xref: /openbmc/linux/mm/memory_hotplug.c (revision 7559e757)
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 #ifdef CONFIG_MHP_MEMMAP_ON_MEMORY
45 /*
46  * memory_hotplug.memmap_on_memory parameter
47  */
48 static bool memmap_on_memory __ro_after_init;
49 module_param(memmap_on_memory, bool, 0444);
50 MODULE_PARM_DESC(memmap_on_memory, "Enable memmap on memory for memory hotplug");
51 
52 static inline bool mhp_memmap_on_memory(void)
53 {
54 	return memmap_on_memory;
55 }
56 #else
57 static inline bool mhp_memmap_on_memory(void)
58 {
59 	return false;
60 }
61 #endif
62 
63 enum {
64 	ONLINE_POLICY_CONTIG_ZONES = 0,
65 	ONLINE_POLICY_AUTO_MOVABLE,
66 };
67 
68 static const char * const online_policy_to_str[] = {
69 	[ONLINE_POLICY_CONTIG_ZONES] = "contig-zones",
70 	[ONLINE_POLICY_AUTO_MOVABLE] = "auto-movable",
71 };
72 
73 static int set_online_policy(const char *val, const struct kernel_param *kp)
74 {
75 	int ret = sysfs_match_string(online_policy_to_str, val);
76 
77 	if (ret < 0)
78 		return ret;
79 	*((int *)kp->arg) = ret;
80 	return 0;
81 }
82 
83 static int get_online_policy(char *buffer, const struct kernel_param *kp)
84 {
85 	return sprintf(buffer, "%s\n", online_policy_to_str[*((int *)kp->arg)]);
86 }
87 
88 /*
89  * memory_hotplug.online_policy: configure online behavior when onlining without
90  * specifying a zone (MMOP_ONLINE)
91  *
92  * "contig-zones": keep zone contiguous
93  * "auto-movable": online memory to ZONE_MOVABLE if the configuration
94  *                 (auto_movable_ratio, auto_movable_numa_aware) allows for it
95  */
96 static int online_policy __read_mostly = ONLINE_POLICY_CONTIG_ZONES;
97 static const struct kernel_param_ops online_policy_ops = {
98 	.set = set_online_policy,
99 	.get = get_online_policy,
100 };
101 module_param_cb(online_policy, &online_policy_ops, &online_policy, 0644);
102 MODULE_PARM_DESC(online_policy,
103 		"Set the online policy (\"contig-zones\", \"auto-movable\") "
104 		"Default: \"contig-zones\"");
105 
106 /*
107  * memory_hotplug.auto_movable_ratio: specify maximum MOVABLE:KERNEL ratio
108  *
109  * The ratio represent an upper limit and the kernel might decide to not
110  * online some memory to ZONE_MOVABLE -- e.g., because hotplugged KERNEL memory
111  * doesn't allow for more MOVABLE memory.
112  */
113 static unsigned int auto_movable_ratio __read_mostly = 301;
114 module_param(auto_movable_ratio, uint, 0644);
115 MODULE_PARM_DESC(auto_movable_ratio,
116 		"Set the maximum ratio of MOVABLE:KERNEL memory in the system "
117 		"in percent for \"auto-movable\" online policy. Default: 301");
118 
119 /*
120  * memory_hotplug.auto_movable_numa_aware: consider numa node stats
121  */
122 #ifdef CONFIG_NUMA
123 static bool auto_movable_numa_aware __read_mostly = true;
124 module_param(auto_movable_numa_aware, bool, 0644);
125 MODULE_PARM_DESC(auto_movable_numa_aware,
126 		"Consider numa node stats in addition to global stats in "
127 		"\"auto-movable\" online policy. Default: true");
128 #endif /* CONFIG_NUMA */
129 
130 /*
131  * online_page_callback contains pointer to current page onlining function.
132  * Initially it is generic_online_page(). If it is required it could be
133  * changed by calling set_online_page_callback() for callback registration
134  * and restore_online_page_callback() for generic callback restore.
135  */
136 
137 static online_page_callback_t online_page_callback = generic_online_page;
138 static DEFINE_MUTEX(online_page_callback_lock);
139 
140 DEFINE_STATIC_PERCPU_RWSEM(mem_hotplug_lock);
141 
142 void get_online_mems(void)
143 {
144 	percpu_down_read(&mem_hotplug_lock);
145 }
146 
147 void put_online_mems(void)
148 {
149 	percpu_up_read(&mem_hotplug_lock);
150 }
151 
152 bool movable_node_enabled = false;
153 
154 #ifndef CONFIG_MEMORY_HOTPLUG_DEFAULT_ONLINE
155 int mhp_default_online_type = MMOP_OFFLINE;
156 #else
157 int mhp_default_online_type = MMOP_ONLINE;
158 #endif
159 
160 static int __init setup_memhp_default_state(char *str)
161 {
162 	const int online_type = mhp_online_type_from_str(str);
163 
164 	if (online_type >= 0)
165 		mhp_default_online_type = online_type;
166 
167 	return 1;
168 }
169 __setup("memhp_default_state=", setup_memhp_default_state);
170 
171 void mem_hotplug_begin(void)
172 {
173 	cpus_read_lock();
174 	percpu_down_write(&mem_hotplug_lock);
175 }
176 
177 void mem_hotplug_done(void)
178 {
179 	percpu_up_write(&mem_hotplug_lock);
180 	cpus_read_unlock();
181 }
182 
183 u64 max_mem_size = U64_MAX;
184 
185 /* add this memory to iomem resource */
186 static struct resource *register_memory_resource(u64 start, u64 size,
187 						 const char *resource_name)
188 {
189 	struct resource *res;
190 	unsigned long flags =  IORESOURCE_SYSTEM_RAM | IORESOURCE_BUSY;
191 
192 	if (strcmp(resource_name, "System RAM"))
193 		flags |= IORESOURCE_SYSRAM_DRIVER_MANAGED;
194 
195 	if (!mhp_range_allowed(start, size, true))
196 		return ERR_PTR(-E2BIG);
197 
198 	/*
199 	 * Make sure value parsed from 'mem=' only restricts memory adding
200 	 * while booting, so that memory hotplug won't be impacted. Please
201 	 * refer to document of 'mem=' in kernel-parameters.txt for more
202 	 * details.
203 	 */
204 	if (start + size > max_mem_size && system_state < SYSTEM_RUNNING)
205 		return ERR_PTR(-E2BIG);
206 
207 	/*
208 	 * Request ownership of the new memory range.  This might be
209 	 * a child of an existing resource that was present but
210 	 * not marked as busy.
211 	 */
212 	res = __request_region(&iomem_resource, start, size,
213 			       resource_name, flags);
214 
215 	if (!res) {
216 		pr_debug("Unable to reserve System RAM region: %016llx->%016llx\n",
217 				start, start + size);
218 		return ERR_PTR(-EEXIST);
219 	}
220 	return res;
221 }
222 
223 static void release_memory_resource(struct resource *res)
224 {
225 	if (!res)
226 		return;
227 	release_resource(res);
228 	kfree(res);
229 }
230 
231 static int check_pfn_span(unsigned long pfn, unsigned long nr_pages)
232 {
233 	/*
234 	 * Disallow all operations smaller than a sub-section and only
235 	 * allow operations smaller than a section for
236 	 * SPARSEMEM_VMEMMAP. Note that check_hotplug_memory_range()
237 	 * enforces a larger memory_block_size_bytes() granularity for
238 	 * memory that will be marked online, so this check should only
239 	 * fire for direct arch_{add,remove}_memory() users outside of
240 	 * add_memory_resource().
241 	 */
242 	unsigned long min_align;
243 
244 	if (IS_ENABLED(CONFIG_SPARSEMEM_VMEMMAP))
245 		min_align = PAGES_PER_SUBSECTION;
246 	else
247 		min_align = PAGES_PER_SECTION;
248 	if (!IS_ALIGNED(pfn | nr_pages, min_align))
249 		return -EINVAL;
250 	return 0;
251 }
252 
253 /*
254  * Return page for the valid pfn only if the page is online. All pfn
255  * walkers which rely on the fully initialized page->flags and others
256  * should use this rather than pfn_valid && pfn_to_page
257  */
258 struct page *pfn_to_online_page(unsigned long pfn)
259 {
260 	unsigned long nr = pfn_to_section_nr(pfn);
261 	struct dev_pagemap *pgmap;
262 	struct mem_section *ms;
263 
264 	if (nr >= NR_MEM_SECTIONS)
265 		return NULL;
266 
267 	ms = __nr_to_section(nr);
268 	if (!online_section(ms))
269 		return NULL;
270 
271 	/*
272 	 * Save some code text when online_section() +
273 	 * pfn_section_valid() are sufficient.
274 	 */
275 	if (IS_ENABLED(CONFIG_HAVE_ARCH_PFN_VALID) && !pfn_valid(pfn))
276 		return NULL;
277 
278 	if (!pfn_section_valid(ms, pfn))
279 		return NULL;
280 
281 	if (!online_device_section(ms))
282 		return pfn_to_page(pfn);
283 
284 	/*
285 	 * Slowpath: when ZONE_DEVICE collides with
286 	 * ZONE_{NORMAL,MOVABLE} within the same section some pfns in
287 	 * the section may be 'offline' but 'valid'. Only
288 	 * get_dev_pagemap() can determine sub-section online status.
