xref: /openbmc/linux/mm/sparse.c (revision 022dacdd)
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 #include <linux/swap.h>
15 #include <linux/swapops.h>
16 
17 #include "internal.h"
18 #include <asm/dma.h>
19 #include <asm/pgalloc.h>
20 #include <asm/pgtable.h>
21 
22 /*
23  * Permanent SPARSEMEM data:
24  *
25  * 1) mem_section	- memory sections, mem_map's for valid memory
26  */
27 #ifdef CONFIG_SPARSEMEM_EXTREME
28 struct mem_section **mem_section;
29 #else
30 struct mem_section mem_section[NR_SECTION_ROOTS][SECTIONS_PER_ROOT]
31 	____cacheline_internodealigned_in_smp;
32 #endif
33 EXPORT_SYMBOL(mem_section);
34 
35 #ifdef NODE_NOT_IN_PAGE_FLAGS
36 /*
37  * If we did not store the node number in the page then we have to
38  * do a lookup in the section_to_node_table in order to find which
39  * node the page belongs to.
40  */
41 #if MAX_NUMNODES <= 256
42 static u8 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned;
43 #else
44 static u16 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned;
45 #endif
46 
47 int page_to_nid(const struct page *page)
48 {
49 	return section_to_node_table[page_to_section(page)];
50 }
51 EXPORT_SYMBOL(page_to_nid);
52 
53 static void set_section_nid(unsigned long section_nr, int nid)
54 {
55 	section_to_node_table[section_nr] = nid;
56 }
57 #else /* !NODE_NOT_IN_PAGE_FLAGS */
58 static inline void set_section_nid(unsigned long section_nr, int nid)
59 {
60 }
61 #endif
62 
63 #ifdef CONFIG_SPARSEMEM_EXTREME
64 static noinline struct mem_section __ref *sparse_index_alloc(int nid)
65 {
66 	struct mem_section *section = NULL;
67 	unsigned long array_size = SECTIONS_PER_ROOT *
68 				   sizeof(struct mem_section);
69 
70 	if (slab_is_available()) {
71 		section = kzalloc_node(array_size, GFP_KERNEL, nid);
72 	} else {
73 		section = memblock_alloc_node(array_size, SMP_CACHE_BYTES,
74 					      nid);
75 		if (!section)
76 			panic("%s: Failed to allocate %lu bytes nid=%d\n",
77 			      __func__, array_size, nid);
78 	}
79 
80 	return section;
81 }
82 
83 static int __meminit sparse_index_init(unsigned long section_nr, int nid)
84 {
85 	unsigned long root = SECTION_NR_TO_ROOT(section_nr);
86 	struct mem_section *section;
87 
88 	/*
89 	 * An existing section is possible in the sub-section hotplug
90 	 * case. First hot-add instantiates, follow-on hot-add reuses
91 	 * the existing section.
92 	 *
93 	 * The mem_hotplug_lock resolves the apparent race below.
94 	 */
95 	if (mem_section[root])
96 		return 0;
97 
98 	section = sparse_index_alloc(nid);
99 	if (!section)
100 		return -ENOMEM;
101 
102 	mem_section[root] = section;
103 
104 	return 0;
105 }
106 #else /* !SPARSEMEM_EXTREME */
107 static inline int sparse_index_init(unsigned long section_nr, int nid)
108 {
109 	return 0;
110 }
111 #endif
112 
113 #ifdef CONFIG_SPARSEMEM_EXTREME
114 unsigned long __section_nr(struct mem_section *ms)
115 {
116 	unsigned long root_nr;
117 	struct mem_section *root = NULL;
118 
119 	for (root_nr = 0; root_nr < NR_SECTION_ROOTS; root_nr++) {
120 		root = __nr_to_section(root_nr * SECTIONS_PER_ROOT);
121 		if (!root)
122 			continue;
123 
124 		if ((ms >= root) && (ms < (root + SECTIONS_PER_ROOT)))
125 		     break;
126 	}
127 
128 	VM_BUG_ON(!root);
129 
130 	return (root_nr * SECTIONS_PER_ROOT) + (ms - root);
131 }
132 #else
133 unsigned long __section_nr(struct mem_section *ms)
134 {
135 	return (unsigned long)(ms - mem_section[0]);
136 }
137 #endif
138 
139 /*
140  * During early boot, before section_mem_map is used for an actual
141  * mem_map, we use section_mem_map to store the section's NUMA
142  * node.  This keeps us from having to use another data structure.  The
143  * node information is cleared just before we store the real mem_map.
