Linux Audio

Check our new training course

Loading...
v6.2
  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/*
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 = 1UL << (MAX_PHYSMEM_BITS-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 <= __highest_present_section_nr));	\
177	     section_nr = next_present_section_nr(section_nr))
178
179static inline unsigned long first_present_section_nr(void)
180{
181	return next_present_section_nr(-1);
182}
183
184#ifdef CONFIG_SPARSEMEM_VMEMMAP
185static void subsection_mask_set(unsigned long *map, unsigned long pfn,
186		unsigned long nr_pages)
187{
188	int idx = subsection_map_index(pfn);
189	int end = subsection_map_index(pfn + nr_pages - 1);
190
191	bitmap_set(map, idx, end - idx + 1);
192}
193
194void __init subsection_map_init(unsigned long pfn, unsigned long nr_pages)
195{
196	int end_sec = pfn_to_section_nr(pfn + nr_pages - 1);
197	unsigned long nr, start_sec = pfn_to_section_nr(pfn);
198
199	if (!nr_pages)
200		return;
201
202	for (nr = start_sec; nr <= end_sec; nr++) {
203		struct mem_section *ms;
204		unsigned long pfns;
205
206		pfns = min(nr_pages, PAGES_PER_SECTION
207				- (pfn & ~PAGE_SECTION_MASK));
208		ms = __nr_to_section(nr);
209		subsection_mask_set(ms->usage->subsection_map, pfn, pfns);
210
211		pr_debug("%s: sec: %lu pfns: %lu set(%d, %d)\n", __func__, nr,
212				pfns, subsection_map_index(pfn),
213				subsection_map_index(pfn + pfns - 1));
214
215		pfn += pfns;
216		nr_pages -= pfns;
217	}
218}
219#else
220void __init subsection_map_init(unsigned long pfn, unsigned long nr_pages)
221{
222}
223#endif
224
225/* Record a memory area against a node. */
226static void __init memory_present(int nid, unsigned long start, unsigned long end)
227{
228	unsigned long pfn;
229
230#ifdef CONFIG_SPARSEMEM_EXTREME
231	if (unlikely(!mem_section)) {
232		unsigned long size, align;
233
234		size = sizeof(struct mem_section *) * NR_SECTION_ROOTS;
235		align = 1 << (INTERNODE_CACHE_SHIFT);
236		mem_section = memblock_alloc(size, align);
237		if (!mem_section)
238			panic("%s: Failed to allocate %lu bytes align=0x%lx\n",
239			      __func__, size, align);
240	}
241#endif
242
243	start &= PAGE_SECTION_MASK;
244	mminit_validate_memmodel_limits(&start, &end);
245	for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) {
246		unsigned long section = pfn_to_section_nr(pfn);
247		struct mem_section *ms;
248
249		sparse_index_init(section, nid);
250		set_section_nid(section, nid);
251
252		ms = __nr_to_section(section);
253		if (!ms->section_mem_map) {
254			ms->section_mem_map = sparse_encode_early_nid(nid) |
255							SECTION_IS_ONLINE;
256			__section_mark_present(ms, section);
257		}
258	}
259}
260
261/*
262 * Mark all memblocks as present using memory_present().
263 * This is a convenience function that is useful to mark all of the systems
264 * memory as present during initialization.
265 */
266static void __init memblocks_present(void)
267{
268	unsigned long start, end;
269	int i, nid;
270
271	for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid)
272		memory_present(nid, start, end);
273}
274
275/*
276 * Subtle, we encode the real pfn into the mem_map such that
277 * the identity pfn - section_mem_map will return the actual
278 * physical page frame number.
279 */
280static unsigned long sparse_encode_mem_map(struct page *mem_map, unsigned long pnum)
281{
282	unsigned long coded_mem_map =
283		(unsigned long)(mem_map - (section_nr_to_pfn(pnum)));
284	BUILD_BUG_ON(SECTION_MAP_LAST_BIT > PFN_SECTION_SHIFT);
285	BUG_ON(coded_mem_map & ~SECTION_MAP_MASK);
286	return coded_mem_map;
287}
288
289#ifdef CONFIG_MEMORY_HOTPLUG
290/*
291 * Decode mem_map from the coded memmap
292 */
293struct page *sparse_decode_mem_map(unsigned long coded_mem_map, unsigned long pnum)
294{
295	/* mask off the extra low bits of information */
296	coded_mem_map &= SECTION_MAP_MASK;
297	return ((struct page *)coded_mem_map) + section_nr_to_pfn(pnum);
298}
299#endif /* CONFIG_MEMORY_HOTPLUG */
300
301static void __meminit sparse_init_one_section(struct mem_section *ms,
302		unsigned long pnum, struct page *mem_map,
303		struct mem_section_usage *usage, unsigned long flags)
304{
 
 
 
305	ms->section_mem_map &= ~SECTION_MAP_MASK;
306	ms->section_mem_map |= sparse_encode_mem_map(mem_map, pnum)
307		| SECTION_HAS_MEM_MAP | flags;
308	ms->usage = usage;
309}
310
311static unsigned long usemap_size(void)
312{
313	return BITS_TO_LONGS(SECTION_BLOCKFLAGS_BITS) * sizeof(unsigned long);
314}
315
316size_t mem_section_usage_size(void)
317{
318	return sizeof(struct mem_section_usage) + usemap_size();
319}
320
321static inline phys_addr_t pgdat_to_phys(struct pglist_data *pgdat)
 
322{
323#ifndef CONFIG_NUMA
324	VM_BUG_ON(pgdat != &contig_page_data);
325	return __pa_symbol(&contig_page_data);
326#else
327	return __pa(pgdat);
328#endif
329}
 
330
331#ifdef CONFIG_MEMORY_HOTREMOVE
332static struct mem_section_usage * __init
333sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
334					 unsigned long size)
335{
336	struct mem_section_usage *usage;
337	unsigned long goal, limit;
 
338	int nid;
339	/*
340	 * A page may contain usemaps for other sections preventing the
341	 * page being freed and making a section unremovable while
342	 * other sections referencing the usemap remain active. Similarly,
343	 * a pgdat can prevent a section being removed. If section A
344	 * contains a pgdat and section B contains the usemap, both
345	 * sections become inter-dependent. This allocates usemaps
346	 * from the same section as the pgdat where possible to avoid
347	 * this problem.
348	 */
349	goal = pgdat_to_phys(pgdat) & (PAGE_SECTION_MASK << PAGE_SHIFT);
350	limit = goal + (1UL << PA_SECTION_SHIFT);
351	nid = early_pfn_to_nid(goal >> PAGE_SHIFT);
352again:
353	usage = memblock_alloc_try_nid(size, SMP_CACHE_BYTES, goal, limit, nid);
354	if (!usage && limit) {
 
