Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * sparse memory mappings.
4 */
5#include <linux/mm.h>
6#include <linux/slab.h>
7#include <linux/mmzone.h>
8#include <linux/memblock.h>
9#include <linux/compiler.h>
10#include <linux/highmem.h>
11#include <linux/export.h>
12#include <linux/spinlock.h>
13#include <linux/vmalloc.h>
14#include <linux/swap.h>
15#include <linux/swapops.h>
16
17#include "internal.h"
18#include <asm/dma.h>
19
20/*
21 * Permanent SPARSEMEM data:
22 *
23 * 1) mem_section - memory sections, mem_map's for valid memory
24 */
25#ifdef CONFIG_SPARSEMEM_EXTREME
26struct mem_section **mem_section;
27#else
28struct mem_section mem_section[NR_SECTION_ROOTS][SECTIONS_PER_ROOT]
29 ____cacheline_internodealigned_in_smp;
30#endif
31EXPORT_SYMBOL(mem_section);
32
33#ifdef NODE_NOT_IN_PAGE_FLAGS
34/*
35 * If we did not store the node number in the page then we have to
36 * do a lookup in the section_to_node_table in order to find which
37 * node the page belongs to.
38 */
39#if MAX_NUMNODES <= 256
40static u8 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned;
41#else
42static u16 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned;
43#endif
44
45int page_to_nid(const struct page *page)
46{
47 return section_to_node_table[page_to_section(page)];
48}
49EXPORT_SYMBOL(page_to_nid);
50
51static void set_section_nid(unsigned long section_nr, int nid)
52{
53 section_to_node_table[section_nr] = nid;
54}
55#else /* !NODE_NOT_IN_PAGE_FLAGS */
56static inline void set_section_nid(unsigned long section_nr, int nid)
57{
58}
59#endif
60
61#ifdef CONFIG_SPARSEMEM_EXTREME
62static noinline struct mem_section __ref *sparse_index_alloc(int nid)
63{
64 struct mem_section *section = NULL;
65 unsigned long array_size = SECTIONS_PER_ROOT *
66 sizeof(struct mem_section);
67
68 if (slab_is_available()) {
69 section = kzalloc_node(array_size, GFP_KERNEL, nid);
70 } else {
71 section = memblock_alloc_node(array_size, SMP_CACHE_BYTES,
72 nid);
73 if (!section)
74 panic("%s: Failed to allocate %lu bytes nid=%d\n",
75 __func__, array_size, nid);
76 }
77
78 return section;
79}
80
81static int __meminit sparse_index_init(unsigned long section_nr, int nid)
82{
83 unsigned long root = SECTION_NR_TO_ROOT(section_nr);
84 struct mem_section *section;
85
86 /*
87 * An existing section is possible in the sub-section hotplug
88 * case. First hot-add instantiates, follow-on hot-add reuses
89 * the existing section.
90 *
91 * The mem_hotplug_lock resolves the apparent race below.
92 */
93 if (mem_section[root])
94 return 0;
95
96 section = sparse_index_alloc(nid);
97 if (!section)
98 return -ENOMEM;
99
100 mem_section[root] = section;
101
102 return 0;
103}
104#else /* !SPARSEMEM_EXTREME */
105static inline int sparse_index_init(unsigned long section_nr, int nid)
106{
107 return 0;
108}
109#endif
110
111#ifdef CONFIG_SPARSEMEM_EXTREME
112unsigned long __section_nr(struct mem_section *ms)
113{
114 unsigned long root_nr;
115 struct mem_section *root = NULL;
116
117 for (root_nr = 0; root_nr < NR_SECTION_ROOTS; root_nr++) {
118 root = __nr_to_section(root_nr * SECTIONS_PER_ROOT);
119 if (!root)
120 continue;
121
122 if ((ms >= root) && (ms < (root + SECTIONS_PER_ROOT)))
123 break;
124 }
125
126 VM_BUG_ON(!root);
127
128 return (root_nr * SECTIONS_PER_ROOT) + (ms - root);
129}
130#else
131unsigned long __section_nr(struct mem_section *ms)
132{
133 return (unsigned long)(ms - mem_section[0]);
134}
135#endif
136
137/*
138 * During early boot, before section_mem_map is used for an actual
139 * mem_map, we use section_mem_map to store the section's NUMA
140 * node. This keeps us from having to use another data structure. The
141 * node information is cleared just before we store the real mem_map.
