xref: /openbmc/linux/mm/sparse.c (revision 305c8388)
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 
21 /*
22  * Permanent SPARSEMEM data:
23  *
24  * 1) mem_section	- memory sections, mem_map's for valid memory
25  */
26 #ifdef CONFIG_SPARSEMEM_EXTREME
27 struct mem_section **mem_section;
28 #else
29 struct mem_section mem_section[NR_SECTION_ROOTS][SECTIONS_PER_ROOT]
30 	____cacheline_internodealigned_in_smp;
31 #endif
32 EXPORT_SYMBOL(mem_section);
33 
34 #ifdef NODE_NOT_IN_PAGE_FLAGS
35 /*
36  * If we did not store the node number in the page then we have to
37  * do a lookup in the section_to_node_table in order to find which
38  * node the page belongs to.
39  */
40 #if MAX_NUMNODES <= 256
41 static u8 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned;
42 #else
43 static u16 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned;
44 #endif
45 
46 int page_to_nid(const struct page *page)
47 {
48 	return section_to_node_table[page_to_section(page)];
49 }
50 EXPORT_SYMBOL(page_to_nid);
51 
52 static void set_section_nid(unsigned long section_nr, int nid)
53 {
54 	section_to_node_table[section_nr] = nid;
55 }
56 #else /* !NODE_NOT_IN_PAGE_FLAGS */
57 static inline void set_section_nid(unsigned long section_nr, int nid)
58 {
59 }
60 #endif
61 
62 #ifdef CONFIG_SPARSEMEM_EXTREME
63 static noinline struct mem_section __ref *sparse_index_alloc(int nid)
64 {
65 	struct mem_section *section = NULL;
66 	unsigned long array_size = SECTIONS_PER_ROOT *
67 				   sizeof(struct mem_section);
68 
69 	if (slab_is_available()) {
70 		section = kzalloc_node(array_size, GFP_KERNEL, nid);
71 	} else {
72 		section = memblock_alloc_node(array_size, SMP_CACHE_BYTES,
73 					      nid);
74 		if (!section)
75 			panic("%s: Failed to allocate %lu bytes nid=%d\n",
76 			      __func__, array_size, nid);
77 	}
78 
79 	return section;
80 }
81 
82 static int __meminit sparse_index_init(unsigned long section_nr, int nid)
83 {
84 	unsigned long root = SECTION_NR_TO_ROOT(section_nr);
85 	struct mem_section *section;
86 
87 	/*
88 	 * An existing section is possible in the sub-section hotplug
89 	 * case. First hot-add instantiates, follow-on hot-add reuses
90 	 * the existing section.
91 	 *
92 	 * The mem_hotplug_lock resolves the apparent race below.
93 	 */
94 	if (mem_section[root])
95 		return 0;
96 
97 	section = sparse_index_alloc(nid);
98 	if (!section)
99 		return -ENOMEM;
100 
101 	mem_section[root] = section;
102 
103 	return 0;
104 }
105 #else /* !SPARSEMEM_EXTREME */
106 static inline int sparse_index_init(unsigned long section_nr, int nid)
107 {
108 	return 0;
109 }
110 #endif
111 
112 #ifdef CONFIG_SPARSEMEM_EXTREME
113 unsigned long __section_nr(struct mem_section *ms)
114 {
115 	unsigned long root_nr;
116 	struct mem_section *root = NULL;
117 
118 	for (root_nr = 0; root_nr < NR_SECTION_ROOTS; root_nr++) {
119 		root = __nr_to_section(root_nr * SECTIONS_PER_ROOT);
120 		if (!root)
121 			continue;
122 
123 		if ((ms >= root) && (ms < (root + SECTIONS_PER_ROOT)))
124 		     break;
125 	}
126 
127 	VM_BUG_ON(!root);
128 
129 	return (root_nr * SECTIONS_PER_ROOT) + (ms - root);
130 }
131 #else
132 unsigned long __section_nr(struct mem_section *ms)
133 {
134 	return (unsigned long)(ms - mem_section[0]);
135 }
136 #endif
137 
138 /*
139  * During early boot, before section_mem_map is used for an actual
140  * mem_map, we use section_mem_map to store the section's NUMA
141  * node.  This keeps us from having to use another data structure.  The
142  * node information is cleared just before we store the real mem_map.
