xref: /openbmc/linux/mm/memory_hotplug.c (revision 0b46b4ac)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  *  linux/mm/memory_hotplug.c
4  *
5  *  Copyright (C)
6  */
7 
8 #include <linux/stddef.h>
9 #include <linux/mm.h>
10 #include <linux/sched/signal.h>
11 #include <linux/swap.h>
12 #include <linux/interrupt.h>
13 #include <linux/pagemap.h>
14 #include <linux/compiler.h>
15 #include <linux/export.h>
16 #include <linux/writeback.h>
17 #include <linux/slab.h>
18 #include <linux/sysctl.h>
19 #include <linux/cpu.h>
20 #include <linux/memory.h>
21 #include <linux/memremap.h>
22 #include <linux/memory_hotplug.h>
23 #include <linux/vmalloc.h>
24 #include <linux/ioport.h>
25 #include <linux/delay.h>
26 #include <linux/migrate.h>
27 #include <linux/page-isolation.h>
28 #include <linux/pfn.h>
29 #include <linux/suspend.h>
30 #include <linux/mm_inline.h>
31 #include <linux/firmware-map.h>
32 #include <linux/stop_machine.h>
33 #include <linux/hugetlb.h>
34 #include <linux/memblock.h>
35 #include <linux/compaction.h>
36 #include <linux/rmap.h>
37 #include <linux/module.h>
38 
39 #include <asm/tlbflush.h>
40 
41 #include "internal.h"
42 #include "shuffle.h"
43 
44 enum {
45 	MEMMAP_ON_MEMORY_DISABLE = 0,
46 	MEMMAP_ON_MEMORY_ENABLE,
47 	MEMMAP_ON_MEMORY_FORCE,
48 };
49 
50 static int memmap_mode __read_mostly = MEMMAP_ON_MEMORY_DISABLE;
51 
memory_block_memmap_size(void)52 static inline unsigned long memory_block_memmap_size(void)
53 {
54 	return PHYS_PFN(memory_block_size_bytes()) * sizeof(struct page);
55 }
56 
memory_block_memmap_on_memory_pages(void)57 static inline unsigned long memory_block_memmap_on_memory_pages(void)
58 {
59 	unsigned long nr_pages = PFN_UP(memory_block_memmap_size());
60 
61 	/*
62 	 * In "forced" memmap_on_memory mode, we add extra pages to align the
63 	 * vmemmap size to cover full pageblocks. That way, we can add memory
64 	 * even if the vmemmap size is not properly aligned, however, we might waste
65 	 * memory.
66 	 */
67 	if (memmap_mode == MEMMAP_ON_MEMORY_FORCE)
68 		return pageblock_align(nr_pages);
69 	return nr_pages;
70 }
71 
72 #ifdef CONFIG_MHP_MEMMAP_ON_MEMORY
73 /*
74  * memory_hotplug.memmap_on_memory parameter
75  */
set_memmap_mode(const char * val,const struct kernel_param * kp)76 static int set_memmap_mode(const char *val, const struct kernel_param *kp)
77 {
78 	int ret, mode;
79 	bool enabled;
80 
81 	if (sysfs_streq(val, "force") ||  sysfs_streq(val, "FORCE")) {
82 		mode = MEMMAP_ON_MEMORY_FORCE;
83 	} else {
84 		ret = kstrtobool(val, &enabled);
85 		if (ret < 0)
86 			return ret;
87 		if (enabled)
88 			mode = MEMMAP_ON_MEMORY_ENABLE;
89 		else
90 			mode = MEMMAP_ON_MEMORY_DISABLE;
91 	}
92 	*((int *)kp->arg) = mode;
93 	if (mode == MEMMAP_ON_MEMORY_FORCE) {
94 		unsigned long memmap_pages = memory_block_memmap_on_memory_pages();
95 
96 		pr_info_once("Memory hotplug will waste %ld pages in each memory block\n",
97 			     memmap_pages - PFN_UP(memory_block_memmap_size()));
98 	}
99 	return 0;
100 }
101 
get_memmap_mode(char * buffer,const struct kernel_param * kp)102 static int get_memmap_mode(char *buffer, const struct kernel_param *kp)
103 {
104 	int mode = *((int *)kp->arg);
105 
106 	if (mode == MEMMAP_ON_MEMORY_FORCE)
107 		return sprintf(buffer, "force\n");
108 	return sprintf(buffer, "%c\n", mode ? 'Y' : 'N');
109 }
110 
111 static const struct kernel_param_ops memmap_mode_ops = {
112 	.set = set_memmap_mode,
113 	.get = get_memmap_mode,
114 };
115 module_param_cb(memmap_on_memory, &memmap_mode_ops, &memmap_mode, 0444);
116 MODULE_PARM_DESC(memmap_on_memory, "Enable memmap on memory for memory hotplug\n"
117 		 "With value \"force\" it could result in memory wastage due "
118 		 "to memmap size limitations (Y/N/force)");
119 
mhp_memmap_on_memory(void)120 static inline bool mhp_memmap_on_memory(void)
121 {
122 	return memmap_mode != MEMMAP_ON_MEMORY_DISABLE;
123 }
124 #else
mhp_memmap_on_memory(void)125 static inline bool mhp_memmap_on_memory(void)
126 {
127 	return false;
128 }
129 #endif
130 
131 enum {
132 	ONLINE_POLICY_CONTIG_ZONES = 0,
133 	ONLINE_POLICY_AUTO_MOVABLE,
134 };
135 
136 static const char * const online_policy_to_str[] = {
137 	[ONLINE_POLICY_CONTIG_ZONES] = "contig-zones",
138 	[ONLINE_POLICY_AUTO_MOVABLE] = "auto-movable",
139 };
140 
set_online_policy(const char * val,const struct kernel_param * kp)141 static int set_online_policy(const char *val, const struct kernel_param *kp)
142 {
143 	int ret = sysfs_match_string(online_policy_to_str, val);
144 
145 	if (ret < 0)
146 		return ret;
147 	*((int *)kp->arg) = ret;
148 	return 0;
149 }
150 
get_online_policy(char * buffer,const struct kernel_param * kp)151 static int get_online_policy(char *buffer, const struct kernel_param *kp)
152 {
153 	return sprintf(buffer, "%s\n", online_policy_to_str[*((int *)kp->arg)]);
154 }
155 
156 /*
157  * memory_hotplug.online_policy: configure online behavior when onlining without
158  * specifying a zone (MMOP_ONLINE)
159  *
160  * "contig-zones": keep zone contiguous
161  * "auto-movable": online memory to ZONE_MOVABLE if the configuration
162  *                 (auto_movable_ratio, auto_movable_numa_aware) allows for it
163  */
164 static int online_policy __read_mostly = ONLINE_POLICY_CONTIG_ZONES;
165 static const struct kernel_param_ops online_policy_ops = {
166 	.set = set_online_policy,
167 	.get = get_online_policy,
168 };
169 module_param_cb(online_policy, &online_policy_ops, &online_policy, 0644);
170 MODULE_PARM_DESC(online_policy,
171 		"Set the online policy (\"contig-zones\", \"auto-movable\") "
172 		"Default: \"contig-zones\"");
173 
174 /*
175  * memory_hotplug.auto_movable_ratio: specify maximum MOVABLE:KERNEL ratio
176  *
177  * The ratio represent an upper limit and the kernel might decide to not
178  * online some memory to ZONE_MOVABLE -- e.g., because hotplugged KERNEL memory
179  * doesn't allow for more MOVABLE memory.
180  */
181 static unsigned int auto_movable_ratio __read_mostly = 301;
182 module_param(auto_movable_ratio, uint, 0644);
183 MODULE_PARM_DESC(auto_movable_ratio,
184 		"Set the maximum ratio of MOVABLE:KERNEL memory in the system "
185 		"in percent for \"auto-movable\" online policy. Default: 301");
186 
187 /*
188  * memory_hotplug.auto_movable_numa_aware: consider numa node stats
189  */
190 #ifdef CONFIG_NUMA
191 static bool auto_movable_numa_aware __read_mostly = true;
192 module_param(auto_movable_numa_aware, bool, 0644);
193 MODULE_PARM_DESC(auto_movable_numa_aware,
194 		"Consider numa node stats in addition to global stats in "
195 		"\"auto-movable\" online policy. Default: true");
196 #endif /* CONFIG_NUMA */
197 
198 /*
199  * online_page_callback contains pointer to current page onlining function.
200  * Initially it is generic_online_page(). If it is required it could be
201  * changed by calling set_online_page_callback() for callback registration
202  * and restore_online_page_callback() for generic callback restore.
203  */
204 
205 static online_page_callback_t online_page_callback = generic_online_page;
206 static DEFINE_MUTEX(online_page_callback_lock);
207 
208 DEFINE_STATIC_PERCPU_RWSEM(mem_hotplug_lock);
209 
get_online_mems(void)210 void get_online_mems(void)
211 {
212 	percpu_down_read(&mem_hotplug_lock);
213 }
214 
put_online_mems(void)215 void put_online_mems(void)
216 {
217 	percpu_up_read(&mem_hotplug_lock);
218 }
219 
220 bool movable_node_enabled = false;
221 
222 #ifndef CONFIG_MEMORY_HOTPLUG_DEFAULT_ONLINE
223 int mhp_default_online_type = MMOP_OFFLINE;
224 #else
225 int mhp_default_online_type = MMOP_ONLINE;
226 #endif
227 
setup_memhp_default_state(char * str)228 static int __init setup_memhp_default_state(char *str)
229 {
230 	const int online_type = mhp_online_type_from_str(str);
231 
232 	if (online_type >= 0)
233 		mhp_default_online_type = online_type;
234 
235 	return 1;
236 }
237 __setup("memhp_default_state=", setup_memhp_default_state);
238 
mem_hotplug_begin(void)239 void mem_hotplug_begin(void)
240 {
241 	cpus_read_lock();
242 	percpu_down_write(&mem_hotplug_lock);
243 }
244 
mem_hotplug_done(void)245 void mem_hotplug_done(void)
246 {
247 	percpu_up_write(&mem_hotplug_lock);
248 	cpus_read_unlock();
249 }
250 
251 u64 max_mem_size = U64_MAX;
252 
253 /* add this memory to iomem resource */
register_memory_resource(u64 start,u64 size,const char * resource_name)254 static struct resource *register_memory_resource(u64 start, u64 size,
255 						 const char *resource_name)
256 {
257 	struct resource *res;
258 	unsigned long flags =  IORESOURCE_SYSTEM_RAM | IORESOURCE_BUSY;
259 
260 	if (strcmp(resource_name, "System RAM"))
261 		flags |= IORESOURCE_SYSRAM_DRIVER_MANAGED;
262 
263 	if (!mhp_range_allowed(start, size, true))
264 		return ERR_PTR(-E2BIG);
265 
266 	/*
267 	 * Make sure value parsed from 'mem=' only restricts memory adding
268 	 * while booting, so that memory hotplug won't be impacted. Please
269 	 * refer to document of 'mem=' in kernel-parameters.txt for more
270 	 * details.
271 	 */
272 	if (start + size > max_mem_size && system_state < SYSTEM_RUNNING)
273 		return ERR_PTR(-E2BIG);
274 
275 	/*
276 	 * Request ownership of the new memory range.  This might be
277 	 * a child of an existing resource that was present but
278 	 * not marked as busy.
279 	 */
280 	res = __request_region(&iomem_resource, start, size,
281 			       resource_name, flags);
282 
283 	if (!res) {
284 		pr_debug("Unable to reserve System RAM region: %016llx->%016llx\n",
285 				start, start + size);
286 		return ERR_PTR(-EEXIST);
287 	}
288 	return res;
289 }
290 
release_memory_resource(struct resource * res)291 static void release_memory_resource(struct resource *res)
292 {
293 	if (!res)
294 		return;
295 	release_resource(res);
296 	kfree(res);
297 }
298 
check_pfn_span(unsigned long pfn,unsigned long nr_pages)299 static int check_pfn_span(unsigned long pfn, unsigned long nr_pages)
300 {
301 	/*
302 	 * Disallow all operations smaller than a sub-section and only
303 	 * allow operations smaller than a section for
304 	 * SPARSEMEM_VMEMMAP. Note that check_hotplug_memory_range()
305 	 * enforces a larger memory_block_size_bytes() granularity for
306 	 * memory that will be marked online, so this check should only
307 	 * fire for direct arch_{add,remove}_memory() users outside of
308 	 * add_memory_resource().
