xref: /openbmc/linux/mm/sparse.c (revision 5a170e9e)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * sparse memory mappings.
4  */
5 #include <linux/mm.h>
6 #include <linux/slab.h>
7 #include <linux/mmzone.h>
8 #include <linux/memblock.h>
9 #include <linux/compiler.h>
10 #include <linux/highmem.h>
11 #include <linux/export.h>
12 #include <linux/spinlock.h>
13 #include <linux/vmalloc.h>
14 
15 #include "internal.h"
16 #include <asm/dma.h>
17 #include <asm/pgalloc.h>
18 #include <asm/pgtable.h>
19 
20 /*
21  * Permanent SPARSEMEM data:
22  *
23  * 1) mem_section	- memory sections, mem_map's for valid memory
24  */
25 #ifdef CONFIG_SPARSEMEM_EXTREME
26 struct mem_section **mem_section;
27 #else
28 struct mem_section mem_section[NR_SECTION_ROOTS][SECTIONS_PER_ROOT]
29 	____cacheline_internodealigned_in_smp;
30 #endif
31 EXPORT_SYMBOL(mem_section);
32 
33 #ifdef NODE_NOT_IN_PAGE_FLAGS
34 /*
35  * If we did not store the node number in the page then we have to
36  * do a lookup in the section_to_node_table in order to find which
37  * node the page belongs to.
38  */
39 #if MAX_NUMNODES <= 256
40 static u8 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned;
41 #else
42 static u16 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned;
43 #endif
44 
45 int page_to_nid(const struct page *page)
46 {
47 	return section_to_node_table[page_to_section(page)];
48 }
49 EXPORT_SYMBOL(page_to_nid);
50 
51 static void set_section_nid(unsigned long section_nr, int nid)
52 {
53 	section_to_node_table[section_nr] = nid;
54 }
55 #else /* !NODE_NOT_IN_PAGE_FLAGS */
56 static inline void set_section_nid(unsigned long section_nr, int nid)
57 {
58 }
59 #endif
60 
61 #ifdef CONFIG_SPARSEMEM_EXTREME
62 static noinline struct mem_section __ref *sparse_index_alloc(int nid)
63 {
64 	struct mem_section *section = NULL;
65 	unsigned long array_size = SECTIONS_PER_ROOT *
66 				   sizeof(struct mem_section);
67 
68 	if (slab_is_available())
69 		section = kzalloc_node(array_size, GFP_KERNEL, nid);
70 	else
71 		section = memblock_alloc_node(array_size, SMP_CACHE_BYTES,
72 					      nid);
73 
74 	return section;
75 }
76 
77 static int __meminit sparse_index_init(unsigned long section_nr, int nid)
78 {
79 	unsigned long root = SECTION_NR_TO_ROOT(section_nr);
80 	struct mem_section *section;
81 
82 	if (mem_section[root])
83 		return -EEXIST;
84 
85 	section = sparse_index_alloc(nid);
86 	if (!section)
87 		return -ENOMEM;
88 
89 	mem_section[root] = section;
90 
91 	return 0;
92 }
93 #else /* !SPARSEMEM_EXTREME */
94 static inline int sparse_index_init(unsigned long section_nr, int nid)
95 {
96 	return 0;
97 }
98 #endif
99 
100 #ifdef CONFIG_SPARSEMEM_EXTREME
101 int __section_nr(struct mem_section* ms)
102 {
103 	unsigned long root_nr;
104 	struct mem_section *root = NULL;
105 
106 	for (root_nr = 0; root_nr < NR_SECTION_ROOTS; root_nr++) {
107 		root = __nr_to_section(root_nr * SECTIONS_PER_ROOT);
108 		if (!root)
109 			continue;
110 
111 		if ((ms >= root) && (ms < (root + SECTIONS_PER_ROOT)))
112 		     break;
113 	}
114 
115 	VM_BUG_ON(!root);
116 
117 	return (root_nr * SECTIONS_PER_ROOT) + (ms - root);
118 }
119 #else
120 int __section_nr(struct mem_section* ms)
121 {
122 	return (int)(ms - mem_section[0]);
123 }
124 #endif
125 
126 /*
127  * During early boot, before section_mem_map is used for an actual
128  * mem_map, we use section_mem_map to store the section's NUMA
129  * node.  This keeps us from having to use another data structure.  The
130  * node information is cleared just before we store the real mem_map.