143  */
144 static inline unsigned long sparse_encode_early_nid(int nid)
145 {
146 	return (nid << SECTION_NID_SHIFT);
147 }
148 
149 static inline int sparse_early_nid(struct mem_section *section)
150 {
151 	return (section->section_mem_map >> SECTION_NID_SHIFT);
152 }
153 
154 /* Validate the physical addressing limitations of the model */
155 void __meminit mminit_validate_memmodel_limits(unsigned long *start_pfn,
156 						unsigned long *end_pfn)
157 {
158 	unsigned long max_sparsemem_pfn = 1UL << (MAX_PHYSMEM_BITS-PAGE_SHIFT);
159 
160 	/*
161 	 * Sanity checks - do not allow an architecture to pass
162 	 * in larger pfns than the maximum scope of sparsemem:
163 	 */
164 	if (*start_pfn > max_sparsemem_pfn) {
165 		mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
166 			"Start of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
167 			*start_pfn, *end_pfn, max_sparsemem_pfn);
168 		WARN_ON_ONCE(1);
169 		*start_pfn = max_sparsemem_pfn;
170 		*end_pfn = max_sparsemem_pfn;
171 	} else if (*end_pfn > max_sparsemem_pfn) {
172 		mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
173 			"End of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
174 			*start_pfn, *end_pfn, max_sparsemem_pfn);
175 		WARN_ON_ONCE(1);
176 		*end_pfn = max_sparsemem_pfn;
177 	}
178 }
179 
180 /*
181  * There are a number of times that we loop over NR_MEM_SECTIONS,
182  * looking for section_present() on each.  But, when we have very
183  * large physical address spaces, NR_MEM_SECTIONS can also be
184  * very large which makes the loops quite long.
185  *
186  * Keeping track of this gives us an easy way to break out of
187  * those loops early.
188  */
189 unsigned long __highest_present_section_nr;
190 static void section_mark_present(struct mem_section *ms)
191 {
192 	unsigned long section_nr = __section_nr(ms);
193 
194 	if (section_nr > __highest_present_section_nr)
195 		__highest_present_section_nr = section_nr;
196 
197 	ms->section_mem_map |= SECTION_MARKED_PRESENT;
198 }
199 
200 #define for_each_present_section_nr(start, section_nr)		\
201 	for (section_nr = next_present_section_nr(start-1);	\
202 	     ((section_nr != -1) &&				\
203 	      (section_nr <= __highest_present_section_nr));	\
204 	     section_nr = next_present_section_nr(section_nr))
205 
206 static inline unsigned long first_present_section_nr(void)
207 {
208 	return next_present_section_nr(-1);
209 }
210 
211 #ifdef CONFIG_SPARSEMEM_VMEMMAP
212 static void subsection_mask_set(unsigned long *map, unsigned long pfn,
213 		unsigned long nr_pages)
214 {
215 	int idx = subsection_map_index(pfn);
216 	int end = subsection_map_index(pfn + nr_pages - 1);
217 
218 	bitmap_set(map, idx, end - idx + 1);
219 }
220 
221 void __init subsection_map_init(unsigned long pfn, unsigned long nr_pages)
222 {
223 	int end_sec = pfn_to_section_nr(pfn + nr_pages - 1);
224 	unsigned long nr, start_sec = pfn_to_section_nr(pfn);
225 
226 	if (!nr_pages)
227 		return;
228 
229 	for (nr = start_sec; nr <= end_sec; nr++) {
230 		struct mem_section *ms;
231 		unsigned long pfns;
232 
233 		pfns = min(nr_pages, PAGES_PER_SECTION
234 				- (pfn & ~PAGE_SECTION_MASK));
235 		ms = __nr_to_section(nr);
236 		subsection_mask_set(ms->usage->subsection_map, pfn, pfns);
237 
238 		pr_debug("%s: sec: %lu pfns: %lu set(%d, %d)\n", __func__, nr,
239 				pfns, subsection_map_index(pfn),
240 				subsection_map_index(pfn + pfns - 1));
241 
242 		pfn += pfns;
243 		nr_pages -= pfns;
244 	}
245 }
246 #else
247 void __init subsection_map_init(unsigned long pfn, unsigned long nr_pages)
248 {
249 }
250 #endif
251 
252 /* Record a memory area against a node. */
253 void __init memory_present(int nid, unsigned long start, unsigned long end)
254 {
255 	unsigned long pfn;
256 
257 #ifdef CONFIG_SPARSEMEM_EXTREME
258 	if (unlikely(!mem_section)) {
259 		unsigned long size, align;
260 
261 		size = sizeof(struct mem_section*) * NR_SECTION_ROOTS;
262 		align = 1 << (INTERNODE_CACHE_SHIFT);
263 		mem_section = memblock_alloc(size, align);
264 		if (!mem_section)
265 			panic("%s: Failed to allocate %lu bytes align=0x%lx\n",
266 			      __func__, size, align);
267 	}
268 #endif
269 
270 	start &= PAGE_SECTION_MASK;
271 	mminit_validate_memmodel_limits(&start, &end);
272 	for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) {
273 		unsigned long section = pfn_to_section_nr(pfn);
274 		struct mem_section *ms;
275 
276 		sparse_index_init(section, nid);
277 		set_section_nid(section, nid);
278 
279 		ms = __nr_to_section(section);
280 		if (!ms->section_mem_map) {
281 			ms->section_mem_map = sparse_encode_early_nid(nid) |
282 							SECTION_IS_ONLINE;
283 			section_mark_present(ms);
284 		}
285 	}
286 }
287 
288 /*
289  * Mark all memblocks as present using memory_present(). This is a
290  * convenience function that is useful for a number of arches
291  * to mark all of the systems memory as present during initialization.