309 	 */
310 	unsigned long min_align;
311 
312 	if (IS_ENABLED(CONFIG_SPARSEMEM_VMEMMAP))
313 		min_align = PAGES_PER_SUBSECTION;
314 	else
315 		min_align = PAGES_PER_SECTION;
316 	if (!IS_ALIGNED(pfn | nr_pages, min_align))
317 		return -EINVAL;
318 	return 0;
319 }
320 
321 /*
322  * Return page for the valid pfn only if the page is online. All pfn
323  * walkers which rely on the fully initialized page->flags and others
324  * should use this rather than pfn_valid && pfn_to_page
325  */
pfn_to_online_page(unsigned long pfn)326 struct page *pfn_to_online_page(unsigned long pfn)
327 {
328 	unsigned long nr = pfn_to_section_nr(pfn);
329 	struct dev_pagemap *pgmap;
330 	struct mem_section *ms;
331 
332 	if (nr >= NR_MEM_SECTIONS)
333 		return NULL;
334 
335 	ms = __nr_to_section(nr);
336 	if (!online_section(ms))
337 		return NULL;
338 
339 	/*
340 	 * Save some code text when online_section() +
341 	 * pfn_section_valid() are sufficient.
342 	 */
343 	if (IS_ENABLED(CONFIG_HAVE_ARCH_PFN_VALID) && !pfn_valid(pfn))
344 		return NULL;
345 
346 	if (!pfn_section_valid(ms, pfn))
347 		return NULL;
348 
349 	if (!online_device_section(ms))
350 		return pfn_to_page(pfn);
351 
352 	/*
353 	 * Slowpath: when ZONE_DEVICE collides with
354 	 * ZONE_{NORMAL,MOVABLE} within the same section some pfns in
355 	 * the section may be 'offline' but 'valid'. Only
356 	 * get_dev_pagemap() can determine sub-section online status.
357 	 */
358 	pgmap = get_dev_pagemap(pfn, NULL);
359 	put_dev_pagemap(pgmap);
360 
361 	/* The presence of a pgmap indicates ZONE_DEVICE offline pfn */
362 	if (pgmap)
363 		return NULL;
364 
365 	return pfn_to_page(pfn);
366 }
367 EXPORT_SYMBOL_GPL(pfn_to_online_page);
368 
__add_pages(int nid,unsigned long pfn,unsigned long nr_pages,struct mhp_params * params)369 int __ref __add_pages(int nid, unsigned long pfn, unsigned long nr_pages,
370 		struct mhp_params *params)
371 {
372 	const unsigned long end_pfn = pfn + nr_pages;
373 	unsigned long cur_nr_pages;
374 	int err;
375 	struct vmem_altmap *altmap = params->altmap;
376 
377 	if (WARN_ON_ONCE(!pgprot_val(params->pgprot)))
378 		return -EINVAL;
379 
380 	VM_BUG_ON(!mhp_range_allowed(PFN_PHYS(pfn), nr_pages * PAGE_SIZE, false));
381 
382 	if (altmap) {
383 		/*
384 		 * Validate altmap is within bounds of the total request
385 		 */
386 		if (altmap->base_pfn != pfn
387 				|| vmem_altmap_offset(altmap) > nr_pages) {
388 			pr_warn_once("memory add fail, invalid altmap\n");
389 			return -EINVAL;
390 		}
391 		altmap->alloc = 0;
392 	}
393 
394 	if (check_pfn_span(pfn, nr_pages)) {
395 		WARN(1, "Misaligned %s start: %#lx end: %#lx\n", __func__, pfn, pfn + nr_pages - 1);
396 		return -EINVAL;
397 	}
398 
399 	for (; pfn < end_pfn; pfn += cur_nr_pages) {
400 		/* Select all remaining pages up to the next section boundary */
401 		cur_nr_pages = min(end_pfn - pfn,
402 				   SECTION_ALIGN_UP(pfn + 1) - pfn);
403 		err = sparse_add_section(nid, pfn, cur_nr_pages, altmap,
404 					 params->pgmap);
405 		if (err)
406 			break;
407 		cond_resched();
408 	}
409 	vmemmap_populate_print_last();
410 	return err;
411 }
412 
413 /* find the smallest valid pfn in the range [start_pfn, end_pfn) */
find_smallest_section_pfn(int nid,struct zone * zone,unsigned long start_pfn,unsigned long end_pfn)414 static unsigned long find_smallest_section_pfn(int nid, struct zone *zone,
415 				     unsigned long start_pfn,
416 				     unsigned long end_pfn)
417 {
418 	for (; start_pfn < end_pfn; start_pfn += PAGES_PER_SUBSECTION) {
419 		if (unlikely(!pfn_to_online_page(start_pfn)))
420 			continue;
421 
422 		if (unlikely(pfn_to_nid(start_pfn) != nid))
423 			continue;
424 
425 		if (zone != page_zone(pfn_to_page(start_pfn)))
426 			continue;
427 
428 		return start_pfn;
429 	}
430 
431 	return 0;
432 }
433 
434 /* find the biggest valid pfn in the range [start_pfn, end_pfn). */
find_biggest_section_pfn(int nid,struct zone * zone,unsigned long start_pfn,unsigned long end_pfn)435 static unsigned long find_biggest_section_pfn(int nid, struct zone *zone,
436 				    unsigned long start_pfn,
437 				    unsigned long end_pfn)
438 {
439 	unsigned long pfn;
440 
441 	/* pfn is the end pfn of a memory section. */
442 	pfn = end_pfn - 1;
443 	for (; pfn >= start_pfn; pfn -= PAGES_PER_SUBSECTION) {
444 		if (unlikely(!pfn_to_online_page(pfn)))
445 			continue;
446 
447 		if (unlikely(pfn_to_nid(pfn) != nid))
448 			continue;
449 
450 		if (zone != page_zone(pfn_to_page(pfn)))
451 			continue;
452 
453 		return pfn;
454 	}
455 
456 	return 0;
457 }
458 
shrink_zone_span(struct zone * zone,unsigned long start_pfn,unsigned long end_pfn)459 static void shrink_zone_span(struct zone *zone, unsigned long start_pfn,
460 			     unsigned long end_pfn)
461 {
462 	unsigned long pfn;
463 	int nid = zone_to_nid(zone);
464 
465 	if (zone->zone_start_pfn == start_pfn) {
466 		/*
467 		 * If the section is smallest section in the zone, it need
468 		 * shrink zone->zone_start_pfn and zone->zone_spanned_pages.
469 		 * In this case, we find second smallest valid mem_section
470 		 * for shrinking zone.
471 		 */
472 		pfn = find_smallest_section_pfn(nid, zone, end_pfn,
473 						zone_end_pfn(zone));
474 		if (pfn) {
475 			zone->spanned_pages = zone_end_pfn(zone) - pfn;
476 			zone->zone_start_pfn = pfn;
477 		} else {
478 			zone->zone_start_pfn = 0;
479 			zone->spanned_pages = 0;
480 		}
481 	} else if (zone_end_pfn(zone) == end_pfn) {
482 		/*
483 		 * If the section is biggest section in the zone, it need
484 		 * shrink zone->spanned_pages.
485 		 * In this case, we find second biggest valid mem_section for
486 		 * shrinking zone.
487 		 */
488 		pfn = find_biggest_section_pfn(nid, zone, zone->zone_start_pfn,
489 					       start_pfn);
490 		if (pfn)
491 			zone->spanned_pages = pfn - zone->zone_start_pfn + 1;
492 		else {
493 			zone->zone_start_pfn = 0;
494 			zone->spanned_pages = 0;
495 		}
496 	}
497 }
498 
update_pgdat_span(struct pglist_data * pgdat)499 static void update_pgdat_span(struct pglist_data *pgdat)
500 {
501 	unsigned long node_start_pfn = 0, node_end_pfn = 0;
502 	struct zone *zone;
503 
504 	for (zone = pgdat->node_zones;
505 	     zone < pgdat->node_zones + MAX_NR_ZONES; zone++) {
506 		unsigned long end_pfn = zone_end_pfn(zone);
507 
508 		/* No need to lock the zones, they can't change. */
509 		if (!zone->spanned_pages)
510 			continue;
511 		if (!node_end_pfn) {
512 			node_start_pfn = zone->zone_start_pfn;
513 			node_end_pfn = end_pfn;
514 			continue;
515 		}
516 
517 		if (end_pfn > node_end_pfn)
518 			node_end_pfn = end_pfn;
519 		if (zone->zone_start_pfn < node_start_pfn)
520 			node_start_pfn = zone->zone_start_pfn;
521 	}
522 
523 	pgdat->node_start_pfn = node_start_pfn;
524 	pgdat->node_spanned_pages = node_end_pfn - node_start_pfn;
525 }
526 
remove_pfn_range_from_zone(struct zone * zone,unsigned long start_pfn,unsigned long nr_pages)527 void __ref remove_pfn_range_from_zone(struct zone *zone,
528 				      unsigned long start_pfn,
529 				      unsigned long nr_pages)
530 {
531 	const unsigned long end_pfn = start_pfn + nr_pages;
532 	struct pglist_data *pgdat = zone->zone_pgdat;
533 	unsigned long pfn, cur_nr_pages;
534 
535 	/* Poison struct pages because they are now uninitialized again. */
536 	for (pfn = start_pfn; pfn < end_pfn; pfn += cur_nr_pages) {
537 		cond_resched();
538 
539 		/* Select all remaining pages up to the next section boundary */
540 		cur_nr_pages =
541 			min(end_pfn - pfn, SECTION_ALIGN_UP(pfn + 1) - pfn);
542 		page_init_poison(pfn_to_page(pfn),
543 				 sizeof(struct page) * cur_nr_pages);
544 	}
545 
546 	/*
547 	 * Zone shrinking code cannot properly deal with ZONE_DEVICE. So
548 	 * we will not try to shrink the zones - which is okay as
549 	 * set_zone_contiguous() cannot deal with ZONE_DEVICE either way.
550 	 */
551 	if (zone_is_zone_device(zone))
552 		return;
553 
554 	clear_zone_contiguous(zone);
555 
556 	shrink_zone_span(zone, start_pfn, start_pfn + nr_pages);
557 	update_pgdat_span(pgdat);
558 
559 	set_zone_contiguous(zone);
560 }
561 
562 /**
563  * __remove_pages() - remove sections of pages
564  * @pfn: starting pageframe (must be aligned to start of a section)
565  * @nr_pages: number of pages to remove (must be multiple of section size)
566  * @altmap: alternative device page map or %NULL if default memmap is used
567  *
568  * Generic helper function to remove section mappings and sysfs entries
569  * for the section of the memory we are removing. Caller needs to make
570  * sure that pages are marked reserved and zones are adjust properly by
571  * calling offline_pages().
572  */
__remove_pages(unsigned long pfn,unsigned long nr_pages,struct vmem_altmap * altmap)573 void __remove_pages(unsigned long pfn, unsigned long nr_pages,
574 		    struct vmem_altmap *altmap)
575 {
576 	const unsigned long end_pfn = pfn + nr_pages;
577 	unsigned long cur_nr_pages;
578 
579 	if (check_pfn_span(pfn, nr_pages)) {
580 		WARN(1, "Misaligned %s start: %#lx end: %#lx\n", __func__, pfn, pfn + nr_pages - 1);
581 		return;
582 	}
583 
584 	for (; pfn < end_pfn; pfn += cur_nr_pages) {
585 		cond_resched();
586 		/* Select all remaining pages up to the next section boundary */
587 		cur_nr_pages = min(end_pfn - pfn,
588 				   SECTION_ALIGN_UP(pfn + 1) - pfn);
589 		sparse_remove_section(pfn, cur_nr_pages, altmap);
590 	}
591 }
592 
set_online_page_callback(online_page_callback_t callback)593 int set_online_page_callback(online_page_callback_t callback)
594 {
595 	int rc = -EINVAL;
596 
597 	get_online_mems();
598 	mutex_lock(&online_page_callback_lock);
599 
600 	if (online_page_callback == generic_online_page) {
601 		online_page_callback = callback;
602 		rc = 0;
603 	}
604 
605 	mutex_unlock(&online_page_callback_lock);
606 	put_online_mems();
607 
608 	return rc;
609 }
610 EXPORT_SYMBOL_GPL(set_online_page_callback);
611 
restore_online_page_callback(online_page_callback_t callback)612 int restore_online_page_callback(online_page_callback_t callback)
613 {
614 	int rc = -EINVAL;
615 
616 	get_online_mems();
617 	mutex_lock(&online_page_callback_lock);
618 
619 	if (online_page_callback == callback) {
620 		online_page_callback = generic_online_page;
621 		rc = 0;
622 	}
623 
624 	mutex_unlock(&online_page_callback_lock);
625 	put_online_mems();
626 
627 	return rc;
628 }
629 EXPORT_SYMBOL_GPL(restore_online_page_callback);
630 
generic_online_page(struct page * page,unsigned int order)631 void generic_online_page(struct page *page, unsigned int order)
632 {
633 	/*
634 	 * Freeing the page with debug_pagealloc enabled will try to unmap it,
635 	 * so we should map it first. This is better than introducing a special
636 	 * case in page freeing fast path.