131  */
132 static inline unsigned long sparse_encode_early_nid(int nid)
133 {
134 	return (nid << SECTION_NID_SHIFT);
135 }
136 
137 static inline int sparse_early_nid(struct mem_section *section)
138 {
139 	return (section->section_mem_map >> SECTION_NID_SHIFT);
140 }
141 
142 /* Validate the physical addressing limitations of the model */
143 void __meminit mminit_validate_memmodel_limits(unsigned long *start_pfn,
144 						unsigned long *end_pfn)
145 {
146 	unsigned long max_sparsemem_pfn = 1UL << (MAX_PHYSMEM_BITS-PAGE_SHIFT);
147 
148 	/*
149 	 * Sanity checks - do not allow an architecture to pass
150 	 * in larger pfns than the maximum scope of sparsemem:
151 	 */
152 	if (*start_pfn > max_sparsemem_pfn) {
153 		mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
154 			"Start of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
155 			*start_pfn, *end_pfn, max_sparsemem_pfn);
156 		WARN_ON_ONCE(1);
157 		*start_pfn = max_sparsemem_pfn;
158 		*end_pfn = max_sparsemem_pfn;
159 	} else if (*end_pfn > max_sparsemem_pfn) {
160 		mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
161 			"End of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
162 			*start_pfn, *end_pfn, max_sparsemem_pfn);
163 		WARN_ON_ONCE(1);
164 		*end_pfn = max_sparsemem_pfn;
165 	}
166 }
167 
168 /*
169  * There are a number of times that we loop over NR_MEM_SECTIONS,
170  * looking for section_present() on each.  But, when we have very
171  * large physical address spaces, NR_MEM_SECTIONS can also be
172  * very large which makes the loops quite long.
173  *
174  * Keeping track of this gives us an easy way to break out of
175  * those loops early.
176  */
177 int __highest_present_section_nr;
178 static void section_mark_present(struct mem_section *ms)
179 {
180 	int section_nr = __section_nr(ms);
181 
182 	if (section_nr > __highest_present_section_nr)
183 		__highest_present_section_nr = section_nr;
184 
185 	ms->section_mem_map |= SECTION_MARKED_PRESENT;
186 }
187 
188 static inline int next_present_section_nr(int section_nr)
189 {
190 	do {
191 		section_nr++;
192 		if (present_section_nr(section_nr))
193 			return section_nr;
194 	} while ((section_nr <= __highest_present_section_nr));
195 
196 	return -1;
197 }
198 #define for_each_present_section_nr(start, section_nr)		\
199 	for (section_nr = next_present_section_nr(start-1);	\
200 	     ((section_nr >= 0) &&				\
201 	      (section_nr <= __highest_present_section_nr));	\
202 	     section_nr = next_present_section_nr(section_nr))
203 
204 static inline unsigned long first_present_section_nr(void)
205 {
206 	return next_present_section_nr(-1);
207 }
208 
209 /* Record a memory area against a node. */
210 void __init memory_present(int nid, unsigned long start, unsigned long end)
211 {
212 	unsigned long pfn;
213 
214 #ifdef CONFIG_SPARSEMEM_EXTREME
215 	if (unlikely(!mem_section)) {
216 		unsigned long size, align;
217 
218 		size = sizeof(struct mem_section*) * NR_SECTION_ROOTS;
219 		align = 1 << (INTERNODE_CACHE_SHIFT);
220 		mem_section = memblock_alloc(size, align);
221 	}
222 #endif
223 
224 	start &= PAGE_SECTION_MASK;
225 	mminit_validate_memmodel_limits(&start, &end);
226 	for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) {
227 		unsigned long section = pfn_to_section_nr(pfn);
228 		struct mem_section *ms;
229 
230 		sparse_index_init(section, nid);
231 		set_section_nid(section, nid);
232 
233 		ms = __nr_to_section(section);
234 		if (!ms->section_mem_map) {
235 			ms->section_mem_map = sparse_encode_early_nid(nid) |
236 							SECTION_IS_ONLINE;
237 			section_mark_present(ms);
238 		}
239 	}
240 }
241 
242 /*
243  * Mark all memblocks as present using memory_present(). This is a
244  * convienence function that is useful for a number of arches
245  * to mark all of the systems memory as present during initialization.