292  */
293 void __init memblocks_present(void)
294 {
295 	struct memblock_region *reg;
296 
297 	for_each_memblock(memory, reg) {
298 		memory_present(memblock_get_region_node(reg),
299 			       memblock_region_memory_base_pfn(reg),
300 			       memblock_region_memory_end_pfn(reg));
301 	}
302 }
303 
304 /*
305  * Subtle, we encode the real pfn into the mem_map such that
306  * the identity pfn - section_mem_map will return the actual
307  * physical page frame number.
308  */
309 static unsigned long sparse_encode_mem_map(struct page *mem_map, unsigned long pnum)
310 {
311 	unsigned long coded_mem_map =
312 		(unsigned long)(mem_map - (section_nr_to_pfn(pnum)));
313 	BUILD_BUG_ON(SECTION_MAP_LAST_BIT > (1UL<<PFN_SECTION_SHIFT));
314 	BUG_ON(coded_mem_map & ~SECTION_MAP_MASK);
315 	return coded_mem_map;
316 }
317 
318 /*
319  * Decode mem_map from the coded memmap
320  */
321 struct page *sparse_decode_mem_map(unsigned long coded_mem_map, unsigned long pnum)
322 {
323 	/* mask off the extra low bits of information */
324 	coded_mem_map &= SECTION_MAP_MASK;
325 	return ((struct page *)coded_mem_map) + section_nr_to_pfn(pnum);
326 }
327 
328 static void __meminit sparse_init_one_section(struct mem_section *ms,
329 		unsigned long pnum, struct page *mem_map,
330 		struct mem_section_usage *usage, unsigned long flags)
331 {
332 	ms->section_mem_map &= ~SECTION_MAP_MASK;
333 	ms->section_mem_map |= sparse_encode_mem_map(mem_map, pnum)
334 		| SECTION_HAS_MEM_MAP | flags;
335 	ms->usage = usage;
336 }
337 
338 static unsigned long usemap_size(void)
339 {
340 	return BITS_TO_LONGS(SECTION_BLOCKFLAGS_BITS) * sizeof(unsigned long);
341 }
342 
343 size_t mem_section_usage_size(void)
344 {
345 	return sizeof(struct mem_section_usage) + usemap_size();
346 }
347 
348 #ifdef CONFIG_MEMORY_HOTREMOVE
349 static struct mem_section_usage * __init
350 sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
351 					 unsigned long size)
352 {
353 	struct mem_section_usage *usage;
354 	unsigned long goal, limit;
355 	int nid;
356 	/*
357 	 * A page may contain usemaps for other sections preventing the
358 	 * page being freed and making a section unremovable while
359 	 * other sections referencing the usemap remain active. Similarly,
360 	 * a pgdat can prevent a section being removed. If section A
361 	 * contains a pgdat and section B contains the usemap, both
362 	 * sections become inter-dependent. This allocates usemaps
363 	 * from the same section as the pgdat where possible to avoid
364 	 * this problem.
