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 */
v3.15
 
  1/*
  2 * sparse memory mappings.
  3 */
  4#include <linux/mm.h>
  5#include <linux/slab.h>
  6#include <linux/mmzone.h>
  7#include <linux/bootmem.h>
  8#include <linux/compiler.h>
  9#include <linux/highmem.h>
 10#include <linux/export.h>
 11#include <linux/spinlock.h>
 12#include <linux/vmalloc.h>
 
 
 
 13
 14#include "internal.h"
 15#include <asm/dma.h>
 16#include <asm/pgalloc.h>
 17#include <asm/pgtable.h>
 18
 19/*
 20 * Permanent SPARSEMEM data:
 21 *
 22 * 1) mem_section	- memory sections, mem_map's for valid memory
 23 */
 24#ifdef CONFIG_SPARSEMEM_EXTREME
 25struct mem_section *mem_section[NR_SECTION_ROOTS]
 26	____cacheline_internodealigned_in_smp;
 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 struct mem_section noinline __init_refok *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		if (node_state(nid, N_HIGH_MEMORY))
 70			section = kzalloc_node(array_size, GFP_KERNEL, nid);
 71		else
 72			section = kzalloc(array_size, GFP_KERNEL);
 73	} else {
 74		section = memblock_virt_alloc_node(array_size, nid);
 
 
 
 
 75	}
 76
 77	return section;
 78}
 79
 80static int __meminit sparse_index_init(unsigned long section_nr, int nid)
 81{
 82	unsigned long root = SECTION_NR_TO_ROOT(section_nr);
 83	struct mem_section *section;
 84
 
 
 
 
 
 
 
 85	if (mem_section[root])
 86		return -EEXIST;
 87
 88	section = sparse_index_alloc(nid);
 89	if (!section)
 90		return -ENOMEM;
 91
 92	mem_section[root] = section;
 93
 94	return 0;
 95}
 96#else /* !SPARSEMEM_EXTREME */
 97static inline int sparse_index_init(unsigned long section_nr, int nid)
 98{
 99	return 0;
100}
101#endif
102
103/*
104 * Although written for the SPARSEMEM_EXTREME case, this happens
105 * to also work for the flat array case because
106 * NR_SECTION_ROOTS==NR_MEM_SECTIONS.
107 */
108int __section_nr(struct mem_section* ms)
109{
110	unsigned long root_nr;
111	struct mem_section* root;
112
113	for (root_nr = 0; root_nr < NR_SECTION_ROOTS; root_nr++) {
114		root = __nr_to_section(root_nr * SECTIONS_PER_ROOT);
115		if (!root)
116			continue;
117
118		if ((ms >= root) && (ms < (root + SECTIONS_PER_ROOT)))
119		     break;
120	}
121
122	VM_BUG_ON(root_nr == NR_SECTION_ROOTS);
123
124	return (root_nr * SECTIONS_PER_ROOT) + (ms - root);
125}
 
 
 
 
 
 
126
127/*
128 * During early boot, before section_mem_map is used for an actual
129 * mem_map, we use section_mem_map to store the section's NUMA
130 * node.  This keeps us from having to use another data structure.  The
131 * node information is cleared just before we store the real mem_map.
132 */
133static inline unsigned long sparse_encode_early_nid(int nid)
134{
135	return (nid << SECTION_NID_SHIFT);
136}
137
138static inline int sparse_early_nid(struct mem_section *section)
139{
140	return (section->section_mem_map >> SECTION_NID_SHIFT);
141}
142
143/* Validate the physical addressing limitations of the model */
144void __meminit mminit_validate_memmodel_limits(unsigned long *start_pfn,
145						unsigned long *end_pfn)
146{
147	unsigned long max_sparsemem_pfn = 1UL << (MAX_PHYSMEM_BITS-PAGE_SHIFT);
148
149	/*
150	 * Sanity checks - do not allow an architecture to pass
151	 * in larger pfns than the maximum scope of sparsemem:
152	 */
153	if (*start_pfn > max_sparsemem_pfn) {
154		mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
155			"Start of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
156			*start_pfn, *end_pfn, max_sparsemem_pfn);
157		WARN_ON_ONCE(1);
158		*start_pfn = max_sparsemem_pfn;
159		*end_pfn = max_sparsemem_pfn;
160	} else if (*end_pfn > max_sparsemem_pfn) {
161		mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
162			"End of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
163			*start_pfn, *end_pfn, max_sparsemem_pfn);
164		WARN_ON_ONCE(1);
165		*end_pfn = max_sparsemem_pfn;
166	}
167}
168
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
169/* Record a memory area against a node. */
170void __init memory_present(int nid, unsigned long start, unsigned long end)
171{
172	unsigned long pfn;
173
 
