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
44	enum {
45	MEMMAP_ON_MEMORY_DISABLE = 0,
46	MEMMAP_ON_MEMORY_ENABLE,
47	MEMMAP_ON_MEMORY_FORCE,
48	};
49
50	static int memmap_mode __read_mostly = MEMMAP_ON_MEMORY_DISABLE;
51
52	static inline unsigned long memory_block_memmap_size(void)
53	{
54	return PHYS_PFN(memory_block_size_bytes()) * sizeof(struct page);
55	}
56
57	static 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	*/
76	static 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(s1: val, s2: "force") || sysfs_streq(s1: val, s2: "FORCE")) {
82	mode = MEMMAP_ON_MEMORY_FORCE;
83	} else {
84	ret = kstrtobool(s: val, res: &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
102	static int get_memmap_mode(char *buffer, const struct kernel_param *kp)
103	{
104	if (*((int *)kp->arg) == MEMMAP_ON_MEMORY_FORCE)
105	return sprintf(buf: buffer, fmt: "force\n");
106	return param_get_bool(buffer, kp);
107	}
108
109	static const struct kernel_param_ops memmap_mode_ops = {
110	.set = set_memmap_mode,
111	.get = get_memmap_mode,
112	};
113	module_param_cb(memmap_on_memory, &memmap_mode_ops, &memmap_mode, 0444);
114	MODULE_PARM_DESC(memmap_on_memory, "Enable memmap on memory for memory hotplug\n"
115	"With value \"force\" it could result in memory wastage due "
116	"to memmap size limitations (Y/N/force)");
117
118	static inline bool mhp_memmap_on_memory(void)
119	{
120	return memmap_mode != MEMMAP_ON_MEMORY_DISABLE;
121	}
122	#else
123	static inline bool mhp_memmap_on_memory(void)
124	{
125	return false;
126	}
127	#endif
128
129	enum {
130	ONLINE_POLICY_CONTIG_ZONES = 0,
131	ONLINE_POLICY_AUTO_MOVABLE,
132	};
133
134	static const char * const online_policy_to_str[] = {
135	[ONLINE_POLICY_CONTIG_ZONES] = "contig-zones",
136	[ONLINE_POLICY_AUTO_MOVABLE] = "auto-movable",
137	};
138
139	static int set_online_policy(const char *val, const struct kernel_param *kp)
140	{
141	int ret = sysfs_match_string(online_policy_to_str, val);
142
143	if (ret < 0)
144	return ret;
145	*((int *)kp->arg) = ret;
146	return 0;
147	}
148
149	static int get_online_policy(char *buffer, const struct kernel_param *kp)
150	{
151	return sprintf(buf: buffer, fmt: "%s\n", online_policy_to_str[*((int *)kp->arg)]);
152	}
153
154	/*
155	* memory_hotplug.online_policy: configure online behavior when onlining without
156	* specifying a zone (MMOP_ONLINE)
157	*
158	* "contig-zones": keep zone contiguous
159	* "auto-movable": online memory to ZONE_MOVABLE if the configuration
160	* (auto_movable_ratio, auto_movable_numa_aware) allows for it
161	*/
162	static int online_policy __read_mostly = ONLINE_POLICY_CONTIG_ZONES;
163	static const struct kernel_param_ops online_policy_ops = {
164	.set = set_online_policy,
165	.get = get_online_policy,
166	};
167	module_param_cb(online_policy, &online_policy_ops, &online_policy, 0644);
168	MODULE_PARM_DESC(online_policy,
169	"Set the online policy (\"contig-zones\", \"auto-movable\") "
170	"Default: \"contig-zones\"");
171
172	/*
173	* memory_hotplug.auto_movable_ratio: specify maximum MOVABLE:KERNEL ratio
174	*
175	* The ratio represent an upper limit and the kernel might decide to not
176	* online some memory to ZONE_MOVABLE -- e.g., because hotplugged KERNEL memory
177	* doesn't allow for more MOVABLE memory.
178	*/
179	static unsigned int auto_movable_ratio __read_mostly = 301;
180	module_param(auto_movable_ratio, uint, 0644);
181	MODULE_PARM_DESC(auto_movable_ratio,
182	"Set the maximum ratio of MOVABLE:KERNEL memory in the system "
183	"in percent for \"auto-movable\" online policy. Default: 301");
184
185	/*
186	* memory_hotplug.auto_movable_numa_aware: consider numa node stats
187	*/
188	#ifdef CONFIG_NUMA
189	static bool auto_movable_numa_aware __read_mostly = true;
190	module_param(auto_movable_numa_aware, bool, 0644);
191	MODULE_PARM_DESC(auto_movable_numa_aware,
192	"Consider numa node stats in addition to global stats in "
193	"\"auto-movable\" online policy. Default: true");
194	#endif /* CONFIG_NUMA */
195
196	/*
197	* online_page_callback contains pointer to current page onlining function.
198	* Initially it is generic_online_page(). If it is required it could be
199	* changed by calling set_online_page_callback() for callback registration
200	* and restore_online_page_callback() for generic callback restore.
201	*/
202
203	static online_page_callback_t online_page_callback = generic_online_page;
204	static DEFINE_MUTEX(online_page_callback_lock);
205
206	DEFINE_STATIC_PERCPU_RWSEM(mem_hotplug_lock);
207
208	void get_online_mems(void)
209	{
210	percpu_down_read(sem: &mem_hotplug_lock);
211	}
212
213	void put_online_mems(void)
214	{
215	percpu_up_read(sem: &mem_hotplug_lock);
216	}
217
218	bool movable_node_enabled = false;
219
220	#ifndef CONFIG_MEMORY_HOTPLUG_DEFAULT_ONLINE
221	int mhp_default_online_type = MMOP_OFFLINE;
222	#else
223	int mhp_default_online_type = MMOP_ONLINE;
224	#endif
225
226	static int __init setup_memhp_default_state(char *str)
227	{
228	const int online_type = mhp_online_type_from_str(str);
229
230	if (online_type >= 0)
231	mhp_default_online_type = online_type;
232
233	return 1;
234	}
235	__setup("memhp_default_state=", setup_memhp_default_state);
236
237	void mem_hotplug_begin(void)
238	{
239	cpus_read_lock();
240	percpu_down_write(&mem_hotplug_lock);
241	}
242
243	void mem_hotplug_done(void)
244	{
245	percpu_up_write(&mem_hotplug_lock);
246	cpus_read_unlock();
247	}
248
249	u64 max_mem_size = U64_MAX;
250
251	/* add this memory to iomem resource */
252	static struct resource *register_memory_resource(u64 start, u64 size,
253	const char *resource_name)
254	{
255	struct resource *res;
256	unsigned long flags = IORESOURCE_SYSTEM_RAM | IORESOURCE_BUSY;
257
258	if (strcmp(resource_name, "System RAM"))
259	flags |= IORESOURCE_SYSRAM_DRIVER_MANAGED;
260
261	if (!mhp_range_allowed(start, size, need_mapping: true))
262	return ERR_PTR(error: -E2BIG);
263
264	/*
265	* Make sure value parsed from 'mem=' only restricts memory adding
266	* while booting, so that memory hotplug won't be impacted. Please
267	* refer to document of 'mem=' in kernel-parameters.txt for more
268	* details.
269	*/
270	if (start + size > max_mem_size && system_state < SYSTEM_RUNNING)
271	return ERR_PTR(error: -E2BIG);
272
273	/*
274	* Request ownership of the new memory range. This might be
275	* a child of an existing resource that was present but
276	* not marked as busy.
277	*/
278	res = __request_region(&iomem_resource, start, n: size,
279	name: resource_name, flags);
280
281	if (!res) {
282	pr_debug("Unable to reserve System RAM region: %016llx->%016llx\n",
283	start, start + size);
284	return ERR_PTR(error: -EEXIST);
285	}
286	return res;
287	}
288
289	static void release_memory_resource(struct resource *res)
290	{
291	if (!res)
292	return;
293	release_resource(new: res);
294	kfree(objp: res);
295	}
296
297	static int check_pfn_span(unsigned long pfn, unsigned long nr_pages)
298	{
299	/*
300	* Disallow all operations smaller than a sub-section and only
301	* allow operations smaller than a section for
302	* SPARSEMEM_VMEMMAP. Note that check_hotplug_memory_range()
303	* enforces a larger memory_block_size_bytes() granularity for
304	* memory that will be marked online, so this check should only
305	* fire for direct arch_{add,remove}_memory() users outside of
306	* add_memory_resource().
307	*/
308	unsigned long min_align;
309
310	if (IS_ENABLED(CONFIG_SPARSEMEM_VMEMMAP))
311	min_align = PAGES_PER_SUBSECTION;
312	else
313	min_align = PAGES_PER_SECTION;
314	if (!IS_ALIGNED(pfn | nr_pages, min_align))
315	return -EINVAL;
316	return 0;
317	}
318
319	/*
320	* Return page for the valid pfn only if the page is online. All pfn
321	* walkers which rely on the fully initialized page->flags and others
322	* should use this rather than pfn_valid && pfn_to_page
323	*/
324	struct page *pfn_to_online_page(unsigned long pfn)
325	{
326	unsigned long nr = pfn_to_section_nr(pfn);
327	struct dev_pagemap *pgmap;
328	struct mem_section *ms;
329
330	if (nr >= NR_MEM_SECTIONS)
331	return NULL;
332
333	ms = __nr_to_section(nr);
334	if (!online_section(section: ms))
335	return NULL;
336
337	/*
338	* Save some code text when online_section() +
339	* pfn_section_valid() are sufficient.
340	*/
341	if (IS_ENABLED(CONFIG_HAVE_ARCH_PFN_VALID) && !pfn_valid(pfn))
342	return NULL;
343
344	if (!pfn_section_valid(ms, pfn))
345	return NULL;
346
347	if (!online_device_section(section: ms))
348	return pfn_to_page(pfn);
349
350	/*
351	* Slowpath: when ZONE_DEVICE collides with
352	* ZONE_{NORMAL,MOVABLE} within the same section some pfns in
353	* the section may be 'offline' but 'valid'. Only
354	* get_dev_pagemap() can determine sub-section online status.