365 	 */
366 	goal = __pa(pgdat) & (PAGE_SECTION_MASK << PAGE_SHIFT);
367 	limit = goal + (1UL << PA_SECTION_SHIFT);
368 	nid = early_pfn_to_nid(goal >> PAGE_SHIFT);
369 again:
370 	usage = memblock_alloc_try_nid(size, SMP_CACHE_BYTES, goal, limit, nid);
371 	if (!usage && limit) {
372 		limit = 0;
373 		goto again;
374 	}
375 	return usage;
376 }
377 
378 static void __init check_usemap_section_nr(int nid,
379 		struct mem_section_usage *usage)
380 {
381 	unsigned long usemap_snr, pgdat_snr;
382 	static unsigned long old_usemap_snr;
383 	static unsigned long old_pgdat_snr;
384 	struct pglist_data *pgdat = NODE_DATA(nid);
385 	int usemap_nid;
386 
387 	/* First call */
388 	if (!old_usemap_snr) {
389 		old_usemap_snr = NR_MEM_SECTIONS;
390 		old_pgdat_snr = NR_MEM_SECTIONS;
391 	}
392 
393 	usemap_snr = pfn_to_section_nr(__pa(usage) >> PAGE_SHIFT);
394 	pgdat_snr = pfn_to_section_nr(__pa(pgdat) >> PAGE_SHIFT);
395 	if (usemap_snr == pgdat_snr)
396 		return;
397 
398 	if (old_usemap_snr == usemap_snr && old_pgdat_snr == pgdat_snr)
399 		/* skip redundant message */
400 		return;
401 
402 	old_usemap_snr = usemap_snr;
403 	old_pgdat_snr = pgdat_snr;
404 
405 	usemap_nid = sparse_early_nid(__nr_to_section(usemap_snr));
406 	if (usemap_nid != nid) {
407 		pr_info("node %d must be removed before remove section %ld\n",
408 			nid, usemap_snr);
409 		return;
410 	}
411 	/*
412 	 * There is a circular dependency.
413 	 * Some platforms allow un-removable section because they will just
414 	 * gather other removable sections for dynamic partitioning.
415 	 * Just notify un-removable section's number here.
416 	 */
417 	pr_info("Section %ld and %ld (node %d) have a circular dependency on usemap and pgdat allocations\n",
418 		usemap_snr, pgdat_snr, nid);
419 }
420 #else
421 static struct mem_section_usage * __init
422 sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
423 					 unsigned long size)
424 {
425 	return memblock_alloc_node(size, SMP_CACHE_BYTES, pgdat->node_id);
426 }
427 
428 static void __init check_usemap_section_nr(int nid,
429 		struct mem_section_usage *usage)
430 {
431 }
432 #endif /* CONFIG_MEMORY_HOTREMOVE */
433 
434 #ifdef CONFIG_SPARSEMEM_VMEMMAP
435 static unsigned long __init section_map_size(void)
436 {
437 	return ALIGN(sizeof(struct page) * PAGES_PER_SECTION, PMD_SIZE);
438 }
439 
440 #else
441 static unsigned long __init section_map_size(void)
442 {
443 	return PAGE_ALIGN(sizeof(struct page) * PAGES_PER_SECTION);
444 }
445 
446 struct page __init *__populate_section_memmap(unsigned long pfn,
447 		unsigned long nr_pages, int nid, struct vmem_altmap *altmap)
448 {
449 	unsigned long size = section_map_size();
450 	struct page *map = sparse_buffer_alloc(size);
451 	phys_addr_t addr = __pa(MAX_DMA_ADDRESS);
452 
453 	if (map)
454 		return map;
455 
456 	map = memblock_alloc_try_nid_raw(size, size, addr,
457 					  MEMBLOCK_ALLOC_ACCESSIBLE, nid);
458 	if (!map)
459 		panic("%s: Failed to allocate %lu bytes align=0x%lx nid=%d from=%pa\n",
460 		      __func__, size, PAGE_SIZE, nid, &addr);
461 
462 	return map;
463 }
464 #endif /* !CONFIG_SPARSEMEM_VMEMMAP */
465 
466 static void *sparsemap_buf __meminitdata;
467 static void *sparsemap_buf_end __meminitdata;
468 
469 static inline void __meminit sparse_buffer_free(unsigned long size)
470 {
471 	WARN_ON(!sparsemap_buf || size == 0);
472 	memblock_free_early(__pa(sparsemap_buf), size);
473 }
474 
475 static void __init sparse_buffer_init(unsigned long size, int nid)
476 {
477 	phys_addr_t addr = __pa(MAX_DMA_ADDRESS);
478 	WARN_ON(sparsemap_buf);	/* forgot to call sparse_buffer_fini()? */
479 	/*
480 	 * Pre-allocated buffer is mainly used by __populate_section_memmap
481 	 * and we want it to be properly aligned to the section size - this is
482 	 * especially the case for VMEMMAP which maps memmap to PMDs
483 	 */
484 	sparsemap_buf = memblock_alloc_exact_nid_raw(size, section_map_size(),
485 					addr, MEMBLOCK_ALLOC_ACCESSIBLE, nid);
486 	sparsemap_buf_end = sparsemap_buf + size;
487 }
488 
489 static void __init sparse_buffer_fini(void)
490 {
491 	unsigned long size = sparsemap_buf_end - sparsemap_buf;
492 
493 	if (sparsemap_buf && size > 0)
494 		sparse_buffer_free(size);
495 	sparsemap_buf = NULL;
496 }
497 
498 void * __meminit sparse_buffer_alloc(unsigned long size)
499 {
500 	void *ptr = NULL;
501 
502 	if (sparsemap_buf) {
503 		ptr = (void *) roundup((unsigned long)sparsemap_buf, size);
504 		if (ptr + size > sparsemap_buf_end)
505 			ptr = NULL;
506 		else {
507 			/* Free redundant aligned space */
508 			if ((unsigned long)(ptr - sparsemap_buf) > 0)
509 				sparse_buffer_free((unsigned long)(ptr - sparsemap_buf));
510 			sparsemap_buf = ptr + size;
511 		}
512 	}
513 	return ptr;
514 }
515 
516 void __weak __meminit vmemmap_populate_print_last(void)
517 {
518 }
519 
520 /*
521  * Initialize sparse on a specific node. The node spans [pnum_begin, pnum_end)
522  * And number of present sections in this node is map_count.