 
355		limit = 0;
356		goto again;
357	}
358	return usage;
359}
360
361static void __init check_usemap_section_nr(int nid,
362		struct mem_section_usage *usage)
363{
364	unsigned long usemap_snr, pgdat_snr;
365	static unsigned long old_usemap_snr;
366	static unsigned long old_pgdat_snr;
367	struct pglist_data *pgdat = NODE_DATA(nid);
368	int usemap_nid;
369
370	/* First call */
371	if (!old_usemap_snr) {
372		old_usemap_snr = NR_MEM_SECTIONS;
373		old_pgdat_snr = NR_MEM_SECTIONS;
374	}
375
376	usemap_snr = pfn_to_section_nr(__pa(usage) >> PAGE_SHIFT);
377	pgdat_snr = pfn_to_section_nr(pgdat_to_phys(pgdat) >> PAGE_SHIFT);
378	if (usemap_snr == pgdat_snr)
379		return;
380
381	if (old_usemap_snr == usemap_snr && old_pgdat_snr == pgdat_snr)
382		/* skip redundant message */
383		return;
384
385	old_usemap_snr = usemap_snr;
386	old_pgdat_snr = pgdat_snr;
387
388	usemap_nid = sparse_early_nid(__nr_to_section(usemap_snr));
389	if (usemap_nid != nid) {
390		pr_info("node %d must be removed before remove section %ld\n",
391			nid, usemap_snr);
392		return;
393	}
394	/*
395	 * There is a circular dependency.
396	 * Some platforms allow un-removable section because they will just
397	 * gather other removable sections for dynamic partitioning.
398	 * Just notify un-removable section's number here.
399	 */
400	pr_info("Section %ld and %ld (node %d) have a circular dependency on usemap and pgdat allocations\n",
401		usemap_snr, pgdat_snr, nid);
402}
403#else
404static struct mem_section_usage * __init
405sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
406					 unsigned long size)
407{
408	return memblock_alloc_node(size, SMP_CACHE_BYTES, pgdat->node_id);
409}
410
411static void __init check_usemap_section_nr(int nid,
412		struct mem_section_usage *usage)
413{
414}
415#endif /* CONFIG_MEMORY_HOTREMOVE */
416
417#ifdef CONFIG_SPARSEMEM_VMEMMAP
418static unsigned long __init section_map_size(void)
 
 
419{
420	return ALIGN(sizeof(struct page) * PAGES_PER_SECTION, PMD_SIZE);
421}
 
 
422
423#else
424static unsigned long __init section_map_size(void)
425{
426	return PAGE_ALIGN(sizeof(struct page) * PAGES_PER_SECTION);
 
 
 
 
 
 
 
 
 
 
427}
428
429struct page __init *__populate_section_memmap(unsigned long pfn,
430		unsigned long nr_pages, int nid, struct vmem_altmap *altmap,
431		struct dev_pagemap *pgmap)
432{
433	unsigned long size = section_map_size();
434	struct page *map = sparse_buffer_alloc(size);
435	phys_addr_t addr = __pa(MAX_DMA_ADDRESS);
436
437	if (map)
438		return map;
439
440	map = memmap_alloc(size, size, addr, nid, false);
441	if (!map)
442		panic("%s: Failed to allocate %lu bytes align=0x%lx nid=%d from=%pa\n",
443		      __func__, size, PAGE_SIZE, nid, &addr);
444
 
 
 
 
445	return map;
446}
447#endif /* !CONFIG_SPARSEMEM_VMEMMAP */
448
449static void *sparsemap_buf __meminitdata;
450static void *sparsemap_buf_end __meminitdata;
451
452static inline void __meminit sparse_buffer_free(unsigned long size)
453{
454	WARN_ON(!sparsemap_buf || size == 0);
455	memblock_free(sparsemap_buf, size);
456}
457
458static void __init sparse_buffer_init(unsigned long size, int nid)
459{
460	phys_addr_t addr = __pa(MAX_DMA_ADDRESS);
461	WARN_ON(sparsemap_buf);	/* forgot to call sparse_buffer_fini()? */
462	/*
463	 * Pre-allocated buffer is mainly used by __populate_section_memmap
464	 * and we want it to be properly aligned to the section size - this is
465	 * especially the case for VMEMMAP which maps memmap to PMDs
466	 */
467	sparsemap_buf = memmap_alloc(size, section_map_size(), addr, nid, true);
468	sparsemap_buf_end = sparsemap_buf + size;
469}
 
470
471static void __init sparse_buffer_fini(void)
472{
473	unsigned long size = sparsemap_buf_end - sparsemap_buf;
474
475	if (sparsemap_buf && size > 0)
476		sparse_buffer_free(size);
477	sparsemap_buf = NULL;
 
 
 
 
 
 
 
478}
 
479
480void * __meminit sparse_buffer_alloc(unsigned long size)
 
 
 
 
 
 
 
 
 
 
 
481{
482	void *ptr = NULL;
 
 
483
484	if (sparsemap_buf) {
485		ptr = (void *) roundup((unsigned long)sparsemap_buf, size);
486		if (ptr + size > sparsemap_buf_end)
487			ptr = NULL;
488		else {
489			/* Free redundant aligned space */
490			if ((unsigned long)(ptr - sparsemap_buf) > 0)
491				sparse_buffer_free((unsigned long)(ptr - sparsemap_buf));
492			sparsemap_buf = ptr + size;
493		}
494	}
495	return ptr;
496}
 
497
498void __weak __meminit vmemmap_populate_print_last(void)
499{
500}
501
502/*
503 * Initialize sparse on a specific node. The node spans [pnum_begin, pnum_end)
504 * And number of present sections in this node is map_count.
505 */
506static void __init sparse_init_nid(int nid, unsigned long pnum_begin,
507				   unsigned long pnum_end,
508				   unsigned long map_count)
509{
510	struct mem_section_usage *usage;
511	unsigned long pnum;
512	struct page *map;
 
 
513
514	usage = sparse_early_usemaps_alloc_pgdat_section(NODE_DATA(nid),
515			mem_section_usage_size() * map_count);
516	if (!usage) {
517		pr_err("%s: node[%d] usemap allocation failed", __func__, nid);
518		goto failed;
519	}
520	sparse_buffer_init(map_count * section_map_size(), nid);
521	for_each_present_section_nr(pnum_begin, pnum) {
522		unsigned long pfn = section_nr_to_pfn(pnum);
523
524		if (pnum >= pnum_end)
525			break;
526
527		map = __populate_section_memmap(pfn, PAGES_PER_SECTION,
528				nid, NULL, NULL);
529		if (!map) {
530			pr_err("%s: node[%d] memory map backing failed. Some memory will not be available.",
531			       __func__, nid);
532			pnum_begin = pnum;
533			sparse_buffer_fini();
534			goto failed;
535		}
536		check_usemap_section_nr(nid, usage);
537		sparse_init_one_section(__nr_to_section(pnum), pnum, map, usage,
538				SECTION_IS_EARLY);
539		usage = (void *) usage + mem_section_usage_size();
540	}
541	sparse_buffer_fini();
542	return;
543failed:
544	/* We failed to allocate, mark all the following pnums as not present */
545	for_each_present_section_nr(pnum_begin, pnum) {
546		struct mem_section *ms;
547
548		if (pnum >= pnum_end)
549			break;
550		ms = __nr_to_section(pnum);
551		ms->section_mem_map = 0;
 
 
552	}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
553}
554
555/*
556 * Allocate the accumulated non-linear sections, allocate a mem_map
557 * for each and record the physical to section mapping.
558 */
559void __init sparse_init(void)
560{
561	unsigned long pnum_end, pnum_begin, map_count = 1;
562	int nid_begin;
563
564	memblocks_present();
 
 
 
 
 
565
566	pnum_begin = first_present_section_nr();
567	nid_begin = sparse_early_nid(__nr_to_section(pnum_begin));
568
569	/* Setup pageblock_order for HUGETLB_PAGE_SIZE_VARIABLE */
570	set_pageblock_order();
571
572	for_each_present_section_nr(pnum_begin + 1, pnum_end) {
573		int nid = sparse_early_nid(__nr_to_section(pnum_end));
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
574
575		if (nid == nid_begin) {
576			map_count++;
 
