xref: /openbmc/linux/mm/sparse.c (revision cc1e6315)
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/bootmem.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_virt_alloc_node(array_size, nid);
72 
73 	return section;
74 }
75 
76 static int __meminit sparse_index_init(unsigned long section_nr, int nid)
77 {
78 	unsigned long root = SECTION_NR_TO_ROOT(section_nr);
79 	struct mem_section *section;
80 
81 	if (mem_section[root])
82 		return -EEXIST;
83 
84 	section = sparse_index_alloc(nid);
85 	if (!section)
86 		return -ENOMEM;
87 
88 	mem_section[root] = section;
89 
90 	return 0;
91 }
92 #else /* !SPARSEMEM_EXTREME */
93 static inline int sparse_index_init(unsigned long section_nr, int nid)
94 {
95 	return 0;
96 }
97 #endif
98 
99 #ifdef CONFIG_SPARSEMEM_EXTREME
100 int __section_nr(struct mem_section* ms)
101 {
102 	unsigned long root_nr;
103 	struct mem_section *root = NULL;
104 
105 	for (root_nr = 0; root_nr < NR_SECTION_ROOTS; root_nr++) {
106 		root = __nr_to_section(root_nr * SECTIONS_PER_ROOT);
107 		if (!root)
108 			continue;
109 
110 		if ((ms >= root) && (ms < (root + SECTIONS_PER_ROOT)))
111 		     break;
112 	}
113 
114 	VM_BUG_ON(!root);
115 
116 	return (root_nr * SECTIONS_PER_ROOT) + (ms - root);
117 }
118 #else
119 int __section_nr(struct mem_section* ms)
120 {
121 	return (int)(ms - mem_section[0]);
122 }
123 #endif
124 
125 /*
126  * During early boot, before section_mem_map is used for an actual
127  * mem_map, we use section_mem_map to store the section's NUMA
128  * node.  This keeps us from having to use another data structure.  The
129  * node information is cleared just before we store the real mem_map.
130  */
131 static inline unsigned long sparse_encode_early_nid(int nid)
132 {
133 	return (nid << SECTION_NID_SHIFT);
134 }
135 
136 static inline int sparse_early_nid(struct mem_section *section)
137 {
138 	return (section->section_mem_map >> SECTION_NID_SHIFT);
139 }
140 
141 /* Validate the physical addressing limitations of the model */
142 void __meminit mminit_validate_memmodel_limits(unsigned long *start_pfn,
143 						unsigned long *end_pfn)
144 {
145 	unsigned long max_sparsemem_pfn = 1UL << (MAX_PHYSMEM_BITS-PAGE_SHIFT);
146 
147 	/*
148 	 * Sanity checks - do not allow an architecture to pass
149 	 * in larger pfns than the maximum scope of sparsemem:
150 	 */
151 	if (*start_pfn > max_sparsemem_pfn) {
152 		mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
153 			"Start of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
154 			*start_pfn, *end_pfn, max_sparsemem_pfn);
155 		WARN_ON_ONCE(1);
156 		*start_pfn = max_sparsemem_pfn;
157 		*end_pfn = max_sparsemem_pfn;
158 	} else if (*end_pfn > max_sparsemem_pfn) {
159 		mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
160 			"End of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
161 			*start_pfn, *end_pfn, max_sparsemem_pfn);
162 		WARN_ON_ONCE(1);
163 		*end_pfn = max_sparsemem_pfn;
164 	}
165 }
166 
167 /*
168  * There are a number of times that we loop over NR_MEM_SECTIONS,
169  * looking for section_present() on each.  But, when we have very
170  * large physical address spaces, NR_MEM_SECTIONS can also be
171  * very large which makes the loops quite long.
172  *
173  * Keeping track of this gives us an easy way to break out of
174  * those loops early.