637 	 */
638 	debug_pagealloc_map_pages(page, 1 << order);
639 	__free_pages_core(page, order);
640 	totalram_pages_add(1UL << order);
641 }
642 EXPORT_SYMBOL_GPL(generic_online_page);
643 
online_pages_range(unsigned long start_pfn,unsigned long nr_pages)644 static void online_pages_range(unsigned long start_pfn, unsigned long nr_pages)
645 {
646 	const unsigned long end_pfn = start_pfn + nr_pages;
647 	unsigned long pfn;
648 
649 	/*
650 	 * Online the pages in MAX_ORDER aligned chunks. The callback might
651 	 * decide to not expose all pages to the buddy (e.g., expose them
652 	 * later). We account all pages as being online and belonging to this
653 	 * zone ("present").
654 	 * When using memmap_on_memory, the range might not be aligned to
655 	 * MAX_ORDER_NR_PAGES - 1, but pageblock aligned. __ffs() will detect
656 	 * this and the first chunk to online will be pageblock_nr_pages.
657 	 */
658 	for (pfn = start_pfn; pfn < end_pfn;) {
659 		int order;
660 
661 		/*
662 		 * Free to online pages in the largest chunks alignment allows.
663 		 *
664 		 * __ffs() behaviour is undefined for 0. start == 0 is
665 		 * MAX_ORDER-aligned, Set order to MAX_ORDER for the case.
666 		 */
667 		if (pfn)
668 			order = min_t(int, MAX_ORDER, __ffs(pfn));
669 		else
670 			order = MAX_ORDER;
671 
672 		(*online_page_callback)(pfn_to_page(pfn), order);
673 		pfn += (1UL << order);
674 	}
675 
676 	/* mark all involved sections as online */
677 	online_mem_sections(start_pfn, end_pfn);
678 }
679 
680 /* check which state of node_states will be changed when online memory */
node_states_check_changes_online(unsigned long nr_pages,struct zone * zone,struct memory_notify * arg)681 static void node_states_check_changes_online(unsigned long nr_pages,
682 	struct zone *zone, struct memory_notify *arg)
683 {
684 	int nid = zone_to_nid(zone);
685 
686 	arg->status_change_nid = NUMA_NO_NODE;
687 	arg->status_change_nid_normal = NUMA_NO_NODE;
688 
689 	if (!node_state(nid, N_MEMORY))
690 		arg->status_change_nid = nid;
691 	if (zone_idx(zone) <= ZONE_NORMAL && !node_state(nid, N_NORMAL_MEMORY))
692 		arg->status_change_nid_normal = nid;
693 }
694 
node_states_set_node(int node,struct memory_notify * arg)695 static void node_states_set_node(int node, struct memory_notify *arg)
696 {
697 	if (arg->status_change_nid_normal >= 0)
698 		node_set_state(node, N_NORMAL_MEMORY);
699 
700 	if (arg->status_change_nid >= 0)
701 		node_set_state(node, N_MEMORY);
702 }
703 
resize_zone_range(struct zone * zone,unsigned long start_pfn,unsigned long nr_pages)704 static void __meminit resize_zone_range(struct zone *zone, unsigned long start_pfn,
705 		unsigned long nr_pages)
706 {
707 	unsigned long old_end_pfn = zone_end_pfn(zone);
708 
709 	if (zone_is_empty(zone) || start_pfn < zone->zone_start_pfn)
710 		zone->zone_start_pfn = start_pfn;
711 
712 	zone->spanned_pages = max(start_pfn + nr_pages, old_end_pfn) - zone->zone_start_pfn;
713 }
714 
resize_pgdat_range(struct pglist_data * pgdat,unsigned long start_pfn,unsigned long nr_pages)715 static void __meminit resize_pgdat_range(struct pglist_data *pgdat, unsigned long start_pfn,
716                                      unsigned long nr_pages)
717 {
718 	unsigned long old_end_pfn = pgdat_end_pfn(pgdat);
719 
720 	if (!pgdat->node_spanned_pages || start_pfn < pgdat->node_start_pfn)
721 		pgdat->node_start_pfn = start_pfn;
722 
723 	pgdat->node_spanned_pages = max(start_pfn + nr_pages, old_end_pfn) - pgdat->node_start_pfn;
724 
725 }
726 
727 #ifdef CONFIG_ZONE_DEVICE
section_taint_zone_device(unsigned long pfn)728 static void section_taint_zone_device(unsigned long pfn)
729 {
730 	struct mem_section *ms = __pfn_to_section(pfn);
731 
732 	ms->section_mem_map |= SECTION_TAINT_ZONE_DEVICE;
733 }
734 #else
section_taint_zone_device(unsigned long pfn)735 static inline void section_taint_zone_device(unsigned long pfn)
736 {
737 }
738 #endif
739 
740 /*
741  * Associate the pfn range with the given zone, initializing the memmaps
742  * and resizing the pgdat/zone data to span the added pages. After this
743  * call, all affected pages are PG_reserved.
744  *
745  * All aligned pageblocks are initialized to the specified migratetype
746  * (usually MIGRATE_MOVABLE). Besides setting the migratetype, no related
747  * zone stats (e.g., nr_isolate_pageblock) are touched.
748  */
move_pfn_range_to_zone(struct zone * zone,unsigned long start_pfn,unsigned long nr_pages,struct vmem_altmap * altmap,int migratetype)749 void __ref move_pfn_range_to_zone(struct zone *zone, unsigned long start_pfn,
750 				  unsigned long nr_pages,
751 				  struct vmem_altmap *altmap, int migratetype)
752 {
753 	struct pglist_data *pgdat = zone->zone_pgdat;
754 	int nid = pgdat->node_id;
755 
756 	clear_zone_contiguous(zone);
757 
758 	if (zone_is_empty(zone))
759 		init_currently_empty_zone(zone, start_pfn, nr_pages);
760 	resize_zone_range(zone, start_pfn, nr_pages);
761 	resize_pgdat_range(pgdat, start_pfn, nr_pages);
762 
763 	/*
764 	 * Subsection population requires care in pfn_to_online_page().
765 	 * Set the taint to enable the slow path detection of
766 	 * ZONE_DEVICE pages in an otherwise  ZONE_{NORMAL,MOVABLE}
767 	 * section.
768 	 */
769 	if (zone_is_zone_device(zone)) {
770 		if (!IS_ALIGNED(start_pfn, PAGES_PER_SECTION))
771 			section_taint_zone_device(start_pfn);
772 		if (!IS_ALIGNED(start_pfn + nr_pages, PAGES_PER_SECTION))
773 			section_taint_zone_device(start_pfn + nr_pages);
774 	}
775 
776 	/*
777 	 * TODO now we have a visible range of pages which are not associated
778 	 * with their zone properly. Not nice but set_pfnblock_flags_mask
779 	 * expects the zone spans the pfn range. All the pages in the range
780 	 * are reserved so nobody should be touching them so we should be safe
781 	 */
782 	memmap_init_range(nr_pages, nid, zone_idx(zone), start_pfn, 0,
783 			 MEMINIT_HOTPLUG, altmap, migratetype);
784 
785 	set_zone_contiguous(zone);
786 }
787 
788 struct auto_movable_stats {
789 	unsigned long kernel_early_pages;
790 	unsigned long movable_pages;
791 };
792 
auto_movable_stats_account_zone(struct auto_movable_stats * stats,struct zone * zone)793 static void auto_movable_stats_account_zone(struct auto_movable_stats *stats,
794 					    struct zone *zone)
795 {
796 	if (zone_idx(zone) == ZONE_MOVABLE) {
797 		stats->movable_pages += zone->present_pages;
798 	} else {
799 		stats->kernel_early_pages += zone->present_early_pages;
800 #ifdef CONFIG_CMA
801 		/*
802 		 * CMA pages (never on hotplugged memory) behave like
803 		 * ZONE_MOVABLE.
804 		 */
805 		stats->movable_pages += zone->cma_pages;
806 		stats->kernel_early_pages -= zone->cma_pages;
807 #endif /* CONFIG_CMA */
808 	}
809 }
810 struct auto_movable_group_stats {
811 	unsigned long movable_pages;
812 	unsigned long req_kernel_early_pages;
813 };
814 
auto_movable_stats_account_group(struct memory_group * group,void * arg)815 static int auto_movable_stats_account_group(struct memory_group *group,
816 					   void *arg)
817 {
818 	const int ratio = READ_ONCE(auto_movable_ratio);
819 	struct auto_movable_group_stats *stats = arg;
820 	long pages;
821 
822 	/*
823 	 * We don't support modifying the config while the auto-movable online
824 	 * policy is already enabled. Just avoid the division by zero below.
825 	 */
826 	if (!ratio)
827 		return 0;
828 
829 	/*
830 	 * Calculate how many early kernel pages this group requires to
831 	 * satisfy the configured zone ratio.
832 	 */
833 	pages = group->present_movable_pages * 100 / ratio;
834 	pages -= group->present_kernel_pages;
835 
836 	if (pages > 0)
837 		stats->req_kernel_early_pages += pages;
838 	stats->movable_pages += group->present_movable_pages;
839 	return 0;
840 }
841 
auto_movable_can_online_movable(int nid,struct memory_group * group,unsigned long nr_pages)842 static bool auto_movable_can_online_movable(int nid, struct memory_group *group,
843 					    unsigned long nr_pages)
844 {
845 	unsigned long kernel_early_pages, movable_pages;
846 	struct auto_movable_group_stats group_stats = {};
847 	struct auto_movable_stats stats = {};
848 	pg_data_t *pgdat = NODE_DATA(nid);
849 	struct zone *zone;
850 	int i;
851 
852 	/* Walk all relevant zones and collect MOVABLE vs. KERNEL stats. */
853 	if (nid == NUMA_NO_NODE) {
854 		/* TODO: cache values */
855 		for_each_populated_zone(zone)
856 			auto_movable_stats_account_zone(&stats, zone);
857 	} else {
858 		for (i = 0; i < MAX_NR_ZONES; i++) {
859 			zone = pgdat->node_zones + i;
860 			if (populated_zone(zone))
861 				auto_movable_stats_account_zone(&stats, zone);
862 		}
863 	}
864 
865 	kernel_early_pages = stats.kernel_early_pages;
866 	movable_pages = stats.movable_pages;
867 
868 	/*
869 	 * Kernel memory inside dynamic memory group allows for more MOVABLE
870 	 * memory within the same group. Remove the effect of all but the
871 	 * current group from the stats.
872 	 */
873 	walk_dynamic_memory_groups(nid, auto_movable_stats_account_group,
874 				   group, &group_stats);
875 	if (kernel_early_pages <= group_stats.req_kernel_early_pages)
876 		return false;
877 	kernel_early_pages -= group_stats.req_kernel_early_pages;
878 	movable_pages -= group_stats.movable_pages;
879 
880 	if (group && group->is_dynamic)
881 		kernel_early_pages += group->present_kernel_pages;
882 
883 	/*
884 	 * Test if we could online the given number of pages to ZONE_MOVABLE
885 	 * and still stay in the configured ratio.
886 	 */
887 	movable_pages += nr_pages;
888 	return movable_pages <= (auto_movable_ratio * kernel_early_pages) / 100;
889 }
890 
891 /*
892  * Returns a default kernel memory zone for the given pfn range.
893  * If no kernel zone covers this pfn range it will automatically go
894  * to the ZONE_NORMAL.
