Linux Audio

Check our new training course

Loading...
v5.14.15
  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#include <linux/bootmem_info.h>
 17
 18#include "internal.h"
 19#include <asm/dma.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
 27struct mem_section **mem_section;
 28#else
 29struct mem_section mem_section[NR_SECTION_ROOTS][SECTIONS_PER_ROOT]
 30	____cacheline_internodealigned_in_smp;
 31#endif
 32EXPORT_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
 41static u8 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned;
 42#else
 43static u16 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned;
 44#endif
 45
 46int page_to_nid(const struct page *page)
 47{
 48	return section_to_node_table[page_to_section(page)];
 49}
 50EXPORT_SYMBOL(page_to_nid);
 51
 52static 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 */
 57static inline void set_section_nid(unsigned long section_nr, int nid)
 58{
 59}
 60#endif
 61
 62#ifdef CONFIG_SPARSEMEM_EXTREME
 63static 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
 82static 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 */
106static inline int sparse_index_init(unsigned long section_nr, int nid)
107{
108	return 0;
109}
110#endif
111
112#ifdef CONFIG_SPARSEMEM_EXTREME
113unsigned 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
132unsigned 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 */
144static inline unsigned long sparse_encode_early_nid(int nid)
145{
146	return (nid << SECTION_NID_SHIFT);
147}
148
149static 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 */
155void __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 */
189unsigned long __highest_present_section_nr;
190static 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
206static inline unsigned long first_present_section_nr(void)
207{
208	return next_present_section_nr(-1);
209}
210
211#ifdef CONFIG_SPARSEMEM_VMEMMAP
212static 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
221void __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
247void __init subsection_map_init(unsigned long pfn, unsigned long nr_pages)
248{
249}
250#endif
251
252/* Record a memory area against a node. */
253static 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().
290 * This is a convenience function that is useful to mark all of the systems
291 * memory as present during initialization.
292 */
293static void __init memblocks_present(void)
294{
295	unsigned long start, end;
296	int i, nid;
297
 
 
 
 
 
 
 
 
 
 
 
 
 
298	for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid)
299		memory_present(nid, start, end);
300}
301
302/*
303 * Subtle, we encode the real pfn into the mem_map such that
304 * the identity pfn - section_mem_map will return the actual
305 * physical page frame number.
306 */
307static unsigned long sparse_encode_mem_map(struct page *mem_map, unsigned long pnum)
308{
309	unsigned long coded_mem_map =
310		(unsigned long)(mem_map - (section_nr_to_pfn(pnum)));
311	BUILD_BUG_ON(SECTION_MAP_LAST_BIT > (1UL<<PFN_SECTION_SHIFT));
312	BUG_ON(coded_mem_map & ~SECTION_MAP_MASK);
313	return coded_mem_map;
314}
315
316#ifdef CONFIG_MEMORY_HOTPLUG
317/*
318 * Decode mem_map from the coded memmap
319 */
320struct page *sparse_decode_mem_map(unsigned long coded_mem_map, unsigned long pnum)
321{
322	/* mask off the extra low bits of information */
323	coded_mem_map &= SECTION_MAP_MASK;
324	return ((struct page *)coded_mem_map) + section_nr_to_pfn(pnum);
325}
326#endif /* CONFIG_MEMORY_HOTPLUG */
327
328static 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
338static unsigned long usemap_size(void)
339{
340	return BITS_TO_LONGS(SECTION_BLOCKFLAGS_BITS) * sizeof(unsigned long);
341}
342
343size_t mem_section_usage_size(void)
344{
345	return sizeof(struct mem_section_usage) + usemap_size();
346}
347
 
348static inline phys_addr_t pgdat_to_phys(struct pglist_data *pgdat)
349{
350#ifndef CONFIG_NUMA
351	return __pa_symbol(pgdat);
 
352#else
353	return __pa(pgdat);
354#endif
355}
356
357#ifdef CONFIG_MEMORY_HOTREMOVE
358static struct mem_section_usage * __init
359sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
360					 unsigned long size)
361{
362	struct mem_section_usage *usage;
363	unsigned long goal, limit;
364	int nid;
365	/*
366	 * A page may contain usemaps for other sections preventing the
367	 * page being freed and making a section unremovable while
368	 * other sections referencing the usemap remain active. Similarly,
369	 * a pgdat can prevent a section being removed. If section A
370	 * contains a pgdat and section B contains the usemap, both
371	 * sections become inter-dependent. This allocates usemaps
372	 * from the same section as the pgdat where possible to avoid
373	 * this problem.
374	 */
375	goal = pgdat_to_phys(pgdat) & (PAGE_SECTION_MASK << PAGE_SHIFT);
376	limit = goal + (1UL << PA_SECTION_SHIFT);
377	nid = early_pfn_to_nid(goal >> PAGE_SHIFT);
378again:
379	usage = memblock_alloc_try_nid(size, SMP_CACHE_BYTES, goal, limit, nid);
380	if (!usage && limit) {
381		limit = 0;
382		goto again;
383	}
384	return usage;
385}
386
387static void __init check_usemap_section_nr(int nid,
388		struct mem_section_usage *usage)
389{
390	unsigned long usemap_snr, pgdat_snr;
391	static unsigned long old_usemap_snr;
392	static unsigned long old_pgdat_snr;
393	struct pglist_data *pgdat = NODE_DATA(nid);
394	int usemap_nid;
395
396	/* First call */
397	if (!old_usemap_snr) {
398		old_usemap_snr = NR_MEM_SECTIONS;
399		old_pgdat_snr = NR_MEM_SECTIONS;
400	}
401
402	usemap_snr = pfn_to_section_nr(__pa(usage) >> PAGE_SHIFT);
403	pgdat_snr = pfn_to_section_nr(pgdat_to_phys(pgdat) >> PAGE_SHIFT);
404	if (usemap_snr == pgdat_snr)
405		return;
406
407	if (old_usemap_snr == usemap_snr && old_pgdat_snr == pgdat_snr)
408		/* skip redundant message */
409		return;
410
411	old_usemap_snr = usemap_snr;
412	old_pgdat_snr = pgdat_snr;
413
414	usemap_nid = sparse_early_nid(__nr_to_section(usemap_snr));
415	if (usemap_nid != nid) {
416		pr_info("node %d must be removed before remove section %ld\n",
417			nid, usemap_snr);
418		return;
419	}
420	/*
421	 * There is a circular dependency.
422	 * Some platforms allow un-removable section because they will just
423	 * gather other removable sections for dynamic partitioning.
424	 * Just notify un-removable section's number here.
425	 */
426	pr_info("Section %ld and %ld (node %d) have a circular dependency on usemap and pgdat allocations\n",
427		usemap_snr, pgdat_snr, nid);
428}
429#else
430static struct mem_section_usage * __init
431sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
432					 unsigned long size)
433{
434	return memblock_alloc_node(size, SMP_CACHE_BYTES, pgdat->node_id);
435}
436
437static void __init check_usemap_section_nr(int nid,
438		struct mem_section_usage *usage)
439{
440}
441#endif /* CONFIG_MEMORY_HOTREMOVE */
442
443#ifdef CONFIG_SPARSEMEM_VMEMMAP
444static unsigned long __init section_map_size(void)
445{
446	return ALIGN(sizeof(struct page) * PAGES_PER_SECTION, PMD_SIZE);
447}
448
449#else
450static unsigned long __init section_map_size(void)
451{
452	return PAGE_ALIGN(sizeof(struct page) * PAGES_PER_SECTION);
453}
454
455struct page __init *__populate_section_memmap(unsigned long pfn,
456		unsigned long nr_pages, int nid, struct vmem_altmap *altmap)
 