142 */
143static inline unsigned long sparse_encode_early_nid(int nid)
144{
145 return (nid << SECTION_NID_SHIFT);
146}
147
148static inline int sparse_early_nid(struct mem_section *section)
149{
150 return (section->section_mem_map >> SECTION_NID_SHIFT);
151}
152
153/* Validate the physical addressing limitations of the model */
154void __meminit mminit_validate_memmodel_limits(unsigned long *start_pfn,
155 unsigned long *end_pfn)
156{
157 unsigned long max_sparsemem_pfn = 1UL << (MAX_PHYSMEM_BITS-PAGE_SHIFT);
158
159 /*
160 * Sanity checks - do not allow an architecture to pass
161 * in larger pfns than the maximum scope of sparsemem:
162 */
163 if (*start_pfn > max_sparsemem_pfn) {
164 mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
165 "Start of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
166 *start_pfn, *end_pfn, max_sparsemem_pfn);
167 WARN_ON_ONCE(1);
168 *start_pfn = max_sparsemem_pfn;
169 *end_pfn = max_sparsemem_pfn;
170 } else if (*end_pfn > max_sparsemem_pfn) {
171 mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
172 "End of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
173 *start_pfn, *end_pfn, max_sparsemem_pfn);
174 WARN_ON_ONCE(1);
175 *end_pfn = max_sparsemem_pfn;
176 }
177}
178
179/*
180 * There are a number of times that we loop over NR_MEM_SECTIONS,
181 * looking for section_present() on each. But, when we have very
182 * large physical address spaces, NR_MEM_SECTIONS can also be
183 * very large which makes the loops quite long.
184 *
185 * Keeping track of this gives us an easy way to break out of
186 * those loops early.
187 */
188unsigned long __highest_present_section_nr;
189static void section_mark_present(struct mem_section *ms)
190{
191 unsigned long section_nr = __section_nr(ms);
192
193 if (section_nr > __highest_present_section_nr)
194 __highest_present_section_nr = section_nr;
195
196 ms->section_mem_map |= SECTION_MARKED_PRESENT;
197}
198
199#define for_each_present_section_nr(start, section_nr) \
200 for (section_nr = next_present_section_nr(start-1); \
201 ((section_nr != -1) && \
202 (section_nr <= __highest_present_section_nr)); \
203 section_nr = next_present_section_nr(section_nr))
204
205static inline unsigned long first_present_section_nr(void)
206{
207 return next_present_section_nr(-1);
208}
209
210#ifdef CONFIG_SPARSEMEM_VMEMMAP
211static void subsection_mask_set(unsigned long *map, unsigned long pfn,
212 unsigned long nr_pages)
213{
214 int idx = subsection_map_index(pfn);
215 int end = subsection_map_index(pfn + nr_pages - 1);
216
217 bitmap_set(map, idx, end - idx + 1);
218}
219
220void __init subsection_map_init(unsigned long pfn, unsigned long nr_pages)
221{
222 int end_sec = pfn_to_section_nr(pfn + nr_pages - 1);
223 unsigned long nr, start_sec = pfn_to_section_nr(pfn);
224
225 if (!nr_pages)
226 return;
227
228 for (nr = start_sec; nr <= end_sec; nr++) {
229 struct mem_section *ms;
230 unsigned long pfns;
231
232 pfns = min(nr_pages, PAGES_PER_SECTION
233 - (pfn & ~PAGE_SECTION_MASK));
234 ms = __nr_to_section(nr);
235 subsection_mask_set(ms->usage->subsection_map, pfn, pfns);
236
237 pr_debug("%s: sec: %lu pfns: %lu set(%d, %d)\n", __func__, nr,
238 pfns, subsection_map_index(pfn),
239 subsection_map_index(pfn + pfns - 1));
240
241 pfn += pfns;
242 nr_pages -= pfns;
243 }
244}
245#else
246void __init subsection_map_init(unsigned long pfn, unsigned long nr_pages)
247{
248}
249#endif
250
251/* Record a memory area against a node. */
252static void __init memory_present(int nid, unsigned long start, unsigned long end)
253{
254 unsigned long pfn;
255
256#ifdef CONFIG_SPARSEMEM_EXTREME
257 if (unlikely(!mem_section)) {
258 unsigned long size, align;
259
260 size = sizeof(struct mem_section*) * NR_SECTION_ROOTS;
261 align = 1 << (INTERNODE_CACHE_SHIFT);
262 mem_section = memblock_alloc(size, align);
263 if (!mem_section)
264 panic("%s: Failed to allocate %lu bytes align=0x%lx\n",
265 __func__, size, align);
266 }
267#endif
268
269 start &= PAGE_SECTION_MASK;
270 mminit_validate_memmodel_limits(&start, &end);
271 for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) {
272 unsigned long section = pfn_to_section_nr(pfn);
273 struct mem_section *ms;
274
275 sparse_index_init(section, nid);
276 set_section_nid(section, nid);
277
278 ms = __nr_to_section(section);
279 if (!ms->section_mem_map) {
280 ms->section_mem_map = sparse_encode_early_nid(nid) |
281 SECTION_IS_ONLINE;
282 section_mark_present(ms);
283 }
284 }
285}
286
287/*
288 * Mark all memblocks as present using memory_present().