895  */
default_kernel_zone_for_pfn(int nid,unsigned long start_pfn,unsigned long nr_pages)896 static struct zone *default_kernel_zone_for_pfn(int nid, unsigned long start_pfn,
897 		unsigned long nr_pages)
898 {
899 	struct pglist_data *pgdat = NODE_DATA(nid);
900 	int zid;
901 
902 	for (zid = 0; zid < ZONE_NORMAL; zid++) {
903 		struct zone *zone = &pgdat->node_zones[zid];
904 
905 		if (zone_intersects(zone, start_pfn, nr_pages))
906 			return zone;
907 	}
908 
909 	return &pgdat->node_zones[ZONE_NORMAL];
910 }
911 
912 /*
913  * Determine to which zone to online memory dynamically based on user
914  * configuration and system stats. We care about the following ratio:
915  *
916  *   MOVABLE : KERNEL
917  *
918  * Whereby MOVABLE is memory in ZONE_MOVABLE and KERNEL is memory in
919  * one of the kernel zones. CMA pages inside one of the kernel zones really
920  * behaves like ZONE_MOVABLE, so we treat them accordingly.
921  *
922  * We don't allow for hotplugged memory in a KERNEL zone to increase the
923  * amount of MOVABLE memory we can have, so we end up with:
924  *
925  *   MOVABLE : KERNEL_EARLY
926  *
927  * Whereby KERNEL_EARLY is memory in one of the kernel zones, available sinze
928  * boot. We base our calculation on KERNEL_EARLY internally, because:
929  *
930  * a) Hotplugged memory in one of the kernel zones can sometimes still get
931  *    hotunplugged, especially when hot(un)plugging individual memory blocks.
932  *    There is no coordination across memory devices, therefore "automatic"
933  *    hotunplugging, as implemented in hypervisors, could result in zone
934  *    imbalances.
935  * b) Early/boot memory in one of the kernel zones can usually not get
936  *    hotunplugged again (e.g., no firmware interface to unplug, fragmented
937  *    with unmovable allocations). While there are corner cases where it might
938  *    still work, it is barely relevant in practice.
939  *
940  * Exceptions are dynamic memory groups, which allow for more MOVABLE
941  * memory within the same memory group -- because in that case, there is
942  * coordination within the single memory device managed by a single driver.
943  *
944  * We rely on "present pages" instead of "managed pages", as the latter is
945  * highly unreliable and dynamic in virtualized environments, and does not
946  * consider boot time allocations. For example, memory ballooning adjusts the
947  * managed pages when inflating/deflating the balloon, and balloon compaction
948  * can even migrate inflated pages between zones.
949  *
950  * Using "present pages" is better but some things to keep in mind are:
951  *
952  * a) Some memblock allocations, such as for the crashkernel area, are
953  *    effectively unused by the kernel, yet they account to "present pages".
954  *    Fortunately, these allocations are comparatively small in relevant setups
955  *    (e.g., fraction of system memory).
956  * b) Some hotplugged memory blocks in virtualized environments, esecially
957  *    hotplugged by virtio-mem, look like they are completely present, however,
958  *    only parts of the memory block are actually currently usable.
959  *    "present pages" is an upper limit that can get reached at runtime. As
960  *    we base our calculations on KERNEL_EARLY, this is not an issue.
961  */
auto_movable_zone_for_pfn(int nid,struct memory_group * group,unsigned long pfn,unsigned long nr_pages)962 static struct zone *auto_movable_zone_for_pfn(int nid,
963 					      struct memory_group *group,
964 					      unsigned long pfn,
965 					      unsigned long nr_pages)
966 {
967 	unsigned long online_pages = 0, max_pages, end_pfn;
968 	struct page *page;
969 
970 	if (!auto_movable_ratio)
971 		goto kernel_zone;
972 
973 	if (group && !group->is_dynamic) {
974 		max_pages = group->s.max_pages;
975 		online_pages = group->present_movable_pages;
976 
977 		/* If anything is !MOVABLE online the rest !MOVABLE. */
978 		if (group->present_kernel_pages)
979 			goto kernel_zone;
980 	} else if (!group || group->d.unit_pages == nr_pages) {
981 		max_pages = nr_pages;
982 	} else {
983 		max_pages = group->d.unit_pages;
984 		/*
985 		 * Take a look at all online sections in the current unit.
986 		 * We can safely assume that all pages within a section belong
987 		 * to the same zone, because dynamic memory groups only deal
988 		 * with hotplugged memory.
989 		 */
990 		pfn = ALIGN_DOWN(pfn, group->d.unit_pages);
991 		end_pfn = pfn + group->d.unit_pages;
992 		for (; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
993 			page = pfn_to_online_page(pfn);
994 			if (!page)
995 				continue;
996 			/* If anything is !MOVABLE online the rest !MOVABLE. */
997 			if (!is_zone_movable_page(page))
998 				goto kernel_zone;
999 			online_pages += PAGES_PER_SECTION;
1000 		}
1001 	}
1002 
1003 	/*
1004 	 * Online MOVABLE if we could *currently* online all remaining parts
1005 	 * MOVABLE. We expect to (add+) online them immediately next, so if
1006 	 * nobody interferes, all will be MOVABLE if possible.
1007 	 */
1008 	nr_pages = max_pages - online_pages;
1009 	if (!auto_movable_can_online_movable(NUMA_NO_NODE, group, nr_pages))
1010 		goto kernel_zone;
1011 
1012 #ifdef CONFIG_NUMA
1013 	if (auto_movable_numa_aware &&
1014 	    !auto_movable_can_online_movable(nid, group, nr_pages))
1015 		goto kernel_zone;
1016 #endif /* CONFIG_NUMA */
1017 
1018 	return &NODE_DATA(nid)->node_zones[ZONE_MOVABLE];
1019 kernel_zone:
1020 	return default_kernel_zone_for_pfn(nid, pfn, nr_pages);
1021 }
1022 
default_zone_for_pfn(int nid,unsigned long start_pfn,unsigned long nr_pages)1023 static inline struct zone *default_zone_for_pfn(int nid, unsigned long start_pfn,
1024 		unsigned long nr_pages)
1025 {
1026 	struct zone *kernel_zone = default_kernel_zone_for_pfn(nid, start_pfn,
1027 			nr_pages);
1028 	struct zone *movable_zone = &NODE_DATA(nid)->node_zones[ZONE_MOVABLE];
1029 	bool in_kernel = zone_intersects(kernel_zone, start_pfn, nr_pages);
1030 	bool in_movable = zone_intersects(movable_zone, start_pfn, nr_pages);
1031 
1032 	/*
1033 	 * We inherit the existing zone in a simple case where zones do not
1034 	 * overlap in the given range
1035 	 */
1036 	if (in_kernel ^ in_movable)
1037 		return (in_kernel) ? kernel_zone : movable_zone;
1038 
1039 	/*
1040 	 * If the range doesn't belong to any zone or two zones overlap in the
1041 	 * given range then we use movable zone only if movable_node is
1042 	 * enabled because we always online to a kernel zone by default.
1043 	 */
1044 	return movable_node_enabled ? movable_zone : kernel_zone;
1045 }
1046 
zone_for_pfn_range(int online_type,int nid,struct memory_group * group,unsigned long start_pfn,unsigned long nr_pages)1047 struct zone *zone_for_pfn_range(int online_type, int nid,
1048 		struct memory_group *group, unsigned long start_pfn,
1049 		unsigned long nr_pages)
1050 {
1051 	if (online_type == MMOP_ONLINE_KERNEL)
1052 		return default_kernel_zone_for_pfn(nid, start_pfn, nr_pages);
1053 
1054 	if (online_type == MMOP_ONLINE_MOVABLE)
1055 		return &NODE_DATA(nid)->node_zones[ZONE_MOVABLE];
1056 
1057 	if (online_policy == ONLINE_POLICY_AUTO_MOVABLE)
1058 		return auto_movable_zone_for_pfn(nid, group, start_pfn, nr_pages);
1059 
1060 	return default_zone_for_pfn(nid, start_pfn, nr_pages);
1061 }
1062 
1063 /*
1064  * This function should only be called by memory_block_{online,offline},
1065  * and {online,offline}_pages.
1066  */
adjust_present_page_count(struct page * page,struct memory_group * group,long nr_pages)1067 void adjust_present_page_count(struct page *page, struct memory_group *group,
1068 			       long nr_pages)
1069 {
1070 	struct zone *zone = page_zone(page);
1071 	const bool movable = zone_idx(zone) == ZONE_MOVABLE;
1072 
1073 	/*
1074 	 * We only support onlining/offlining/adding/removing of complete
1075 	 * memory blocks; therefore, either all is either early or hotplugged.
1076 	 */
1077 	if (early_section(__pfn_to_section(page_to_pfn(page))))
1078 		zone->present_early_pages += nr_pages;
1079 	zone->present_pages += nr_pages;
1080 	zone->zone_pgdat->node_present_pages += nr_pages;
1081 
1082 	if (group && movable)
1083 		group->present_movable_pages += nr_pages;
1084 	else if (group && !movable)
1085 		group->present_kernel_pages += nr_pages;
1086 }
1087 
mhp_init_memmap_on_memory(unsigned long pfn,unsigned long nr_pages,struct zone * zone)1088 int mhp_init_memmap_on_memory(unsigned long pfn, unsigned long nr_pages,
1089 			      struct zone *zone)
1090 {
1091 	unsigned long end_pfn = pfn + nr_pages;
1092 	int ret, i;
1093 
1094 	ret = kasan_add_zero_shadow(__va(PFN_PHYS(pfn)), PFN_PHYS(nr_pages));
1095 	if (ret)
1096 		return ret;
1097 
1098 	move_pfn_range_to_zone(zone, pfn, nr_pages, NULL, MIGRATE_UNMOVABLE);
1099 
1100 	for (i = 0; i < nr_pages; i++)
1101 		SetPageVmemmapSelfHosted(pfn_to_page(pfn + i));
1102 
1103 	/*
1104 	 * It might be that the vmemmap_pages fully span sections. If that is
1105 	 * the case, mark those sections online here as otherwise they will be
1106 	 * left offline.
1107 	 */
1108 	if (nr_pages >= PAGES_PER_SECTION)
1109 	        online_mem_sections(pfn, ALIGN_DOWN(end_pfn, PAGES_PER_SECTION));
1110 
1111 	return ret;
1112 }
1113 
mhp_deinit_memmap_on_memory(unsigned long pfn,unsigned long nr_pages)1114 void mhp_deinit_memmap_on_memory(unsigned long pfn, unsigned long nr_pages)
1115 {
1116 	unsigned long end_pfn = pfn + nr_pages;
1117 
1118 	/*
1119 	 * It might be that the vmemmap_pages fully span sections. If that is
1120 	 * the case, mark those sections offline here as otherwise they will be
1121 	 * left online.
1122 	 */
1123 	if (nr_pages >= PAGES_PER_SECTION)
1124 		offline_mem_sections(pfn, ALIGN_DOWN(end_pfn, PAGES_PER_SECTION));
1125 
1126         /*
1127 	 * The pages associated with this vmemmap have been offlined, so
1128 	 * we can reset its state here.
1129 	 */
1130 	remove_pfn_range_from_zone(page_zone(pfn_to_page(pfn)), pfn, nr_pages);
1131 	kasan_remove_zero_shadow(__va(PFN_PHYS(pfn)), PFN_PHYS(nr_pages));
1132 }
1133 
1134 /*
1135  * Must be called with mem_hotplug_lock in write mode.
1136  */
online_pages(unsigned long pfn,unsigned long nr_pages,struct zone * zone,struct memory_group * group)1137 int __ref online_pages(unsigned long pfn, unsigned long nr_pages,
1138 		       struct zone *zone, struct memory_group *group)
1139 {
1140 	unsigned long flags;
1141 	int need_zonelists_rebuild = 0;
1142 	const int nid = zone_to_nid(zone);
1143 	int ret;
1144 	struct memory_notify arg;
1145 
1146 	/*
1147 	 * {on,off}lining is constrained to full memory sections (or more
1148 	 * precisely to memory blocks from the user space POV).
1149 	 * memmap_on_memory is an exception because it reserves initial part
1150 	 * of the physical memory space for vmemmaps. That space is pageblock
1151 	 * aligned.