175  */
176 int __highest_present_section_nr;
177 static void section_mark_present(struct mem_section *ms)
178 {
179 	int section_nr = __section_nr(ms);
180 
181 	if (section_nr > __highest_present_section_nr)
182 		__highest_present_section_nr = section_nr;
183 
184 	ms->section_mem_map |= SECTION_MARKED_PRESENT;
185 }
186 
187 static inline int next_present_section_nr(int section_nr)
188 {
189 	do {
190 		section_nr++;
191 		if (present_section_nr(section_nr))
192 			return section_nr;
193 	} while ((section_nr <= __highest_present_section_nr));
194 
195 	return -1;
196 }
197 #define for_each_present_section_nr(start, section_nr)		\
198 	for (section_nr = next_present_section_nr(start-1);	\
199 	     ((section_nr >= 0) &&				\
200 	      (section_nr <= __highest_present_section_nr));	\
201 	     section_nr = next_present_section_nr(section_nr))
202 
203 static inline unsigned long first_present_section_nr(void)
204 {
205 	return next_present_section_nr(-1);
206 }
207 
208 /* Record a memory area against a node. */
209 void __init memory_present(int nid, unsigned long start, unsigned long end)
210 {
211 	unsigned long pfn;
212 
213 #ifdef CONFIG_SPARSEMEM_EXTREME
214 	if (unlikely(!mem_section)) {
215 		unsigned long size, align;
216 
217 		size = sizeof(struct mem_section*) * NR_SECTION_ROOTS;
218 		align = 1 << (INTERNODE_CACHE_SHIFT);
219 		mem_section = memblock_virt_alloc(size, align);
220 	}
221 #endif
222 
223 	start &= PAGE_SECTION_MASK;
224 	mminit_validate_memmodel_limits(&start, &end);
225 	for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) {
226 		unsigned long section = pfn_to_section_nr(pfn);
227 		struct mem_section *ms;
228 
229 		sparse_index_init(section, nid);
230 		set_section_nid(section, nid);
231 
232 		ms = __nr_to_section(section);
233 		if (!ms->section_mem_map) {
234 			ms->section_mem_map = sparse_encode_early_nid(nid) |
235 							SECTION_IS_ONLINE;
236 			section_mark_present(ms);
237 		}
238 	}
239 }
240 
241 /*
242  * Subtle, we encode the real pfn into the mem_map such that
243  * the identity pfn - section_mem_map will return the actual
244  * physical page frame number.
245  */
246 static unsigned long sparse_encode_mem_map(struct page *mem_map, unsigned long pnum)
247 {
248 	unsigned long coded_mem_map =
249 		(unsigned long)(mem_map - (section_nr_to_pfn(pnum)));
250 	BUILD_BUG_ON(SECTION_MAP_LAST_BIT > (1UL<<PFN_SECTION_SHIFT));
251 	BUG_ON(coded_mem_map & ~SECTION_MAP_MASK);
252 	return coded_mem_map;
253 }
254 
255 /*
256  * Decode mem_map from the coded memmap
257  */
258 struct page *sparse_decode_mem_map(unsigned long coded_mem_map, unsigned long pnum)
259 {
260 	/* mask off the extra low bits of information */
261 	coded_mem_map &= SECTION_MAP_MASK;
262 	return ((struct page *)coded_mem_map) + section_nr_to_pfn(pnum);
263 }
264 
265 static void __meminit sparse_init_one_section(struct mem_section *ms,
266 		unsigned long pnum, struct page *mem_map,
267 		unsigned long *pageblock_bitmap)
268 {
269 	ms->section_mem_map &= ~SECTION_MAP_MASK;
270 	ms->section_mem_map |= sparse_encode_mem_map(mem_map, pnum) |
271 							SECTION_HAS_MEM_MAP;
272  	ms->pageblock_flags = pageblock_bitmap;
273 }
274 
275 unsigned long usemap_size(void)
276 {
277 	return BITS_TO_LONGS(SECTION_BLOCKFLAGS_BITS) * sizeof(unsigned long);
278 }
279 
280 #ifdef CONFIG_MEMORY_HOTPLUG
281 static unsigned long *__kmalloc_section_usemap(void)
282 {
283 	return kmalloc(usemap_size(), GFP_KERNEL);
284 }
285 #endif /* CONFIG_MEMORY_HOTPLUG */
286 
287 #ifdef CONFIG_MEMORY_HOTREMOVE
288 static unsigned long * __init
289 sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
290 					 unsigned long size)
291 {
292 	unsigned long goal, limit;
293 	unsigned long *p;
294 	int nid;
295 	/*
296 	 * A page may contain usemaps for other sections preventing the
297 	 * page being freed and making a section unremovable while
298 	 * other sections referencing the usemap remain active. Similarly,
299 	 * a pgdat can prevent a section being removed. If section A
300 	 * contains a pgdat and section B contains the usemap, both
301 	 * sections become inter-dependent. This allocates usemaps
302 	 * from the same section as the pgdat where possible to avoid
303 	 * this problem.
