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#include <asm/pgalloc.h>
20#include <asm/pgtable.h>
21
22/*
23 * Permanent SPARSEMEM data:
24 *
25 * 1) mem_section - memory sections, mem_map's for valid memory
26 */
27#ifdef CONFIG_SPARSEMEM_EXTREME
28struct mem_section **mem_section;
29#else
30struct mem_section mem_section[NR_SECTION_ROOTS][SECTIONS_PER_ROOT]
31 ____cacheline_internodealigned_in_smp;
32#endif
33EXPORT_SYMBOL(mem_section);
34
35#ifdef NODE_NOT_IN_PAGE_FLAGS
36/*
37 * If we did not store the node number in the page then we have to
38 * do a lookup in the section_to_node_table in order to find which
39 * node the page belongs to.
40 */
41#if MAX_NUMNODES <= 256
42static u8 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned;
43#else
44static u16 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned;
45#endif
46
47int page_to_nid(const struct page *page)
48{
49 return section_to_node_table[page_to_section(page)];
50}
51EXPORT_SYMBOL(page_to_nid);
52
53static void set_section_nid(unsigned long section_nr, int nid)
54{
55 section_to_node_table[section_nr] = nid;
56}
57#else /* !NODE_NOT_IN_PAGE_FLAGS */
58static inline void set_section_nid(unsigned long section_nr, int nid)
59{
60}
61#endif
62
63#ifdef CONFIG_SPARSEMEM_EXTREME
64static noinline struct mem_section __ref *sparse_index_alloc(int nid)
65{
66 struct mem_section *section = NULL;
67 unsigned long array_size = SECTIONS_PER_ROOT *
68 sizeof(struct mem_section);
69
70 if (slab_is_available()) {
71 section = kzalloc_node(array_size, GFP_KERNEL, nid);
72 } else {
73 section = memblock_alloc_node(array_size, SMP_CACHE_BYTES,
74 nid);
75 if (!section)
76 panic("%s: Failed to allocate %lu bytes nid=%d\n",
77 __func__, array_size, nid);
78 }
79
80 return section;
81}
82
83static int __meminit sparse_index_init(unsigned long section_nr, int nid)
84{
85 unsigned long root = SECTION_NR_TO_ROOT(section_nr);
86 struct mem_section *section;
87
88 /*
89 * An existing section is possible in the sub-section hotplug
90 * case. First hot-add instantiates, follow-on hot-add reuses
91 * the existing section.
92 *
93 * The mem_hotplug_lock resolves the apparent race below.
94 */
95 if (mem_section[root])
96 return 0;
97
98 section = sparse_index_alloc(nid);
99 if (!section)
100 return -ENOMEM;
101
102 mem_section[root] = section;
103
104 return 0;
105}
106#else /* !SPARSEMEM_EXTREME */
107static inline int sparse_index_init(unsigned long section_nr, int nid)
108{
109 return 0;
110}
111#endif
112
113#ifdef CONFIG_SPARSEMEM_EXTREME
114unsigned long __section_nr(struct mem_section *ms)
115{
116 unsigned long root_nr;
117 struct mem_section *root = NULL;
118
119 for (root_nr = 0; root_nr < NR_SECTION_ROOTS; root_nr++) {
120 root = __nr_to_section(root_nr * SECTIONS_PER_ROOT);
121 if (!root)
122 continue;
123
124 if ((ms >= root) && (ms < (root + SECTIONS_PER_ROOT)))
125 break;
126 }
127
128 VM_BUG_ON(!root);
129
130 return (root_nr * SECTIONS_PER_ROOT) + (ms - root);
131}
132#else
133unsigned long __section_nr(struct mem_section *ms)
134{
135 return (unsigned long)(ms - mem_section[0]);
136}
137#endif
138
139/*
140 * During early boot, before section_mem_map is used for an actual
141 * mem_map, we use section_mem_map to store the section's NUMA
142 * node. This keeps us from having to use another data structure. The
143 * node information is cleared just before we store the real mem_map.
144 */
145static inline unsigned long sparse_encode_early_nid(int nid)
146{
147 return (nid << SECTION_NID_SHIFT);
148}
149
150static inline int sparse_early_nid(struct mem_section *section)
151{
152 return (section->section_mem_map >> SECTION_NID_SHIFT);
153}
154
155/* Validate the physical addressing limitations of the model */
156void __meminit mminit_validate_memmodel_limits(unsigned long *start_pfn,
157 unsigned long *end_pfn)
158{
159 unsigned long max_sparsemem_pfn = 1UL << (MAX_PHYSMEM_BITS-PAGE_SHIFT);
160
161 /*
162 * Sanity checks - do not allow an architecture to pass
163 * in larger pfns than the maximum scope of sparsemem:
164 */
165 if (*start_pfn > max_sparsemem_pfn) {
166 mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
167 "Start of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
168 *start_pfn, *end_pfn, max_sparsemem_pfn);
169 WARN_ON_ONCE(1);
170 *start_pfn = max_sparsemem_pfn;
171 *end_pfn = max_sparsemem_pfn;
172 } else if (*end_pfn > max_sparsemem_pfn) {
173 mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
174 "End of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
175 *start_pfn, *end_pfn, max_sparsemem_pfn);
176 WARN_ON_ONCE(1);
177 *end_pfn = max_sparsemem_pfn;
178 }
179}
180
181/*
182 * There are a number of times that we loop over NR_MEM_SECTIONS,
183 * looking for section_present() on each. But, when we have very
184 * large physical address spaces, NR_MEM_SECTIONS can also be
185 * very large which makes the loops quite long.
186 *
187 * Keeping track of this gives us an easy way to break out of
188 * those loops early.