355	*/
356	pgmap = get_dev_pagemap(pfn, NULL);
357	put_dev_pagemap(pgmap);
358
359	/* The presence of a pgmap indicates ZONE_DEVICE offline pfn */
360	if (pgmap)
361	return NULL;
362
363	return pfn_to_page(pfn);
364	}
365	EXPORT_SYMBOL_GPL(pfn_to_online_page);
366
367	int __ref __add_pages(int nid, unsigned long pfn, unsigned long nr_pages,
368	struct mhp_params *params)
369	{
370	const unsigned long end_pfn = pfn + nr_pages;
371	unsigned long cur_nr_pages;
372	int err;
373	struct vmem_altmap *altmap = params->altmap;
374
375	if (WARN_ON_ONCE(!pgprot_val(params->pgprot)))
376	return -EINVAL;
377
378	VM_BUG_ON(!mhp_range_allowed(PFN_PHYS(pfn), nr_pages * PAGE_SIZE, false));
379
380	if (altmap) {
381	/*
382	* Validate altmap is within bounds of the total request
383	*/
384	if (altmap->base_pfn != pfn
385	|| vmem_altmap_offset(altmap) > nr_pages) {
386	pr_warn_once("memory add fail, invalid altmap\n");
387	return -EINVAL;
388	}
389	altmap->alloc = 0;
390	}
391
392	if (check_pfn_span(pfn, nr_pages)) {
393	WARN(1, "Misaligned %s start: %#lx end: %#lx\n", __func__, pfn, pfn + nr_pages - 1);
394	return -EINVAL;
395	}
396
397	for (; pfn < end_pfn; pfn += cur_nr_pages) {
398	/* Select all remaining pages up to the next section boundary */
399	cur_nr_pages = min(end_pfn - pfn,
400	SECTION_ALIGN_UP(pfn + 1) - pfn);
401	err = sparse_add_section(nid, pfn, nr_pages: cur_nr_pages, altmap,
402	pgmap: params->pgmap);
403	if (err)
404	break;
405	cond_resched();
406	}
407	vmemmap_populate_print_last();
408	return err;
409	}
410
411	/* find the smallest valid pfn in the range [start_pfn, end_pfn) */
412	static unsigned long find_smallest_section_pfn(int nid, struct zone *zone,
413	unsigned long start_pfn,
414	unsigned long end_pfn)
415	{
416	for (; start_pfn < end_pfn; start_pfn += PAGES_PER_SUBSECTION) {
417	if (unlikely(!pfn_to_online_page(start_pfn)))
418	continue;
419
420	if (unlikely(pfn_to_nid(start_pfn) != nid))
421	continue;
422
423	if (zone != page_zone(pfn_to_page(start_pfn)))
424	continue;
425
426	return start_pfn;
427	}
428
429	return 0;
430	}
431
432	/* find the biggest valid pfn in the range [start_pfn, end_pfn). */
433	static unsigned long find_biggest_section_pfn(int nid, struct zone *zone,
434	unsigned long start_pfn,
435	unsigned long end_pfn)
436	{
437	unsigned long pfn;
438
439	/* pfn is the end pfn of a memory section. */
440	pfn = end_pfn - 1;
441	for (; pfn >= start_pfn; pfn -= PAGES_PER_SUBSECTION) {
442	if (unlikely(!pfn_to_online_page(pfn)))
443	continue;
444
445	if (unlikely(pfn_to_nid(pfn) != nid))
446	continue;
447
448	if (zone != page_zone(pfn_to_page(pfn)))
449	continue;
450
451	return pfn;
452	}
453
454	return 0;
455	}
456
457	static void shrink_zone_span(struct zone *zone, unsigned long start_pfn,
458	unsigned long end_pfn)
459	{
460	unsigned long pfn;
461	int nid = zone_to_nid(zone);
462
463	if (zone->zone_start_pfn == start_pfn) {
464	/*
465	* If the section is smallest section in the zone, it need
466	* shrink zone->zone_start_pfn and zone->zone_spanned_pages.
467	* In this case, we find second smallest valid mem_section
468	* for shrinking zone.
469	*/
470	pfn = find_smallest_section_pfn(nid, zone, start_pfn: end_pfn,
471	end_pfn: zone_end_pfn(zone));
472	if (pfn) {
473	zone->spanned_pages = zone_end_pfn(zone) - pfn;
474	zone->zone_start_pfn = pfn;
475	} else {
476	zone->zone_start_pfn = 0;
477	zone->spanned_pages = 0;
478	}
479	} else if (zone_end_pfn(zone) == end_pfn) {
480	/*
481	* If the section is biggest section in the zone, it need
482	* shrink zone->spanned_pages.
483	* In this case, we find second biggest valid mem_section for
484	* shrinking zone.
485	*/
486	pfn = find_biggest_section_pfn(nid, zone, start_pfn: zone->zone_start_pfn,
487	end_pfn: start_pfn);
488	if (pfn)
489	zone->spanned_pages = pfn - zone->zone_start_pfn + 1;
490	else {
491	zone->zone_start_pfn = 0;
492	zone->spanned_pages = 0;
493	}
494	}
495	}
496
497	static void update_pgdat_span(struct pglist_data *pgdat)
498	{
499	unsigned long node_start_pfn = 0, node_end_pfn = 0;
500	struct zone *zone;
501
502	for (zone = pgdat->node_zones;
503	zone < pgdat->node_zones + MAX_NR_ZONES; zone++) {
504	unsigned long end_pfn = zone_end_pfn(zone);
505
506	/* No need to lock the zones, they can't change. */
507	if (!zone->spanned_pages)
508	continue;
509	if (!node_end_pfn) {
510	node_start_pfn = zone->zone_start_pfn;
511	node_end_pfn = end_pfn;
512	continue;
513	}
514
515	if (end_pfn > node_end_pfn)
516	node_end_pfn = end_pfn;
517	if (zone->zone_start_pfn < node_start_pfn)
518	node_start_pfn = zone->zone_start_pfn;
519	}
520
521	pgdat->node_start_pfn = node_start_pfn;
522	pgdat->node_spanned_pages = node_end_pfn - node_start_pfn;
523	}
524
525	void __ref remove_pfn_range_from_zone(struct zone *zone,
526	unsigned long start_pfn,
527	unsigned long nr_pages)
528	{
529	const unsigned long end_pfn = start_pfn + nr_pages;
530	struct pglist_data *pgdat = zone->zone_pgdat;
531	unsigned long pfn, cur_nr_pages;
532
533	/* Poison struct pages because they are now uninitialized again. */
534	for (pfn = start_pfn; pfn < end_pfn; pfn += cur_nr_pages) {
535	cond_resched();
536
537	/* Select all remaining pages up to the next section boundary */
538	cur_nr_pages =
539	min(end_pfn - pfn, SECTION_ALIGN_UP(pfn + 1) - pfn);
540	page_init_poison(pfn_to_page(pfn),
541	size: sizeof(struct page) * cur_nr_pages);
542	}
543
544	/*
545	* Zone shrinking code cannot properly deal with ZONE_DEVICE. So
546	* we will not try to shrink the zones - which is okay as
547	* set_zone_contiguous() cannot deal with ZONE_DEVICE either way.
548	*/
549	if (zone_is_zone_device(zone))
550	return;
551
552	clear_zone_contiguous(zone);
553
554	shrink_zone_span(zone, start_pfn, end_pfn: start_pfn + nr_pages);
555	update_pgdat_span(pgdat);
556
557	set_zone_contiguous(zone);
558	}
559
560	/**
561	* __remove_pages() - remove sections of pages
562	* @pfn: starting pageframe (must be aligned to start of a section)
563	* @nr_pages: number of pages to remove (must be multiple of section size)
564	* @altmap: alternative device page map or %NULL if default memmap is used
565	*
566	* Generic helper function to remove section mappings and sysfs entries
567	* for the section of the memory we are removing. Caller needs to make
568	* sure that pages are marked reserved and zones are adjust properly by
569	* calling offline_pages().
570	*/
571	void __remove_pages(unsigned long pfn, unsigned long nr_pages,
572	struct vmem_altmap *altmap)
573	{
574	const unsigned long end_pfn = pfn + nr_pages;
575	unsigned long cur_nr_pages;
576
577	if (check_pfn_span(pfn, nr_pages)) {
578	WARN(1, "Misaligned %s start: %#lx end: %#lx\n", __func__, pfn, pfn + nr_pages - 1);
579	return;
580	}
581
582	for (; pfn < end_pfn; pfn += cur_nr_pages) {
583	cond_resched();
584	/* Select all remaining pages up to the next section boundary */
585	cur_nr_pages = min(end_pfn - pfn,
586	SECTION_ALIGN_UP(pfn + 1) - pfn);
587	sparse_remove_section(pfn, nr_pages: cur_nr_pages, altmap);
588	}
589	}
590
591	int set_online_page_callback(online_page_callback_t callback)
592	{
593	int rc = -EINVAL;
594
595	get_online_mems();
596	mutex_lock(&online_page_callback_lock);
597
598	if (online_page_callback == generic_online_page) {
599	online_page_callback = callback;
600	rc = 0;
601	}
602
603	mutex_unlock(lock: &online_page_callback_lock);
604	put_online_mems();
605
606	return rc;
607	}
608	EXPORT_SYMBOL_GPL(set_online_page_callback);
609
610	int restore_online_page_callback(online_page_callback_t callback)
611	{
612	int rc = -EINVAL;
613
614	get_online_mems();
615	mutex_lock(&online_page_callback_lock);
616
617	if (online_page_callback == callback) {
618	online_page_callback = generic_online_page;
619	rc = 0;
620	}
621
622	mutex_unlock(lock: &online_page_callback_lock);
623	put_online_mems();
624
625	return rc;
626	}
627	EXPORT_SYMBOL_GPL(restore_online_page_callback);
628
629	void generic_online_page(struct page *page, unsigned int order)
630	{
631	/*
632	* Freeing the page with debug_pagealloc enabled will try to unmap it,
633	* so we should map it first. This is better than introducing a special
634	* case in page freeing fast path.
635	*/
636	debug_pagealloc_map_pages(page, numpages: 1 << order);
637	__free_pages_core(page, order);
638	totalram_pages_add(count: 1UL << order);
639	}
640	EXPORT_SYMBOL_GPL(generic_online_page);
641
642	static void online_pages_range(unsigned long start_pfn, unsigned long nr_pages)
643	{
644	const unsigned long end_pfn = start_pfn + nr_pages;
645	unsigned long pfn;
646
647	/*
648	* Online the pages in MAX_ORDER aligned chunks. The callback might
649	* decide to not expose all pages to the buddy (e.g., expose them
650	* later). We account all pages as being online and belonging to this
651	* zone ("present").
652	* When using memmap_on_memory, the range might not be aligned to
653	* MAX_ORDER_NR_PAGES - 1, but pageblock aligned. __ffs() will detect
654	* this and the first chunk to online will be pageblock_nr_pages.
655	*/
656	for (pfn = start_pfn; pfn < end_pfn;) {
657	int order;
658
659	/*
660	* Free to online pages in the largest chunks alignment allows.
661	*
662	* __ffs() behaviour is undefined for 0. start == 0 is
663	* MAX_ORDER-aligned, Set order to MAX_ORDER for the case.