289 	 */
290 	pgmap = get_dev_pagemap(pfn, NULL);
291 	put_dev_pagemap(pgmap);
292 
293 	/* The presence of a pgmap indicates ZONE_DEVICE offline pfn */
294 	if (pgmap)
295 		return NULL;
296 
297 	return pfn_to_page(pfn);
298 }
299 EXPORT_SYMBOL_GPL(pfn_to_online_page);
300 
301 int __ref __add_pages(int nid, unsigned long pfn, unsigned long nr_pages,
302 		struct mhp_params *params)
303 {
304 	const unsigned long end_pfn = pfn + nr_pages;
305 	unsigned long cur_nr_pages;
306 	int err;
307 	struct vmem_altmap *altmap = params->altmap;
308 
309 	if (WARN_ON_ONCE(!pgprot_val(params->pgprot)))
310 		return -EINVAL;
311 
312 	VM_BUG_ON(!mhp_range_allowed(PFN_PHYS(pfn), nr_pages * PAGE_SIZE, false));
313 
314 	if (altmap) {
315 		/*
316 		 * Validate altmap is within bounds of the total request
317 		 */
318 		if (altmap->base_pfn != pfn
319 				|| vmem_altmap_offset(altmap) > nr_pages) {
320 			pr_warn_once("memory add fail, invalid altmap\n");
321 			return -EINVAL;
322 		}
323 		altmap->alloc = 0;
324 	}
325 
326 	if (check_pfn_span(pfn, nr_pages)) {
327 		WARN(1, "Misaligned %s start: %#lx end: %#lx\n", __func__, pfn, pfn + nr_pages - 1);
328 		return -EINVAL;
329 	}
330 
331 	for (; pfn < end_pfn; pfn += cur_nr_pages) {
332 		/* Select all remaining pages up to the next section boundary */
333 		cur_nr_pages = min(end_pfn - pfn,
334 				   SECTION_ALIGN_UP(pfn + 1) - pfn);
335 		err = sparse_add_section(nid, pfn, cur_nr_pages, altmap,
336 					 params->pgmap);
337 		if (err)
338 			break;
339 		cond_resched();
340 	}
341 	vmemmap_populate_print_last();
342 	return err;
343 }
344 
345 /* find the smallest valid pfn in the range [start_pfn, end_pfn) */
346 static unsigned long find_smallest_section_pfn(int nid, struct zone *zone,
347 				     unsigned long start_pfn,
348 				     unsigned long end_pfn)
349 {
350 	for (; start_pfn < end_pfn; start_pfn += PAGES_PER_SUBSECTION) {
351 		if (unlikely(!pfn_to_online_page(start_pfn)))
352 			continue;
353 
354 		if (unlikely(pfn_to_nid(start_pfn) != nid))
355 			continue;
356 
357 		if (zone != page_zone(pfn_to_page(start_pfn)))
358 			continue;
359 
360 		return start_pfn;
361 	}
362 
363 	return 0;
364 }
365 
366 /* find the biggest valid pfn in the range [start_pfn, end_pfn). */
367 static unsigned long find_biggest_section_pfn(int nid, struct zone *zone,
368 				    unsigned long start_pfn,
369 				    unsigned long end_pfn)
370 {
371 	unsigned long pfn;
372 
373 	/* pfn is the end pfn of a memory section. */
374 	pfn = end_pfn - 1;
375 	for (; pfn >= start_pfn; pfn -= PAGES_PER_SUBSECTION) {
376 		if (unlikely(!pfn_to_online_page(pfn)))
377 			continue;
378 
379 		if (unlikely(pfn_to_nid(pfn) != nid))
380 			continue;
381 
382 		if (zone != page_zone(pfn_to_page(pfn)))
383 			continue;
384 
385 		return pfn;
386 	}
387 
388 	return 0;
389 }
390 
391 static void shrink_zone_span(struct zone *zone, unsigned long start_pfn,
392 			     unsigned long end_pfn)
393 {
394 	unsigned long pfn;
395 	int nid = zone_to_nid(zone);
396 
397 	if (zone->zone_start_pfn == start_pfn) {
398 		/*
399 		 * If the section is smallest section in the zone, it need
400 		 * shrink zone->zone_start_pfn and zone->zone_spanned_pages.
401 		 * In this case, we find second smallest valid mem_section
402 		 * for shrinking zone.
403 		 */
404 		pfn = find_smallest_section_pfn(nid, zone, end_pfn,
405 						zone_end_pfn(zone));
406 		if (pfn) {
407 			zone->spanned_pages = zone_end_pfn(zone) - pfn;
408 			zone->zone_start_pfn = pfn;
409 		} else {
410 			zone->zone_start_pfn = 0;
411 			zone->spanned_pages = 0;
412 		}
413 	} else if (zone_end_pfn(zone) == end_pfn) {
414 		/*
415 		 * If the section is biggest section in the zone, it need
416 		 * shrink zone->spanned_pages.
417 		 * In this case, we find second biggest valid mem_section for
418 		 * shrinking zone.
419 		 */
420 		pfn = find_biggest_section_pfn(nid, zone, zone->zone_start_pfn,
421 					       start_pfn);
422 		if (pfn)
423 			zone->spanned_pages = pfn - zone->zone_start_pfn + 1;
424 		else {
425 			zone->zone_start_pfn = 0;
426 			zone->spanned_pages = 0;
427 		}
428 	}
429 }
430 
431 static void update_pgdat_span(struct pglist_data *pgdat)
432 {
433 	unsigned long node_start_pfn = 0, node_end_pfn = 0;
434 	struct zone *zone;
435 
436 	for (zone = pgdat->node_zones;
437 	     zone < pgdat->node_zones + MAX_NR_ZONES; zone++) {
438 		unsigned long end_pfn = zone_end_pfn(zone);
439 
440 		/* No need to lock the zones, they can't change. */
441 		if (!zone->spanned_pages)
442 			continue;
443 		if (!node_end_pfn) {
444 			node_start_pfn = zone->zone_start_pfn;
445 			node_end_pfn = end_pfn;
446 			continue;
447 		}
448 
449 		if (end_pfn > node_end_pfn)
450 			node_end_pfn = end_pfn;
451 		if (zone->zone_start_pfn < node_start_pfn)
452 			node_start_pfn = zone->zone_start_pfn;
453 	}
454 
455 	pgdat->node_start_pfn = node_start_pfn;
456 	pgdat->node_spanned_pages = node_end_pfn - node_start_pfn;
457 }
458 
459 void __ref remove_pfn_range_from_zone(struct zone *zone,
460 				      unsigned long start_pfn,
461 				      unsigned long nr_pages)
462 {
463 	const unsigned long end_pfn = start_pfn + nr_pages;
464 	struct pglist_data *pgdat = zone->zone_pgdat;
465 	unsigned long pfn, cur_nr_pages;
466 
467 	/* Poison struct pages because they are now uninitialized again. */
468 	for (pfn = start_pfn; pfn < end_pfn; pfn += cur_nr_pages) {
469 		cond_resched();
470 
471 		/* Select all remaining pages up to the next section boundary */
472 		cur_nr_pages =
473 			min(end_pfn - pfn, SECTION_ALIGN_UP(pfn + 1) - pfn);
474 		page_init_poison(pfn_to_page(pfn),
475 				 sizeof(struct page) * cur_nr_pages);
476 	}
477 
478 	/*
479 	 * Zone shrinking code cannot properly deal with ZONE_DEVICE. So
480 	 * we will not try to shrink the zones - which is okay as
481 	 * set_zone_contiguous() cannot deal with ZONE_DEVICE either way.
482 	 */
483 	if (zone_is_zone_device(zone))
484 		return;
485 
486 	clear_zone_contiguous(zone);
487 
488 	shrink_zone_span(zone, start_pfn, start_pfn + nr_pages);
489 	update_pgdat_span(pgdat);
490 
491 	set_zone_contiguous(zone);
492 }
493 
494 /**
495  * __remove_pages() - remove sections of pages
496  * @pfn: starting pageframe (must be aligned to start of a section)
497  * @nr_pages: number of pages to remove (must be multiple of section size)
498  * @altmap: alternative device page map or %NULL if default memmap is used
499  *
500  * Generic helper function to remove section mappings and sysfs entries
501  * for the section of the memory we are removing. Caller needs to make
502  * sure that pages are marked reserved and zones are adjust properly by
503  * calling offline_pages().
504  */
505 void __remove_pages(unsigned long pfn, unsigned long nr_pages,
506 		    struct vmem_altmap *altmap)
507 {
508 	const unsigned long end_pfn = pfn + nr_pages;
509 	unsigned long cur_nr_pages;
510 
511 	if (check_pfn_span(pfn, nr_pages)) {
512 		WARN(1, "Misaligned %s start: %#lx end: %#lx\n", __func__, pfn, pfn + nr_pages - 1);
513 		return;
514 	}
515 
516 	for (; pfn < end_pfn; pfn += cur_nr_pages) {
517 		cond_resched();
518 		/* Select all remaining pages up to the next section boundary */
519 		cur_nr_pages = min(end_pfn - pfn,
520 				   SECTION_ALIGN_UP(pfn + 1) - pfn);
521 		sparse_remove_section(pfn, cur_nr_pages, altmap);
522 	}
523 }
524 
525 int set_online_page_callback(online_page_callback_t callback)
526 {
527 	int rc = -EINVAL;
528 
529 	get_online_mems();
530 	mutex_lock(&online_page_callback_lock);
531 
532 	if (online_page_callback == generic_online_page) {
533 		online_page_callback = callback;
534 		rc = 0;
535 	}
536 
537 	mutex_unlock(&online_page_callback_lock);
538 	put_online_mems();
539 
540 	return rc;
541 }
542 EXPORT_SYMBOL_GPL(set_online_page_callback);
543 
544 int restore_online_page_callback(online_page_callback_t callback)
545 {
546 	int rc = -EINVAL;
547 
548 	get_online_mems();
549 	mutex_lock(&online_page_callback_lock);
550 
551 	if (online_page_callback == callback) {
552 		online_page_callback = generic_online_page;
553 		rc = 0;
554 	}
555 
556 	mutex_unlock(&online_page_callback_lock);
557 	put_online_mems();
558 
559 	return rc;
560 }
561 EXPORT_SYMBOL_GPL(restore_online_page_callback);
562 
563 void generic_online_page(struct page *page, unsigned int order)
564 {
565 	/*
566 	 * Freeing the page with debug_pagealloc enabled will try to unmap it,
567 	 * so we should map it first. This is better than introducing a special
568 	 * case in page freeing fast path.
569 	 */
570 	debug_pagealloc_map_pages(page, 1 << order);
571 	__free_pages_core(page, order);
572 	totalram_pages_add(1UL << order);
573 }
574 EXPORT_SYMBOL_GPL(generic_online_page);
575 
576 static void online_pages_range(unsigned long start_pfn, unsigned long nr_pages)
577 {
578 	const unsigned long end_pfn = start_pfn + nr_pages;
579 	unsigned long pfn;
580 
581 	/*
582 	 * Online the pages in MAX_ORDER aligned chunks. The callback might
583 	 * decide to not expose all pages to the buddy (e.g., expose them
584 	 * later). We account all pages as being online and belonging to this
585 	 * zone ("present").
586 	 * When using memmap_on_memory, the range might not be aligned to
587 	 * MAX_ORDER_NR_PAGES - 1, but pageblock aligned. __ffs() will detect
588 	 * this and the first chunk to online will be pageblock_nr_pages.
589 	 */
590 	for (pfn = start_pfn; pfn < end_pfn;) {
591 		int order;
592 
593 		/*
594 		 * Free to online pages in the largest chunks alignment allows.
595 		 *
596 		 * __ffs() behaviour is undefined for 0. start == 0 is
597 		 * MAX_ORDER-aligned, Set order to MAX_ORDER for the case.