144  */
145 static inline unsigned long sparse_encode_early_nid(int nid)
146 {
147 	return (nid << SECTION_NID_SHIFT);
148 }
149 
150 static inline int sparse_early_nid(struct mem_section *section)
151 {
152 	return (section->section_mem_map >> SECTION_NID_SHIFT);
153 }
154 
155 /* Validate the physical addressing limitations of the model */
156 void __meminit mminit_validate_memmodel_limits(unsigned long *start_pfn,
157 						unsigned long *end_pfn)
158 {
159 	unsigned long max_sparsemem_pfn = 1UL << (MAX_PHYSMEM_BITS-PAGE_SHIFT);
160 
161 	/*
162 	 * Sanity checks - do not allow an architecture to pass
163 	 * in larger pfns than the maximum scope of sparsemem:
164 	 */
165 	if (*start_pfn > max_sparsemem_pfn) {
166 		mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
167 			"Start of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
168 			*start_pfn, *end_pfn, max_sparsemem_pfn);
169 		WARN_ON_ONCE(1);
170 		*start_pfn = max_sparsemem_pfn;
171 		*end_pfn = max_sparsemem_pfn;
172 	} else if (*end_pfn > max_sparsemem_pfn) {
173 		mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
174 			"End of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
175 			*start_pfn, *end_pfn, max_sparsemem_pfn);
176 		WARN_ON_ONCE(1);
177 		*end_pfn = max_sparsemem_pfn;
178 	}
179 }
180 
181 /*
182  * There are a number of times that we loop over NR_MEM_SECTIONS,
183  * looking for section_present() on each.  But, when we have very
184  * large physical address spaces, NR_MEM_SECTIONS can also be
185  * very large which makes the loops quite long.
186  *
187  * Keeping track of this gives us an easy way to break out of
188  * those loops early.
189  */
190 unsigned long __highest_present_section_nr;
191 static void section_mark_present(struct mem_section *ms)
192 {
193 	unsigned long section_nr = __section_nr(ms);
194 
195 	if (section_nr > __highest_present_section_nr)
196 		__highest_present_section_nr = section_nr;
197 
198 	ms->section_mem_map |= SECTION_MARKED_PRESENT;
199 }
200 
201 #define for_each_present_section_nr(start, section_nr)		\
202 	for (section_nr = next_present_section_nr(start-1);	\
203 	     ((section_nr != -1) &&				\
204 	      (section_nr <= __highest_present_section_nr));	\
205 	     section_nr = next_present_section_nr(section_nr))
206 
207 static inline unsigned long first_present_section_nr(void)
208 {
209 	return next_present_section_nr(-1);
210 }
211 
212 #ifdef CONFIG_SPARSEMEM_VMEMMAP
213 static void subsection_mask_set(unsigned long *map, unsigned long pfn,
214 		unsigned long nr_pages)
215 {
216 	int idx = subsection_map_index(pfn);
217 	int end = subsection_map_index(pfn + nr_pages - 1);
218 
219 	bitmap_set(map, idx, end - idx + 1);
220 }
221 
222 void __init subsection_map_init(unsigned long pfn, unsigned long nr_pages)
223 {
224 	int end_sec = pfn_to_section_nr(pfn + nr_pages - 1);
225 	unsigned long nr, start_sec = pfn_to_section_nr(pfn);
226 
227 	if (!nr_pages)
228 		return;
229 
230 	for (nr = start_sec; nr <= end_sec; nr++) {
231 		struct mem_section *ms;
232 		unsigned long pfns;
233 
234 		pfns = min(nr_pages, PAGES_PER_SECTION
235 				- (pfn & ~PAGE_SECTION_MASK));
236 		ms = __nr_to_section(nr);
237 		subsection_mask_set(ms->usage->subsection_map, pfn, pfns);
238 
239 		pr_debug("%s: sec: %lu pfns: %lu set(%d, %d)\n", __func__, nr,
240 				pfns, subsection_map_index(pfn),
241 				subsection_map_index(pfn + pfns - 1));
242 
243 		pfn += pfns;
244 		nr_pages -= pfns;
245 	}
246 }
247 #else
248 void __init subsection_map_init(unsigned long pfn, unsigned long nr_pages)
249 {
250 }
251 #endif
252 
253 /* Record a memory area against a node. */
254 void __init memory_present(int nid, unsigned long start, unsigned long end)
255 {
256 	unsigned long pfn;
257 
258 #ifdef CONFIG_SPARSEMEM_EXTREME
259 	if (unlikely(!mem_section)) {
260 		unsigned long size, align;
261 
262 		size = sizeof(struct mem_section*) * NR_SECTION_ROOTS;
263 		align = 1 << (INTERNODE_CACHE_SHIFT);
264 		mem_section = memblock_alloc(size, align);
265 		if (!mem_section)
266 			panic("%s: Failed to allocate %lu bytes align=0x%lx\n",
267 			      __func__, size, align);
268 	}
269 #endif
270 
271 	start &= PAGE_SECTION_MASK;
272 	mminit_validate_memmodel_limits(&start, &end);
273 	for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) {
274 		unsigned long section = pfn_to_section_nr(pfn);
275 		struct mem_section *ms;
276 
277 		sparse_index_init(section, nid);
278 		set_section_nid(section, nid);
279 
280 		ms = __nr_to_section(section);
281 		if (!ms->section_mem_map) {
282 			ms->section_mem_map = sparse_encode_early_nid(nid) |
283 							SECTION_IS_ONLINE;
284 			section_mark_present(ms);
285 		}
286 	}
287 }
288 
289 /*
290  * Mark all memblocks as present using memory_present(). This is a
291  * convienence function that is useful for a number of arches
292  * to mark all of the systems memory as present during initialization.