577			continue;
578		}
579		/* Init node with sections in range [pnum_begin, pnum_end) */
580		sparse_init_nid(nid_begin, pnum_begin, pnum_end, map_count);
581		nid_begin = nid;
582		pnum_begin = pnum_end;
583		map_count = 1;
 
 
 
 
 
584	}
585	/* cover the last node */
586	sparse_init_nid(nid_begin, pnum_begin, pnum_end, map_count);
587	vmemmap_populate_print_last();
 
 
 
 
 
588}
589
590#ifdef CONFIG_MEMORY_HOTPLUG
591
592/* Mark all memory sections within the pfn range as online */
593void online_mem_sections(unsigned long start_pfn, unsigned long end_pfn)
594{
595	unsigned long pfn;
596
597	for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
598		unsigned long section_nr = pfn_to_section_nr(pfn);
599		struct mem_section *ms;
600
601		/* onlining code should never touch invalid ranges */
602		if (WARN_ON(!valid_section_nr(section_nr)))
603			continue;
604
605		ms = __nr_to_section(section_nr);
606		ms->section_mem_map |= SECTION_IS_ONLINE;
607	}
608}
609
610/* Mark all memory sections within the pfn range as offline */
 
611void offline_mem_sections(unsigned long start_pfn, unsigned long end_pfn)
612{
613	unsigned long pfn;
614
615	for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
616		unsigned long section_nr = pfn_to_section_nr(pfn);
617		struct mem_section *ms;
618
619		/*
620		 * TODO this needs some double checking. Offlining code makes
621		 * sure to check pfn_valid but those checks might be just bogus
622		 */
623		if (WARN_ON(!valid_section_nr(section_nr)))
624			continue;
625
626		ms = __nr_to_section(section_nr);
627		ms->section_mem_map &= ~SECTION_IS_ONLINE;
628	}
629}
 
630
631#ifdef CONFIG_SPARSEMEM_VMEMMAP
632static struct page * __meminit populate_section_memmap(unsigned long pfn,
633		unsigned long nr_pages, int nid, struct vmem_altmap *altmap,
634		struct dev_pagemap *pgmap)
635{
636	return __populate_section_memmap(pfn, nr_pages, nid, altmap, pgmap);
 
637}
638
639static void depopulate_section_memmap(unsigned long pfn, unsigned long nr_pages,
640		struct vmem_altmap *altmap)
641{
642	unsigned long start = (unsigned long) pfn_to_page(pfn);
643	unsigned long end = start + nr_pages * sizeof(struct page);
644
645	vmemmap_free(start, end, altmap);
646}
 
647static void free_map_bootmem(struct page *memmap)
648{
649	unsigned long start = (unsigned long)memmap;
650	unsigned long end = (unsigned long)(memmap + PAGES_PER_SECTION);
651
652	vmemmap_free(start, end, NULL);
653}
654
655static int clear_subsection_map(unsigned long pfn, unsigned long nr_pages)
656{
657	DECLARE_BITMAP(map, SUBSECTIONS_PER_SECTION) = { 0 };
658	DECLARE_BITMAP(tmp, SUBSECTIONS_PER_SECTION) = { 0 };
659	struct mem_section *ms = __pfn_to_section(pfn);
660	unsigned long *subsection_map = ms->usage
661		? &ms->usage->subsection_map[0] : NULL;
662
663	subsection_mask_set(map, pfn, nr_pages);
664	if (subsection_map)
665		bitmap_and(tmp, map, subsection_map, SUBSECTIONS_PER_SECTION);
666
667	if (WARN(!subsection_map || !bitmap_equal(tmp, map, SUBSECTIONS_PER_SECTION),
668				"section already deactivated (%#lx + %ld)\n",
669				pfn, nr_pages))
670		return -EINVAL;
671
672	bitmap_xor(subsection_map, map, subsection_map, SUBSECTIONS_PER_SECTION);
673	return 0;
674}
675
676static bool is_subsection_map_empty(struct mem_section *ms)
677{
678	return bitmap_empty(&ms->usage->subsection_map[0],
679			    SUBSECTIONS_PER_SECTION);
680}
681
682static int fill_subsection_map(unsigned long pfn, unsigned long nr_pages)
683{
684	struct mem_section *ms = __pfn_to_section(pfn);
685	DECLARE_BITMAP(map, SUBSECTIONS_PER_SECTION) = { 0 };
686	unsigned long *subsection_map;
687	int rc = 0;
688
689	subsection_mask_set(map, pfn, nr_pages);
 
 
690
691	subsection_map = &ms->usage->subsection_map[0];
 
 
692
693	if (bitmap_empty(map, SUBSECTIONS_PER_SECTION))
694		rc = -EINVAL;
695	else if (bitmap_intersects(map, subsection_map, SUBSECTIONS_PER_SECTION))
696		rc = -EEXIST;
697	else
698		bitmap_or(subsection_map, map, subsection_map,
699				SUBSECTIONS_PER_SECTION);
700
701	return rc;
702}
703#else
704struct page * __meminit populate_section_memmap(unsigned long pfn,
705		unsigned long nr_pages, int nid, struct vmem_altmap *altmap,
706		struct dev_pagemap *pgmap)
707{
708	return kvmalloc_node(array_size(sizeof(struct page),
709					PAGES_PER_SECTION), GFP_KERNEL, nid);
710}
711
712static void depopulate_section_memmap(unsigned long pfn, unsigned long nr_pages,
713		struct vmem_altmap *altmap)
714{
715	kvfree(pfn_to_page(pfn));
 
 
 
 
716}
717
 
718static void free_map_bootmem(struct page *memmap)
719{
720	unsigned long maps_section_nr, removing_section_nr, i;
721	unsigned long magic, nr_pages;
722	struct page *page = virt_to_page(memmap);
723
724	nr_pages = PAGE_ALIGN(PAGES_PER_SECTION * sizeof(struct page))
725		>> PAGE_SHIFT;
726
727	for (i = 0; i < nr_pages; i++, page++) {
728		magic = page->index;
729
730		BUG_ON(magic == NODE_INFO);
731
732		maps_section_nr = pfn_to_section_nr(page_to_pfn(page));
733		removing_section_nr = page_private(page);
734
735		/*
736		 * When this function is called, the removing section is
737		 * logical offlined state. This means all pages are isolated
738		 * from page allocator. If removing section's memmap is placed
739		 * on the same section, it must not be freed.
740		 * If it is freed, page allocator may allocate it which will
741		 * be removed physically soon.
742		 */
743		if (maps_section_nr != removing_section_nr)
744			put_page_bootmem(page);
745	}
746}
747
748static int clear_subsection_map(unsigned long pfn, unsigned long nr_pages)
749{
750	return 0;
751}
752
753static bool is_subsection_map_empty(struct mem_section *ms)
754{
755	return true;
756}
757
758static int fill_subsection_map(unsigned long pfn, unsigned long nr_pages)
759{
760	return 0;
761}
762#endif /* CONFIG_SPARSEMEM_VMEMMAP */
763
764/*
765 * To deactivate a memory region, there are 3 cases to handle across
766 * two configurations (SPARSEMEM_VMEMMAP={y,n}):
767 *
768 * 1. deactivation of a partial hot-added section (only possible in
769 *    the SPARSEMEM_VMEMMAP=y case).
770 *      a) section was present at memory init.
771 *      b) section was hot-added post memory init.
772 * 2. deactivation of a complete hot-added section.
773 * 3. deactivation of a complete section from memory init.
774 *
775 * For 1, when subsection_map does not empty we will not be freeing the
776 * usage map, but still need to free the vmemmap range.
777 *
778 * For 2 and 3, the SPARSEMEM_VMEMMAP={y,n} cases are unified
779 */
780static void section_deactivate(unsigned long pfn, unsigned long nr_pages,
781		struct vmem_altmap *altmap)
782{
783	struct mem_section *ms = __pfn_to_section(pfn);
784	bool section_is_early = early_section(ms);
785	struct page *memmap = NULL;
786	bool empty;
 