598 		 */
599 		if (pfn)
600 			order = min_t(int, MAX_ORDER, __ffs(pfn));
601 		else
602 			order = MAX_ORDER;
603 
604 		(*online_page_callback)(pfn_to_page(pfn), order);
605 		pfn += (1UL << order);
606 	}
607 
608 	/* mark all involved sections as online */
609 	online_mem_sections(start_pfn, end_pfn);
610 }
611 
612 /* check which state of node_states will be changed when online memory */
613 static void node_states_check_changes_online(unsigned long nr_pages,
614 	struct zone *zone, struct memory_notify *arg)
615 {
616 	int nid = zone_to_nid(zone);
617 
618 	arg->status_change_nid = NUMA_NO_NODE;
619 	arg->status_change_nid_normal = NUMA_NO_NODE;
620 
621 	if (!node_state(nid, N_MEMORY))
622 		arg->status_change_nid = nid;
623 	if (zone_idx(zone) <= ZONE_NORMAL && !node_state(nid, N_NORMAL_MEMORY))
624 		arg->status_change_nid_normal = nid;
625 }
626 
627 static void node_states_set_node(int node, struct memory_notify *arg)
628 {
629 	if (arg->status_change_nid_normal >= 0)
630 		node_set_state(node, N_NORMAL_MEMORY);
631 
632 	if (arg->status_change_nid >= 0)
633 		node_set_state(node, N_MEMORY);
634 }
635 
636 static void __meminit resize_zone_range(struct zone *zone, unsigned long start_pfn,
637 		unsigned long nr_pages)
638 {
639 	unsigned long old_end_pfn = zone_end_pfn(zone);
640 
641 	if (zone_is_empty(zone) || start_pfn < zone->zone_start_pfn)
642 		zone->zone_start_pfn = start_pfn;
643 
644 	zone->spanned_pages = max(start_pfn + nr_pages, old_end_pfn) - zone->zone_start_pfn;
645 }
646 
647 static void __meminit resize_pgdat_range(struct pglist_data *pgdat, unsigned long start_pfn,
648                                      unsigned long nr_pages)
649 {
650 	unsigned long old_end_pfn = pgdat_end_pfn(pgdat);
651 
652 	if (!pgdat->node_spanned_pages || start_pfn < pgdat->node_start_pfn)
653 		pgdat->node_start_pfn = start_pfn;
654 
655 	pgdat->node_spanned_pages = max(start_pfn + nr_pages, old_end_pfn) - pgdat->node_start_pfn;
656 
657 }
658 
659 #ifdef CONFIG_ZONE_DEVICE
660 static void section_taint_zone_device(unsigned long pfn)
661 {
662 	struct mem_section *ms = __pfn_to_section(pfn);
663 
664 	ms->section_mem_map |= SECTION_TAINT_ZONE_DEVICE;
665 }
666 #else
667 static inline void section_taint_zone_device(unsigned long pfn)
668 {
669 }
670 #endif
671 
672 /*
673  * Associate the pfn range with the given zone, initializing the memmaps
674  * and resizing the pgdat/zone data to span the added pages. After this
675  * call, all affected pages are PG_reserved.
676  *
677  * All aligned pageblocks are initialized to the specified migratetype
678  * (usually MIGRATE_MOVABLE). Besides setting the migratetype, no related
679  * zone stats (e.g., nr_isolate_pageblock) are touched.
680  */
681 void __ref move_pfn_range_to_zone(struct zone *zone, unsigned long start_pfn,
682 				  unsigned long nr_pages,
683 				  struct vmem_altmap *altmap, int migratetype)
684 {
685 	struct pglist_data *pgdat = zone->zone_pgdat;
686 	int nid = pgdat->node_id;
687 
688 	clear_zone_contiguous(zone);
689 
690 	if (zone_is_empty(zone))
691 		init_currently_empty_zone(zone, start_pfn, nr_pages);
692 	resize_zone_range(zone, start_pfn, nr_pages);
693 	resize_pgdat_range(pgdat, start_pfn, nr_pages);
694 
695 	/*
696 	 * Subsection population requires care in pfn_to_online_page().
697 	 * Set the taint to enable the slow path detection of
698 	 * ZONE_DEVICE pages in an otherwise  ZONE_{NORMAL,MOVABLE}
699 	 * section.
700 	 */
701 	if (zone_is_zone_device(zone)) {
702 		if (!IS_ALIGNED(start_pfn, PAGES_PER_SECTION))
703 			section_taint_zone_device(start_pfn);
704 		if (!IS_ALIGNED(start_pfn + nr_pages, PAGES_PER_SECTION))
705 			section_taint_zone_device(start_pfn + nr_pages);
706 	}
707 
708 	/*
709 	 * TODO now we have a visible range of pages which are not associated
710 	 * with their zone properly. Not nice but set_pfnblock_flags_mask
711 	 * expects the zone spans the pfn range. All the pages in the range
712 	 * are reserved so nobody should be touching them so we should be safe
713 	 */
714 	memmap_init_range(nr_pages, nid, zone_idx(zone), start_pfn, 0,
715 			 MEMINIT_HOTPLUG, altmap, migratetype);
716 
717 	set_zone_contiguous(zone);
718 }
719 
720 struct auto_movable_stats {
721 	unsigned long kernel_early_pages;
722 	unsigned long movable_pages;
723 };
724 
725 static void auto_movable_stats_account_zone(struct auto_movable_stats *stats,
726 					    struct zone *zone)
727 {
728 	if (zone_idx(zone) == ZONE_MOVABLE) {
729 		stats->movable_pages += zone->present_pages;
730 	} else {
731 		stats->kernel_early_pages += zone->present_early_pages;
732 #ifdef CONFIG_CMA
733 		/*
734 		 * CMA pages (never on hotplugged memory) behave like
735 		 * ZONE_MOVABLE.
736 		 */
737 		stats->movable_pages += zone->cma_pages;
738 		stats->kernel_early_pages -= zone->cma_pages;
739 #endif /* CONFIG_CMA */
740 	}
741 }
742 struct auto_movable_group_stats {
743 	unsigned long movable_pages;
744 	unsigned long req_kernel_early_pages;
745 };
746 
747 static int auto_movable_stats_account_group(struct memory_group *group,
748 					   void *arg)
749 {
750 	const int ratio = READ_ONCE(auto_movable_ratio);
751 	struct auto_movable_group_stats *stats = arg;
752 	long pages;
753 
754 	/*
755 	 * We don't support modifying the config while the auto-movable online
756 	 * policy is already enabled. Just avoid the division by zero below.
757 	 */
758 	if (!ratio)
759 		return 0;
760 
761 	/*
762 	 * Calculate how many early kernel pages this group requires to
763 	 * satisfy the configured zone ratio.
764 	 */
765 	pages = group->present_movable_pages * 100 / ratio;
766 	pages -= group->present_kernel_pages;
767 
768 	if (pages > 0)
769 		stats->req_kernel_early_pages += pages;
770 	stats->movable_pages += group->present_movable_pages;
771 	return 0;
772 }
773 
774 static bool auto_movable_can_online_movable(int nid, struct memory_group *group,
775 					    unsigned long nr_pages)
776 {
777 	unsigned long kernel_early_pages, movable_pages;
778 	struct auto_movable_group_stats group_stats = {};
779 	struct auto_movable_stats stats = {};
780 	pg_data_t *pgdat = NODE_DATA(nid);
781 	struct zone *zone;
782 	int i;
783 
784 	/* Walk all relevant zones and collect MOVABLE vs. KERNEL stats. */
785 	if (nid == NUMA_NO_NODE) {
786 		/* TODO: cache values */
787 		for_each_populated_zone(zone)
788 			auto_movable_stats_account_zone(&stats, zone);
789 	} else {
790 		for (i = 0; i < MAX_NR_ZONES; i++) {
791 			zone = pgdat->node_zones + i;
792 			if (populated_zone(zone))
793 				auto_movable_stats_account_zone(&stats, zone);
794 		}
795 	}
796 
797 	kernel_early_pages = stats.kernel_early_pages;
798 	movable_pages = stats.movable_pages;
799 
800 	/*
801 	 * Kernel memory inside dynamic memory group allows for more MOVABLE
802 	 * memory within the same group. Remove the effect of all but the
803 	 * current group from the stats.
804 	 */
805 	walk_dynamic_memory_groups(nid, auto_movable_stats_account_group,
806 				   group, &group_stats);
807 	if (kernel_early_pages <= group_stats.req_kernel_early_pages)
808 		return false;
809 	kernel_early_pages -= group_stats.req_kernel_early_pages;
810 	movable_pages -= group_stats.movable_pages;
811 
812 	if (group && group->is_dynamic)
813 		kernel_early_pages += group->present_kernel_pages;
814 
815 	/*
816 	 * Test if we could online the given number of pages to ZONE_MOVABLE
817 	 * and still stay in the configured ratio.
818 	 */
819 	movable_pages += nr_pages;
820 	return movable_pages <= (auto_movable_ratio * kernel_early_pages) / 100;
821 }
822 
823 /*
824  * Returns a default kernel memory zone for the given pfn range.
825  * If no kernel zone covers this pfn range it will automatically go
826  * to the ZONE_NORMAL.
827  */
828 static struct zone *default_kernel_zone_for_pfn(int nid, unsigned long start_pfn,
829 		unsigned long nr_pages)
830 {
831 	struct pglist_data *pgdat = NODE_DATA(nid);
832 	int zid;
833 
834 	for (zid = 0; zid < ZONE_NORMAL; zid++) {
835 		struct zone *zone = &pgdat->node_zones[zid];
836 
837 		if (zone_intersects(zone, start_pfn, nr_pages))
838 			return zone;
839 	}
840 
841 	return &pgdat->node_zones[ZONE_NORMAL];
842 }
843 
844 /*
845  * Determine to which zone to online memory dynamically based on user
846  * configuration and system stats. We care about the following ratio:
847  *
848  *   MOVABLE : KERNEL
849  *
850  * Whereby MOVABLE is memory in ZONE_MOVABLE and KERNEL is memory in
851  * one of the kernel zones. CMA pages inside one of the kernel zones really
852  * behaves like ZONE_MOVABLE, so we treat them accordingly.
853  *
854  * We don't allow for hotplugged memory in a KERNEL zone to increase the
855  * amount of MOVABLE memory we can have, so we end up with:
856  *
857  *   MOVABLE : KERNEL_EARLY
858  *
859  * Whereby KERNEL_EARLY is memory in one of the kernel zones, available sinze
860  * boot. We base our calculation on KERNEL_EARLY internally, because:
861  *
862  * a) Hotplugged memory in one of the kernel zones can sometimes still get
863  *    hotunplugged, especially when hot(un)plugging individual memory blocks.
864  *    There is no coordination across memory devices, therefore "automatic"
865  *    hotunplugging, as implemented in hypervisors, could result in zone
866  *    imbalances.
867  * b) Early/boot memory in one of the kernel zones can usually not get
868  *    hotunplugged again (e.g., no firmware interface to unplug, fragmented
869  *    with unmovable allocations). While there are corner cases where it might
870  *    still work, it is barely relevant in practice.
871  *
872  * Exceptions are dynamic memory groups, which allow for more MOVABLE
873  * memory within the same memory group -- because in that case, there is
874  * coordination within the single memory device managed by a single driver.
875  *
876  * We rely on "present pages" instead of "managed pages", as the latter is
877  * highly unreliable and dynamic in virtualized environments, and does not
878  * consider boot time allocations. For example, memory ballooning adjusts the
879  * managed pages when inflating/deflating the balloon, and balloon compaction
880  * can even migrate inflated pages between zones.
881  *
882  * Using "present pages" is better but some things to keep in mind are:
883  *
884  * a) Some memblock allocations, such as for the crashkernel area, are
885  *    effectively unused by the kernel, yet they account to "present pages".