289 * This is a convenience function that is useful to mark all of the systems
290 * memory as present during initialization.
291 */
292static void __init memblocks_present(void)
293{
294 struct memblock_region *reg;
295
296 for_each_memblock(memory, reg) {
297 memory_present(memblock_get_region_node(reg),
298 memblock_region_memory_base_pfn(reg),
299 memblock_region_memory_end_pfn(reg));
300 }
301}
302
303/*
304 * Subtle, we encode the real pfn into the mem_map such that
305 * the identity pfn - section_mem_map will return the actual
306 * physical page frame number.
307 */
308static unsigned long sparse_encode_mem_map(struct page *mem_map, unsigned long pnum)
309{
310 unsigned long coded_mem_map =
311 (unsigned long)(mem_map - (section_nr_to_pfn(pnum)));
312 BUILD_BUG_ON(SECTION_MAP_LAST_BIT > (1UL<<PFN_SECTION_SHIFT));
313 BUG_ON(coded_mem_map & ~SECTION_MAP_MASK);
314 return coded_mem_map;
315}
316
317/*
318 * Decode mem_map from the coded memmap
319 */
320struct page *sparse_decode_mem_map(unsigned long coded_mem_map, unsigned long pnum)
321{
322 /* mask off the extra low bits of information */
323 coded_mem_map &= SECTION_MAP_MASK;
324 return ((struct page *)coded_mem_map) + section_nr_to_pfn(pnum);
325}
326
327static void __meminit sparse_init_one_section(struct mem_section *ms,
328 unsigned long pnum, struct page *mem_map,
329 struct mem_section_usage *usage, unsigned long flags)
330{
331 ms->section_mem_map &= ~SECTION_MAP_MASK;
332 ms->section_mem_map |= sparse_encode_mem_map(mem_map, pnum)
333 | SECTION_HAS_MEM_MAP | flags;
334 ms->usage = usage;
335}
336
337static unsigned long usemap_size(void)
338{
339 return BITS_TO_LONGS(SECTION_BLOCKFLAGS_BITS) * sizeof(unsigned long);
340}
341
342size_t mem_section_usage_size(void)
343{
344 return sizeof(struct mem_section_usage) + usemap_size();
345}
346
347#ifdef CONFIG_MEMORY_HOTREMOVE
348static struct mem_section_usage * __init
349sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
350 unsigned long size)
351{
352 struct mem_section_usage *usage;
353 unsigned long goal, limit;
354 int nid;
355 /*
356 * A page may contain usemaps for other sections preventing the
357 * page being freed and making a section unremovable while
358 * other sections referencing the usemap remain active. Similarly,
359 * a pgdat can prevent a section being removed. If section A
360 * contains a pgdat and section B contains the usemap, both
361 * sections become inter-dependent. This allocates usemaps
362 * from the same section as the pgdat where possible to avoid
363 * this problem.
