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// SPDX-License-Identifier: GPL-2.0
2/*
3 * sparse memory mappings.
4 */
5#include <linux/mm.h>
6#include <linux/slab.h>
7#include <linux/mmzone.h>
8#include <linux/memblock.h>
9#include <linux/compiler.h>
10#include <linux/highmem.h>
11#include <linux/export.h>
12#include <linux/spinlock.h>
13#include <linux/vmalloc.h>
14#include <linux/swap.h>
15#include <linux/swapops.h>
16#include <linux/bootmem_info.h>
17#include <linux/vmstat.h>
18#include "internal.h"
19#include <asm/dma.h>
20
21/*
22 * Permanent SPARSEMEM data:
23 *
24 * 1) mem_section - memory sections, mem_map's for valid memory
25 */
26#ifdef CONFIG_SPARSEMEM_EXTREME
27struct mem_section **mem_section;
28#else
29struct mem_section mem_section[NR_SECTION_ROOTS][SECTIONS_PER_ROOT]
30 ____cacheline_internodealigned_in_smp;
31#endif
32EXPORT_SYMBOL(mem_section);
33
34#ifdef NODE_NOT_IN_PAGE_FLAGS
35/*
36 * If we did not store the node number in the page then we have to
37 * do a lookup in the section_to_node_table in order to find which
38 * node the page belongs to.
39 */
40#if MAX_NUMNODES <= 256
41static u8 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned;
42#else
43static u16 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned;
44#endif
45
46int page_to_nid(const struct page *page)
47{
48 return section_to_node_table[page_to_section(page)];
49}
50EXPORT_SYMBOL(page_to_nid);
51
52static void set_section_nid(unsigned long section_nr, int nid)
53{
54 section_to_node_table[section_nr] = nid;
55}
56#else /* !NODE_NOT_IN_PAGE_FLAGS */
57static inline void set_section_nid(unsigned long section_nr, int nid)
58{
59}
60#endif
61
62#ifdef CONFIG_SPARSEMEM_EXTREME
63static noinline struct mem_section __ref *sparse_index_alloc(int nid)
64{
65 struct mem_section *section = NULL;
66 unsigned long array_size = SECTIONS_PER_ROOT *
67 sizeof(struct mem_section);
68
69 if (slab_is_available()) {
70 section = kzalloc_node(array_size, GFP_KERNEL, nid);
71 } else {
72 section = memblock_alloc_node(array_size, SMP_CACHE_BYTES,
73 nid);
74 if (!section)
75 panic("%s: Failed to allocate %lu bytes nid=%d\n",
76 __func__, array_size, nid);
77 }
78
79 return section;
80}
81
82static int __meminit sparse_index_init(unsigned long section_nr, int nid)
83{
84 unsigned long root = SECTION_NR_TO_ROOT(section_nr);
85 struct mem_section *section;
86
87 /*
88 * An existing section is possible in the sub-section hotplug
89 * case. First hot-add instantiates, follow-on hot-add reuses
90 * the existing section.
91 *
92 * The mem_hotplug_lock resolves the apparent race below.
93 */
94 if (mem_section[root])
95 return 0;
96
97 section = sparse_index_alloc(nid);
98 if (!section)
99 return -ENOMEM;
100
101 mem_section[root] = section;
102
103 return 0;
104}
105#else /* !SPARSEMEM_EXTREME */
106static inline int sparse_index_init(unsigned long section_nr, int nid)
107{
108 return 0;
109}
110#endif
111
112/*
113 * During early boot, before section_mem_map is used for an actual
114 * mem_map, we use section_mem_map to store the section's NUMA
115 * node. This keeps us from having to use another data structure. The
116 * node information is cleared just before we store the real mem_map.