886  *    Fortunately, these allocations are comparatively small in relevant setups
887  *    (e.g., fraction of system memory).
888  * b) Some hotplugged memory blocks in virtualized environments, esecially
889  *    hotplugged by virtio-mem, look like they are completely present, however,
890  *    only parts of the memory block are actually currently usable.
891  *    "present pages" is an upper limit that can get reached at runtime. As
892  *    we base our calculations on KERNEL_EARLY, this is not an issue.
893  */
894 static struct zone *auto_movable_zone_for_pfn(int nid,
895 					      struct memory_group *group,
896 					      unsigned long pfn,
897 					      unsigned long nr_pages)
898 {
899 	unsigned long online_pages = 0, max_pages, end_pfn;
900 	struct page *page;
901 
902 	if (!auto_movable_ratio)
903 		goto kernel_zone;
904 
905 	if (group && !group->is_dynamic) {
906 		max_pages = group->s.max_pages;
907 		online_pages = group->present_movable_pages;
908 
909 		/* If anything is !MOVABLE online the rest !MOVABLE. */
910 		if (group->present_kernel_pages)
911 			goto kernel_zone;
912 	} else if (!group || group->d.unit_pages == nr_pages) {
913 		max_pages = nr_pages;
914 	} else {
915 		max_pages = group->d.unit_pages;
916 		/*
917 		 * Take a look at all online sections in the current unit.
918 		 * We can safely assume that all pages within a section belong
919 		 * to the same zone, because dynamic memory groups only deal
920 		 * with hotplugged memory.
921 		 */
922 		pfn = ALIGN_DOWN(pfn, group->d.unit_pages);
923 		end_pfn = pfn + group->d.unit_pages;
924 		for (; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
925 			page = pfn_to_online_page(pfn);
926 			if (!page)
927 				continue;
928 			/* If anything is !MOVABLE online the rest !MOVABLE. */
929 			if (!is_zone_movable_page(page))
930 				goto kernel_zone;
931 			online_pages += PAGES_PER_SECTION;
932 		}
933 	}
934 
935 	/*
936 	 * Online MOVABLE if we could *currently* online all remaining parts
937 	 * MOVABLE. We expect to (add+) online them immediately next, so if
938 	 * nobody interferes, all will be MOVABLE if possible.
939 	 */
940 	nr_pages = max_pages - online_pages;
941 	if (!auto_movable_can_online_movable(NUMA_NO_NODE, group, nr_pages))
942 		goto kernel_zone;
943 
944 #ifdef CONFIG_NUMA
945 	if (auto_movable_numa_aware &&
946 	    !auto_movable_can_online_movable(nid, group, nr_pages))
947 		goto kernel_zone;
948 #endif /* CONFIG_NUMA */
949 
950 	return &NODE_DATA(nid)->node_zones[ZONE_MOVABLE];
951 kernel_zone:
952 	return default_kernel_zone_for_pfn(nid, pfn, nr_pages);
953 }
954 
955 static inline struct zone *default_zone_for_pfn(int nid, unsigned long start_pfn,
956 		unsigned long nr_pages)
957 {
958 	struct zone *kernel_zone = default_kernel_zone_for_pfn(nid, start_pfn,
959 			nr_pages);
960 	struct zone *movable_zone = &NODE_DATA(nid)->node_zones[ZONE_MOVABLE];
961 	bool in_kernel = zone_intersects(kernel_zone, start_pfn, nr_pages);
962 	bool in_movable = zone_intersects(movable_zone, start_pfn, nr_pages);
963 
964 	/*
965 	 * We inherit the existing zone in a simple case where zones do not
966 	 * overlap in the given range
967 	 */
968 	if (in_kernel ^ in_movable)
969 		return (in_kernel) ? kernel_zone : movable_zone;
970 
971 	/*
972 	 * If the range doesn't belong to any zone or two zones overlap in the
973 	 * given range then we use movable zone only if movable_node is
974 	 * enabled because we always online to a kernel zone by default.
975 	 */
976 	return movable_node_enabled ? movable_zone : kernel_zone;
977 }
978 
979 struct zone *zone_for_pfn_range(int online_type, int nid,
980 		struct memory_group *group, unsigned long start_pfn,
981 		unsigned long nr_pages)
982 {
983 	if (online_type == MMOP_ONLINE_KERNEL)
984 		return default_kernel_zone_for_pfn(nid, start_pfn, nr_pages);
985 
986 	if (online_type == MMOP_ONLINE_MOVABLE)
987 		return &NODE_DATA(nid)->node_zones[ZONE_MOVABLE];
988 
989 	if (online_policy == ONLINE_POLICY_AUTO_MOVABLE)
990 		return auto_movable_zone_for_pfn(nid, group, start_pfn, nr_pages);
991 
992 	return default_zone_for_pfn(nid, start_pfn, nr_pages);
993 }
994 
995 /*
996  * This function should only be called by memory_block_{online,offline},
997  * and {online,offline}_pages.
998  */
999 void adjust_present_page_count(struct page *page, struct memory_group *group,
1000 			       long nr_pages)
1001 {
1002 	struct zone *zone = page_zone(page);
1003 	const bool movable = zone_idx(zone) == ZONE_MOVABLE;
1004 
1005 	/*
1006 	 * We only support onlining/offlining/adding/removing of complete
1007 	 * memory blocks; therefore, either all is either early or hotplugged.
1008 	 */
1009 	if (early_section(__pfn_to_section(page_to_pfn(page))))
1010 		zone->present_early_pages += nr_pages;
1011 	zone->present_pages += nr_pages;
1012 	zone->zone_pgdat->node_present_pages += nr_pages;
1013 
1014 	if (group && movable)
1015 		group->present_movable_pages += nr_pages;
1016 	else if (group && !movable)
1017 		group->present_kernel_pages += nr_pages;
1018 }
1019 
1020 int mhp_init_memmap_on_memory(unsigned long pfn, unsigned long nr_pages,
1021 			      struct zone *zone)
1022 {
1023 	unsigned long end_pfn = pfn + nr_pages;
1024 	int ret, i;
1025 
1026 	ret = kasan_add_zero_shadow(__va(PFN_PHYS(pfn)), PFN_PHYS(nr_pages));
1027 	if (ret)
1028 		return ret;
1029 
1030 	move_pfn_range_to_zone(zone, pfn, nr_pages, NULL, MIGRATE_UNMOVABLE);
1031 
1032 	for (i = 0; i < nr_pages; i++)
1033 		SetPageVmemmapSelfHosted(pfn_to_page(pfn + i));
1034 
1035 	/*
1036 	 * It might be that the vmemmap_pages fully span sections. If that is
1037 	 * the case, mark those sections online here as otherwise they will be
1038 	 * left offline.
1039 	 */
1040 	if (nr_pages >= PAGES_PER_SECTION)
1041 	        online_mem_sections(pfn, ALIGN_DOWN(end_pfn, PAGES_PER_SECTION));
1042 
1043 	return ret;
1044 }
1045 
1046 void mhp_deinit_memmap_on_memory(unsigned long pfn, unsigned long nr_pages)
1047 {
1048 	unsigned long end_pfn = pfn + nr_pages;
1049 
1050 	/*
1051 	 * It might be that the vmemmap_pages fully span sections. If that is
1052 	 * the case, mark those sections offline here as otherwise they will be
1053 	 * left online.
1054 	 */
1055 	if (nr_pages >= PAGES_PER_SECTION)
1056 		offline_mem_sections(pfn, ALIGN_DOWN(end_pfn, PAGES_PER_SECTION));
1057 
1058         /*
1059 	 * The pages associated with this vmemmap have been offlined, so
1060 	 * we can reset its state here.
1061 	 */
1062 	remove_pfn_range_from_zone(page_zone(pfn_to_page(pfn)), pfn, nr_pages);
1063 	kasan_remove_zero_shadow(__va(PFN_PHYS(pfn)), PFN_PHYS(nr_pages));
1064 }
1065 
1066 int __ref online_pages(unsigned long pfn, unsigned long nr_pages,
1067 		       struct zone *zone, struct memory_group *group)
1068 {
1069 	unsigned long flags;
1070 	int need_zonelists_rebuild = 0;
1071 	const int nid = zone_to_nid(zone);
1072 	int ret;
1073 	struct memory_notify arg;
1074 
1075 	/*
1076 	 * {on,off}lining is constrained to full memory sections (or more
1077 	 * precisely to memory blocks from the user space POV).
1078 	 * memmap_on_memory is an exception because it reserves initial part
1079 	 * of the physical memory space for vmemmaps. That space is pageblock
1080 	 * aligned.
1081 	 */
1082 	if (WARN_ON_ONCE(!nr_pages || !pageblock_aligned(pfn) ||
1083 			 !IS_ALIGNED(pfn + nr_pages, PAGES_PER_SECTION)))
1084 		return -EINVAL;
1085 
1086 	mem_hotplug_begin();
1087 
1088 	/* associate pfn range with the zone */
1089 	move_pfn_range_to_zone(zone, pfn, nr_pages, NULL, MIGRATE_ISOLATE);
1090 
1091 	arg.start_pfn = pfn;
1092 	arg.nr_pages = nr_pages;
1093 	node_states_check_changes_online(nr_pages, zone, &arg);
1094 
1095 	ret = memory_notify(MEM_GOING_ONLINE, &arg);
1096 	ret = notifier_to_errno(ret);
1097 	if (ret)
1098 		goto failed_addition;
1099 
1100 	/*
1101 	 * Fixup the number of isolated pageblocks before marking the sections
1102 	 * onlining, such that undo_isolate_page_range() works correctly.
1103 	 */
1104 	spin_lock_irqsave(&zone->lock, flags);
1105 	zone->nr_isolate_pageblock += nr_pages / pageblock_nr_pages;
1106 	spin_unlock_irqrestore(&zone->lock, flags);
1107 
1108 	/*
1109 	 * If this zone is not populated, then it is not in zonelist.
1110 	 * This means the page allocator ignores this zone.
1111 	 * So, zonelist must be updated after online.
1112 	 */
1113 	if (!populated_zone(zone)) {
1114 		need_zonelists_rebuild = 1;
1115 		setup_zone_pageset(zone);
1116 	}
1117 
1118 	online_pages_range(pfn, nr_pages);
1119 	adjust_present_page_count(pfn_to_page(pfn), group, nr_pages);
1120 
1121 	node_states_set_node(nid, &arg);
1122 	if (need_zonelists_rebuild)
1123 		build_all_zonelists(NULL);
1124 
1125 	/* Basic onlining is complete, allow allocation of onlined pages. */
1126 	undo_isolate_page_range(pfn, pfn + nr_pages, MIGRATE_MOVABLE);
1127 
1128 	/*
1129 	 * Freshly onlined pages aren't shuffled (e.g., all pages are placed to
1130 	 * the tail of the freelist when undoing isolation). Shuffle the whole
1131 	 * zone to make sure the just onlined pages are properly distributed
1132 	 * across the whole freelist - to create an initial shuffle.