189 */
190unsigned long __highest_present_section_nr;
191static void section_mark_present(struct mem_section *ms)
192{
193 unsigned long section_nr = __section_nr(ms);
194
195 if (section_nr > __highest_present_section_nr)
196 __highest_present_section_nr = section_nr;
197
198 ms->section_mem_map |= SECTION_MARKED_PRESENT;
199}
200
201static inline unsigned long next_present_section_nr(unsigned long section_nr)
202{
203 do {
204 section_nr++;
205 if (present_section_nr(section_nr))
206 return section_nr;
207 } while ((section_nr <= __highest_present_section_nr));
208
209 return -1;
210}
211#define for_each_present_section_nr(start, section_nr) \
212 for (section_nr = next_present_section_nr(start-1); \
213 ((section_nr != -1) && \
214 (section_nr <= __highest_present_section_nr)); \
215 section_nr = next_present_section_nr(section_nr))
216
217static inline unsigned long first_present_section_nr(void)
218{
219 return next_present_section_nr(-1);
220}
221
222static void subsection_mask_set(unsigned long *map, unsigned long pfn,
223 unsigned long nr_pages)
224{
225 int idx = subsection_map_index(pfn);
226 int end = subsection_map_index(pfn + nr_pages - 1);
227
228 bitmap_set(map, idx, end - idx + 1);
229}
230
231void __init subsection_map_init(unsigned long pfn, unsigned long nr_pages)
232{
233 int end_sec = pfn_to_section_nr(pfn + nr_pages - 1);
234 unsigned long nr, start_sec = pfn_to_section_nr(pfn);
235
236 if (!nr_pages)
237 return;
238
239 for (nr = start_sec; nr <= end_sec; nr++) {
240 struct mem_section *ms;
241 unsigned long pfns;
242
243 pfns = min(nr_pages, PAGES_PER_SECTION
244 - (pfn & ~PAGE_SECTION_MASK));
245 ms = __nr_to_section(nr);
246 subsection_mask_set(ms->usage->subsection_map, pfn, pfns);
247
248 pr_debug("%s: sec: %lu pfns: %lu set(%d, %d)\n", __func__, nr,
249 pfns, subsection_map_index(pfn),
250 subsection_map_index(pfn + pfns - 1));
251
252 pfn += pfns;
253 nr_pages -= pfns;
254 }
255}
256
257/* Record a memory area against a node. */
258void __init memory_present(int nid, unsigned long start, unsigned long end)
259{
260 unsigned long pfn;
261
262#ifdef CONFIG_SPARSEMEM_EXTREME
263 if (unlikely(!mem_section)) {
264 unsigned long size, align;
265
266 size = sizeof(struct mem_section*) * NR_SECTION_ROOTS;
267 align = 1 << (INTERNODE_CACHE_SHIFT);
268 mem_section = memblock_alloc(size, align);
269 if (!mem_section)
270 panic("%s: Failed to allocate %lu bytes align=0x%lx\n",
271 __func__, size, align);
272 }
273#endif
274
275 start &= PAGE_SECTION_MASK;
276 mminit_validate_memmodel_limits(&start, &end);
277 for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) {
278 unsigned long section = pfn_to_section_nr(pfn);
279 struct mem_section *ms;
280
281 sparse_index_init(section, nid);
282 set_section_nid(section, nid);
283
284 ms = __nr_to_section(section);
285 if (!ms->section_mem_map) {
286 ms->section_mem_map = sparse_encode_early_nid(nid) |
287 SECTION_IS_ONLINE;
288 section_mark_present(ms);
289 }
290 }
291}
292
293/*
294 * Mark all memblocks as present using memory_present(). This is a
295 * convienence function that is useful for a number of arches
296 * to mark all of the systems memory as present during initialization.
297 */
298void __init memblocks_present(void)
299{
300 struct memblock_region *reg;
301
302 for_each_memblock(memory, reg) {
303 memory_present(memblock_get_region_node(reg),
304 memblock_region_memory_base_pfn(reg),
305 memblock_region_memory_end_pfn(reg));
306 }
307}
308
309/*
310 * Subtle, we encode the real pfn into the mem_map such that
311 * the identity pfn - section_mem_map will return the actual
312 * physical page frame number.
313 */
314static unsigned long sparse_encode_mem_map(struct page *mem_map, unsigned long pnum)
315{
316 unsigned long coded_mem_map =
317 (unsigned long)(mem_map - (section_nr_to_pfn(pnum)));
318 BUILD_BUG_ON(SECTION_MAP_LAST_BIT > (1UL<<PFN_SECTION_SHIFT));
319 BUG_ON(coded_mem_map & ~SECTION_MAP_MASK);
320 return coded_mem_map;
321}
322
323/*
324 * Decode mem_map from the coded memmap
325 */
326struct page *sparse_decode_mem_map(unsigned long coded_mem_map, unsigned long pnum)
327{
328 /* mask off the extra low bits of information */
329 coded_mem_map &= SECTION_MAP_MASK;
330 return ((struct page *)coded_mem_map) + section_nr_to_pfn(pnum);
331}
332
333static void __meminit sparse_init_one_section(struct mem_section *ms,
334 unsigned long pnum, struct page *mem_map,
335 struct mem_section_usage *usage, unsigned long flags)
336{
337 ms->section_mem_map &= ~SECTION_MAP_MASK;
338 ms->section_mem_map |= sparse_encode_mem_map(mem_map, pnum)
339 | SECTION_HAS_MEM_MAP | flags;
340 ms->usage = usage;
341}
342
343static unsigned long usemap_size(void)
344{
345 return BITS_TO_LONGS(SECTION_BLOCKFLAGS_BITS) * sizeof(unsigned long);
346}
347
348size_t mem_section_usage_size(void)
349{
350 return sizeof(struct mem_section_usage) + usemap_size();
351}
352
353#ifdef CONFIG_MEMORY_HOTREMOVE
354static struct mem_section_usage * __init
355sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
356 unsigned long size)
357{
358 struct mem_section_usage *usage;
359 unsigned long goal, limit;
360 int nid;
361 /*
362 * A page may contain usemaps for other sections preventing the
363 * page being freed and making a section unremovable while
364 * other sections referencing the usemap remain active. Similarly,
365 * a pgdat can prevent a section being removed. If section A
366 * contains a pgdat and section B contains the usemap, both
367 * sections become inter-dependent. This allocates usemaps
368 * from the same section as the pgdat where possible to avoid
369 * this problem.
370 */
371 goal = __pa(pgdat) & (PAGE_SECTION_MASK << PAGE_SHIFT);
372 limit = goal + (1UL << PA_SECTION_SHIFT);
373 nid = early_pfn_to_nid(goal >> PAGE_SHIFT);
374again:
375 usage = memblock_alloc_try_nid(size, SMP_CACHE_BYTES, goal, limit, nid);
376 if (!usage && limit) {
377 limit = 0;
378 goto again;
379 }
380 return usage;
381}
382
383static void __init check_usemap_section_nr(int nid,
384 struct mem_section_usage *usage)
385{
386 unsigned long usemap_snr, pgdat_snr;
387 static unsigned long old_usemap_snr;
388 static unsigned long old_pgdat_snr;
389 struct pglist_data *pgdat = NODE_DATA(nid);
390 int usemap_nid;
391
392 /* First call */
393 if (!old_usemap_snr) {
394 old_usemap_snr = NR_MEM_SECTIONS;
395 old_pgdat_snr = NR_MEM_SECTIONS;
396 }
397
398 usemap_snr = pfn_to_section_nr(__pa(usage) >> PAGE_SHIFT);
399 pgdat_snr = pfn_to_section_nr(__pa(pgdat) >> PAGE_SHIFT);
400 if (usemap_snr == pgdat_snr)
401 return;
402
403 if (old_usemap_snr == usemap_snr && old_pgdat_snr == pgdat_snr)
404 /* skip redundant message */
405 return;
406
407 old_usemap_snr = usemap_snr;
408 old_pgdat_snr = pgdat_snr;
409
410 usemap_nid = sparse_early_nid(__nr_to_section(usemap_snr));
411 if (usemap_nid != nid) {
412 pr_info("node %d must be removed before remove section %ld\n",
413 nid, usemap_snr);
414 return;
415 }
416 /*
417 * There is a circular dependency.