 
 
 
 
 
 
 
 
 
 
 
 
174	start &= PAGE_SECTION_MASK;
175	mminit_validate_memmodel_limits(&start, &end);
176	for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) {
177		unsigned long section = pfn_to_section_nr(pfn);
178		struct mem_section *ms;
179
180		sparse_index_init(section, nid);
181		set_section_nid(section, nid);
182
183		ms = __nr_to_section(section);
184		if (!ms->section_mem_map)
185			ms->section_mem_map = sparse_encode_early_nid(nid) |
186							SECTION_MARKED_PRESENT;
 
 
187	}
188}
189
190/*
191 * Only used by the i386 NUMA architecures, but relatively
192 * generic code.
 
193 */
194unsigned long __init node_memmap_size_bytes(int nid, unsigned long start_pfn,
195						     unsigned long end_pfn)
196{
197	unsigned long pfn;
198	unsigned long nr_pages = 0;
199
200	mminit_validate_memmodel_limits(&start_pfn, &end_pfn);
201	for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
202		if (nid != early_pfn_to_nid(pfn))
203			continue;
204
205		if (pfn_present(pfn))
206			nr_pages += PAGES_PER_SECTION;
207	}
208
209	return nr_pages * sizeof(struct page);
 
210}
211
212/*
213 * Subtle, we encode the real pfn into the mem_map such that
214 * the identity pfn - section_mem_map will return the actual
215 * physical page frame number.
216 */
217static unsigned long sparse_encode_mem_map(struct page *mem_map, unsigned long pnum)
218{
219	return (unsigned long)(mem_map - (section_nr_to_pfn(pnum)));
 
 
 
 
220}
221
 
222/*
223 * Decode mem_map from the coded memmap
224 */
225struct page *sparse_decode_mem_map(unsigned long coded_mem_map, unsigned long pnum)
226{
227	/* mask off the extra low bits of information */
228	coded_mem_map &= SECTION_MAP_MASK;
229	return ((struct page *)coded_mem_map) + section_nr_to_pfn(pnum);
230}
 
231
232static int __meminit sparse_init_one_section(struct mem_section *ms,
233		unsigned long pnum, struct page *mem_map,
234		unsigned long *pageblock_bitmap)
235{
236	if (!present_section(ms))
237		return -EINVAL;
238
239	ms->section_mem_map &= ~SECTION_MAP_MASK;
240	ms->section_mem_map |= sparse_encode_mem_map(mem_map, pnum) |
241							SECTION_HAS_MEM_MAP;
242 	ms->pageblock_flags = pageblock_bitmap;
 
243
244	return 1;
 
 
245}
246
247unsigned long usemap_size(void)
248{
249	unsigned long size_bytes;
250	size_bytes = roundup(SECTION_BLOCKFLAGS_BITS, 8) / 8;
251	size_bytes = roundup(size_bytes, sizeof(unsigned long));
252	return size_bytes;
253}
254
255#ifdef CONFIG_MEMORY_HOTPLUG
256static unsigned long *__kmalloc_section_usemap(void)
257{
258	return kmalloc(usemap_size(), GFP_KERNEL);
 
 
 