1133 	 */
1134 	shuffle_zone(zone);
1135 
1136 	/* reinitialise watermarks and update pcp limits */
1137 	init_per_zone_wmark_min();
1138 
1139 	kswapd_run(nid);
1140 	kcompactd_run(nid);
1141 
1142 	writeback_set_ratelimit();
1143 
1144 	memory_notify(MEM_ONLINE, &arg);
1145 	mem_hotplug_done();
1146 	return 0;
1147 
1148 failed_addition:
1149 	pr_debug("online_pages [mem %#010llx-%#010llx] failed\n",
1150 		 (unsigned long long) pfn << PAGE_SHIFT,
1151 		 (((unsigned long long) pfn + nr_pages) << PAGE_SHIFT) - 1);
1152 	memory_notify(MEM_CANCEL_ONLINE, &arg);
1153 	remove_pfn_range_from_zone(zone, pfn, nr_pages);
1154 	mem_hotplug_done();
1155 	return ret;
1156 }
1157 
1158 /* we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG */
1159 static pg_data_t __ref *hotadd_init_pgdat(int nid)
1160 {
1161 	struct pglist_data *pgdat;
1162 
1163 	/*
1164 	 * NODE_DATA is preallocated (free_area_init) but its internal
1165 	 * state is not allocated completely. Add missing pieces.
1166 	 * Completely offline nodes stay around and they just need
1167 	 * reintialization.
1168 	 */
1169 	pgdat = NODE_DATA(nid);
1170 
1171 	/* init node's zones as empty zones, we don't have any present pages.*/
1172 	free_area_init_core_hotplug(pgdat);
1173 
1174 	/*
1175 	 * The node we allocated has no zone fallback lists. For avoiding
1176 	 * to access not-initialized zonelist, build here.
1177 	 */
1178 	build_all_zonelists(pgdat);
1179 
1180 	return pgdat;
1181 }
1182 
1183 /*
1184  * __try_online_node - online a node if offlined
1185  * @nid: the node ID
1186  * @set_node_online: Whether we want to online the node
1187  * called by cpu_up() to online a node without onlined memory.
1188  *
1189  * Returns:
1190  * 1 -> a new node has been allocated
1191  * 0 -> the node is already online
1192  * -ENOMEM -> the node could not be allocated
1193  */
1194 static int __try_online_node(int nid, bool set_node_online)
1195 {
1196 	pg_data_t *pgdat;
1197 	int ret = 1;
1198 
1199 	if (node_online(nid))
1200 		return 0;
1201 
1202 	pgdat = hotadd_init_pgdat(nid);
1203 	if (!pgdat) {
1204 		pr_err("Cannot online node %d due to NULL pgdat\n", nid);
1205 		ret = -ENOMEM;
1206 		goto out;
1207 	}
1208 
1209 	if (set_node_online) {
1210 		node_set_online(nid);
1211 		ret = register_one_node(nid);
1212 		BUG_ON(ret);
1213 	}
1214 out:
1215 	return ret;
1216 }
1217 
1218 /*
1219  * Users of this function always want to online/register the node
1220  */
1221 int try_online_node(int nid)
1222 {
1223 	int ret;
1224 
1225 	mem_hotplug_begin();
1226 	ret =  __try_online_node(nid, true);
1227 	mem_hotplug_done();
1228 	return ret;
1229 }
1230 
1231 static int check_hotplug_memory_range(u64 start, u64 size)
1232 {
1233 	/* memory range must be block size aligned */
1234 	if (!size || !IS_ALIGNED(start, memory_block_size_bytes()) ||
1235 	    !IS_ALIGNED(size, memory_block_size_bytes())) {
1236 		pr_err("Block size [%#lx] unaligned hotplug range: start %#llx, size %#llx",
1237 		       memory_block_size_bytes(), start, size);
1238 		return -EINVAL;
1239 	}
1240 
1241 	return 0;
1242 }
1243 
1244 static int online_memory_block(struct memory_block *mem, void *arg)
1245 {
1246 	mem->online_type = mhp_default_online_type;
1247 	return device_online(&mem->dev);
1248 }
1249 
1250 bool mhp_supports_memmap_on_memory(unsigned long size)
1251 {
1252 	unsigned long nr_vmemmap_pages = size / PAGE_SIZE;
1253 	unsigned long vmemmap_size = nr_vmemmap_pages * sizeof(struct page);
1254 	unsigned long remaining_size = size - vmemmap_size;
1255 
1256 	/*
1257 	 * Besides having arch support and the feature enabled at runtime, we
1258 	 * need a few more assumptions to hold true:
1259 	 *
1260 	 * a) We span a single memory block: memory onlining/offlinin;g happens
1261 	 *    in memory block granularity. We don't want the vmemmap of online
1262 	 *    memory blocks to reside on offline memory blocks. In the future,
1263 	 *    we might want to support variable-sized memory blocks to make the
1264 	 *    feature more versatile.
1265 	 *
1266 	 * b) The vmemmap pages span complete PMDs: We don't want vmemmap code
1267 	 *    to populate memory from the altmap for unrelated parts (i.e.,
1268 	 *    other memory blocks)
1269 	 *
1270 	 * c) The vmemmap pages (and thereby the pages that will be exposed to
1271 	 *    the buddy) have to cover full pageblocks: memory onlining/offlining
1272 	 *    code requires applicable ranges to be page-aligned, for example, to
1273 	 *    set the migratetypes properly.
1274 	 *
1275 	 * TODO: Although we have a check here to make sure that vmemmap pages
1276 	 *       fully populate a PMD, it is not the right place to check for
1277 	 *       this. A much better solution involves improving vmemmap code
1278 	 *       to fallback to base pages when trying to populate vmemmap using
1279 	 *       altmap as an alternative source of memory, and we do not exactly
1280 	 *       populate a single PMD.
1281 	 */
1282 	return mhp_memmap_on_memory() &&
1283 	       size == memory_block_size_bytes() &&
1284 	       IS_ALIGNED(vmemmap_size, PMD_SIZE) &&
1285 	       IS_ALIGNED(remaining_size, (pageblock_nr_pages << PAGE_SHIFT));
1286 }
1287 
1288 /*
1289  * NOTE: The caller must call lock_device_hotplug() to serialize hotplug
1290  * and online/offline operations (triggered e.g. by sysfs).
1291  *
1292  * we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG
1293  */
1294 int __ref add_memory_resource(int nid, struct resource *res, mhp_t mhp_flags)
1295 {
1296 	struct mhp_params params = { .pgprot = pgprot_mhp(PAGE_KERNEL) };
1297 	enum memblock_flags memblock_flags = MEMBLOCK_NONE;
1298 	struct vmem_altmap mhp_altmap = {};
1299 	struct memory_group *group = NULL;
1300 	u64 start, size;
1301 	bool new_node = false;
1302 	int ret;
1303 
1304 	start = res->start;
1305 	size = resource_size(res);
1306 
1307 	ret = check_hotplug_memory_range(start, size);
1308 	if (ret)
1309 		return ret;
1310 
1311 	if (mhp_flags & MHP_NID_IS_MGID) {
1312 		group = memory_group_find_by_id(nid);
1313 		if (!group)
1314 			return -EINVAL;
1315 		nid = group->nid;
1316 	}
1317 
1318 	if (!node_possible(nid)) {
1319 		WARN(1, "node %d was absent from the node_possible_map\n", nid);
1320 		return -EINVAL;
1321 	}
1322 
1323 	mem_hotplug_begin();
1324 
1325 	if (IS_ENABLED(CONFIG_ARCH_KEEP_MEMBLOCK)) {
1326 		if (res->flags & IORESOURCE_SYSRAM_DRIVER_MANAGED)
1327 			memblock_flags = MEMBLOCK_DRIVER_MANAGED;
1328 		ret = memblock_add_node(start, size, nid, memblock_flags);
1329 		if (ret)
1330 			goto error_mem_hotplug_end;
1331 	}
1332 
1333 	ret = __try_online_node(nid, false);
1334 	if (ret < 0)
1335 		goto error;
1336 	new_node = ret;
1337 
1338 	/*
1339 	 * Self hosted memmap array
1340 	 */
1341 	if (mhp_flags & MHP_MEMMAP_ON_MEMORY) {
1342 		if (!mhp_supports_memmap_on_memory(size)) {
1343 			ret = -EINVAL;
1344 			goto error;
1345 		}
1346 		mhp_altmap.free = PHYS_PFN(size);
1347 		mhp_altmap.base_pfn = PHYS_PFN(start);
1348 		params.altmap = &mhp_altmap;
1349 	}
1350 
1351 	/* call arch's memory hotadd */
1352 	ret = arch_add_memory(nid, start, size, &params);
1353 	if (ret < 0)
1354 		goto error;
1355 
1356 	/* create memory block devices after memory was added */
1357 	ret = create_memory_block_devices(start, size, mhp_altmap.alloc,
1358 					  group);
1359 	if (ret) {
1360 		arch_remove_memory(start, size, NULL);
1361 		goto error;
1362 	}
1363 
1364 	if (new_node) {
1365 		/* If sysfs file of new node can't be created, cpu on the node
1366 		 * can't be hot-added. There is no rollback way now.
1367 		 * So, check by BUG_ON() to catch it reluctantly..
1368 		 * We online node here. We can't roll back from here.
1369 		 */
1370 		node_set_online(nid);
1371 		ret = __register_one_node(nid);
1372 		BUG_ON(ret);
1373 	}
1374 
1375 	register_memory_blocks_under_node(nid, PFN_DOWN(start),
1376 					  PFN_UP(start + size - 1),
1377 					  MEMINIT_HOTPLUG);
1378 
1379 	/* create new memmap entry */
1380 	if (!strcmp(res->name, "System RAM"))
1381 		firmware_map_add_hotplug(start, start + size, "System RAM");
1382 
1383 	/* device_online() will take the lock when calling online_pages() */
1384 	mem_hotplug_done();
1385 
1386 	/*
1387 	 * In case we're allowed to merge the resource, flag it and trigger
1388 	 * merging now that adding succeeded.
1389 	 */
1390 	if (mhp_flags & MHP_MERGE_RESOURCE)
1391 		merge_system_ram_resource(res);
1392 
1393 	/* online pages if requested */
1394 	if (mhp_default_online_type != MMOP_OFFLINE)
1395 		walk_memory_blocks(start, size, NULL, online_memory_block);
1396 
1397 	return ret;
1398 error:
1399 	if (IS_ENABLED(CONFIG_ARCH_KEEP_MEMBLOCK))
1400 		memblock_remove(start, size);
1401 error_mem_hotplug_end:
1402 	mem_hotplug_done();
1403 	return ret;
1404 }
1405 
1406 /* requires device_hotplug_lock, see add_memory_resource() */
1407 int __ref __add_memory(int nid, u64 start, u64 size, mhp_t mhp_flags)
1408 {
1409 	struct resource *res;
1410 	int ret;
1411 
1412 	res = register_memory_resource(start, size, "System RAM");
1413 	if (IS_ERR(res))
1414 		return PTR_ERR(res);
1415 
1416 	ret = add_memory_resource(nid, res, mhp_flags);
1417 	if (ret < 0)
1418 		release_memory_resource(res);
1419 	return ret;
1420 }
1421 
1422 int add_memory(int nid, u64 start, u64 size, mhp_t mhp_flags)
1423 {
1424 	int rc;
1425 
1426 	lock_device_hotplug();
1427 	rc = __add_memory(nid, start, size, mhp_flags);
1428 	unlock_device_hotplug();
1429 
1430 	return rc;
1431 }
1432 EXPORT_SYMBOL_GPL(add_memory);
1433 
1434 /*
1435  * Add special, driver-managed memory to the system as system RAM. Such
1436  * memory is not exposed via the raw firmware-provided memmap as system
1437  * RAM, instead, it is detected and added by a driver - during cold boot,
1438  * after a reboot, and after kexec.