117 */
118static inline unsigned long sparse_encode_early_nid(int nid)
119{
120 return ((unsigned long)nid << SECTION_NID_SHIFT);
121}
122
123static inline int sparse_early_nid(struct mem_section *section)
124{
125 return (section->section_mem_map >> SECTION_NID_SHIFT);
126}
127
128/* Validate the physical addressing limitations of the model */
129static void __meminit mminit_validate_memmodel_limits(unsigned long *start_pfn,
130 unsigned long *end_pfn)
131{
132 unsigned long max_sparsemem_pfn = (DIRECT_MAP_PHYSMEM_END + 1) >> PAGE_SHIFT;
133
134 /*
135 * Sanity checks - do not allow an architecture to pass
136 * in larger pfns than the maximum scope of sparsemem:
137 */
138 if (*start_pfn > max_sparsemem_pfn) {
139 mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
140 "Start of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
141 *start_pfn, *end_pfn, max_sparsemem_pfn);
142 WARN_ON_ONCE(1);
143 *start_pfn = max_sparsemem_pfn;
144 *end_pfn = max_sparsemem_pfn;
145 } else if (*end_pfn > max_sparsemem_pfn) {
146 mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
147 "End of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
148 *start_pfn, *end_pfn, max_sparsemem_pfn);
149 WARN_ON_ONCE(1);
150 *end_pfn = max_sparsemem_pfn;
151 }
152}
153
154/*
155 * There are a number of times that we loop over NR_MEM_SECTIONS,
156 * looking for section_present() on each. But, when we have very
157 * large physical address spaces, NR_MEM_SECTIONS can also be
158 * very large which makes the loops quite long.
159 *
160 * Keeping track of this gives us an easy way to break out of
161 * those loops early.
162 */
163unsigned long __highest_present_section_nr;
164static void __section_mark_present(struct mem_section *ms,
165 unsigned long section_nr)
166{
167 if (section_nr > __highest_present_section_nr)
168 __highest_present_section_nr = section_nr;
169
170 ms->section_mem_map |= SECTION_MARKED_PRESENT;
171}
172
173#define for_each_present_section_nr(start, section_nr) \
174 for (section_nr = next_present_section_nr(start-1); \
175 section_nr != -1; \
176 section_nr = next_present_section_nr(section_nr))
177
178static inline unsigned long first_present_section_nr(void)
179{
180 return next_present_section_nr(-1);
181}
182
183#ifdef CONFIG_SPARSEMEM_VMEMMAP
184static void subsection_mask_set(unsigned long *map, unsigned long pfn,
185 unsigned long nr_pages)
186{
187 int idx = subsection_map_index(pfn);
188 int end = subsection_map_index(pfn + nr_pages - 1);
189
190 bitmap_set(map, idx, end - idx + 1);
191}
192
193void __init subsection_map_init(unsigned long pfn, unsigned long nr_pages)
194{
195 int end_sec_nr = pfn_to_section_nr(pfn + nr_pages - 1);
196 unsigned long nr, start_sec_nr = pfn_to_section_nr(pfn);
197
198 for (nr = start_sec_nr; nr <= end_sec_nr; nr++) {
199 struct mem_section *ms;
200 unsigned long pfns;
201
202 pfns = min(nr_pages, PAGES_PER_SECTION
203 - (pfn & ~PAGE_SECTION_MASK));
204 ms = __nr_to_section(nr);
205 subsection_mask_set(ms->usage->subsection_map, pfn, pfns);
206
207 pr_debug("%s: sec: %lu pfns: %lu set(%d, %d)\n", __func__, nr,
208 pfns, subsection_map_index(pfn),
209 subsection_map_index(pfn + pfns - 1));
210
211 pfn += pfns;
212 nr_pages -= pfns;
213 }
214}
215#else
216void __init subsection_map_init(unsigned long pfn, unsigned long nr_pages)
217{
218}
219#endif
220
221/* Record a memory area against a node. */
222static void __init memory_present(int nid, unsigned long start, unsigned long end)
223{
224 unsigned long pfn;
225
226 start &= PAGE_SECTION_MASK;
227 mminit_validate_memmodel_limits(&start, &end);
228 for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) {
229 unsigned long section_nr = pfn_to_section_nr(pfn);
230 struct mem_section *ms;
231
232 sparse_index_init(section_nr, nid);
233 set_section_nid(section_nr, nid);
234
235 ms = __nr_to_section(section_nr);
236 if (!ms->section_mem_map) {
237 ms->section_mem_map = sparse_encode_early_nid(nid) |
238 SECTION_IS_ONLINE;
239 __section_mark_present(ms, section_nr);
240 }
241 }
242}
243
244/*
245 * Mark all memblocks as present using memory_present().
246 * This is a convenience function that is useful to mark all of the systems
247 * memory as present during initialization.