1152 	 */
1153 	if (WARN_ON_ONCE(!nr_pages || !pageblock_aligned(pfn) ||
1154 			 !IS_ALIGNED(pfn + nr_pages, PAGES_PER_SECTION)))
1155 		return -EINVAL;
1156 
1157 
1158 	/* associate pfn range with the zone */
1159 	move_pfn_range_to_zone(zone, pfn, nr_pages, NULL, MIGRATE_ISOLATE);
1160 
1161 	arg.start_pfn = pfn;
1162 	arg.nr_pages = nr_pages;
1163 	node_states_check_changes_online(nr_pages, zone, &arg);
1164 
1165 	ret = memory_notify(MEM_GOING_ONLINE, &arg);
1166 	ret = notifier_to_errno(ret);
1167 	if (ret)
1168 		goto failed_addition;
1169 
1170 	/*
1171 	 * Fixup the number of isolated pageblocks before marking the sections
1172 	 * onlining, such that undo_isolate_page_range() works correctly.
1173 	 */
1174 	spin_lock_irqsave(&zone->lock, flags);
1175 	zone->nr_isolate_pageblock += nr_pages / pageblock_nr_pages;
1176 	spin_unlock_irqrestore(&zone->lock, flags);
1177 
1178 	/*
1179 	 * If this zone is not populated, then it is not in zonelist.
1180 	 * This means the page allocator ignores this zone.
1181 	 * So, zonelist must be updated after online.
1182 	 */
1183 	if (!populated_zone(zone)) {
1184 		need_zonelists_rebuild = 1;
1185 		setup_zone_pageset(zone);
1186 	}
1187 
1188 	online_pages_range(pfn, nr_pages);
1189 	adjust_present_page_count(pfn_to_page(pfn), group, nr_pages);
1190 
1191 	node_states_set_node(nid, &arg);
1192 	if (need_zonelists_rebuild)
1193 		build_all_zonelists(NULL);
1194 
1195 	/* Basic onlining is complete, allow allocation of onlined pages. */
1196 	undo_isolate_page_range(pfn, pfn + nr_pages, MIGRATE_MOVABLE);
1197 
1198 	/*
1199 	 * Freshly onlined pages aren't shuffled (e.g., all pages are placed to
1200 	 * the tail of the freelist when undoing isolation). Shuffle the whole
1201 	 * zone to make sure the just onlined pages are properly distributed
1202 	 * across the whole freelist - to create an initial shuffle.
1203 	 */
1204 	shuffle_zone(zone);
1205 
1206 	/* reinitialise watermarks and update pcp limits */
1207 	init_per_zone_wmark_min();
1208 
1209 	kswapd_run(nid);
1210 	kcompactd_run(nid);
1211 
1212 	writeback_set_ratelimit();
1213 
1214 	memory_notify(MEM_ONLINE, &arg);
1215 	return 0;
1216 
1217 failed_addition:
1218 	pr_debug("online_pages [mem %#010llx-%#010llx] failed\n",
1219 		 (unsigned long long) pfn << PAGE_SHIFT,
1220 		 (((unsigned long long) pfn + nr_pages) << PAGE_SHIFT) - 1);
1221 	memory_notify(MEM_CANCEL_ONLINE, &arg);
1222 	remove_pfn_range_from_zone(zone, pfn, nr_pages);
1223 	return ret;
1224 }
1225 
1226 /* we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG */
hotadd_init_pgdat(int nid)1227 static pg_data_t __ref *hotadd_init_pgdat(int nid)
1228 {
1229 	struct pglist_data *pgdat;
1230 
1231 	/*
1232 	 * NODE_DATA is preallocated (free_area_init) but its internal
1233 	 * state is not allocated completely. Add missing pieces.
1234 	 * Completely offline nodes stay around and they just need
1235 	 * reintialization.
1236 	 */
1237 	pgdat = NODE_DATA(nid);
1238 
1239 	/* init node's zones as empty zones, we don't have any present pages.*/
1240 	free_area_init_core_hotplug(pgdat);
1241 
1242 	/*
1243 	 * The node we allocated has no zone fallback lists. For avoiding
1244 	 * to access not-initialized zonelist, build here.
1245 	 */
1246 	build_all_zonelists(pgdat);
1247 
1248 	return pgdat;
1249 }
1250 
1251 /*
1252  * __try_online_node - online a node if offlined
1253  * @nid: the node ID
1254  * @set_node_online: Whether we want to online the node
1255  * called by cpu_up() to online a node without onlined memory.
1256  *
1257  * Returns:
1258  * 1 -> a new node has been allocated
1259  * 0 -> the node is already online
1260  * -ENOMEM -> the node could not be allocated
1261  */
__try_online_node(int nid,bool set_node_online)1262 static int __try_online_node(int nid, bool set_node_online)
1263 {
1264 	pg_data_t *pgdat;
1265 	int ret = 1;
1266 
1267 	if (node_online(nid))
1268 		return 0;
1269 
1270 	pgdat = hotadd_init_pgdat(nid);
1271 	if (!pgdat) {
1272 		pr_err("Cannot online node %d due to NULL pgdat\n", nid);
1273 		ret = -ENOMEM;
1274 		goto out;
1275 	}
1276 
1277 	if (set_node_online) {
1278 		node_set_online(nid);
1279 		ret = register_one_node(nid);
1280 		BUG_ON(ret);
1281 	}
1282 out:
1283 	return ret;
1284 }
1285 
1286 /*
1287  * Users of this function always want to online/register the node
1288  */
try_online_node(int nid)1289 int try_online_node(int nid)
1290 {
1291 	int ret;
1292 
1293 	mem_hotplug_begin();
1294 	ret =  __try_online_node(nid, true);
1295 	mem_hotplug_done();
1296 	return ret;
1297 }
1298 
check_hotplug_memory_range(u64 start,u64 size)1299 static int check_hotplug_memory_range(u64 start, u64 size)
1300 {
1301 	/* memory range must be block size aligned */
1302 	if (!size || !IS_ALIGNED(start, memory_block_size_bytes()) ||
1303 	    !IS_ALIGNED(size, memory_block_size_bytes())) {
1304 		pr_err("Block size [%#lx] unaligned hotplug range: start %#llx, size %#llx",
1305 		       memory_block_size_bytes(), start, size);
1306 		return -EINVAL;
1307 	}
1308 
1309 	return 0;
1310 }
1311 
online_memory_block(struct memory_block * mem,void * arg)1312 static int online_memory_block(struct memory_block *mem, void *arg)
1313 {
1314 	mem->online_type = mhp_default_online_type;
1315 	return device_online(&mem->dev);
1316 }
1317 
1318 #ifndef arch_supports_memmap_on_memory
arch_supports_memmap_on_memory(unsigned long vmemmap_size)1319 static inline bool arch_supports_memmap_on_memory(unsigned long vmemmap_size)
1320 {
1321 	/*
1322 	 * As default, we want the vmemmap to span a complete PMD such that we
1323 	 * can map the vmemmap using a single PMD if supported by the
1324 	 * architecture.
1325 	 */
1326 	return IS_ALIGNED(vmemmap_size, PMD_SIZE);
1327 }
1328 #endif
1329 
mhp_supports_memmap_on_memory(unsigned long size)1330 static bool mhp_supports_memmap_on_memory(unsigned long size)
1331 {
1332 	unsigned long vmemmap_size = memory_block_memmap_size();
1333 	unsigned long memmap_pages = memory_block_memmap_on_memory_pages();
1334 
1335 	/*
1336 	 * Besides having arch support and the feature enabled at runtime, we
1337 	 * need a few more assumptions to hold true:
1338 	 *
1339 	 * a) We span a single memory block: memory onlining/offlinin;g happens
1340 	 *    in memory block granularity. We don't want the vmemmap of online
1341 	 *    memory blocks to reside on offline memory blocks. In the future,
1342 	 *    we might want to support variable-sized memory blocks to make the
1343 	 *    feature more versatile.
1344 	 *
1345 	 * b) The vmemmap pages span complete PMDs: We don't want vmemmap code
1346 	 *    to populate memory from the altmap for unrelated parts (i.e.,
1347 	 *    other memory blocks)
1348 	 *
1349 	 * c) The vmemmap pages (and thereby the pages that will be exposed to
1350 	 *    the buddy) have to cover full pageblocks: memory onlining/offlining
1351 	 *    code requires applicable ranges to be page-aligned, for example, to
1352 	 *    set the migratetypes properly.
1353 	 *
1354 	 * TODO: Although we have a check here to make sure that vmemmap pages
1355 	 *       fully populate a PMD, it is not the right place to check for
1356 	 *       this. A much better solution involves improving vmemmap code
1357 	 *       to fallback to base pages when trying to populate vmemmap using
1358 	 *       altmap as an alternative source of memory, and we do not exactly
1359 	 *       populate a single PMD.
1360 	 */
1361 	if (!mhp_memmap_on_memory() || size != memory_block_size_bytes())
1362 		return false;
1363 
1364 	/*
1365 	 * Make sure the vmemmap allocation is fully contained
1366 	 * so that we always allocate vmemmap memory from altmap area.
1367 	 */
1368 	if (!IS_ALIGNED(vmemmap_size, PAGE_SIZE))
1369 		return false;
1370 
1371 	/*
1372 	 * start pfn should be pageblock_nr_pages aligned for correctly
1373 	 * setting migrate types
1374 	 */
1375 	if (!pageblock_aligned(memmap_pages))
1376 		return false;
1377 
1378 	if (memmap_pages == PHYS_PFN(memory_block_size_bytes()))
1379 		/* No effective hotplugged memory doesn't make sense. */
1380 		return false;
1381 
1382 	return arch_supports_memmap_on_memory(vmemmap_size);
1383 }
1384 
1385 /*
1386  * NOTE: The caller must call lock_device_hotplug() to serialize hotplug
1387  * and online/offline operations (triggered e.g. by sysfs).
1388  *
1389  * we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG
1390  */
add_memory_resource(int nid,struct resource * res,mhp_t mhp_flags)1391 int __ref add_memory_resource(int nid, struct resource *res, mhp_t mhp_flags)
1392 {
1393 	struct mhp_params params = { .pgprot = pgprot_mhp(PAGE_KERNEL) };
1394 	enum memblock_flags memblock_flags = MEMBLOCK_NONE;
1395 	struct vmem_altmap mhp_altmap = {
1396 		.base_pfn =  PHYS_PFN(res->start),
1397 		.end_pfn  =  PHYS_PFN(res->end),
1398 	};
1399 	struct memory_group *group = NULL;
1400 	u64 start, size;
1401 	bool new_node = false;
1402 	int ret;
1403 
1404 	start = res->start;
1405 	size = resource_size(res);
1406 
1407 	ret = check_hotplug_memory_range(start, size);
1408 	if (ret)
1409 		return ret;
1410 
1411 	if (mhp_flags & MHP_NID_IS_MGID) {
1412 		group = memory_group_find_by_id(nid);
1413 		if (!group)
1414 			return -EINVAL;
1415 		nid = group->nid;
1416 	}
1417 
1418 	if (!node_possible(nid)) {
1419 		WARN(1, "node %d was absent from the node_possible_map\n", nid);
1420 		return -EINVAL;
1421 	}
1422 
1423 	mem_hotplug_begin();
1424 
1425 	if (IS_ENABLED(CONFIG_ARCH_KEEP_MEMBLOCK)) {
1426 		if (res->flags & IORESOURCE_SYSRAM_DRIVER_MANAGED)
1427 			memblock_flags = MEMBLOCK_DRIVER_MANAGED;
1428 		ret = memblock_add_node(start, size, nid, memblock_flags);
1429 		if (ret)
1430 			goto error_mem_hotplug_end;
1431 	}
1432 
1433 	ret = __try_online_node(nid, false);
1434 	if (ret < 0)
1435 		goto error;
1436 	new_node = ret;
1437 
1438 	/*
1439 	 * Self hosted memmap array
1440 	 */
1441 	if (mhp_flags & MHP_MEMMAP_ON_MEMORY) {
1442 		if (mhp_supports_memmap_on_memory(size)) {
1443 			mhp_altmap.free = memory_block_memmap_on_memory_pages();
1444 			params.altmap = kmalloc(sizeof(struct vmem_altmap), GFP_KERNEL);
1445 			if (!params.altmap) {
1446 				ret = -ENOMEM;
1447 				goto error;
1448 			}
1449 
1450 			memcpy(params.altmap, &mhp_altmap, sizeof(mhp_altmap));
1451 		}
1452 		/* fallback to not using altmap  */
1453 	}
1454 
1455 	/* call arch's memory hotadd */
1456 	ret = arch_add_memory(nid, start, size, &params);
1457 	if (ret < 0)
1458 		goto error_free;
1459 
1460 	/* create memory block devices after memory was added */
1461 	ret = create_memory_block_devices(start, size, params.altmap, group);
1462 	if (ret) {
1463 		arch_remove_memory(start, size, params.altmap);
1464 		goto error_free;
1465 	}
1466 
1467 	if (new_node) {
1468 		/* If sysfs file of new node can't be created, cpu on the node
1469 		 * can't be hot-added. There is no rollback way now.