 
787
788	if (clear_subsection_map(pfn, nr_pages))
789		return;
 
 
 
 
 
 
 
 
 
 
 
 
 
790
791	empty = is_subsection_map_empty(ms);
792	if (empty) {
793		unsigned long section_nr = pfn_to_section_nr(pfn);
794
795		/*
796		 * When removing an early section, the usage map is kept (as the
797		 * usage maps of other sections fall into the same page). It
798		 * will be re-used when re-adding the section - which is then no
799		 * longer an early section. If the usage map is PageReserved, it
800		 * was allocated during boot.
801		 */
802		if (!PageReserved(virt_to_page(ms->usage))) {
803			kfree(ms->usage);
804			ms->usage = NULL;
805		}
806		memmap = sparse_decode_mem_map(ms->section_mem_map, section_nr);
807		/*
808		 * Mark the section invalid so that valid_section()
809		 * return false. This prevents code from dereferencing
810		 * ms->usage array.
811		 */
812		ms->section_mem_map &= ~SECTION_HAS_MEM_MAP;
813	}
814
 
815	/*
816	 * The memmap of early sections is always fully populated. See
817	 * section_activate() and pfn_valid() .
818	 */
819	if (!section_is_early)
820		depopulate_section_memmap(pfn, nr_pages, altmap);
821	else if (memmap)
822		free_map_bootmem(memmap);
823
824	if (empty)
825		ms->section_mem_map = (unsigned long)NULL;
826}
827
828static struct page * __meminit section_activate(int nid, unsigned long pfn,
829		unsigned long nr_pages, struct vmem_altmap *altmap,
830		struct dev_pagemap *pgmap)
831{
832	struct mem_section *ms = __pfn_to_section(pfn);
833	struct mem_section_usage *usage = NULL;
834	struct page *memmap;
835	int rc = 0;
836
837	if (!ms->usage) {
838		usage = kzalloc(mem_section_usage_size(), GFP_KERNEL);
839		if (!usage)
840			return ERR_PTR(-ENOMEM);
841		ms->usage = usage;
842	}
 
 
843
844	rc = fill_subsection_map(pfn, nr_pages);
845	if (rc) {
846		if (usage)
847			ms->usage = NULL;
848		kfree(usage);
849		return ERR_PTR(rc);
850	}
851
852	/*
853	 * The early init code does not consider partially populated
854	 * initial sections, it simply assumes that memory will never be
855	 * referenced.  If we hot-add memory into such a section then we
856	 * do not need to populate the memmap and can simply reuse what
857	 * is already there.
858	 */
859	if (nr_pages < PAGES_PER_SECTION && early_section(ms))
860		return pfn_to_page(pfn);
861
862	memmap = populate_section_memmap(pfn, nr_pages, nid, altmap, pgmap);
863	if (!memmap) {
864		section_deactivate(pfn, nr_pages, altmap);
865		return ERR_PTR(-ENOMEM);
 
866	}
867
868	return memmap;
869}
870
871/**
872 * sparse_add_section - add a memory section, or populate an existing one
873 * @nid: The node to add section on
874 * @start_pfn: start pfn of the memory range
875 * @nr_pages: number of pfns to add in the section
876 * @altmap: alternate pfns to allocate the memmap backing store
877 * @pgmap: alternate compound page geometry for devmap mappings
878 *
879 * This is only intended for hotplug.
880 *
881 * Note that only VMEMMAP supports sub-section aligned hotplug,
882 * the proper alignment and size are gated by check_pfn_span().
883 *
884 *
885 * Return:
886 * * 0		- On success.
887 * * -EEXIST	- Section has been present.
888 * * -ENOMEM	- Out of memory.
889 */
890int __meminit sparse_add_section(int nid, unsigned long start_pfn,
891		unsigned long nr_pages, struct vmem_altmap *altmap,
892		struct dev_pagemap *pgmap)
893{
894	unsigned long section_nr = pfn_to_section_nr(start_pfn);
895	struct mem_section *ms;
896	struct page *memmap;
897	int ret;
898
899	ret = sparse_index_init(section_nr, nid);
900	if (ret < 0)
901		return ret;
 
902
903	memmap = section_activate(nid, start_pfn, nr_pages, altmap, pgmap);
904	if (IS_ERR(memmap))
905		return PTR_ERR(memmap);
906
 
907	/*
908	 * Poison uninitialized struct pages in order to catch invalid flags
909	 * combinations.
910	 */
911	page_init_poison(memmap, sizeof(struct page) * nr_pages);
 
 
 
 
 
912
913	ms = __nr_to_section(section_nr);
914	set_section_nid(section_nr, nid);
915	__section_mark_present(ms, section_nr);
916
917	/* Align memmap to section boundary in the subsection case */
918	if (section_nr_to_pfn(section_nr) != start_pfn)
919		memmap = pfn_to_page(section_nr_to_pfn(section_nr));
920	sparse_init_one_section(ms, section_nr, memmap, ms->usage, 0);
921
922	return 0;
 
923}
924
925void sparse_remove_section(struct mem_section *ms, unsigned long pfn,
926		unsigned long nr_pages, unsigned long map_offset,
927		struct vmem_altmap *altmap)
928{
929	section_deactivate(pfn, nr_pages, altmap);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
930}
 
931#endif /* CONFIG_MEMORY_HOTPLUG */
v4.17
  1// SPDX-License-Identifier: GPL-2.0
  2/*
  3 * sparse memory mappings.
  4 */
  5#include <linux/mm.h>
  6#include <linux/slab.h>
  7#include <linux/mmzone.h>
  8#include <linux/bootmem.h>
  9#include <linux/compiler.h>
 10#include <linux/highmem.h>
 11#include <linux/export.h>
 12#include <linux/spinlock.h>
 13#include <linux/vmalloc.h>
 
 
 