293  */
294 void __init memblocks_present(void)
295 {
296 	struct memblock_region *reg;
297 
298 	for_each_memblock(memory, reg) {
299 		memory_present(memblock_get_region_node(reg),
300 			       memblock_region_memory_base_pfn(reg),
301 			       memblock_region_memory_end_pfn(reg));
302 	}
303 }
304 
305 /*
306  * Subtle, we encode the real pfn into the mem_map such that
307  * the identity pfn - section_mem_map will return the actual
308  * physical page frame number.
309  */
310 static unsigned long sparse_encode_mem_map(struct page *mem_map, unsigned long pnum)
311 {
312 	unsigned long coded_mem_map =
313 		(unsigned long)(mem_map - (section_nr_to_pfn(pnum)));
314 	BUILD_BUG_ON(SECTION_MAP_LAST_BIT > (1UL<<PFN_SECTION_SHIFT));
315 	BUG_ON(coded_mem_map & ~SECTION_MAP_MASK);
316 	return coded_mem_map;
317 }
318 
319 /*
320  * Decode mem_map from the coded memmap
321  */
322 struct page *sparse_decode_mem_map(unsigned long coded_mem_map, unsigned long pnum)
323 {
324 	/* mask off the extra low bits of information */
325 	coded_mem_map &= SECTION_MAP_MASK;
326 	return ((struct page *)coded_mem_map) + section_nr_to_pfn(pnum);
327 }
328 
329 static void __meminit sparse_init_one_section(struct mem_section *ms,
330 		unsigned long pnum, struct page *mem_map,
331 		struct mem_section_usage *usage, unsigned long flags)
332 {
333 	ms->section_mem_map &= ~SECTION_MAP_MASK;
334 	ms->section_mem_map |= sparse_encode_mem_map(mem_map, pnum)
335 		| SECTION_HAS_MEM_MAP | flags;
336 	ms->usage = usage;
337 }
338 
339 static unsigned long usemap_size(void)
340 {
341 	return BITS_TO_LONGS(SECTION_BLOCKFLAGS_BITS) * sizeof(unsigned long);
342 }
343 
344 size_t mem_section_usage_size(void)
345 {
346 	return sizeof(struct mem_section_usage) + usemap_size();
347 }
348 
349 #ifdef CONFIG_MEMORY_HOTREMOVE
350 static struct mem_section_usage * __init
351 sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
352 					 unsigned long size)
353 {
354 	struct mem_section_usage *usage;
355 	unsigned long goal, limit;
356 	int nid;
357 	/*
358 	 * A page may contain usemaps for other sections preventing the
359 	 * page being freed and making a section unremovable while
360 	 * other sections referencing the usemap remain active. Similarly,
361 	 * a pgdat can prevent a section being removed. If section A
362 	 * contains a pgdat and section B contains the usemap, both
363 	 * sections become inter-dependent. This allocates usemaps
364 	 * from the same section as the pgdat where possible to avoid
365 	 * this problem.