248 */
249static void __init memblocks_present(void)
250{
251 unsigned long start, end;
252 int i, nid;
253
254#ifdef CONFIG_SPARSEMEM_EXTREME
255 if (unlikely(!mem_section)) {
256 unsigned long size, align;
257
258 size = sizeof(struct mem_section *) * NR_SECTION_ROOTS;
259 align = 1 << (INTERNODE_CACHE_SHIFT);
260 mem_section = memblock_alloc(size, align);
261 if (!mem_section)
262 panic("%s: Failed to allocate %lu bytes align=0x%lx\n",
263 __func__, size, align);
264 }
265#endif
266
267 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid)
268 memory_present(nid, start, end);
269}
270
271/*
272 * Subtle, we encode the real pfn into the mem_map such that
273 * the identity pfn - section_mem_map will return the actual
274 * physical page frame number.
275 */
276static unsigned long sparse_encode_mem_map(struct page *mem_map, unsigned long pnum)
277{
278 unsigned long coded_mem_map =
279 (unsigned long)(mem_map - (section_nr_to_pfn(pnum)));
280 BUILD_BUG_ON(SECTION_MAP_LAST_BIT > PFN_SECTION_SHIFT);
281 BUG_ON(coded_mem_map & ~SECTION_MAP_MASK);
282 return coded_mem_map;
283}
284
285#ifdef CONFIG_MEMORY_HOTPLUG
286/*
287 * Decode mem_map from the coded memmap
288 */
289struct page *sparse_decode_mem_map(unsigned long coded_mem_map, unsigned long pnum)
290{
291 /* mask off the extra low bits of information */
292 coded_mem_map &= SECTION_MAP_MASK;
293 return ((struct page *)coded_mem_map) + section_nr_to_pfn(pnum);
294}
295#endif /* CONFIG_MEMORY_HOTPLUG */
296
297static void __meminit sparse_init_one_section(struct mem_section *ms,
298 unsigned long pnum, struct page *mem_map,
299 struct mem_section_usage *usage, unsigned long flags)
300{
301 ms->section_mem_map &= ~SECTION_MAP_MASK;
302 ms->section_mem_map |= sparse_encode_mem_map(mem_map, pnum)
303 | SECTION_HAS_MEM_MAP | flags;
304 ms->usage = usage;
305}
306
307static unsigned long usemap_size(void)
308{
309 return BITS_TO_LONGS(SECTION_BLOCKFLAGS_BITS) * sizeof(unsigned long);
310}
311
312size_t mem_section_usage_size(void)
313{
314 return sizeof(struct mem_section_usage) + usemap_size();
315}
316
317#ifdef CONFIG_MEMORY_HOTREMOVE
318static inline phys_addr_t pgdat_to_phys(struct pglist_data *pgdat)
319{
320#ifndef CONFIG_NUMA
321 VM_BUG_ON(pgdat != &contig_page_data);
322 return __pa_symbol(&contig_page_data);
323#else
324 return __pa(pgdat);
325#endif
326}
327
328static struct mem_section_usage * __init
329sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
330 unsigned long size)
331{
332 struct mem_section_usage *usage;
333 unsigned long goal, limit;
334 int nid;
335 /*
336 * A page may contain usemaps for other sections preventing the
337 * page being freed and making a section unremovable while
338 * other sections referencing the usemap remain active. Similarly,
339 * a pgdat can prevent a section being removed. If section A
340 * contains a pgdat and section B contains the usemap, both
341 * sections become inter-dependent. This allocates usemaps
342 * from the same section as the pgdat where possible to avoid
343 * this problem.