304 	 */
305 	goal = __pa(pgdat) & (PAGE_SECTION_MASK << PAGE_SHIFT);
306 	limit = goal + (1UL << PA_SECTION_SHIFT);
307 	nid = early_pfn_to_nid(goal >> PAGE_SHIFT);
308 again:
309 	p = memblock_virt_alloc_try_nid_nopanic(size,
310 						SMP_CACHE_BYTES, goal, limit,
311 						nid);
312 	if (!p && limit) {
313 		limit = 0;
314 		goto again;
315 	}
316 	return p;
317 }
318 
319 static void __init check_usemap_section_nr(int nid, unsigned long *usemap)
320 {
321 	unsigned long usemap_snr, pgdat_snr;
322 	static unsigned long old_usemap_snr;
323 	static unsigned long old_pgdat_snr;
324 	struct pglist_data *pgdat = NODE_DATA(nid);
325 	int usemap_nid;
326 
327 	/* First call */
328 	if (!old_usemap_snr) {
329 		old_usemap_snr = NR_MEM_SECTIONS;
330 		old_pgdat_snr = NR_MEM_SECTIONS;
331 	}
332 
333 	usemap_snr = pfn_to_section_nr(__pa(usemap) >> PAGE_SHIFT);
334 	pgdat_snr = pfn_to_section_nr(__pa(pgdat) >> PAGE_SHIFT);
335 	if (usemap_snr == pgdat_snr)
336 		return;
337 
338 	if (old_usemap_snr == usemap_snr && old_pgdat_snr == pgdat_snr)
339 		/* skip redundant message */
340 		return;
341 
342 	old_usemap_snr = usemap_snr;
343 	old_pgdat_snr = pgdat_snr;
344 
345 	usemap_nid = sparse_early_nid(__nr_to_section(usemap_snr));
346 	if (usemap_nid != nid) {
347 		pr_info("node %d must be removed before remove section %ld\n",
348 			nid, usemap_snr);
349 		return;
350 	}
351 	/*
352 	 * There is a circular dependency.
353 	 * Some platforms allow un-removable section because they will just
354 	 * gather other removable sections for dynamic partitioning.
355 	 * Just notify un-removable section's number here.
