Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * sparse memory mappings.
4 */
5#include <linux/mm.h>
6#include <linux/slab.h>
7#include <linux/mmzone.h>
8#include <linux/memblock.h>
9#include <linux/compiler.h>
10#include <linux/highmem.h>
11#include <linux/export.h>
12#include <linux/spinlock.h>
13#include <linux/vmalloc.h>
14#include <linux/swap.h>
15#include <linux/swapops.h>
16
17#include "internal.h"
18#include <asm/dma.h>
19
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_alloc_node(array_size, SMP_CACHE_BYTES,
72 nid);
73 if (!section)
74 panic("%s: Failed to allocate %lu bytes nid=%d\n",
75 __func__, array_size, nid);
76 }
77
78 return section;
79}
80
81static int __meminit sparse_index_init(unsigned long section_nr, int nid)
82{
83 unsigned long root = SECTION_NR_TO_ROOT(section_nr);
84 struct mem_section *section;
85
86 /*
87 * An existing section is possible in the sub-section hotplug
88 * case. First hot-add instantiates, follow-on hot-add reuses
89 * the existing section.
90 *
91 * The mem_hotplug_lock resolves the apparent race below.
92 */
93 if (mem_section[root])
94 return 0;
95
96 section = sparse_index_alloc(nid);
97 if (!section)
98 return -ENOMEM;
99
100 mem_section[root] = section;
101
102 return 0;
103}
104#else /* !SPARSEMEM_EXTREME */
105static inline int sparse_index_init(unsigned long section_nr, int nid)
106{
107 return 0;
108}
109#endif
110
111#ifdef CONFIG_SPARSEMEM_EXTREME
112unsigned long __section_nr(struct mem_section *ms)
113{
114 unsigned long root_nr;
115 struct mem_section *root = NULL;
116
117 for (root_nr = 0; root_nr < NR_SECTION_ROOTS; root_nr++) {
118 root = __nr_to_section(root_nr * SECTIONS_PER_ROOT);
119 if (!root)
120 continue;
121
122 if ((ms >= root) && (ms < (root + SECTIONS_PER_ROOT)))
123 break;
124 }
125
126 VM_BUG_ON(!root);
127
128 return (root_nr * SECTIONS_PER_ROOT) + (ms - root);
129}
130#else
131unsigned long __section_nr(struct mem_section *ms)
132{
133 return (unsigned long)(ms - mem_section[0]);
134}
135#endif
136
137/*
138 * During early boot, before section_mem_map is used for an actual
139 * mem_map, we use section_mem_map to store the section's NUMA
140 * node. This keeps us from having to use another data structure. The
141 * node information is cleared just before we store the real mem_map.
142 */
143static inline unsigned long sparse_encode_early_nid(int nid)
144{
145 return (nid << SECTION_NID_SHIFT);
146}
147
148static inline int sparse_early_nid(struct mem_section *section)
149{
150 return (section->section_mem_map >> SECTION_NID_SHIFT);
151}
152
153/* Validate the physical addressing limitations of the model */
154void __meminit mminit_validate_memmodel_limits(unsigned long *start_pfn,
155 unsigned long *end_pfn)
156{
157 unsigned long max_sparsemem_pfn = 1UL << (MAX_PHYSMEM_BITS-PAGE_SHIFT);
158
159 /*
160 * Sanity checks - do not allow an architecture to pass
161 * in larger pfns than the maximum scope of sparsemem:
162 */
163 if (*start_pfn > max_sparsemem_pfn) {
164 mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
165 "Start of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
166 *start_pfn, *end_pfn, max_sparsemem_pfn);
167 WARN_ON_ONCE(1);
168 *start_pfn = max_sparsemem_pfn;
169 *end_pfn = max_sparsemem_pfn;
170 } else if (*end_pfn > max_sparsemem_pfn) {
171 mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
172 "End of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
173 *start_pfn, *end_pfn, max_sparsemem_pfn);
174 WARN_ON_ONCE(1);
175 *end_pfn = max_sparsemem_pfn;
176 }
177}
178
179/*
180 * There are a number of times that we loop over NR_MEM_SECTIONS,
181 * looking for section_present() on each. But, when we have very
182 * large physical address spaces, NR_MEM_SECTIONS can also be
183 * very large which makes the loops quite long.
184 *
185 * Keeping track of this gives us an easy way to break out of
186 * those loops early.
187 */
188unsigned long __highest_present_section_nr;
189static void section_mark_present(struct mem_section *ms)
190{
191 unsigned long section_nr = __section_nr(ms);
192
193 if (section_nr > __highest_present_section_nr)
194 __highest_present_section_nr = section_nr;
195
196 ms->section_mem_map |= SECTION_MARKED_PRESENT;
197}
198
199#define for_each_present_section_nr(start, section_nr) \
200 for (section_nr = next_present_section_nr(start-1); \
201 ((section_nr != -1) && \
202 (section_nr <= __highest_present_section_nr)); \
203 section_nr = next_present_section_nr(section_nr))
204
205static inline unsigned long first_present_section_nr(void)
206{
207 return next_present_section_nr(-1);
208}
209
210#ifdef CONFIG_SPARSEMEM_VMEMMAP
211static void subsection_mask_set(unsigned long *map, unsigned long pfn,
212 unsigned long nr_pages)
213{
214 int idx = subsection_map_index(pfn);
215 int end = subsection_map_index(pfn + nr_pages - 1);
216
217 bitmap_set(map, idx, end - idx + 1);
218}
219
220void __init subsection_map_init(unsigned long pfn, unsigned long nr_pages)
221{
222 int end_sec = pfn_to_section_nr(pfn + nr_pages - 1);
223 unsigned long nr, start_sec = pfn_to_section_nr(pfn);
224
225 if (!nr_pages)
226 return;
227
228 for (nr = start_sec; nr <= end_sec; nr++) {
229 struct mem_section *ms;
230 unsigned long pfns;
231
232 pfns = min(nr_pages, PAGES_PER_SECTION
233 - (pfn & ~PAGE_SECTION_MASK));
234 ms = __nr_to_section(nr);
235 subsection_mask_set(ms->usage->subsection_map, pfn, pfns);
236
237 pr_debug("%s: sec: %lu pfns: %lu set(%d, %d)\n", __func__, nr,
238 pfns, subsection_map_index(pfn),
239 subsection_map_index(pfn + pfns - 1));
240
241 pfn += pfns;
242 nr_pages -= pfns;
243 }
244}
245#else
246void __init subsection_map_init(unsigned long pfn, unsigned long nr_pages)
247{
248}
249#endif
250
251/* Record a memory area against a node. */
252static void __init memory_present(int nid, unsigned long start, unsigned long end)
253{
254 unsigned long pfn;
255
256#ifdef CONFIG_SPARSEMEM_EXTREME
257 if (unlikely(!mem_section)) {
258 unsigned long size, align;
259
260 size = sizeof(struct mem_section*) * NR_SECTION_ROOTS;
261 align = 1 << (INTERNODE_CACHE_SHIFT);
262 mem_section = memblock_alloc(size, align);
263 if (!mem_section)
264 panic("%s: Failed to allocate %lu bytes align=0x%lx\n",
265 __func__, size, align);
266 }
267#endif
268
269 start &= PAGE_SECTION_MASK;
270 mminit_validate_memmodel_limits(&start, &end);
271 for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) {
272 unsigned long section = pfn_to_section_nr(pfn);
273 struct mem_section *ms;
274
275 sparse_index_init(section, nid);
276 set_section_nid(section, nid);
277
278 ms = __nr_to_section(section);
279 if (!ms->section_mem_map) {
280 ms->section_mem_map = sparse_encode_early_nid(nid) |
281 SECTION_IS_ONLINE;
282 section_mark_present(ms);
283 }
284 }
285}
286
287/*
288 * Mark all memblocks as present using memory_present().