664	*/
665	if (pfn)
666	order = min_t(int, MAX_ORDER, __ffs(pfn));
667	else
668	order = MAX_ORDER;
669
670	(*online_page_callback)(pfn_to_page(pfn), order);
671	pfn += (1UL << order);
672	}
673
674	/* mark all involved sections as online */
675	online_mem_sections(start_pfn, end_pfn);
676	}
677
678	/* check which state of node_states will be changed when online memory */
679	static void node_states_check_changes_online(unsigned long nr_pages,
680	struct zone *zone, struct memory_notify *arg)
681	{
682	int nid = zone_to_nid(zone);
683
684	arg->status_change_nid = NUMA_NO_NODE;
685	arg->status_change_nid_normal = NUMA_NO_NODE;
686
687	if (!node_state(node: nid, state: N_MEMORY))
688	arg->status_change_nid = nid;
689	if (zone_idx(zone) <= ZONE_NORMAL && !node_state(node: nid, state: N_NORMAL_MEMORY))
690	arg->status_change_nid_normal = nid;
691	}
692
693	static void node_states_set_node(int node, struct memory_notify *arg)
694	{
695	if (arg->status_change_nid_normal >= 0)
696	node_set_state(node, state: N_NORMAL_MEMORY);
697
698	if (arg->status_change_nid >= 0)
699	node_set_state(node, state: N_MEMORY);
700	}
701
702	static void __meminit resize_zone_range(struct zone *zone, unsigned long start_pfn,
703	unsigned long nr_pages)
704	{
705	unsigned long old_end_pfn = zone_end_pfn(zone);
706
707	if (zone_is_empty(zone) || start_pfn < zone->zone_start_pfn)
708	zone->zone_start_pfn = start_pfn;
709
710	zone->spanned_pages = max(start_pfn + nr_pages, old_end_pfn) - zone->zone_start_pfn;
711	}
712
713	static void __meminit resize_pgdat_range(struct pglist_data *pgdat, unsigned long start_pfn,
714	unsigned long nr_pages)
715	{
716	unsigned long old_end_pfn = pgdat_end_pfn(pgdat);
717
718	if (!pgdat->node_spanned_pages || start_pfn < pgdat->node_start_pfn)
719	pgdat->node_start_pfn = start_pfn;
720
721	pgdat->node_spanned_pages = max(start_pfn + nr_pages, old_end_pfn) - pgdat->node_start_pfn;
722
723	}
724
725	#ifdef CONFIG_ZONE_DEVICE
726	static void section_taint_zone_device(unsigned long pfn)
727	{
728	struct mem_section *ms = __pfn_to_section(pfn);
729
730	ms->section_mem_map |= SECTION_TAINT_ZONE_DEVICE;
731	}
732	#else
733	static inline void section_taint_zone_device(unsigned long pfn)
734	{
735	}
736	#endif
737
738	/*
739	* Associate the pfn range with the given zone, initializing the memmaps
740	* and resizing the pgdat/zone data to span the added pages. After this
741	* call, all affected pages are PG_reserved.
742	*
743	* All aligned pageblocks are initialized to the specified migratetype
744	* (usually MIGRATE_MOVABLE). Besides setting the migratetype, no related
745	* zone stats (e.g., nr_isolate_pageblock) are touched.
746	*/
747	void __ref move_pfn_range_to_zone(struct zone *zone, unsigned long start_pfn,
748	unsigned long nr_pages,
749	struct vmem_altmap *altmap, int migratetype)
750	{
751	struct pglist_data *pgdat = zone->zone_pgdat;
752	int nid = pgdat->node_id;
753
754	clear_zone_contiguous(zone);
755
756	if (zone_is_empty(zone))
757	init_currently_empty_zone(zone, start_pfn, size: nr_pages);
758	resize_zone_range(zone, start_pfn, nr_pages);
759	resize_pgdat_range(pgdat, start_pfn, nr_pages);
760
761	/*
762	* Subsection population requires care in pfn_to_online_page().
763	* Set the taint to enable the slow path detection of
764	* ZONE_DEVICE pages in an otherwise ZONE_{NORMAL,MOVABLE}
765	* section.
766	*/
767	if (zone_is_zone_device(zone)) {
768	if (!IS_ALIGNED(start_pfn, PAGES_PER_SECTION))
769	section_taint_zone_device(pfn: start_pfn);
770	if (!IS_ALIGNED(start_pfn + nr_pages, PAGES_PER_SECTION))
771	section_taint_zone_device(pfn: start_pfn + nr_pages);
772	}
773
774	/*
775	* TODO now we have a visible range of pages which are not associated
776	* with their zone properly. Not nice but set_pfnblock_flags_mask
777	* expects the zone spans the pfn range. All the pages in the range
778	* are reserved so nobody should be touching them so we should be safe
779	*/
780	memmap_init_range(nr_pages, nid, zone_idx(zone), start_pfn, 0,
781	MEMINIT_HOTPLUG, altmap, migratetype);
782
783	set_zone_contiguous(zone);
784	}
785
786	struct auto_movable_stats {
787	unsigned long kernel_early_pages;
788	unsigned long movable_pages;
789	};
790
791	static void auto_movable_stats_account_zone(struct auto_movable_stats *stats,
792	struct zone *zone)
793	{
794	if (zone_idx(zone) == ZONE_MOVABLE) {
795	stats->movable_pages += zone->present_pages;
796	} else {
797	stats->kernel_early_pages += zone->present_early_pages;
798	#ifdef CONFIG_CMA
799	/*
800	* CMA pages (never on hotplugged memory) behave like
801	* ZONE_MOVABLE.
802	*/
803	stats->movable_pages += zone->cma_pages;
804	stats->kernel_early_pages -= zone->cma_pages;
805	#endif /* CONFIG_CMA */
806	}
807	}
808	struct auto_movable_group_stats {
809	unsigned long movable_pages;
810	unsigned long req_kernel_early_pages;
811	};
812
813	static int auto_movable_stats_account_group(struct memory_group *group,
814	void *arg)
815	{
816	const int ratio = READ_ONCE(auto_movable_ratio);
817	struct auto_movable_group_stats *stats = arg;
818	long pages;
819
820	/*
821	* We don't support modifying the config while the auto-movable online
822	* policy is already enabled. Just avoid the division by zero below.
823	*/
824	if (!ratio)
825	return 0;
826
827	/*
828	* Calculate how many early kernel pages this group requires to
829	* satisfy the configured zone ratio.
830	*/
831	pages = group->present_movable_pages * 100 / ratio;
832	pages -= group->present_kernel_pages;
833
834	if (pages > 0)
835	stats->req_kernel_early_pages += pages;
836	stats->movable_pages += group->present_movable_pages;
837	return 0;
838	}
839
840	static bool auto_movable_can_online_movable(int nid, struct memory_group *group,
841	unsigned long nr_pages)
842	{
843	unsigned long kernel_early_pages, movable_pages;
844	struct auto_movable_group_stats group_stats = {};
845	struct auto_movable_stats stats = {};
846	pg_data_t *pgdat = NODE_DATA(nid);
847	struct zone *zone;
848	int i;
849
850	/* Walk all relevant zones and collect MOVABLE vs. KERNEL stats. */
851	if (nid == NUMA_NO_NODE) {
852	/* TODO: cache values */
853	for_each_populated_zone(zone)
854	auto_movable_stats_account_zone(stats: &stats, zone);
855	} else {
856	for (i = 0; i < MAX_NR_ZONES; i++) {
857	zone = pgdat->node_zones + i;
858	if (populated_zone(zone))
859	auto_movable_stats_account_zone(stats: &stats, zone);
860	}
861	}
862
863	kernel_early_pages = stats.kernel_early_pages;
864	movable_pages = stats.movable_pages;
865
866	/*
867	* Kernel memory inside dynamic memory group allows for more MOVABLE
868	* memory within the same group. Remove the effect of all but the
869	* current group from the stats.
870	*/
871	walk_dynamic_memory_groups(nid, func: auto_movable_stats_account_group,
872	excluded: group, arg: &group_stats);
873	if (kernel_early_pages <= group_stats.req_kernel_early_pages)
874	return false;
875	kernel_early_pages -= group_stats.req_kernel_early_pages;
876	movable_pages -= group_stats.movable_pages;
877
878	if (group && group->is_dynamic)
879	kernel_early_pages += group->present_kernel_pages;
880
881	/*
882	* Test if we could online the given number of pages to ZONE_MOVABLE
883	* and still stay in the configured ratio.
884	*/
885	movable_pages += nr_pages;
886	return movable_pages <= (auto_movable_ratio * kernel_early_pages) / 100;
887	}
888
889	/*
890	* Returns a default kernel memory zone for the given pfn range.
891	* If no kernel zone covers this pfn range it will automatically go
892	* to the ZONE_NORMAL.
893	*/
894	static struct zone *default_kernel_zone_for_pfn(int nid, unsigned long start_pfn,
895	unsigned long nr_pages)
896	{
897	struct pglist_data *pgdat = NODE_DATA(nid);
898	int zid;
899
900	for (zid = 0; zid < ZONE_NORMAL; zid++) {
901	struct zone *zone = &pgdat->node_zones[zid];
902
903	if (zone_intersects(zone, start_pfn, nr_pages))
904	return zone;
905	}
906
907	return &pgdat->node_zones[ZONE_NORMAL];
908	}
909
910	/*
911	* Determine to which zone to online memory dynamically based on user
912	* configuration and system stats. We care about the following ratio:
913	*
914	* MOVABLE : KERNEL
915	*
916	* Whereby MOVABLE is memory in ZONE_MOVABLE and KERNEL is memory in
917	* one of the kernel zones. CMA pages inside one of the kernel zones really
918	* behaves like ZONE_MOVABLE, so we treat them accordingly.
919	*
920	* We don't allow for hotplugged memory in a KERNEL zone to increase the
921	* amount of MOVABLE memory we can have, so we end up with:
922	*
923	* MOVABLE : KERNEL_EARLY
924	*
925	* Whereby KERNEL_EARLY is memory in one of the kernel zones, available sinze
926	* boot. We base our calculation on KERNEL_EARLY internally, because:
927	*
928	* a) Hotplugged memory in one of the kernel zones can sometimes still get
929	* hotunplugged, especially when hot(un)plugging individual memory blocks.
930	* There is no coordination across memory devices, therefore "automatic"
931	* hotunplugging, as implemented in hypervisors, could result in zone
932	* imbalances.
933	* b) Early/boot memory in one of the kernel zones can usually not get
934	* hotunplugged again (e.g., no firmware interface to unplug, fragmented
935	* with unmovable allocations). While there are corner cases where it might
936	* still work, it is barely relevant in practice.
937	*
938	* Exceptions are dynamic memory groups, which allow for more MOVABLE
939	* memory within the same memory group -- because in that case, there is
940	* coordination within the single memory device managed by a single driver.
941	*
942	* We rely on "present pages" instead of "managed pages", as the latter is
943	* highly unreliable and dynamic in virtualized environments, and does not
944	* consider boot time allocations. For example, memory ballooning adjusts the
945	* managed pages when inflating/deflating the balloon, and balloon compaction
946	* can even migrate inflated pages between zones.
947	*
948	* Using "present pages" is better but some things to keep in mind are:
949	*
950	* a) Some memblock allocations, such as for the crashkernel area, are
951	* effectively unused by the kernel, yet they account to "present pages".