364 */
365 goal = __pa(pgdat) & (PAGE_SECTION_MASK << PAGE_SHIFT);
366 limit = goal + (1UL << PA_SECTION_SHIFT);
367 nid = early_pfn_to_nid(goal >> PAGE_SHIFT);
368again:
369 usage = memblock_alloc_try_nid(size, SMP_CACHE_BYTES, goal, limit, nid);
370 if (!usage && limit) {
371 limit = 0;
372 goto again;
373 }
374 return usage;
375}
376
377static void __init check_usemap_section_nr(int nid,
378 struct mem_section_usage *usage)
379{
380 unsigned long usemap_snr, pgdat_snr;
381 static unsigned long old_usemap_snr;
382 static unsigned long old_pgdat_snr;
383 struct pglist_data *pgdat = NODE_DATA(nid);
384 int usemap_nid;
385
386 /* First call */
387 if (!old_usemap_snr) {
388 old_usemap_snr = NR_MEM_SECTIONS;
389 old_pgdat_snr = NR_MEM_SECTIONS;
390 }
391
392 usemap_snr = pfn_to_section_nr(__pa(usage) >> PAGE_SHIFT);
393 pgdat_snr = pfn_to_section_nr(__pa(pgdat) >> PAGE_SHIFT);
394 if (usemap_snr == pgdat_snr)
395 return;
396
397 if (old_usemap_snr == usemap_snr && old_pgdat_snr == pgdat_snr)
398 /* skip redundant message */
399 return;
400
401 old_usemap_snr = usemap_snr;
402 old_pgdat_snr = pgdat_snr;
403
404 usemap_nid = sparse_early_nid(__nr_to_section(usemap_snr));
405 if (usemap_nid != nid) {
406 pr_info("node %d must be removed before remove section %ld\n",
407 nid, usemap_snr);
408 return;
409 }
410 /*
411 * There is a circular dependency.
412 * Some platforms allow un-removable section because they will just
413 * gather other removable sections for dynamic partitioning.
414 * Just notify un-removable section's number here.
415 */
416 pr_info("Section %ld and %ld (node %d) have a circular dependency on usemap and pgdat allocations\n",
417 usemap_snr, pgdat_snr, nid);
418}
419#else
420static struct mem_section_usage * __init
421sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
422 unsigned long size)
423{
424 return memblock_alloc_node(size, SMP_CACHE_BYTES, pgdat->node_id);
425}
426
427static void __init check_usemap_section_nr(int nid,
428 struct mem_section_usage *usage)
429{
430}
431#endif /* CONFIG_MEMORY_HOTREMOVE */
432
433#ifdef CONFIG_SPARSEMEM_VMEMMAP
434static unsigned long __init section_map_size(void)
435{
436 return ALIGN(sizeof(struct page) * PAGES_PER_SECTION, PMD_SIZE);
437}
438
439#else
440static unsigned long __init section_map_size(void)
441{
442 return PAGE_ALIGN(sizeof(struct page) * PAGES_PER_SECTION);
443}
444
445struct page __init *__populate_section_memmap(unsigned long pfn,
446 unsigned long nr_pages, int nid, struct vmem_altmap *altmap)
447{
448 unsigned long size = section_map_size();
449 struct page *map = sparse_buffer_alloc(size);
450 phys_addr_t addr = __pa(MAX_DMA_ADDRESS);
451
452 if (map)
453 return map;
454
455 map = memblock_alloc_try_nid_raw(size, size, addr,
456 MEMBLOCK_ALLOC_ACCESSIBLE, nid);
457 if (!map)
458 panic("%s: Failed to allocate %lu bytes align=0x%lx nid=%d from=%pa\n",
459 __func__, size, PAGE_SIZE, nid, &addr);
460
461 return map;
462}
463#endif /* !CONFIG_SPARSEMEM_VMEMMAP */
464
465static void *sparsemap_buf __meminitdata;
466static void *sparsemap_buf_end __meminitdata;
467
468static inline void __meminit sparse_buffer_free(unsigned long size)
469{
470 WARN_ON(!sparsemap_buf || size == 0);
471 memblock_free_early(__pa(sparsemap_buf), size);
472}
473
474static void __init sparse_buffer_init(unsigned long size, int nid)
475{
476 phys_addr_t addr = __pa(MAX_DMA_ADDRESS);
477 WARN_ON(sparsemap_buf); /* forgot to call sparse_buffer_fini()? */
478 /*
479 * Pre-allocated buffer is mainly used by __populate_section_memmap
480 * and we want it to be properly aligned to the section size - this is
481 * especially the case for VMEMMAP which maps memmap to PMDs
482 */
483 sparsemap_buf = memblock_alloc_exact_nid_raw(size, section_map_size(),
484 addr, MEMBLOCK_ALLOC_ACCESSIBLE, nid);
485 sparsemap_buf_end = sparsemap_buf + size;
486}
487
488static void __init sparse_buffer_fini(void)
489{
490 unsigned long size = sparsemap_buf_end - sparsemap_buf;
491
492 if (sparsemap_buf && size > 0)
493 sparse_buffer_free(size);
494 sparsemap_buf = NULL;
495}
496
497void * __meminit sparse_buffer_alloc(unsigned long size)
498{
499 void *ptr = NULL;
500
501 if (sparsemap_buf) {
502 ptr = (void *) roundup((unsigned long)sparsemap_buf, size);
503 if (ptr + size > sparsemap_buf_end)
504 ptr = NULL;
505 else {
506 /* Free redundant aligned space */
507 if ((unsigned long)(ptr - sparsemap_buf) > 0)
508 sparse_buffer_free((unsigned long)(ptr - sparsemap_buf));
509 sparsemap_buf = ptr + size;
510 }
511 }
512 return ptr;
513}
514
515void __weak __meminit vmemmap_populate_print_last(void)
516{
517}
518
519/*
520 * Initialize sparse on a specific node. The node spans [pnum_begin, pnum_end)
521 * And number of present sections in this node is map_count.