356 	 */
357 	pr_info("Section %ld and %ld (node %d) have a circular dependency on usemap and pgdat allocations\n",
358 		usemap_snr, pgdat_snr, nid);
359 }
360 #else
361 static unsigned long * __init
362 sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
363 					 unsigned long size)
364 {
365 	return memblock_virt_alloc_node_nopanic(size, pgdat->node_id);
366 }
367 
368 static void __init check_usemap_section_nr(int nid, unsigned long *usemap)
369 {
370 }
371 #endif /* CONFIG_MEMORY_HOTREMOVE */
372 
373 #ifdef CONFIG_SPARSEMEM_VMEMMAP
374 static unsigned long __init section_map_size(void)
375 {
376 	return ALIGN(sizeof(struct page) * PAGES_PER_SECTION, PMD_SIZE);
377 }
378 
379 #else
380 static unsigned long __init section_map_size(void)
381 {
382 	return PAGE_ALIGN(sizeof(struct page) * PAGES_PER_SECTION);
383 }
384 
385 struct page __init *sparse_mem_map_populate(unsigned long pnum, int nid,
386 		struct vmem_altmap *altmap)
387 {
388 	unsigned long size = section_map_size();
389 	struct page *map = sparse_buffer_alloc(size);
390 
391 	if (map)
392 		return map;
393 
394 	map = memblock_virt_alloc_try_nid(size,
395 					  PAGE_SIZE, __pa(MAX_DMA_ADDRESS),
396 					  BOOTMEM_ALLOC_ACCESSIBLE, nid);
397 	return map;
398 }
399 #endif /* !CONFIG_SPARSEMEM_VMEMMAP */
400 
401 static void *sparsemap_buf __meminitdata;
402 static void *sparsemap_buf_end __meminitdata;
403 
404 static void __init sparse_buffer_init(unsigned long size, int nid)
405 {
406 	WARN_ON(sparsemap_buf);	/* forgot to call sparse_buffer_fini()? */
407 	sparsemap_buf =
408 		memblock_virt_alloc_try_nid_raw(size, PAGE_SIZE,
409 						__pa(MAX_DMA_ADDRESS),
410 						BOOTMEM_ALLOC_ACCESSIBLE, nid);
411 	sparsemap_buf_end = sparsemap_buf + size;
412 }
413 
414 static void __init sparse_buffer_fini(void)
415 {
416 	unsigned long size = sparsemap_buf_end - sparsemap_buf;
417 
418 	if (sparsemap_buf && size > 0)
419 		memblock_free_early(__pa(sparsemap_buf), size);
420 	sparsemap_buf = NULL;
421 }
422 
423 void * __meminit sparse_buffer_alloc(unsigned long size)
424 {
425 	void *ptr = NULL;
426 
427 	if (sparsemap_buf) {
428 		ptr = PTR_ALIGN(sparsemap_buf, size);
429 		if (ptr + size > sparsemap_buf_end)
430 			ptr = NULL;
431 		else
432 			sparsemap_buf = ptr + size;
433 	}
434 	return ptr;
435 }
436 
437 void __weak __meminit vmemmap_populate_print_last(void)
438 {
439 }
440 
441 /*
442  * Initialize sparse on a specific node. The node spans [pnum_begin, pnum_end)
443  * And number of present sections in this node is map_count.
444  */
445 static void __init sparse_init_nid(int nid, unsigned long pnum_begin,
446 				   unsigned long pnum_end,
447 				   unsigned long map_count)
448 {
449 	unsigned long pnum, usemap_longs, *usemap;
450 	struct page *map;
451 
452 	usemap_longs = BITS_TO_LONGS(SECTION_BLOCKFLAGS_BITS);
453 	usemap = sparse_early_usemaps_alloc_pgdat_section(NODE_DATA(nid),
454 							  usemap_size() *
455 							  map_count);
456 	if (!usemap) {
457 		pr_err("%s: node[%d] usemap allocation failed", __func__, nid);
458 		goto failed;
459 	}
460 	sparse_buffer_init(map_count * section_map_size(), nid);
461 	for_each_present_section_nr(pnum_begin, pnum) {
462 		if (pnum >= pnum_end)
463 			break;
464 
465 		map = sparse_mem_map_populate(pnum, nid, NULL);
466 		if (!map) {
467 			pr_err("%s: node[%d] memory map backing failed. Some memory will not be available.",
468 			       __func__, nid);
469 			pnum_begin = pnum;
470 			goto failed;
471 		}
472 		check_usemap_section_nr(nid, usemap);
473 		sparse_init_one_section(__nr_to_section(pnum), pnum, map, usemap);
474 		usemap += usemap_longs;
475 	}
476 	sparse_buffer_fini();
477 	return;
478 failed:
479 	/* We failed to allocate, mark all the following pnums as not present */
480 	for_each_present_section_nr(pnum_begin, pnum) {
481 		struct mem_section *ms;
482 
483 		if (pnum >= pnum_end)
484 			break;
485 		ms = __nr_to_section(pnum);
486 		ms->section_mem_map = 0;
487 	}
488 }
489 
490 /*
491  * Allocate the accumulated non-linear sections, allocate a mem_map
492  * for each and record the physical to section mapping.