1439  *
1440  * Reasons why this memory should not be used for the initial memmap of a
1441  * kexec kernel or for placing kexec images:
1442  * - The booting kernel is in charge of determining how this memory will be
1443  *   used (e.g., use persistent memory as system RAM)
1444  * - Coordination with a hypervisor is required before this memory
1445  *   can be used (e.g., inaccessible parts).
1446  *
1447  * For this memory, no entries in /sys/firmware/memmap ("raw firmware-provided
1448  * memory map") are created. Also, the created memory resource is flagged
1449  * with IORESOURCE_SYSRAM_DRIVER_MANAGED, so in-kernel users can special-case
1450  * this memory as well (esp., not place kexec images onto it).
1451  *
1452  * The resource_name (visible via /proc/iomem) has to have the format
1453  * "System RAM ($DRIVER)".
1454  */
1455 int add_memory_driver_managed(int nid, u64 start, u64 size,
1456 			      const char *resource_name, mhp_t mhp_flags)
1457 {
1458 	struct resource *res;
1459 	int rc;
1460 
1461 	if (!resource_name ||
1462 	    strstr(resource_name, "System RAM (") != resource_name ||
1463 	    resource_name[strlen(resource_name) - 1] != ')')
1464 		return -EINVAL;
1465 
1466 	lock_device_hotplug();
1467 
1468 	res = register_memory_resource(start, size, resource_name);
1469 	if (IS_ERR(res)) {
1470 		rc = PTR_ERR(res);
1471 		goto out_unlock;
1472 	}
1473 
1474 	rc = add_memory_resource(nid, res, mhp_flags);
1475 	if (rc < 0)
1476 		release_memory_resource(res);
1477 
1478 out_unlock:
1479 	unlock_device_hotplug();
1480 	return rc;
1481 }
1482 EXPORT_SYMBOL_GPL(add_memory_driver_managed);
1483 
1484 /*
1485  * Platforms should define arch_get_mappable_range() that provides
1486  * maximum possible addressable physical memory range for which the
1487  * linear mapping could be created. The platform returned address
1488  * range must adhere to these following semantics.
1489  *
1490  * - range.start <= range.end
1491  * - Range includes both end points [range.start..range.end]
1492  *
1493  * There is also a fallback definition provided here, allowing the
1494  * entire possible physical address range in case any platform does
1495  * not define arch_get_mappable_range().
1496  */
1497 struct range __weak arch_get_mappable_range(void)
1498 {
1499 	struct range mhp_range = {
1500 		.start = 0UL,
1501 		.end = -1ULL,
1502 	};
1503 	return mhp_range;
1504 }
1505 
1506 struct range mhp_get_pluggable_range(bool need_mapping)
1507 {
1508 	const u64 max_phys = (1ULL << MAX_PHYSMEM_BITS) - 1;
1509 	struct range mhp_range;
1510 
1511 	if (need_mapping) {
1512 		mhp_range = arch_get_mappable_range();
1513 		if (mhp_range.start > max_phys) {
1514 			mhp_range.start = 0;
1515 			mhp_range.end = 0;
1516 		}
1517 		mhp_range.end = min_t(u64, mhp_range.end, max_phys);
1518 	} else {
1519 		mhp_range.start = 0;
1520 		mhp_range.end = max_phys;
1521 	}
1522 	return mhp_range;
1523 }
1524 EXPORT_SYMBOL_GPL(mhp_get_pluggable_range);
1525 
1526 bool mhp_range_allowed(u64 start, u64 size, bool need_mapping)
1527 {
1528 	struct range mhp_range = mhp_get_pluggable_range(need_mapping);
1529 	u64 end = start + size;
1530 
1531 	if (start < end && start >= mhp_range.start && (end - 1) <= mhp_range.end)
1532 		return true;
1533 
1534 	pr_warn("Hotplug memory [%#llx-%#llx] exceeds maximum addressable range [%#llx-%#llx]\n",
1535 		start, end, mhp_range.start, mhp_range.end);
1536 	return false;
1537 }
1538 
1539 #ifdef CONFIG_MEMORY_HOTREMOVE
1540 /*
1541  * Scan pfn range [start,end) to find movable/migratable pages (LRU pages,
1542  * non-lru movable pages and hugepages). Will skip over most unmovable
1543  * pages (esp., pages that can be skipped when offlining), but bail out on
1544  * definitely unmovable pages.
1545  *
1546  * Returns:
1547  *	0 in case a movable page is found and movable_pfn was updated.
1548  *	-ENOENT in case no movable page was found.
1549  *	-EBUSY in case a definitely unmovable page was found.
1550  */
1551 static int scan_movable_pages(unsigned long start, unsigned long end,
1552 			      unsigned long *movable_pfn)
1553 {
1554 	unsigned long pfn;
1555 
1556 	for (pfn = start; pfn < end; pfn++) {
1557 		struct page *page, *head;
1558 		unsigned long skip;
1559 
1560 		if (!pfn_valid(pfn))
1561 			continue;
1562 		page = pfn_to_page(pfn);
1563 		if (PageLRU(page))
1564 			goto found;
1565 		if (__PageMovable(page))
1566 			goto found;
1567 
1568 		/*
1569 		 * PageOffline() pages that are not marked __PageMovable() and
1570 		 * have a reference count > 0 (after MEM_GOING_OFFLINE) are
1571 		 * definitely unmovable. If their reference count would be 0,
1572 		 * they could at least be skipped when offlining memory.
1573 		 */
1574 		if (PageOffline(page) && page_count(page))
1575 			return -EBUSY;
1576 
1577 		if (!PageHuge(page))
1578 			continue;
1579 		head = compound_head(page);
1580 		/*
1581 		 * This test is racy as we hold no reference or lock.  The
1582 		 * hugetlb page could have been free'ed and head is no longer
1583 		 * a hugetlb page before the following check.  In such unlikely
1584 		 * cases false positives and negatives are possible.  Calling
1585 		 * code must deal with these scenarios.
1586 		 */
1587 		if (HPageMigratable(head))
1588 			goto found;
1589 		skip = compound_nr(head) - (page - head);
1590 		pfn += skip - 1;
1591 	}
1592 	return -ENOENT;
1593 found:
1594 	*movable_pfn = pfn;
1595 	return 0;
1596 }
1597 
1598 static void do_migrate_range(unsigned long start_pfn, unsigned long end_pfn)
1599 {
1600 	unsigned long pfn;
1601 	struct page *page, *head;
1602 	LIST_HEAD(source);
1603 	static DEFINE_RATELIMIT_STATE(migrate_rs, DEFAULT_RATELIMIT_INTERVAL,
1604 				      DEFAULT_RATELIMIT_BURST);
1605 
1606 	for (pfn = start_pfn; pfn < end_pfn; pfn++) {
1607 		struct folio *folio;
1608 		bool isolated;
1609 
1610 		if (!pfn_valid(pfn))
1611 			continue;
1612 		page = pfn_to_page(pfn);
1613 		folio = page_folio(page);
1614 		head = &folio->page;
1615 
1616 		if (PageHuge(page)) {
1617 			pfn = page_to_pfn(head) + compound_nr(head) - 1;
1618 			isolate_hugetlb(folio, &source);
1619 			continue;
1620 		} else if (PageTransHuge(page))
1621 			pfn = page_to_pfn(head) + thp_nr_pages(page) - 1;
1622 
1623 		/*
1624 		 * HWPoison pages have elevated reference counts so the migration would
1625 		 * fail on them. It also doesn't make any sense to migrate them in the
1626 		 * first place. Still try to unmap such a page in case it is still mapped
1627 		 * (e.g. current hwpoison implementation doesn't unmap KSM pages but keep
1628 		 * the unmap as the catch all safety net).
1629 		 */
1630 		if (PageHWPoison(page)) {
1631 			if (WARN_ON(folio_test_lru(folio)))
1632 				folio_isolate_lru(folio);
1633 			if (folio_mapped(folio))
1634 				try_to_unmap(folio, TTU_IGNORE_MLOCK);
1635 			continue;
1636 		}
1637 
1638 		if (!get_page_unless_zero(page))
1639 			continue;
1640 		/*
1641 		 * We can skip free pages. And we can deal with pages on
1642 		 * LRU and non-lru movable pages.
1643 		 */
1644 		if (PageLRU(page))
1645 			isolated = isolate_lru_page(page);
1646 		else
1647 			isolated = isolate_movable_page(page, ISOLATE_UNEVICTABLE);
1648 		if (isolated) {
1649 			list_add_tail(&page->lru, &source);
1650 			if (!__PageMovable(page))
1651 				inc_node_page_state(page, NR_ISOLATED_ANON +
1652 						    page_is_file_lru(page));
1653 
1654 		} else {
1655 			if (__ratelimit(&migrate_rs)) {
1656 				pr_warn("failed to isolate pfn %lx\n", pfn);
1657 				dump_page(page, "isolation failed");
1658 			}
1659 		}
1660 		put_page(page);
1661 	}
1662 	if (!list_empty(&source)) {
1663 		nodemask_t nmask = node_states[N_MEMORY];
1664 		struct migration_target_control mtc = {
1665 			.nmask = &nmask,
1666 			.gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
1667 		};
1668 		int ret;
1669 
1670 		/*
1671 		 * We have checked that migration range is on a single zone so
1672 		 * we can use the nid of the first page to all the others.