246  */
247 void __init memblocks_present(void)
248 {
249 	struct memblock_region *reg;
250 
251 	for_each_memblock(memory, reg) {
252 		memory_present(memblock_get_region_node(reg),
253 			       memblock_region_memory_base_pfn(reg),
254 			       memblock_region_memory_end_pfn(reg));
255 	}
256 }
257 
258 /*
259  * Subtle, we encode the real pfn into the mem_map such that
260  * the identity pfn - section_mem_map will return the actual
261  * physical page frame number.
262  */
263 static unsigned long sparse_encode_mem_map(struct page *mem_map, unsigned long pnum)
264 {
265 	unsigned long coded_mem_map =
266 		(unsigned long)(mem_map - (section_nr_to_pfn(pnum)));
267 	BUILD_BUG_ON(SECTION_MAP_LAST_BIT > (1UL<<PFN_SECTION_SHIFT));
268 	BUG_ON(coded_mem_map & ~SECTION_MAP_MASK);
269 	return coded_mem_map;
270 }
271 
272 /*
273  * Decode mem_map from the coded memmap
274  */
275 struct page *sparse_decode_mem_map(unsigned long coded_mem_map, unsigned long pnum)
276 {
277 	/* mask off the extra low bits of information */
278 	coded_mem_map &= SECTION_MAP_MASK;
279 	return ((struct page *)coded_mem_map) + section_nr_to_pfn(pnum);
280 }
281 
282 static void __meminit sparse_init_one_section(struct mem_section *ms,
283 		unsigned long pnum, struct page *mem_map,
284 		unsigned long *pageblock_bitmap)
285 {
286 	ms->section_mem_map &= ~SECTION_MAP_MASK;
287 	ms->section_mem_map |= sparse_encode_mem_map(mem_map, pnum) |
288 							SECTION_HAS_MEM_MAP;
289  	ms->pageblock_flags = pageblock_bitmap;
290 }
291 
292 unsigned long usemap_size(void)
293 {
294 	return BITS_TO_LONGS(SECTION_BLOCKFLAGS_BITS) * sizeof(unsigned long);
295 }
296 
297 #ifdef CONFIG_MEMORY_HOTPLUG
298 static unsigned long *__kmalloc_section_usemap(void)
299 {
300 	return kmalloc(usemap_size(), GFP_KERNEL);
301 }
302 #endif /* CONFIG_MEMORY_HOTPLUG */
303 
304 #ifdef CONFIG_MEMORY_HOTREMOVE
305 static unsigned long * __init
306 sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
307 					 unsigned long size)
308 {
309 	unsigned long goal, limit;
310 	unsigned long *p;
311 	int nid;
312 	/*
313 	 * A page may contain usemaps for other sections preventing the
314 	 * page being freed and making a section unremovable while
315 	 * other sections referencing the usemap remain active. Similarly,
316 	 * a pgdat can prevent a section being removed. If section A
317 	 * contains a pgdat and section B contains the usemap, both
318 	 * sections become inter-dependent. This allocates usemaps
319 	 * from the same section as the pgdat where possible to avoid
320 	 * this problem.
321 	 */
322 	goal = __pa(pgdat) & (PAGE_SECTION_MASK << PAGE_SHIFT);
323 	limit = goal + (1UL << PA_SECTION_SHIFT);
324 	nid = early_pfn_to_nid(goal >> PAGE_SHIFT);
325 again:
326 	p = memblock_alloc_try_nid_nopanic(size,
327 						SMP_CACHE_BYTES, goal, limit,
328 						nid);
329 	if (!p && limit) {
330 		limit = 0;
331 		goto again;
332 	}
333 	return p;
334 }
335 
336 static void __init check_usemap_section_nr(int nid, unsigned long *usemap)
337 {
338 	unsigned long usemap_snr, pgdat_snr;
339 	static unsigned long old_usemap_snr;
340 	static unsigned long old_pgdat_snr;
341 	struct pglist_data *pgdat = NODE_DATA(nid);
342 	int usemap_nid;
343 
344 	/* First call */
345 	if (!old_usemap_snr) {
346 		old_usemap_snr = NR_MEM_SECTIONS;
347 		old_pgdat_snr = NR_MEM_SECTIONS;
348 	}
349 
350 	usemap_snr = pfn_to_section_nr(__pa(usemap) >> PAGE_SHIFT);
351 	pgdat_snr = pfn_to_section_nr(__pa(pgdat) >> PAGE_SHIFT);
352 	if (usemap_snr == pgdat_snr)
353 		return;
354 
355 	if (old_usemap_snr == usemap_snr && old_pgdat_snr == pgdat_snr)
356 		/* skip redundant message */
357 		return;
358 
359 	old_usemap_snr = usemap_snr;
360 	old_pgdat_snr = pgdat_snr;
361 
362 	usemap_nid = sparse_early_nid(__nr_to_section(usemap_snr));
363 	if (usemap_nid != nid) {
364 		pr_info("node %d must be removed before remove section %ld\n",
365 			nid, usemap_snr);
366 		return;
367 	}
368 	/*
369 	 * There is a circular dependency.