289 * This is a convenience function that is useful to mark all of the systems
290 * memory as present during initialization.
291 */
292static void __init memblocks_present(void)
293{
294 struct memblock_region *reg;
295
296 for_each_memblock(memory, reg) {
297 memory_present(memblock_get_region_node(reg),
298 memblock_region_memory_base_pfn(reg),
299 memblock_region_memory_end_pfn(reg));
300 }
301}
302
303/*
304 * Subtle, we encode the real pfn into the mem_map such that
305 * the identity pfn - section_mem_map will return the actual
306 * physical page frame number.
307 */
308static unsigned long sparse_encode_mem_map(struct page *mem_map, unsigned long pnum)
309{
310 unsigned long coded_mem_map =
311 (unsigned long)(mem_map - (section_nr_to_pfn(pnum)));
312 BUILD_BUG_ON(SECTION_MAP_LAST_BIT > (1UL<<PFN_SECTION_SHIFT));
313 BUG_ON(coded_mem_map & ~SECTION_MAP_MASK);
314 return coded_mem_map;
315}
316
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
327static void __meminit sparse_init_one_section(struct mem_section *ms,
328 unsigned long pnum, struct page *mem_map,
329 struct mem_section_usage *usage, unsigned long flags)
330{
331 ms->section_mem_map &= ~SECTION_MAP_MASK;
332 ms->section_mem_map |= sparse_encode_mem_map(mem_map, pnum)
333 | SECTION_HAS_MEM_MAP | flags;
334 ms->usage = usage;
335}
336
337static unsigned long usemap_size(void)
338{
339 return BITS_TO_LONGS(SECTION_BLOCKFLAGS_BITS) * sizeof(unsigned long);
340}
341
342size_t mem_section_usage_size(void)
343{
344 return sizeof(struct mem_section_usage) + usemap_size();
345}
346
347#ifdef CONFIG_MEMORY_HOTREMOVE
348static struct mem_section_usage * __init
349sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
350 unsigned long size)
351{
352 struct mem_section_usage *usage;
353 unsigned long goal, limit;
354 int nid;
355 /*
356 * A page may contain usemaps for other sections preventing the
357 * page being freed and making a section unremovable while
358 * other sections referencing the usemap remain active. Similarly,
359 * a pgdat can prevent a section being removed. If section A
360 * contains a pgdat and section B contains the usemap, both
361 * sections become inter-dependent. This allocates usemaps
362 * from the same section as the pgdat where possible to avoid
363 * this problem.
364 */
365 goal = __pa(pgdat) & (PAGE_SECTION_MASK << PAGE_SHIFT);
366 limit = goal + (1UL << PA_SECTION_SHIFT);
367 nid = early_pfn_to_nid(goal >> PAGE_SHIFT);
368again:
369 usage = memblock_alloc_try_nid(size, SMP_CACHE_BYTES, goal, limit, nid);
370 if (!usage && limit) {
371 limit = 0;
372 goto again;
373 }
374 return usage;
375}
376
377static void __init check_usemap_section_nr(int nid,
378 struct mem_section_usage *usage)
379{
380 unsigned long usemap_snr, pgdat_snr;
381 static unsigned long old_usemap_snr;
382 static unsigned long old_pgdat_snr;
383 struct pglist_data *pgdat = NODE_DATA(nid);
384 int usemap_nid;
385
386 /* First call */
387 if (!old_usemap_snr) {
388 old_usemap_snr = NR_MEM_SECTIONS;
389 old_pgdat_snr = NR_MEM_SECTIONS;
390 }
391
392 usemap_snr = pfn_to_section_nr(__pa(usage) >> PAGE_SHIFT);
393 pgdat_snr = pfn_to_section_nr(__pa(pgdat) >> PAGE_SHIFT);
394 if (usemap_snr == pgdat_snr)
395 return;
396
397 if (old_usemap_snr == usemap_snr && old_pgdat_snr == pgdat_snr)
398 /* skip redundant message */
399 return;
400
401 old_usemap_snr = usemap_snr;
402 old_pgdat_snr = pgdat_snr;
403
404 usemap_nid = sparse_early_nid(__nr_to_section(usemap_snr));
405 if (usemap_nid != nid) {
406 pr_info("node %d must be removed before remove section %ld\n",
407 nid, usemap_snr);
408 return;
409 }
410 /*
411 * There is a circular dependency.