418 * Some platforms allow un-removable section because they will just
419 * gather other removable sections for dynamic partitioning.
420 * Just notify un-removable section's number here.
421 */
422 pr_info("Section %ld and %ld (node %d) have a circular dependency on usemap and pgdat allocations\n",
423 usemap_snr, pgdat_snr, nid);
424}
425#else
426static struct mem_section_usage * __init
427sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
428 unsigned long size)
429{
430 return memblock_alloc_node(size, SMP_CACHE_BYTES, pgdat->node_id);
431}
432
433static void __init check_usemap_section_nr(int nid,
434 struct mem_section_usage *usage)
435{
436}
437#endif /* CONFIG_MEMORY_HOTREMOVE */
438
439#ifdef CONFIG_SPARSEMEM_VMEMMAP
440static unsigned long __init section_map_size(void)
441{
442 return ALIGN(sizeof(struct page) * PAGES_PER_SECTION, PMD_SIZE);
443}
444
445#else
446static unsigned long __init section_map_size(void)
447{
448 return PAGE_ALIGN(sizeof(struct page) * PAGES_PER_SECTION);
449}
450
451struct page __init *__populate_section_memmap(unsigned long pfn,
452 unsigned long nr_pages, int nid, struct vmem_altmap *altmap)
453{
454 unsigned long size = section_map_size();
455 struct page *map = sparse_buffer_alloc(size);
456 phys_addr_t addr = __pa(MAX_DMA_ADDRESS);
457
458 if (map)
459 return map;
460
461 map = memblock_alloc_try_nid(size,
462 PAGE_SIZE, addr,
463 MEMBLOCK_ALLOC_ACCESSIBLE, nid);
464 if (!map)
465 panic("%s: Failed to allocate %lu bytes align=0x%lx nid=%d from=%pa\n",
466 __func__, size, PAGE_SIZE, nid, &addr);
467
468 return map;
469}
470#endif /* !CONFIG_SPARSEMEM_VMEMMAP */
471
472static void *sparsemap_buf __meminitdata;
473static void *sparsemap_buf_end __meminitdata;
474
475static inline void __meminit sparse_buffer_free(unsigned long size)
476{
477 WARN_ON(!sparsemap_buf || size == 0);
478 memblock_free_early(__pa(sparsemap_buf), size);
479}
480
481static void __init sparse_buffer_init(unsigned long size, int nid)
482{
483 phys_addr_t addr = __pa(MAX_DMA_ADDRESS);
484 WARN_ON(sparsemap_buf); /* forgot to call sparse_buffer_fini()? */
485 sparsemap_buf =
486 memblock_alloc_try_nid_raw(size, PAGE_SIZE,
487 addr,
488 MEMBLOCK_ALLOC_ACCESSIBLE, nid);
489 sparsemap_buf_end = sparsemap_buf + size;
490}
491
492static void __init sparse_buffer_fini(void)
493{
494 unsigned long size = sparsemap_buf_end - sparsemap_buf;
495
496 if (sparsemap_buf && size > 0)
497 sparse_buffer_free(size);
498 sparsemap_buf = NULL;
499}
500
501void * __meminit sparse_buffer_alloc(unsigned long size)
502{
503 void *ptr = NULL;
504
505 if (sparsemap_buf) {
506 ptr = (void *) roundup((unsigned long)sparsemap_buf, size);
507 if (ptr + size > sparsemap_buf_end)
508 ptr = NULL;
509 else {
510 /* Free redundant aligned space */
511 if ((unsigned long)(ptr - sparsemap_buf) > 0)
512 sparse_buffer_free((unsigned long)(ptr - sparsemap_buf));
513 sparsemap_buf = ptr + size;
514 }
515 }
516 return ptr;
517}
518
519void __weak __meminit vmemmap_populate_print_last(void)
520{
521}
522
523/*
524 * Initialize sparse on a specific node. The node spans [pnum_begin, pnum_end)
525 * And number of present sections in this node is map_count.
526 */
527static void __init sparse_init_nid(int nid, unsigned long pnum_begin,
528 unsigned long pnum_end,
529 unsigned long map_count)
530{
531 struct mem_section_usage *usage;
532 unsigned long pnum;
533 struct page *map;
534
535 usage = sparse_early_usemaps_alloc_pgdat_section(NODE_DATA(nid),
536 mem_section_usage_size() * map_count);
537 if (!usage) {
538 pr_err("%s: node[%d] usemap allocation failed", __func__, nid);
539 goto failed;
540 }
541 sparse_buffer_init(map_count * section_map_size(), nid);
542 for_each_present_section_nr(pnum_begin, pnum) {
543 unsigned long pfn = section_nr_to_pfn(pnum);
544
545 if (pnum >= pnum_end)
546 break;
547
548 map = __populate_section_memmap(pfn, PAGES_PER_SECTION,
549 nid, NULL);
550 if (!map) {
551 pr_err("%s: node[%d] memory map backing failed. Some memory will not be available.",
552 __func__, nid);
553 pnum_begin = pnum;
554 goto failed;
555 }
556 check_usemap_section_nr(nid, usage);
557 sparse_init_one_section(__nr_to_section(pnum), pnum, map, usage,
558 SECTION_IS_EARLY);
559 usage = (void *) usage + mem_section_usage_size();
560 }
561 sparse_buffer_fini();
562 return;
563failed:
564 /* We failed to allocate, mark all the following pnums as not present */
565 for_each_present_section_nr(pnum_begin, pnum) {
566 struct mem_section *ms;
567
568 if (pnum >= pnum_end)
569 break;
570 ms = __nr_to_section(pnum);
571 ms->section_mem_map = 0;
572 }
573}
574
575/*
576 * Allocate the accumulated non-linear sections, allocate a mem_map
577 * for each and record the physical to section mapping.