493  */
494 void __init sparse_init(void)
495 {
496 	unsigned long pnum_begin = first_present_section_nr();
497 	int nid_begin = sparse_early_nid(__nr_to_section(pnum_begin));
498 	unsigned long pnum_end, map_count = 1;
499 
500 	/* Setup pageblock_order for HUGETLB_PAGE_SIZE_VARIABLE */
501 	set_pageblock_order();
502 
503 	for_each_present_section_nr(pnum_begin + 1, pnum_end) {
504 		int nid = sparse_early_nid(__nr_to_section(pnum_end));
505 
506 		if (nid == nid_begin) {
507 			map_count++;
508 			continue;
509 		}
510 		/* Init node with sections in range [pnum_begin, pnum_end) */
511 		sparse_init_nid(nid_begin, pnum_begin, pnum_end, map_count);
512 		nid_begin = nid;
513 		pnum_begin = pnum_end;
514 		map_count = 1;
515 	}
516 	/* cover the last node */
517 	sparse_init_nid(nid_begin, pnum_begin, pnum_end, map_count);
518 	vmemmap_populate_print_last();
519 }
520 
521 #ifdef CONFIG_MEMORY_HOTPLUG
522 
523 /* Mark all memory sections within the pfn range as online */
524 void online_mem_sections(unsigned long start_pfn, unsigned long end_pfn)
525 {
526 	unsigned long pfn;
527 
528 	for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
529 		unsigned long section_nr = pfn_to_section_nr(pfn);
530 		struct mem_section *ms;
531 
532 		/* onlining code should never touch invalid ranges */
533 		if (WARN_ON(!valid_section_nr(section_nr)))
534 			continue;
535 
536 		ms = __nr_to_section(section_nr);
537 		ms->section_mem_map |= SECTION_IS_ONLINE;
538 	}
539 }
540 
541 #ifdef CONFIG_MEMORY_HOTREMOVE
542 /* Mark all memory sections within the pfn range as online */
543 void offline_mem_sections(unsigned long start_pfn, unsigned long end_pfn)
544 {
545 	unsigned long pfn;
546 
547 	for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
548 		unsigned long section_nr = pfn_to_section_nr(pfn);
549 		struct mem_section *ms;
550 
551 		/*
552 		 * TODO this needs some double checking. Offlining code makes
553 		 * sure to check pfn_valid but those checks might be just bogus
554 		 */
555 		if (WARN_ON(!valid_section_nr(section_nr)))
556 			continue;
557 
558 		ms = __nr_to_section(section_nr);
559 		ms->section_mem_map &= ~SECTION_IS_ONLINE;
560 	}
561 }
562 #endif
563 
564 #ifdef CONFIG_SPARSEMEM_VMEMMAP
565 static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid,
566 		struct vmem_altmap *altmap)
567 {
568 	/* This will make the necessary allocations eventually. */
569 	return sparse_mem_map_populate(pnum, nid, altmap);
570 }
571 static void __kfree_section_memmap(struct page *memmap,
572 		struct vmem_altmap *altmap)
573 {
574 	unsigned long start = (unsigned long)memmap;
575 	unsigned long end = (unsigned long)(memmap + PAGES_PER_SECTION);
576 
577 	vmemmap_free(start, end, altmap);
578 }
579 #ifdef CONFIG_MEMORY_HOTREMOVE
580 static void free_map_bootmem(struct page *memmap)
581 {
582 	unsigned long start = (unsigned long)memmap;
583 	unsigned long end = (unsigned long)(memmap + PAGES_PER_SECTION);
584 
585 	vmemmap_free(start, end, NULL);
586 }
587 #endif /* CONFIG_MEMORY_HOTREMOVE */
588 #else
589 static struct page *__kmalloc_section_memmap(void)
590 {
591 	struct page *page, *ret;
592 	unsigned long memmap_size = sizeof(struct page) * PAGES_PER_SECTION;
593 