457{
458	unsigned long size = section_map_size();
459	struct page *map = sparse_buffer_alloc(size);
460	phys_addr_t addr = __pa(MAX_DMA_ADDRESS);
461
462	if (map)
463		return map;
464
465	map = memblock_alloc_try_nid_raw(size, size, addr,
466					  MEMBLOCK_ALLOC_ACCESSIBLE, nid);
467	if (!map)
468		panic("%s: Failed to allocate %lu bytes align=0x%lx nid=%d from=%pa\n",
469		      __func__, size, PAGE_SIZE, nid, &addr);
470
471	return map;
472}
473#endif /* !CONFIG_SPARSEMEM_VMEMMAP */
474
475static void *sparsemap_buf __meminitdata;
476static void *sparsemap_buf_end __meminitdata;
477
478static inline void __meminit sparse_buffer_free(unsigned long size)
479{
480	WARN_ON(!sparsemap_buf || size == 0);
481	memblock_free_early(__pa(sparsemap_buf), size);
482}
483
484static void __init sparse_buffer_init(unsigned long size, int nid)
485{
486	phys_addr_t addr = __pa(MAX_DMA_ADDRESS);
487	WARN_ON(sparsemap_buf);	/* forgot to call sparse_buffer_fini()? */
488	/*
489	 * Pre-allocated buffer is mainly used by __populate_section_memmap
490	 * and we want it to be properly aligned to the section size - this is
491	 * especially the case for VMEMMAP which maps memmap to PMDs
492	 */
493	sparsemap_buf = memblock_alloc_exact_nid_raw(size, section_map_size(),
494					addr, MEMBLOCK_ALLOC_ACCESSIBLE, nid);
495	sparsemap_buf_end = sparsemap_buf + size;
 
 
 
496}
497
498static void __init sparse_buffer_fini(void)
499{
500	unsigned long size = sparsemap_buf_end - sparsemap_buf;
501
502	if (sparsemap_buf && size > 0)
503		sparse_buffer_free(size);
504	sparsemap_buf = NULL;
505}
506
507void * __meminit sparse_buffer_alloc(unsigned long size)
508{
509	void *ptr = NULL;
510
511	if (sparsemap_buf) {
512		ptr = (void *) roundup((unsigned long)sparsemap_buf, size);
513		if (ptr + size > sparsemap_buf_end)
514			ptr = NULL;
515		else {
516			/* Free redundant aligned space */
517			if ((unsigned long)(ptr - sparsemap_buf) > 0)
518				sparse_buffer_free((unsigned long)(ptr - sparsemap_buf));
519			sparsemap_buf = ptr + size;
520		}
521	}
522	return ptr;
523}
524
525void __weak __meminit vmemmap_populate_print_last(void)
526{
527}
528
529/*
530 * Initialize sparse on a specific node. The node spans [pnum_begin, pnum_end)
531 * And number of present sections in this node is map_count.
532 */
533static void __init sparse_init_nid(int nid, unsigned long pnum_begin,
534				   unsigned long pnum_end,
535				   unsigned long map_count)
536{
537	struct mem_section_usage *usage;
538	unsigned long pnum;
539	struct page *map;
540
541	usage = sparse_early_usemaps_alloc_pgdat_section(NODE_DATA(nid),
542			mem_section_usage_size() * map_count);
543	if (!usage) {
544		pr_err("%s: node[%d] usemap allocation failed", __func__, nid);
545		goto failed;
546	}
547	sparse_buffer_init(map_count * section_map_size(), nid);
548	for_each_present_section_nr(pnum_begin, pnum) {
549		unsigned long pfn = section_nr_to_pfn(pnum);
550
551		if (pnum >= pnum_end)
552			break;
553
554		map = __populate_section_memmap(pfn, PAGES_PER_SECTION,
555				nid, NULL);
556		if (!map) {
557			pr_err("%s: node[%d] memory map backing failed. Some memory will not be available.",
558			       __func__, nid);
559			pnum_begin = pnum;
560			sparse_buffer_fini();
561			goto failed;
562		}
563		check_usemap_section_nr(nid, usage);
564		sparse_init_one_section(__nr_to_section(pnum), pnum, map, usage,
565				SECTION_IS_EARLY);
566		usage = (void *) usage + mem_section_usage_size();
567	}
568	sparse_buffer_fini();
569	return;
570failed:
571	/* We failed to allocate, mark all the following pnums as not present */
572	for_each_present_section_nr(pnum_begin, pnum) {
573		struct mem_section *ms;
574
575		if (pnum >= pnum_end)
576			break;
577		ms = __nr_to_section(pnum);
578		ms->section_mem_map = 0;
579	}
580}
581
582/*
583 * Allocate the accumulated non-linear sections, allocate a mem_map
584 * for each and record the physical to section mapping.
585 */
586void __init sparse_init(void)
587{
588	unsigned long pnum_end, pnum_begin, map_count = 1;
589	int nid_begin;
590
 
 
591	memblocks_present();
592
593	pnum_begin = first_present_section_nr();
594	nid_begin = sparse_early_nid(__nr_to_section(pnum_begin));
595
596	/* Setup pageblock_order for HUGETLB_PAGE_SIZE_VARIABLE */
597	set_pageblock_order();
598
599	for_each_present_section_nr(pnum_begin + 1, pnum_end) {
600		int nid = sparse_early_nid(__nr_to_section(pnum_end));
601
602		if (nid == nid_begin) {
603			map_count++;
604			continue;
605		}
606		/* Init node with sections in range [pnum_begin, pnum_end) */
607		sparse_init_nid(nid_begin, pnum_begin, pnum_end, map_count);
608		nid_begin = nid;
609		pnum_begin = pnum_end;
610		map_count = 1;
611	}
612	/* cover the last node */
613	sparse_init_nid(nid_begin, pnum_begin, pnum_end, map_count);
614	vmemmap_populate_print_last();
615}
616
617#ifdef CONFIG_MEMORY_HOTPLUG
618
619/* Mark all memory sections within the pfn range as online */
620void online_mem_sections(unsigned long start_pfn, unsigned long end_pfn)
621{
622	unsigned long pfn;
623
624	for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
625		unsigned long section_nr = pfn_to_section_nr(pfn);
626		struct mem_section *ms;
627
628		/* onlining code should never touch invalid ranges */
629		if (WARN_ON(!valid_section_nr(section_nr)))
630			continue;
631
632		ms = __nr_to_section(section_nr);
633		ms->section_mem_map |= SECTION_IS_ONLINE;
634	}
635}
636
637/* Mark all memory sections within the pfn range as offline */
638void offline_mem_sections(unsigned long start_pfn, unsigned long end_pfn)
639{
640	unsigned long pfn;
641
642	for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
643		unsigned long section_nr = pfn_to_section_nr(pfn);
644		struct mem_section *ms;
645
646		/*
647		 * TODO this needs some double checking. Offlining code makes
648		 * sure to check pfn_valid but those checks might be just bogus
649		 */
650		if (WARN_ON(!valid_section_nr(section_nr)))
651			continue;
652
653		ms = __nr_to_section(section_nr);
654		ms->section_mem_map &= ~SECTION_IS_ONLINE;
655	}
656}
657
658#ifdef CONFIG_SPARSEMEM_VMEMMAP
659static struct page * __meminit populate_section_memmap(unsigned long pfn,
660		unsigned long nr_pages, int nid, struct vmem_altmap *altmap)
 