578 */
579void __init sparse_init(void)
580{
581 unsigned long pnum_begin = first_present_section_nr();
582 int nid_begin = sparse_early_nid(__nr_to_section(pnum_begin));
583 unsigned long pnum_end, map_count = 1;
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 *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#else
672struct page *populate_section_memmap(unsigned long pfn,
673 unsigned long nr_pages, int nid, struct vmem_altmap *altmap)
674{
675 struct page *page, *ret;
676 unsigned long memmap_size = sizeof(struct page) * PAGES_PER_SECTION;
677
678 page = alloc_pages(GFP_KERNEL|__GFP_NOWARN, get_order(memmap_size));
679 if (page)
680 goto got_map_page;
681
682 ret = vmalloc(memmap_size);
683 if (ret)
684 goto got_map_ptr;
685
686 return NULL;
687got_map_page:
688 ret = (struct page *)pfn_to_kaddr(page_to_pfn(page));
689got_map_ptr:
690
691 return ret;
692}
693
694static void depopulate_section_memmap(unsigned long pfn, unsigned long nr_pages,
695 struct vmem_altmap *altmap)
696{
697 struct page *memmap = pfn_to_page(pfn);
698
699 if (is_vmalloc_addr(memmap))
700 vfree(memmap);
701 else
702 free_pages((unsigned long)memmap,
703 get_order(sizeof(struct page) * PAGES_PER_SECTION));
704}
705
706static void free_map_bootmem(struct page *memmap)
707{
708 unsigned long maps_section_nr, removing_section_nr, i;
709 unsigned long magic, nr_pages;
710 struct page *page = virt_to_page(memmap);
711
712 nr_pages = PAGE_ALIGN(PAGES_PER_SECTION * sizeof(struct page))
713 >> PAGE_SHIFT;
714
715 for (i = 0; i < nr_pages; i++, page++) {
716 magic = (unsigned long) page->freelist;
717
718 BUG_ON(magic == NODE_INFO);
719
720 maps_section_nr = pfn_to_section_nr(page_to_pfn(page));
721 removing_section_nr = page_private(page);
722
723 /*
724 * When this function is called, the removing section is
725 * logical offlined state. This means all pages are isolated
726 * from page allocator. If removing section's memmap is placed
727 * on the same section, it must not be freed.
728 * If it is freed, page allocator may allocate it which will
729 * be removed physically soon.
730 */
731 if (maps_section_nr != removing_section_nr)
732 put_page_bootmem(page);
733 }
734}
735#endif /* CONFIG_SPARSEMEM_VMEMMAP */
736
737static void section_deactivate(unsigned long pfn, unsigned long nr_pages,
738 struct vmem_altmap *altmap)
739{
740 DECLARE_BITMAP(map, SUBSECTIONS_PER_SECTION) = { 0 };
741 DECLARE_BITMAP(tmp, SUBSECTIONS_PER_SECTION) = { 0 };
742 struct mem_section *ms = __pfn_to_section(pfn);
743 bool section_is_early = early_section(ms);
744 struct page *memmap = NULL;
745 unsigned long *subsection_map = ms->usage
746 ? &ms->usage->subsection_map[0] : NULL;
747
748 subsection_mask_set(map, pfn, nr_pages);
749 if (subsection_map)
750 bitmap_and(tmp, map, subsection_map, SUBSECTIONS_PER_SECTION);
751
752 if (WARN(!subsection_map || !bitmap_equal(tmp, map, SUBSECTIONS_PER_SECTION),
753 "section already deactivated (%#lx + %ld)\n",
754 pfn, nr_pages))
755 return;
756
757 /*
758 * There are 3 cases to handle across two configurations
759 * (SPARSEMEM_VMEMMAP={y,n}):
760 *
761 * 1/ deactivation of a partial hot-added section (only possible
762 * in the SPARSEMEM_VMEMMAP=y case).
763 * a/ section was present at memory init
764 * b/ section was hot-added post memory init
765 * 2/ deactivation of a complete hot-added section
766 * 3/ deactivation of a complete section from memory init
767 *
768 * For 1/, when subsection_map does not empty we will not be
769 * freeing the usage map, but still need to free the vmemmap
770 * range.
771 *
772 * For 2/ and 3/ the SPARSEMEM_VMEMMAP={y,n} cases are unified
773 */
774 bitmap_xor(subsection_map, map, subsection_map, SUBSECTIONS_PER_SECTION);
775 if (bitmap_empty(subsection_map, SUBSECTIONS_PER_SECTION)) {
776 unsigned long section_nr = pfn_to_section_nr(pfn);
777
778 if (!section_is_early) {
779 kfree(ms->usage);
780 ms->usage = NULL;
781 }
782 memmap = sparse_decode_mem_map(ms->section_mem_map, section_nr);
783 ms->section_mem_map = sparse_encode_mem_map(NULL, section_nr);
784 }
785
786 if (section_is_early && memmap)
787 free_map_bootmem(memmap);
788 else
789 depopulate_section_memmap(pfn, nr_pages, altmap);
790}
791
792static struct page * __meminit section_activate(int nid, unsigned long pfn,
793 unsigned long nr_pages, struct vmem_altmap *altmap)
794{
795 DECLARE_BITMAP(map, SUBSECTIONS_PER_SECTION) = { 0 };
796 struct mem_section *ms = __pfn_to_section(pfn);
797 struct mem_section_usage *usage = NULL;
798 unsigned long *subsection_map;
799 struct page *memmap;
800 int rc = 0;
801
802 subsection_mask_set(map, pfn, nr_pages);
803
804 if (!ms->usage) {
805 usage = kzalloc(mem_section_usage_size(), GFP_KERNEL);
806 if (!usage)
807 return ERR_PTR(-ENOMEM);
808 ms->usage = usage;
809 }
810 subsection_map = &ms->usage->subsection_map[0];
811
812 if (bitmap_empty(map, SUBSECTIONS_PER_SECTION))
813 rc = -EINVAL;
814 else if (bitmap_intersects(map, subsection_map, SUBSECTIONS_PER_SECTION))
815 rc = -EEXIST;
816 else
817 bitmap_or(subsection_map, map, subsection_map,
818 SUBSECTIONS_PER_SECTION);
819
820 if (rc) {
821 if (usage)
822 ms->usage = NULL;
823 kfree(usage);
824 return ERR_PTR(rc);
825 }
826
827 /*
828 * The early init code does not consider partially populated
829 * initial sections, it simply assumes that memory will never be
830 * referenced. If we hot-add memory into such a section then we
831 * do not need to populate the memmap and can simply reuse what
832 * is already there.
833 */
834 if (nr_pages < PAGES_PER_SECTION && early_section(ms))
835 return pfn_to_page(pfn);
836
837 memmap = populate_section_memmap(pfn, nr_pages, nid, altmap);
838 if (!memmap) {
839 section_deactivate(pfn, nr_pages, altmap);
840 return ERR_PTR(-ENOMEM);
841 }
842
843 return memmap;
844}
845
846/**
847 * sparse_add_section - add a memory section, or populate an existing one
848 * @nid: The node to add section on
849 * @start_pfn: start pfn of the memory range
850 * @nr_pages: number of pfns to add in the section
851 * @altmap: device page map
852 *
853 * This is only intended for hotplug.