412 * Some platforms allow un-removable section because they will just
413 * gather other removable sections for dynamic partitioning.
414 * Just notify un-removable section's number here.
415 */
416 pr_info("Section %ld and %ld (node %d) have a circular dependency on usemap and pgdat allocations\n",
417 usemap_snr, pgdat_snr, nid);
418}
419#else
420static struct mem_section_usage * __init
421sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
422 unsigned long size)
423{
424 return memblock_alloc_node(size, SMP_CACHE_BYTES, pgdat->node_id);
425}
426
427static void __init check_usemap_section_nr(int nid,
428 struct mem_section_usage *usage)
429{
430}
431#endif /* CONFIG_MEMORY_HOTREMOVE */
432
433#ifdef CONFIG_SPARSEMEM_VMEMMAP
434static unsigned long __init section_map_size(void)
435{
436 return ALIGN(sizeof(struct page) * PAGES_PER_SECTION, PMD_SIZE);
437}
438
439#else
440static unsigned long __init section_map_size(void)
441{
442 return PAGE_ALIGN(sizeof(struct page) * PAGES_PER_SECTION);
443}
444
445struct page __init *__populate_section_memmap(unsigned long pfn,
446 unsigned long nr_pages, int nid, struct vmem_altmap *altmap)
447{
448 unsigned long size = section_map_size();
449 struct page *map = sparse_buffer_alloc(size);
450 phys_addr_t addr = __pa(MAX_DMA_ADDRESS);
451
452 if (map)
453 return map;
454
455 map = memblock_alloc_try_nid_raw(size, size, addr,
456 MEMBLOCK_ALLOC_ACCESSIBLE, nid);
457 if (!map)
458 panic("%s: Failed to allocate %lu bytes align=0x%lx nid=%d from=%pa\n",
459 __func__, size, PAGE_SIZE, nid, &addr);
460
461 return map;
462}
463#endif /* !CONFIG_SPARSEMEM_VMEMMAP */
464
465static void *sparsemap_buf __meminitdata;
466static void *sparsemap_buf_end __meminitdata;
467
468static inline void __meminit sparse_buffer_free(unsigned long size)
469{
470 WARN_ON(!sparsemap_buf || size == 0);
471 memblock_free_early(__pa(sparsemap_buf), size);
472}
473
474static void __init sparse_buffer_init(unsigned long size, int nid)
475{
476 phys_addr_t addr = __pa(MAX_DMA_ADDRESS);
477 WARN_ON(sparsemap_buf); /* forgot to call sparse_buffer_fini()? */
478 /*
479 * Pre-allocated buffer is mainly used by __populate_section_memmap
480 * and we want it to be properly aligned to the section size - this is
481 * especially the case for VMEMMAP which maps memmap to PMDs
482 */
483 sparsemap_buf = memblock_alloc_exact_nid_raw(size, section_map_size(),
484 addr, MEMBLOCK_ALLOC_ACCESSIBLE, nid);
485 sparsemap_buf_end = sparsemap_buf + size;
486}
487
488static void __init sparse_buffer_fini(void)
489{
490 unsigned long size = sparsemap_buf_end - sparsemap_buf;
491
492 if (sparsemap_buf && size > 0)
493 sparse_buffer_free(size);
494 sparsemap_buf = NULL;
495}
496
497void * __meminit sparse_buffer_alloc(unsigned long size)
498{
499 void *ptr = NULL;
500
501 if (sparsemap_buf) {
502 ptr = (void *) roundup((unsigned long)sparsemap_buf, size);
503 if (ptr + size > sparsemap_buf_end)
504 ptr = NULL;
505 else {
506 /* Free redundant aligned space */
507 if ((unsigned long)(ptr - sparsemap_buf) > 0)
508 sparse_buffer_free((unsigned long)(ptr - sparsemap_buf));
509 sparsemap_buf = ptr + size;
510 }
511 }
512 return ptr;
513}
514
515void __weak __meminit vmemmap_populate_print_last(void)
516{
517}
518
519/*
520 * Initialize sparse on a specific node. The node spans [pnum_begin, pnum_end)
521 * And number of present sections in this node is map_count.
522 */
523static void __init sparse_init_nid(int nid, unsigned long pnum_begin,
524 unsigned long pnum_end,
525 unsigned long map_count)
526{
527 struct mem_section_usage *usage;
528 unsigned long pnum;
529 struct page *map;
530
531 usage = sparse_early_usemaps_alloc_pgdat_section(NODE_DATA(nid),
532 mem_section_usage_size() * map_count);
533 if (!usage) {
534 pr_err("%s: node[%d] usemap allocation failed", __func__, nid);
535 goto failed;
536 }
537 sparse_buffer_init(map_count * section_map_size(), nid);
538 for_each_present_section_nr(pnum_begin, pnum) {
539 unsigned long pfn = section_nr_to_pfn(pnum);
540
541 if (pnum >= pnum_end)
542 break;
543
544 map = __populate_section_memmap(pfn, PAGES_PER_SECTION,
545 nid, NULL);
546 if (!map) {
547 pr_err("%s: node[%d] memory map backing failed. Some memory will not be available.",
548 __func__, nid);
549 pnum_begin = pnum;
550 goto failed;
551 }
552 check_usemap_section_nr(nid, usage);
553 sparse_init_one_section(__nr_to_section(pnum), pnum, map, usage,
554 SECTION_IS_EARLY);
555 usage = (void *) usage + mem_section_usage_size();
556 }
557 sparse_buffer_fini();
558 return;
559failed:
560 /* We failed to allocate, mark all the following pnums as not present */
561 for_each_present_section_nr(pnum_begin, pnum) {
562 struct mem_section *ms;
563
564 if (pnum >= pnum_end)
565 break;
566 ms = __nr_to_section(pnum);
567 ms->section_mem_map = 0;
568 }
569}
570
571/*
572 * Allocate the accumulated non-linear sections, allocate a mem_map
573 * for each and record the physical to section mapping.