 14
 15#include "internal.h"
 16#include <asm/dma.h>
 17#include <asm/pgalloc.h>
 18#include <asm/pgtable.h>
 19
 20/*
 21 * Permanent SPARSEMEM data:
 22 *
 23 * 1) mem_section	- memory sections, mem_map's for valid memory
 24 */
 25#ifdef CONFIG_SPARSEMEM_EXTREME
 26struct mem_section **mem_section;
 27#else
 28struct mem_section mem_section[NR_SECTION_ROOTS][SECTIONS_PER_ROOT]
 29	____cacheline_internodealigned_in_smp;
 30#endif
 31EXPORT_SYMBOL(mem_section);
 32
 33#ifdef NODE_NOT_IN_PAGE_FLAGS
 34/*
 35 * If we did not store the node number in the page then we have to
 36 * do a lookup in the section_to_node_table in order to find which
 37 * node the page belongs to.
 38 */
 39#if MAX_NUMNODES <= 256
 40static u8 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned;
 41#else
 42static u16 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned;
 43#endif
 44
 45int page_to_nid(const struct page *page)
 46{
 47	return section_to_node_table[page_to_section(page)];
 48}
 49EXPORT_SYMBOL(page_to_nid);
 50
 51static void set_section_nid(unsigned long section_nr, int nid)
 52{
 53	section_to_node_table[section_nr] = nid;
 54}
 55#else /* !NODE_NOT_IN_PAGE_FLAGS */
 56static inline void set_section_nid(unsigned long section_nr, int nid)
 57{
 58}
 59#endif
 60
 61#ifdef CONFIG_SPARSEMEM_EXTREME
 62static noinline struct mem_section __ref *sparse_index_alloc(int nid)
 63{
 64	struct mem_section *section = NULL;
 65	unsigned long array_size = SECTIONS_PER_ROOT *
 66				   sizeof(struct mem_section);
 67
 68	if (slab_is_available())
 69		section = kzalloc_node(array_size, GFP_KERNEL, nid);
 70	else
 71		section = memblock_virt_alloc_node(array_size, nid);
 
 
 
 
 
 72
 73	return section;
 74}
 75
 76static int __meminit sparse_index_init(unsigned long section_nr, int nid)
 77{
 78	unsigned long root = SECTION_NR_TO_ROOT(section_nr);
 79	struct mem_section *section;
 80
 
 
 
 
 
 
 
 81	if (mem_section[root])
 82		return -EEXIST;
 83
 84	section = sparse_index_alloc(nid);
 85	if (!section)
 86		return -ENOMEM;
 87
 88	mem_section[root] = section;
 89
 90	return 0;
 91}
 92#else /* !SPARSEMEM_EXTREME */
 93static inline int sparse_index_init(unsigned long section_nr, int nid)
 94{
 95	return 0;
 96}
 97#endif
 98
 99#ifdef CONFIG_SPARSEMEM_EXTREME
100int __section_nr(struct mem_section* ms)
101{
102	unsigned long root_nr;
103	struct mem_section *root = NULL;
104
105	for (root_nr = 0; root_nr < NR_SECTION_ROOTS; root_nr++) {
106		root = __nr_to_section(root_nr * SECTIONS_PER_ROOT);
107		if (!root)
108			continue;
109
110		if ((ms >= root) && (ms < (root + SECTIONS_PER_ROOT)))
111		     break;
112	}
113
114	VM_BUG_ON(!root);
115
116	return (root_nr * SECTIONS_PER_ROOT) + (ms - root);
117}
118#else
119int __section_nr(struct mem_section* ms)
120{
121	return (int)(ms - mem_section[0]);
122}
123#endif
124
125/*
126 * During early boot, before section_mem_map is used for an actual
127 * mem_map, we use section_mem_map to store the section's NUMA
128 * node.  This keeps us from having to use another data structure.  The
129 * node information is cleared just before we store the real mem_map.
130 */
131static inline unsigned long sparse_encode_early_nid(int nid)
132{
133	return (nid << SECTION_NID_SHIFT);
134}
135
136static inline int sparse_early_nid(struct mem_section *section)
137{
138	return (section->section_mem_map >> SECTION_NID_SHIFT);
139}
140
141/* Validate the physical addressing limitations of the model */
142void __meminit mminit_validate_memmodel_limits(unsigned long *start_pfn,
143						unsigned long *end_pfn)
144{
145	unsigned long max_sparsemem_pfn = 1UL << (MAX_PHYSMEM_BITS-PAGE_SHIFT);
146
147	/*
148	 * Sanity checks - do not allow an architecture to pass
149	 * in larger pfns than the maximum scope of sparsemem:
150	 */
151	if (*start_pfn > max_sparsemem_pfn) {
152		mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
153			"Start of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
154			*start_pfn, *end_pfn, max_sparsemem_pfn);
155		WARN_ON_ONCE(1);
156		*start_pfn = max_sparsemem_pfn;
157		*end_pfn = max_sparsemem_pfn;
158	} else if (*end_pfn > max_sparsemem_pfn) {
159		mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
160			"End of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
161			*start_pfn, *end_pfn, max_sparsemem_pfn);
162		WARN_ON_ONCE(1);
163		*end_pfn = max_sparsemem_pfn;
164	}
165}
166
167/*
168 * There are a number of times that we loop over NR_MEM_SECTIONS,
169 * looking for section_present() on each.  But, when we have very
170 * large physical address spaces, NR_MEM_SECTIONS can also be
171 * very large which makes the loops quite long.
172 *
173 * Keeping track of this gives us an easy way to break out of
174 * those loops early.
175 */
176int __highest_present_section_nr;
177static void section_mark_present(struct mem_section *ms)
 
178{
179	int section_nr = __section_nr(ms);
180
181	if (section_nr > __highest_present_section_nr)
182		__highest_present_section_nr = section_nr;
183
184	ms->section_mem_map |= SECTION_MARKED_PRESENT;
185}
186
187static inline int next_present_section_nr(int section_nr)
188{
189	do {
190		section_nr++;
191		if (present_section_nr(section_nr))
192			return section_nr;
193	} while ((section_nr < NR_MEM_SECTIONS) &&
194		 (section_nr <= __highest_present_section_nr));
195
196	return -1;
197}
198#define for_each_present_section_nr(start, section_nr)		\
199	for (section_nr = next_present_section_nr(start-1);	\
200	     ((section_nr >= 0) &&				\
201	      (section_nr < NR_MEM_SECTIONS) &&			\
202	      (section_nr <= __highest_present_section_nr));	\
203	     section_nr = next_present_section_nr(section_nr))
204
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
205/* Record a memory area against a node. */
206void __init memory_present(int nid, unsigned long start, unsigned long end)
207{
208	unsigned long pfn;
209
210#ifdef CONFIG_SPARSEMEM_EXTREME
211	if (unlikely(!mem_section)) {
212		unsigned long size, align;
213
214		size = sizeof(struct mem_section*) * NR_SECTION_ROOTS;
215		align = 1 << (INTERNODE_CACHE_SHIFT);
216		mem_section = memblock_virt_alloc(size, align);
 
 
 
217	}
218#endif
219
220	start &= PAGE_SECTION_MASK;
221	mminit_validate_memmodel_limits(&start, &end);
222	for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) {
223		unsigned long section = pfn_to_section_nr(pfn);
224		struct mem_section *ms;
225
226		sparse_index_init(section, nid);
227		set_section_nid(section, nid);
228
229		ms = __nr_to_section(section);
230		if (!ms->section_mem_map) {
231			ms->section_mem_map = sparse_encode_early_nid(nid) |
232							SECTION_IS_ONLINE;
233			section_mark_present(ms);
234		}
235	}
236}
237
238/*
 
 
 
 
 
 
 
 
 
 
 
 
 
 
239 * Subtle, we encode the real pfn into the mem_map such that
240 * the identity pfn - section_mem_map will return the actual
241 * physical page frame number.
242 */
243static unsigned long sparse_encode_mem_map(struct page *mem_map, unsigned long pnum)
244{
245	unsigned long coded_mem_map =
246		(unsigned long)(mem_map - (section_nr_to_pfn(pnum)));
247	BUILD_BUG_ON(SECTION_MAP_LAST_BIT > (1UL<<PFN_SECTION_SHIFT));
248	BUG_ON(coded_mem_map & ~SECTION_MAP_MASK);
249	return coded_mem_map;
250}
251
 