1673 		 */
1674 		mtc.nid = page_to_nid(list_first_entry(&source, struct page, lru));
1675 
1676 		/*
1677 		 * try to allocate from a different node but reuse this node
1678 		 * if there are no other online nodes to be used (e.g. we are
1679 		 * offlining a part of the only existing node)
1680 		 */
1681 		node_clear(mtc.nid, nmask);
1682 		if (nodes_empty(nmask))
1683 			node_set(mtc.nid, nmask);
1684 		ret = migrate_pages(&source, alloc_migration_target, NULL,
1685 			(unsigned long)&mtc, MIGRATE_SYNC, MR_MEMORY_HOTPLUG, NULL);
1686 		if (ret) {
1687 			list_for_each_entry(page, &source, lru) {
1688 				if (__ratelimit(&migrate_rs)) {
1689 					pr_warn("migrating pfn %lx failed ret:%d\n",
1690 						page_to_pfn(page), ret);
1691 					dump_page(page, "migration failure");
1692 				}
1693 			}
1694 			putback_movable_pages(&source);
1695 		}
1696 	}
1697 }
1698 
1699 static int __init cmdline_parse_movable_node(char *p)
1700 {
1701 	movable_node_enabled = true;
1702 	return 0;
1703 }
1704 early_param("movable_node", cmdline_parse_movable_node);
1705 
1706 /* check which state of node_states will be changed when offline memory */
1707 static void node_states_check_changes_offline(unsigned long nr_pages,
1708 		struct zone *zone, struct memory_notify *arg)
1709 {
1710 	struct pglist_data *pgdat = zone->zone_pgdat;
1711 	unsigned long present_pages = 0;
1712 	enum zone_type zt;
1713 
1714 	arg->status_change_nid = NUMA_NO_NODE;
1715 	arg->status_change_nid_normal = NUMA_NO_NODE;
1716 
1717 	/*
1718 	 * Check whether node_states[N_NORMAL_MEMORY] will be changed.
1719 	 * If the memory to be offline is within the range
1720 	 * [0..ZONE_NORMAL], and it is the last present memory there,
1721 	 * the zones in that range will become empty after the offlining,
1722 	 * thus we can determine that we need to clear the node from
1723 	 * node_states[N_NORMAL_MEMORY].
1724 	 */
1725 	for (zt = 0; zt <= ZONE_NORMAL; zt++)
1726 		present_pages += pgdat->node_zones[zt].present_pages;
1727 	if (zone_idx(zone) <= ZONE_NORMAL && nr_pages >= present_pages)
1728 		arg->status_change_nid_normal = zone_to_nid(zone);
1729 
1730 	/*
1731 	 * We have accounted the pages from [0..ZONE_NORMAL); ZONE_HIGHMEM
1732 	 * does not apply as we don't support 32bit.
1733 	 * Here we count the possible pages from ZONE_MOVABLE.
1734 	 * If after having accounted all the pages, we see that the nr_pages
1735 	 * to be offlined is over or equal to the accounted pages,
1736 	 * we know that the node will become empty, and so, we can clear
1737 	 * it for N_MEMORY as well.
1738 	 */
1739 	present_pages += pgdat->node_zones[ZONE_MOVABLE].present_pages;
1740 
1741 	if (nr_pages >= present_pages)
1742 		arg->status_change_nid = zone_to_nid(zone);
1743 }
1744 
1745 static void node_states_clear_node(int node, struct memory_notify *arg)
1746 {
1747 	if (arg->status_change_nid_normal >= 0)
1748 		node_clear_state(node, N_NORMAL_MEMORY);
1749 
1750 	if (arg->status_change_nid >= 0)
1751 		node_clear_state(node, N_MEMORY);
1752 }
1753 
1754 static int count_system_ram_pages_cb(unsigned long start_pfn,
1755 				     unsigned long nr_pages, void *data)
1756 {
1757 	unsigned long *nr_system_ram_pages = data;
1758 
1759 	*nr_system_ram_pages += nr_pages;
1760 	return 0;
1761 }
1762 
1763 int __ref offline_pages(unsigned long start_pfn, unsigned long nr_pages,
1764 			struct zone *zone, struct memory_group *group)
1765 {
1766 	const unsigned long end_pfn = start_pfn + nr_pages;
1767 	unsigned long pfn, system_ram_pages = 0;
1768 	const int node = zone_to_nid(zone);
1769 	unsigned long flags;
1770 	struct memory_notify arg;
1771 	char *reason;
1772 	int ret;
1773 
1774 	/*
1775 	 * {on,off}lining is constrained to full memory sections (or more
1776 	 * precisely to memory blocks from the user space POV).
1777 	 * memmap_on_memory is an exception because it reserves initial part
1778 	 * of the physical memory space for vmemmaps. That space is pageblock
1779 	 * aligned.
1780 	 */
1781 	if (WARN_ON_ONCE(!nr_pages || !pageblock_aligned(start_pfn) ||
1782 			 !IS_ALIGNED(start_pfn + nr_pages, PAGES_PER_SECTION)))
1783 		return -EINVAL;
1784 
1785 	mem_hotplug_begin();
1786 
1787 	/*
1788 	 * Don't allow to offline memory blocks that contain holes.
1789 	 * Consequently, memory blocks with holes can never get onlined
1790 	 * via the hotplug path - online_pages() - as hotplugged memory has
1791 	 * no holes. This way, we e.g., don't have to worry about marking
1792 	 * memory holes PG_reserved, don't need pfn_valid() checks, and can
1793 	 * avoid using walk_system_ram_range() later.
1794 	 */
1795 	walk_system_ram_range(start_pfn, nr_pages, &system_ram_pages,
1796 			      count_system_ram_pages_cb);
1797 	if (system_ram_pages != nr_pages) {
1798 		ret = -EINVAL;
1799 		reason = "memory holes";
1800 		goto failed_removal;
1801 	}
1802 
1803 	/*
1804 	 * We only support offlining of memory blocks managed by a single zone,
1805 	 * checked by calling code. This is just a sanity check that we might
1806 	 * want to remove in the future.
1807 	 */
1808 	if (WARN_ON_ONCE(page_zone(pfn_to_page(start_pfn)) != zone ||
1809 			 page_zone(pfn_to_page(end_pfn - 1)) != zone)) {
1810 		ret = -EINVAL;
1811 		reason = "multizone range";
1812 		goto failed_removal;
1813 	}
1814 
1815 	/*
1816 	 * Disable pcplists so that page isolation cannot race with freeing
1817 	 * in a way that pages from isolated pageblock are left on pcplists.
1818 	 */
1819 	zone_pcp_disable(zone);
1820 	lru_cache_disable();
1821 
1822 	/* set above range as isolated */
1823 	ret = start_isolate_page_range(start_pfn, end_pfn,
1824 				       MIGRATE_MOVABLE,
1825 				       MEMORY_OFFLINE | REPORT_FAILURE,
1826 				       GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL);
1827 	if (ret) {
1828 		reason = "failure to isolate range";
1829 		goto failed_removal_pcplists_disabled;
1830 	}
1831 
1832 	arg.start_pfn = start_pfn;
1833 	arg.nr_pages = nr_pages;
1834 	node_states_check_changes_offline(nr_pages, zone, &arg);
1835 
1836 	ret = memory_notify(MEM_GOING_OFFLINE, &arg);
1837 	ret = notifier_to_errno(ret);
1838 	if (ret) {
1839 		reason = "notifier failure";
1840 		goto failed_removal_isolated;
1841 	}
1842 
1843 	do {
1844 		pfn = start_pfn;
1845 		do {
1846 			if (signal_pending(current)) {
1847 				ret = -EINTR;
1848 				reason = "signal backoff";
1849 				goto failed_removal_isolated;
1850 			}
1851 
1852 			cond_resched();
1853 
1854 			ret = scan_movable_pages(pfn, end_pfn, &pfn);
1855 			if (!ret) {
1856 				/*
1857 				 * TODO: fatal migration failures should bail
1858 				 * out
1859 				 */
1860 				do_migrate_range(pfn, end_pfn);
1861 			}
1862 		} while (!ret);
1863 
1864 		if (ret != -ENOENT) {
1865 			reason = "unmovable page";
1866 			goto failed_removal_isolated;
1867 		}
1868 
1869 		/*
1870 		 * Dissolve free hugepages in the memory block before doing
1871 		 * offlining actually in order to make hugetlbfs's object
1872 		 * counting consistent.
1873 		 */
1874 		ret = dissolve_free_huge_pages(start_pfn, end_pfn);
1875 		if (ret) {
1876 			reason = "failure to dissolve huge pages";
1877 			goto failed_removal_isolated;
1878 		}
1879 
1880 		ret = test_pages_isolated(start_pfn, end_pfn, MEMORY_OFFLINE);
1881 
1882 	} while (ret);
1883 
1884 	/* Mark all sections offline and remove free pages from the buddy. */
1885 	__offline_isolated_pages(start_pfn, end_pfn);
1886 	pr_debug("Offlined Pages %ld\n", nr_pages);
1887 
1888 	/*
1889 	 * The memory sections are marked offline, and the pageblock flags
1890 	 * effectively stale; nobody should be touching them. Fixup the number
1891 	 * of isolated pageblocks, memory onlining will properly revert this.
1892 	 */
1893 	spin_lock_irqsave(&zone->lock, flags);
1894 	zone->nr_isolate_pageblock -= nr_pages / pageblock_nr_pages;
1895 	spin_unlock_irqrestore(&zone->lock, flags);
1896 
1897 	lru_cache_enable();
1898 	zone_pcp_enable(zone);
1899 
1900 	/* removal success */
1901 	adjust_managed_page_count(pfn_to_page(start_pfn), -nr_pages);
1902 	adjust_present_page_count(pfn_to_page(start_pfn), group, -nr_pages);
1903 
1904 	/* reinitialise watermarks and update pcp limits */
1905 	init_per_zone_wmark_min();
1906 
1907 	if (!populated_zone(zone)) {
1908 		zone_pcp_reset(zone);
1909 		build_all_zonelists(NULL);
1910 	}
1911 
1912 	node_states_clear_node(node, &arg);
1913 	if (arg.status_change_nid >= 0) {
1914 		kcompactd_stop(node);
1915 		kswapd_stop(node);
1916 	}
1917 
1918 	writeback_set_ratelimit();
1919 
1920 	memory_notify(MEM_OFFLINE, &arg);
1921 	remove_pfn_range_from_zone(zone, start_pfn, nr_pages);
1922 	mem_hotplug_done();
1923 	return 0;
1924 
1925 failed_removal_isolated:
1926 	/* pushback to free area */
1927 	undo_isolate_page_range(start_pfn, end_pfn, MIGRATE_MOVABLE);
1928 	memory_notify(MEM_CANCEL_OFFLINE, &arg);
1929 failed_removal_pcplists_disabled:
1930 	lru_cache_enable();
1931 	zone_pcp_enable(zone);
1932 failed_removal:
1933 	pr_debug("memory offlining [mem %#010llx-%#010llx] failed due to %s\n",
1934 		 (unsigned long long) start_pfn << PAGE_SHIFT,
1935 		 ((unsigned long long) end_pfn << PAGE_SHIFT) - 1,
1936 		 reason);
1937 	mem_hotplug_done();
1938 	return ret;
1939 }
1940 
1941 static int check_memblock_offlined_cb(struct memory_block *mem, void *arg)
1942 {
1943 	int *nid = arg;
1944 
1945 	*nid = mem->nid;
1946 	if (unlikely(mem->state != MEM_OFFLINE)) {
1947 		phys_addr_t beginpa, endpa;
1948 
1949 		beginpa = PFN_PHYS(section_nr_to_pfn(mem->start_section_nr));
1950 		endpa = beginpa + memory_block_size_bytes() - 1;
1951 		pr_warn("removing memory fails, because memory [%pa-%pa] is onlined\n",
1952 			&beginpa, &endpa);
1953 
1954 		return -EBUSY;
1955 	}
1956 	return 0;
1957 }
1958 
1959 static int get_nr_vmemmap_pages_cb(struct memory_block *mem, void *arg)
1960 {
1961 	/*
1962 	 * If not set, continue with the next block.