574 */
575void __init sparse_init(void)
576{
577 unsigned long pnum_end, pnum_begin, map_count = 1;
578 int nid_begin;
579
580 memblocks_present();
581
582 pnum_begin = first_present_section_nr();
583 nid_begin = sparse_early_nid(__nr_to_section(pnum_begin));
584
585 /* Setup pageblock_order for HUGETLB_PAGE_SIZE_VARIABLE */
586 set_pageblock_order();
587
588 for_each_present_section_nr(pnum_begin + 1, pnum_end) {
589 int nid = sparse_early_nid(__nr_to_section(pnum_end));
590
591 if (nid == nid_begin) {
592 map_count++;
593 continue;
594 }
595 /* Init node with sections in range [pnum_begin, pnum_end) */
596 sparse_init_nid(nid_begin, pnum_begin, pnum_end, map_count);
597 nid_begin = nid;
598 pnum_begin = pnum_end;
599 map_count = 1;
600 }
601 /* cover the last node */
602 sparse_init_nid(nid_begin, pnum_begin, pnum_end, map_count);
603 vmemmap_populate_print_last();
604}
605
606#ifdef CONFIG_MEMORY_HOTPLUG
607
608/* Mark all memory sections within the pfn range as online */
609void online_mem_sections(unsigned long start_pfn, unsigned long end_pfn)
610{
611 unsigned long pfn;
612
613 for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
614 unsigned long section_nr = pfn_to_section_nr(pfn);
615 struct mem_section *ms;
616
617 /* onlining code should never touch invalid ranges */
618 if (WARN_ON(!valid_section_nr(section_nr)))
619 continue;
620
621 ms = __nr_to_section(section_nr);
622 ms->section_mem_map |= SECTION_IS_ONLINE;
623 }
624}
625
626#ifdef CONFIG_MEMORY_HOTREMOVE
627/* Mark all memory sections within the pfn range as offline */
628void offline_mem_sections(unsigned long start_pfn, unsigned long end_pfn)
629{
630 unsigned long pfn;
631
632 for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
633 unsigned long section_nr = pfn_to_section_nr(pfn);
634 struct mem_section *ms;
635
636 /*
637 * TODO this needs some double checking. Offlining code makes
638 * sure to check pfn_valid but those checks might be just bogus
639 */
640 if (WARN_ON(!valid_section_nr(section_nr)))
641 continue;
642
643 ms = __nr_to_section(section_nr);
644 ms->section_mem_map &= ~SECTION_IS_ONLINE;
645 }
646}
647#endif
648
649#ifdef CONFIG_SPARSEMEM_VMEMMAP
650static struct page * __meminit populate_section_memmap(unsigned long pfn,
651 unsigned long nr_pages, int nid, struct vmem_altmap *altmap)
652{
653 return __populate_section_memmap(pfn, nr_pages, nid, altmap);
654}
655
656static void depopulate_section_memmap(unsigned long pfn, unsigned long nr_pages,
657 struct vmem_altmap *altmap)
658{
659 unsigned long start = (unsigned long) pfn_to_page(pfn);
660 unsigned long end = start + nr_pages * sizeof(struct page);
661
662 vmemmap_free(start, end, altmap);
663}
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
672static int clear_subsection_map(unsigned long pfn, unsigned long nr_pages)
673{
674 DECLARE_BITMAP(map, SUBSECTIONS_PER_SECTION) = { 0 };
675 DECLARE_BITMAP(tmp, SUBSECTIONS_PER_SECTION) = { 0 };
676 struct mem_section *ms = __pfn_to_section(pfn);
677 unsigned long *subsection_map = ms->usage
678 ? &ms->usage->subsection_map[0] : NULL;
679
680 subsection_mask_set(map, pfn, nr_pages);
681 if (subsection_map)
682 bitmap_and(tmp, map, subsection_map, SUBSECTIONS_PER_SECTION);
683
684 if (WARN(!subsection_map || !bitmap_equal(tmp, map, SUBSECTIONS_PER_SECTION),
685 "section already deactivated (%#lx + %ld)\n",
686 pfn, nr_pages))
687 return -EINVAL;
688
689 bitmap_xor(subsection_map, map, subsection_map, SUBSECTIONS_PER_SECTION);
690 return 0;
691}
692
693static bool is_subsection_map_empty(struct mem_section *ms)
694{
695 return bitmap_empty(&ms->usage->subsection_map[0],
696 SUBSECTIONS_PER_SECTION);
697}
698
699static int fill_subsection_map(unsigned long pfn, unsigned long nr_pages)
700{
701 struct mem_section *ms = __pfn_to_section(pfn);
702 DECLARE_BITMAP(map, SUBSECTIONS_PER_SECTION) = { 0 };
703 unsigned long *subsection_map;
704 int rc = 0;
705
706 subsection_mask_set(map, pfn, nr_pages);
707
708 subsection_map = &ms->usage->subsection_map[0];
709
710 if (bitmap_empty(map, SUBSECTIONS_PER_SECTION))
711 rc = -EINVAL;
712 else if (bitmap_intersects(map, subsection_map, SUBSECTIONS_PER_SECTION))
713 rc = -EEXIST;
714 else
715 bitmap_or(subsection_map, map, subsection_map,
716 SUBSECTIONS_PER_SECTION);
717
718 return rc;
719}
720#else
721struct page * __meminit populate_section_memmap(unsigned long pfn,
722 unsigned long nr_pages, int nid, struct vmem_altmap *altmap)
723{
724 return kvmalloc_node(array_size(sizeof(struct page),
725 PAGES_PER_SECTION), GFP_KERNEL, nid);
726}
727
728static void depopulate_section_memmap(unsigned long pfn, unsigned long nr_pages,
729 struct vmem_altmap *altmap)
730{
731 kvfree(pfn_to_page(pfn));
732}
733
734static void free_map_bootmem(struct page *memmap)
735{
736 unsigned long maps_section_nr, removing_section_nr, i;
737 unsigned long magic, nr_pages;
738 struct page *page = virt_to_page(memmap);
739
740 nr_pages = PAGE_ALIGN(PAGES_PER_SECTION * sizeof(struct page))
741 >> PAGE_SHIFT;
742
743 for (i = 0; i < nr_pages; i++, page++) {
744 magic = (unsigned long) page->freelist;
745
746 BUG_ON(magic == NODE_INFO);
747
748 maps_section_nr = pfn_to_section_nr(page_to_pfn(page));
749 removing_section_nr = page_private(page);
750
751 /*
752 * When this function is called, the removing section is
753 * logical offlined state. This means all pages are isolated
754 * from page allocator. If removing section's memmap is placed
755 * on the same section, it must not be freed.