523  */
524 static void __init sparse_init_nid(int nid, unsigned long pnum_begin,
525 				   unsigned long pnum_end,
526 				   unsigned long map_count)
527 {
528 	struct mem_section_usage *usage;
529 	unsigned long pnum;
530 	struct page *map;
531 
532 	usage = sparse_early_usemaps_alloc_pgdat_section(NODE_DATA(nid),
533 			mem_section_usage_size() * map_count);
534 	if (!usage) {
535 		pr_err("%s: node[%d] usemap allocation failed", __func__, nid);
536 		goto failed;
537 	}
538 	sparse_buffer_init(map_count * section_map_size(), nid);
539 	for_each_present_section_nr(pnum_begin, pnum) {
540 		unsigned long pfn = section_nr_to_pfn(pnum);
541 
542 		if (pnum >= pnum_end)
543 			break;
544 
545 		map = __populate_section_memmap(pfn, PAGES_PER_SECTION,
546 				nid, NULL);
547 		if (!map) {
548 			pr_err("%s: node[%d] memory map backing failed. Some memory will not be available.",
549 			       __func__, nid);
550 			pnum_begin = pnum;
551 			goto failed;
552 		}
553 		check_usemap_section_nr(nid, usage);
554 		sparse_init_one_section(__nr_to_section(pnum), pnum, map, usage,
555 				SECTION_IS_EARLY);
556 		usage = (void *) usage + mem_section_usage_size();
557 	}
558 	sparse_buffer_fini();
559 	return;
560 failed:
561 	/* We failed to allocate, mark all the following pnums as not present */
562 	for_each_present_section_nr(pnum_begin, pnum) {
563 		struct mem_section *ms;
564 
565 		if (pnum >= pnum_end)
566 			break;
567 		ms = __nr_to_section(pnum);
568 		ms->section_mem_map = 0;
569 	}
570 }
571 
572 /*
573  * Allocate the accumulated non-linear sections, allocate a mem_map
574  * for each and record the physical to section mapping.
575  */
576 void __init sparse_init(void)
577 {
578 	unsigned long pnum_begin = first_present_section_nr();
579 	int nid_begin = sparse_early_nid(__nr_to_section(pnum_begin));
580 	unsigned long pnum_end, map_count = 1;
581 
582 	/* Setup pageblock_order for HUGETLB_PAGE_SIZE_VARIABLE */
583 	set_pageblock_order();
584 
585 	for_each_present_section_nr(pnum_begin + 1, pnum_end) {
586 		int nid = sparse_early_nid(__nr_to_section(pnum_end));
587 
588 		if (nid == nid_begin) {
589 			map_count++;
590 			continue;
591 		}
592 		/* Init node with sections in range [pnum_begin, pnum_end) */
593 		sparse_init_nid(nid_begin, pnum_begin, pnum_end, map_count);
594 		nid_begin = nid;
595 		pnum_begin = pnum_end;
596 		map_count = 1;
597 	}
598 	/* cover the last node */
599 	sparse_init_nid(nid_begin, pnum_begin, pnum_end, map_count);
600 	vmemmap_populate_print_last();
601 }
602 
603 #ifdef CONFIG_MEMORY_HOTPLUG
604 
605 /* Mark all memory sections within the pfn range as online */
606 void online_mem_sections(unsigned long start_pfn, unsigned long end_pfn)
607 {
608 	unsigned long pfn;
609 
610 	for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
611 		unsigned long section_nr = pfn_to_section_nr(pfn);
612 		struct mem_section *ms;
613 
614 		/* onlining code should never touch invalid ranges */
615 		if (WARN_ON(!valid_section_nr(section_nr)))
616 			continue;
617 
618 		ms = __nr_to_section(section_nr);
619 		ms->section_mem_map |= SECTION_IS_ONLINE;
620 	}
621 }
622 
623 #ifdef CONFIG_MEMORY_HOTREMOVE
624 /* Mark all memory sections within the pfn range as offline */
625 void offline_mem_sections(unsigned long start_pfn, unsigned long end_pfn)
626 {
627 	unsigned long pfn;
628 
629 	for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
630 		unsigned long section_nr = pfn_to_section_nr(pfn);
631 		struct mem_section *ms;
632 
633 		/*
634 		 * TODO this needs some double checking. Offlining code makes
635 		 * sure to check pfn_valid but those checks might be just bogus
636 		 */
637 		if (WARN_ON(!valid_section_nr(section_nr)))
638 			continue;
639 
640 		ms = __nr_to_section(section_nr);