952	* Fortunately, these allocations are comparatively small in relevant setups
953	* (e.g., fraction of system memory).
954	* b) Some hotplugged memory blocks in virtualized environments, esecially
955	* hotplugged by virtio-mem, look like they are completely present, however,
956	* only parts of the memory block are actually currently usable.
957	* "present pages" is an upper limit that can get reached at runtime. As
958	* we base our calculations on KERNEL_EARLY, this is not an issue.
959	*/
960	static struct zone *auto_movable_zone_for_pfn(int nid,
961	struct memory_group *group,
962	unsigned long pfn,
963	unsigned long nr_pages)
964	{
965	unsigned long online_pages = 0, max_pages, end_pfn;
966	struct page *page;
967
968	if (!auto_movable_ratio)
969	goto kernel_zone;
970
971	if (group && !group->is_dynamic) {
972	max_pages = group->s.max_pages;
973	online_pages = group->present_movable_pages;
974
975	/* If anything is !MOVABLE online the rest !MOVABLE. */
976	if (group->present_kernel_pages)
977	goto kernel_zone;
978	} else if (!group || group->d.unit_pages == nr_pages) {
979	max_pages = nr_pages;
980	} else {
981	max_pages = group->d.unit_pages;
982	/*
983	* Take a look at all online sections in the current unit.
984	* We can safely assume that all pages within a section belong
985	* to the same zone, because dynamic memory groups only deal
986	* with hotplugged memory.
987	*/
988	pfn = ALIGN_DOWN(pfn, group->d.unit_pages);
989	end_pfn = pfn + group->d.unit_pages;
990	for (; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
991	page = pfn_to_online_page(pfn);
992	if (!page)
993	continue;
994	/* If anything is !MOVABLE online the rest !MOVABLE. */
995	if (!is_zone_movable_page(page))
996	goto kernel_zone;
997	online_pages += PAGES_PER_SECTION;
998	}
999	}
1000
1001	/*
1002	* Online MOVABLE if we could *currently* online all remaining parts
1003	* MOVABLE. We expect to (add+) online them immediately next, so if
1004	* nobody interferes, all will be MOVABLE if possible.
1005	*/
1006	nr_pages = max_pages - online_pages;
1007	if (!auto_movable_can_online_movable(NUMA_NO_NODE, group, nr_pages))
1008	goto kernel_zone;
1009
1010	#ifdef CONFIG_NUMA
1011	if (auto_movable_numa_aware &&
1012	!auto_movable_can_online_movable(nid, group, nr_pages))
1013	goto kernel_zone;
1014	#endif /* CONFIG_NUMA */
1015
1016	return &NODE_DATA(nid)->node_zones[ZONE_MOVABLE];
1017	kernel_zone:
1018	return default_kernel_zone_for_pfn(nid, start_pfn: pfn, nr_pages);
1019	}
1020
1021	static inline struct zone *default_zone_for_pfn(int nid, unsigned long start_pfn,
1022	unsigned long nr_pages)
1023	{
1024	struct zone *kernel_zone = default_kernel_zone_for_pfn(nid, start_pfn,
1025	nr_pages);
1026	struct zone *movable_zone = &NODE_DATA(nid)->node_zones[ZONE_MOVABLE];
1027	bool in_kernel = zone_intersects(zone: kernel_zone, start_pfn, nr_pages);
1028	bool in_movable = zone_intersects(zone: movable_zone, start_pfn, nr_pages);
1029
1030	/*
1031	* We inherit the existing zone in a simple case where zones do not
1032	* overlap in the given range
1033	*/
1034	if (in_kernel ^ in_movable)
1035	return (in_kernel) ? kernel_zone : movable_zone;
1036
1037	/*
1038	* If the range doesn't belong to any zone or two zones overlap in the
1039	* given range then we use movable zone only if movable_node is
1040	* enabled because we always online to a kernel zone by default.
1041	*/
1042	return movable_node_enabled ? movable_zone : kernel_zone;
1043	}
1044
1045	struct zone *zone_for_pfn_range(int online_type, int nid,
1046	struct memory_group *group, unsigned long start_pfn,
1047	unsigned long nr_pages)
1048	{
1049	if (online_type == MMOP_ONLINE_KERNEL)
1050	return default_kernel_zone_for_pfn(nid, start_pfn, nr_pages);
1051
1052	if (online_type == MMOP_ONLINE_MOVABLE)
1053	return &NODE_DATA(nid)->node_zones[ZONE_MOVABLE];
1054
1055	if (online_policy == ONLINE_POLICY_AUTO_MOVABLE)
1056	return auto_movable_zone_for_pfn(nid, group, pfn: start_pfn, nr_pages);
1057
1058	return default_zone_for_pfn(nid, start_pfn, nr_pages);
1059	}
1060
1061	/*
1062	* This function should only be called by memory_block_{online,offline},
1063	* and {online,offline}_pages.
1064	*/
1065	void adjust_present_page_count(struct page *page, struct memory_group *group,
1066	long nr_pages)
1067	{
1068	struct zone *zone = page_zone(page);
1069	const bool movable = zone_idx(zone) == ZONE_MOVABLE;
1070
1071	/*
1072	* We only support onlining/offlining/adding/removing of complete
1073	* memory blocks; therefore, either all is either early or hotplugged.
1074	*/
1075	if (early_section(section: __pfn_to_section(page_to_pfn(page))))
1076	zone->present_early_pages += nr_pages;
1077	zone->present_pages += nr_pages;
1078	zone->zone_pgdat->node_present_pages += nr_pages;
1079
1080	if (group && movable)
1081	group->present_movable_pages += nr_pages;
1082	else if (group && !movable)
1083	group->present_kernel_pages += nr_pages;
1084	}
1085
1086	int mhp_init_memmap_on_memory(unsigned long pfn, unsigned long nr_pages,
1087	struct zone *zone)
1088	{
1089	unsigned long end_pfn = pfn + nr_pages;
1090	int ret, i;
1091
1092	ret = kasan_add_zero_shadow(__va(PFN_PHYS(pfn)), PFN_PHYS(nr_pages));
1093	if (ret)
1094	return ret;
1095
1096	move_pfn_range_to_zone(zone, start_pfn: pfn, nr_pages, NULL, migratetype: MIGRATE_UNMOVABLE);
1097
1098	for (i = 0; i < nr_pages; i++)
1099	SetPageVmemmapSelfHosted(pfn_to_page(pfn + i));
1100
1101	/*
1102	* It might be that the vmemmap_pages fully span sections. If that is
1103	* the case, mark those sections online here as otherwise they will be
1104	* left offline.
1105	*/
1106	if (nr_pages >= PAGES_PER_SECTION)
1107	online_mem_sections(start_pfn: pfn, ALIGN_DOWN(end_pfn, PAGES_PER_SECTION));
1108
1109	return ret;
1110	}
1111
1112	void mhp_deinit_memmap_on_memory(unsigned long pfn, unsigned long nr_pages)
1113	{
1114	unsigned long end_pfn = pfn + nr_pages;
1115
1116	/*
1117	* It might be that the vmemmap_pages fully span sections. If that is
1118	* the case, mark those sections offline here as otherwise they will be
1119	* left online.
1120	*/
1121	if (nr_pages >= PAGES_PER_SECTION)
1122	offline_mem_sections(start_pfn: pfn, ALIGN_DOWN(end_pfn, PAGES_PER_SECTION));
1123
1124	/*
1125	* The pages associated with this vmemmap have been offlined, so
1126	* we can reset its state here.
1127	*/
1128	remove_pfn_range_from_zone(zone: page_zone(pfn_to_page(pfn)), start_pfn: pfn, nr_pages);
1129	kasan_remove_zero_shadow(__va(PFN_PHYS(pfn)), PFN_PHYS(nr_pages));
1130	}
1131
1132	int __ref online_pages(unsigned long pfn, unsigned long nr_pages,
1133	struct zone *zone, struct memory_group *group)
1134	{
1135	unsigned long flags;
1136	int need_zonelists_rebuild = 0;
1137	const int nid = zone_to_nid(zone);
1138	int ret;
1139	struct memory_notify arg;
1140
1141	/*
1142	* {on,off}lining is constrained to full memory sections (or more
1143	* precisely to memory blocks from the user space POV).
1144	* memmap_on_memory is an exception because it reserves initial part
1145	* of the physical memory space for vmemmaps. That space is pageblock
1146	* aligned.
1147	*/
1148	if (WARN_ON_ONCE(!nr_pages || !pageblock_aligned(pfn) ||
1149	!IS_ALIGNED(pfn + nr_pages, PAGES_PER_SECTION)))
1150	return -EINVAL;
1151
1152	mem_hotplug_begin();
1153
1154	/* associate pfn range with the zone */
1155	move_pfn_range_to_zone(zone, start_pfn: pfn, nr_pages, NULL, migratetype: MIGRATE_ISOLATE);
1156
1157	arg.start_pfn = pfn;
1158	arg.nr_pages = nr_pages;
1159	node_states_check_changes_online(nr_pages, zone, arg: &arg);
1160
1161	ret = memory_notify(MEM_GOING_ONLINE, v: &arg);
1162	ret = notifier_to_errno(ret);
1163	if (ret)
1164	goto failed_addition;
1165
1166	/*
1167	* Fixup the number of isolated pageblocks before marking the sections
1168	* onlining, such that undo_isolate_page_range() works correctly.
1169	*/
1170	spin_lock_irqsave(&zone->lock, flags);
1171	zone->nr_isolate_pageblock += nr_pages / pageblock_nr_pages;
1172	spin_unlock_irqrestore(lock: &zone->lock, flags);
1173
1174	/*
1175	* If this zone is not populated, then it is not in zonelist.
1176	* This means the page allocator ignores this zone.
1177	* So, zonelist must be updated after online.
1178	*/
1179	if (!populated_zone(zone)) {
1180	need_zonelists_rebuild = 1;
1181	setup_zone_pageset(zone);
1182	}
1183
1184	online_pages_range(start_pfn: pfn, nr_pages);
1185	adjust_present_page_count(pfn_to_page(pfn), group, nr_pages);
1186
1187	node_states_set_node(node: nid, arg: &arg);
1188	if (need_zonelists_rebuild)
1189	build_all_zonelists(NULL);
1190
1191	/* Basic onlining is complete, allow allocation of onlined pages. */
1192	undo_isolate_page_range(start_pfn: pfn, end_pfn: pfn + nr_pages, migratetype: MIGRATE_MOVABLE);
1193
1194	/*
1195	* Freshly onlined pages aren't shuffled (e.g., all pages are placed to
1196	* the tail of the freelist when undoing isolation). Shuffle the whole
1197	* zone to make sure the just onlined pages are properly distributed
1198	* across the whole freelist - to create an initial shuffle.