756 * If it is freed, page allocator may allocate it which will
757 * be removed physically soon.
758 */
759 if (maps_section_nr != removing_section_nr)
760 put_page_bootmem(page);
761 }
762}
763
764static int clear_subsection_map(unsigned long pfn, unsigned long nr_pages)
765{
766 return 0;
767}
768
769static bool is_subsection_map_empty(struct mem_section *ms)
770{
771 return true;
772}
773
774static int fill_subsection_map(unsigned long pfn, unsigned long nr_pages)
775{
776 return 0;
777}
778#endif /* CONFIG_SPARSEMEM_VMEMMAP */
779
780/*
781 * To deactivate a memory region, there are 3 cases to handle across
782 * two configurations (SPARSEMEM_VMEMMAP={y,n}):
783 *
784 * 1. deactivation of a partial hot-added section (only possible in
785 * the SPARSEMEM_VMEMMAP=y case).
786 * a) section was present at memory init.
787 * b) section was hot-added post memory init.
788 * 2. deactivation of a complete hot-added section.
789 * 3. deactivation of a complete section from memory init.
790 *
791 * For 1, when subsection_map does not empty we will not be freeing the
792 * usage map, but still need to free the vmemmap range.
793 *
794 * For 2 and 3, the SPARSEMEM_VMEMMAP={y,n} cases are unified
795 */
796static void section_deactivate(unsigned long pfn, unsigned long nr_pages,
797 struct vmem_altmap *altmap)
798{
799 struct mem_section *ms = __pfn_to_section(pfn);
800 bool section_is_early = early_section(ms);
801 struct page *memmap = NULL;
802 bool empty;
803
804 if (clear_subsection_map(pfn, nr_pages))
805 return;
806
807 empty = is_subsection_map_empty(ms);
808 if (empty) {
809 unsigned long section_nr = pfn_to_section_nr(pfn);
810
811 /*
812 * When removing an early section, the usage map is kept (as the
813 * usage maps of other sections fall into the same page). It
814 * will be re-used when re-adding the section - which is then no
815 * longer an early section. If the usage map is PageReserved, it
816 * was allocated during boot.
817 */
818 if (!PageReserved(virt_to_page(ms->usage))) {
819 kfree(ms->usage);
820 ms->usage = NULL;
821 }
822 memmap = sparse_decode_mem_map(ms->section_mem_map, section_nr);
823 /*
824 * Mark the section invalid so that valid_section()
825 * return false. This prevents code from dereferencing
826 * ms->usage array.
827 */
828 ms->section_mem_map &= ~SECTION_HAS_MEM_MAP;
829 }
830
831 /*
832 * The memmap of early sections is always fully populated. See
833 * section_activate() and pfn_valid() .
834 */
835 if (!section_is_early)
836 depopulate_section_memmap(pfn, nr_pages, altmap);
837 else if (memmap)
838 free_map_bootmem(memmap);
839
840 if (empty)
841 ms->section_mem_map = (unsigned long)NULL;
842}
843
844static struct page * __meminit section_activate(int nid, unsigned long pfn,
845 unsigned long nr_pages, struct vmem_altmap *altmap)
846{
847 struct mem_section *ms = __pfn_to_section(pfn);
848 struct mem_section_usage *usage = NULL;
849 struct page *memmap;
850 int rc = 0;
851
852 if (!ms->usage) {
853 usage = kzalloc(mem_section_usage_size(), GFP_KERNEL);
854 if (!usage)
855 return ERR_PTR(-ENOMEM);
856 ms->usage = usage;
857 }
858
859 rc = fill_subsection_map(pfn, nr_pages);
860 if (rc) {
861 if (usage)
862 ms->usage = NULL;
863 kfree(usage);
864 return ERR_PTR(rc);
865 }
866
867 /*
868 * The early init code does not consider partially populated
869 * initial sections, it simply assumes that memory will never be
870 * referenced. If we hot-add memory into such a section then we
871 * do not need to populate the memmap and can simply reuse what
872 * is already there.
873 */
874 if (nr_pages < PAGES_PER_SECTION && early_section(ms))
875 return pfn_to_page(pfn);
876
877 memmap = populate_section_memmap(pfn, nr_pages, nid, altmap);
878 if (!memmap) {
879 section_deactivate(pfn, nr_pages, altmap);
880 return ERR_PTR(-ENOMEM);
881 }
882
883 return memmap;
884}
885
886/**
887 * sparse_add_section - add a memory section, or populate an existing one
888 * @nid: The node to add section on
889 * @start_pfn: start pfn of the memory range
890 * @nr_pages: number of pfns to add in the section
891 * @altmap: device page map
892 *
893 * This is only intended for hotplug.
894 *
895 * Note that only VMEMMAP supports sub-section aligned hotplug,
896 * the proper alignment and size are gated by check_pfn_span().
897 *
898 *
899 * Return:
900 * * 0 - On success.
901 * * -EEXIST - Section has been present.
902 * * -ENOMEM - Out of memory.
903 */
904int __meminit sparse_add_section(int nid, unsigned long start_pfn,
905 unsigned long nr_pages, struct vmem_altmap *altmap)
906{
907 unsigned long section_nr = pfn_to_section_nr(start_pfn);
908 struct mem_section *ms;
909 struct page *memmap;
910 int ret;
911
912 ret = sparse_index_init(section_nr, nid);
913 if (ret < 0)
914 return ret;
915
916 memmap = section_activate(nid, start_pfn, nr_pages, altmap);
917 if (IS_ERR(memmap))
918 return PTR_ERR(memmap);
919
920 /*
921 * Poison uninitialized struct pages in order to catch invalid flags
922 * combinations.