366 	 */
367 	goal = __pa(pgdat) & (PAGE_SECTION_MASK << PAGE_SHIFT);
368 	limit = goal + (1UL << PA_SECTION_SHIFT);
369 	nid = early_pfn_to_nid(goal >> PAGE_SHIFT);
370 again:
371 	usage = memblock_alloc_try_nid(size, SMP_CACHE_BYTES, goal, limit, nid);
372 	if (!usage && limit) {
373 		limit = 0;
374 		goto again;
375 	}
376 	return usage;
377 }
378 
379 static void __init check_usemap_section_nr(int nid,
380 		struct mem_section_usage *usage)
381 {
382 	unsigned long usemap_snr, pgdat_snr;
383 	static unsigned long old_usemap_snr;
384 	static unsigned long old_pgdat_snr;
385 	struct pglist_data *pgdat = NODE_DATA(nid);
386 	int usemap_nid;
387 
388 	/* First call */
389 	if (!old_usemap_snr) {
390 		old_usemap_snr = NR_MEM_SECTIONS;
391 		old_pgdat_snr = NR_MEM_SECTIONS;
392 	}
393 
394 	usemap_snr = pfn_to_section_nr(__pa(usage) >> PAGE_SHIFT);
395 	pgdat_snr = pfn_to_section_nr(__pa(pgdat) >> PAGE_SHIFT);
396 	if (usemap_snr == pgdat_snr)
397 		return;
398 
399 	if (old_usemap_snr == usemap_snr && old_pgdat_snr == pgdat_snr)
400 		/* skip redundant message */
401 		return;
402 
403 	old_usemap_snr = usemap_snr;
404 	old_pgdat_snr = pgdat_snr;
405 
406 	usemap_nid = sparse_early_nid(__nr_to_section(usemap_snr));
407 	if (usemap_nid != nid) {
408 		pr_info("node %d must be removed before remove section %ld\n",
409 			nid, usemap_snr);
410 		return;
411 	}
412 	/*
413 	 * There is a circular dependency.
414 	 * Some platforms allow un-removable section because they will just
415 	 * gather other removable sections for dynamic partitioning.
416 	 * Just notify un-removable section's number here.
417 	 */
418 	pr_info("Section %ld and %ld (node %d) have a circular dependency on usemap and pgdat allocations\n",
419 		usemap_snr, pgdat_snr, nid);
420 }
421 #else
422 static struct mem_section_usage * __init
423 sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
424 					 unsigned long size)
425 {
426 	return memblock_alloc_node(size, SMP_CACHE_BYTES, pgdat->node_id);
427 }
428 
429 static void __init check_usemap_section_nr(int nid,
430 		struct mem_section_usage *usage)
431 {
432 }
433 #endif /* CONFIG_MEMORY_HOTREMOVE */
434 
435 #ifdef CONFIG_SPARSEMEM_VMEMMAP
436 static unsigned long __init section_map_size(void)
437 {
438 	return ALIGN(sizeof(struct page) * PAGES_PER_SECTION, PMD_SIZE);
439 }
440 
441 #else
442 static unsigned long __init section_map_size(void)
443 {
444 	return PAGE_ALIGN(sizeof(struct page) * PAGES_PER_SECTION);
445 }
446 
447 struct page __init *__populate_section_memmap(unsigned long pfn,
448 		unsigned long nr_pages, int nid, struct vmem_altmap *altmap)
449 {
450 	unsigned long size = section_map_size();
451 	struct page *map = sparse_buffer_alloc(size);
452 	phys_addr_t addr = __pa(MAX_DMA_ADDRESS);
453 
454 	if (map)
455 		return map;
456 
457 	map = memblock_alloc_try_nid_raw(size, size, addr,
458 					  MEMBLOCK_ALLOC_ACCESSIBLE, nid);
459 	if (!map)
460 		panic("%s: Failed to allocate %lu bytes align=0x%lx nid=%d from=%pa\n",
461 		      __func__, size, PAGE_SIZE, nid, &addr);
462 
463 	return map;
464 }
465 #endif /* !CONFIG_SPARSEMEM_VMEMMAP */
466 
467 static void *sparsemap_buf __meminitdata;
468 static void *sparsemap_buf_end __meminitdata;
469 
470 static inline void __meminit sparse_buffer_free(unsigned long size)
471 {
472 	WARN_ON(!sparsemap_buf || size == 0);
473 	memblock_free_early(__pa(sparsemap_buf), size);
474 }
475 
476 static void __init sparse_buffer_init(unsigned long size, int nid)
477 {
478 	phys_addr_t addr = __pa(MAX_DMA_ADDRESS);
479 	WARN_ON(sparsemap_buf);	/* forgot to call sparse_buffer_fini()? */
480 	/*
481 	 * Pre-allocated buffer is mainly used by __populate_section_memmap
482 	 * and we want it to be properly aligned to the section size - this is
483 	 * especially the case for VMEMMAP which maps memmap to PMDs
484 	 */
485 	sparsemap_buf = memblock_alloc_exact_nid_raw(size, section_map_size(),
486 					addr, MEMBLOCK_ALLOC_ACCESSIBLE, nid);
487 	sparsemap_buf_end = sparsemap_buf + size;
488 }
489 
490 static void __init sparse_buffer_fini(void)
491 {
492 	unsigned long size = sparsemap_buf_end - sparsemap_buf;
493 
494 	if (sparsemap_buf && size > 0)
495 		sparse_buffer_free(size);
496 	sparsemap_buf = NULL;
497 }
498 
499 void * __meminit sparse_buffer_alloc(unsigned long size)
500 {
501 	void *ptr = NULL;
502 
503 	if (sparsemap_buf) {
504 		ptr = (void *) roundup((unsigned long)sparsemap_buf, size);
505 		if (ptr + size > sparsemap_buf_end)
506 			ptr = NULL;
507 		else {
508 			/* Free redundant aligned space */
509 			if ((unsigned long)(ptr - sparsemap_buf) > 0)
510 				sparse_buffer_free((unsigned long)(ptr - sparsemap_buf));
511 			sparsemap_buf = ptr + size;
512 		}
513 	}
514 	return ptr;
515 }
516 
517 void __weak __meminit vmemmap_populate_print_last(void)
518 {
519 }
520 
521 /*
522  * Initialize sparse on a specific node. The node spans [pnum_begin, pnum_end)
523  * And number of present sections in this node is map_count.