 
259}
260#endif /* CONFIG_MEMORY_HOTPLUG */
261
262#ifdef CONFIG_MEMORY_HOTREMOVE
263static unsigned long * __init
264sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
265					 unsigned long size)
266{
 
267	unsigned long goal, limit;
268	unsigned long *p;
269	int nid;
270	/*
271	 * A page may contain usemaps for other sections preventing the
272	 * page being freed and making a section unremovable while
273	 * other sections referencing the usemap remain active. Similarly,
274	 * a pgdat can prevent a section being removed. If section A
275	 * contains a pgdat and section B contains the usemap, both
276	 * sections become inter-dependent. This allocates usemaps
277	 * from the same section as the pgdat where possible to avoid
278	 * this problem.
279	 */
280	goal = __pa(pgdat) & (PAGE_SECTION_MASK << PAGE_SHIFT);
281	limit = goal + (1UL << PA_SECTION_SHIFT);
282	nid = early_pfn_to_nid(goal >> PAGE_SHIFT);
283again:
284	p = memblock_virt_alloc_try_nid_nopanic(size,
285						SMP_CACHE_BYTES, goal, limit,
286						nid);
287	if (!p && limit) {
288		limit = 0;
289		goto again;
290	}
291	return p;
292}
293
294static void __init check_usemap_section_nr(int nid, unsigned long *usemap)
 
295{
296	unsigned long usemap_snr, pgdat_snr;
297	static unsigned long old_usemap_snr = NR_MEM_SECTIONS;
298	static unsigned long old_pgdat_snr = NR_MEM_SECTIONS;
299	struct pglist_data *pgdat = NODE_DATA(nid);
300	int usemap_nid;
301
302	usemap_snr = pfn_to_section_nr(__pa(usemap) >> PAGE_SHIFT);
303	pgdat_snr = pfn_to_section_nr(__pa(pgdat) >> PAGE_SHIFT);
 
 
 
 
 
 
304	if (usemap_snr == pgdat_snr)
305		return;
306
307	if (old_usemap_snr == usemap_snr && old_pgdat_snr == pgdat_snr)
308		/* skip redundant message */
309		return;
310
311	old_usemap_snr = usemap_snr;
312	old_pgdat_snr = pgdat_snr;
313
314	usemap_nid = sparse_early_nid(__nr_to_section(usemap_snr));
315	if (usemap_nid != nid) {
316		printk(KERN_INFO
317		       "node %d must be removed before remove section %ld\n",
318		       nid, usemap_snr);
319		return;
320	}
321	/*
322	 * There is a circular dependency.
323	 * Some platforms allow un-removable section because they will just
324	 * gather other removable sections for dynamic partitioning.
325	 * Just notify un-removable section's number here.
326	 */
327	printk(KERN_INFO "Section %ld and %ld (node %d)", usemap_snr,
328	       pgdat_snr, nid);
329	printk(KERN_CONT
330	       " have a circular dependency on usemap and pgdat allocations\n");
331}
332#else
333static unsigned long * __init
334sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
335					 unsigned long size)
336{
337	return memblock_virt_alloc_node_nopanic(size, pgdat->node_id);
338}
339
340static void __init check_usemap_section_nr(int nid, unsigned long *usemap)
 
341{
342}
343#endif /* CONFIG_MEMORY_HOTREMOVE */
344
345static void __init sparse_early_usemaps_alloc_node(void *data,
346				 unsigned long pnum_begin,
347				 unsigned long pnum_end,
348				 unsigned long usemap_count, int nodeid)
349{
350	void *usemap;
351	unsigned long pnum;
352	unsigned long **usemap_map = (unsigned long **)data;
353	int size = usemap_size();
354
355	usemap = sparse_early_usemaps_alloc_pgdat_section(NODE_DATA(nodeid),
356							  size * usemap_count);
357	if (!usemap) {
358		printk(KERN_WARNING "%s: allocation failed\n", __func__);
359		return;
360	}
361
362	for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
363		if (!present_section_nr(pnum))
364			continue;
365		usemap_map[pnum] = usemap;
366		usemap += size;
367		check_usemap_section_nr(nodeid, usemap_map[pnum]);
368	}
369}
370
371#ifndef CONFIG_SPARSEMEM_VMEMMAP
372struct page __init *sparse_mem_map_populate(unsigned long pnum, int nid)
373{
374	struct page *map;
375	unsigned long size;
 