923 */
924 page_init_poison(memmap, sizeof(struct page) * nr_pages);
925
926 ms = __nr_to_section(section_nr);
927 set_section_nid(section_nr, nid);
928 section_mark_present(ms);
929
930 /* Align memmap to section boundary in the subsection case */
931 if (section_nr_to_pfn(section_nr) != start_pfn)
932 memmap = pfn_to_page(section_nr_to_pfn(section_nr));
933 sparse_init_one_section(ms, section_nr, memmap, ms->usage, 0);
934
935 return 0;
936}
937
938#ifdef CONFIG_MEMORY_FAILURE
939static void clear_hwpoisoned_pages(struct page *memmap, int nr_pages)
940{
941 int i;
942
943 /*
944 * A further optimization is to have per section refcounted
945 * num_poisoned_pages. But that would need more space per memmap, so
946 * for now just do a quick global check to speed up this routine in the
947 * absence of bad pages.
948 */
949 if (atomic_long_read(&num_poisoned_pages) == 0)
950 return;
951
952 for (i = 0; i < nr_pages; i++) {
953 if (PageHWPoison(&memmap[i])) {
954 num_poisoned_pages_dec();
955 ClearPageHWPoison(&memmap[i]);
956 }
957 }
958}
959#else
960static inline void clear_hwpoisoned_pages(struct page *memmap, int nr_pages)
961{
962}
963#endif
964
965void sparse_remove_section(struct mem_section *ms, unsigned long pfn,
966 unsigned long nr_pages, unsigned long map_offset,
967 struct vmem_altmap *altmap)
968{
969 clear_hwpoisoned_pages(pfn_to_page(pfn) + map_offset,
970 nr_pages - map_offset);
971 section_deactivate(pfn, nr_pages, altmap);
972}
973#endif /* CONFIG_MEMORY_HOTPLUG */
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 pr_info("node %d must be removed before remove section %ld\n",
317 nid, usemap_snr);
318 return;
319 }
320 /*
321 * There is a circular dependency.
322 * Some platforms allow un-removable section because they will just
323 * gather other removable sections for dynamic partitioning.
324 * Just notify un-removable section's number here.
325 */
326 pr_info("Section %ld and %ld (node %d) have a circular dependency on usemap and pgdat allocations\n",
327 usemap_snr, pgdat_snr, nid);
328}
329#else
330static unsigned long * __init
331sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
332 unsigned long size)
333{
334 return memblock_virt_alloc_node_nopanic(size, pgdat->node_id);
335}
336
337static void __init check_usemap_section_nr(int nid, unsigned long *usemap)
338{
339}
340#endif /* CONFIG_MEMORY_HOTREMOVE */
341
342static void __init sparse_early_usemaps_alloc_node(void *data,
343 unsigned long pnum_begin,
344 unsigned long pnum_end,
345 unsigned long usemap_count, int nodeid)
346{
347 void *usemap;
348 unsigned long pnum;
349 unsigned long **usemap_map = (unsigned long **)data;
350 int size = usemap_size();
351
352 usemap = sparse_early_usemaps_alloc_pgdat_section(NODE_DATA(nodeid),
353 size * usemap_count);
354 if (!usemap) {
355 pr_warn("%s: allocation failed\n", __func__);
356 return;
357 }
358
359 for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
360 if (!present_section_nr(pnum))
361 continue;
362 usemap_map[pnum] = usemap;
363 usemap += size;
364 check_usemap_section_nr(nodeid, usemap_map[pnum]);
365 }
366}
367
368#ifndef CONFIG_SPARSEMEM_VMEMMAP
369struct page __init *sparse_mem_map_populate(unsigned long pnum, int nid)
370{
371 struct page *map;
372 unsigned long size;
373
374 map = alloc_remap(nid, sizeof(struct page) * PAGES_PER_SECTION);
375 if (map)
376 return map;
377
378 size = PAGE_ALIGN(sizeof(struct page) * PAGES_PER_SECTION);
379 map = memblock_virt_alloc_try_nid(size,
380 PAGE_SIZE, __pa(MAX_DMA_ADDRESS),
381 BOOTMEM_ALLOC_ACCESSIBLE, nid);
382 return map;
383}
384void __init sparse_mem_maps_populate_node(struct page **map_map,
385 unsigned long pnum_begin,
386 unsigned long pnum_end,
387 unsigned long map_count, int nodeid)
388{
389 void *map;
390 unsigned long pnum;
391 unsigned long size = sizeof(struct page) * PAGES_PER_SECTION;
392
393 map = alloc_remap(nodeid, size * map_count);
394 if (map) {
395 for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
396 if (!present_section_nr(pnum))
397 continue;
398 map_map[pnum] = map;
399 map += size;
400 }
401 return;
402 }
403
404 size = PAGE_ALIGN(size);
405 map = memblock_virt_alloc_try_nid(size * map_count,
406 PAGE_SIZE, __pa(MAX_DMA_ADDRESS),
407 BOOTMEM_ALLOC_ACCESSIBLE, nodeid);
408 if (map) {
409 for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
410 if (!present_section_nr(pnum))
411 continue;
412 map_map[pnum] = map;
413 map += size;
414 }
415 return;
416 }
417
418 /* fallback */
419 for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
420 struct mem_section *ms;
421
422 if (!present_section_nr(pnum))
423 continue;
424 map_map[pnum] = sparse_mem_map_populate(pnum, nodeid);
425 if (map_map[pnum])
426 continue;
427 ms = __nr_to_section(pnum);
428 pr_err("%s: sparsemem memory map backing failed some memory will not be available\n",
429 __func__);
430 ms->section_mem_map = 0;
431 }
432}
433#endif /* !CONFIG_SPARSEMEM_VMEMMAP */
434
435#ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
436static void __init sparse_early_mem_maps_alloc_node(void *data,
437 unsigned long pnum_begin,
438 unsigned long pnum_end,
439 unsigned long map_count, int nodeid)
440{
441 struct page **map_map = (struct page **)data;
442 sparse_mem_maps_populate_node(map_map, pnum_begin, pnum_end,
443 map_count, nodeid);
444}
445#else
446static struct page __init *sparse_early_mem_map_alloc(unsigned long pnum)
447{
448 struct page *map;
449 struct mem_section *ms = __nr_to_section(pnum);
450 int nid = sparse_early_nid(ms);
451
452 map = sparse_mem_map_populate(pnum, nid);
453 if (map)
454 return map;
455
456 pr_err("%s: sparsemem memory map backing failed some memory will not be available\n",
457 __func__);
458 ms->section_mem_map = 0;
459 return NULL;
460}
461#endif
462
463void __weak __meminit vmemmap_populate_print_last(void)
464{
465}
466
467/**
468 * alloc_usemap_and_memmap - memory alloction for pageblock flags and vmemmap
469 * @map: usemap_map for pageblock flags or mmap_map for vmemmap