641 		ms->section_mem_map &= ~SECTION_IS_ONLINE;
642 	}
643 }
644 #endif
645 
646 #ifdef CONFIG_SPARSEMEM_VMEMMAP
647 static struct page * __meminit populate_section_memmap(unsigned long pfn,
648 		unsigned long nr_pages, int nid, struct vmem_altmap *altmap)
649 {
650 	return __populate_section_memmap(pfn, nr_pages, nid, altmap);
651 }
652 
653 static void depopulate_section_memmap(unsigned long pfn, unsigned long nr_pages,
654 		struct vmem_altmap *altmap)
655 {
656 	unsigned long start = (unsigned long) pfn_to_page(pfn);
657 	unsigned long end = start + nr_pages * sizeof(struct page);
658 
659 	vmemmap_free(start, end, altmap);
660 }
661 static void free_map_bootmem(struct page *memmap)
662 {
663 	unsigned long start = (unsigned long)memmap;
664 	unsigned long end = (unsigned long)(memmap + PAGES_PER_SECTION);
665 
666 	vmemmap_free(start, end, NULL);
667 }
668 
669 static int clear_subsection_map(unsigned long pfn, unsigned long nr_pages)
670 {
671 	DECLARE_BITMAP(map, SUBSECTIONS_PER_SECTION) = { 0 };
672 	DECLARE_BITMAP(tmp, SUBSECTIONS_PER_SECTION) = { 0 };
673 	struct mem_section *ms = __pfn_to_section(pfn);
674 	unsigned long *subsection_map = ms->usage
675 		? &ms->usage->subsection_map[0] : NULL;
676 
677 	subsection_mask_set(map, pfn, nr_pages);
678 	if (subsection_map)
679 		bitmap_and(tmp, map, subsection_map, SUBSECTIONS_PER_SECTION);
680 
681 	if (WARN(!subsection_map || !bitmap_equal(tmp, map, SUBSECTIONS_PER_SECTION),
682 				"section already deactivated (%#lx + %ld)\n",
683 				pfn, nr_pages))
684 		return -EINVAL;
685 
686 	bitmap_xor(subsection_map, map, subsection_map, SUBSECTIONS_PER_SECTION);
687 	return 0;
688 }
689 
690 static bool is_subsection_map_empty(struct mem_section *ms)
691 {
692 	return bitmap_empty(&ms->usage->subsection_map[0],
693 			    SUBSECTIONS_PER_SECTION);
694 }
695 
696 static int fill_subsection_map(unsigned long pfn, unsigned long nr_pages)
697 {
698 	struct mem_section *ms = __pfn_to_section(pfn);
699 	DECLARE_BITMAP(map, SUBSECTIONS_PER_SECTION) = { 0 };
700 	unsigned long *subsection_map;
701 	int rc = 0;
702 
703 	subsection_mask_set(map, pfn, nr_pages);
704 
705 	subsection_map = &ms->usage->subsection_map[0];
706 
707 	if (bitmap_empty(map, SUBSECTIONS_PER_SECTION))
708 		rc = -EINVAL;
709 	else if (bitmap_intersects(map, subsection_map, SUBSECTIONS_PER_SECTION))
710 		rc = -EEXIST;
711 	else
712 		bitmap_or(subsection_map, map, subsection_map,
713 				SUBSECTIONS_PER_SECTION);
714 
715 	return rc;
716 }
717 #else
718 struct page * __meminit populate_section_memmap(unsigned long pfn,
719 		unsigned long nr_pages, int nid, struct vmem_altmap *altmap)
720 {
721 	return kvmalloc_node(array_size(sizeof(struct page),
722 					PAGES_PER_SECTION), GFP_KERNEL, nid);
723 }
724 
725 static void depopulate_section_memmap(unsigned long pfn, unsigned long nr_pages,
726 		struct vmem_altmap *altmap)
727 {
728 	kvfree(pfn_to_page(pfn));
729 }
730 
731 static void free_map_bootmem(struct page *memmap)
732 {
733 	unsigned long maps_section_nr, removing_section_nr, i;
734 	unsigned long magic, nr_pages;
735 	struct page *page = virt_to_page(memmap);
736 
737 	nr_pages = PAGE_ALIGN(PAGES_PER_SECTION * sizeof(struct page))
738 		>> PAGE_SHIFT;
739 
740 	for (i = 0; i < nr_pages; i++, page++) {
741 		magic = (unsigned long) page->freelist;
742 
743 		BUG_ON(magic == NODE_INFO);
744 
745 		maps_section_nr = pfn_to_section_nr(page_to_pfn(page));
746 		removing_section_nr = page_private(page);
747 
748 		/*
749 		 * When this function is called, the removing section is
750 		 * logical offlined state. This means all pages are isolated
751 		 * from page allocator. If removing section's memmap is placed
752 		 * on the same section, it must not be freed.