1470 		 * So, check by BUG_ON() to catch it reluctantly..
1471 		 * We online node here. We can't roll back from here.
1472 		 */
1473 		node_set_online(nid);
1474 		ret = __register_one_node(nid);
1475 		BUG_ON(ret);
1476 	}
1477 
1478 	register_memory_blocks_under_node(nid, PFN_DOWN(start),
1479 					  PFN_UP(start + size - 1),
1480 					  MEMINIT_HOTPLUG);
1481 
1482 	/* create new memmap entry */
1483 	if (!strcmp(res->name, "System RAM"))
1484 		firmware_map_add_hotplug(start, start + size, "System RAM");
1485 
1486 	/* device_online() will take the lock when calling online_pages() */
1487 	mem_hotplug_done();
1488 
1489 	/*
1490 	 * In case we're allowed to merge the resource, flag it and trigger
1491 	 * merging now that adding succeeded.
1492 	 */
1493 	if (mhp_flags & MHP_MERGE_RESOURCE)
1494 		merge_system_ram_resource(res);
1495 
1496 	/* online pages if requested */
1497 	if (mhp_default_online_type != MMOP_OFFLINE)
1498 		walk_memory_blocks(start, size, NULL, online_memory_block);
1499 
1500 	return ret;
1501 error_free:
1502 	kfree(params.altmap);
1503 error:
1504 	if (IS_ENABLED(CONFIG_ARCH_KEEP_MEMBLOCK))
1505 		memblock_remove(start, size);
1506 error_mem_hotplug_end:
1507 	mem_hotplug_done();
1508 	return ret;
1509 }
1510 
1511 /* requires device_hotplug_lock, see add_memory_resource() */
__add_memory(int nid,u64 start,u64 size,mhp_t mhp_flags)1512 int __ref __add_memory(int nid, u64 start, u64 size, mhp_t mhp_flags)
1513 {
1514 	struct resource *res;
1515 	int ret;
1516 
1517 	res = register_memory_resource(start, size, "System RAM");
1518 	if (IS_ERR(res))
1519 		return PTR_ERR(res);
1520 
1521 	ret = add_memory_resource(nid, res, mhp_flags);
1522 	if (ret < 0)
1523 		release_memory_resource(res);
1524 	return ret;
1525 }
1526 
add_memory(int nid,u64 start,u64 size,mhp_t mhp_flags)1527 int add_memory(int nid, u64 start, u64 size, mhp_t mhp_flags)
1528 {
1529 	int rc;
1530 
1531 	lock_device_hotplug();
1532 	rc = __add_memory(nid, start, size, mhp_flags);
1533 	unlock_device_hotplug();
1534 
1535 	return rc;
1536 }
1537 EXPORT_SYMBOL_GPL(add_memory);
1538 
1539 /*
1540  * Add special, driver-managed memory to the system as system RAM. Such
1541  * memory is not exposed via the raw firmware-provided memmap as system
1542  * RAM, instead, it is detected and added by a driver - during cold boot,
1543  * after a reboot, and after kexec.
1544  *
1545  * Reasons why this memory should not be used for the initial memmap of a
1546  * kexec kernel or for placing kexec images:
1547  * - The booting kernel is in charge of determining how this memory will be
1548  *   used (e.g., use persistent memory as system RAM)
1549  * - Coordination with a hypervisor is required before this memory
1550  *   can be used (e.g., inaccessible parts).
1551  *
1552  * For this memory, no entries in /sys/firmware/memmap ("raw firmware-provided
1553  * memory map") are created. Also, the created memory resource is flagged
1554  * with IORESOURCE_SYSRAM_DRIVER_MANAGED, so in-kernel users can special-case
1555  * this memory as well (esp., not place kexec images onto it).
1556  *
1557  * The resource_name (visible via /proc/iomem) has to have the format
1558  * "System RAM ($DRIVER)".
1559  */
add_memory_driver_managed(int nid,u64 start,u64 size,const char * resource_name,mhp_t mhp_flags)1560 int add_memory_driver_managed(int nid, u64 start, u64 size,
1561 			      const char *resource_name, mhp_t mhp_flags)
1562 {
1563 	struct resource *res;
1564 	int rc;
1565 
1566 	if (!resource_name ||
1567 	    strstr(resource_name, "System RAM (") != resource_name ||
1568 	    resource_name[strlen(resource_name) - 1] != ')')
1569 		return -EINVAL;
1570 
1571 	lock_device_hotplug();
1572 
1573 	res = register_memory_resource(start, size, resource_name);
1574 	if (IS_ERR(res)) {
1575 		rc = PTR_ERR(res);
1576 		goto out_unlock;
1577 	}
1578 
1579 	rc = add_memory_resource(nid, res, mhp_flags);
1580 	if (rc < 0)
1581 		release_memory_resource(res);
1582 
1583 out_unlock:
1584 	unlock_device_hotplug();
1585 	return rc;
1586 }
1587 EXPORT_SYMBOL_GPL(add_memory_driver_managed);
1588 
1589 /*
1590  * Platforms should define arch_get_mappable_range() that provides
1591  * maximum possible addressable physical memory range for which the
1592  * linear mapping could be created. The platform returned address
1593  * range must adhere to these following semantics.
1594  *
1595  * - range.start <= range.end
1596  * - Range includes both end points [range.start..range.end]
1597  *
1598  * There is also a fallback definition provided here, allowing the
1599  * entire possible physical address range in case any platform does
1600  * not define arch_get_mappable_range().
1601  */
arch_get_mappable_range(void)1602 struct range __weak arch_get_mappable_range(void)
1603 {
1604 	struct range mhp_range = {
1605 		.start = 0UL,
1606 		.end = -1ULL,
1607 	};
1608 	return mhp_range;
1609 }
1610 
mhp_get_pluggable_range(bool need_mapping)1611 struct range mhp_get_pluggable_range(bool need_mapping)
1612 {
1613 	const u64 max_phys = PHYSMEM_END;
1614 	struct range mhp_range;
1615 
1616 	if (need_mapping) {
1617 		mhp_range = arch_get_mappable_range();
1618 		if (mhp_range.start > max_phys) {
1619 			mhp_range.start = 0;
1620 			mhp_range.end = 0;
1621 		}
1622 		mhp_range.end = min_t(u64, mhp_range.end, max_phys);
1623 	} else {
1624 		mhp_range.start = 0;
1625 		mhp_range.end = max_phys;
1626 	}
1627 	return mhp_range;
1628 }
1629 EXPORT_SYMBOL_GPL(mhp_get_pluggable_range);
1630 
mhp_range_allowed(u64 start,u64 size,bool need_mapping)1631 bool mhp_range_allowed(u64 start, u64 size, bool need_mapping)
1632 {
1633 	struct range mhp_range = mhp_get_pluggable_range(need_mapping);
1634 	u64 end = start + size;
1635 
1636 	if (start < end && start >= mhp_range.start && (end - 1) <= mhp_range.end)
1637 		return true;
1638 
1639 	pr_warn("Hotplug memory [%#llx-%#llx] exceeds maximum addressable range [%#llx-%#llx]\n",
1640 		start, end, mhp_range.start, mhp_range.end);
1641 	return false;
1642 }
1643 
1644 #ifdef CONFIG_MEMORY_HOTREMOVE
1645 /*
1646  * Scan pfn range [start,end) to find movable/migratable pages (LRU pages,
1647  * non-lru movable pages and hugepages). Will skip over most unmovable
1648  * pages (esp., pages that can be skipped when offlining), but bail out on
1649  * definitely unmovable pages.
1650  *
1651  * Returns:
1652  *	0 in case a movable page is found and movable_pfn was updated.
1653  *	-ENOENT in case no movable page was found.
1654  *	-EBUSY in case a definitely unmovable page was found.
1655  */
scan_movable_pages(unsigned long start,unsigned long end,unsigned long * movable_pfn)1656 static int scan_movable_pages(unsigned long start, unsigned long end,
1657 			      unsigned long *movable_pfn)
1658 {
1659 	unsigned long pfn;
1660 
1661 	for (pfn = start; pfn < end; pfn++) {
1662 		struct page *page, *head;
1663 		unsigned long skip;
1664 
1665 		if (!pfn_valid(pfn))
1666 			continue;
1667 		page = pfn_to_page(pfn);
1668 		if (PageLRU(page))
1669 			goto found;
1670 		if (__PageMovable(page))
1671 			goto found;
1672 
1673 		/*
1674 		 * PageOffline() pages that are not marked __PageMovable() and
1675 		 * have a reference count > 0 (after MEM_GOING_OFFLINE) are
1676 		 * definitely unmovable. If their reference count would be 0,
1677 		 * they could at least be skipped when offlining memory.
1678 		 */
1679 		if (PageOffline(page) && page_count(page))
1680 			return -EBUSY;
1681 
1682 		if (!PageHuge(page))
1683 			continue;
1684 		head = compound_head(page);
1685 		/*
1686 		 * This test is racy as we hold no reference or lock.  The
1687 		 * hugetlb page could have been free'ed and head is no longer
1688 		 * a hugetlb page before the following check.  In such unlikely
1689 		 * cases false positives and negatives are possible.  Calling
1690 		 * code must deal with these scenarios.
1691 		 */
1692 		if (HPageMigratable(head))
1693 			goto found;
1694 		skip = compound_nr(head) - (pfn - page_to_pfn(head));
1695 		pfn += skip - 1;
1696 	}
1697 	return -ENOENT;
1698 found:
1699 	*movable_pfn = pfn;
1700 	return 0;
1701 }
1702 
do_migrate_range(unsigned long start_pfn,unsigned long end_pfn)1703 static void do_migrate_range(unsigned long start_pfn, unsigned long end_pfn)
1704 {
1705 	unsigned long pfn;
1706 	struct page *page, *head;
1707 	LIST_HEAD(source);
1708 	static DEFINE_RATELIMIT_STATE(migrate_rs, DEFAULT_RATELIMIT_INTERVAL,
1709 				      DEFAULT_RATELIMIT_BURST);
1710 
1711 	for (pfn = start_pfn; pfn < end_pfn; pfn++) {
1712 		struct folio *folio;
1713 		bool isolated;
1714 
1715 		if (!pfn_valid(pfn))
1716 			continue;
1717 		page = pfn_to_page(pfn);
1718 		folio = page_folio(page);
1719 		head = &folio->page;
1720 
1721 		if (PageHuge(page)) {
1722 			pfn = page_to_pfn(head) + compound_nr(head) - 1;
1723 			isolate_hugetlb(folio, &source);
1724 			continue;
1725 		} else if (PageTransHuge(page))
1726 			pfn = page_to_pfn(head) + thp_nr_pages(page) - 1;
1727 
1728 		/*
1729 		 * HWPoison pages have elevated reference counts so the migration would
1730 		 * fail on them. It also doesn't make any sense to migrate them in the
1731 		 * first place. Still try to unmap such a page in case it is still mapped
1732 		 * (e.g. current hwpoison implementation doesn't unmap KSM pages but keep
1733 		 * the unmap as the catch all safety net).
1734 		 */
1735 		if (PageHWPoison(page)) {
1736 			if (WARN_ON(folio_test_lru(folio)))
1737 				folio_isolate_lru(folio);
1738 			if (folio_mapped(folio))
1739 				try_to_unmap(folio, TTU_IGNORE_MLOCK);
1740 			continue;
1741 		}
1742 
1743 		if (!get_page_unless_zero(page))
1744 			continue;
1745 		/*
1746 		 * We can skip free pages. And we can deal with pages on
1747 		 * LRU and non-lru movable pages.