344 */
345 goal = pgdat_to_phys(pgdat) & (PAGE_SECTION_MASK << PAGE_SHIFT);
346 limit = goal + (1UL << PA_SECTION_SHIFT);
347 nid = early_pfn_to_nid(goal >> PAGE_SHIFT);
348again:
349 usage = memblock_alloc_try_nid(size, SMP_CACHE_BYTES, goal, limit, nid);
350 if (!usage && limit) {
351 limit = MEMBLOCK_ALLOC_ACCESSIBLE;
352 goto again;
353 }
354 return usage;
355}
356
357static void __init check_usemap_section_nr(int nid,
358 struct mem_section_usage *usage)
359{
360 unsigned long usemap_snr, pgdat_snr;
361 static unsigned long old_usemap_snr;
362 static unsigned long old_pgdat_snr;
363 struct pglist_data *pgdat = NODE_DATA(nid);
364 int usemap_nid;
365
366 /* First call */
367 if (!old_usemap_snr) {
368 old_usemap_snr = NR_MEM_SECTIONS;
369 old_pgdat_snr = NR_MEM_SECTIONS;
370 }
371
372 usemap_snr = pfn_to_section_nr(__pa(usage) >> PAGE_SHIFT);
373 pgdat_snr = pfn_to_section_nr(pgdat_to_phys(pgdat) >> PAGE_SHIFT);
374 if (usemap_snr == pgdat_snr)
375 return;
376
377 if (old_usemap_snr == usemap_snr && old_pgdat_snr == pgdat_snr)
378 /* skip redundant message */
379 return;
380
381 old_usemap_snr = usemap_snr;
382 old_pgdat_snr = pgdat_snr;
383
384 usemap_nid = sparse_early_nid(__nr_to_section(usemap_snr));
385 if (usemap_nid != nid) {
386 pr_info("node %d must be removed before remove section %ld\n",
387 nid, usemap_snr);
388 return;
389 }
390 /*
391 * There is a circular dependency.
392 * Some platforms allow un-removable section because they will just
393 * gather other removable sections for dynamic partitioning.
394 * Just notify un-removable section's number here.
395 */
396 pr_info("Section %ld and %ld (node %d) have a circular dependency on usemap and pgdat allocations\n",
397 usemap_snr, pgdat_snr, nid);
398}
399#else
400static struct mem_section_usage * __init
401sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
402 unsigned long size)
403{
404 return memblock_alloc_node(size, SMP_CACHE_BYTES, pgdat->node_id);
405}
406
407static void __init check_usemap_section_nr(int nid,
408 struct mem_section_usage *usage)
409{
410}
411#endif /* CONFIG_MEMORY_HOTREMOVE */
412
413#ifdef CONFIG_SPARSEMEM_VMEMMAP
414static unsigned long __init section_map_size(void)
415{
416 return ALIGN(sizeof(struct page) * PAGES_PER_SECTION, PMD_SIZE);
417}
418
419#else
420static unsigned long __init section_map_size(void)
421{
422 return PAGE_ALIGN(sizeof(struct page) * PAGES_PER_SECTION);
423}
424
425struct page __init *__populate_section_memmap(unsigned long pfn,
426 unsigned long nr_pages, int nid, struct vmem_altmap *altmap,
427 struct dev_pagemap *pgmap)
428{
429 unsigned long size = section_map_size();
430 struct page *map = sparse_buffer_alloc(size);
431 phys_addr_t addr = __pa(MAX_DMA_ADDRESS);
432
433 if (map)
434 return map;
435
436 map = memmap_alloc(size, size, addr, nid, false);
437 if (!map)
438 panic("%s: Failed to allocate %lu bytes align=0x%lx nid=%d from=%pa\n",
439 __func__, size, PAGE_SIZE, nid, &addr);
440
441 return map;
442}
443#endif /* !CONFIG_SPARSEMEM_VMEMMAP */
444
445static void *sparsemap_buf __meminitdata;
446static void *sparsemap_buf_end __meminitdata;
447
448static inline void __meminit sparse_buffer_free(unsigned long size)
449{
450 WARN_ON(!sparsemap_buf || size == 0);
451 memblock_free(sparsemap_buf, size);
452}
453
454static void __init sparse_buffer_init(unsigned long size, int nid)
455{
456 phys_addr_t addr = __pa(MAX_DMA_ADDRESS);
457 WARN_ON(sparsemap_buf); /* forgot to call sparse_buffer_fini()? */
458 /*
459 * Pre-allocated buffer is mainly used by __populate_section_memmap
460 * and we want it to be properly aligned to the section size - this is
461 * especially the case for VMEMMAP which maps memmap to PMDs
462 */
463 sparsemap_buf = memmap_alloc(size, section_map_size(), addr, nid, true);
464 sparsemap_buf_end = sparsemap_buf + size;
465#ifndef CONFIG_SPARSEMEM_VMEMMAP
466 memmap_boot_pages_add(DIV_ROUND_UP(size, PAGE_SIZE));
467#endif
468}
469
470static void __init sparse_buffer_fini(void)
471{
472 unsigned long size = sparsemap_buf_end - sparsemap_buf;
473
474 if (sparsemap_buf && size > 0)
475 sparse_buffer_free(size);
476 sparsemap_buf = NULL;
477}
478
479void * __meminit sparse_buffer_alloc(unsigned long size)
480{
481 void *ptr = NULL;
482
483 if (sparsemap_buf) {
484 ptr = (void *) roundup((unsigned long)sparsemap_buf, size);
485 if (ptr + size > sparsemap_buf_end)
486 ptr = NULL;
487 else {
488 /* Free redundant aligned space */
489 if ((unsigned long)(ptr - sparsemap_buf) > 0)
490 sparse_buffer_free((unsigned long)(ptr - sparsemap_buf));
491 sparsemap_buf = ptr + size;
492 }
493 }
494 return ptr;
495}
496
497void __weak __meminit vmemmap_populate_print_last(void)
498{
499}
500
501/*
502 * Initialize sparse on a specific node. The node spans [pnum_begin, pnum_end)
503 * And number of present sections in this node is map_count.