1199	*/
1200	shuffle_zone(z: zone);
1201
1202	/* reinitialise watermarks and update pcp limits */
1203	init_per_zone_wmark_min();
1204
1205	kswapd_run(nid);
1206	kcompactd_run(nid);
1207
1208	writeback_set_ratelimit();
1209
1210	memory_notify(MEM_ONLINE, v: &arg);
1211	mem_hotplug_done();
1212	return 0;
1213
1214	failed_addition:
1215	pr_debug("online_pages [mem %#010llx-%#010llx] failed\n",
1216	(unsigned long long) pfn << PAGE_SHIFT,
1217	(((unsigned long long) pfn + nr_pages) << PAGE_SHIFT) - 1);
1218	memory_notify(MEM_CANCEL_ONLINE, v: &arg);
1219	remove_pfn_range_from_zone(zone, start_pfn: pfn, nr_pages);
1220	mem_hotplug_done();
1221	return ret;
1222	}
1223
1224	/* we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG */
1225	static pg_data_t __ref *hotadd_init_pgdat(int nid)
1226	{
1227	struct pglist_data *pgdat;
1228
1229	/*
1230	* NODE_DATA is preallocated (free_area_init) but its internal
1231	* state is not allocated completely. Add missing pieces.
1232	* Completely offline nodes stay around and they just need
1233	* reintialization.
1234	*/
1235	pgdat = NODE_DATA(nid);
1236
1237	/* init node's zones as empty zones, we don't have any present pages.*/
1238	free_area_init_core_hotplug(pgdat);
1239
1240	/*
1241	* The node we allocated has no zone fallback lists. For avoiding
1242	* to access not-initialized zonelist, build here.
1243	*/
1244	build_all_zonelists(pgdat);
1245
1246	return pgdat;
1247	}
1248
1249	/*
1250	* __try_online_node - online a node if offlined
1251	* @nid: the node ID
1252	* @set_node_online: Whether we want to online the node
1253	* called by cpu_up() to online a node without onlined memory.
1254	*
1255	* Returns:
1256	* 1 -> a new node has been allocated
1257	* 0 -> the node is already online
1258	* -ENOMEM -> the node could not be allocated
1259	*/
1260	static int __try_online_node(int nid, bool set_node_online)
1261	{
1262	pg_data_t *pgdat;
1263	int ret = 1;
1264
1265	if (node_online(nid))
1266	return 0;
1267
1268	pgdat = hotadd_init_pgdat(nid);
1269	if (!pgdat) {
1270	pr_err("Cannot online node %d due to NULL pgdat\n", nid);
1271	ret = -ENOMEM;
1272	goto out;
1273	}
1274
1275	if (set_node_online) {
1276	node_set_online(nid);
1277	ret = register_one_node(nid);
1278	BUG_ON(ret);
1279	}
1280	out:
1281	return ret;
1282	}
1283
1284	/*
1285	* Users of this function always want to online/register the node
1286	*/
1287	int try_online_node(int nid)
1288	{
1289	int ret;
1290
1291	mem_hotplug_begin();
1292	ret = __try_online_node(nid, set_node_online: true);
1293	mem_hotplug_done();
1294	return ret;
1295	}
1296
1297	static int check_hotplug_memory_range(u64 start, u64 size)
1298	{
1299	/* memory range must be block size aligned */
1300	if (!size || !IS_ALIGNED(start, memory_block_size_bytes()) ||
1301	!IS_ALIGNED(size, memory_block_size_bytes())) {
1302	pr_err("Block size [%#lx] unaligned hotplug range: start %#llx, size %#llx",
1303	memory_block_size_bytes(), start, size);
1304	return -EINVAL;
1305	}
1306
1307	return 0;
1308	}
1309
1310	static int online_memory_block(struct memory_block *mem, void *arg)
1311	{
1312	mem->online_type = mhp_default_online_type;
1313	return device_online(dev: &mem->dev);
1314	}
1315
1316	#ifndef arch_supports_memmap_on_memory
1317	static inline bool arch_supports_memmap_on_memory(unsigned long vmemmap_size)
1318	{
1319	/*
1320	* As default, we want the vmemmap to span a complete PMD such that we
1321	* can map the vmemmap using a single PMD if supported by the
1322	* architecture.
1323	*/
1324	return IS_ALIGNED(vmemmap_size, PMD_SIZE);
1325	}
1326	#endif
1327
1328	static bool mhp_supports_memmap_on_memory(unsigned long size)
1329	{
1330	unsigned long vmemmap_size = memory_block_memmap_size();
1331	unsigned long memmap_pages = memory_block_memmap_on_memory_pages();
1332
1333	/*
1334	* Besides having arch support and the feature enabled at runtime, we
1335	* need a few more assumptions to hold true:
1336	*
1337	* a) We span a single memory block: memory onlining/offlinin;g happens
1338	* in memory block granularity. We don't want the vmemmap of online
1339	* memory blocks to reside on offline memory blocks. In the future,
1340	* we might want to support variable-sized memory blocks to make the
1341	* feature more versatile.
1342	*
1343	* b) The vmemmap pages span complete PMDs: We don't want vmemmap code
1344	* to populate memory from the altmap for unrelated parts (i.e.,
1345	* other memory blocks)
1346	*
1347	* c) The vmemmap pages (and thereby the pages that will be exposed to
1348	* the buddy) have to cover full pageblocks: memory onlining/offlining
1349	* code requires applicable ranges to be page-aligned, for example, to
1350	* set the migratetypes properly.
1351	*
1352	* TODO: Although we have a check here to make sure that vmemmap pages
1353	* fully populate a PMD, it is not the right place to check for
1354	* this. A much better solution involves improving vmemmap code
1355	* to fallback to base pages when trying to populate vmemmap using
1356	* altmap as an alternative source of memory, and we do not exactly
1357	* populate a single PMD.
1358	*/
1359	if (!mhp_memmap_on_memory() || size != memory_block_size_bytes())
1360	return false;
1361
1362	/*
1363	* Make sure the vmemmap allocation is fully contained
1364	* so that we always allocate vmemmap memory from altmap area.
1365	*/
1366	if (!IS_ALIGNED(vmemmap_size, PAGE_SIZE))
1367	return false;
1368
1369	/*
1370	* start pfn should be pageblock_nr_pages aligned for correctly
1371	* setting migrate types
1372	*/
1373	if (!pageblock_aligned(memmap_pages))
1374	return false;
1375
1376	if (memmap_pages == PHYS_PFN(memory_block_size_bytes()))
1377	/* No effective hotplugged memory doesn't make sense. */
1378	return false;
1379
1380	return arch_supports_memmap_on_memory(vmemmap_size);
1381	}
1382
1383	/*
1384	* NOTE: The caller must call lock_device_hotplug() to serialize hotplug
1385	* and online/offline operations (triggered e.g. by sysfs).
1386	*
1387	* we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG
1388	*/
1389	int __ref add_memory_resource(int nid, struct resource *res, mhp_t mhp_flags)
1390	{
1391	struct mhp_params params = { .pgprot = pgprot_mhp(PAGE_KERNEL) };
1392	enum memblock_flags memblock_flags = MEMBLOCK_NONE;
1393	struct vmem_altmap mhp_altmap = {
1394	.base_pfn = PHYS_PFN(res->start),
1395	.end_pfn = PHYS_PFN(res->end),
1396	};
1397	struct memory_group *group = NULL;
1398	u64 start, size;
1399	bool new_node = false;
1400	int ret;
1401
1402	start = res->start;
1403	size = resource_size(res);
1404
1405	ret = check_hotplug_memory_range(start, size);
1406	if (ret)
1407	return ret;
1408
1409	if (mhp_flags & MHP_NID_IS_MGID) {
1410	group = memory_group_find_by_id(mgid: nid);
1411	if (!group)
1412	return -EINVAL;
1413	nid = group->nid;
1414	}
1415
1416	if (!node_possible(nid)) {
1417	WARN(1, "node %d was absent from the node_possible_map\n", nid);
1418	return -EINVAL;
1419	}
1420
1421	mem_hotplug_begin();
1422
1423	if (IS_ENABLED(CONFIG_ARCH_KEEP_MEMBLOCK)) {
1424	if (res->flags & IORESOURCE_SYSRAM_DRIVER_MANAGED)
1425	memblock_flags = MEMBLOCK_DRIVER_MANAGED;
1426	ret = memblock_add_node(base: start, size, nid, flags: memblock_flags);
1427	if (ret)
1428	goto error_mem_hotplug_end;
1429	}
1430
1431	ret = __try_online_node(nid, set_node_online: false);
1432	if (ret < 0)
1433	goto error;
1434	new_node = ret;
1435
1436	/*
1437	* Self hosted memmap array
1438	*/
1439	if (mhp_flags & MHP_MEMMAP_ON_MEMORY) {
1440	if (mhp_supports_memmap_on_memory(size)) {
1441	mhp_altmap.free = memory_block_memmap_on_memory_pages();
1442	params.altmap = kmalloc(size: sizeof(struct vmem_altmap), GFP_KERNEL);
1443	if (!params.altmap) {
1444	ret = -ENOMEM;
1445	goto error;
1446	}
1447
1448	memcpy(params.altmap, &mhp_altmap, sizeof(mhp_altmap));
1449	}
1450	/* fallback to not using altmap */
1451	}
1452
1453	/* call arch's memory hotadd */
1454	ret = arch_add_memory(nid, start, size, params: &params);
1455	if (ret < 0)
1456	goto error_free;
1457
1458	/* create memory block devices after memory was added */
1459	ret = create_memory_block_devices(start, size, altmap: params.altmap, group);
1460	if (ret) {
1461	arch_remove_memory(start, size, NULL);
1462	goto error_free;
1463	}
1464
1465	if (new_node) {
1466	/* If sysfs file of new node can't be created, cpu on the node
1467	* can't be hot-added. There is no rollback way now.
1468	* So, check by BUG_ON() to catch it reluctantly..
1469	* We online node here. We can't roll back from here.
1470	*/
1471	node_set_online(nid);
1472	ret = __register_one_node(nid);
1473	BUG_ON(ret);
1474	}
1475
1476	register_memory_blocks_under_node(nid, PFN_DOWN(start),
1477	PFN_UP(start + size - 1),
1478	context: MEMINIT_HOTPLUG);
1479
1480	/* create new memmap entry */
1481	if (!strcmp(res->name, "System RAM"))
1482	firmware_map_add_hotplug(start, end: start + size, type: "System RAM");
1483
1484	/* device_online() will take the lock when calling online_pages() */
1485	mem_hotplug_done();
1486
1487	/*
1488	* In case we're allowed to merge the resource, flag it and trigger
1489	* merging now that adding succeeded.