753 		 * If it is freed, page allocator may allocate it which will
754 		 * be removed physically soon.
755 		 */
756 		if (maps_section_nr != removing_section_nr)
757 			put_page_bootmem(page);
758 	}
759 }
760 
761 static int clear_subsection_map(unsigned long pfn, unsigned long nr_pages)
762 {
763 	return 0;
764 }
765 
766 static bool is_subsection_map_empty(struct mem_section *ms)
767 {
768 	return true;
769 }
770 
771 static int fill_subsection_map(unsigned long pfn, unsigned long nr_pages)
772 {
773 	return 0;
774 }
775 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
776 
777 /*
778  * To deactivate a memory region, there are 3 cases to handle across
779  * two configurations (SPARSEMEM_VMEMMAP={y,n}):
780  *
781  * 1. deactivation of a partial hot-added section (only possible in
782  *    the SPARSEMEM_VMEMMAP=y case).
783  *      a) section was present at memory init.
784  *      b) section was hot-added post memory init.
785  * 2. deactivation of a complete hot-added section.
786  * 3. deactivation of a complete section from memory init.
787  *
788  * For 1, when subsection_map does not empty we will not be freeing the
789  * usage map, but still need to free the vmemmap range.
790  *
791  * For 2 and 3, the SPARSEMEM_VMEMMAP={y,n} cases are unified
792  */
793 static void section_deactivate(unsigned long pfn, unsigned long nr_pages,
794 		struct vmem_altmap *altmap)
795 {
796 	struct mem_section *ms = __pfn_to_section(pfn);
797 	bool section_is_early = early_section(ms);
798 	struct page *memmap = NULL;
799 	bool empty;
800 
801 	if (clear_subsection_map(pfn, nr_pages))
802 		return;
803 
804 	empty = is_subsection_map_empty(ms);
805 	if (empty) {
806 		unsigned long section_nr = pfn_to_section_nr(pfn);
807 
808 		/*
809 		 * When removing an early section, the usage map is kept (as the
810 		 * usage maps of other sections fall into the same page). It
811 		 * will be re-used when re-adding the section - which is then no
812 		 * longer an early section. If the usage map is PageReserved, it
813 		 * was allocated during boot.
814 		 */
815 		if (!PageReserved(virt_to_page(ms->usage))) {
816 			kfree(ms->usage);
817 			ms->usage = NULL;
818 		}
819 		memmap = sparse_decode_mem_map(ms->section_mem_map, section_nr);
820 		/*
821 		 * Mark the section invalid so that valid_section()
822 		 * return false. This prevents code from dereferencing
823 		 * ms->usage array.