252/*
253 * Decode mem_map from the coded memmap
254 */
255struct page *sparse_decode_mem_map(unsigned long coded_mem_map, unsigned long pnum)
256{
257	/* mask off the extra low bits of information */
258	coded_mem_map &= SECTION_MAP_MASK;
259	return ((struct page *)coded_mem_map) + section_nr_to_pfn(pnum);
260}
 
261
262static int __meminit sparse_init_one_section(struct mem_section *ms,
263		unsigned long pnum, struct page *mem_map,
264		unsigned long *pageblock_bitmap)
265{
266	if (!present_section(ms))
267		return -EINVAL;
268
269	ms->section_mem_map &= ~SECTION_MAP_MASK;
270	ms->section_mem_map |= sparse_encode_mem_map(mem_map, pnum) |
271							SECTION_HAS_MEM_MAP;
272 	ms->pageblock_flags = pageblock_bitmap;
 
273
274	return 1;
 
 
275}
276
277unsigned long usemap_size(void)
278{
279	return BITS_TO_LONGS(SECTION_BLOCKFLAGS_BITS) * sizeof(unsigned long);
280}
281
282#ifdef CONFIG_MEMORY_HOTPLUG
283static unsigned long *__kmalloc_section_usemap(void)
284{
285	return kmalloc(usemap_size(), GFP_KERNEL);
 
 
 
 
 
286}
287#endif /* CONFIG_MEMORY_HOTPLUG */
288
289#ifdef CONFIG_MEMORY_HOTREMOVE
290static unsigned long * __init
291sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
292					 unsigned long size)
293{
 
294	unsigned long goal, limit;
295	unsigned long *p;
296	int nid;
297	/*
298	 * A page may contain usemaps for other sections preventing the
299	 * page being freed and making a section unremovable while
300	 * other sections referencing the usemap remain active. Similarly,
301	 * a pgdat can prevent a section being removed. If section A
302	 * contains a pgdat and section B contains the usemap, both
303	 * sections become inter-dependent. This allocates usemaps
304	 * from the same section as the pgdat where possible to avoid
305	 * this problem.
306	 */
307	goal = __pa(pgdat) & (PAGE_SECTION_MASK << PAGE_SHIFT);
308	limit = goal + (1UL << PA_SECTION_SHIFT);
309	nid = early_pfn_to_nid(goal >> PAGE_SHIFT);
310again:
311	p = memblock_virt_alloc_try_nid_nopanic(size,
312						SMP_CACHE_BYTES, goal, limit,
313						nid);
314	if (!p && limit) {
315		limit = 0;
316		goto again;
317	}
318	return p;
319}
320
321static void __init check_usemap_section_nr(int nid, unsigned long *usemap)
 
322{
323	unsigned long usemap_snr, pgdat_snr;
324	static unsigned long old_usemap_snr;
325	static unsigned long old_pgdat_snr;
326	struct pglist_data *pgdat = NODE_DATA(nid);
327	int usemap_nid;
328
329	/* First call */
330	if (!old_usemap_snr) {
331		old_usemap_snr = NR_MEM_SECTIONS;
332		old_pgdat_snr = NR_MEM_SECTIONS;
333	}
334
335	usemap_snr = pfn_to_section_nr(__pa(usemap) >> PAGE_SHIFT);
336	pgdat_snr = pfn_to_section_nr(__pa(pgdat) >> PAGE_SHIFT);
337	if (usemap_snr == pgdat_snr)
338		return;
339
340	if (old_usemap_snr == usemap_snr && old_pgdat_snr == pgdat_snr)
341		/* skip redundant message */
342		return;
343
344	old_usemap_snr = usemap_snr;
345	old_pgdat_snr = pgdat_snr;
346
347	usemap_nid = sparse_early_nid(__nr_to_section(usemap_snr));
348	if (usemap_nid != nid) {
349		pr_info("node %d must be removed before remove section %ld\n",
350			nid, usemap_snr);
351		return;
352	}
353	/*
354	 * There is a circular dependency.
355	 * Some platforms allow un-removable section because they will just
356	 * gather other removable sections for dynamic partitioning.
357	 * Just notify un-removable section's number here.
358	 */
359	pr_info("Section %ld and %ld (node %d) have a circular dependency on usemap and pgdat allocations\n",
360		usemap_snr, pgdat_snr, nid);
361}
362#else
363static unsigned long * __init
364sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
365					 unsigned long size)
366{
367	return memblock_virt_alloc_node_nopanic(size, pgdat->node_id);
368}
369
370static void __init check_usemap_section_nr(int nid, unsigned long *usemap)
 
371{
372}
373#endif /* CONFIG_MEMORY_HOTREMOVE */
374
375static void __init sparse_early_usemaps_alloc_node(void *data,
376				 unsigned long pnum_begin,
377				 unsigned long pnum_end,
378				 unsigned long usemap_count, int nodeid)
379{
380	void *usemap;
381	unsigned long pnum;
382	unsigned long **usemap_map = (unsigned long **)data;
383	int size = usemap_size();
384
385	usemap = sparse_early_usemaps_alloc_pgdat_section(NODE_DATA(nodeid),
386							  size * usemap_count);
387	if (!usemap) {
388		pr_warn("%s: allocation failed\n", __func__);
389		return;
390	}
391
392	for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
393		if (!present_section_nr(pnum))
394			continue;
395		usemap_map[pnum] = usemap;
396		usemap += size;
397		check_usemap_section_nr(nodeid, usemap_map[pnum]);
398	}
399}
400
401#ifndef CONFIG_SPARSEMEM_VMEMMAP
402struct page __init *sparse_mem_map_populate(unsigned long pnum, int nid,
403		struct vmem_altmap *altmap)
404{
405	struct page *map;
406	unsigned long size;
 
 
 
 
 
 
 
 
 
407
408	size = PAGE_ALIGN(sizeof(struct page) * PAGES_PER_SECTION);
409	map = memblock_virt_alloc_try_nid(size,
410					  PAGE_SIZE, __pa(MAX_DMA_ADDRESS),
411					  BOOTMEM_ALLOC_ACCESSIBLE, nid);
412	return map;
413}
414void __init sparse_mem_maps_populate_node(struct page **map_map,
415					  unsigned long pnum_begin,
416					  unsigned long pnum_end,
417					  unsigned long map_count, int nodeid)
 