1748 		 */
1749 		if (PageLRU(page))
1750 			isolated = isolate_lru_page(page);
1751 		else
1752 			isolated = isolate_movable_page(page, ISOLATE_UNEVICTABLE);
1753 		if (isolated) {
1754 			list_add_tail(&page->lru, &source);
1755 			if (!__PageMovable(page))
1756 				inc_node_page_state(page, NR_ISOLATED_ANON +
1757 						    page_is_file_lru(page));
1758 
1759 		} else {
1760 			if (__ratelimit(&migrate_rs)) {
1761 				pr_warn("failed to isolate pfn %lx\n", pfn);
1762 				dump_page(page, "isolation failed");
1763 			}
1764 		}
1765 		put_page(page);
1766 	}
1767 	if (!list_empty(&source)) {
1768 		nodemask_t nmask = node_states[N_MEMORY];
1769 		struct migration_target_control mtc = {
1770 			.nmask = &nmask,
1771 			.gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
1772 		};
1773 		int ret;
1774 
1775 		/*
1776 		 * We have checked that migration range is on a single zone so
1777 		 * we can use the nid of the first page to all the others.
1778 		 */
1779 		mtc.nid = page_to_nid(list_first_entry(&source, struct page, lru));
1780 
1781 		/*
1782 		 * try to allocate from a different node but reuse this node
1783 		 * if there are no other online nodes to be used (e.g. we are
1784 		 * offlining a part of the only existing node)
1785 		 */
1786 		node_clear(mtc.nid, nmask);
1787 		if (nodes_empty(nmask))
1788 			node_set(mtc.nid, nmask);
1789 		ret = migrate_pages(&source, alloc_migration_target, NULL,
1790 			(unsigned long)&mtc, MIGRATE_SYNC, MR_MEMORY_HOTPLUG, NULL);
1791 		if (ret) {
1792 			list_for_each_entry(page, &source, lru) {
1793 				if (__ratelimit(&migrate_rs)) {
1794 					pr_warn("migrating pfn %lx failed ret:%d\n",
1795 						page_to_pfn(page), ret);
1796 					dump_page(page, "migration failure");
1797 				}
1798 			}
1799 			putback_movable_pages(&source);
1800 		}
1801 	}
1802 }
1803 
cmdline_parse_movable_node(char * p)1804 static int __init cmdline_parse_movable_node(char *p)
1805 {
1806 	movable_node_enabled = true;
1807 	return 0;
1808 }
1809 early_param("movable_node", cmdline_parse_movable_node);
1810 
1811 /* check which state of node_states will be changed when offline memory */
node_states_check_changes_offline(unsigned long nr_pages,struct zone * zone,struct memory_notify * arg)1812 static void node_states_check_changes_offline(unsigned long nr_pages,
1813 		struct zone *zone, struct memory_notify *arg)
1814 {
1815 	struct pglist_data *pgdat = zone->zone_pgdat;
1816 	unsigned long present_pages = 0;
1817 	enum zone_type zt;
1818 
1819 	arg->status_change_nid = NUMA_NO_NODE;
1820 	arg->status_change_nid_normal = NUMA_NO_NODE;
1821 
1822 	/*
1823 	 * Check whether node_states[N_NORMAL_MEMORY] will be changed.
1824 	 * If the memory to be offline is within the range
1825 	 * [0..ZONE_NORMAL], and it is the last present memory there,
1826 	 * the zones in that range will become empty after the offlining,
1827 	 * thus we can determine that we need to clear the node from
1828 	 * node_states[N_NORMAL_MEMORY].
1829 	 */
1830 	for (zt = 0; zt <= ZONE_NORMAL; zt++)
1831 		present_pages += pgdat->node_zones[zt].present_pages;
1832 	if (zone_idx(zone) <= ZONE_NORMAL && nr_pages >= present_pages)
1833 		arg->status_change_nid_normal = zone_to_nid(zone);
1834 
1835 	/*
1836 	 * We have accounted the pages from [0..ZONE_NORMAL); ZONE_HIGHMEM
1837 	 * does not apply as we don't support 32bit.
1838 	 * Here we count the possible pages from ZONE_MOVABLE.
1839 	 * If after having accounted all the pages, we see that the nr_pages
1840 	 * to be offlined is over or equal to the accounted pages,
1841 	 * we know that the node will become empty, and so, we can clear
1842 	 * it for N_MEMORY as well.
1843 	 */
1844 	present_pages += pgdat->node_zones[ZONE_MOVABLE].present_pages;
1845 
1846 	if (nr_pages >= present_pages)
1847 		arg->status_change_nid = zone_to_nid(zone);
1848 }
1849 
node_states_clear_node(int node,struct memory_notify * arg)1850 static void node_states_clear_node(int node, struct memory_notify *arg)
1851 {
1852 	if (arg->status_change_nid_normal >= 0)
1853 		node_clear_state(node, N_NORMAL_MEMORY);
1854 
1855 	if (arg->status_change_nid >= 0)
1856 		node_clear_state(node, N_MEMORY);
1857 }
1858 
count_system_ram_pages_cb(unsigned long start_pfn,unsigned long nr_pages,void * data)1859 static int count_system_ram_pages_cb(unsigned long start_pfn,
1860 				     unsigned long nr_pages, void *data)
1861 {
1862 	unsigned long *nr_system_ram_pages = data;
1863 
1864 	*nr_system_ram_pages += nr_pages;
1865 	return 0;
1866 }
1867 
1868 /*
1869  * Must be called with mem_hotplug_lock in write mode.
1870  */
offline_pages(unsigned long start_pfn,unsigned long nr_pages,struct zone * zone,struct memory_group * group)1871 int __ref offline_pages(unsigned long start_pfn, unsigned long nr_pages,
1872 			struct zone *zone, struct memory_group *group)
1873 {
1874 	const unsigned long end_pfn = start_pfn + nr_pages;
1875 	unsigned long pfn, system_ram_pages = 0;
1876 	const int node = zone_to_nid(zone);
1877 	unsigned long flags;
1878 	struct memory_notify arg;
1879 	char *reason;
1880 	int ret;
1881 
1882 	/*
1883 	 * {on,off}lining is constrained to full memory sections (or more
1884 	 * precisely to memory blocks from the user space POV).
1885 	 * memmap_on_memory is an exception because it reserves initial part
1886 	 * of the physical memory space for vmemmaps. That space is pageblock
1887 	 * aligned.
1888 	 */
1889 	if (WARN_ON_ONCE(!nr_pages || !pageblock_aligned(start_pfn) ||
1890 			 !IS_ALIGNED(start_pfn + nr_pages, PAGES_PER_SECTION)))
1891 		return -EINVAL;
1892 
1893 	/*
1894 	 * Don't allow to offline memory blocks that contain holes.
1895 	 * Consequently, memory blocks with holes can never get onlined
1896 	 * via the hotplug path - online_pages() - as hotplugged memory has
1897 	 * no holes. This way, we e.g., don't have to worry about marking
1898 	 * memory holes PG_reserved, don't need pfn_valid() checks, and can
1899 	 * avoid using walk_system_ram_range() later.
1900 	 */
1901 	walk_system_ram_range(start_pfn, nr_pages, &system_ram_pages,
1902 			      count_system_ram_pages_cb);
1903 	if (system_ram_pages != nr_pages) {
1904 		ret = -EINVAL;
1905 		reason = "memory holes";
1906 		goto failed_removal;
1907 	}
1908 
1909 	/*
1910 	 * We only support offlining of memory blocks managed by a single zone,
1911 	 * checked by calling code. This is just a sanity check that we might
1912 	 * want to remove in the future.
1913 	 */
1914 	if (WARN_ON_ONCE(page_zone(pfn_to_page(start_pfn)) != zone ||
1915 			 page_zone(pfn_to_page(end_pfn - 1)) != zone)) {
1916 		ret = -EINVAL;
1917 		reason = "multizone range";
1918 		goto failed_removal;
1919 	}
1920 
1921 	/*
1922 	 * Disable pcplists so that page isolation cannot race with freeing
1923 	 * in a way that pages from isolated pageblock are left on pcplists.
1924 	 */
1925 	zone_pcp_disable(zone);
1926 	lru_cache_disable();
1927 
1928 	/* set above range as isolated */
1929 	ret = start_isolate_page_range(start_pfn, end_pfn,
1930 				       MIGRATE_MOVABLE,
1931 				       MEMORY_OFFLINE | REPORT_FAILURE,
1932 				       GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL);
1933 	if (ret) {
1934 		reason = "failure to isolate range";
1935 		goto failed_removal_pcplists_disabled;
1936 	}
1937 
1938 	arg.start_pfn = start_pfn;
1939 	arg.nr_pages = nr_pages;
1940 	node_states_check_changes_offline(nr_pages, zone, &arg);
1941 
1942 	ret = memory_notify(MEM_GOING_OFFLINE, &arg);
1943 	ret = notifier_to_errno(ret);
1944 	if (ret) {
1945 		reason = "notifier failure";
1946 		goto failed_removal_isolated;
1947 	}
1948 
1949 	do {
1950 		pfn = start_pfn;
1951 		do {
1952 			/*
1953 			 * Historically we always checked for any signal and
1954 			 * can't limit it to fatal signals without eventually
1955 			 * breaking user space.
1956 			 */
1957 			if (signal_pending(current)) {
1958 				ret = -EINTR;
1959 				reason = "signal backoff";
1960 				goto failed_removal_isolated;
1961 			}
1962 
1963 			cond_resched();
1964 
1965 			ret = scan_movable_pages(pfn, end_pfn, &pfn);
1966 			if (!ret) {
1967 				/*
1968 				 * TODO: fatal migration failures should bail
1969 				 * out
1970 				 */
1971 				do_migrate_range(pfn, end_pfn);
1972 			}
1973 		} while (!ret);
1974 
1975 		if (ret != -ENOENT) {
1976 			reason = "unmovable page";
1977 			goto failed_removal_isolated;
1978 		}
1979 
1980 		/*
1981 		 * Dissolve free hugepages in the memory block before doing
1982 		 * offlining actually in order to make hugetlbfs's object
1983 		 * counting consistent.
1984 		 */
1985 		ret = dissolve_free_huge_pages(start_pfn, end_pfn);
1986 		if (ret) {
1987 			reason = "failure to dissolve huge pages";
1988 			goto failed_removal_isolated;
1989 		}
1990 
1991 		ret = test_pages_isolated(start_pfn, end_pfn, MEMORY_OFFLINE);
1992 
1993 	} while (ret);
1994 
1995 	/* Mark all sections offline and remove free pages from the buddy. */
1996 	__offline_isolated_pages(start_pfn, end_pfn);
1997 	pr_debug("Offlined Pages %ld\n", nr_pages);
1998 
1999 	/*
2000 	 * The memory sections are marked offline, and the pageblock flags
2001 	 * effectively stale; nobody should be touching them. Fixup the number
2002 	 * of isolated pageblocks, memory onlining will properly revert this.
2003 	 */
2004 	spin_lock_irqsave(&zone->lock, flags);
2005 	zone->nr_isolate_pageblock -= nr_pages / pageblock_nr_pages;
2006 	spin_unlock_irqrestore(&zone->lock, flags);
2007 
2008 	lru_cache_enable();
2009 	zone_pcp_enable(zone);
2010 
2011 	/* removal success */
2012 	adjust_managed_page_count(pfn_to_page(start_pfn), -nr_pages);
2013 	adjust_present_page_count(pfn_to_page(start_pfn), group, -nr_pages);
2014 
2015 	/* reinitialise watermarks and update pcp limits */
2016 	init_per_zone_wmark_min();
2017 
2018 	if (!populated_zone(zone)) {
2019 		zone_pcp_reset(zone);
2020 		build_all_zonelists(NULL);
2021 	}
2022 
2023 	node_states_clear_node(node, &arg);
2024 	if (arg.status_change_nid >= 0) {
2025 		kcompactd_stop(node);
2026 		kswapd_stop(node);
2027 	}
2028 
2029 	writeback_set_ratelimit();
2030 
2031 	memory_notify(MEM_OFFLINE, &arg);
2032 	remove_pfn_range_from_zone(zone, start_pfn, nr_pages);
2033 	return 0;
2034 
2035 failed_removal_isolated:
2036 	/* pushback to free area */
2037 	undo_isolate_page_range(start_pfn, end_pfn, MIGRATE_MOVABLE);
2038 	memory_notify(MEM_CANCEL_OFFLINE, &arg);
2039 failed_removal_pcplists_disabled:
2040 	lru_cache_enable();
2041 	zone_pcp_enable(zone);
2042 failed_removal:
2043 	pr_debug("memory offlining [mem %#010llx-%#010llx] failed due to %s\n",
2044 		 (unsigned long long) start_pfn << PAGE_SHIFT,
2045 		 ((unsigned long long) end_pfn << PAGE_SHIFT) - 1,
2046 		 reason);
2047 	return ret;
2048 }
2049 
check_memblock_offlined_cb(struct memory_block * mem,void * arg)2050 static int check_memblock_offlined_cb(struct memory_block *mem, void *arg)
2051 {
2052 	int *nid = arg;
2053 
2054 	*nid = mem->nid;
2055 	if (unlikely(mem->state != MEM_OFFLINE)) {
2056 		phys_addr_t beginpa, endpa;
2057 
2058 		beginpa = PFN_PHYS(section_nr_to_pfn(mem->start_section_nr));
2059 		endpa = beginpa + memory_block_size_bytes() - 1;
2060 		pr_warn("removing memory fails, because memory [%pa-%pa] is onlined\n",
2061 			&beginpa, &endpa);
2062 
2063 		return -EBUSY;
2064 	}
2065 	return 0;
2066 }
2067 
test_has_altmap_cb(struct memory_block * mem,void * arg)2068 static int test_has_altmap_cb(struct memory_block *mem, void *arg)
2069 {
2070 	struct memory_block **mem_ptr = (struct memory_block **)arg;
2071 	/*
2072 	 * return the memblock if we have altmap
2073 	 * and break callback.