1963 	 */
1964 	return mem->nr_vmemmap_pages;
1965 }
1966 
1967 static int check_cpu_on_node(int nid)
1968 {
1969 	int cpu;
1970 
1971 	for_each_present_cpu(cpu) {
1972 		if (cpu_to_node(cpu) == nid)
1973 			/*
1974 			 * the cpu on this node isn't removed, and we can't
1975 			 * offline this node.
1976 			 */
1977 			return -EBUSY;
1978 	}
1979 
1980 	return 0;
1981 }
1982 
1983 static int check_no_memblock_for_node_cb(struct memory_block *mem, void *arg)
1984 {
1985 	int nid = *(int *)arg;
1986 
1987 	/*
1988 	 * If a memory block belongs to multiple nodes, the stored nid is not
1989 	 * reliable. However, such blocks are always online (e.g., cannot get
1990 	 * offlined) and, therefore, are still spanned by the node.
1991 	 */
1992 	return mem->nid == nid ? -EEXIST : 0;
1993 }
1994 
1995 /**
1996  * try_offline_node
1997  * @nid: the node ID
1998  *
1999  * Offline a node if all memory sections and cpus of the node are removed.
2000  *
2001  * NOTE: The caller must call lock_device_hotplug() to serialize hotplug
2002  * and online/offline operations before this call.
2003  */
2004 void try_offline_node(int nid)
2005 {
2006 	int rc;
2007 
2008 	/*
2009 	 * If the node still spans pages (especially ZONE_DEVICE), don't
2010 	 * offline it. A node spans memory after move_pfn_range_to_zone(),
2011 	 * e.g., after the memory block was onlined.
2012 	 */
2013 	if (node_spanned_pages(nid))
2014 		return;
2015 
2016 	/*
2017 	 * Especially offline memory blocks might not be spanned by the
2018 	 * node. They will get spanned by the node once they get onlined.
2019 	 * However, they link to the node in sysfs and can get onlined later.
2020 	 */
2021 	rc = for_each_memory_block(&nid, check_no_memblock_for_node_cb);
2022 	if (rc)
2023 		return;
2024 
2025 	if (check_cpu_on_node(nid))
2026 		return;
2027 
2028 	/*
2029 	 * all memory/cpu of this node are removed, we can offline this
2030 	 * node now.
2031 	 */
2032 	node_set_offline(nid);
2033 	unregister_one_node(nid);
2034 }
2035 EXPORT_SYMBOL(try_offline_node);
2036 
2037 static int __ref try_remove_memory(u64 start, u64 size)
2038 {
2039 	struct vmem_altmap mhp_altmap = {};
2040 	struct vmem_altmap *altmap = NULL;
2041 	unsigned long nr_vmemmap_pages;
2042 	int rc = 0, nid = NUMA_NO_NODE;
2043 
2044 	BUG_ON(check_hotplug_memory_range(start, size));
2045 
2046 	/*
2047 	 * All memory blocks must be offlined before removing memory.  Check
2048 	 * whether all memory blocks in question are offline and return error
2049 	 * if this is not the case.
2050 	 *
2051 	 * While at it, determine the nid. Note that if we'd have mixed nodes,
2052 	 * we'd only try to offline the last determined one -- which is good
2053 	 * enough for the cases we care about.
2054 	 */
2055 	rc = walk_memory_blocks(start, size, &nid, check_memblock_offlined_cb);
2056 	if (rc)
2057 		return rc;
2058 
2059 	/*
2060 	 * We only support removing memory added with MHP_MEMMAP_ON_MEMORY in
2061 	 * the same granularity it was added - a single memory block.
2062 	 */
2063 	if (mhp_memmap_on_memory()) {
2064 		nr_vmemmap_pages = walk_memory_blocks(start, size, NULL,
2065 						      get_nr_vmemmap_pages_cb);
2066 		if (nr_vmemmap_pages) {
2067 			if (size != memory_block_size_bytes()) {
2068 				pr_warn("Refuse to remove %#llx - %#llx,"
2069 					"wrong granularity\n",
2070 					start, start + size);
2071 				return -EINVAL;
2072 			}
2073 
2074 			/*
2075 			 * Let remove_pmd_table->free_hugepage_table do the
2076 			 * right thing if we used vmem_altmap when hot-adding
2077 			 * the range.
2078 			 */
2079 			mhp_altmap.alloc = nr_vmemmap_pages;
2080 			altmap = &mhp_altmap;
2081 		}
2082 	}
2083 
2084 	/* remove memmap entry */
2085 	firmware_map_remove(start, start + size, "System RAM");
2086 
2087 	/*
2088 	 * Memory block device removal under the device_hotplug_lock is
2089 	 * a barrier against racing online attempts.
2090 	 */
2091 	remove_memory_block_devices(start, size);
2092 
2093 	mem_hotplug_begin();
2094 
2095 	arch_remove_memory(start, size, altmap);
2096 
2097 	if (IS_ENABLED(CONFIG_ARCH_KEEP_MEMBLOCK)) {
2098 		memblock_phys_free(start, size);
2099 		memblock_remove(start, size);
2100 	}
2101 
2102 	release_mem_region_adjustable(start, size);
2103 
2104 	if (nid != NUMA_NO_NODE)
2105 		try_offline_node(nid);
2106 
2107 	mem_hotplug_done();
2108 	return 0;
2109 }
2110 
2111 /**
2112  * __remove_memory - Remove memory if every memory block is offline
2113  * @start: physical address of the region to remove
2114  * @size: size of the region to remove
2115  *
2116  * NOTE: The caller must call lock_device_hotplug() to serialize hotplug
2117  * and online/offline operations before this call, as required by
2118  * try_offline_node().
2119  */
2120 void __remove_memory(u64 start, u64 size)
2121 {
2122 
2123 	/*
2124 	 * trigger BUG() if some memory is not offlined prior to calling this
2125 	 * function
2126 	 */
2127 	if (try_remove_memory(start, size))
2128 		BUG();
2129 }
2130 
2131 /*
2132  * Remove memory if every memory block is offline, otherwise return -EBUSY is
2133  * some memory is not offline
2134  */
2135 int remove_memory(u64 start, u64 size)
2136 {
2137 	int rc;
2138 
2139 	lock_device_hotplug();
2140 	rc = try_remove_memory(start, size);
2141 	unlock_device_hotplug();
2142 
2143 	return rc;
2144 }
2145 EXPORT_SYMBOL_GPL(remove_memory);
2146 
2147 static int try_offline_memory_block(struct memory_block *mem, void *arg)
2148 {
2149 	uint8_t online_type = MMOP_ONLINE_KERNEL;
2150 	uint8_t **online_types = arg;
2151 	struct page *page;
2152 	int rc;
2153 
2154 	/*
2155 	 * Sense the online_type via the zone of the memory block. Offlining
2156 	 * with multiple zones within one memory block will be rejected
2157 	 * by offlining code ... so we don't care about that.
2158 	 */
2159 	page = pfn_to_online_page(section_nr_to_pfn(mem->start_section_nr));
2160 	if (page && zone_idx(page_zone(page)) == ZONE_MOVABLE)
2161 		online_type = MMOP_ONLINE_MOVABLE;
2162 
2163 	rc = device_offline(&mem->dev);
2164 	/*
2165 	 * Default is MMOP_OFFLINE - change it only if offlining succeeded,
2166 	 * so try_reonline_memory_block() can do the right thing.
2167 	 */
2168 	if (!rc)
2169 		**online_types = online_type;
2170 
2171 	(*online_types)++;
2172 	/* Ignore if already offline. */
2173 	return rc < 0 ? rc : 0;
2174 }
2175 
2176 static int try_reonline_memory_block(struct memory_block *mem, void *arg)
2177 {
2178 	uint8_t **online_types = arg;
2179 	int rc;
2180 
2181 	if (**online_types != MMOP_OFFLINE) {
2182 		mem->online_type = **online_types;
2183 		rc = device_online(&mem->dev);
2184 		if (rc < 0)
2185 			pr_warn("%s: Failed to re-online memory: %d",
2186 				__func__, rc);
2187 	}
2188 
2189 	/* Continue processing all remaining memory blocks. */
2190 	(*online_types)++;
2191 	return 0;
2192 }
2193 
2194 /*
2195  * Try to offline and remove memory. Might take a long time to finish in case
2196  * memory is still in use. Primarily useful for memory devices that logically
2197  * unplugged all memory (so it's no longer in use) and want to offline + remove
2198  * that memory.
2199  */
2200 int offline_and_remove_memory(u64 start, u64 size)
2201 {
2202 	const unsigned long mb_count = size / memory_block_size_bytes();
2203 	uint8_t *online_types, *tmp;
2204 	int rc;
2205 
2206 	if (!IS_ALIGNED(start, memory_block_size_bytes()) ||
2207 	    !IS_ALIGNED(size, memory_block_size_bytes()) || !size)
2208 		return -EINVAL;
2209 
2210 	/*
2211 	 * We'll remember the old online type of each memory block, so we can
2212 	 * try to revert whatever we did when offlining one memory block fails
2213 	 * after offlining some others succeeded.
2214 	 */
2215 	online_types = kmalloc_array(mb_count, sizeof(*online_types),
2216 				     GFP_KERNEL);
2217 	if (!online_types)
2218 		return -ENOMEM;
2219 	/*
2220 	 * Initialize all states to MMOP_OFFLINE, so when we abort processing in
2221 	 * try_offline_memory_block(), we'll skip all unprocessed blocks in
2222 	 * try_reonline_memory_block().
2223 	 */
2224 	memset(online_types, MMOP_OFFLINE, mb_count);
2225 
2226 	lock_device_hotplug();
2227 
2228 	tmp = online_types;
2229 	rc = walk_memory_blocks(start, size, &tmp, try_offline_memory_block);
2230 
2231 	/*
2232 	 * In case we succeeded to offline all memory, remove it.
2233 	 * This cannot fail as it cannot get onlined in the meantime.
2234 	 */
2235 	if (!rc) {
2236 		rc = try_remove_memory(start, size);
2237 		if (rc)
2238 			pr_err("%s: Failed to remove memory: %d", __func__, rc);
2239 	}
2240 
2241 	/*
2242 	 * Rollback what we did. While memory onlining might theoretically fail
2243 	 * (nacked by a notifier), it barely ever happens.
2244 	 */
2245 	if (rc) {
2246 		tmp = online_types;
2247 		walk_memory_blocks(start, size, &tmp,
2248 				   try_reonline_memory_block);
2249 	}
2250 	unlock_device_hotplug();
2251 
2252 	kfree(online_types);
2253 	return rc;
2254 }
2255 EXPORT_SYMBOL_GPL(offline_and_remove_memory);
2256 #endif /* CONFIG_MEMORY_HOTREMOVE */
2257