661{
662	return __populate_section_memmap(pfn, nr_pages, nid, altmap);
663}
664
665static void depopulate_section_memmap(unsigned long pfn, unsigned long nr_pages,
666		struct vmem_altmap *altmap)
667{
668	unsigned long start = (unsigned long) pfn_to_page(pfn);
669	unsigned long end = start + nr_pages * sizeof(struct page);
670
 
671	vmemmap_free(start, end, altmap);
672}
673static void free_map_bootmem(struct page *memmap)
674{
675	unsigned long start = (unsigned long)memmap;
676	unsigned long end = (unsigned long)(memmap + PAGES_PER_SECTION);
677
678	vmemmap_free(start, end, NULL);
679}
680
681static int clear_subsection_map(unsigned long pfn, unsigned long nr_pages)
682{
683	DECLARE_BITMAP(map, SUBSECTIONS_PER_SECTION) = { 0 };
684	DECLARE_BITMAP(tmp, SUBSECTIONS_PER_SECTION) = { 0 };
685	struct mem_section *ms = __pfn_to_section(pfn);
686	unsigned long *subsection_map = ms->usage
687		? &ms->usage->subsection_map[0] : NULL;
688
689	subsection_mask_set(map, pfn, nr_pages);
690	if (subsection_map)
691		bitmap_and(tmp, map, subsection_map, SUBSECTIONS_PER_SECTION);
692
693	if (WARN(!subsection_map || !bitmap_equal(tmp, map, SUBSECTIONS_PER_SECTION),
694				"section already deactivated (%#lx + %ld)\n",
695				pfn, nr_pages))
696		return -EINVAL;
697
698	bitmap_xor(subsection_map, map, subsection_map, SUBSECTIONS_PER_SECTION);
699	return 0;
700}
701
702static bool is_subsection_map_empty(struct mem_section *ms)
703{
704	return bitmap_empty(&ms->usage->subsection_map[0],
705			    SUBSECTIONS_PER_SECTION);
706}
707
708static int fill_subsection_map(unsigned long pfn, unsigned long nr_pages)
709{
710	struct mem_section *ms = __pfn_to_section(pfn);
711	DECLARE_BITMAP(map, SUBSECTIONS_PER_SECTION) = { 0 };
712	unsigned long *subsection_map;
713	int rc = 0;
714
715	subsection_mask_set(map, pfn, nr_pages);
716
717	subsection_map = &ms->usage->subsection_map[0];
718
719	if (bitmap_empty(map, SUBSECTIONS_PER_SECTION))
720		rc = -EINVAL;
721	else if (bitmap_intersects(map, subsection_map, SUBSECTIONS_PER_SECTION))
722		rc = -EEXIST;
723	else
724		bitmap_or(subsection_map, map, subsection_map,
725				SUBSECTIONS_PER_SECTION);
726
727	return rc;
728}
729#else
730struct page * __meminit populate_section_memmap(unsigned long pfn,
731		unsigned long nr_pages, int nid, struct vmem_altmap *altmap)
 
732{
733	return kvmalloc_node(array_size(sizeof(struct page),
734					PAGES_PER_SECTION), GFP_KERNEL, nid);
735}
736
737static void depopulate_section_memmap(unsigned long pfn, unsigned long nr_pages,
738		struct vmem_altmap *altmap)
739{
740	kvfree(pfn_to_page(pfn));
741}
742
743static void free_map_bootmem(struct page *memmap)
744{
745	unsigned long maps_section_nr, removing_section_nr, i;
746	unsigned long magic, nr_pages;
747	struct page *page = virt_to_page(memmap);
748
749	nr_pages = PAGE_ALIGN(PAGES_PER_SECTION * sizeof(struct page))
750		>> PAGE_SHIFT;
751
752	for (i = 0; i < nr_pages; i++, page++) {
753		magic = (unsigned long) page->freelist;
754
755		BUG_ON(magic == NODE_INFO);
756
757		maps_section_nr = pfn_to_section_nr(page_to_pfn(page));
758		removing_section_nr = page_private(page);
759
760		/*
761		 * When this function is called, the removing section is
762		 * logical offlined state. This means all pages are isolated
763		 * from page allocator. If removing section's memmap is placed
764		 * on the same section, it must not be freed.
765		 * If it is freed, page allocator may allocate it which will
766		 * be removed physically soon.
767		 */
768		if (maps_section_nr != removing_section_nr)
769			put_page_bootmem(page);
770	}
771}
772
773static int clear_subsection_map(unsigned long pfn, unsigned long nr_pages)
774{
775	return 0;
776}
777
778static bool is_subsection_map_empty(struct mem_section *ms)
779{
780	return true;
781}
782
783static int fill_subsection_map(unsigned long pfn, unsigned long nr_pages)
784{
785	return 0;
786}
787#endif /* CONFIG_SPARSEMEM_VMEMMAP */
788
789/*
790 * To deactivate a memory region, there are 3 cases to handle across
791 * two configurations (SPARSEMEM_VMEMMAP={y,n}):
792 *
793 * 1. deactivation of a partial hot-added section (only possible in
794 *    the SPARSEMEM_VMEMMAP=y case).
795 *      a) section was present at memory init.
796 *      b) section was hot-added post memory init.
797 * 2. deactivation of a complete hot-added section.
798 * 3. deactivation of a complete section from memory init.
799 *
800 * For 1, when subsection_map does not empty we will not be freeing the
801 * usage map, but still need to free the vmemmap range.
802 *
803 * For 2 and 3, the SPARSEMEM_VMEMMAP={y,n} cases are unified
804 */
805static void section_deactivate(unsigned long pfn, unsigned long nr_pages,
806		struct vmem_altmap *altmap)
807{
808	struct mem_section *ms = __pfn_to_section(pfn);
809	bool section_is_early = early_section(ms);
810	struct page *memmap = NULL;
811	bool empty;
812
813	if (clear_subsection_map(pfn, nr_pages))
814		return;
815
816	empty = is_subsection_map_empty(ms);
817	if (empty) {
818		unsigned long section_nr = pfn_to_section_nr(pfn);
819
820		/*
 
 
 
 
 
 
 
821		 * When removing an early section, the usage map is kept (as the
822		 * usage maps of other sections fall into the same page). It
823		 * will be re-used when re-adding the section - which is then no
824		 * longer an early section. If the usage map is PageReserved, it
825		 * was allocated during boot.
826		 */
827		if (!PageReserved(virt_to_page(ms->usage))) {
828			kfree(ms->usage);
829			ms->usage = NULL;
830		}
831		memmap = sparse_decode_mem_map(ms->section_mem_map, section_nr);
832		/*
833		 * Mark the section invalid so that valid_section()
834		 * return false. This prevents code from dereferencing
835		 * ms->usage array.
836		 */
837		ms->section_mem_map &= ~SECTION_HAS_MEM_MAP;
838	}
839
840	/*
841	 * The memmap of early sections is always fully populated. See
842	 * section_activate() and pfn_valid() .
843	 */
844	if (!section_is_early)
845		depopulate_section_memmap(pfn, nr_pages, altmap);
846	else if (memmap)
847		free_map_bootmem(memmap);
848
849	if (empty)
850		ms->section_mem_map = (unsigned long)NULL;
851}
852
853static struct page * __meminit section_activate(int nid, unsigned long pfn,
854		unsigned long nr_pages, struct vmem_altmap *altmap)
 
