after having accounted all the pages, we see that the nr_pages
1420 	 * to be offlined is over or equal to the accounted pages,
1421 	 * we know that the node will become empty, and so, we can clear
1422 	 * it for N_MEMORY as well.
1423 	 */
1424 	present_pages += pgdat->node_zones[ZONE_MOVABLE].present_pages;
1425 
1426 	if (nr_pages >= present_pages)
1427 		arg->status_change_nid = zone_to_nid(zone);
1428 }
1429 
1430 static void node_states_clear_node(int node, struct memory_notify *arg)
1431 {
1432 	if (arg->status_change_nid_normal >= 0)
1433 		node_clear_state(node, N_NORMAL_MEMORY);
1434 
1435 	if (arg->status_change_nid_high >= 0)
1436 		node_clear_state(node, N_HIGH_MEMORY);
1437 
1438 	if (arg->status_change_nid >= 0)
1439 		node_clear_state(node, N_MEMORY);
1440 }
1441 
1442 static int count_system_ram_pages_cb(unsigned long start_pfn,
1443 				     unsigned long nr_pages, void *data)
1444 {
1445 	unsigned long *nr_system_ram_pages = data;
1446 
1447 	*nr_system_ram_pages += nr_pages;
1448 	return 0;
1449 }
1450 
1451 static int __ref __offline_pages(unsigned long start_pfn,
1452 		  unsigned long end_pfn)
1453 {
1454 	unsigned long pfn, nr_pages = 0;
1455 	unsigned long offlined_pages = 0;
1456 	int ret, node, nr_isolate_pageblock;
1457 	unsigned long flags;
1458 	struct zone *zone;
1459 	struct memory_notify arg;
1460 	char *reason;
1461 
1462 	mem_hotplug_begin();
1463 
1464 	/*
1465 	 * Don't allow to offline memory blocks that contain holes.
1466 	 * Consequently, memory blocks with holes can never get onlined
1467 	 * via the hotplug path - online_pages() - as hotplugged memory has
1468 	 * no holes. This way, we e.g., don't have to worry about marking
1469 	 * memory holes PG_reserved, don't need pfn_valid() checks, and can
1470 	 * avoid using walk_system_ram_range() later.
1471 	 */
1472 	walk_system_ram_range(start_pfn, end_pfn - start_pfn, &nr_pages,
1473 			      count_system_ram_pages_cb);
1474 	if (nr_pages != end_pfn - start_pfn) {
1475 		ret = -EINVAL;
1476 		reason = "memory holes";
1477 		goto failed_removal;
1478 	}
1479 
1480 	/* This makes hotplug much easier...and readable.
1481 	   we assume this for now. .*/
1482 	zone = test_pages_in_a_zone(start_pfn, end_pfn);
1483 	if (!zone) {
1484 		ret = -EINVAL;
1485 		reason = "multizone range";
1486 		goto failed_removal;
1487 	}
1488 	node = zone_to_nid(zone);
1489 
1490 	/* set above range as isolated */
1491 	ret = start_isolate_page_range(start_pfn, end_pfn,
1492 				       MIGRATE_MOVABLE,
1493 				       MEMORY_OFFLINE | REPORT_FAILURE);
1494 	if (ret < 0) {
1495 		reason = "failure to isolate range";
1496 		goto failed_removal;
1497 	}
1498 	nr_isolate_pageblock = ret;
1499 
1500 	arg.start_pfn = start_pfn;
1501 	arg.nr_pages = nr_pages;
1502 	node_states_check_changes_offline(nr_pages, zone, &arg);
1503 
1504 	ret = memory_notify(MEM_GOING_OFFLINE, &arg);
1505 	ret = notifier_to_errno(ret);
1506 	if (ret) {
1507 		reason = "notifier failure";
1508 		goto failed_removal_isolated;
1509 	}
1510 
1511 	do {
1512 		pfn = start_pfn;
1513 		do {
1514 			if (signal_pending(current)) {
1515 				ret = -EINTR;
1516 				reason = "signal backoff";
1517 				goto failed_removal_isolated;
1518 			}
1519 
1520 			cond_resched();
1521 			lru_add_drain_all();
1522 
1523 			ret = scan_movable_pages(pfn, end_pfn, &pfn);
1524 			if (!ret) {
1525 				/*
1526 				 * TODO: fatal migration failures should bail
1527 				 * out
1528 				 */
1529 				do_migrate_range(pfn, end_pfn);
1530 			}
1531 		} while (!ret);
1532 
1533 		if (ret != -ENOENT) {
1534 			reason = "unmovable page";
1535 			goto failed_removal_isolated;
1536 		}
1537 
1538 		/*
1539 		 * Dissolve free hugepages in the memory block before doing
1540 		 * offlining actually in order to make hugetlbfs's object
1541 		 * counting consistent.