504 */
505static void __init sparse_init_nid(int nid, unsigned long pnum_begin,
506 unsigned long pnum_end,
507 unsigned long map_count)
508{
509 struct mem_section_usage *usage;
510 unsigned long pnum;
511 struct page *map;
512
513 usage = sparse_early_usemaps_alloc_pgdat_section(NODE_DATA(nid),
514 mem_section_usage_size() * map_count);
515 if (!usage) {
516 pr_err("%s: node[%d] usemap allocation failed", __func__, nid);
517 goto failed;
518 }
519 sparse_buffer_init(map_count * section_map_size(), nid);
520 for_each_present_section_nr(pnum_begin, pnum) {
521 unsigned long pfn = section_nr_to_pfn(pnum);
522
523 if (pnum >= pnum_end)
524 break;
525
526 map = __populate_section_memmap(pfn, PAGES_PER_SECTION,
527 nid, NULL, NULL);
528 if (!map) {
529 pr_err("%s: node[%d] memory map backing failed. Some memory will not be available.",
530 __func__, nid);
531 pnum_begin = pnum;
532 sparse_buffer_fini();
533 goto failed;
534 }
535 check_usemap_section_nr(nid, usage);
536 sparse_init_one_section(__nr_to_section(pnum), pnum, map, usage,
537 SECTION_IS_EARLY);
538 usage = (void *) usage + mem_section_usage_size();
539 }
540 sparse_buffer_fini();
541 return;
542failed:
543 /* We failed to allocate, mark all the following pnums as not present */
544 for_each_present_section_nr(pnum_begin, pnum) {
545 struct mem_section *ms;
546
547 if (pnum >= pnum_end)
548 break;
549 ms = __nr_to_section(pnum);
550 ms->section_mem_map = 0;
551 }
552}
553
554/*
555 * Allocate the accumulated non-linear sections, allocate a mem_map
556 * for each and record the physical to section mapping.
557 */
558void __init sparse_init(void)
559{
560 unsigned long pnum_end, pnum_begin, map_count = 1;
561 int nid_begin;
562
563 /* see include/linux/mmzone.h 'struct mem_section' definition */
564 BUILD_BUG_ON(!is_power_of_2(sizeof(struct mem_section)));
565 memblocks_present();
566
567 pnum_begin = first_present_section_nr();
568 nid_begin = sparse_early_nid(__nr_to_section(pnum_begin));
569
570 /* Setup pageblock_order for HUGETLB_PAGE_SIZE_VARIABLE */
571 set_pageblock_order();
572
573 for_each_present_section_nr(pnum_begin + 1, pnum_end) {
574 int nid = sparse_early_nid(__nr_to_section(pnum_end));
575
576 if (nid == nid_begin) {
577 map_count++;
578 continue;
579 }
580 /* Init node with sections in range [pnum_begin, pnum_end) */
581 sparse_init_nid(nid_begin, pnum_begin, pnum_end, map_count);
582 nid_begin = nid;
583 pnum_begin = pnum_end;
584 map_count = 1;
585 }
586 /* cover the last node */
587 sparse_init_nid(nid_begin, pnum_begin, pnum_end, map_count);
588 vmemmap_populate_print_last();
589}
590
591#ifdef CONFIG_MEMORY_HOTPLUG
592
593/* Mark all memory sections within the pfn range as online */
594void online_mem_sections(unsigned long start_pfn, unsigned long end_pfn)
595{
596 unsigned long pfn;
597
598 for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
599 unsigned long section_nr = pfn_to_section_nr(pfn);
600 struct mem_section *ms;
601
602 /* onlining code should never touch invalid ranges */
603 if (WARN_ON(!valid_section_nr(section_nr)))
604 continue;
605
606 ms = __nr_to_section(section_nr);
607 ms->section_mem_map |= SECTION_IS_ONLINE;
608 }
609}
610
611/* Mark all memory sections within the pfn range as offline */
612void offline_mem_sections(unsigned long start_pfn, unsigned long end_pfn)