370 	 * Some platforms allow un-removable section because they will just
371 	 * gather other removable sections for dynamic partitioning.
372 	 * Just notify un-removable section's number here.
373 	 */
374 	pr_info("Section %ld and %ld (node %d) have a circular dependency on usemap and pgdat allocations\n",
375 		usemap_snr, pgdat_snr, nid);
376 }
377 #else
378 static unsigned long * __init
379 sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
380 					 unsigned long size)
381 {
382 	return memblock_alloc_node_nopanic(size, pgdat->node_id);
383 }
384 
385 static void __init check_usemap_section_nr(int nid, unsigned long *usemap)
386 {
387 }
388 #endif /* CONFIG_MEMORY_HOTREMOVE */
389 
390 #ifdef CONFIG_SPARSEMEM_VMEMMAP
391 static unsigned long __init section_map_size(void)
392 {
393 	return ALIGN(sizeof(struct page) * PAGES_PER_SECTION, PMD_SIZE);
394 }
395 
396 #else
397 static unsigned long __init section_map_size(void)
398 {
399 	return PAGE_ALIGN(sizeof(struct page) * PAGES_PER_SECTION);
400 }
401 
402 struct page __init *sparse_mem_map_populate(unsigned long pnum, int nid,
403 		struct vmem_altmap *altmap)
404 {
405 	unsigned long size = section_map_size();
406 	struct page *map = sparse_buffer_alloc(size);
407 
408 	if (map)
409 		return map;
410 
411 	map = memblock_alloc_try_nid(size,
412 					  PAGE_SIZE, __pa(MAX_DMA_ADDRESS),
413 					  MEMBLOCK_ALLOC_ACCESSIBLE, nid);
414 	return map;
415 }
416 #endif /* !CONFIG_SPARSEMEM_VMEMMAP */
417 
418 static void *sparsemap_buf __meminitdata;
419 static void *sparsemap_buf_end __meminitdata;
420 
421 static void __init sparse_buffer_init(unsigned long size, int nid)
422 {
423 	WARN_ON(sparsemap_buf);	/* forgot to call sparse_buffer_fini()? */
424 	sparsemap_buf =
425 		memblock_alloc_try_nid_raw(size, PAGE_SIZE,
426 						__pa(MAX_DMA_ADDRESS),
427 						MEMBLOCK_ALLOC_ACCESSIBLE, nid);
428 	sparsemap_buf_end = sparsemap_buf + size;
429 }
430 
431 static void __init sparse_buffer_fini(void)
432 {
433 	unsigned long size = sparsemap_buf_end - sparsemap_buf;
434 
435 	if (sparsemap_buf && size > 0)
436 		memblock_free_early(__pa(sparsemap_buf), size);
437 	sparsemap_buf = NULL;
438 }
439 
440 void * __meminit sparse_buffer_alloc(unsigned long size)
441 {
442 	void *ptr = NULL;
443 
444 	if (sparsemap_buf) {
445 		ptr = PTR_ALIGN(sparsemap_buf, size);
446 		if (ptr + size > sparsemap_buf_end)
447 			ptr = NULL;
448 		else
449 			sparsemap_buf = ptr + size;
450 	}
451 	return ptr;
452 }
453 
454 void __weak __meminit vmemmap_populate_print_last(void)
455 {
456 }
457 
458 /*
459  * Initialize sparse on a specific node. The node spans [pnum_begin, pnum_end)
460  * And number of present sections in this node is map_count.