524  */
525 static void __init sparse_init_nid(int nid, unsigned long pnum_begin,
526 				   unsigned long pnum_end,
527 				   unsigned long map_count)
528 {
529 	struct mem_section_usage *usage;
530 	unsigned long pnum;
531 	struct page *map;
532 
533 	usage = sparse_early_usemaps_alloc_pgdat_section(NODE_DATA(nid),
534 			mem_section_usage_size() * map_count);
535 	if (!usage) {
536 		pr_err("%s: node[%d] usemap allocation failed", __func__, nid);
537 		goto failed;
538 	}
539 	sparse_buffer_init(map_count * section_map_size(), nid);
540 	for_each_present_section_nr(pnum_begin, pnum) {
541 		unsigned long pfn = section_nr_to_pfn(pnum);
542 
543 		if (pnum >= pnum_end)
544 			break;
545 
546 		map = __populate_section_memmap(pfn, PAGES_PER_SECTION,
547 				nid, NULL);
548 		if (!map) {
549 			pr_err("%s: node[%d] memory map backing failed. Some memory will not be available.",
550 			       __func__, nid);
551 			pnum_begin = pnum;
552 			goto failed;
553 		}
554 		check_usemap_section_nr(nid, usage);
555 		sparse_init_one_section(__nr_to_section(pnum), pnum, map, usage,
556 				SECTION_IS_EARLY);
557 		usage = (void *) usage + mem_section_usage_size();
558 	}
559 	sparse_buffer_fini();
560 	return;
561 failed:
562 	/* We failed to allocate, mark all the following pnums as not present */
563 	for_each_present_section_nr(pnum_begin, pnum) {
564 		struct mem_section *ms;
565 
566 		if (pnum >= pnum_end)
567 			break;
568 		ms = __nr_to_section(pnum);
569 		ms->section_mem_map = 0;
570 	}
571 }
572 
573 /*
574  * Allocate the accumulated non-linear sections, allocate a mem_map
575  * for each and record the physical to section mapping.
576  */
577 void __init sparse_init(void)
578 {
579 	unsigned long pnum_begin = first_present_section_nr();
580 	int nid_begin = sparse_early_nid(__nr_to_section(pnum_begin));
581 	unsigned long pnum_end, map_count = 1;
582 
583 	/* Setup pageblock_order for HUGETLB_PAGE_SIZE_VARIABLE */
584 	set_pageblock_order();
585 
586 	for_each_present_section_nr(pnum_begin + 1, pnum_end) {
587 		int nid = sparse_early_nid(__nr_to_section(pnum_end));
588 
589 		if (nid == nid_begin) {
590 			map_count++;
591 			continue;
592 		}
593 		/* Init node with sections in range [pnum_begin, pnum_end) */
594 		sparse_init_nid(nid_begin, pnum_begin, pnum_end, map_count);
595 		nid_begin = nid;
596 		pnum_begin = pnum_end;
597 		map_count = 1;
598 	}
599 	/* cover the last node */
600 	sparse_init_nid(nid_begin, pnum_begin, pnum_end, map_count);
601 	vmemmap_populate_print_last();
602 }
603 
604 #ifdef CONFIG_MEMORY_HOTPLUG
605 
606 /* Mark all memory sections within the pfn range as online */
607 void online_mem_sections(unsigned long start_pfn, unsigned long end_pfn)
608 {
609 	unsigned long pfn;
610 
611 	for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
612 		unsigned long section_nr = pfn_to_section_nr(pfn);
613 		struct mem_section *ms;
614 
615 		/* onlining code should never touch invalid ranges */
616 		if (WARN_ON(!valid_section_nr(section_nr)))
617 			continue;
618 
619 		ms = __nr_to_section(section_nr);
620 		ms->section_mem_map |= SECTION_IS_ONLINE;
621 	}
622 }
623 
624 #ifdef CONFIG_MEMORY_HOTREMOVE
625 /* Mark all memory sections within the pfn range as offline */
626 void offline_mem_sections(unsigned long start_pfn, unsigned long end_pfn)
627 {
628 	unsigned long pfn;
629 
630 	for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
631 		unsigned long section_nr = pfn_to_section_nr(pfn);
632 		struct mem_section *ms;
633 
634 		/*
635 		 * TODO this needs some double checking. Offlining code makes
636 		 * sure to check pfn_valid but those checks might be just bogus
637 		 */
638 		if (WARN_ON(!valid_section_nr(section_nr)))
639 			continue;
640 
641 		ms = __nr_to_section(section_nr);