2074 	 */
2075 	if (mem->altmap) {
2076 		*mem_ptr = mem;
2077 		return 1;
2078 	}
2079 	return 0;
2080 }
2081 
check_cpu_on_node(int nid)2082 static int check_cpu_on_node(int nid)
2083 {
2084 	int cpu;
2085 
2086 	for_each_present_cpu(cpu) {
2087 		if (cpu_to_node(cpu) == nid)
2088 			/*
2089 			 * the cpu on this node isn't removed, and we can't
2090 			 * offline this node.
2091 			 */
2092 			return -EBUSY;
2093 	}
2094 
2095 	return 0;
2096 }
2097 
check_no_memblock_for_node_cb(struct memory_block * mem,void * arg)2098 static int check_no_memblock_for_node_cb(struct memory_block *mem, void *arg)
2099 {
2100 	int nid = *(int *)arg;
2101 
2102 	/*
2103 	 * If a memory block belongs to multiple nodes, the stored nid is not
2104 	 * reliable. However, such blocks are always online (e.g., cannot get
2105 	 * offlined) and, therefore, are still spanned by the node.
2106 	 */
2107 	return mem->nid == nid ? -EEXIST : 0;
2108 }
2109 
2110 /**
2111  * try_offline_node
2112  * @nid: the node ID
2113  *
2114  * Offline a node if all memory sections and cpus of the node are removed.
2115  *
2116  * NOTE: The caller must call lock_device_hotplug() to serialize hotplug
2117  * and online/offline operations before this call.
2118  */
try_offline_node(int nid)2119 void try_offline_node(int nid)
2120 {
2121 	int rc;
2122 
2123 	/*
2124 	 * If the node still spans pages (especially ZONE_DEVICE), don't
2125 	 * offline it. A node spans memory after move_pfn_range_to_zone(),
2126 	 * e.g., after the memory block was onlined.
2127 	 */
2128 	if (node_spanned_pages(nid))
2129 		return;
2130 
2131 	/*
2132 	 * Especially offline memory blocks might not be spanned by the
2133 	 * node. They will get spanned by the node once they get onlined.
2134 	 * However, they link to the node in sysfs and can get onlined later.
2135 	 */
2136 	rc = for_each_memory_block(&nid, check_no_memblock_for_node_cb);
2137 	if (rc)
2138 		return;
2139 
2140 	if (check_cpu_on_node(nid))
2141 		return;
2142 
2143 	/*
2144 	 * all memory/cpu of this node are removed, we can offline this
2145 	 * node now.
2146 	 */
2147 	node_set_offline(nid);
2148 	unregister_one_node(nid);
2149 }
2150 EXPORT_SYMBOL(try_offline_node);
2151 
try_remove_memory(u64 start,u64 size)2152 static int __ref try_remove_memory(u64 start, u64 size)
2153 {
2154 	struct memory_block *mem;
2155 	int rc = 0, nid = NUMA_NO_NODE;
2156 	struct vmem_altmap *altmap = NULL;
2157 
2158 	BUG_ON(check_hotplug_memory_range(start, size));
2159 
2160 	/*
2161 	 * All memory blocks must be offlined before removing memory.  Check
2162 	 * whether all memory blocks in question are offline and return error
2163 	 * if this is not the case.
2164 	 *
2165 	 * While at it, determine the nid. Note that if we'd have mixed nodes,
2166 	 * we'd only try to offline the last determined one -- which is good
2167 	 * enough for the cases we care about.
2168 	 */
2169 	rc = walk_memory_blocks(start, size, &nid, check_memblock_offlined_cb);
2170 	if (rc)
2171 		return rc;
2172 
2173 	/*
2174 	 * We only support removing memory added with MHP_MEMMAP_ON_MEMORY in
2175 	 * the same granularity it was added - a single memory block.
2176 	 */
2177 	if (mhp_memmap_on_memory()) {
2178 		rc = walk_memory_blocks(start, size, &mem, test_has_altmap_cb);
2179 		if (rc) {
2180 			if (size != memory_block_size_bytes()) {
2181 				pr_warn("Refuse to remove %#llx - %#llx,"
2182 					"wrong granularity\n",
2183 					start, start + size);
2184 				return -EINVAL;
2185 			}
2186 			altmap = mem->altmap;
2187 			/*
2188 			 * Mark altmap NULL so that we can add a debug
2189 			 * check on memblock free.
2190 			 */
2191 			mem->altmap = NULL;
2192 		}
2193 	}
2194 
2195 	/* remove memmap entry */
2196 	firmware_map_remove(start, start + size, "System RAM");
2197 
2198 	/*
2199 	 * Memory block device removal under the device_hotplug_lock is
2200 	 * a barrier against racing online attempts.
2201 	 */
2202 	remove_memory_block_devices(start, size);
2203 
2204 	mem_hotplug_begin();
2205 
2206 	arch_remove_memory(start, size, altmap);
2207 
2208 	/* Verify that all vmemmap pages have actually been freed. */
2209 	if (altmap) {
2210 		WARN(altmap->alloc, "Altmap not fully unmapped");
2211 		kfree(altmap);
2212 	}
2213 
2214 	if (IS_ENABLED(CONFIG_ARCH_KEEP_MEMBLOCK)) {
2215 		memblock_phys_free(start, size);
2216 		memblock_remove(start, size);
2217 	}
2218 
2219 	release_mem_region_adjustable(start, size);
2220 
2221 	if (nid != NUMA_NO_NODE)
2222 		try_offline_node(nid);
2223 
2224 	mem_hotplug_done();
2225 	return 0;
2226 }
2227 
2228 /**
2229  * __remove_memory - Remove memory if every memory block is offline
2230  * @start: physical address of the region to remove
2231  * @size: size of the region to remove
2232  *
2233  * NOTE: The caller must call lock_device_hotplug() to serialize hotplug
2234  * and online/offline operations before this call, as required by
2235  * try_offline_node().
2236  */
__remove_memory(u64 start,u64 size)2237 void __remove_memory(u64 start, u64 size)
2238 {
2239 
2240 	/*
2241 	 * trigger BUG() if some memory is not offlined prior to calling this
2242 	 * function
2243 	 */
2244 	if (try_remove_memory(start, size))
2245 		BUG();
2246 }
2247 
2248 /*
2249  * Remove memory if every memory block is offline, otherwise return -EBUSY is
2250  * some memory is not offline
2251  */
remove_memory(u64 start,u64 size)2252 int remove_memory(u64 start, u64 size)
2253 {
2254 	int rc;
2255 
2256 	lock_device_hotplug();
2257 	rc = try_remove_memory(start, size);
2258 	unlock_device_hotplug();
2259 
2260 	return rc;
2261 }
2262 EXPORT_SYMBOL_GPL(remove_memory);
2263 
try_offline_memory_block(struct memory_block * mem,void * arg)2264 static int try_offline_memory_block(struct memory_block *mem, void *arg)
2265 {
2266 	uint8_t online_type = MMOP_ONLINE_KERNEL;
2267 	uint8_t **online_types = arg;
2268 	struct page *page;
2269 	int rc;
2270 
2271 	/*
2272 	 * Sense the online_type via the zone of the memory block. Offlining
2273 	 * with multiple zones within one memory block will be rejected
2274 	 * by offlining code ... so we don't care about that.
2275 	 */
2276 	page = pfn_to_online_page(section_nr_to_pfn(mem->start_section_nr));
2277 	if (page && zone_idx(page_zone(page)) == ZONE_MOVABLE)
2278 		online_type = MMOP_ONLINE_MOVABLE;
2279 
2280 	rc = device_offline(&mem->dev);
2281 	/*
2282 	 * Default is MMOP_OFFLINE - change it only if offlining succeeded,
2283 	 * so try_reonline_memory_block() can do the right thing.
2284 	 */
2285 	if (!rc)
2286 		**online_types = online_type;
2287 
2288 	(*online_types)++;
2289 	/* Ignore if already offline. */
2290 	return rc < 0 ? rc : 0;
2291 }
2292 
try_reonline_memory_block(struct memory_block * mem,void * arg)2293 static int try_reonline_memory_block(struct memory_block *mem, void *arg)
2294 {
2295 	uint8_t **online_types = arg;
2296 	int rc;
2297 
2298 	if (**online_types != MMOP_OFFLINE) {
2299 		mem->online_type = **online_types;
2300 		rc = device_online(&mem->dev);
2301 		if (rc < 0)
2302 			pr_warn("%s: Failed to re-online memory: %d",
2303 				__func__, rc);
2304 	}
2305 
2306 	/* Continue processing all remaining memory blocks. */
2307 	(*online_types)++;
2308 	return 0;
2309 }
2310 
2311 /*
2312  * Try to offline and remove memory. Might take a long time to finish in case
2313  * memory is still in use. Primarily useful for memory devices that logically
2314  * unplugged all memory (so it's no longer in use) and want to offline + remove
2315  * that memory.
2316  */
offline_and_remove_memory(u64 start,u64 size)2317 int offline_and_remove_memory(u64 start, u64 size)
2318 {
2319 	const unsigned long mb_count = size / memory_block_size_bytes();
2320 	uint8_t *online_types, *tmp;
2321 	int rc;
2322 
2323 	if (!IS_ALIGNED(start, memory_block_size_bytes()) ||
2324 	    !IS_ALIGNED(size, memory_block_size_bytes()) || !size)
2325 		return -EINVAL;
2326 
2327 	/*
2328 	 * We'll remember the old online type of each memory block, so we can
2329 	 * try to revert whatever we did when offlining one memory block fails
2330 	 * after offlining some others succeeded.
2331 	 */
2332 	online_types = kmalloc_array(mb_count, sizeof(*online_types),
2333 				     GFP_KERNEL);
2334 	if (!online_types)
2335 		return -ENOMEM;
2336 	/*
2337 	 * Initialize all states to MMOP_OFFLINE, so when we abort processing in
2338 	 * try_offline_memory_block(), we'll skip all unprocessed blocks in
2339 	 * try_reonline_memory_block().
2340 	 */
2341 	memset(online_types, MMOP_OFFLINE, mb_count);
2342 
2343 	lock_device_hotplug();
2344 
2345 	tmp = online_types;
2346 	rc = walk_memory_blocks(start, size, &tmp, try_offline_memory_block);
2347 
2348 	/*
2349 	 * In case we succeeded to offline all memory, remove it.
2350 	 * This cannot fail as it cannot get onlined in the meantime.
2351 	 */
2352 	if (!rc) {
2353 		rc = try_remove_memory(start, size);
2354 		if (rc)
2355 			pr_err("%s: Failed to remove memory: %d", __func__, rc);
2356 	}
2357 
2358 	/*
2359 	 * Rollback what we did. While memory onlining might theoretically fail
2360 	 * (nacked by a notifier), it barely ever happens.
2361 	 */
2362 	if (rc) {
2363 		tmp = online_types;
2364 		walk_memory_blocks(start, size, &tmp,
2365 				   try_reonline_memory_block);
2366 	}
2367 	unlock_device_hotplug();
2368 
2369 	kfree(online_types);
2370 	return rc;
2371 }
2372 EXPORT_SYMBOL_GPL(offline_and_remove_memory);
2373 #endif /* CONFIG_MEMORY_HOTREMOVE */
2374