855{
856	struct mem_section *ms = __pfn_to_section(pfn);
857	struct mem_section_usage *usage = NULL;
858	struct page *memmap;
859	int rc = 0;
860
861	if (!ms->usage) {
862		usage = kzalloc(mem_section_usage_size(), GFP_KERNEL);
863		if (!usage)
864			return ERR_PTR(-ENOMEM);
865		ms->usage = usage;
866	}
867
868	rc = fill_subsection_map(pfn, nr_pages);
869	if (rc) {
870		if (usage)
871			ms->usage = NULL;
872		kfree(usage);
873		return ERR_PTR(rc);
874	}
875
876	/*
877	 * The early init code does not consider partially populated
878	 * initial sections, it simply assumes that memory will never be
879	 * referenced.  If we hot-add memory into such a section then we
880	 * do not need to populate the memmap and can simply reuse what
881	 * is already there.
882	 */
883	if (nr_pages < PAGES_PER_SECTION && early_section(ms))
884		return pfn_to_page(pfn);
885
886	memmap = populate_section_memmap(pfn, nr_pages, nid, altmap);
887	if (!memmap) {
888		section_deactivate(pfn, nr_pages, altmap);
889		return ERR_PTR(-ENOMEM);
890	}
891
892	return memmap;
893}
894
895/**
896 * sparse_add_section - add a memory section, or populate an existing one
897 * @nid: The node to add section on
898 * @start_pfn: start pfn of the memory range
899 * @nr_pages: number of pfns to add in the section
900 * @altmap: device page map
 
901 *
902 * This is only intended for hotplug.
903 *
904 * Note that only VMEMMAP supports sub-section aligned hotplug,
905 * the proper alignment and size are gated by check_pfn_span().
906 *
907 *
908 * Return:
909 * * 0		- On success.
910 * * -EEXIST	- Section has been present.
911 * * -ENOMEM	- Out of memory.
912 */
913int __meminit sparse_add_section(int nid, unsigned long start_pfn,
914		unsigned long nr_pages, struct vmem_altmap *altmap)
 
915{
916	unsigned long section_nr = pfn_to_section_nr(start_pfn);
917	struct mem_section *ms;
918	struct page *memmap;
919	int ret;
920
921	ret = sparse_index_init(section_nr, nid);
922	if (ret < 0)
923		return ret;
924
925	memmap = section_activate(nid, start_pfn, nr_pages, altmap);
926	if (IS_ERR(memmap))
927		return PTR_ERR(memmap);
928
929	/*
930	 * Poison uninitialized struct pages in order to catch invalid flags
931	 * combinations.
932	 */
933	page_init_poison(memmap, sizeof(struct page) * nr_pages);
 
934
935	ms = __nr_to_section(section_nr);
936	set_section_nid(section_nr, nid);
937	section_mark_present(ms);
938
939	/* Align memmap to section boundary in the subsection case */
940	if (section_nr_to_pfn(section_nr) != start_pfn)
941		memmap = pfn_to_page(section_nr_to_pfn(section_nr));
942	sparse_init_one_section(ms, section_nr, memmap, ms->usage, 0);
943
944	return 0;
945}
946
947#ifdef CONFIG_MEMORY_FAILURE
948static void clear_hwpoisoned_pages(struct page *memmap, int nr_pages)
949{
950	int i;
951
952	/*
953	 * A further optimization is to have per section refcounted
954	 * num_poisoned_pages.  But that would need more space per memmap, so
955	 * for now just do a quick global check to speed up this routine in the
956	 * absence of bad pages.
957	 */
958	if (atomic_long_read(&num_poisoned_pages) == 0)
959		return;
960
961	for (i = 0; i < nr_pages; i++) {
962		if (PageHWPoison(&memmap[i])) {
963			num_poisoned_pages_dec();
964			ClearPageHWPoison(&memmap[i]);
965		}
966	}
967}
968#else
969static inline void clear_hwpoisoned_pages(struct page *memmap, int nr_pages)
970{
971}
972#endif
973
974void sparse_remove_section(struct mem_section *ms, unsigned long pfn,
975		unsigned long nr_pages, unsigned long map_offset,
976		struct vmem_altmap *altmap)
977{
978	clear_hwpoisoned_pages(pfn_to_page(pfn) + map_offset,
979			nr_pages - map_offset);
980	section_deactivate(pfn, nr_pages, altmap);
981}
982#endif /* CONFIG_MEMORY_HOTPLUG */
v6.13.7
  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#include <linux/bootmem_info.h>
 17#include <linux/vmstat.h>
 18#include "internal.h"
 19#include <asm/dma.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
 27struct mem_section **mem_section;
 28#else
 29struct mem_section mem_section[NR_SECTION_ROOTS][SECTIONS_PER_ROOT]
 30	____cacheline_internodealigned_in_smp;
 31#endif
 32EXPORT_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
 41static u8 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned;
 42#else
 43static u16 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned;
 44#endif
 45
 46int page_to_nid(const struct page *page)
 47{
 48	return section_to_node_table[page_to_section(page)];
 49}
 50EXPORT_SYMBOL(page_to_nid);
 51
 52static 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 */
 57static inline void set_section_nid(unsigned long section_nr, int nid)
 58{
 59}
 60#endif
 61
 62#ifdef CONFIG_SPARSEMEM_EXTREME
 63static 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
 82static 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 */
106static inline int sparse_index_init(unsigned long section_nr, int nid)
107{
108	return 0;
109}
110#endif
111
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
112/*
113 * During early boot, before section_mem_map is used for an actual
114 * mem_map, we use section_mem_map to store the section's NUMA
115 * node.  This keeps us from having to use another data structure.  The
116 * node information is cleared just before we store the real mem_map.
117 */
118static inline unsigned long sparse_encode_early_nid(int nid)
119{
120	return ((unsigned long)nid << SECTION_NID_SHIFT);
121}
122
123static inline int sparse_early_nid(struct mem_section *section)
124{
125	return (section->section_mem_map >> SECTION_NID_SHIFT);
126}
127
128/* Validate the physical addressing limitations of the model */
129static void __meminit mminit_validate_memmodel_limits(unsigned long *start_pfn,
130						unsigned long *end_pfn)
131{
132	unsigned long max_sparsemem_pfn = (DIRECT_MAP_PHYSMEM_END + 1) >> PAGE_SHIFT;
133
134	/*
135	 * Sanity checks - do not allow an architecture to pass
136	 * in larger pfns than the maximum scope of sparsemem:
137	 */
138	if (*start_pfn > max_sparsemem_pfn) {
139		mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
140			"Start of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
141			*start_pfn, *end_pfn, max_sparsemem_pfn);
142		WARN_ON_ONCE(1);
143		*start_pfn = max_sparsemem_pfn;
144		*end_pfn = max_sparsemem_pfn;
145	} else if (*end_pfn > max_sparsemem_pfn) {
146		mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
147			"End of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
148			*start_pfn, *end_pfn, max_sparsemem_pfn);
149		WARN_ON_ONCE(1);
150		*end_pfn = max_sparsemem_pfn;
151	}
152}
153
154/*
155 * There are a number of times that we loop over NR_MEM_SECTIONS,
156 * looking for section_present() on each.  But, when we have very
157 * large physical address spaces, NR_MEM_SECTIONS can also be
158 * very large which makes the loops quite long.
159 *
160 * Keeping track of this gives us an easy way to break out of
161 * those loops early.
162 */
163unsigned long __highest_present_section_nr;
164static void __section_mark_present(struct mem_section *ms,
165		unsigned long section_nr)
166{
 