642 		ms->section_mem_map &= ~SECTION_IS_ONLINE;
643 	}
644 }
645 #endif
646 
647 #ifdef CONFIG_SPARSEMEM_VMEMMAP
648 static struct page * __meminit populate_section_memmap(unsigned long pfn,
649 		unsigned long nr_pages, int nid, struct vmem_altmap *altmap)
650 {
651 	return __populate_section_memmap(pfn, nr_pages, nid, altmap);
652 }
653 
654 static void depopulate_section_memmap(unsigned long pfn, unsigned long nr_pages,
655 		struct vmem_altmap *altmap)
656 {
657 	unsigned long start = (unsigned long) pfn_to_page(pfn);
658 	unsigned long end = start + nr_pages * sizeof(struct page);
659 
660 	vmemmap_free(start, end, altmap);
661 }
662 static void free_map_bootmem(struct page *memmap)
663 {
664 	unsigned long start = (unsigned long)memmap;
665 	unsigned long end = (unsigned long)(memmap + PAGES_PER_SECTION);
666 
667 	vmemmap_free(start, end, NULL);
668 }
669 
670 static int clear_subsection_map(unsigned long pfn, unsigned long nr_pages)
671 {
672 	DECLARE_BITMAP(map, SUBSECTIONS_PER_SECTION) = { 0 };
673 	DECLARE_BITMAP(tmp, SUBSECTIONS_PER_SECTION) = { 0 };
674 	struct mem_section *ms = __pfn_to_section(pfn);
675 	unsigned long *subsection_map = ms->usage
676 		? &ms->usage->subsection_map[0] : NULL;
677 
678 	subsection_mask_set(map, pfn, nr_pages);
679 	if (subsection_map)
680 		bitmap_and(tmp, map, subsection_map, SUBSECTIONS_PER_SECTION);
681 
682 	if (WARN(!subsection_map || !bitmap_equal(tmp, map, SUBSECTIONS_PER_SECTION),
683 				"section already deactivated (%#lx + %ld)\n",
684 				pfn, nr_pages))
685 		return -EINVAL;
686 
687 	bitmap_xor(subsection_map, map, subsection_map, SUBSECTIONS_PER_SECTION);
688 	return 0;
689 }
690 
691 static bool is_subsection_map_empty(struct mem_section *ms)
692 {
693 	return bitmap_empty(&ms->usage->subsection_map[0],
694 			    SUBSECTIONS_PER_SECTION);
695 }
696 
697 static int fill_subsection_map(unsigned long pfn, unsigned long nr_pages)
698 {
699 	struct mem_section *ms = __pfn_to_section(pfn);
700 	DECLARE_BITMAP(map, SUBSECTIONS_PER_SECTION) = { 0 };
701 	unsigned long *subsection_map;
702 	int rc = 0;
703 
704 	subsection_mask_set(map, pfn, nr_pages);
705 
706 	subsection_map = &ms->usage->subsection_map[0];
707 
708 	if (bitmap_empty(map, SUBSECTIONS_PER_SECTION))
709 		rc = -EINVAL;
710 	else if (bitmap_intersects(map, subsection_map, SUBSECTIONS_PER_SECTION))
711 		rc = -EEXIST;
712 	else
713 		bitmap_or(subsection_map, map, subsection_map,
714 				SUBSECTIONS_PER_SECTION);
715 
716 	return rc;
717 }
718 #else
719 struct page * __meminit populate_section_memmap(unsigned long pfn,
720 		unsigned long nr_pages, int nid, struct vmem_altmap *altmap)
721 {
722 	return kvmalloc_node(array_size(sizeof(struct page),
723 					PAGES_PER_SECTION), GFP_KERNEL, nid);
724 }
725 
726 static void depopulate_section_memmap(unsigned long pfn, unsigned long nr_pages,
727 		struct vmem_altmap *altmap)
728 {
729 	kvfree(pfn_to_page(pfn));
730 }
731 
732 static void free_map_bootmem(struct page *memmap)
733 {
734 	unsigned long maps_section_nr, removing_section_nr, i;
735 	unsigned long magic, nr_pages;
736 	struct page *page = virt_to_page(memmap);
737 
738 	nr_pages = PAGE_ALIGN(PAGES_PER_SECTION * sizeof(struct page))
739 		>> PAGE_SHIFT;
740 
741 	for (i = 0; i < nr_pages; i++, page++) {
742 		magic = (unsigned long) page->freelist;
743 
744 		BUG_ON(magic == NODE_INFO);
745 
746 		maps_section_nr = pfn_to_section_nr(page_to_pfn(page));
747 		removing_section_nr = page_private(page);
748 
749 		/*
750 		 * When this function is called, the removing section is
751 		 * logical offlined state. This means all pages are isolated
752 		 * from page allocator. If removing section's memmap is placed
753 		 * on the same section, it must not be freed.