824 		 */
825 		ms->section_mem_map &= ~SECTION_HAS_MEM_MAP;
826 	}
827 
828 	if (section_is_early && memmap)
829 		free_map_bootmem(memmap);
830 	else
831 		depopulate_section_memmap(pfn, nr_pages, altmap);
832 
833 	if (empty)
834 		ms->section_mem_map = (unsigned long)NULL;
835 }
836 
837 static struct page * __meminit section_activate(int nid, unsigned long pfn,
838 		unsigned long nr_pages, struct vmem_altmap *altmap)
839 {
840 	struct mem_section *ms = __pfn_to_section(pfn);
841 	struct mem_section_usage *usage = NULL;
842 	struct page *memmap;
843 	int rc = 0;
844 
845 	if (!ms->usage) {
846 		usage = kzalloc(mem_section_usage_size(), GFP_KERNEL);
847 		if (!usage)
848 			return ERR_PTR(-ENOMEM);
849 		ms->usage = usage;
850 	}
851 
852 	rc = fill_subsection_map(pfn, nr_pages);
853 	if (rc) {
854 		if (usage)
855 			ms->usage = NULL;
856 		kfree(usage);
857 		return ERR_PTR(rc);
858 	}
859 
860 	/*
861 	 * The early init code does not consider partially populated
862 	 * initial sections, it simply assumes that memory will never be
863 	 * referenced.  If we hot-add memory into such a section then we
864 	 * do not need to populate the memmap and can simply reuse what
865 	 * is already there.
866 	 */
867 	if (nr_pages < PAGES_PER_SECTION && early_section(ms))
868 		return pfn_to_page(pfn);
869 
870 	memmap = populate_section_memmap(pfn, nr_pages, nid, altmap);
871 	if (!memmap) {
872 		section_deactivate(pfn, nr_pages, altmap);
873 		return ERR_PTR(-ENOMEM);
874 	}
875 
876 	return memmap;
877 }
878 
879 /**
880  * sparse_add_section - add a memory section, or populate an existing one
881  * @nid: The node to add section on
882  * @start_pfn: start pfn of the memory range
883  * @nr_pages: number of pfns to add in the section
884  * @altmap: device page map
885  *
886  * This is only intended for hotplug.
887  *
888  * Note that only VMEMMAP supports sub-section aligned hotplug,
889  * the proper alignment and size are gated by check_pfn_span().
890  *
891  *
892  * Return:
893  * * 0		- On success.
894  * * -EEXIST	- Section has been present.
895  * * -ENOMEM	- Out of memory.
896  */
897 int __meminit sparse_add_section(int nid, unsigned long start_pfn,
898 		unsigned long nr_pages, struct vmem_altmap *altmap)
899 {
900 	unsigned long section_nr = pfn_to_section_nr(start_pfn);
901 	struct mem_section *ms;
902 	struct page *memmap;
903 	int ret;
904 
905 	ret = sparse_index_init(section_nr, nid);
906 	if (ret < 0)
907 		return ret;
908 
909 	memmap = section_activate(nid, start_pfn, nr_pages, altmap);
910 	if (IS_ERR(memmap))
911 		return PTR_ERR(memmap);
912 
913 	/*
914 	 * Poison uninitialized struct pages in order to catch invalid flags
915 	 * combinations.
916 	 */
917 	page_init_poison(memmap, sizeof(struct page) * nr_pages);
918 
919 	ms = __nr_to_section(section_nr);
920 	set_section_nid(section_nr, nid);
921 	section_mark_present(ms);
922 
923 	/* Align memmap to section boundary in the subsection case */
924 	if (section_nr_to_pfn(section_nr) != start_pfn)
925 		memmap = pfn_to_page(section_nr_to_pfn(section_nr));
926 	sparse_init_one_section(ms, section_nr, memmap, ms->usage, 0);
927 
928 	return 0;
929 }
930 
931 #ifdef CONFIG_MEMORY_FAILURE
932 static void clear_hwpoisoned_pages(struct page *memmap, int nr_pages)
933 {
934 	int i;
935 
936 	/*
937 	 * A further optimization is to have per section refcounted
938 	 * num_poisoned_pages.  But that would need more space per memmap, so
939 	 * for now just do a quick global check to speed up this routine in the
940 	 * absence of bad pages.
941 	 */
942 	if (atomic_long_read(&num_poisoned_pages) == 0)
943 		return;
944 
945 	for (i = 0; i < nr_pages; i++) {
946 		if (PageHWPoison(&memmap[i])) {
947 			num_poisoned_pages_dec();
948 			ClearPageHWPoison(&memmap[i]);
949 		}
950 	}
951 }
952 #else
953 static inline void clear_hwpoisoned_pages(struct page *memmap, int nr_pages)
954 {
955 }
956 #endif
957 
958 void sparse_remove_section(struct mem_section *ms, unsigned long pfn,
959 		unsigned long nr_pages, unsigned long map_offset,
960 		struct vmem_altmap *altmap)
961 {
962 	clear_hwpoisoned_pages(pfn_to_page(pfn) + map_offset,
963 			nr_pages - map_offset);
964 	section_deactivate(pfn, nr_pages, altmap);
965 }
966 #endif /* CONFIG_MEMORY_HOTPLUG */
967