 
167	if (section_nr > __highest_present_section_nr)
168		__highest_present_section_nr = section_nr;
169
170	ms->section_mem_map |= SECTION_MARKED_PRESENT;
171}
172
173#define for_each_present_section_nr(start, section_nr)		\
174	for (section_nr = next_present_section_nr(start-1);	\
175	     section_nr != -1;								\
 
176	     section_nr = next_present_section_nr(section_nr))
177
178static inline unsigned long first_present_section_nr(void)
179{
180	return next_present_section_nr(-1);
181}
182
183#ifdef CONFIG_SPARSEMEM_VMEMMAP
184static void subsection_mask_set(unsigned long *map, unsigned long pfn,
185		unsigned long nr_pages)
186{
187	int idx = subsection_map_index(pfn);
188	int end = subsection_map_index(pfn + nr_pages - 1);
189
190	bitmap_set(map, idx, end - idx + 1);
191}
192
193void __init subsection_map_init(unsigned long pfn, unsigned long nr_pages)
194{
195	int end_sec_nr = pfn_to_section_nr(pfn + nr_pages - 1);
196	unsigned long nr, start_sec_nr = pfn_to_section_nr(pfn);
 
 
 
197
198	for (nr = start_sec_nr; nr <= end_sec_nr; nr++) {
199		struct mem_section *ms;
200		unsigned long pfns;
201
202		pfns = min(nr_pages, PAGES_PER_SECTION
203				- (pfn & ~PAGE_SECTION_MASK));
204		ms = __nr_to_section(nr);
205		subsection_mask_set(ms->usage->subsection_map, pfn, pfns);
206
207		pr_debug("%s: sec: %lu pfns: %lu set(%d, %d)\n", __func__, nr,
208				pfns, subsection_map_index(pfn),
209				subsection_map_index(pfn + pfns - 1));
210
211		pfn += pfns;
212		nr_pages -= pfns;
213	}
214}
215#else
216void __init subsection_map_init(unsigned long pfn, unsigned long nr_pages)
217{
218}
219#endif
220
221/* Record a memory area against a node. */
222static void __init memory_present(int nid, unsigned long start, unsigned long end)
223{
224	unsigned long pfn;
225
 
 
 
 
 
 
 
 
 
 
 
 
 
226	start &= PAGE_SECTION_MASK;
227	mminit_validate_memmodel_limits(&start, &end);
228	for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) {
229		unsigned long section_nr = pfn_to_section_nr(pfn);
230		struct mem_section *ms;
231
232		sparse_index_init(section_nr, nid);
233		set_section_nid(section_nr, nid);
234
235		ms = __nr_to_section(section_nr);
236		if (!ms->section_mem_map) {
237			ms->section_mem_map = sparse_encode_early_nid(nid) |
238							SECTION_IS_ONLINE;
239			__section_mark_present(ms, section_nr);
240		}
241	}
242}
243
244/*
245 * Mark all memblocks as present using memory_present().
246 * This is a convenience function that is useful to mark all of the systems
247 * memory as present during initialization.
248 */
249static void __init memblocks_present(void)
250{
251	unsigned long start, end;
252	int i, nid;
253
254#ifdef CONFIG_SPARSEMEM_EXTREME
255	if (unlikely(!mem_section)) {
256		unsigned long size, align;
257
258		size = sizeof(struct mem_section *) * NR_SECTION_ROOTS;
259		align = 1 << (INTERNODE_CACHE_SHIFT);
260		mem_section = memblock_alloc(size, align);
261		if (!mem_section)
262			panic("%s: Failed to allocate %lu bytes align=0x%lx\n",
263			      __func__, size, align);
264	}
265#endif
266
267	for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid)
268		memory_present(nid, start, end);
269}
270
271/*
272 * Subtle, we encode the real pfn into the mem_map such that
273 * the identity pfn - section_mem_map will return the actual
274 * physical page frame number.
275 */
276static unsigned long sparse_encode_mem_map(struct page *mem_map, unsigned long pnum)
277{
278	unsigned long coded_mem_map =
279		(unsigned long)(mem_map - (section_nr_to_pfn(pnum)));
280	BUILD_BUG_ON(SECTION_MAP_LAST_BIT > PFN_SECTION_SHIFT);
281	BUG_ON(coded_mem_map & ~SECTION_MAP_MASK);
282	return coded_mem_map;
283}
284
285#ifdef CONFIG_MEMORY_HOTPLUG
286/*
287 * Decode mem_map from the coded memmap
288 */
289struct page *sparse_decode_mem_map(unsigned long coded_mem_map, unsigned long pnum)
290{
291	/* mask off the extra low bits of information */
292	coded_mem_map &= SECTION_MAP_MASK;
293	return ((struct page *)coded_mem_map) + section_nr_to_pfn(pnum);
294}
295#endif /* CONFIG_MEMORY_HOTPLUG */
296
297static void __meminit sparse_init_one_section(struct mem_section *ms,
298		unsigned long pnum, struct page *mem_map,
299		struct mem_section_usage *usage, unsigned long flags)
300{
301	ms->section_mem_map &= ~SECTION_MAP_MASK;
302	ms->section_mem_map |= sparse_encode_mem_map(mem_map, pnum)
303		| SECTION_HAS_MEM_MAP | flags;
304	ms->usage = usage;
305}
306
307static unsigned long usemap_size(void)
308{
309	return BITS_TO_LONGS(SECTION_BLOCKFLAGS_BITS) * sizeof(unsigned long);
310}
311
312size_t mem_section_usage_size(void)
313{
314	return sizeof(struct mem_section_usage) + usemap_size();
315}
316
317#ifdef CONFIG_MEMORY_HOTREMOVE
318static inline phys_addr_t pgdat_to_phys(struct pglist_data *pgdat)
319{
320#ifndef CONFIG_NUMA
321	VM_BUG_ON(pgdat != &contig_page_data);
322	return __pa_symbol(&contig_page_data);
323#else
324	return __pa(pgdat);
325#endif
326}
327
 