754 		 * If it is freed, page allocator may allocate it which will
755 		 * be removed physically soon.
756 		 */
757 		if (maps_section_nr != removing_section_nr)
758 			put_page_bootmem(page);
759 	}
760 }
761 
762 static int clear_subsection_map(unsigned long pfn, unsigned long nr_pages)
763 {
764 	return 0;
765 }
766 
767 static bool is_subsection_map_empty(struct mem_section *ms)
768 {
769 	return true;
770 }
771 
772 static int fill_subsection_map(unsigned long pfn, unsigned long nr_pages)
773 {
774 	return 0;
775 }
776 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
777 
778 /*
779  * To deactivate a memory region, there are 3 cases to handle across
780  * two configurations (SPARSEMEM_VMEMMAP={y,n}):
781  *
782  * 1. deactivation of a partial hot-added section (only possible in
783  *    the SPARSEMEM_VMEMMAP=y case).
784  *      a) section was present at memory init.
785  *      b) section was hot-added post memory init.
786  * 2. deactivation of a complete hot-added section.
787  * 3. deactivation of a complete section from memory init.
788  *
789  * For 1, when subsection_map does not empty we will not be freeing the
790  * usage map, but still need to free the vmemmap range.
791  *
792  * For 2 and 3, the SPARSEMEM_VMEMMAP={y,n} cases are unified
793  */
794 static void section_deactivate(unsigned long pfn, unsigned long nr_pages,
795 		struct vmem_altmap *altmap)
796 {
797 	struct mem_section *ms = __pfn_to_section(pfn);
798 	bool section_is_early = early_section(ms);
799 	struct page *memmap = NULL;
800 	bool empty;
801 
802 	if (clear_subsection_map(pfn, nr_pages))
803 		return;
804 
805 	empty = is_subsection_map_empty(ms);
806 	if (empty) {
807 		unsigned long section_nr = pfn_to_section_nr(pfn);
808 
809 		/*
810 		 * When removing an early section, the usage map is kept (as the
811 		 * usage maps of other sections fall into the same page). It
812 		 * will be re-used when re-adding the section - which is then no
813 		 * longer an early section. If the usage map is PageReserved, it
814 		 * was allocated during boot.
815 		 */
816 		if (!PageReserved(virt_to_page(ms->usage))) {
817 			kfree(ms->usage);
818 			ms->usage = NULL;
819 		}
820 		memmap = sparse_decode_mem_map(ms->section_mem_map, section_nr);
821 		/*
822 		 * Mark the section invalid so that valid_section()
823 		 * return false. This prevents code from dereferencing
824 		 * ms->usage array.