376
377	map = alloc_remap(nid, sizeof(struct page) * PAGES_PER_SECTION);
378	if (map)
379		return map;
380
381	size = PAGE_ALIGN(sizeof(struct page) * PAGES_PER_SECTION);
382	map = memblock_virt_alloc_try_nid(size,
383					  PAGE_SIZE, __pa(MAX_DMA_ADDRESS),
384					  BOOTMEM_ALLOC_ACCESSIBLE, nid);
 
 
385	return map;
386}
387void __init sparse_mem_maps_populate_node(struct page **map_map,
388					  unsigned long pnum_begin,
389					  unsigned long pnum_end,
390					  unsigned long map_count, int nodeid)
 
 
391{
392	void *map;
393	unsigned long pnum;
394	unsigned long size = sizeof(struct page) * PAGES_PER_SECTION;
395
396	map = alloc_remap(nodeid, size * map_count);
397	if (map) {
398		for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
399			if (!present_section_nr(pnum))
400				continue;
401			map_map[pnum] = map;
402			map += size;
403		}
404		return;
405	}
 
 
 
406
407	size = PAGE_ALIGN(size);
408	map = memblock_virt_alloc_try_nid(size * map_count,
409					  PAGE_SIZE, __pa(MAX_DMA_ADDRESS),
410					  BOOTMEM_ALLOC_ACCESSIBLE, nodeid);
411	if (map) {
412		for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
413			if (!present_section_nr(pnum))
414				continue;
415			map_map[pnum] = map;
416			map += size;
417		}
418		return;
419	}
420
421	/* fallback */
422	for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
423		struct mem_section *ms;
424
425		if (!present_section_nr(pnum))
426			continue;
427		map_map[pnum] = sparse_mem_map_populate(pnum, nodeid);
428		if (map_map[pnum])
429			continue;
430		ms = __nr_to_section(pnum);
431		printk(KERN_ERR "%s: sparsemem memory map backing failed "
432			"some memory will not be available.\n", __func__);
433		ms->section_mem_map = 0;
434	}
435}
436#endif /* !CONFIG_SPARSEMEM_VMEMMAP */
437
438#ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
439static void __init sparse_early_mem_maps_alloc_node(void *data,
440				 unsigned long pnum_begin,
441				 unsigned long pnum_end,
442				 unsigned long map_count, int nodeid)
443{
444	struct page **map_map = (struct page **)data;
445	sparse_mem_maps_populate_node(map_map, pnum_begin, pnum_end,
446					 map_count, nodeid);
447}
448#else
449static struct page __init *sparse_early_mem_map_alloc(unsigned long pnum)
450{
451	struct page *map;
452	struct mem_section *ms = __nr_to_section(pnum);
453	int nid = sparse_early_nid(ms);
454
455	map = sparse_mem_map_populate(pnum, nid);
456	if (map)
457		return map;
458
459	printk(KERN_ERR "%s: sparsemem memory map backing failed "
460			"some memory will not be available.\n", __func__);
461	ms->section_mem_map = 0;
462	return NULL;
 
 
 
 
463}
464#endif
465
466void __weak __meminit vmemmap_populate_print_last(void)
467{
468}
469
470/**
471 *  alloc_usemap_and_memmap - memory alloction for pageblock flags and vmemmap
472 *  @map: usemap_map for pageblock flags or mmap_map for vmemmap
473 */
474static void __init alloc_usemap_and_memmap(void (*alloc_func)
475					(void *, unsigned long, unsigned long,
476					unsigned long, int), void *data)
477{
 