522 */
523static void __init sparse_init_nid(int nid, unsigned long pnum_begin,
524 unsigned long pnum_end,
525 unsigned long map_count)
526{
527 struct mem_section_usage *usage;
528 unsigned long pnum;
529 struct page *map;
530
531 usage = sparse_early_usemaps_alloc_pgdat_section(NODE_DATA(nid),
532 mem_section_usage_size() * map_count);
533 if (!usage) {
534 pr_err("%s: node[%d] usemap allocation failed", __func__, nid);
535 goto failed;
536 }
537 sparse_buffer_init(map_count * section_map_size(), nid);
538 for_each_present_section_nr(pnum_begin, pnum) {
539 unsigned long pfn = section_nr_to_pfn(pnum);
540
541 if (pnum >= pnum_end)
542 break;
543
544 map = __populate_section_memmap(pfn, PAGES_PER_SECTION,
545 nid, NULL);
546 if (!map) {
547 pr_err("%s: node[%d] memory map backing failed. Some memory will not be available.",
548 __func__, nid);
549 pnum_begin = pnum;
550 goto failed;
551 }
552 check_usemap_section_nr(nid, usage);
553 sparse_init_one_section(__nr_to_section(pnum), pnum, map, usage,
554 SECTION_IS_EARLY);
555 usage = (void *) usage + mem_section_usage_size();
556 }
557 sparse_buffer_fini();
558 return;
559failed:
560 /* We failed to allocate, mark all the following pnums as not present */
561 for_each_present_section_nr(pnum_begin, pnum) {
562 struct mem_section *ms;
563
564 if (pnum >= pnum_end)
565 break;
566 ms = __nr_to_section(pnum);
567 ms->section_mem_map = 0;
568 }
569}
570
571/*
572 * Allocate the accumulated non-linear sections, allocate a mem_map
573 * for each and record the physical to section mapping.
574 */
575void __init sparse_init(void)
576{
577 unsigned long pnum_end, pnum_begin, map_count = 1;
578 int nid_begin;
579
580 memblocks_present();
581
582 pnum_begin = first_present_section_nr();
583 nid_begin = sparse_early_nid(__nr_to_section(pnum_begin));
584
585 /* Setup pageblock_order for HUGETLB_PAGE_SIZE_VARIABLE */
586 set_pageblock_order();
587
588 for_each_present_section_nr(pnum_begin + 1, pnum_end) {
589 int nid = sparse_early_nid(__nr_to_section(pnum_end));
590
591 if (nid == nid_begin) {
592 map_count++;
593 continue;
594 }
595 /* Init node with sections in range [pnum_begin, pnum_end) */
596 sparse_init_nid(nid_begin, pnum_begin, pnum_end, map_count);
597 nid_begin = nid;
598 pnum_begin = pnum_end;
599 map_count = 1;
600 }
601 /* cover the last node */
602 sparse_init_nid(nid_begin, pnum_begin, pnum_end, map_count);
603 vmemmap_populate_print_last();
604}
605
606#ifdef CONFIG_MEMORY_HOTPLUG
607
608/* Mark all memory sections within the pfn range as online */
609void online_mem_sections(unsigned long start_pfn, unsigned long end_pfn)
610{
611 unsigned long pfn;
612
613 for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
614 unsigned long section_nr = pfn_to_section_nr(pfn);
615 struct mem_section *ms;
616
617 /* onlining code should never touch invalid ranges */
618 if (WARN_ON(!valid_section_nr(section_nr)))
619 continue;
620
621 ms = __nr_to_section(section_nr);
622 ms->section_mem_map |= SECTION_IS_ONLINE;
623 }
624}
625
626#ifdef CONFIG_MEMORY_HOTREMOVE
627/* Mark all memory sections within the pfn range as offline */
628void offline_mem_sections(unsigned long start_pfn, unsigned long end_pfn)
629{
630 unsigned long pfn;
631
632 for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
633 unsigned long section_nr = pfn_to_section_nr(pfn);
634 struct mem_section *ms;
635
636 /*
637 * TODO this needs some double checking. Offlining code makes
638 * sure to check pfn_valid but those checks might be just bogus
639 */
640 if (WARN_ON(!valid_section_nr(section_nr)))
641 continue;
642
643 ms = __nr_to_section(section_nr);