 
418{
419	void *map;
420	unsigned long pnum;
421	unsigned long size = sizeof(struct page) * PAGES_PER_SECTION;
422
423	size = PAGE_ALIGN(size);
424	map = memblock_virt_alloc_try_nid_raw(size * map_count,
425					      PAGE_SIZE, __pa(MAX_DMA_ADDRESS),
426					      BOOTMEM_ALLOC_ACCESSIBLE, nodeid);
427	if (map) {
428		for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
429			if (!present_section_nr(pnum))
430				continue;
431			map_map[pnum] = map;
432			map += size;
433		}
434		return;
435	}
436
437	/* fallback */
438	for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
439		struct mem_section *ms;
440
441		if (!present_section_nr(pnum))
442			continue;
443		map_map[pnum] = sparse_mem_map_populate(pnum, nodeid, NULL);
444		if (map_map[pnum])
445			continue;
446		ms = __nr_to_section(pnum);
447		pr_err("%s: sparsemem memory map backing failed some memory will not be available\n",
448		       __func__);
449		ms->section_mem_map = 0;
450	}
451}
452#endif /* !CONFIG_SPARSEMEM_VMEMMAP */
453
454#ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
455static void __init sparse_early_mem_maps_alloc_node(void *data,
456				 unsigned long pnum_begin,
457				 unsigned long pnum_end,
458				 unsigned long map_count, int nodeid)
459{
460	struct page **map_map = (struct page **)data;
461	sparse_mem_maps_populate_node(map_map, pnum_begin, pnum_end,
462					 map_count, nodeid);
463}
464#else
465static struct page __init *sparse_early_mem_map_alloc(unsigned long pnum)
466{
467	struct page *map;
468	struct mem_section *ms = __nr_to_section(pnum);
469	int nid = sparse_early_nid(ms);
470
471	map = sparse_mem_map_populate(pnum, nid, NULL);
472	if (map)
473		return map;
474
475	pr_err("%s: sparsemem memory map backing failed some memory will not be available\n",
476	       __func__);
477	ms->section_mem_map = 0;
478	return NULL;
 
 
 
 
479}
480#endif
481
482void __weak __meminit vmemmap_populate_print_last(void)
483{
484}
485
486/**
487 *  alloc_usemap_and_memmap - memory alloction for pageblock flags and vmemmap
488 *  @map: usemap_map for pageblock flags or mmap_map for vmemmap
489 */
490static void __init alloc_usemap_and_memmap(void (*alloc_func)
491					(void *, unsigned long, unsigned long,
492					unsigned long, int), void *data)
493{
 
494	unsigned long pnum;
495	unsigned long map_count;
496	int nodeid_begin = 0;
497	unsigned long pnum_begin = 0;
498
499	for_each_present_section_nr(0, pnum) {
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
500		struct mem_section *ms;
501
 
 
502		ms = __nr_to_section(pnum);
503		nodeid_begin = sparse_early_nid(ms);
504		pnum_begin = pnum;
505		break;
506	}
507	map_count = 1;
508	for_each_present_section_nr(pnum_begin + 1, pnum) {
509		struct mem_section *ms;
510		int nodeid;
511
512		ms = __nr_to_section(pnum);
513		nodeid = sparse_early_nid(ms);
514		if (nodeid == nodeid_begin) {
515			map_count++;
516			continue;
517		}
518		/* ok, we need to take cake of from pnum_begin to pnum - 1*/
519		alloc_func(data, pnum_begin, pnum,
520						map_count, nodeid_begin);
521		/* new start, update count etc*/
522		nodeid_begin = nodeid;
523		pnum_begin = pnum;
524		map_count = 1;
525	}
526	/* ok, last chunk */
527	alloc_func(data, pnum_begin, NR_MEM_SECTIONS,
528						map_count, nodeid_begin);
529}
530
531/*
532 * Allocate the accumulated non-linear sections, allocate a mem_map
533 * for each and record the physical to section mapping.
534 */
535void __init sparse_init(void)
536{
537	unsigned long pnum;
538	struct page *map;
539	unsigned long *usemap;
540	unsigned long **usemap_map;
541	int size;
542#ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
543	int size2;
544	struct page **map_map;
545#endif
546
547	/* see include/linux/mmzone.h 'struct mem_section' definition */
548	BUILD_BUG_ON(!is_power_of_2(sizeof(struct mem_section)));
549
550	/* Setup pageblock_order for HUGETLB_PAGE_SIZE_VARIABLE */
551	set_pageblock_order();
552
553	/*
554	 * map is using big page (aka 2M in x86 64 bit)
555	 * usemap is less one page (aka 24 bytes)
556	 * so alloc 2M (with 2M align) and 24 bytes in turn will
557	 * make next 2M slip to one more 2M later.
558	 * then in big system, the memory will have a lot of holes...
559	 * here try to allocate 2M pages continuously.
560	 *
561	 * powerpc need to call sparse_init_one_section right after each
562	 * sparse_early_mem_map_alloc, so allocate usemap_map at first.
563	 */
564	size = sizeof(unsigned long *) * NR_MEM_SECTIONS;
565	usemap_map = memblock_virt_alloc(size, 0);
566	if (!usemap_map)
567		panic("can not allocate usemap_map\n");
568	alloc_usemap_and_memmap(sparse_early_usemaps_alloc_node,
569							(void *)usemap_map);
570
571#ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
572	size2 = sizeof(struct page *) * NR_MEM_SECTIONS;
573	map_map = memblock_virt_alloc(size2, 0);
574	if (!map_map)
575		panic("can not allocate map_map\n");
576	alloc_usemap_and_memmap(sparse_early_mem_maps_alloc_node,
577							(void *)map_map);
578#endif
579
580	for_each_present_section_nr(0, pnum) {
581		usemap = usemap_map[pnum];
582		if (!usemap)
583			continue;
584
585#ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
586		map = map_map[pnum];
587#else
588		map = sparse_early_mem_map_alloc(pnum);
589#endif
590		if (!map)
591			continue;
592
593		sparse_init_one_section(__nr_to_section(pnum), pnum, map,
594								usemap);
595	}
596
 
597	vmemmap_populate_print_last();
598
599#ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
600	memblock_free_early(__pa(map_map), size2);
601#endif
602	memblock_free_early(__pa(usemap_map), size);
603}
604
605#ifdef CONFIG_MEMORY_HOTPLUG
606
607/* Mark all memory sections within the pfn range as online */
608void online_mem_sections(unsigned long start_pfn, unsigned long end_pfn)
609{
610	unsigned long pfn;
611
612	for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
613		unsigned long section_nr = pfn_to_section_nr(pfn);
614		struct mem_section *ms;
615
616		/* onlining code should never touch invalid ranges */
617		if (WARN_ON(!valid_section_nr(section_nr)))
618			continue;
619
620		ms = __nr_to_section(section_nr);
621		ms->section_mem_map |= SECTION_IS_ONLINE;
622	}
623}
624
625#ifdef CONFIG_MEMORY_HOTREMOVE
626/* Mark all memory sections within the pfn range as online */
627void offline_mem_sections(unsigned long start_pfn, unsigned long end_pfn)
628{
629	unsigned long pfn;
630
631	for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
632		unsigned long section_nr = pfn_to_section_nr(pfn);
633		struct mem_section *ms;
634
635		/*
636		 * TODO this needs some double checking. Offlining code makes
637		 * sure to check pfn_valid but those checks might be just bogus
638		 */
639		if (WARN_ON(!valid_section_nr(section_nr)))
640			continue;
641
642		ms = __nr_to_section(section_nr);
643		ms->section_mem_map &= ~SECTION_IS_ONLINE;
644	}
645}
646#endif
647
648#ifdef CONFIG_SPARSEMEM_VMEMMAP
649static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid,
650		struct vmem_altmap *altmap)
 
651{
652	/* This will make the necessary allocations eventually. */
653	return sparse_mem_map_populate(pnum, nid, altmap);
654}
655static void __kfree_section_memmap(struct page *memmap,
 
