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