1542 		 */
1543 		ret = dissolve_free_huge_pages(start_pfn, end_pfn);
1544 		if (ret) {
1545 			reason = "failure to dissolve huge pages";
1546 			goto failed_removal_isolated;
1547 		}
1548 		/* check again */
1549 		ret = walk_system_ram_range(start_pfn, end_pfn - start_pfn,
1550 					    NULL, check_pages_isolated_cb);
1551 	} while (ret);
1552 
1553 	/* Ok, all of our target is isolated.
1554 	   We cannot do rollback at this point. */
1555 	walk_system_ram_range(start_pfn, end_pfn - start_pfn,
1556 			      &offlined_pages, offline_isolated_pages_cb);
1557 	pr_info("Offlined Pages %ld\n", offlined_pages);
1558 	/*
1559 	 * Onlining will reset pagetype flags and makes migrate type
1560 	 * MOVABLE, so just need to decrease the number of isolated
1561 	 * pageblocks zone counter here.
1562 	 */
1563 	spin_lock_irqsave(&zone->lock, flags);
1564 	zone->nr_isolate_pageblock -= nr_isolate_pageblock;
1565 	spin_unlock_irqrestore(&zone->lock, flags);
1566 
1567 	/* removal success */
1568 	adjust_managed_page_count(pfn_to_page(start_pfn), -offlined_pages);
1569 	zone->present_pages -= offlined_pages;
1570 
1571 	pgdat_resize_lock(zone->zone_pgdat, &flags);
1572 	zone->zone_pgdat->node_present_pages -= offlined_pages;
1573 	pgdat_resize_unlock(zone->zone_pgdat, &flags);
1574 
1575 	init_per_zone_wmark_min();
1576 
1577 	if (!populated_zone(zone)) {
1578 		zone_pcp_reset(zone);
1579 		build_all_zonelists(NULL);
1580 	} else
1581 		zone_pcp_update(zone);
1582 
1583 	node_states_clear_node(node, &arg);
1584 	if (arg.status_change_nid >= 0) {
1585 		kswapd_stop(node);
1586 		kcompactd_stop(node);
1587 	}
1588 
1589 	vm_total_pages = nr_free_pagecache_pages();
1590 	writeback_set_ratelimit();
1591 
1592 	memory_notify(MEM_OFFLINE, &arg);
1593 	remove_pfn_range_from_zone(zone, start_pfn, nr_pages);
1594 	mem_hotplug_done();
1595 	return 0;
1596 
1597 failed_removal_isolated:
1598 	undo_isolate_page_range(start_pfn, end_pfn, MIGRATE_MOVABLE);
1599 	memory_notify(MEM_CANCEL_OFFLINE, &arg);
1600 failed_removal:
1601 	pr_debug("memory offlining [mem %#010llx-%#010llx] failed due to %s\n",
1602 		 (unsigned long long) start_pfn << PAGE_SHIFT,
1603 		 ((unsigned long long) end_pfn << PAGE_SHIFT) - 1,
1604 		 reason);
1605 	/* pushback to free area */
1606 	mem_hotplug_done();
1607 	return ret;
1608 }
1609 
1610 int offline_pages(unsigned long start_pfn, unsigned long nr_pages)
1611 {
1612 	return __offline_pages(start_pfn, start_pfn + nr_pages);
1613 }
1614 
1615 static int check_memblock_offlined_cb(struct memory_block *mem, void *arg)
1616 {
1617 	int ret = !is_memblock_offlined(mem);
1618 
1619 	if (unlikely(ret)) {
1620 		phys_addr_t beginpa, endpa;
1621 
1622 		beginpa = PFN_PHYS(section_nr_to_pfn(mem->start_section_nr));
1623 		endpa = beginpa + memory_block_size_bytes() - 1;
1624 		pr_warn("removing memory fails, because memory [%pa-%pa] is onlined\n",
1625 			&beginpa, &endpa);
1626 
1627 		return -EBUSY;
1628 	}
1629 	return 0;
1630 }
1631 
1632 static int check_cpu_on_node(pg_data_t *pgdat)
1633 {
1634 	int cpu;
1635 
1636 	for_each_present_cpu(cpu) {
1637 		if (cpu_to_node(cpu) == pgdat->node_id)
1638 			/*
1639 			 * the cpu on this node isn't removed, and we can't
1640 			 * offline this node.