854 *
855 * Return:
856 * * 0 - On success.
857 * * -EEXIST - Section has been present.
858 * * -ENOMEM - Out of memory.
859 */
860int __meminit sparse_add_section(int nid, unsigned long start_pfn,
861 unsigned long nr_pages, struct vmem_altmap *altmap)
862{
863 unsigned long section_nr = pfn_to_section_nr(start_pfn);
864 struct mem_section *ms;
865 struct page *memmap;
866 int ret;
867
868 ret = sparse_index_init(section_nr, nid);
869 if (ret < 0)
870 return ret;
871
872 memmap = section_activate(nid, start_pfn, nr_pages, altmap);
873 if (IS_ERR(memmap))
874 return PTR_ERR(memmap);
875
876 /*
877 * Poison uninitialized struct pages in order to catch invalid flags
878 * combinations.
879 */
880 page_init_poison(pfn_to_page(start_pfn), sizeof(struct page) * nr_pages);
881
882 ms = __nr_to_section(section_nr);
883 set_section_nid(section_nr, nid);
884 section_mark_present(ms);
885
886 /* Align memmap to section boundary in the subsection case */
887 if (section_nr_to_pfn(section_nr) != start_pfn)
888 memmap = pfn_to_kaddr(section_nr_to_pfn(section_nr));
889 sparse_init_one_section(ms, section_nr, memmap, ms->usage, 0);
890
891 return 0;
892}
893
894#ifdef CONFIG_MEMORY_FAILURE
895static void clear_hwpoisoned_pages(struct page *memmap, int nr_pages)
896{
897 int i;
898
899 /*
900 * A further optimization is to have per section refcounted
901 * num_poisoned_pages. But that would need more space per memmap, so
902 * for now just do a quick global check to speed up this routine in the
903 * absence of bad pages.
904 */
905 if (atomic_long_read(&num_poisoned_pages) == 0)
906 return;
907
908 for (i = 0; i < nr_pages; i++) {
909 if (PageHWPoison(&memmap[i])) {
910 num_poisoned_pages_dec();
911 ClearPageHWPoison(&memmap[i]);
912 }
913 }
914}
915#else
916static inline void clear_hwpoisoned_pages(struct page *memmap, int nr_pages)
917{
918}
919#endif
920
921void sparse_remove_section(struct mem_section *ms, unsigned long pfn,
922 unsigned long nr_pages, unsigned long map_offset,
923 struct vmem_altmap *altmap)
924{
925 clear_hwpoisoned_pages(pfn_to_page(pfn) + map_offset,
926 nr_pages - map_offset);
927 section_deactivate(pfn, nr_pages, altmap);
928}
929#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 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 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#ifdef CONFIG_SPARSEMEM_EXTREME
104int __section_nr(struct mem_section* ms)
105{
106 unsigned long root_nr;
107 struct mem_section* root;
108
109 for (root_nr = 0; root_nr < NR_SECTION_ROOTS; root_nr++) {
110 root = __nr_to_section(root_nr * SECTIONS_PER_ROOT);
111 if (!root)
112 continue;
113
114 if ((ms >= root) && (ms < (root + SECTIONS_PER_ROOT)))
115 break;
116 }
117
118 VM_BUG_ON(root_nr == NR_SECTION_ROOTS);
119
120 return (root_nr * SECTIONS_PER_ROOT) + (ms - root);
121}
122#else
123int __section_nr(struct mem_section* ms)
124{
125 return (int)(ms - mem_section[0]);
126}
127#endif
128
129/*
130 * During early boot, before section_mem_map is used for an actual
131 * mem_map, we use section_mem_map to store the section's NUMA
132 * node. This keeps us from having to use another data structure. The
133 * node information is cleared just before we store the real mem_map.
134 */
135static inline unsigned long sparse_encode_early_nid(int nid)
136{
137 return (nid << SECTION_NID_SHIFT);
138}
139
140static inline int sparse_early_nid(struct mem_section *section)
141{
142 return (section->section_mem_map >> SECTION_NID_SHIFT);
143}
144
145/* Validate the physical addressing limitations of the model */
146void __meminit mminit_validate_memmodel_limits(unsigned long *start_pfn,
147 unsigned long *end_pfn)
148{
149 unsigned long max_sparsemem_pfn = 1UL << (MAX_PHYSMEM_BITS-PAGE_SHIFT);
150
151 /*
152 * Sanity checks - do not allow an architecture to pass
153 * in larger pfns than the maximum scope of sparsemem:
154 */
155 if (*start_pfn > max_sparsemem_pfn) {
156 mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
157 "Start of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
158 *start_pfn, *end_pfn, max_sparsemem_pfn);
159 WARN_ON_ONCE(1);
160 *start_pfn = max_sparsemem_pfn;
161 *end_pfn = max_sparsemem_pfn;
162 } else if (*end_pfn > max_sparsemem_pfn) {
163 mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
164 "End of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
165 *start_pfn, *end_pfn, max_sparsemem_pfn);
166 WARN_ON_ONCE(1);
167 *end_pfn = max_sparsemem_pfn;
168 }
169}
170
171/* Record a memory area against a node. */
172void __init memory_present(int nid, unsigned long start, unsigned long end)
173{
174 unsigned long pfn;
175
176 start &= PAGE_SECTION_MASK;
177 mminit_validate_memmodel_limits(&start, &end);
178 for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) {
179 unsigned long section = pfn_to_section_nr(pfn);
180 struct mem_section *ms;
181
182 sparse_index_init(section, nid);
183 set_section_nid(section, nid);
184
185 ms = __nr_to_section(section);
186 if (!ms->section_mem_map)
187 ms->section_mem_map = sparse_encode_early_nid(nid) |
188 SECTION_MARKED_PRESENT;
189 }
190}
191
192/*
193 * Only used by the i386 NUMA architecures, but relatively
194 * generic code.
195 */
196unsigned long __init node_memmap_size_bytes(int nid, unsigned long start_pfn,
197 unsigned long end_pfn)
198{
199 unsigned long pfn;
200 unsigned long nr_pages = 0;
201
202 mminit_validate_memmodel_limits(&start_pfn, &end_pfn);
203 for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
204 if (nid != early_pfn_to_nid(pfn))
205 continue;
206
207 if (pfn_present(pfn))
208 nr_pages += PAGES_PER_SECTION;
209 }
210
211 return nr_pages * sizeof(struct page);
212}
213
214/*
215 * Subtle, we encode the real pfn into the mem_map such that
216 * the identity pfn - section_mem_map will return the actual
217 * physical page frame number.