594 	page = alloc_pages(GFP_KERNEL|__GFP_NOWARN, get_order(memmap_size));
595 	if (page)
596 		goto got_map_page;
597 
598 	ret = vmalloc(memmap_size);
599 	if (ret)
600 		goto got_map_ptr;
601 
602 	return NULL;
603 got_map_page:
604 	ret = (struct page *)pfn_to_kaddr(page_to_pfn(page));
605 got_map_ptr:
606 
607 	return ret;
608 }
609 
610 static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid,
611 		struct vmem_altmap *altmap)
612 {
613 	return __kmalloc_section_memmap();
614 }
615 
616 static void __kfree_section_memmap(struct page *memmap,
617 		struct vmem_altmap *altmap)
618 {
619 	if (is_vmalloc_addr(memmap))
620 		vfree(memmap);
621 	else
622 		free_pages((unsigned long)memmap,
623 			   get_order(sizeof(struct page) * PAGES_PER_SECTION));
624 }
625 
626 #ifdef CONFIG_MEMORY_HOTREMOVE
627 static void free_map_bootmem(struct page *memmap)
628 {
629 	unsigned long maps_section_nr, removing_section_nr, i;
630 	unsigned long magic, nr_pages;
631 	struct page *page = virt_to_page(memmap);
632 
633 	nr_pages = PAGE_ALIGN(PAGES_PER_SECTION * sizeof(struct page))
634 		>> PAGE_SHIFT;
635 
636 	for (i = 0; i < nr_pages; i++, page++) {
637 		magic = (unsigned long) page->freelist;
638 
639 		BUG_ON(magic == NODE_INFO);
640 
641 		maps_section_nr = pfn_to_section_nr(page_to_pfn(page));
642 		removing_section_nr = page_private(page);
643 
644 		/*
645 		 * When this function is called, the removing section is
646 		 * logical offlined state. This means all pages are isolated
647 		 * from page allocator. If removing section's memmap is placed
648 		 * on the same section, it must not be freed.
649 		 * If it is freed, page allocator may allocate it which will
650 		 * be removed physically soon.
651 		 */
652 		if (maps_section_nr != removing_section_nr)
653 			put_page_bootmem(page);
654 	}
655 }
656 #endif /* CONFIG_MEMORY_HOTREMOVE */
657 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
658 
659 /*
660  * returns the number of sections whose mem_maps were properly
661  * set.  If this is <=0, then that means that the passed-in
662  * map was not consumed and must be freed.
663  */
664 int __meminit sparse_add_one_section(struct pglist_data *pgdat,
665 		unsigned long start_pfn, struct vmem_altmap *altmap)
666 {
667 	unsigned long section_nr = pfn_to_section_nr(start_pfn);
668 	struct mem_section *ms;
669 	struct page *memmap;
670 	unsigned long *usemap;
671 	unsigned long flags;
672 	int ret;
673 
674 	/*
675 	 * no locking for this, because it does its own
676 	 * plus, it does a kmalloc
677 	 */
678 	ret = sparse_index_init(section_nr, pgdat->node_id);
679 	if (ret < 0 && ret != -EEXIST)
680 		return ret;
681 	ret = 0;
682 	memmap = kmalloc_section_memmap(section_nr, pgdat->node_id, altmap);
683 	if (!memmap)
684 		return -ENOMEM;
685 	usemap = __kmalloc_section_usemap();
686 	if (!usemap) {
687 		__kfree_section_memmap(memmap, altmap);
688 		return -ENOMEM;
689 	}
690 
691 	pgdat_resize_lock(pgdat, &flags);
692 
693 	ms = __pfn_to_section(start_pfn);
694 	if (ms->section_mem_map & SECTION_MARKED_PRESENT) {
695 		ret = -EEXIST;
696 		goto out;
697 	}
698 
699 #ifdef CONFIG_DEBUG_VM
700 	/*
701 	 * Poison uninitialized struct pages in order to catch invalid flags
702 	 * combinations.