478	unsigned long pnum;
479	unsigned long map_count;
480	int nodeid_begin = 0;
481	unsigned long pnum_begin = 0;
482
483	for (pnum = 0; pnum < NR_MEM_SECTIONS; pnum++) {
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
484		struct mem_section *ms;
485
486		if (!present_section_nr(pnum))
487			continue;
488		ms = __nr_to_section(pnum);
489		nodeid_begin = sparse_early_nid(ms);
490		pnum_begin = pnum;
491		break;
492	}
493	map_count = 1;
494	for (pnum = pnum_begin + 1; pnum < NR_MEM_SECTIONS; pnum++) {
495		struct mem_section *ms;
496		int nodeid;
497
498		if (!present_section_nr(pnum))
499			continue;
500		ms = __nr_to_section(pnum);
501		nodeid = sparse_early_nid(ms);
502		if (nodeid == nodeid_begin) {
503			map_count++;
504			continue;
505		}
506		/* ok, we need to take cake of from pnum_begin to pnum - 1*/
507		alloc_func(data, pnum_begin, pnum,
508						map_count, nodeid_begin);
509		/* new start, update count etc*/
510		nodeid_begin = nodeid;
511		pnum_begin = pnum;
512		map_count = 1;
513	}
514	/* ok, last chunk */
515	alloc_func(data, pnum_begin, NR_MEM_SECTIONS,
516						map_count, nodeid_begin);
517}
518
519/*
520 * Allocate the accumulated non-linear sections, allocate a mem_map
521 * for each and record the physical to section mapping.
522 */
523void __init sparse_init(void)
524{
525	unsigned long pnum;
526	struct page *map;
527	unsigned long *usemap;
528	unsigned long **usemap_map;
529	int size;
530#ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
531	int size2;
532	struct page **map_map;
533#endif
534
535	/* see include/linux/mmzone.h 'struct mem_section' definition */
536	BUILD_BUG_ON(!is_power_of_2(sizeof(struct mem_section)));
537
538	/* Setup pageblock_order for HUGETLB_PAGE_SIZE_VARIABLE */
539	set_pageblock_order();
540
541	/*
542	 * map is using big page (aka 2M in x86 64 bit)
543	 * usemap is less one page (aka 24 bytes)
544	 * so alloc 2M (with 2M align) and 24 bytes in turn will
545	 * make next 2M slip to one more 2M later.
546	 * then in big system, the memory will have a lot of holes...
547	 * here try to allocate 2M pages continuously.
548	 *
549	 * powerpc need to call sparse_init_one_section right after each
550	 * sparse_early_mem_map_alloc, so allocate usemap_map at first.
551	 */
552	size = sizeof(unsigned long *) * NR_MEM_SECTIONS;
553	usemap_map = memblock_virt_alloc(size, 0);
554	if (!usemap_map)
555		panic("can not allocate usemap_map\n");
556	alloc_usemap_and_memmap(sparse_early_usemaps_alloc_node,
557							(void *)usemap_map);
558
559#ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
560	size2 = sizeof(struct page *) * NR_MEM_SECTIONS;
561	map_map = memblock_virt_alloc(size2, 0);
562	if (!map_map)
563		panic("can not allocate map_map\n");
564	alloc_usemap_and_memmap(sparse_early_mem_maps_alloc_node,
565							(void *)map_map);
566#endif
567
568	for (pnum = 0; pnum < NR_MEM_SECTIONS; pnum++) {
569		if (!present_section_nr(pnum))
570			continue;
 
 
 
 
 
 
 
 
 
 
 
 
 
571
572		usemap = usemap_map[pnum];
573		if (!usemap)
574			continue;
 
 
 
 
 
575
576#ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
577		map = map_map[pnum];
578#else
579		map = sparse_early_mem_map_alloc(pnum);
580#endif
581		if (!map)
582			continue;
583
584		sparse_init_one_section(__nr_to_section(pnum), pnum, map,
585								usemap);
586	}
 
587
588	vmemmap_populate_print_last();
 
 
 
 
 
 
 
 
 
 
 