218 */
219static unsigned long sparse_encode_mem_map(struct page *mem_map, unsigned long pnum)
220{
221 return (unsigned long)(mem_map - (section_nr_to_pfn(pnum)));
222}
223
224/*
225 * Decode mem_map from the coded memmap
226 */
227struct page *sparse_decode_mem_map(unsigned long coded_mem_map, unsigned long pnum)
228{
229 /* mask off the extra low bits of information */
230 coded_mem_map &= SECTION_MAP_MASK;
231 return ((struct page *)coded_mem_map) + section_nr_to_pfn(pnum);
232}
233
234static int __meminit sparse_init_one_section(struct mem_section *ms,
235 unsigned long pnum, struct page *mem_map,
236 unsigned long *pageblock_bitmap)
237{
238 if (!present_section(ms))
239 return -EINVAL;
240
241 ms->section_mem_map &= ~SECTION_MAP_MASK;
242 ms->section_mem_map |= sparse_encode_mem_map(mem_map, pnum) |
243 SECTION_HAS_MEM_MAP;
244 ms->pageblock_flags = pageblock_bitmap;
245
246 return 1;
247}
248
249unsigned long usemap_size(void)
250{
251 unsigned long size_bytes;
252 size_bytes = roundup(SECTION_BLOCKFLAGS_BITS, 8) / 8;
253 size_bytes = roundup(size_bytes, sizeof(unsigned long));
254 return size_bytes;
255}
256
257#ifdef CONFIG_MEMORY_HOTPLUG
258static unsigned long *__kmalloc_section_usemap(void)
259{
260 return kmalloc(usemap_size(), GFP_KERNEL);
261}
262#endif /* CONFIG_MEMORY_HOTPLUG */
263
264#ifdef CONFIG_MEMORY_HOTREMOVE
265static unsigned long * __init
266sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
267 unsigned long size)
268{
269 unsigned long goal, limit;
270 unsigned long *p;
271 int nid;
272 /*
273 * A page may contain usemaps for other sections preventing the
274 * page being freed and making a section unremovable while
275 * other sections referencing the usemap remain active. Similarly,
276 * a pgdat can prevent a section being removed. If section A
277 * contains a pgdat and section B contains the usemap, both
278 * sections become inter-dependent. This allocates usemaps
279 * from the same section as the pgdat where possible to avoid
280 * this problem.
281 */
282 goal = __pa(pgdat) & (PAGE_SECTION_MASK << PAGE_SHIFT);
283 limit = goal + (1UL << PA_SECTION_SHIFT);
284 nid = early_pfn_to_nid(goal >> PAGE_SHIFT);
285again:
286 p = memblock_virt_alloc_try_nid_nopanic(size,
287 SMP_CACHE_BYTES, goal, limit,
288 nid);
289 if (!p && limit) {
290 limit = 0;
291 goto again;
292 }
293 return p;
294}
295
296static void __init check_usemap_section_nr(int nid, unsigned long *usemap)
297{
298 unsigned long usemap_snr, pgdat_snr;
299 static unsigned long old_usemap_snr = NR_MEM_SECTIONS;
300 static unsigned long old_pgdat_snr = NR_MEM_SECTIONS;
301 struct pglist_data *pgdat = NODE_DATA(nid);
302 int usemap_nid;
303
304 usemap_snr = pfn_to_section_nr(__pa(usemap) >> PAGE_SHIFT);
305 pgdat_snr = pfn_to_section_nr(__pa(pgdat) >> PAGE_SHIFT);
306 if (usemap_snr == pgdat_snr)
307 return;
308
309 if (old_usemap_snr == usemap_snr && old_pgdat_snr == pgdat_snr)
310 /* skip redundant message */
311 return;
312
313 old_usemap_snr = usemap_snr;
314 old_pgdat_snr = pgdat_snr;
315
316 usemap_nid = sparse_early_nid(__nr_to_section(usemap_snr));
317 if (usemap_nid != nid) {
318 pr_info("node %d must be removed before remove section %ld\n",
319 nid, usemap_snr);
320 return;
321 }
322 /*
323 * There is a circular dependency.
324 * Some platforms allow un-removable section because they will just
325 * gather other removable sections for dynamic partitioning.
326 * Just notify un-removable section's number here.
327 */
328 pr_info("Section %ld and %ld (node %d) have a circular dependency on usemap and pgdat allocations\n",
329 usemap_snr, pgdat_snr, nid);
330}
331#else
332static unsigned long * __init
333sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
334 unsigned long size)
335{
336 return memblock_virt_alloc_node_nopanic(size, pgdat->node_id);
337}
338
339static void __init check_usemap_section_nr(int nid, unsigned long *usemap)
340{
341}
342#endif /* CONFIG_MEMORY_HOTREMOVE */
343
344static void __init sparse_early_usemaps_alloc_node(void *data,
345 unsigned long pnum_begin,
346 unsigned long pnum_end,
347 unsigned long usemap_count, int nodeid)
348{
349 void *usemap;
350 unsigned long pnum;
351 unsigned long **usemap_map = (unsigned long **)data;
352 int size = usemap_size();
353
354 usemap = sparse_early_usemaps_alloc_pgdat_section(NODE_DATA(nodeid),
355 size * usemap_count);
356 if (!usemap) {
357 pr_warn("%s: allocation failed\n", __func__);
358 return;
359 }
360
361 for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
362 if (!present_section_nr(pnum))
363 continue;
364 usemap_map[pnum] = usemap;
365 usemap += size;
366 check_usemap_section_nr(nodeid, usemap_map[pnum]);
367 }
368}
369
370#ifndef CONFIG_SPARSEMEM_VMEMMAP
371struct page __init *sparse_mem_map_populate(unsigned long pnum, int nid)
372{
373 struct page *map;
374 unsigned long size;
375
376 map = alloc_remap(nid, sizeof(struct page) * PAGES_PER_SECTION);
377 if (map)