613{
614 unsigned long pfn;
615
616 for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
617 unsigned long section_nr = pfn_to_section_nr(pfn);
618 struct mem_section *ms;
619
620 /*
621 * TODO this needs some double checking. Offlining code makes
622 * sure to check pfn_valid but those checks might be just bogus
623 */
624 if (WARN_ON(!valid_section_nr(section_nr)))
625 continue;
626
627 ms = __nr_to_section(section_nr);
628 ms->section_mem_map &= ~SECTION_IS_ONLINE;
629 }
630}
631
632#ifdef CONFIG_SPARSEMEM_VMEMMAP
633static struct page * __meminit populate_section_memmap(unsigned long pfn,
634 unsigned long nr_pages, int nid, struct vmem_altmap *altmap,
635 struct dev_pagemap *pgmap)
636{
637 return __populate_section_memmap(pfn, nr_pages, nid, altmap, pgmap);
638}
639
640static void depopulate_section_memmap(unsigned long pfn, unsigned long nr_pages,
641 struct vmem_altmap *altmap)
642{
643 unsigned long start = (unsigned long) pfn_to_page(pfn);
644 unsigned long end = start + nr_pages * sizeof(struct page);
645
646 memmap_pages_add(-1L * (DIV_ROUND_UP(end - start, PAGE_SIZE)));
647 vmemmap_free(start, end, altmap);
648}
649static void free_map_bootmem(struct page *memmap)
650{
651 unsigned long start = (unsigned long)memmap;
652 unsigned long end = (unsigned long)(memmap + PAGES_PER_SECTION);
653
654 vmemmap_free(start, end, NULL);
655}
656
657static int clear_subsection_map(unsigned long pfn, unsigned long nr_pages)
658{
659 DECLARE_BITMAP(map, SUBSECTIONS_PER_SECTION) = { 0 };
660 DECLARE_BITMAP(tmp, SUBSECTIONS_PER_SECTION) = { 0 };
661 struct mem_section *ms = __pfn_to_section(pfn);
662 unsigned long *subsection_map = ms->usage
663 ? &ms->usage->subsection_map[0] : NULL;
664
665 subsection_mask_set(map, pfn, nr_pages);
666 if (subsection_map)
667 bitmap_and(tmp, map, subsection_map, SUBSECTIONS_PER_SECTION);
668
669 if (WARN(!subsection_map || !bitmap_equal(tmp, map, SUBSECTIONS_PER_SECTION),
670 "section already deactivated (%#lx + %ld)\n",
671 pfn, nr_pages))
672 return -EINVAL;
673
674 bitmap_xor(subsection_map, map, subsection_map, SUBSECTIONS_PER_SECTION);
675 return 0;
676}
677
678static bool is_subsection_map_empty(struct mem_section *ms)
679{
680 return bitmap_empty(&ms->usage->subsection_map[0],
681 SUBSECTIONS_PER_SECTION);
682}
683
684static int fill_subsection_map(unsigned long pfn, unsigned long nr_pages)
685{
686 struct mem_section *ms = __pfn_to_section(pfn);
687 DECLARE_BITMAP(map, SUBSECTIONS_PER_SECTION) = { 0 };
688 unsigned long *subsection_map;
689 int rc = 0;
690
691 subsection_mask_set(map, pfn, nr_pages);
692
693 subsection_map = &ms->usage->subsection_map[0];
694
695 if (bitmap_empty(map, SUBSECTIONS_PER_SECTION))
696 rc = -EINVAL;
697 else if (bitmap_intersects(map, subsection_map, SUBSECTIONS_PER_SECTION))
698 rc = -EEXIST;
699 else
700 bitmap_or(subsection_map, map, subsection_map,
701 SUBSECTIONS_PER_SECTION);
702
703 return rc;
704}
705#else
706static struct page * __meminit populate_section_memmap(unsigned long pfn,
707 unsigned long nr_pages, int nid, struct vmem_altmap *altmap,
708 struct dev_pagemap *pgmap)
709{
710 return kvmalloc_node(array_size(sizeof(struct page),
711 PAGES_PER_SECTION), GFP_KERNEL, nid);
712}
713
714static void depopulate_section_memmap(unsigned long pfn, unsigned long nr_pages,