1490	*/
1491	if (mhp_flags & MHP_MERGE_RESOURCE)
1492	merge_system_ram_resource(res);
1493
1494	/* online pages if requested */
1495	if (mhp_default_online_type != MMOP_OFFLINE)
1496	walk_memory_blocks(start, size, NULL, func: online_memory_block);
1497
1498	return ret;
1499	error_free:
1500	kfree(objp: params.altmap);
1501	error:
1502	if (IS_ENABLED(CONFIG_ARCH_KEEP_MEMBLOCK))
1503	memblock_remove(base: start, size);
1504	error_mem_hotplug_end:
1505	mem_hotplug_done();
1506	return ret;
1507	}
1508
1509	/* requires device_hotplug_lock, see add_memory_resource() */
1510	int __ref __add_memory(int nid, u64 start, u64 size, mhp_t mhp_flags)
1511	{
1512	struct resource *res;
1513	int ret;
1514
1515	res = register_memory_resource(start, size, resource_name: "System RAM");
1516	if (IS_ERR(ptr: res))
1517	return PTR_ERR(ptr: res);
1518
1519	ret = add_memory_resource(nid, res, mhp_flags);
1520	if (ret < 0)
1521	release_memory_resource(res);
1522	return ret;
1523	}
1524
1525	int add_memory(int nid, u64 start, u64 size, mhp_t mhp_flags)
1526	{
1527	int rc;
1528
1529	lock_device_hotplug();
1530	rc = __add_memory(nid, start, size, mhp_flags);
1531	unlock_device_hotplug();
1532
1533	return rc;
1534	}
1535	EXPORT_SYMBOL_GPL(add_memory);
1536
1537	/*
1538	* Add special, driver-managed memory to the system as system RAM. Such
1539	* memory is not exposed via the raw firmware-provided memmap as system
1540	* RAM, instead, it is detected and added by a driver - during cold boot,
1541	* after a reboot, and after kexec.
1542	*
1543	* Reasons why this memory should not be used for the initial memmap of a
1544	* kexec kernel or for placing kexec images:
1545	* - The booting kernel is in charge of determining how this memory will be
1546	* used (e.g., use persistent memory as system RAM)
1547	* - Coordination with a hypervisor is required before this memory
1548	* can be used (e.g., inaccessible parts).
1549	*
1550	* For this memory, no entries in /sys/firmware/memmap ("raw firmware-provided
1551	* memory map") are created. Also, the created memory resource is flagged
1552	* with IORESOURCE_SYSRAM_DRIVER_MANAGED, so in-kernel users can special-case
1553	* this memory as well (esp., not place kexec images onto it).
1554	*
1555	* The resource_name (visible via /proc/iomem) has to have the format
1556	* "System RAM ($DRIVER)".
1557	*/
1558	int add_memory_driver_managed(int nid, u64 start, u64 size,
1559	const char *resource_name, mhp_t mhp_flags)
1560	{
1561	struct resource *res;
1562	int rc;
1563
1564	if (!resource_name ||
1565	strstr(resource_name, "System RAM (") != resource_name ||
1566	resource_name[strlen(resource_name) - 1] != ')')
1567	return -EINVAL;
1568
1569	lock_device_hotplug();
1570
1571	res = register_memory_resource(start, size, resource_name);
1572	if (IS_ERR(ptr: res)) {
1573	rc = PTR_ERR(ptr: res);
1574	goto out_unlock;
1575	}
1576
1577	rc = add_memory_resource(nid, res, mhp_flags);
1578	if (rc < 0)
1579	release_memory_resource(res);
1580
1581	out_unlock:
1582	unlock_device_hotplug();
1583	return rc;
1584	}
1585	EXPORT_SYMBOL_GPL(add_memory_driver_managed);
1586
1587	/*
1588	* Platforms should define arch_get_mappable_range() that provides
1589	* maximum possible addressable physical memory range for which the
1590	* linear mapping could be created. The platform returned address
1591	* range must adhere to these following semantics.
1592	*
1593	* - range.start <= range.end
1594	* - Range includes both end points [range.start..range.end]
1595	*
1596	* There is also a fallback definition provided here, allowing the
1597	* entire possible physical address range in case any platform does
1598	* not define arch_get_mappable_range().
1599	*/
1600	struct range __weak arch_get_mappable_range(void)
1601	{
1602	struct range mhp_range = {
1603	.start = 0UL,
1604	.end = -1ULL,
1605	};
1606	return mhp_range;
1607	}
1608
1609	struct range mhp_get_pluggable_range(bool need_mapping)
1610	{
1611	const u64 max_phys = (1ULL << MAX_PHYSMEM_BITS) - 1;
1612	struct range mhp_range;
1613
1614	if (need_mapping) {
1615	mhp_range = arch_get_mappable_range();
1616	if (mhp_range.start > max_phys) {
1617	mhp_range.start = 0;
1618	mhp_range.end = 0;
1619	}
1620	mhp_range.end = min_t(u64, mhp_range.end, max_phys);
1621	} else {
1622	mhp_range.start = 0;
1623	mhp_range.end = max_phys;
1624	}
1625	return mhp_range;
1626	}
1627	EXPORT_SYMBOL_GPL(mhp_get_pluggable_range);
1628
1629	bool mhp_range_allowed(u64 start, u64 size, bool need_mapping)
1630	{
1631	struct range mhp_range = mhp_get_pluggable_range(need_mapping);
1632	u64 end = start + size;
1633
1634	if (start < end && start >= mhp_range.start && (end - 1) <= mhp_range.end)
1635	return true;
1636
1637	pr_warn("Hotplug memory [%#llx-%#llx] exceeds maximum addressable range [%#llx-%#llx]\n",
1638	start, end, mhp_range.start, mhp_range.end);
1639	return false;
1640	}
1641
1642	#ifdef CONFIG_MEMORY_HOTREMOVE
1643	/*
1644	* Scan pfn range [start,end) to find movable/migratable pages (LRU pages,
1645	* non-lru movable pages and hugepages). Will skip over most unmovable
1646	* pages (esp., pages that can be skipped when offlining), but bail out on
1647	* definitely unmovable pages.
1648	*
1649	* Returns:
1650	* 0 in case a movable page is found and movable_pfn was updated.
1651	* -ENOENT in case no movable page was found.
1652	* -EBUSY in case a definitely unmovable page was found.
1653	*/
1654	static int scan_movable_pages(unsigned long start, unsigned long end,
1655	unsigned long *movable_pfn)
1656	{
1657	unsigned long pfn;
1658
1659	for (pfn = start; pfn < end; pfn++) {
1660	struct page *page, *head;
1661	unsigned long skip;
1662
1663	if (!pfn_valid(pfn))
1664	continue;
1665	page = pfn_to_page(pfn);
1666	if (PageLRU(page))
1667	goto found;
1668	if (__PageMovable(page))
1669	goto found;
1670
1671	/*
1672	* PageOffline() pages that are not marked __PageMovable() and
1673	* have a reference count > 0 (after MEM_GOING_OFFLINE) are
1674	* definitely unmovable. If their reference count would be 0,
1675	* they could at least be skipped when offlining memory.
1676	*/
1677	if (PageOffline(page) && page_count(page))
1678	return -EBUSY;
1679
1680	if (!PageHuge(page))
1681	continue;
1682	head = compound_head(page);
1683	/*
1684	* This test is racy as we hold no reference or lock. The
1685	* hugetlb page could have been free'ed and head is no longer
1686	* a hugetlb page before the following check. In such unlikely
1687	* cases false positives and negatives are possible. Calling
1688	* code must deal with these scenarios.
1689	*/
1690	if (HPageMigratable(page: head))
1691	goto found;
1692	skip = compound_nr(page: head) - (pfn - page_to_pfn(head));
1693	pfn += skip - 1;
1694	}
1695	return -ENOENT;
1696	found:
1697	*movable_pfn = pfn;
1698	return 0;
1699	}
1700
1701	static void do_migrate_range(unsigned long start_pfn, unsigned long end_pfn)
1702	{
1703	unsigned long pfn;
1704	struct page *page, *head;
1705	LIST_HEAD(source);
1706	static DEFINE_RATELIMIT_STATE(migrate_rs, DEFAULT_RATELIMIT_INTERVAL,
1707	DEFAULT_RATELIMIT_BURST);
1708
1709	for (pfn = start_pfn; pfn < end_pfn; pfn++) {
1710	struct folio *folio;
1711	bool isolated;
1712
1713	if (!pfn_valid(pfn))
1714	continue;
1715	page = pfn_to_page(pfn);
1716	folio = page_folio(page);
1717	head = &folio->page;
1718
1719	if (PageHuge(page)) {
1720	pfn = page_to_pfn(head) + compound_nr(page: head) - 1;
1721	isolate_hugetlb(folio, list: &source);
1722	continue;
1723	} else if (PageTransHuge(page))
1724	pfn = page_to_pfn(head) + thp_nr_pages(page) - 1;
1725
1726	/*
1727	* HWPoison pages have elevated reference counts so the migration would
1728	* fail on them. It also doesn't make any sense to migrate them in the
1729	* first place. Still try to unmap such a page in case it is still mapped
1730	* (e.g. current hwpoison implementation doesn't unmap KSM pages but keep
1731	* the unmap as the catch all safety net).
1732	*/
1733	if (PageHWPoison(page)) {
1734	if (WARN_ON(folio_test_lru(folio)))
1735	folio_isolate_lru(folio);
1736	if (folio_mapped(folio))
1737	try_to_unmap(folio, flags: TTU_IGNORE_MLOCK);
1738	continue;
1739	}
1740
1741	if (!get_page_unless_zero(page))
1742	continue;
1743	/*
1744	* We can skip free pages. And we can deal with pages on
1745	* LRU and non-lru movable pages.
1746	*/
1747	if (PageLRU(page))
1748	isolated = isolate_lru_page(page);
1749	else
1750	isolated = isolate_movable_page(page, ISOLATE_UNEVICTABLE);
1751	if (isolated) {
1752	list_add_tail(new: &page->lru, head: &source);
1753	if (!__PageMovable(page))
1754	inc_node_page_state(page, NR_ISOLATED_ANON +
1755	page_is_file_lru(page));
1756
1757	} else {
1758	if (__ratelimit(&migrate_rs)) {
1759	pr_warn("failed to isolate pfn %lx\n", pfn);
1760	dump_page(page, reason: "isolation failed");
1761	}
1762	}
1763	put_page(page);
1764	}
1765	if (!list_empty(head: &source)) {
1766	nodemask_t nmask = node_states[N_MEMORY];
1767	struct migration_target_control mtc = {
1768	.nmask = &nmask,
1769	.gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
1770	};
1771	int ret;
1772
1773	/*
1774	* We have checked that migration range is on a single zone so
1775	* we can use the nid of the first page to all the others.