328static struct mem_section_usage * __init
329sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
330					 unsigned long size)
331{
332	struct mem_section_usage *usage;
333	unsigned long goal, limit;
334	int nid;
335	/*
336	 * A page may contain usemaps for other sections preventing the
337	 * page being freed and making a section unremovable while
338	 * other sections referencing the usemap remain active. Similarly,
339	 * a pgdat can prevent a section being removed. If section A
340	 * contains a pgdat and section B contains the usemap, both
341	 * sections become inter-dependent. This allocates usemaps
342	 * from the same section as the pgdat where possible to avoid
343	 * this problem.
344	 */
345	goal = pgdat_to_phys(pgdat) & (PAGE_SECTION_MASK << PAGE_SHIFT);
346	limit = goal + (1UL << PA_SECTION_SHIFT);
347	nid = early_pfn_to_nid(goal >> PAGE_SHIFT);
348again:
349	usage = memblock_alloc_try_nid(size, SMP_CACHE_BYTES, goal, limit, nid);
350	if (!usage && limit) {
351		limit = MEMBLOCK_ALLOC_ACCESSIBLE;
352		goto again;
353	}
354	return usage;
355}
356
357static void __init check_usemap_section_nr(int nid,
358		struct mem_section_usage *usage)
359{
360	unsigned long usemap_snr, pgdat_snr;
361	static unsigned long old_usemap_snr;
362	static unsigned long old_pgdat_snr;
363	struct pglist_data *pgdat = NODE_DATA(nid);
364	int usemap_nid;
365
366	/* First call */
367	if (!old_usemap_snr) {
368		old_usemap_snr = NR_MEM_SECTIONS;
369		old_pgdat_snr = NR_MEM_SECTIONS;
370	}
371
372	usemap_snr = pfn_to_section_nr(__pa(usage) >> PAGE_SHIFT);
373	pgdat_snr = pfn_to_section_nr(pgdat_to_phys(pgdat) >> PAGE_SHIFT);
374	if (usemap_snr == pgdat_snr)
375		return;
376
377	if (old_usemap_snr == usemap_snr && old_pgdat_snr == pgdat_snr)
378		/* skip redundant message */
379		return;
380
381	old_usemap_snr = usemap_snr;
382	old_pgdat_snr = pgdat_snr;
383
384	usemap_nid = sparse_early_nid(__nr_to_section(usemap_snr));
385	if (usemap_nid != nid) {
386		pr_info("node %d must be removed before remove section %ld\n",
387			nid, usemap_snr);
388		return;
389	}
390	/*
391	 * There is a circular dependency.
392	 * Some platforms allow un-removable section because they will just
393	 * gather other removable sections for dynamic partitioning.
394	 * Just notify un-removable section's number here.
395	 */
396	pr_info("Section %ld and %ld (node %d) have a circular dependency on usemap and pgdat allocations\n",
397		usemap_snr, pgdat_snr, nid);
398}
399#else
400static struct mem_section_usage * __init
401sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
402					 unsigned long size)
403{
404	return memblock_alloc_node(size, SMP_CACHE_BYTES, pgdat->node_id);
405}
406
407static void __init check_usemap_section_nr(int nid,
408		struct mem_section_usage *usage)
409{
410}
411#endif /* CONFIG_MEMORY_HOTREMOVE */
412
413#ifdef CONFIG_SPARSEMEM_VMEMMAP
414static unsigned long __init section_map_size(void)
415{
416	return ALIGN(sizeof(struct page) * PAGES_PER_SECTION, PMD_SIZE);
417}
418
419#else
420static unsigned long __init section_map_size(void)
421{
422	return PAGE_ALIGN(sizeof(struct page) * PAGES_PER_SECTION);
423}
424
425struct page __init *__populate_section_memmap(unsigned long pfn,
426		unsigned long nr_pages, int nid, struct vmem_altmap *altmap,
427		struct dev_pagemap *pgmap)
428{
429	unsigned long size = section_map_size();
430	struct page *map = sparse_buffer_alloc(size);
431	phys_addr_t addr = __pa(MAX_DMA_ADDRESS);
432
433	if (map)
434		return map;
435
436	map = memmap_alloc(size, size, addr, nid, false);
 
437	if (!map)
438		panic("%s: Failed to allocate %lu bytes align=0x%lx nid=%d from=%pa\n",
439		      __func__, size, PAGE_SIZE, nid, &addr);
440
441	return map;
442}
443#endif /* !CONFIG_SPARSEMEM_VMEMMAP */
444
445static void *sparsemap_buf __meminitdata;
446static void *sparsemap_buf_end __meminitdata;
447
448static inline void __meminit sparse_buffer_free(unsigned long size)
449{
450	WARN_ON(!sparsemap_buf || size == 0);
451	memblock_free(sparsemap_buf, size);
452}
453
454static void __init sparse_buffer_init(unsigned long size, int nid)
455{
456	phys_addr_t addr = __pa(MAX_DMA_ADDRESS);
457	WARN_ON(sparsemap_buf);	/* forgot to call sparse_buffer_fini()? */
458	/*
459	 * Pre-allocated buffer is mainly used by __populate_section_memmap
460	 * and we want it to be properly aligned to the section size - this is
461	 * especially the case for VMEMMAP which maps memmap to PMDs
462	 */
463	sparsemap_buf = memmap_alloc(size, section_map_size(), addr, nid, true);
 