715 struct vmem_altmap *altmap)
716{
717 kvfree(pfn_to_page(pfn));
718}
719
720static void free_map_bootmem(struct page *memmap)
721{
722 unsigned long maps_section_nr, removing_section_nr, i;
723 unsigned long type, nr_pages;
724 struct page *page = virt_to_page(memmap);
725
726 nr_pages = PAGE_ALIGN(PAGES_PER_SECTION * sizeof(struct page))
727 >> PAGE_SHIFT;
728
729 for (i = 0; i < nr_pages; i++, page++) {
730 type = bootmem_type(page);
731
732 BUG_ON(type == NODE_INFO);
733
734 maps_section_nr = pfn_to_section_nr(page_to_pfn(page));
735 removing_section_nr = bootmem_info(page);
736
737 /*
738 * When this function is called, the removing section is
739 * logical offlined state. This means all pages are isolated
740 * from page allocator. If removing section's memmap is placed
741 * on the same section, it must not be freed.
742 * If it is freed, page allocator may allocate it which will
743 * be removed physically soon.
744 */
745 if (maps_section_nr != removing_section_nr)
746 put_page_bootmem(page);
747 }
748}
749
750static int clear_subsection_map(unsigned long pfn, unsigned long nr_pages)
751{
752 return 0;
753}
754
755static bool is_subsection_map_empty(struct mem_section *ms)
756{
757 return true;
758}
759
760static int fill_subsection_map(unsigned long pfn, unsigned long nr_pages)
761{
762 return 0;
763}
764#endif /* CONFIG_SPARSEMEM_VMEMMAP */
765
766/*
767 * To deactivate a memory region, there are 3 cases to handle across
768 * two configurations (SPARSEMEM_VMEMMAP={y,n}):
769 *
770 * 1. deactivation of a partial hot-added section (only possible in
771 * the SPARSEMEM_VMEMMAP=y case).
772 * a) section was present at memory init.
773 * b) section was hot-added post memory init.
774 * 2. deactivation of a complete hot-added section.
775 * 3. deactivation of a complete section from memory init.
776 *
777 * For 1, when subsection_map does not empty we will not be freeing the
778 * usage map, but still need to free the vmemmap range.
779 *
780 * For 2 and 3, the SPARSEMEM_VMEMMAP={y,n} cases are unified
781 */
782static void section_deactivate(unsigned long pfn, unsigned long nr_pages,
783 struct vmem_altmap *altmap)
784{
785 struct mem_section *ms = __pfn_to_section(pfn);
786 bool section_is_early = early_section(ms);
787 struct page *memmap = NULL;
788 bool empty;
789
790 if (clear_subsection_map(pfn, nr_pages))
791 return;
792
793 empty = is_subsection_map_empty(ms);
794 if (empty) {
795 unsigned long section_nr = pfn_to_section_nr(pfn);
796
797 /*
798 * Mark the section invalid so that valid_section()
799 * return false. This prevents code from dereferencing
800 * ms->usage array.
801 */
802 ms->section_mem_map &= ~SECTION_HAS_MEM_MAP;
803
804 /*
805 * When removing an early section, the usage map is kept (as the
806 * usage maps of other sections fall into the same page). It
807 * will be re-used when re-adding the section - which is then no
808 * longer an early section. If the usage map is PageReserved, it
809 * was allocated during boot.
810 */
811 if (!PageReserved(virt_to_page(ms->usage))) {
812 kfree_rcu(ms->usage, rcu);
813 WRITE_ONCE(ms->usage, NULL);
814 }
815 memmap = sparse_decode_mem_map(ms->section_mem_map, section_nr);
816 }
817
818 /*
819 * The memmap of early sections is always fully populated. See
820 * section_activate() and pfn_valid() .