 
 
 
589
590#ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
591	memblock_free_early(__pa(map_map), size2);
592#endif
593	memblock_free_early(__pa(usemap_map), size);
594}
595
596#ifdef CONFIG_MEMORY_HOTPLUG
597#ifdef CONFIG_SPARSEMEM_VMEMMAP
598static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid)
 
599{
600	/* This will make the necessary allocations eventually. */
601	return sparse_mem_map_populate(pnum, nid);
602}
603static void __kfree_section_memmap(struct page *memmap)
 
 
604{
605	unsigned long start = (unsigned long)memmap;
606	unsigned long end = (unsigned long)(memmap + PAGES_PER_SECTION);
607
608	vmemmap_free(start, end);
609}
610#ifdef CONFIG_MEMORY_HOTREMOVE
611static void free_map_bootmem(struct page *memmap)
612{
613	unsigned long start = (unsigned long)memmap;
614	unsigned long end = (unsigned long)(memmap + PAGES_PER_SECTION);
615
616	vmemmap_free(start, end);
617}
618#endif /* CONFIG_MEMORY_HOTREMOVE */
619#else
620static struct page *__kmalloc_section_memmap(void)
621{
622	struct page *page, *ret;
623	unsigned long memmap_size = sizeof(struct page) * PAGES_PER_SECTION;
 
 
 
624
625	page = alloc_pages(GFP_KERNEL|__GFP_NOWARN, get_order(memmap_size));
626	if (page)
627		goto got_map_page;
628
629	ret = vmalloc(memmap_size);
630	if (ret)
631		goto got_map_ptr;
 
632
633	return NULL;
634got_map_page:
635	ret = (struct page *)pfn_to_kaddr(page_to_pfn(page));
636got_map_ptr:
637
638	return ret;
 
 
 
639}
640
641static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid)
642{
643	return __kmalloc_section_memmap();
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
644}
645
646static void __kfree_section_memmap(struct page *memmap)
 
647{
648	if (is_vmalloc_addr(memmap))
649		vfree(memmap);
650	else
651		free_pages((unsigned long)memmap,
652			   get_order(sizeof(struct page) * PAGES_PER_SECTION));
653}
654
655#ifdef CONFIG_MEMORY_HOTREMOVE
656static void free_map_bootmem(struct page *memmap)
657{
658	unsigned long maps_section_nr, removing_section_nr, i;
659	unsigned long magic, nr_pages;
660	struct page *page = virt_to_page(memmap);
661
662	nr_pages = PAGE_ALIGN(PAGES_PER_SECTION * sizeof(struct page))
663		>> PAGE_SHIFT;
664
665	for (i = 0; i < nr_pages; i++, page++) {
666		magic = (unsigned long) page->lru.next;
667
668		BUG_ON(magic == NODE_INFO);
669
670		maps_section_nr = pfn_to_section_nr(page_to_pfn(page));
671		removing_section_nr = page->private;
672
673		/*
674		 * When this function is called, the removing section is
675		 * logical offlined state. This means all pages are isolated
676		 * from page allocator. If removing section's memmap is placed
677		 * on the same section, it must not be freed.
678		 * If it is freed, page allocator may allocate it which will
679		 * be removed physically soon.
680		 */
681		if (maps_section_nr != removing_section_nr)
682			put_page_bootmem(page);
683	}
684}
685#endif /* CONFIG_MEMORY_HOTREMOVE */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
686#endif /* CONFIG_SPARSEMEM_VMEMMAP */
687
688/*
689 * returns the number of sections whose mem_maps were properly
690 * set.  If this is <=0, then that means that the passed-in
691 * map was not consumed and must be freed.
 