464	sparsemap_buf_end = sparsemap_buf + size;
465#ifndef CONFIG_SPARSEMEM_VMEMMAP
466	memmap_boot_pages_add(DIV_ROUND_UP(size, PAGE_SIZE));
467#endif
468}
469
470static void __init sparse_buffer_fini(void)
471{
472	unsigned long size = sparsemap_buf_end - sparsemap_buf;
473
474	if (sparsemap_buf && size > 0)
475		sparse_buffer_free(size);
476	sparsemap_buf = NULL;
477}
478
479void * __meminit sparse_buffer_alloc(unsigned long size)
480{
481	void *ptr = NULL;
482
483	if (sparsemap_buf) {
484		ptr = (void *) roundup((unsigned long)sparsemap_buf, size);
485		if (ptr + size > sparsemap_buf_end)
486			ptr = NULL;
487		else {
488			/* Free redundant aligned space */
489			if ((unsigned long)(ptr - sparsemap_buf) > 0)
490				sparse_buffer_free((unsigned long)(ptr - sparsemap_buf));
491			sparsemap_buf = ptr + size;
492		}
493	}
494	return ptr;
495}
496
497void __weak __meminit vmemmap_populate_print_last(void)
498{
499}
500
501/*
502 * Initialize sparse on a specific node. The node spans [pnum_begin, pnum_end)
503 * And number of present sections in this node is map_count.
504 */
505static void __init sparse_init_nid(int nid, unsigned long pnum_begin,
506				   unsigned long pnum_end,
507				   unsigned long map_count)
508{
509	struct mem_section_usage *usage;
510	unsigned long pnum;
511	struct page *map;
512
513	usage = sparse_early_usemaps_alloc_pgdat_section(NODE_DATA(nid),
514			mem_section_usage_size() * map_count);
515	if (!usage) {
516		pr_err("%s: node[%d] usemap allocation failed", __func__, nid);
517		goto failed;
518	}
519	sparse_buffer_init(map_count * section_map_size(), nid);
520	for_each_present_section_nr(pnum_begin, pnum) {
521		unsigned long pfn = section_nr_to_pfn(pnum);
522
523		if (pnum >= pnum_end)
524			break;
525
526		map = __populate_section_memmap(pfn, PAGES_PER_SECTION,
527				nid, NULL, NULL);
528		if (!map) {
529			pr_err("%s: node[%d] memory map backing failed. Some memory will not be available.",
530			       __func__, nid);
531			pnum_begin = pnum;
532			sparse_buffer_fini();
533			goto failed;
534		}
535		check_usemap_section_nr(nid, usage);
536		sparse_init_one_section(__nr_to_section(pnum), pnum, map, usage,
537				SECTION_IS_EARLY);
538		usage = (void *) usage + mem_section_usage_size();
539	}
540	sparse_buffer_fini();
541	return;
542failed:
543	/* We failed to allocate, mark all the following pnums as not present */
544	for_each_present_section_nr(pnum_begin, pnum) {
545		struct mem_section *ms;
546
547		if (pnum >= pnum_end)
548			break;
549		ms = __nr_to_section(pnum);
550		ms->section_mem_map = 0;
551	}
552}
553
554/*
555 * Allocate the accumulated non-linear sections, allocate a mem_map
556 * for each and record the physical to section mapping.
557 */
558void __init sparse_init(void)
559{
560	unsigned long pnum_end, pnum_begin, map_count = 1;
561	int nid_begin;
562
563	/* see include/linux/mmzone.h 'struct mem_section' definition */
564	BUILD_BUG_ON(!is_power_of_2(sizeof(struct mem_section)));
565	memblocks_present();
566
567	pnum_begin = first_present_section_nr();
568	nid_begin = sparse_early_nid(__nr_to_section(pnum_begin));
569
570	/* Setup pageblock_order for HUGETLB_PAGE_SIZE_VARIABLE */
571	set_pageblock_order();
572
573	for_each_present_section_nr(pnum_begin + 1, pnum_end) {
574		int nid = sparse_early_nid(__nr_to_section(pnum_end));
575
576		if (nid == nid_begin) {
577			map_count++;
578			continue;
579		}
580		/* Init node with sections in range [pnum_begin, pnum_end) */
581		sparse_init_nid(nid_begin, pnum_begin, pnum_end, map_count);
582		nid_begin = nid;
583		pnum_begin = pnum_end;
584		map_count = 1;
585	}
586	/* cover the last node */
587	sparse_init_nid(nid_begin, pnum_begin, pnum_end, map_count);
588	vmemmap_populate_print_last();
589}
590
591#ifdef CONFIG_MEMORY_HOTPLUG
592
593/* Mark all memory sections within the pfn range as online */
594void online_mem_sections(unsigned long start_pfn, unsigned long end_pfn)
595{
596	unsigned long pfn;
597
598	for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
599		unsigned long section_nr = pfn_to_section_nr(pfn);
600		struct mem_section *ms;
601
602		/* onlining code should never touch invalid ranges */
603		if (WARN_ON(!valid_section_nr(section_nr)))
604			continue;
605
606		ms = __nr_to_section(section_nr);
607		ms->section_mem_map |= SECTION_IS_ONLINE;
608	}
609}
610
611/* Mark all memory sections within the pfn range as offline */
612void offline_mem_sections(unsigned long start_pfn, unsigned long end_pfn)
613{
614	unsigned long pfn;
615
616	for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
617		unsigned long section_nr = pfn_to_section_nr(pfn);
618		struct mem_section *ms;
619
620		/*
621		 * TODO this needs some double checking. Offlining code makes
622		 * sure to check pfn_valid but those checks might be just bogus
623		 */
624		if (WARN_ON(!valid_section_nr(section_nr)))
625			continue;
626
627		ms = __nr_to_section(section_nr);
628		ms->section_mem_map &= ~SECTION_IS_ONLINE;
629	}
630}
631
632#ifdef CONFIG_SPARSEMEM_VMEMMAP
633static struct page * __meminit populate_section_memmap(unsigned long pfn,
634		unsigned long nr_pages, int nid, struct vmem_altmap *altmap,
635		struct dev_pagemap *pgmap)
636{
637	return __populate_section_memmap(pfn, nr_pages, nid, altmap, pgmap);
638}
639
640static void depopulate_section_memmap(unsigned long pfn, unsigned long nr_pages,
641		struct vmem_altmap *altmap)
642{
643	unsigned long start = (unsigned long) pfn_to_page(pfn);
644	unsigned long end = start + nr_pages * sizeof(struct page);
645
646	memmap_pages_add(-1L * (DIV_ROUND_UP(end - start, PAGE_SIZE)));
647	vmemmap_free(start, end, altmap);
648}
649static void free_map_bootmem(struct page *memmap)
650{
651	unsigned long start = (unsigned long)memmap;
652	unsigned long end = (unsigned long)(memmap + PAGES_PER_SECTION);
653
654	vmemmap_free(start, end, NULL);
655}
656
657static int clear_subsection_map(unsigned long pfn, unsigned long nr_pages)
658{
659	DECLARE_BITMAP(map, SUBSECTIONS_PER_SECTION) = { 0 };
660	DECLARE_BITMAP(tmp, SUBSECTIONS_PER_SECTION) = { 0 };
661	struct mem_section *ms = __pfn_to_section(pfn);
662	unsigned long *subsection_map = ms->usage
663		? &ms->usage->subsection_map[0] : NULL;
664
665	subsection_mask_set(map, pfn, nr_pages);
666	if (subsection_map)
667		bitmap_and(tmp, map, subsection_map, SUBSECTIONS_PER_SECTION);
668
669	if (WARN(!subsection_map || !bitmap_equal(tmp, map, SUBSECTIONS_PER_SECTION),
670				"section already deactivated (%#lx + %ld)\n",
671				pfn, nr_pages))
672		return -EINVAL;
673
674	bitmap_xor(subsection_map, map, subsection_map, SUBSECTIONS_PER_SECTION);
675	return 0;
676}
677
678static bool is_subsection_map_empty(struct mem_section *ms)
679{
680	return bitmap_empty(&ms->usage->subsection_map[0],
681			    SUBSECTIONS_PER_SECTION);
682}
683
684static int fill_subsection_map(unsigned long pfn, unsigned long nr_pages)
685{
686	struct mem_section *ms = __pfn_to_section(pfn);
687	DECLARE_BITMAP(map, SUBSECTIONS_PER_SECTION) = { 0 };
688	unsigned long *subsection_map;
689	int rc = 0;
690
691	subsection_mask_set(map, pfn, nr_pages);
692
693	subsection_map = &ms->usage->subsection_map[0];
694
695	if (bitmap_empty(map, SUBSECTIONS_PER_SECTION))
696		rc = -EINVAL;
697	else if (bitmap_intersects(map, subsection_map, SUBSECTIONS_PER_SECTION))
698		rc = -EEXIST;
699	else
700		bitmap_or(subsection_map, map, subsection_map,
701				SUBSECTIONS_PER_SECTION);
702
703	return rc;
704}
705#else
706static struct page * __meminit populate_section_memmap(unsigned long pfn,
707		unsigned long nr_pages, int nid, struct vmem_altmap *altmap,
708		struct dev_pagemap *pgmap)
709{
710	return kvmalloc_node(array_size(sizeof(struct page),
711					PAGES_PER_SECTION), GFP_KERNEL, nid);
712}
713
714static void depopulate_section_memmap(unsigned long pfn, unsigned long nr_pages,
715		struct vmem_altmap *altmap)
716{
717	kvfree(pfn_to_page(pfn));
718}
719
720static void free_map_bootmem(struct page *memmap)
721{
722	unsigned long maps_section_nr, removing_section_nr, i;
723	unsigned long type, nr_pages;
724	struct page *page = virt_to_page(memmap);
725
726	nr_pages = PAGE_ALIGN(PAGES_PER_SECTION * sizeof(struct page))
727		>> PAGE_SHIFT;
728
729	for (i = 0; i < nr_pages; i++, page++) {
730		type = bootmem_type(page);
731
732		BUG_ON(type == NODE_INFO);
733
734		maps_section_nr = pfn_to_section_nr(page_to_pfn(page));
735		removing_section_nr = bootmem_info(page);
736
737		/*
738		 * When this function is called, the removing section is
739		 * logical offlined state. This means all pages are isolated
740		 * from page allocator. If removing section's memmap is placed
741		 * on the same section, it must not be freed.
742		 * If it is freed, page allocator may allocate it which will
743		 * be removed physically soon.
744		 */
745		if (maps_section_nr != removing_section_nr)
746			put_page_bootmem(page);
747	}
748}
749
750static int clear_subsection_map(unsigned long pfn, unsigned long nr_pages)
751{
752	return 0;
753}
754
755static bool is_subsection_map_empty(struct mem_section *ms)
756{
757	return true;
758}
759
760static int fill_subsection_map(unsigned long pfn, unsigned long nr_pages)
761{
762	return 0;
763}
764#endif /* CONFIG_SPARSEMEM_VMEMMAP */
765
766/*
767 * To deactivate a memory region, there are 3 cases to handle across
768 * two configurations (SPARSEMEM_VMEMMAP={y,n}):
769 *
770 * 1. deactivation of a partial hot-added section (only possible in
771 *    the SPARSEMEM_VMEMMAP=y case).
772 *      a) section was present at memory init.
773 *      b) section was hot-added post memory init.
774 * 2. deactivation of a complete hot-added section.
775 * 3. deactivation of a complete section from memory init.
776 *
777 * For 1, when subsection_map does not empty we will not be freeing the
778 * usage map, but still need to free the vmemmap range.
779 *
780 * For 2 and 3, the SPARSEMEM_VMEMMAP={y,n} cases are unified
781 */
782static void section_deactivate(unsigned long pfn, unsigned long nr_pages,
783		struct vmem_altmap *altmap)
784{
785	struct mem_section *ms = __pfn_to_section(pfn);
786	bool section_is_early = early_section(ms);
787	struct page *memmap = NULL;
788	bool empty;
789
790	if (clear_subsection_map(pfn, nr_pages))
791		return;
792
793	empty = is_subsection_map_empty(ms);
794	if (empty) {
795		unsigned long section_nr = pfn_to_section_nr(pfn);
796
797		/*
798		 * Mark the section invalid so that valid_section()
799		 * return false. This prevents code from dereferencing
800		 * ms->usage array.
801		 */
802		ms->section_mem_map &= ~SECTION_HAS_MEM_MAP;
803
804		/*
805		 * When removing an early section, the usage map is kept (as the
806		 * usage maps of other sections fall into the same page). It
807		 * will be re-used when re-adding the section - which is then no
808		 * longer an early section. If the usage map is PageReserved, it
809		 * was allocated during boot.
810		 */
811		if (!PageReserved(virt_to_page(ms->usage))) {
812			kfree_rcu(ms->usage, rcu);
813			WRITE_ONCE(ms->usage, NULL);
814		}
815		memmap = sparse_decode_mem_map(ms->section_mem_map, section_nr);
 