644 ms->section_mem_map &= ~SECTION_IS_ONLINE;
645 }
646}
647#endif
648
649#ifdef CONFIG_SPARSEMEM_VMEMMAP
650static struct page * __meminit populate_section_memmap(unsigned long pfn,
651 unsigned long nr_pages, int nid, struct vmem_altmap *altmap)
652{
653 return __populate_section_memmap(pfn, nr_pages, nid, altmap);
654}
655
656static void depopulate_section_memmap(unsigned long pfn, unsigned long nr_pages,
657 struct vmem_altmap *altmap)
658{
659 unsigned long start = (unsigned long) pfn_to_page(pfn);
660 unsigned long end = start + nr_pages * sizeof(struct page);
661
662 vmemmap_free(start, end, altmap);
663}
664static void free_map_bootmem(struct page *memmap)
665{
666 unsigned long start = (unsigned long)memmap;
667 unsigned long end = (unsigned long)(memmap + PAGES_PER_SECTION);
668
669 vmemmap_free(start, end, NULL);
670}
671
672static int clear_subsection_map(unsigned long pfn, unsigned long nr_pages)
673{
674 DECLARE_BITMAP(map, SUBSECTIONS_PER_SECTION) = { 0 };
675 DECLARE_BITMAP(tmp, SUBSECTIONS_PER_SECTION) = { 0 };
676 struct mem_section *ms = __pfn_to_section(pfn);
677 unsigned long *subsection_map = ms->usage
678 ? &ms->usage->subsection_map[0] : NULL;
679
680 subsection_mask_set(map, pfn, nr_pages);
681 if (subsection_map)
682 bitmap_and(tmp, map, subsection_map, SUBSECTIONS_PER_SECTION);
683
684 if (WARN(!subsection_map || !bitmap_equal(tmp, map, SUBSECTIONS_PER_SECTION),
685 "section already deactivated (%#lx + %ld)\n",
686 pfn, nr_pages))
687 return -EINVAL;
688
689 bitmap_xor(subsection_map, map, subsection_map, SUBSECTIONS_PER_SECTION);
690 return 0;
691}
692
693static bool is_subsection_map_empty(struct mem_section *ms)
694{
695 return bitmap_empty(&ms->usage->subsection_map[0],
696 SUBSECTIONS_PER_SECTION);
697}
698
699static int fill_subsection_map(unsigned long pfn, unsigned long nr_pages)
700{
701 struct mem_section *ms = __pfn_to_section(pfn);
702 DECLARE_BITMAP(map, SUBSECTIONS_PER_SECTION) = { 0 };
703 unsigned long *subsection_map;
704 int rc = 0;
705
706 subsection_mask_set(map, pfn, nr_pages);
707
708 subsection_map = &ms->usage->subsection_map[0];
709
710 if (bitmap_empty(map, SUBSECTIONS_PER_SECTION))
711 rc = -EINVAL;
712 else if (bitmap_intersects(map, subsection_map, SUBSECTIONS_PER_SECTION))
713 rc = -EEXIST;
714 else
715 bitmap_or(subsection_map, map, subsection_map,
716 SUBSECTIONS_PER_SECTION);
717
718 return rc;
719}
720#else
721struct page * __meminit populate_section_memmap(unsigned long pfn,
722 unsigned long nr_pages, int nid, struct vmem_altmap *altmap)
723{
724 return kvmalloc_node(array_size(sizeof(struct page),
725 PAGES_PER_SECTION), GFP_KERNEL, nid);
726}
727
728static void depopulate_section_memmap(unsigned long pfn, unsigned long nr_pages,
729 struct vmem_altmap *altmap)
730{
731 kvfree(pfn_to_page(pfn));
732}
733
734static void free_map_bootmem(struct page *memmap)
735{
736 unsigned long maps_section_nr, removing_section_nr, i;
737 unsigned long magic, nr_pages;
738 struct page *page = virt_to_page(memmap);
739
740 nr_pages = PAGE_ALIGN(PAGES_PER_SECTION * sizeof(struct page))
741 >> PAGE_SHIFT;
742
743 for (i = 0; i < nr_pages; i++, page++) {
744 magic = (unsigned long) page->freelist;
745
746 BUG_ON(magic == NODE_INFO);
747
748 maps_section_nr = pfn_to_section_nr(page_to_pfn(page));
749 removing_section_nr = page_private(page);
750
751 /*
752 * When this function is called, the removing section is
753 * logical offlined state. This means all pages are isolated
754 * from page allocator. If removing section's memmap is placed
755 * on the same section, it must not be freed.