470 */
471static void __init alloc_usemap_and_memmap(void (*alloc_func)
472 (void *, unsigned long, unsigned long,
473 unsigned long, int), void *data)
474{
475 unsigned long pnum;
476 unsigned long map_count;
477 int nodeid_begin = 0;
478 unsigned long pnum_begin = 0;
479
480 for (pnum = 0; pnum < NR_MEM_SECTIONS; pnum++) {
481 struct mem_section *ms;
482
483 if (!present_section_nr(pnum))
484 continue;
485 ms = __nr_to_section(pnum);
486 nodeid_begin = sparse_early_nid(ms);
487 pnum_begin = pnum;
488 break;
489 }
490 map_count = 1;
491 for (pnum = pnum_begin + 1; pnum < NR_MEM_SECTIONS; pnum++) {
492 struct mem_section *ms;
493 int nodeid;
494
495 if (!present_section_nr(pnum))
496 continue;
497 ms = __nr_to_section(pnum);
498 nodeid = sparse_early_nid(ms);
499 if (nodeid == nodeid_begin) {
500 map_count++;
501 continue;
502 }
503 /* ok, we need to take cake of from pnum_begin to pnum - 1*/
504 alloc_func(data, pnum_begin, pnum,
505 map_count, nodeid_begin);
506 /* new start, update count etc*/
507 nodeid_begin = nodeid;
508 pnum_begin = pnum;
509 map_count = 1;
510 }
511 /* ok, last chunk */
512 alloc_func(data, pnum_begin, NR_MEM_SECTIONS,
513 map_count, nodeid_begin);
514}
515
516/*
517 * Allocate the accumulated non-linear sections, allocate a mem_map
518 * for each and record the physical to section mapping.
519 */
520void __init sparse_init(void)
521{
522 unsigned long pnum;
523 struct page *map;
524 unsigned long *usemap;
525 unsigned long **usemap_map;
526 int size;
527#ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
528 int size2;
529 struct page **map_map;
530#endif
531
532 /* see include/linux/mmzone.h 'struct mem_section' definition */
533 BUILD_BUG_ON(!is_power_of_2(sizeof(struct mem_section)));
534
535 /* Setup pageblock_order for HUGETLB_PAGE_SIZE_VARIABLE */
536 set_pageblock_order();
537
538 /*
539 * map is using big page (aka 2M in x86 64 bit)
540 * usemap is less one page (aka 24 bytes)
541 * so alloc 2M (with 2M align) and 24 bytes in turn will
542 * make next 2M slip to one more 2M later.
543 * then in big system, the memory will have a lot of holes...
544 * here try to allocate 2M pages continuously.
545 *
546 * powerpc need to call sparse_init_one_section right after each
547 * sparse_early_mem_map_alloc, so allocate usemap_map at first.
548 */
549 size = sizeof(unsigned long *) * NR_MEM_SECTIONS;
550 usemap_map = memblock_virt_alloc(size, 0);
551 if (!usemap_map)
552 panic("can not allocate usemap_map\n");
553 alloc_usemap_and_memmap(sparse_early_usemaps_alloc_node,
554 (void *)usemap_map);
555
556#ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
557 size2 = sizeof(struct page *) * NR_MEM_SECTIONS;
558 map_map = memblock_virt_alloc(size2, 0);
559 if (!map_map)
560 panic("can not allocate map_map\n");
561 alloc_usemap_and_memmap(sparse_early_mem_maps_alloc_node,
562 (void *)map_map);
563#endif
564
565 for (pnum = 0; pnum < NR_MEM_SECTIONS; pnum++) {
566 if (!present_section_nr(pnum))
567 continue;
568
569 usemap = usemap_map[pnum];
570 if (!usemap)
571 continue;
572
573#ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
574 map = map_map[pnum];
575#else
576 map = sparse_early_mem_map_alloc(pnum);
577#endif
578 if (!map)
579 continue;
580
581 sparse_init_one_section(__nr_to_section(pnum), pnum, map,
582 usemap);
583 }
584
585 vmemmap_populate_print_last();
586
587#ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
588 memblock_free_early(__pa(map_map), size2);
589#endif
590 memblock_free_early(__pa(usemap_map), size);
591}
592
593#ifdef CONFIG_MEMORY_HOTPLUG
594#ifdef CONFIG_SPARSEMEM_VMEMMAP
595static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid)
596{
597 /* This will make the necessary allocations eventually. */
598 return sparse_mem_map_populate(pnum, nid);
599}
600static void __kfree_section_memmap(struct page *memmap)
601{
602 unsigned long start = (unsigned long)memmap;
603 unsigned long end = (unsigned long)(memmap + PAGES_PER_SECTION);
604
605 vmemmap_free(start, end);
606}
607#ifdef CONFIG_MEMORY_HOTREMOVE
608static void free_map_bootmem(struct page *memmap)
609{
610 unsigned long start = (unsigned long)memmap;
611 unsigned long end = (unsigned long)(memmap + PAGES_PER_SECTION);
612
613 vmemmap_free(start, end);
614}
615#endif /* CONFIG_MEMORY_HOTREMOVE */
616#else
617static struct page *__kmalloc_section_memmap(void)
618{
619 struct page *page, *ret;
620 unsigned long memmap_size = sizeof(struct page) * PAGES_PER_SECTION;
621
622 page = alloc_pages(GFP_KERNEL|__GFP_NOWARN, get_order(memmap_size));
623 if (page)
624 goto got_map_page;
625
626 ret = vmalloc(memmap_size);
627 if (ret)
628 goto got_map_ptr;
629
630 return NULL;
631got_map_page:
632 ret = (struct page *)pfn_to_kaddr(page_to_pfn(page));
633got_map_ptr:
634
635 return ret;
636}
637
638static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid)
639{
640 return __kmalloc_section_memmap();
641}
642
643static void __kfree_section_memmap(struct page *memmap)
644{
645 if (is_vmalloc_addr(memmap))
646 vfree(memmap);
647 else
648 free_pages((unsigned long)memmap,
649 get_order(sizeof(struct page) * PAGES_PER_SECTION));
650}
651
652#ifdef CONFIG_MEMORY_HOTREMOVE
653static void free_map_bootmem(struct page *memmap)
654{
655 unsigned long maps_section_nr, removing_section_nr, i;
656 unsigned long magic, nr_pages;
657 struct page *page = virt_to_page(memmap);
658
659 nr_pages = PAGE_ALIGN(PAGES_PER_SECTION * sizeof(struct page))
660 >> PAGE_SHIFT;
661
662 for (i = 0; i < nr_pages; i++, page++) {
663 magic = (unsigned long) page->lru.next;
664
665 BUG_ON(magic == NODE_INFO);
666
667 maps_section_nr = pfn_to_section_nr(page_to_pfn(page));
668 removing_section_nr = page->private;
669
670 /*
671 * When this function is called, the removing section is
672 * logical offlined state. This means all pages are isolated
673 * from page allocator. If removing section's memmap is placed
674 * on the same section, it must not be freed.