378 return map;
379
380 size = PAGE_ALIGN(sizeof(struct page) * PAGES_PER_SECTION);
381 map = memblock_virt_alloc_try_nid(size,
382 PAGE_SIZE, __pa(MAX_DMA_ADDRESS),
383 BOOTMEM_ALLOC_ACCESSIBLE, nid);
384 return map;
385}
386void __init sparse_mem_maps_populate_node(struct page **map_map,
387 unsigned long pnum_begin,
388 unsigned long pnum_end,
389 unsigned long map_count, int nodeid)
390{
391 void *map;
392 unsigned long pnum;
393 unsigned long size = sizeof(struct page) * PAGES_PER_SECTION;
394
395 map = alloc_remap(nodeid, size * map_count);
396 if (map) {
397 for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
398 if (!present_section_nr(pnum))
399 continue;
400 map_map[pnum] = map;
401 map += size;
402 }
403 return;
404 }
405
406 size = PAGE_ALIGN(size);
407 map = memblock_virt_alloc_try_nid(size * map_count,
408 PAGE_SIZE, __pa(MAX_DMA_ADDRESS),
409 BOOTMEM_ALLOC_ACCESSIBLE, nodeid);
410 if (map) {
411 for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
412 if (!present_section_nr(pnum))
413 continue;
414 map_map[pnum] = map;
415 map += size;
416 }
417 return;
418 }
419
420 /* fallback */
421 for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
422 struct mem_section *ms;
423
424 if (!present_section_nr(pnum))
425 continue;
426 map_map[pnum] = sparse_mem_map_populate(pnum, nodeid);
427 if (map_map[pnum])
428 continue;
429 ms = __nr_to_section(pnum);
430 pr_err("%s: sparsemem memory map backing failed some memory will not be available\n",
431 __func__);
432 ms->section_mem_map = 0;
433 }
434}
435#endif /* !CONFIG_SPARSEMEM_VMEMMAP */
436
437#ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
438static void __init sparse_early_mem_maps_alloc_node(void *data,
439 unsigned long pnum_begin,
440 unsigned long pnum_end,
441 unsigned long map_count, int nodeid)
442{
443 struct page **map_map = (struct page **)data;
444 sparse_mem_maps_populate_node(map_map, pnum_begin, pnum_end,
445 map_count, nodeid);
446}
447#else
448static struct page __init *sparse_early_mem_map_alloc(unsigned long pnum)
449{
450 struct page *map;
451 struct mem_section *ms = __nr_to_section(pnum);
452 int nid = sparse_early_nid(ms);
453
454 map = sparse_mem_map_populate(pnum, nid);
455 if (map)
456 return map;
457
458 pr_err("%s: sparsemem memory map backing failed some memory will not be available\n",
459 __func__);
460 ms->section_mem_map = 0;
461 return NULL;
462}
463#endif
464
465void __weak __meminit vmemmap_populate_print_last(void)
466{
467}
468
469/**
470 * alloc_usemap_and_memmap - memory alloction for pageblock flags and vmemmap
471 * @map: usemap_map for pageblock flags or mmap_map for vmemmap
472 */
473static void __init alloc_usemap_and_memmap(void (*alloc_func)
474 (void *, unsigned long, unsigned long,
475 unsigned long, int), void *data)
476{
477 unsigned long pnum;
478 unsigned long map_count;
479 int nodeid_begin = 0;
480 unsigned long pnum_begin = 0;
481
482 for (pnum = 0; pnum < NR_MEM_SECTIONS; pnum++) {
483 struct mem_section *ms;
484
485 if (!present_section_nr(pnum))
486 continue;
487 ms = __nr_to_section(pnum);
488 nodeid_begin = sparse_early_nid(ms);
489 pnum_begin = pnum;
490 break;
491 }
492 map_count = 1;
493 for (pnum = pnum_begin + 1; pnum < NR_MEM_SECTIONS; pnum++) {
494 struct mem_section *ms;
495 int nodeid;
496
497 if (!present_section_nr(pnum))
498 continue;
499 ms = __nr_to_section(pnum);
500 nodeid = sparse_early_nid(ms);
501 if (nodeid == nodeid_begin) {
502 map_count++;
503 continue;
504 }
505 /* ok, we need to take cake of from pnum_begin to pnum - 1*/
506 alloc_func(data, pnum_begin, pnum,
507 map_count, nodeid_begin);
508 /* new start, update count etc*/
509 nodeid_begin = nodeid;
510 pnum_begin = pnum;
511 map_count = 1;
512 }
513 /* ok, last chunk */
514 alloc_func(data, pnum_begin, NR_MEM_SECTIONS,
515 map_count, nodeid_begin);
516}
517
518/*
519 * Allocate the accumulated non-linear sections, allocate a mem_map
520 * for each and record the physical to section mapping.
521 */
522void __init sparse_init(void)
523{
524 unsigned long pnum;
525 struct page *map;
526 unsigned long *usemap;
527 unsigned long **usemap_map;
528 int size;
529#ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
530 int size2;
531 struct page **map_map;
532#endif
533
534 /* see include/linux/mmzone.h 'struct mem_section' definition */
535 BUILD_BUG_ON(!is_power_of_2(sizeof(struct mem_section)));
536
537 /* Setup pageblock_order for HUGETLB_PAGE_SIZE_VARIABLE */
538 set_pageblock_order();
539
540 /*
541 * map is using big page (aka 2M in x86 64 bit)
542 * usemap is less one page (aka 24 bytes)
543 * so alloc 2M (with 2M align) and 24 bytes in turn will
544 * make next 2M slip to one more 2M later.
545 * then in big system, the memory will have a lot of holes...
546 * here try to allocate 2M pages continuously.
547 *
548 * powerpc need to call sparse_init_one_section right after each
549 * sparse_early_mem_map_alloc, so allocate usemap_map at first.