756 * If it is freed, page allocator may allocate it which will
757 * be removed physically soon.
758 */
759 if (maps_section_nr != removing_section_nr)
760 put_page_bootmem(page);
761 }
762}
763
764static int clear_subsection_map(unsigned long pfn, unsigned long nr_pages)
765{
766 return 0;
767}
768
769static bool is_subsection_map_empty(struct mem_section *ms)
770{
771 return true;
772}
773
774static int fill_subsection_map(unsigned long pfn, unsigned long nr_pages)
775{
776 return 0;
777}
778#endif /* CONFIG_SPARSEMEM_VMEMMAP */
779
780/*
781 * To deactivate a memory region, there are 3 cases to handle across
782 * two configurations (SPARSEMEM_VMEMMAP={y,n}):
783 *
784 * 1. deactivation of a partial hot-added section (only possible in
785 * the SPARSEMEM_VMEMMAP=y case).
786 * a) section was present at memory init.
787 * b) section was hot-added post memory init.
788 * 2. deactivation of a complete hot-added section.
789 * 3. deactivation of a complete section from memory init.
790 *
791 * For 1, when subsection_map does not empty we will not be freeing the
792 * usage map, but still need to free the vmemmap range.
793 *
794 * For 2 and 3, the SPARSEMEM_VMEMMAP={y,n} cases are unified
795 */
796static void section_deactivate(unsigned long pfn, unsigned long nr_pages,
797 struct vmem_altmap *altmap)
798{
799 struct mem_section *ms = __pfn_to_section(pfn);
800 bool section_is_early = early_section(ms);
801 struct page *memmap = NULL;
802 bool empty;
803
804 if (clear_subsection_map(pfn, nr_pages))
805 return;
806
807 empty = is_subsection_map_empty(ms);
808 if (empty) {
809 unsigned long section_nr = pfn_to_section_nr(pfn);
810
811 /*
812 * When removing an early section, the usage map is kept (as the
813 * usage maps of other sections fall into the same page). It
814 * will be re-used when re-adding the section - which is then no
815 * longer an early section. If the usage map is PageReserved, it
816 * was allocated during boot.
817 */
818 if (!PageReserved(virt_to_page(ms->usage))) {
819 kfree(ms->usage);
820 ms->usage = NULL;
821 }
822 memmap = sparse_decode_mem_map(ms->section_mem_map, section_nr);
823 /*
824 * Mark the section invalid so that valid_section()
825 * return false. This prevents code from dereferencing
826 * ms->usage array.
827 */
828 ms->section_mem_map &= ~SECTION_HAS_MEM_MAP;
829 }
830
831 /*
832 * The memmap of early sections is always fully populated. See
833 * section_activate() and pfn_valid() .