825 		 */
826 		ms->section_mem_map &= ~SECTION_HAS_MEM_MAP;
827 	}
828 
829 	if (section_is_early && memmap)
830 		free_map_bootmem(memmap);
831 	else
832 		depopulate_section_memmap(pfn, nr_pages, altmap);
833 
834 	if (empty)
835 		ms->section_mem_map = (unsigned long)NULL;
836 }
837 
838 static struct page * __meminit section_activate(int nid, unsigned long pfn,
839 		unsigned long nr_pages, struct vmem_altmap *altmap)
840 {
841 	struct mem_section *ms = __pfn_to_section(pfn);
842 	struct mem_section_usage *usage = NULL;
843 	struct page *memmap;
844 	int rc = 0;
845 
846 	if (!ms->usage) {
847 		usage = kzalloc(mem_section_usage_size(), GFP_KERNEL);
848 		if (!usage)
849 			return ERR_PTR(-ENOMEM);
850 		ms->usage = usage;
851 	}
852 
853 	rc = fill_subsection_map(pfn, nr_pages);
854 	if (rc) {
855 		if (usage)
856 			ms->usage = NULL;
857 		kfree(usage);
858 		return ERR_PTR(rc);
859 	}
860 
861 	/*
862 	 * The early init code does not consider partially populated
863 	 * initial sections, it simply assumes that memory will never be
864 	 * referenced.  If we hot-add memory into such a section then we
865 	 * do not need to populate the memmap and can simply reuse what
866 	 * is already there.
867 	 */
868 	if (nr_pages < PAGES_PER_SECTION && early_section(ms))
869 		return pfn_to_page(pfn);
870 
871 	memmap = populate_section_memmap(pfn, nr_pages, nid, altmap);
872 	if (!memmap) {
873 		section_deactivate(pfn, nr_pages, altmap);
874 		return ERR_PTR(-ENOMEM);
875 	}
876 
877 	return memmap;
878 }
879 
880 /**
881  * sparse_add_section - add a memory section, or populate an existing one
882  * @nid: The node to add section on
883  * @start_pfn: start pfn of the memory range
884  * @nr_pages: number of pfns to add in the section
885  * @altmap: device page map
886  *
887  * This is only intended for hotplug.
888  *
889  * Note that only VMEMMAP supports sub-section aligned hotplug,
890  * the proper alignment and size are gated by check_pfn_span().
891  *
892  *
893  * Return:
894  * * 0		- On success.
895  * * -EEXIST	- Section has been present.
896  * * -ENOMEM	- Out of memory.
897  */
898 int __meminit sparse_add_section(int nid, unsigned long start_pfn,
899 		unsigned long nr_pages, struct vmem_altmap *altmap)
900 {
901 	unsigned long section_nr = pfn_to_section_nr(start_pfn);
902 	struct mem_section *ms;
903 	struct page *memmap;
904 	int ret;
905 
906 	ret = sparse_index_init(section_nr, nid);
907 	if (ret < 0)
908 		return ret;
909 
910 	memmap = section_activate(nid, start_pfn, nr_pages, altmap);
911 	if (IS_ERR(memmap))
912 		return PTR_ERR(memmap);
913 
914 	/*
915 	 * Poison uninitialized struct pages in order to catch invalid flags
916 	 * combinations.
917 	 */
918 	page_init_poison(memmap, sizeof(struct page) * nr_pages);
919 
920 	ms = __nr_to_section(section_nr);
921 	set_section_nid(section_nr, nid);
922 	section_mark_present(ms);
923 
924 	/* Align memmap to section boundary in the subsection case */
925 	if (section_nr_to_pfn(section_nr) != start_pfn)
926 		memmap = pfn_to_page(section_nr_to_pfn(section_nr));
927 	sparse_init_one_section(ms, section_nr, memmap, ms->usage, 0);
928 
929 	return 0;
930 }
931 
932 #ifdef CONFIG_MEMORY_FAILURE
933 static void clear_hwpoisoned_pages(struct page *memmap, int nr_pages)
934 {
935 	int i;
936 
937 	/*
938 	 * A further optimization is to have per section refcounted
939 	 * num_poisoned_pages.  But that would need more space per memmap, so
940 	 * for now just do a quick global check to speed up this routine in the
941 	 * absence of bad pages.
942 	 */
943 	if (atomic_long_read(&num_poisoned_pages) == 0)
944 		return;
945 
946 	for (i = 0; i < nr_pages; i++) {
947 		if (PageHWPoison(&memmap[i])) {
948 			num_poisoned_pages_dec();
949 			ClearPageHWPoison(&memmap[i]);
950 		}
951 	}
952 }
953 #else
954 static inline void clear_hwpoisoned_pages(struct page *memmap, int nr_pages)
955 {
956 }
957 #endif
958 
959 void sparse_remove_section(struct mem_section *ms, unsigned long pfn,
960 		unsigned long nr_pages, unsigned long map_offset,
961 		struct vmem_altmap *altmap)
962 {
963 	clear_hwpoisoned_pages(pfn_to_page(pfn) + map_offset,
964 			nr_pages - map_offset);
965 	section_deactivate(pfn, nr_pages, altmap);
966 }
967 #endif /* CONFIG_MEMORY_HOTPLUG */
968