550 */
551 size = sizeof(unsigned long *) * NR_MEM_SECTIONS;
552 usemap_map = memblock_virt_alloc(size, 0);
553 if (!usemap_map)
554 panic("can not allocate usemap_map\n");
555 alloc_usemap_and_memmap(sparse_early_usemaps_alloc_node,
556 (void *)usemap_map);
557
558#ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
559 size2 = sizeof(struct page *) * NR_MEM_SECTIONS;
560 map_map = memblock_virt_alloc(size2, 0);
561 if (!map_map)
562 panic("can not allocate map_map\n");
563 alloc_usemap_and_memmap(sparse_early_mem_maps_alloc_node,
564 (void *)map_map);
565#endif
566
567 for (pnum = 0; pnum < NR_MEM_SECTIONS; pnum++) {
568 if (!present_section_nr(pnum))
569 continue;
570
571 usemap = usemap_map[pnum];
572 if (!usemap)
573 continue;
574
575#ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
576 map = map_map[pnum];
577#else
578 map = sparse_early_mem_map_alloc(pnum);
579#endif
580 if (!map)
581 continue;
582
583 sparse_init_one_section(__nr_to_section(pnum), pnum, map,
584 usemap);
585 }
586
587 vmemmap_populate_print_last();
588
589#ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
590 memblock_free_early(__pa(map_map), size2);
591#endif
592 memblock_free_early(__pa(usemap_map), size);
593}
594
595#ifdef CONFIG_MEMORY_HOTPLUG
596#ifdef CONFIG_SPARSEMEM_VMEMMAP
597static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid)
598{
599 /* This will make the necessary allocations eventually. */
600 return sparse_mem_map_populate(pnum, nid);
601}
602static void __kfree_section_memmap(struct page *memmap)
603{
604 unsigned long start = (unsigned long)memmap;
605 unsigned long end = (unsigned long)(memmap + PAGES_PER_SECTION);
606
607 vmemmap_free(start, end);
608}
609#ifdef CONFIG_MEMORY_HOTREMOVE
610static void free_map_bootmem(struct page *memmap)
611{
612 unsigned long start = (unsigned long)memmap;
613 unsigned long end = (unsigned long)(memmap + PAGES_PER_SECTION);
614
615 vmemmap_free(start, end);
616}
617#endif /* CONFIG_MEMORY_HOTREMOVE */
618#else
619static struct page *__kmalloc_section_memmap(void)
620{
621 struct page *page, *ret;
622 unsigned long memmap_size = sizeof(struct page) * PAGES_PER_SECTION;
623
624 page = alloc_pages(GFP_KERNEL|__GFP_NOWARN, get_order(memmap_size));
625 if (page)
626 goto got_map_page;
627
628 ret = vmalloc(memmap_size);
629 if (ret)
630 goto got_map_ptr;
631
632 return NULL;
633got_map_page:
634 ret = (struct page *)pfn_to_kaddr(page_to_pfn(page));
635got_map_ptr:
636
637 return ret;
638}
639
640static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid)
641{
642 return __kmalloc_section_memmap();
643}
644
645static void __kfree_section_memmap(struct page *memmap)
646{
647 if (is_vmalloc_addr(memmap))
648 vfree(memmap);
649 else
650 free_pages((unsigned long)memmap,
651 get_order(sizeof(struct page) * PAGES_PER_SECTION));
652}
653
654#ifdef CONFIG_MEMORY_HOTREMOVE
655static void free_map_bootmem(struct page *memmap)
656{
657 unsigned long maps_section_nr, removing_section_nr, i;
658 unsigned long magic, nr_pages;
659 struct page *page = virt_to_page(memmap);
660
661 nr_pages = PAGE_ALIGN(PAGES_PER_SECTION * sizeof(struct page))
662 >> PAGE_SHIFT;
663
664 for (i = 0; i < nr_pages; i++, page++) {
665 magic = (unsigned long) page->lru.next;
666
667 BUG_ON(magic == NODE_INFO);
668
669 maps_section_nr = pfn_to_section_nr(page_to_pfn(page));
670 removing_section_nr = page->private;
671
672 /*
673 * When this function is called, the removing section is
674 * logical offlined state. This means all pages are isolated
675 * from page allocator. If removing section's memmap is placed
676 * on the same section, it must not be freed.
677 * If it is freed, page allocator may allocate it which will
678 * be removed physically soon.
679 */
680 if (maps_section_nr != removing_section_nr)
681 put_page_bootmem(page);
682 }
683}
684#endif /* CONFIG_MEMORY_HOTREMOVE */
685#endif /* CONFIG_SPARSEMEM_VMEMMAP */
686
687/*
688 * returns the number of sections whose mem_maps were properly
689 * set. If this is <=0, then that means that the passed-in
690 * map was not consumed and must be freed.
691 */
692int __meminit sparse_add_one_section(struct zone *zone, unsigned long start_pfn)
693{
694 unsigned long section_nr = pfn_to_section_nr(start_pfn);
695 struct pglist_data *pgdat = zone->zone_pgdat;
696 struct mem_section *ms;
697 struct page *memmap;
698 unsigned long *usemap;
699 unsigned long flags;
700 int ret;
701
702 /*
703 * no locking for this, because it does its own
704 * plus, it does a kmalloc
705 */
706 ret = sparse_index_init(section_nr, pgdat->node_id);
707 if (ret < 0 && ret != -EEXIST)
708 return ret;
709 memmap = kmalloc_section_memmap(section_nr, pgdat->node_id);
710 if (!memmap)
711 return -ENOMEM;
712 usemap = __kmalloc_section_usemap();
713 if (!usemap) {
714 __kfree_section_memmap(memmap);
715 return -ENOMEM;
716 }
717
718 pgdat_resize_lock(pgdat, &flags);
719
720 ms = __pfn_to_section(start_pfn);
721 if (ms->section_mem_map & SECTION_MARKED_PRESENT) {
722 ret = -EEXIST;
723 goto out;
724 }
725
726 memset(memmap, 0, sizeof(struct page) * PAGES_PER_SECTION);
727
728 ms->section_mem_map |= SECTION_MARKED_PRESENT;
729
730 ret = sparse_init_one_section(ms, section_nr, memmap, usemap);
731
732out:
733 pgdat_resize_unlock(pgdat, &flags);
734 if (ret <= 0) {
735 kfree(usemap);
736 __kfree_section_memmap(memmap);
737 }
738 return ret;
739}
740
741#ifdef CONFIG_MEMORY_HOTREMOVE
742#ifdef CONFIG_MEMORY_FAILURE
743static void clear_hwpoisoned_pages(struct page *memmap, int nr_pages)
744{
745 int i;
746
747 if (!memmap)
748 return;
749
750 for (i = 0; i < nr_pages; i++) {
751 if (PageHWPoison(&memmap[i])) {
752 atomic_long_sub(1, &num_poisoned_pages);
753 ClearPageHWPoison(&memmap[i]);
754 }
755 }
756}
757#else
758static inline void clear_hwpoisoned_pages(struct page *memmap, int nr_pages)
759{
760}
761#endif
762
763static void free_section_usemap(struct page *memmap, unsigned long *usemap)
764{
765 struct page *usemap_page;
766
767 if (!usemap)
768 return;
769
770 usemap_page = virt_to_page(usemap);
771 /*
772 * Check to see if allocation came from hot-plug-add
773 */
774 if (PageSlab(usemap_page) || PageCompound(usemap_page)) {
775 kfree(usemap);
776 if (memmap)
777 __kfree_section_memmap(memmap);
778 return;
779 }
780
781 /*
782 * The usemap came from bootmem. This is packed with other usemaps
783 * on the section which has pgdat at boot time. Just keep it as is now.
784 */
785
786 if (memmap)
787 free_map_bootmem(memmap);
788}
789
790void sparse_remove_one_section(struct zone *zone, struct mem_section *ms,
791 unsigned long map_offset)
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 + map_offset,
808 PAGES_PER_SECTION - map_offset);
809 free_section_usemap(memmap, usemap);
810}
811#endif /* CONFIG_MEMORY_HOTREMOVE */
812#endif /* CONFIG_MEMORY_HOTPLUG */