461  */
462 static void __init sparse_init_nid(int nid, unsigned long pnum_begin,
463 				   unsigned long pnum_end,
464 				   unsigned long map_count)
465 {
466 	unsigned long pnum, usemap_longs, *usemap;
467 	struct page *map;
468 
469 	usemap_longs = BITS_TO_LONGS(SECTION_BLOCKFLAGS_BITS);
470 	usemap = sparse_early_usemaps_alloc_pgdat_section(NODE_DATA(nid),
471 							  usemap_size() *
472 							  map_count);
473 	if (!usemap) {
474 		pr_err("%s: node[%d] usemap allocation failed", __func__, nid);
475 		goto failed;
476 	}
477 	sparse_buffer_init(map_count * section_map_size(), nid);
478 	for_each_present_section_nr(pnum_begin, pnum) {
479 		if (pnum >= pnum_end)
480 			break;
481 
482 		map = sparse_mem_map_populate(pnum, nid, NULL);
483 		if (!map) {
484 			pr_err("%s: node[%d] memory map backing failed. Some memory will not be available.",
485 			       __func__, nid);
486 			pnum_begin = pnum;
487 			goto failed;
488 		}
489 		check_usemap_section_nr(nid, usemap);
490 		sparse_init_one_section(__nr_to_section(pnum), pnum, map, usemap);
491 		usemap += usemap_longs;
492 	}
493 	sparse_buffer_fini();
494 	return;
495 failed:
496 	/* We failed to allocate, mark all the following pnums as not present */
497 	for_each_present_section_nr(pnum_begin, pnum) {
498 		struct mem_section *ms;
499 
500 		if (pnum >= pnum_end)
501 			break;
502 		ms = __nr_to_section(pnum);
503 		ms->section_mem_map = 0;
504 	}
505 }
506 
507 /*
508  * Allocate the accumulated non-linear sections, allocate a mem_map
509  * for each and record the physical to section mapping.
510  */
511 void __init sparse_init(void)
512 {
513 	unsigned long pnum_begin = first_present_section_nr();
514 	int nid_begin = sparse_early_nid(__nr_to_section(pnum_begin));
515 	unsigned long pnum_end, map_count = 1;
516 
517 	/* Setup pageblock_order for HUGETLB_PAGE_SIZE_VARIABLE */
518 	set_pageblock_order();
519 
520 	for_each_present_section_nr(pnum_begin + 1, pnum_end) {
521 		int nid = sparse_early_nid(__nr_to_section(pnum_end));
522 
523 		if (nid == nid_begin) {
524 			map_count++;
525 			continue;
526 		}
527 		/* Init node with sections in range [pnum_begin, pnum_end) */
528 		sparse_init_nid(nid_begin, pnum_begin, pnum_end, map_count);
529 		nid_begin = nid;
530 		pnum_begin = pnum_end;
531 		map_count = 1;
532 	}
533 	/* cover the last node */
534 	sparse_init_nid(nid_begin, pnum_begin, pnum_end, map_count);
535 	vmemmap_populate_print_last();
536 }
537 
538 #ifdef CONFIG_MEMORY_HOTPLUG
539 
540 /* Mark all memory sections within the pfn range as online */
541 void online_mem_sections(unsigned long start_pfn, unsigned long end_pfn)
542 {
543 	unsigned long pfn;
544 
545 	for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
546 		unsigned long section_nr = pfn_to_section_nr(pfn);
547 		struct mem_section *ms;
548 
549 		/* onlining code should never touch invalid ranges */
550 		if (WARN_ON(!valid_section_nr(section_nr)))
551 			continue;
552 
553 		ms = __nr_to_section(section_nr);
554 		ms->section_mem_map |= SECTION_IS_ONLINE;
555 	}
556 }
557 
558 #ifdef CONFIG_MEMORY_HOTREMOVE
559 /* Mark all memory sections within the pfn range as online */
560 void offline_mem_sections(unsigned long start_pfn, unsigned long end_pfn)
561 {
562 	unsigned long pfn;
563 
564 	for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
565 		unsigned long section_nr = pfn_to_section_nr(pfn);
566 		struct mem_section *ms;
567 
568 		/*