821 */
822 if (!section_is_early)
823 depopulate_section_memmap(pfn, nr_pages, altmap);
824 else if (memmap)
825 free_map_bootmem(memmap);
826
827 if (empty)
828 ms->section_mem_map = (unsigned long)NULL;
829}
830
831static struct page * __meminit section_activate(int nid, unsigned long pfn,
832 unsigned long nr_pages, struct vmem_altmap *altmap,
833 struct dev_pagemap *pgmap)
834{
835 struct mem_section *ms = __pfn_to_section(pfn);
836 struct mem_section_usage *usage = NULL;
837 struct page *memmap;
838 int rc;
839
840 if (!ms->usage) {
841 usage = kzalloc(mem_section_usage_size(), GFP_KERNEL);
842 if (!usage)
843 return ERR_PTR(-ENOMEM);
844 ms->usage = usage;
845 }
846
847 rc = fill_subsection_map(pfn, nr_pages);
848 if (rc) {
849 if (usage)
850 ms->usage = NULL;
851 kfree(usage);
852 return ERR_PTR(rc);
853 }
854
855 /*
856 * The early init code does not consider partially populated
857 * initial sections, it simply assumes that memory will never be
858 * referenced. If we hot-add memory into such a section then we
859 * do not need to populate the memmap and can simply reuse what
860 * is already there.
861 */
862 if (nr_pages < PAGES_PER_SECTION && early_section(ms))
863 return pfn_to_page(pfn);
864
865 memmap = populate_section_memmap(pfn, nr_pages, nid, altmap, pgmap);
866 if (!memmap) {
867 section_deactivate(pfn, nr_pages, altmap);
868 return ERR_PTR(-ENOMEM);
869 }
870
871 return memmap;
872}
873
874/**
875 * sparse_add_section - add a memory section, or populate an existing one
876 * @nid: The node to add section on
877 * @start_pfn: start pfn of the memory range
878 * @nr_pages: number of pfns to add in the section
879 * @altmap: alternate pfns to allocate the memmap backing store
880 * @pgmap: alternate compound page geometry for devmap mappings
881 *
882 * This is only intended for hotplug.
883 *
884 * Note that only VMEMMAP supports sub-section aligned hotplug,
885 * the proper alignment and size are gated by check_pfn_span().
886 *
887 *
888 * Return:
889 * * 0 - On success.
890 * * -EEXIST - Section has been present.
891 * * -ENOMEM - Out of memory.
892 */
893int __meminit sparse_add_section(int nid, unsigned long start_pfn,
894 unsigned long nr_pages, struct vmem_altmap *altmap,
895 struct dev_pagemap *pgmap)
896{
897 unsigned long section_nr = pfn_to_section_nr(start_pfn);
898 struct mem_section *ms;
899 struct page *memmap;
900 int ret;
901
902 ret = sparse_index_init(section_nr, nid);
903 if (ret < 0)
904 return ret;
905
906 memmap = section_activate(nid, start_pfn, nr_pages, altmap, pgmap);
907 if (IS_ERR(memmap))
908 return PTR_ERR(memmap);
909
910 /*
911 * Poison uninitialized struct pages in order to catch invalid flags
912 * combinations.
913 */
914 if (!altmap || !altmap->inaccessible)
915 page_init_poison(memmap, sizeof(struct page) * nr_pages);
916
917 ms = __nr_to_section(section_nr);
918 set_section_nid(section_nr, nid);
919 __section_mark_present(ms, section_nr);
920
921 /* Align memmap to section boundary in the subsection case */
922 if (section_nr_to_pfn(section_nr) != start_pfn)
923 memmap = pfn_to_page(section_nr_to_pfn(section_nr));
924 sparse_init_one_section(ms, section_nr, memmap, ms->usage, 0);
925
926 return 0;
927}
928
929void sparse_remove_section(unsigned long pfn, unsigned long nr_pages,
930 struct vmem_altmap *altmap)
931{
932 struct mem_section *ms = __pfn_to_section(pfn);
933
934 if (WARN_ON_ONCE(!valid_section(ms)))
935 return;
936
937 section_deactivate(pfn, nr_pages, altmap);
938}
939#endif /* CONFIG_MEMORY_HOTPLUG */