1641 			 */
1642 			return -EBUSY;
1643 	}
1644 
1645 	return 0;
1646 }
1647 
1648 static int check_no_memblock_for_node_cb(struct memory_block *mem, void *arg)
1649 {
1650 	int nid = *(int *)arg;
1651 
1652 	/*
1653 	 * If a memory block belongs to multiple nodes, the stored nid is not
1654 	 * reliable. However, such blocks are always online (e.g., cannot get
1655 	 * offlined) and, therefore, are still spanned by the node.
1656 	 */
1657 	return mem->nid == nid ? -EEXIST : 0;
1658 }
1659 
1660 /**
1661  * try_offline_node
1662  * @nid: the node ID
1663  *
1664  * Offline a node if all memory sections and cpus of the node are removed.
1665  *
1666  * NOTE: The caller must call lock_device_hotplug() to serialize hotplug
1667  * and online/offline operations before this call.
1668  */
1669 void try_offline_node(int nid)
1670 {
1671 	pg_data_t *pgdat = NODE_DATA(nid);
1672 	int rc;
1673 
1674 	/*
1675 	 * If the node still spans pages (especially ZONE_DEVICE), don't
1676 	 * offline it. A node spans memory after move_pfn_range_to_zone(),
1677 	 * e.g., after the memory block was onlined.
1678 	 */
1679 	if (pgdat->node_spanned_pages)
1680 		return;
1681 
1682 	/*
1683 	 * Especially offline memory blocks might not be spanned by the
1684 	 * node. They will get spanned by the node once they get onlined.
1685 	 * However, they link to the node in sysfs and can get onlined later.
1686 	 */
1687 	rc = for_each_memory_block(&nid, check_no_memblock_for_node_cb);
1688 	if (rc)
1689 		return;
1690 
1691 	if (check_cpu_on_node(pgdat))
1692 		return;
1693 
1694 	/*
1695 	 * all memory/cpu of this node are removed, we can offline this
1696 	 * node now.
1697 	 */
1698 	node_set_offline(nid);
1699 	unregister_one_node(nid);
1700 }
1701 EXPORT_SYMBOL(try_offline_node);
1702 
1703 static void __release_memory_resource(resource_size_t start,
1704 				      resource_size_t size)
1705 {
1706 	int ret;
1707 
1708 	/*
1709 	 * When removing memory in the same granularity as it was added,
1710 	 * this function never fails. It might only fail if resources
1711 	 * have to be adjusted or split. We'll ignore the error, as
1712 	 * removing of memory cannot fail.
1713 	 */
1714 	ret = release_mem_region_adjustable(&iomem_resource, start, size);
1715 	if (ret) {
1716 		resource_size_t endres = start + size - 1;
1717 
1718 		pr_warn("Unable to release resource <%pa-%pa> (%d)\n",
1719 			&start, &endres, ret);
1720 	}
1721 }
1722 
1723 static int __ref try_remove_memory(int nid, u64 start, u64 size)
1724 {
1725 	int rc = 0;
1726 
1727 	BUG_ON(check_hotplug_memory_range(start, size));
1728 
1729 	/*
1730 	 * All memory blocks must be offlined before removing memory.  Check
1731 	 * whether all memory blocks in question are offline and return error
1732 	 * if this is not the case.