834 */
835 if (!section_is_early)
836 depopulate_section_memmap(pfn, nr_pages, altmap);
837 else if (memmap)
838 free_map_bootmem(memmap);
839
840 if (empty)
841 ms->section_mem_map = (unsigned long)NULL;
842}
843
844static struct page * __meminit section_activate(int nid, unsigned long pfn,
845 unsigned long nr_pages, struct vmem_altmap *altmap)
846{
847 struct mem_section *ms = __pfn_to_section(pfn);
848 struct mem_section_usage *usage = NULL;
849 struct page *memmap;
850 int rc = 0;
851
852 if (!ms->usage) {
853 usage = kzalloc(mem_section_usage_size(), GFP_KERNEL);
854 if (!usage)
855 return ERR_PTR(-ENOMEM);
856 ms->usage = usage;
857 }
858
859 rc = fill_subsection_map(pfn, nr_pages);
860 if (rc) {
861 if (usage)
862 ms->usage = NULL;
863 kfree(usage);
864 return ERR_PTR(rc);
865 }
866
867 /*
868 * The early init code does not consider partially populated
869 * initial sections, it simply assumes that memory will never be
870 * referenced. If we hot-add memory into such a section then we
871 * do not need to populate the memmap and can simply reuse what
872 * is already there.
873 */
874 if (nr_pages < PAGES_PER_SECTION && early_section(ms))
875 return pfn_to_page(pfn);
876
877 memmap = populate_section_memmap(pfn, nr_pages, nid, altmap);
878 if (!memmap) {
879 section_deactivate(pfn, nr_pages, altmap);
880 return ERR_PTR(-ENOMEM);
881 }
882
883 return memmap;
884}
885
886/**
887 * sparse_add_section - add a memory section, or populate an existing one
888 * @nid: The node to add section on
889 * @start_pfn: start pfn of the memory range
890 * @nr_pages: number of pfns to add in the section
891 * @altmap: device page map
892 *
893 * This is only intended for hotplug.
894 *
895 * Note that only VMEMMAP supports sub-section aligned hotplug,
896 * the proper alignment and size are gated by check_pfn_span().
897 *
898 *
899 * Return:
900 * * 0 - On success.
901 * * -EEXIST - Section has been present.
902 * * -ENOMEM - Out of memory.
903 */
904int __meminit sparse_add_section(int nid, unsigned long start_pfn,
905 unsigned long nr_pages, struct vmem_altmap *altmap)
906{
907 unsigned long section_nr = pfn_to_section_nr(start_pfn);
908 struct mem_section *ms;
909 struct page *memmap;
910 int ret;
911
912 ret = sparse_index_init(section_nr, nid);
913 if (ret < 0)
914 return ret;
915
916 memmap = section_activate(nid, start_pfn, nr_pages, altmap);
917 if (IS_ERR(memmap))
918 return PTR_ERR(memmap);
919
920 /*
921 * Poison uninitialized struct pages in order to catch invalid flags
922 * combinations.
923 */
924 page_init_poison(memmap, sizeof(struct page) * nr_pages);
925
926 ms = __nr_to_section(section_nr);
927 set_section_nid(section_nr, nid);
928 section_mark_present(ms);
929
930 /* Align memmap to section boundary in the subsection case */
931 if (section_nr_to_pfn(section_nr) != start_pfn)
932 memmap = pfn_to_page(section_nr_to_pfn(section_nr));
933 sparse_init_one_section(ms, section_nr, memmap, ms->usage, 0);
934
935 return 0;
936}
937
938#ifdef CONFIG_MEMORY_FAILURE
939static void clear_hwpoisoned_pages(struct page *memmap, int nr_pages)
940{
941 int i;
942
943 /*
944 * A further optimization is to have per section refcounted
945 * num_poisoned_pages. But that would need more space per memmap, so
946 * for now just do a quick global check to speed up this routine in the
947 * absence of bad pages.
948 */
949 if (atomic_long_read(&num_poisoned_pages) == 0)
950 return;
951
952 for (i = 0; i < nr_pages; i++) {
953 if (PageHWPoison(&memmap[i])) {
954 num_poisoned_pages_dec();
955 ClearPageHWPoison(&memmap[i]);
956 }
957 }
958}
959#else
960static inline void clear_hwpoisoned_pages(struct page *memmap, int nr_pages)
961{
962}
963#endif
964
965void sparse_remove_section(struct mem_section *ms, unsigned long pfn,
966 unsigned long nr_pages, unsigned long map_offset,
967 struct vmem_altmap *altmap)
968{
969 clear_hwpoisoned_pages(pfn_to_page(pfn) + map_offset,
970 nr_pages - map_offset);
971 section_deactivate(pfn, nr_pages, altmap);
972}
973#endif /* CONFIG_MEMORY_HOTPLUG */
1// 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 */