656		struct vmem_altmap *altmap)
657{
658	unsigned long start = (unsigned long)memmap;
659	unsigned long end = (unsigned long)(memmap + PAGES_PER_SECTION);
660
661	vmemmap_free(start, end, altmap);
662}
663#ifdef CONFIG_MEMORY_HOTREMOVE
664static void free_map_bootmem(struct page *memmap)
665{
666	unsigned long start = (unsigned long)memmap;
667	unsigned long end = (unsigned long)(memmap + PAGES_PER_SECTION);
668
669	vmemmap_free(start, end, NULL);
670}
671#endif /* CONFIG_MEMORY_HOTREMOVE */
672#else
673static struct page *__kmalloc_section_memmap(void)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
674{
675	struct page *page, *ret;
676	unsigned long memmap_size = sizeof(struct page) * PAGES_PER_SECTION;
 
 
677
678	page = alloc_pages(GFP_KERNEL|__GFP_NOWARN, get_order(memmap_size));
679	if (page)
680		goto got_map_page;
681
682	ret = vmalloc(memmap_size);
683	if (ret)
684		goto got_map_ptr;
685
686	return NULL;
687got_map_page:
688	ret = (struct page *)pfn_to_kaddr(page_to_pfn(page));
689got_map_ptr:
 
 
 
690
691	return ret;
692}
693
694static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid,
695		struct vmem_altmap *altmap)
 
696{
697	return __kmalloc_section_memmap();
 
698}
699
700static void __kfree_section_memmap(struct page *memmap,
701		struct vmem_altmap *altmap)
702{
703	if (is_vmalloc_addr(memmap))
704		vfree(memmap);
705	else
706		free_pages((unsigned long)memmap,
707			   get_order(sizeof(struct page) * PAGES_PER_SECTION));
708}
709
710#ifdef CONFIG_MEMORY_HOTREMOVE
711static void free_map_bootmem(struct page *memmap)
712{
713	unsigned long maps_section_nr, removing_section_nr, i;
714	unsigned long magic, nr_pages;
715	struct page *page = virt_to_page(memmap);
716
717	nr_pages = PAGE_ALIGN(PAGES_PER_SECTION * sizeof(struct page))
718		>> PAGE_SHIFT;
719
720	for (i = 0; i < nr_pages; i++, page++) {
721		magic = (unsigned long) page->freelist;
722
723		BUG_ON(magic == NODE_INFO);
724
725		maps_section_nr = pfn_to_section_nr(page_to_pfn(page));
726		removing_section_nr = page_private(page);
727
728		/*
729		 * When this function is called, the removing section is
730		 * logical offlined state. This means all pages are isolated
731		 * from page allocator. If removing section's memmap is placed
732		 * on the same section, it must not be freed.
733		 * If it is freed, page allocator may allocate it which will
734		 * be removed physically soon.
735		 */
736		if (maps_section_nr != removing_section_nr)
737			put_page_bootmem(page);
738	}
739}
740#endif /* CONFIG_MEMORY_HOTREMOVE */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
741#endif /* CONFIG_SPARSEMEM_VMEMMAP */
742
743/*
744 * returns the number of sections whose mem_maps were properly
745 * set.  If this is <=0, then that means that the passed-in
746 * map was not consumed and must be freed.
 
 
 
 
 
 
 
 
 
 
 
747 */
748int __meminit sparse_add_one_section(struct pglist_data *pgdat,
749		unsigned long start_pfn, struct vmem_altmap *altmap)
750{
751	unsigned long section_nr = pfn_to_section_nr(start_pfn);
752	struct mem_section *ms;
753	struct page *memmap;
754	unsigned long *usemap;
755	unsigned long flags;
756	int ret;
757
758	/*
759	 * no locking for this, because it does its own
760	 * plus, it does a kmalloc
761	 */
762	ret = sparse_index_init(section_nr, pgdat->node_id);
763	if (ret < 0 && ret != -EEXIST)
764		return ret;
765	memmap = kmalloc_section_memmap(section_nr, pgdat->node_id, altmap);
766	if (!memmap)
767		return -ENOMEM;
768	usemap = __kmalloc_section_usemap();
769	if (!usemap) {
770		__kfree_section_memmap(memmap, altmap);
771		return -ENOMEM;
772	}
773
774	pgdat_resize_lock(pgdat, &flags);
 
 
775
776	ms = __pfn_to_section(start_pfn);
777	if (ms->section_mem_map & SECTION_MARKED_PRESENT) {
778		ret = -EEXIST;
779		goto out;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
780	}
781
782#ifdef CONFIG_DEBUG_VM
783	/*
784	 * Poison uninitialized struct pages in order to catch invalid flags
785	 * combinations.
786	 */
787	memset(memmap, PAGE_POISON_PATTERN, sizeof(struct page) * PAGES_PER_SECTION);
788#endif
 
 
789
790	section_mark_present(ms);
 
 
791
792	ret = sparse_init_one_section(ms, section_nr, memmap, usemap);
 
 
 
 
 
 
 
793
794out:
795	pgdat_resize_unlock(pgdat, &flags);
796	if (ret <= 0) {
797		kfree(usemap);
798		__kfree_section_memmap(memmap, altmap);
799	}
800	return ret;
801}
802
803#ifdef CONFIG_MEMORY_HOTREMOVE
804#ifdef CONFIG_MEMORY_FAILURE
805static void clear_hwpoisoned_pages(struct page *memmap, int nr_pages)
806{
807	int i;
 
 
808
809	if (!memmap)
810		return;
 
 
 
 
 
 
 
811
812	for (i = 0; i < nr_pages; i++) {
813		if (PageHWPoison(&memmap[i])) {
814			atomic_long_sub(1, &num_poisoned_pages);
815			ClearPageHWPoison(&memmap[i]);
816		}
817	}
 
 
818}
819#else
820static inline void clear_hwpoisoned_pages(struct page *memmap, int nr_pages)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
821{
822}
823#endif
 
 
824
825static void free_section_usemap(struct page *memmap, unsigned long *usemap,
826		struct vmem_altmap *altmap)
827{
828	struct page *usemap_page;
829
830	if (!usemap)
831		return;
 
832
833	usemap_page = virt_to_page(usemap);
834	/*
835	 * Check to see if allocation came from hot-plug-add
 
836	 */
837	if (PageSlab(usemap_page) || PageCompound(usemap_page)) {
838		kfree(usemap);
839		if (memmap)
840			__kfree_section_memmap(memmap, altmap);
841		return;
842	}
843
844	/*
845	 * The usemap came from bootmem. This is packed with other usemaps
846	 * on the section which has pgdat at boot time. Just keep it as is now.
847	 */
 
 
 
 
848
849	if (memmap)
850		free_map_bootmem(memmap);
851}
852
853void sparse_remove_one_section(struct zone *zone, struct mem_section *ms,
854		unsigned long map_offset, struct vmem_altmap *altmap)
 
855{
856	struct page *memmap = NULL;
857	unsigned long *usemap = NULL, flags;
858	struct pglist_data *pgdat = zone->zone_pgdat;
859
860	pgdat_resize_lock(pgdat, &flags);
861	if (ms->section_mem_map) {
862		usemap = ms->pageblock_flags;
863		memmap = sparse_decode_mem_map(ms->section_mem_map,
864						__section_nr(ms));
865		ms->section_mem_map = 0;
866		ms->pageblock_flags = NULL;
867	}
868	pgdat_resize_unlock(pgdat, &flags);
869
870	clear_hwpoisoned_pages(memmap + map_offset,
871			PAGES_PER_SECTION - map_offset);
872	free_section_usemap(memmap, usemap, altmap);
873}
874#endif /* CONFIG_MEMORY_HOTREMOVE */
875#endif /* CONFIG_MEMORY_HOTPLUG */