 
 
 
 
 
816	}
817
818	/*
819	 * The memmap of early sections is always fully populated. See
820	 * section_activate() and pfn_valid() .
821	 */
822	if (!section_is_early)
823		depopulate_section_memmap(pfn, nr_pages, altmap);
824	else if (memmap)
825		free_map_bootmem(memmap);
826
827	if (empty)
828		ms->section_mem_map = (unsigned long)NULL;
829}
830
831static struct page * __meminit section_activate(int nid, unsigned long pfn,
832		unsigned long nr_pages, struct vmem_altmap *altmap,
833		struct dev_pagemap *pgmap)
834{
835	struct mem_section *ms = __pfn_to_section(pfn);
836	struct mem_section_usage *usage = NULL;
837	struct page *memmap;
838	int rc;
839
840	if (!ms->usage) {
841		usage = kzalloc(mem_section_usage_size(), GFP_KERNEL);
842		if (!usage)
843			return ERR_PTR(-ENOMEM);
844		ms->usage = usage;
845	}
846
847	rc = fill_subsection_map(pfn, nr_pages);
848	if (rc) {
849		if (usage)
850			ms->usage = NULL;
851		kfree(usage);
852		return ERR_PTR(rc);
853	}
854
855	/*
856	 * The early init code does not consider partially populated
857	 * initial sections, it simply assumes that memory will never be
858	 * referenced.  If we hot-add memory into such a section then we
859	 * do not need to populate the memmap and can simply reuse what
860	 * is already there.
861	 */
862	if (nr_pages < PAGES_PER_SECTION && early_section(ms))
863		return pfn_to_page(pfn);
864
865	memmap = populate_section_memmap(pfn, nr_pages, nid, altmap, pgmap);
866	if (!memmap) {
867		section_deactivate(pfn, nr_pages, altmap);
868		return ERR_PTR(-ENOMEM);
869	}
870
871	return memmap;
872}
873
874/**
875 * sparse_add_section - add a memory section, or populate an existing one
876 * @nid: The node to add section on
877 * @start_pfn: start pfn of the memory range
878 * @nr_pages: number of pfns to add in the section
879 * @altmap: alternate pfns to allocate the memmap backing store
880 * @pgmap: alternate compound page geometry for devmap mappings
881 *
882 * This is only intended for hotplug.
883 *
884 * Note that only VMEMMAP supports sub-section aligned hotplug,
885 * the proper alignment and size are gated by check_pfn_span().
886 *
887 *
888 * Return:
889 * * 0		- On success.
890 * * -EEXIST	- Section has been present.
891 * * -ENOMEM	- Out of memory.
892 */
893int __meminit sparse_add_section(int nid, unsigned long start_pfn,
894		unsigned long nr_pages, struct vmem_altmap *altmap,
895		struct dev_pagemap *pgmap)
896{
897	unsigned long section_nr = pfn_to_section_nr(start_pfn);
898	struct mem_section *ms;
899	struct page *memmap;
900	int ret;
901
902	ret = sparse_index_init(section_nr, nid);
903	if (ret < 0)
904		return ret;
905
906	memmap = section_activate(nid, start_pfn, nr_pages, altmap, pgmap);
907	if (IS_ERR(memmap))
908		return PTR_ERR(memmap);
909
910	/*
911	 * Poison uninitialized struct pages in order to catch invalid flags
912	 * combinations.
913	 */
914	if (!altmap || !altmap->inaccessible)
915		page_init_poison(memmap, sizeof(struct page) * nr_pages);
916
917	ms = __nr_to_section(section_nr);
918	set_section_nid(section_nr, nid);
919	__section_mark_present(ms, section_nr);
920
921	/* Align memmap to section boundary in the subsection case */
922	if (section_nr_to_pfn(section_nr) != start_pfn)
923		memmap = pfn_to_page(section_nr_to_pfn(section_nr));
924	sparse_init_one_section(ms, section_nr, memmap, ms->usage, 0);
925
926	return 0;
927}
928
929void sparse_remove_section(unsigned long pfn, unsigned long nr_pages,
930			   struct vmem_altmap *altmap)
931{
932	struct mem_section *ms = __pfn_to_section(pfn);
933
934	if (WARN_ON_ONCE(!valid_section(ms)))
 
 
 
 
 
 
935		return;
936
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
937	section_deactivate(pfn, nr_pages, altmap);
938}
939#endif /* CONFIG_MEMORY_HOTPLUG */