703 	 */
704 	memset(memmap, PAGE_POISON_PATTERN, sizeof(struct page) * PAGES_PER_SECTION);
705 #endif
706 
707 	section_mark_present(ms);
708 	sparse_init_one_section(ms, section_nr, memmap, usemap);
709 
710 out:
711 	pgdat_resize_unlock(pgdat, &flags);
712 	if (ret < 0) {
713 		kfree(usemap);
714 		__kfree_section_memmap(memmap, altmap);
715 	}
716 	return ret;
717 }
718 
719 #ifdef CONFIG_MEMORY_HOTREMOVE
720 #ifdef CONFIG_MEMORY_FAILURE
721 static void clear_hwpoisoned_pages(struct page *memmap, int nr_pages)
722 {
723 	int i;
724 
725 	if (!memmap)
726 		return;
727 
728 	for (i = 0; i < nr_pages; i++) {
729 		if (PageHWPoison(&memmap[i])) {
730 			atomic_long_sub(1, &num_poisoned_pages);
731 			ClearPageHWPoison(&memmap[i]);
732 		}
733 	}
734 }
735 #else
736 static inline void clear_hwpoisoned_pages(struct page *memmap, int nr_pages)
737 {
738 }
739 #endif
740 
741 static void free_section_usemap(struct page *memmap, unsigned long *usemap,
742 		struct vmem_altmap *altmap)
743 {
744 	struct page *usemap_page;
745 
746 	if (!usemap)
747 		return;
748 
749 	usemap_page = virt_to_page(usemap);
750 	/*
751 	 * Check to see if allocation came from hot-plug-add
752 	 */
753 	if (PageSlab(usemap_page) || PageCompound(usemap_page)) {
754 		kfree(usemap);
755 		if (memmap)
756 			__kfree_section_memmap(memmap, altmap);
757 		return;
758 	}
759 
760 	/*
761 	 * The usemap came from bootmem. This is packed with other usemaps
762 	 * on the section which has pgdat at boot time. Just keep it as is now.
763 	 */
764 
765 	if (memmap)
766 		free_map_bootmem(memmap);
767 }
768 
769 void sparse_remove_one_section(struct zone *zone, struct mem_section *ms,
770 		unsigned long map_offset, struct vmem_altmap *altmap)
771 {
772 	struct page *memmap = NULL;
773 	unsigned long *usemap = NULL, flags;
774 	struct pglist_data *pgdat = zone->zone_pgdat;
775 
776 	pgdat_resize_lock(pgdat, &flags);
777 	if (ms->section_mem_map) {
778 		usemap = ms->pageblock_flags;
779 		memmap = sparse_decode_mem_map(ms->section_mem_map,
780 						__section_nr(ms));
781 		ms->section_mem_map = 0;
782 		ms->pageblock_flags = NULL;
783 	}
784 	pgdat_resize_unlock(pgdat, &flags);
785 
786 	clear_hwpoisoned_pages(memmap + map_offset,
787 			PAGES_PER_SECTION - map_offset);
788 	free_section_usemap(memmap, usemap, altmap);
789 }
790 #endif /* CONFIG_MEMORY_HOTREMOVE */
791 #endif /* CONFIG_MEMORY_HOTPLUG */
792