675 * If it is freed, page allocator may allocate it which will
676 * be removed physically soon.
677 */
678 if (maps_section_nr != removing_section_nr)
679 put_page_bootmem(page);
680 }
681}
682#endif /* CONFIG_MEMORY_HOTREMOVE */
683#endif /* CONFIG_SPARSEMEM_VMEMMAP */
684
685/*
686 * returns the number of sections whose mem_maps were properly
687 * set. If this is <=0, then that means that the passed-in
688 * map was not consumed and must be freed.
689 */
690int __meminit sparse_add_one_section(struct zone *zone, unsigned long start_pfn)
691{
692 unsigned long section_nr = pfn_to_section_nr(start_pfn);
693 struct pglist_data *pgdat = zone->zone_pgdat;
694 struct mem_section *ms;
695 struct page *memmap;
696 unsigned long *usemap;
697 unsigned long flags;
698 int ret;
699
700 /*
701 * no locking for this, because it does its own
702 * plus, it does a kmalloc
703 */
704 ret = sparse_index_init(section_nr, pgdat->node_id);
705 if (ret < 0 && ret != -EEXIST)
706 return ret;
707 memmap = kmalloc_section_memmap(section_nr, pgdat->node_id);
708 if (!memmap)
709 return -ENOMEM;
710 usemap = __kmalloc_section_usemap();
711 if (!usemap) {
712 __kfree_section_memmap(memmap);
713 return -ENOMEM;
714 }
715
716 pgdat_resize_lock(pgdat, &flags);
717
718 ms = __pfn_to_section(start_pfn);
719 if (ms->section_mem_map & SECTION_MARKED_PRESENT) {
720 ret = -EEXIST;
721 goto out;
722 }
723
724 memset(memmap, 0, sizeof(struct page) * PAGES_PER_SECTION);
725
726 ms->section_mem_map |= SECTION_MARKED_PRESENT;
727
728 ret = sparse_init_one_section(ms, section_nr, memmap, usemap);
729
730out:
731 pgdat_resize_unlock(pgdat, &flags);
732 if (ret <= 0) {
733 kfree(usemap);
734 __kfree_section_memmap(memmap);
735 }
736 return ret;
737}
738
739#ifdef CONFIG_MEMORY_HOTREMOVE
740#ifdef CONFIG_MEMORY_FAILURE
741static void clear_hwpoisoned_pages(struct page *memmap, int nr_pages)
742{
743 int i;
744
745 if (!memmap)
746 return;
747
748 for (i = 0; i < nr_pages; i++) {
749 if (PageHWPoison(&memmap[i])) {
750 atomic_long_sub(1, &num_poisoned_pages);
751 ClearPageHWPoison(&memmap[i]);
752 }
753 }
754}
755#else
756static inline void clear_hwpoisoned_pages(struct page *memmap, int nr_pages)
757{
758}
759#endif
760
761static void free_section_usemap(struct page *memmap, unsigned long *usemap)
762{
763 struct page *usemap_page;
764
765 if (!usemap)
766 return;
767
768 usemap_page = virt_to_page(usemap);
769 /*
770 * Check to see if allocation came from hot-plug-add
771 */
772 if (PageSlab(usemap_page) || PageCompound(usemap_page)) {
773 kfree(usemap);
774 if (memmap)
775 __kfree_section_memmap(memmap);
776 return;
777 }
778
779 /*
780 * The usemap came from bootmem. This is packed with other usemaps
781 * on the section which has pgdat at boot time. Just keep it as is now.
782 */
783
784 if (memmap)
785 free_map_bootmem(memmap);
786}
787
788void sparse_remove_one_section(struct zone *zone, struct mem_section *ms,
789 unsigned long map_offset)
790{
791 struct page *memmap = NULL;
792 unsigned long *usemap = NULL, flags;
793 struct pglist_data *pgdat = zone->zone_pgdat;
794
795 pgdat_resize_lock(pgdat, &flags);
796 if (ms->section_mem_map) {
797 usemap = ms->pageblock_flags;
798 memmap = sparse_decode_mem_map(ms->section_mem_map,
799 __section_nr(ms));
800 ms->section_mem_map = 0;
801 ms->pageblock_flags = NULL;
802 }
803 pgdat_resize_unlock(pgdat, &flags);
804
805 clear_hwpoisoned_pages(memmap + map_offset,
806 PAGES_PER_SECTION - map_offset);
807 free_section_usemap(memmap, usemap);
808}
809#endif /* CONFIG_MEMORY_HOTREMOVE */
810#endif /* CONFIG_MEMORY_HOTPLUG */