1733 	 */
1734 	rc = walk_memory_blocks(start, size, NULL, check_memblock_offlined_cb);
1735 	if (rc)
1736 		goto done;
1737 
1738 	/* remove memmap entry */
1739 	firmware_map_remove(start, start + size, "System RAM");
1740 
1741 	/*
1742 	 * Memory block device removal under the device_hotplug_lock is
1743 	 * a barrier against racing online attempts.
1744 	 */
1745 	remove_memory_block_devices(start, size);
1746 
1747 	mem_hotplug_begin();
1748 
1749 	arch_remove_memory(nid, start, size, NULL);
1750 
1751 	if (IS_ENABLED(CONFIG_ARCH_KEEP_MEMBLOCK)) {
1752 		memblock_free(start, size);
1753 		memblock_remove(start, size);
1754 	}
1755 
1756 	__release_memory_resource(start, size);
1757 
1758 	try_offline_node(nid);
1759 
1760 done:
1761 	mem_hotplug_done();
1762 	return rc;
1763 }
1764 
1765 /**
1766  * remove_memory
1767  * @nid: the node ID
1768  * @start: physical address of the region to remove
1769  * @size: size of the region to remove
1770  *
1771  * NOTE: The caller must call lock_device_hotplug() to serialize hotplug
1772  * and online/offline operations before this call, as required by
1773  * try_offline_node().
1774  */
1775 void __remove_memory(int nid, u64 start, u64 size)
1776 {
1777 
1778 	/*
1779 	 * trigger BUG() if some memory is not offlined prior to calling this
1780 	 * function
1781 	 */
1782 	if (try_remove_memory(nid, start, size))
1783 		BUG();
1784 }
1785 
1786 /*
1787  * Remove memory if every memory block is offline, otherwise return -EBUSY is
1788  * some memory is not offline
1789  */
1790 int remove_memory(int nid, u64 start, u64 size)
1791 {
1792 	int rc;
1793 
1794 	lock_device_hotplug();
1795 	rc  = try_remove_memory(nid, start, size);
1796 	unlock_device_hotplug();
1797 
1798 	return rc;
1799 }
1800 EXPORT_SYMBOL_GPL(remove_memory);
1801 
1802 /*
1803  * Try to offline and remove a memory block. Might take a long time to
1804  * finish in case memory is still in use. Primarily useful for memory devices
1805  * that logically unplugged all memory (so it's no longer in use) and want to
1806  * offline + remove the memory block.
1807  */
1808 int offline_and_remove_memory(int nid, u64 start, u64 size)
1809 {
1810 	struct memory_block *mem;
1811 	int rc = -EINVAL;
1812 
1813 	if (!IS_ALIGNED(start, memory_block_size_bytes()) ||
1814 	    size != memory_block_size_bytes())
1815 		return rc;
1816 
1817 	lock_device_hotplug();
1818 	mem = find_memory_block(__pfn_to_section(PFN_DOWN(start)));
1819 	if (mem)
1820 		rc = device_offline(&mem->dev);
1821 	/* Ignore if the device is already offline. */
1822 	if (rc > 0)
1823 		rc = 0;
1824 
1825 	/*
1826 	 * In case we succeeded to offline the memory block, remove it.
1827 	 * This cannot fail as it cannot get onlined in the meantime.
1828 	 */
1829 	if (!rc) {
1830 		rc = try_remove_memory(nid, start, size);
1831 		WARN_ON_ONCE(rc);
1832 	}
1833 	unlock_device_hotplug();
1834 
1835 	return rc;
1836 }
1837 EXPORT_SYMBOL_GPL(offline_and_remove_memory);
1838 #endif /* CONFIG_MEMORY_HOTREMOVE */
1839