 
 
 
 
 
 
 
 
 
 
692 */
693int __meminit sparse_add_one_section(struct zone *zone, unsigned long start_pfn)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
694{
695	unsigned long section_nr = pfn_to_section_nr(start_pfn);
696	struct pglist_data *pgdat = zone->zone_pgdat;
697	struct mem_section *ms;
698	struct page *memmap;
699	unsigned long *usemap;
700	unsigned long flags;
701	int ret;
 
 
 
 
 
 
 
 
 
 
 
 
 
702
703	/*
704	 * no locking for this, because it does its own
705	 * plus, it does a kmalloc
 
 
 
706	 */
707	ret = sparse_index_init(section_nr, pgdat->node_id);
708	if (ret < 0 && ret != -EEXIST)
709		return ret;
710	memmap = kmalloc_section_memmap(section_nr, pgdat->node_id);
711	if (!memmap)
712		return -ENOMEM;
713	usemap = __kmalloc_section_usemap();
714	if (!usemap) {
715		__kfree_section_memmap(memmap);
716		return -ENOMEM;
717	}
718
719	pgdat_resize_lock(pgdat, &flags);
 
720
721	ms = __pfn_to_section(start_pfn);
722	if (ms->section_mem_map & SECTION_MARKED_PRESENT) {
723		ret = -EEXIST;
724		goto out;
725	}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
726
727	memset(memmap, 0, sizeof(struct page) * PAGES_PER_SECTION);
 
 
728
729	ms->section_mem_map |= SECTION_MARKED_PRESENT;
 
 
 
 
730
731	ret = sparse_init_one_section(ms, section_nr, memmap, usemap);
 
 
 
 
 
 
 
732
733out:
734	pgdat_resize_unlock(pgdat, &flags);
735	if (ret <= 0) {
736		kfree(usemap);
737		__kfree_section_memmap(memmap);
738	}
739	return ret;
740}
741
742#ifdef CONFIG_MEMORY_HOTREMOVE
743#ifdef CONFIG_MEMORY_FAILURE
744static void clear_hwpoisoned_pages(struct page *memmap, int nr_pages)
745{
746	int i;
747
748	if (!memmap)
 
 
 
 
 
 
749		return;
750
751	for (i = 0; i < PAGES_PER_SECTION; i++) {
752		if (PageHWPoison(&memmap[i])) {
753			atomic_long_sub(1, &num_poisoned_pages);
754			ClearPageHWPoison(&memmap[i]);
755		}
756	}
757}
758#else
759static inline void clear_hwpoisoned_pages(struct page *memmap, int nr_pages)
760{
761}
762#endif
763
764static void free_section_usemap(struct page *memmap, unsigned long *usemap)
765{
766	struct page *usemap_page;
767
768	if (!usemap)
769		return;
770
771	usemap_page = virt_to_page(usemap);
772	/*
773	 * Check to see if allocation came from hot-plug-add
774	 */
775	if (PageSlab(usemap_page) || PageCompound(usemap_page)) {
776		kfree(usemap);
777		if (memmap)
778			__kfree_section_memmap(memmap);
779		return;
780	}
781
782	/*
783	 * The usemap came from bootmem. This is packed with other usemaps
784	 * on the section which has pgdat at boot time. Just keep it as is now.
785	 */
786
787	if (memmap)
788		free_map_bootmem(memmap);
789}
790
791void sparse_remove_one_section(struct zone *zone, struct mem_section *ms)
792{
793	struct page *memmap = NULL;
794	unsigned long *usemap = NULL, flags;
795	struct pglist_data *pgdat = zone->zone_pgdat;
796
797	pgdat_resize_lock(pgdat, &flags);
798	if (ms->section_mem_map) {
799		usemap = ms->pageblock_flags;
800		memmap = sparse_decode_mem_map(ms->section_mem_map,
801						__section_nr(ms));
802		ms->section_mem_map = 0;
803		ms->pageblock_flags = NULL;
804	}
805	pgdat_resize_unlock(pgdat, &flags);
806
807	clear_hwpoisoned_pages(memmap, PAGES_PER_SECTION);
808	free_section_usemap(memmap, usemap);
809}
810#endif /* CONFIG_MEMORY_HOTREMOVE */
811#endif /* CONFIG_MEMORY_HOTPLUG */