1776	*/
1777	mtc.nid = page_to_nid(list_first_entry(&source, struct page, lru));
1778
1779	/*
1780	* try to allocate from a different node but reuse this node
1781	* if there are no other online nodes to be used (e.g. we are
1782	* offlining a part of the only existing node)
1783	*/
1784	node_clear(mtc.nid, nmask);
1785	if (nodes_empty(nmask))
1786	node_set(mtc.nid, nmask);
1787	ret = migrate_pages(l: &source, new: alloc_migration_target, NULL,
1788	private: (unsigned long)&mtc, mode: MIGRATE_SYNC, reason: MR_MEMORY_HOTPLUG, NULL);
1789	if (ret) {
1790	list_for_each_entry(page, &source, lru) {
1791	if (__ratelimit(&migrate_rs)) {
1792	pr_warn("migrating pfn %lx failed ret:%d\n",
1793	page_to_pfn(page), ret);
1794	dump_page(page, reason: "migration failure");
1795	}
1796	}
1797	putback_movable_pages(l: &source);
1798	}
1799	}
1800	}
1801
1802	static int __init cmdline_parse_movable_node(char *p)
1803	{
1804	movable_node_enabled = true;
1805	return 0;
1806	}
1807	early_param("movable_node", cmdline_parse_movable_node);
1808
1809	/* check which state of node_states will be changed when offline memory */
1810	static void node_states_check_changes_offline(unsigned long nr_pages,
1811	struct zone *zone, struct memory_notify *arg)
1812	{
1813	struct pglist_data *pgdat = zone->zone_pgdat;
1814	unsigned long present_pages = 0;
1815	enum zone_type zt;
1816
1817	arg->status_change_nid = NUMA_NO_NODE;
1818	arg->status_change_nid_normal = NUMA_NO_NODE;
1819
1820	/*
1821	* Check whether node_states[N_NORMAL_MEMORY] will be changed.
1822	* If the memory to be offline is within the range
1823	* [0..ZONE_NORMAL], and it is the last present memory there,
1824	* the zones in that range will become empty after the offlining,
1825	* thus we can determine that we need to clear the node from
1826	* node_states[N_NORMAL_MEMORY].
1827	*/
1828	for (zt = 0; zt <= ZONE_NORMAL; zt++)
1829	present_pages += pgdat->node_zones[zt].present_pages;
1830	if (zone_idx(zone) <= ZONE_NORMAL && nr_pages >= present_pages)
1831	arg->status_change_nid_normal = zone_to_nid(zone);
1832
1833	/*
1834	* We have accounted the pages from [0..ZONE_NORMAL); ZONE_HIGHMEM
1835	* does not apply as we don't support 32bit.
1836	* Here we count the possible pages from ZONE_MOVABLE.
1837	* If after having accounted all the pages, we see that the nr_pages
1838	* to be offlined is over or equal to the accounted pages,
1839	* we know that the node will become empty, and so, we can clear
1840	* it for N_MEMORY as well.
1841	*/
1842	present_pages += pgdat->node_zones[ZONE_MOVABLE].present_pages;
1843
1844	if (nr_pages >= present_pages)
1845	arg->status_change_nid = zone_to_nid(zone);
1846	}
1847
1848	static void node_states_clear_node(int node, struct memory_notify *arg)
1849	{
1850	if (arg->status_change_nid_normal >= 0)
1851	node_clear_state(node, state: N_NORMAL_MEMORY);
1852
1853	if (arg->status_change_nid >= 0)
1854	node_clear_state(node, state: N_MEMORY);
1855	}
1856
1857	static int count_system_ram_pages_cb(unsigned long start_pfn,
1858	unsigned long nr_pages, void *data)
1859	{
1860	unsigned long *nr_system_ram_pages = data;
1861
1862	*nr_system_ram_pages += nr_pages;
1863	return 0;
1864	}
1865
1866	int __ref offline_pages(unsigned long start_pfn, unsigned long nr_pages,
1867	struct zone *zone, struct memory_group *group)
1868	{
1869	const unsigned long end_pfn = start_pfn + nr_pages;
1870	unsigned long pfn, system_ram_pages = 0;
1871	const int node = zone_to_nid(zone);
1872	unsigned long flags;
1873	struct memory_notify arg;
1874	char *reason;
1875	int ret;
1876
1877	/*
1878	* {on,off}lining is constrained to full memory sections (or more
1879	* precisely to memory blocks from the user space POV).
1880	* memmap_on_memory is an exception because it reserves initial part
1881	* of the physical memory space for vmemmaps. That space is pageblock
1882	* aligned.
1883	*/
1884	if (WARN_ON_ONCE(!nr_pages || !pageblock_aligned(start_pfn) ||
1885	!IS_ALIGNED(start_pfn + nr_pages, PAGES_PER_SECTION)))
1886	return -EINVAL;
1887
1888	mem_hotplug_begin();
1889
1890	/*
1891	* Don't allow to offline memory blocks that contain holes.
1892	* Consequently, memory blocks with holes can never get onlined
1893	* via the hotplug path - online_pages() - as hotplugged memory has
1894	* no holes. This way, we e.g., don't have to worry about marking
1895	* memory holes PG_reserved, don't need pfn_valid() checks, and can
1896	* avoid using walk_system_ram_range() later.
1897	*/
1898	walk_system_ram_range(start_pfn, nr_pages, arg: &system_ram_pages,
1899	func: count_system_ram_pages_cb);
1900	if (system_ram_pages != nr_pages) {
1901	ret = -EINVAL;
1902	reason = "memory holes";
1903	goto failed_removal;
1904	}
1905
1906	/*
1907	* We only support offlining of memory blocks managed by a single zone,
1908	* checked by calling code. This is just a sanity check that we might
1909	* want to remove in the future.
1910	*/
1911	if (WARN_ON_ONCE(page_zone(pfn_to_page(start_pfn)) != zone ||
1912	page_zone(pfn_to_page(end_pfn - 1)) != zone)) {
1913	ret = -EINVAL;
1914	reason = "multizone range";
1915	goto failed_removal;
1916	}
1917
1918	/*
1919	* Disable pcplists so that page isolation cannot race with freeing
1920	* in a way that pages from isolated pageblock are left on pcplists.
1921	*/
1922	zone_pcp_disable(zone);
1923	lru_cache_disable();
1924
1925	/* set above range as isolated */
1926	ret = start_isolate_page_range(start_pfn, end_pfn,
1927	migratetype: MIGRATE_MOVABLE,
1928	MEMORY_OFFLINE | REPORT_FAILURE,
1929	GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL);
1930	if (ret) {
1931	reason = "failure to isolate range";
1932	goto failed_removal_pcplists_disabled;
1933	}
1934
1935	arg.start_pfn = start_pfn;
1936	arg.nr_pages = nr_pages;
1937	node_states_check_changes_offline(nr_pages, zone, arg: &arg);
1938
1939	ret = memory_notify(MEM_GOING_OFFLINE, v: &arg);
1940	ret = notifier_to_errno(ret);
1941	if (ret) {
1942	reason = "notifier failure";
1943	goto failed_removal_isolated;
1944	}
1945
1946	do {
1947	pfn = start_pfn;
1948	do {
1949	/*
1950	* Historically we always checked for any signal and
1951	* can't limit it to fatal signals without eventually
1952	* breaking user space.
1953	*/
1954	if (signal_pending(current)) {
1955	ret = -EINTR;
1956	reason = "signal backoff";
1957	goto failed_removal_isolated;
1958	}
1959
1960	cond_resched();
1961
1962	ret = scan_movable_pages(start: pfn, end: end_pfn, movable_pfn: &pfn);
1963	if (!ret) {
1964	/*
1965	* TODO: fatal migration failures should bail
1966	* out
1967	*/
1968	do_migrate_range(start_pfn: pfn, end_pfn);
1969	}
1970	} while (!ret);
1971
1972	if (ret != -ENOENT) {
1973	reason = "unmovable page";
1974	goto failed_removal_isolated;
1975	}
1976
1977	/*
1978	* Dissolve free hugepages in the memory block before doing
1979	* offlining actually in order to make hugetlbfs's object
1980	* counting consistent.
1981	*/
1982	ret = dissolve_free_huge_pages(start_pfn, end_pfn);
1983	if (ret) {
1984	reason = "failure to dissolve huge pages";
1985	goto failed_removal_isolated;
1986	}
1987
1988	ret = test_pages_isolated(start_pfn, end_pfn, MEMORY_OFFLINE);
1989
1990	} while (ret);
1991
1992	/* Mark all sections offline and remove free pages from the buddy. */
1993	__offline_isolated_pages(start_pfn, end_pfn);
1994	pr_debug("Offlined Pages %ld\n", nr_pages);
1995
1996	/*
1997	* The memory sections are marked offline, and the pageblock flags
1998	* effectively stale; nobody should be touching them. Fixup the number
1999	* of isolated pageblocks, memory onlining will properly revert this.
2000	*/
2001	spin_lock_irqsave(&zone->lock, flags);
2002	zone->nr_isolate_pageblock -= nr_pages / pageblock_nr_pages;
2003	spin_unlock_irqrestore(lock: &zone->lock, flags);
2004
2005	lru_cache_enable();
2006	zone_pcp_enable(zone);
2007
2008	/* removal success */
2009	adjust_managed_page_count(pfn_to_page(start_pfn), count: -nr_pages);
2010	adjust_present_page_count(pfn_to_page(start_pfn), group, nr_pages: -nr_pages);
2011
2012	/* reinitialise watermarks and update pcp limits */
2013	init_per_zone_wmark_min();
2014
2015	/*
2016	* Make sure to mark the node as memory-less before rebuilding the zone
2017	* list. Otherwise this node would still appear in the fallback lists.
2018	*/
2019	node_states_clear_node(node, arg: &arg);
2020	if (!populated_zone(zone)) {
2021	zone_pcp_reset(zone);
2022	build_all_zonelists(NULL);
2023	}
2024
2025	if (arg.status_change_nid >= 0) {
2026	kcompactd_stop(nid: node);
2027	kswapd_stop(nid: node);
2028	}
2029
2030	writeback_set_ratelimit();
2031
2032	memory_notify(MEM_OFFLINE, v: &arg);
2033	remove_pfn_range_from_zone(zone, start_pfn, nr_pages);
2034	mem_hotplug_done();
2035	return 0;
2036
2037	failed_removal_isolated:
2038	/* pushback to free area */
2039	undo_isolate_page_range(start_pfn, end_pfn, migratetype: MIGRATE_MOVABLE);
2040	memory_notify(MEM_CANCEL_OFFLINE, v: &arg);
2041	failed_removal_pcplists_disabled:
2042	lru_cache_enable();
2043	zone_pcp_enable(zone);
2044	failed_removal:
2045	pr_debug("memory offlining [mem %#010llx-%#010llx] failed due to %s\n",
2046	(unsigned long long) start_pfn << PAGE_SHIFT,
2047	((unsigned long long) end_pfn << PAGE_SHIFT) - 1,
2048	reason);
2049	mem_hotplug_done();
2050	return ret;
2051	}
2052
2053	static int check_memblock_offlined_cb(struct memory_block *mem, void *arg)
2054	{
2055	int *nid = arg;
2056
2057	*nid = mem->nid;
2058	if (unlikely(mem->state != MEM_OFFLINE)) {
2059	phys_addr_t beginpa, endpa;
2060
2061	beginpa = PFN_PHYS(section_nr_to_pfn(mem->start_section_nr));
2062	endpa = beginpa + memory_block_size_bytes() - 1;
2063	pr_warn("removing memory fails, because memory [%pa-%pa] is onlined\n",
2064	&beginpa, &endpa);
2065
2066	return -EBUSY;
2067	}
2068	return 0;
2069	}
2070
2071	static int test_has_altmap_cb(struct memory_block *mem, void *arg)
2072	{
2073	struct memory_block **mem_ptr = (struct memory_block **)arg;
2074	/*
2075	* return the memblock if we have altmap
2076	* and break callback.