569 		 * TODO this needs some double checking. Offlining code makes
570 		 * sure to check pfn_valid but those checks might be just bogus
571 		 */
572 		if (WARN_ON(!valid_section_nr(section_nr)))
573 			continue;
574 
575 		ms = __nr_to_section(section_nr);
576 		ms->section_mem_map &= ~SECTION_IS_ONLINE;
577 	}
578 }
579 #endif
580 
581 #ifdef CONFIG_SPARSEMEM_VMEMMAP
582 static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid,
583 		struct vmem_altmap *altmap)
584 {
585 	/* This will make the necessary allocations eventually. */
586 	return sparse_mem_map_populate(pnum, nid, altmap);
587 }
588 static void __kfree_section_memmap(struct page *memmap,
589 		struct vmem_altmap *altmap)
590 {
591 	unsigned long start = (unsigned long)memmap;
592 	unsigned long end = (unsigned long)(memmap + PAGES_PER_SECTION);
593 
594 	vmemmap_free(start, end, altmap);
595 }
596 #ifdef CONFIG_MEMORY_HOTREMOVE
597 static void free_map_bootmem(struct page *memmap)
598 {
599 	unsigned long start = (unsigned long)memmap;
600 	unsigned long end = (unsigned long)(memmap + PAGES_PER_SECTION);
601 
602 	vmemmap_free(start, end, NULL);
603 }
604 #endif /* CONFIG_MEMORY_HOTREMOVE */
605 #else
606 static struct page *__kmalloc_section_memmap(void)
607 {
608 	struct page *page, *ret;
609 	unsigned long memmap_size = sizeof(struct page) * PAGES_PER_SECTION;
610 
611 	page = alloc_pages(GFP_KERNEL|__GFP_NOWARN, get_order(memmap_size));
612 	if (page)
613 		goto got_map_page;
614 
615 	ret = vmalloc(memmap_size);
616 	if (ret)
617 		goto got_map_ptr;
618 
619 	return NULL;
620 got_map_page:
621 	ret = (struct page *)pfn_to_kaddr(page_to_pfn(page));
622 got_map_ptr:
623 
624 	return ret;
625 }
626 
627 static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid,
628 		struct vmem_altmap *altmap)
629 {
630 	return __kmalloc_section_memmap();
631 }
632 
633 static void __kfree_section_memmap(struct page *memmap,
634 		struct vmem_altmap *altmap)
635 {
636 	if (is_vmalloc_addr(memmap))
637 		vfree(memmap);
638 	else
639 		free_pages((unsigned long)memmap,
640 			   get_order(sizeof(struct page) * PAGES_PER_SECTION));
641 }
642 
643 #ifdef CONFIG_MEMORY_HOTREMOVE
644 static void free_map_bootmem(struct page *memmap)
645 {
646 	unsigned long maps_section_nr, removing_section_nr, i;
647 	unsigned long magic, nr_pages;
648 	struct page *page = virt_to_page(memmap);
649 
650 	nr_pages = PAGE_ALIGN(PAGES_PER_SECTION * sizeof(struct page))
651 		>> PAGE_SHIFT;
652 
653 	for (i = 0; i < nr_pages; i++, page++) {
654 		magic = (unsigned long) page->freelist;
655 
656 		BUG_ON(magic == NODE_INFO);
657 
658 		maps_section_nr = pfn_to_section_nr(page_to_pfn(page));
659 		removing_section_nr = page_private(page);
660 
661 		/*
662 		 * When this function is called, the removing section is
663 		 * logical offlined state. This means all pages are isolated
664 		 * from page allocator. If removing section's memmap is placed
665 		 * on the same section, it must not be freed.
666 		 * If it is freed, page allocator may allocate it which will
667 		 * be removed physically soon.
668 		 */
669 		if (maps_section_nr != removing_section_nr)
670 			put_page_bootmem(page);
671 	}
672 }
673 #endif /* CONFIG_MEMORY_HOTREMOVE */
674 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
675 
676 /*
677  * returns the number of sections whose mem_maps were properly
678  * set.  If this is <=0, then that means that the passed-in
679  * map was not consumed and must be freed.
680  */
681 int __meminit sparse_add_one_section(int nid, unsigned long start_pfn,
682 				     struct vmem_altmap *altmap)
683 {
684 	unsigned long section_nr = pfn_to_section_nr(start_pfn);
685 	struct mem_section *ms;
686 	struct page *memmap;
687 	unsigned long *usemap;
688 	int ret;
689 
690 	/*
691 	 * no locking for this, because it does its own
692 	 * plus, it does a kmalloc
693 	 */
694 	ret = sparse_index_init(section_nr, nid);
695 	if (ret < 0 && ret != -EEXIST)
696 		return ret;
697 	ret = 0;
698 	memmap = kmalloc_section_memmap(section_nr, nid, altmap);
699 	if (!memmap)
700 		return -ENOMEM;
701 	usemap = __kmalloc_section_usemap();
702 	if (!usemap) {
703 		__kfree_section_memmap(memmap, altmap);
704 		return -ENOMEM;
705 	}
706 
707 	ms = __pfn_to_section(start_pfn);
708 	if (ms->section_mem_map & SECTION_MARKED_PRESENT) {
709 		ret = -EEXIST;
710 		goto out;
711 	}
712 
713 	/*
714 	 * Poison uninitialized struct pages in order to catch invalid flags
715 	 * combinations.
716 	 */
717 	page_init_poison(memmap, sizeof(struct page) * PAGES_PER_SECTION);
718 
719 	section_mark_present(ms);
720 	sparse_init_one_section(ms, section_nr, memmap, usemap);
721 
722 out:
723 	if (ret < 0) {
724 		kfree(usemap);
725 		__kfree_section_memmap(memmap, altmap);
726 	}
727 	return ret;
728 }
729 
730 #ifdef CONFIG_MEMORY_HOTREMOVE
731 #ifdef CONFIG_MEMORY_FAILURE
732 static void clear_hwpoisoned_pages(struct page *memmap, int nr_pages)
733 {
734 	int i;
735 
736 	if (!memmap)
737 		return;
738 
739 	/*
740 	 * A further optimization is to have per section refcounted
741 	 * num_poisoned_pages.  But that would need more space per memmap, so
742 	 * for now just do a quick global check to speed up this routine in the
743 	 * absence of bad pages.
744 	 */
745 	if (atomic_long_read(&num_poisoned_pages) == 0)
746 		return;
747 
748 	for (i = 0; i < nr_pages; i++) {
749 		if (PageHWPoison(&memmap[i])) {
750 			atomic_long_sub(1, &num_poisoned_pages);
751 			ClearPageHWPoison(&memmap[i]);
752 		}
753 	}
754 }
755 #else
756 static inline void clear_hwpoisoned_pages(struct page *memmap, int nr_pages)
757 {
758 }
759 #endif
760 
761 static void free_section_usemap(struct page *memmap, unsigned long *usemap,
762 		struct vmem_altmap *altmap)
763 {
764 	struct page *usemap_page;
765 
766 	if (!usemap)
767 		return;
768 
769 	usemap_page = virt_to_page(usemap);
770 	/*
771 	 * Check to see if allocation came from hot-plug-add
772 	 */
773 	if (PageSlab(usemap_page) || PageCompound(usemap_page)) {
774 		kfree(usemap);
775 		if (memmap)
776 			__kfree_section_memmap(memmap, altmap);
777 		return;
778 	}
779 
780 	/*
781 	 * The usemap came from bootmem. This is packed with other usemaps
782 	 * on the section which has pgdat at boot time. Just keep it as is now.
783 	 */
784 
785 	if (memmap)
786 		free_map_bootmem(memmap);
787 }
788 
789 void sparse_remove_one_section(struct zone *zone, struct mem_section *ms,
790 		unsigned long map_offset, struct vmem_altmap *altmap)
791 {
792 	struct page *memmap = NULL;
793 	unsigned long *usemap = NULL;
794 
795 	if (ms->section_mem_map) {
796 		usemap = ms->pageblock_flags;
797 		memmap = sparse_decode_mem_map(ms->section_mem_map,
798 						__section_nr(ms));
799 		ms->section_mem_map = 0;
800 		ms->pageblock_flags = NULL;
801 	}
802 
803 	clear_hwpoisoned_pages(memmap + map_offset,
804 			PAGES_PER_SECTION - map_offset);
805 	free_section_usemap(memmap, usemap, altmap);
806 }
807 #endif /* CONFIG_MEMORY_HOTREMOVE */
808 #endif /* CONFIG_MEMORY_HOTPLUG */
809