2077	*/
2078	if (mem->altmap) {
2079	*mem_ptr = mem;
2080	return 1;
2081	}
2082	return 0;
2083	}
2084
2085	static int check_cpu_on_node(int nid)
2086	{
2087	int cpu;
2088
2089	for_each_present_cpu(cpu) {
2090	if (cpu_to_node(cpu) == nid)
2091	/*
2092	* the cpu on this node isn't removed, and we can't
2093	* offline this node.
2094	*/
2095	return -EBUSY;
2096	}
2097
2098	return 0;
2099	}
2100
2101	static int check_no_memblock_for_node_cb(struct memory_block *mem, void *arg)
2102	{
2103	int nid = *(int *)arg;
2104
2105	/*
2106	* If a memory block belongs to multiple nodes, the stored nid is not
2107	* reliable. However, such blocks are always online (e.g., cannot get
2108	* offlined) and, therefore, are still spanned by the node.
2109	*/
2110	return mem->nid == nid ? -EEXIST : 0;
2111	}
2112
2113	/**
2114	* try_offline_node
2115	* @nid: the node ID
2116	*
2117	* Offline a node if all memory sections and cpus of the node are removed.
2118	*
2119	* NOTE: The caller must call lock_device_hotplug() to serialize hotplug
2120	* and online/offline operations before this call.
2121	*/
2122	void try_offline_node(int nid)
2123	{
2124	int rc;
2125
2126	/*
2127	* If the node still spans pages (especially ZONE_DEVICE), don't
2128	* offline it. A node spans memory after move_pfn_range_to_zone(),
2129	* e.g., after the memory block was onlined.
2130	*/
2131	if (node_spanned_pages(nid))
2132	return;
2133
2134	/*
2135	* Especially offline memory blocks might not be spanned by the
2136	* node. They will get spanned by the node once they get onlined.
2137	* However, they link to the node in sysfs and can get onlined later.
2138	*/
2139	rc = for_each_memory_block(arg: &nid, func: check_no_memblock_for_node_cb);
2140	if (rc)
2141	return;
2142
2143	if (check_cpu_on_node(nid))
2144	return;
2145
2146	/*
2147	* all memory/cpu of this node are removed, we can offline this
2148	* node now.
2149	*/
2150	node_set_offline(nid);
2151	unregister_one_node(nid);
2152	}
2153	EXPORT_SYMBOL(try_offline_node);
2154
2155	static int __ref try_remove_memory(u64 start, u64 size)
2156	{
2157	struct memory_block *mem;
2158	int rc = 0, nid = NUMA_NO_NODE;
2159	struct vmem_altmap *altmap = NULL;
2160
2161	BUG_ON(check_hotplug_memory_range(start, size));
2162
2163	/*
2164	* All memory blocks must be offlined before removing memory. Check
2165	* whether all memory blocks in question are offline and return error
2166	* if this is not the case.
2167	*
2168	* While at it, determine the nid. Note that if we'd have mixed nodes,
2169	* we'd only try to offline the last determined one -- which is good
2170	* enough for the cases we care about.
2171	*/
2172	rc = walk_memory_blocks(start, size, arg: &nid, func: check_memblock_offlined_cb);
2173	if (rc)
2174	return rc;
2175
2176	/*
2177	* We only support removing memory added with MHP_MEMMAP_ON_MEMORY in
2178	* the same granularity it was added - a single memory block.
2179	*/
2180	if (mhp_memmap_on_memory()) {
2181	rc = walk_memory_blocks(start, size, arg: &mem, func: test_has_altmap_cb);
2182	if (rc) {
2183	if (size != memory_block_size_bytes()) {
2184	pr_warn("Refuse to remove %#llx - %#llx,"
2185	"wrong granularity\n",
2186	start, start + size);
2187	return -EINVAL;
2188	}
2189	altmap = mem->altmap;
2190	/*
2191	* Mark altmap NULL so that we can add a debug
2192	* check on memblock free.
2193	*/
2194	mem->altmap = NULL;
2195	}
2196	}
2197
2198	/* remove memmap entry */
2199	firmware_map_remove(start, end: start + size, type: "System RAM");
2200
2201	/*
2202	* Memory block device removal under the device_hotplug_lock is
2203	* a barrier against racing online attempts.
2204	*/
2205	remove_memory_block_devices(start, size);
2206
2207	mem_hotplug_begin();
2208
2209	arch_remove_memory(start, size, altmap);
2210
2211	/* Verify that all vmemmap pages have actually been freed. */
2212	if (altmap) {
2213	WARN(altmap->alloc, "Altmap not fully unmapped");
2214	kfree(objp: altmap);
2215	}
2216
2217	if (IS_ENABLED(CONFIG_ARCH_KEEP_MEMBLOCK)) {
2218	memblock_phys_free(base: start, size);
2219	memblock_remove(base: start, size);
2220	}
2221
2222	release_mem_region_adjustable(start, size);
2223
2224	if (nid != NUMA_NO_NODE)
2225	try_offline_node(nid);
2226
2227	mem_hotplug_done();
2228	return 0;
2229	}
2230
2231	/**
2232	* __remove_memory - Remove memory if every memory block is offline
2233	* @start: physical address of the region to remove
2234	* @size: size of the region to remove
2235	*
2236	* NOTE: The caller must call lock_device_hotplug() to serialize hotplug
2237	* and online/offline operations before this call, as required by
2238	* try_offline_node().
2239	*/
2240	void __remove_memory(u64 start, u64 size)
2241	{
2242
2243	/*
2244	* trigger BUG() if some memory is not offlined prior to calling this
2245	* function
2246	*/
2247	if (try_remove_memory(start, size))
2248	BUG();
2249	}
2250
2251	/*
2252	* Remove memory if every memory block is offline, otherwise return -EBUSY is
2253	* some memory is not offline
2254	*/
2255	int remove_memory(u64 start, u64 size)
2256	{
2257	int rc;
2258
2259	lock_device_hotplug();
2260	rc = try_remove_memory(start, size);
2261	unlock_device_hotplug();
2262
2263	return rc;
2264	}
2265	EXPORT_SYMBOL_GPL(remove_memory);
2266
2267	static int try_offline_memory_block(struct memory_block *mem, void *arg)
2268	{
2269	uint8_t online_type = MMOP_ONLINE_KERNEL;
2270	uint8_t **online_types = arg;
2271	struct page *page;
2272	int rc;
2273
2274	/*
2275	* Sense the online_type via the zone of the memory block. Offlining
2276	* with multiple zones within one memory block will be rejected
2277	* by offlining code ... so we don't care about that.
2278	*/
2279	page = pfn_to_online_page(section_nr_to_pfn(sec: mem->start_section_nr));
2280	if (page && zone_idx(page_zone(page)) == ZONE_MOVABLE)
2281	online_type = MMOP_ONLINE_MOVABLE;
2282
2283	rc = device_offline(dev: &mem->dev);
2284	/*
2285	* Default is MMOP_OFFLINE - change it only if offlining succeeded,
2286	* so try_reonline_memory_block() can do the right thing.
2287	*/
2288	if (!rc)
2289	**online_types = online_type;
2290
2291	(*online_types)++;
2292	/* Ignore if already offline. */
2293	return rc < 0 ? rc : 0;
2294	}
2295
2296	static int try_reonline_memory_block(struct memory_block *mem, void *arg)
2297	{
2298	uint8_t **online_types = arg;
2299	int rc;
2300
2301	if (**online_types != MMOP_OFFLINE) {
2302	mem->online_type = **online_types;
2303	rc = device_online(dev: &mem->dev);
2304	if (rc < 0)
2305	pr_warn("%s: Failed to re-online memory: %d",
2306	__func__, rc);
2307	}
2308
2309	/* Continue processing all remaining memory blocks. */
2310	(*online_types)++;
2311	return 0;
2312	}
2313
2314	/*
2315	* Try to offline and remove memory. Might take a long time to finish in case
2316	* memory is still in use. Primarily useful for memory devices that logically
2317	* unplugged all memory (so it's no longer in use) and want to offline + remove
2318	* that memory.
2319	*/
2320	int offline_and_remove_memory(u64 start, u64 size)
2321	{
2322	const unsigned long mb_count = size / memory_block_size_bytes();
2323	uint8_t *online_types, *tmp;
2324	int rc;
2325
2326	if (!IS_ALIGNED(start, memory_block_size_bytes()) ||
2327	!IS_ALIGNED(size, memory_block_size_bytes()) || !size)
2328	return -EINVAL;
2329
2330	/*
2331	* We'll remember the old online type of each memory block, so we can
2332	* try to revert whatever we did when offlining one memory block fails
2333	* after offlining some others succeeded.
2334	*/
2335	online_types = kmalloc_array(n: mb_count, size: sizeof(*online_types),
2336	GFP_KERNEL);
2337	if (!online_types)
2338	return -ENOMEM;
2339	/*
2340	* Initialize all states to MMOP_OFFLINE, so when we abort processing in
2341	* try_offline_memory_block(), we'll skip all unprocessed blocks in
2342	* try_reonline_memory_block().
2343	*/
2344	memset(online_types, MMOP_OFFLINE, mb_count);
2345
2346	lock_device_hotplug();
2347
2348	tmp = online_types;
2349	rc = walk_memory_blocks(start, size, arg: &tmp, func: try_offline_memory_block);
2350
2351	/*
2352	* In case we succeeded to offline all memory, remove it.
2353	* This cannot fail as it cannot get onlined in the meantime.
2354	*/
2355	if (!rc) {
2356	rc = try_remove_memory(start, size);
2357	if (rc)
2358	pr_err("%s: Failed to remove memory: %d", __func__, rc);
2359	}
2360
2361	/*
2362	* Rollback what we did. While memory onlining might theoretically fail
2363	* (nacked by a notifier), it barely ever happens.
2364	*/
2365	if (rc) {
2366	tmp = online_types;
2367	walk_memory_blocks(start, size, arg: &tmp,
2368	func: try_reonline_memory_block);
2369	}
2370	unlock_device_hotplug();
2371
2372	kfree(objp: online_types);
2373	return rc;
2374	}
2375	EXPORT_SYMBOL_GPL(offline_and_remove_memory);
2376	#endif /* CONFIG_MEMORY_HOTREMOVE */
2377