projects/linux·blob v5.13 · refs

981 lines · 26.6 KiB

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 unsigned long start, end;
295 int i, nid;
296
297 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid)
298 memory_present(nid, start, end);
299}
300
301/*
302 * Subtle, we encode the real pfn into the mem_map such that
303 * the identity pfn - section_mem_map will return the actual
304 * physical page frame number.
305 */
306static unsigned long sparse_encode_mem_map(struct page *mem_map, unsigned long pnum)
307{
308 unsigned long coded_mem_map =
309 (unsigned long)(mem_map - (section_nr_to_pfn(pnum)));
310 BUILD_BUG_ON(SECTION_MAP_LAST_BIT > (1UL<<PFN_SECTION_SHIFT));
311 BUG_ON(coded_mem_map & ~SECTION_MAP_MASK);
312 return coded_mem_map;
313}
314
315#ifdef CONFIG_MEMORY_HOTPLUG
316/*
317 * Decode mem_map from the coded memmap
318 */
319struct page *sparse_decode_mem_map(unsigned long coded_mem_map, unsigned long pnum)
320{
321 /* mask off the extra low bits of information */
322 coded_mem_map &= SECTION_MAP_MASK;
323 return ((struct page *)coded_mem_map) + section_nr_to_pfn(pnum);
324}
325#endif /* CONFIG_MEMORY_HOTPLUG */
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
347static inline phys_addr_t pgdat_to_phys(struct pglist_data *pgdat)
348{
349#ifndef CONFIG_NEED_MULTIPLE_NODES
350 return __pa_symbol(pgdat);
351#else
352 return __pa(pgdat);
353#endif
354}
355
356#ifdef CONFIG_MEMORY_HOTREMOVE
357static struct mem_section_usage * __init
358sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
359 unsigned long size)
360{
361 struct mem_section_usage *usage;
362 unsigned long goal, limit;
363 int nid;
364 /*
365 * A page may contain usemaps for other sections preventing the
366 * page being freed and making a section unremovable while
367 * other sections referencing the usemap remain active. Similarly,
368 * a pgdat can prevent a section being removed. If section A
369 * contains a pgdat and section B contains the usemap, both
370 * sections become inter-dependent. This allocates usemaps
371 * from the same section as the pgdat where possible to avoid
372 * this problem.
373 */
374 goal = pgdat_to_phys(pgdat) & (PAGE_SECTION_MASK << PAGE_SHIFT);
375 limit = goal + (1UL << PA_SECTION_SHIFT);
376 nid = early_pfn_to_nid(goal >> PAGE_SHIFT);
377again:
378 usage = memblock_alloc_try_nid(size, SMP_CACHE_BYTES, goal, limit, nid);
379 if (!usage && limit) {
380 limit = 0;
381 goto again;
382 }
383 return usage;
384}
385
386static void __init check_usemap_section_nr(int nid,
387 struct mem_section_usage *usage)
388{
389 unsigned long usemap_snr, pgdat_snr;
390 static unsigned long old_usemap_snr;
391 static unsigned long old_pgdat_snr;
392 struct pglist_data *pgdat = NODE_DATA(nid);
393 int usemap_nid;
394
395 /* First call */
396 if (!old_usemap_snr) {
397 old_usemap_snr = NR_MEM_SECTIONS;
398 old_pgdat_snr = NR_MEM_SECTIONS;
399 }
400
401 usemap_snr = pfn_to_section_nr(__pa(usage) >> PAGE_SHIFT);
402 pgdat_snr = pfn_to_section_nr(pgdat_to_phys(pgdat) >> PAGE_SHIFT);
403 if (usemap_snr == pgdat_snr)
404 return;
405
406 if (old_usemap_snr == usemap_snr && old_pgdat_snr == pgdat_snr)
407 /* skip redundant message */
408 return;
409
410 old_usemap_snr = usemap_snr;
411 old_pgdat_snr = pgdat_snr;
412
413 usemap_nid = sparse_early_nid(__nr_to_section(usemap_snr));
414 if (usemap_nid != nid) {
415 pr_info("node %d must be removed before remove section %ld\n",
416 nid, usemap_snr);
417 return;
418 }
419 /*
420 * There is a circular dependency.
421 * Some platforms allow un-removable section because they will just
422 * gather other removable sections for dynamic partitioning.
423 * Just notify un-removable section's number here.
424 */
425 pr_info("Section %ld and %ld (node %d) have a circular dependency on usemap and pgdat allocations\n",
426 usemap_snr, pgdat_snr, nid);
427}
428#else
429static struct mem_section_usage * __init
430sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
431 unsigned long size)
432{
433 return memblock_alloc_node(size, SMP_CACHE_BYTES, pgdat->node_id);
434}
435
436static void __init check_usemap_section_nr(int nid,
437 struct mem_section_usage *usage)
438{
439}
440#endif /* CONFIG_MEMORY_HOTREMOVE */
441
442#ifdef CONFIG_SPARSEMEM_VMEMMAP
443static unsigned long __init section_map_size(void)
444{
445 return ALIGN(sizeof(struct page) * PAGES_PER_SECTION, PMD_SIZE);
446}
447
448#else
449static unsigned long __init section_map_size(void)
450{
451 return PAGE_ALIGN(sizeof(struct page) * PAGES_PER_SECTION);
452}
453
454struct page __init *__populate_section_memmap(unsigned long pfn,
455 unsigned long nr_pages, int nid, struct vmem_altmap *altmap)
456{
457 unsigned long size = section_map_size();
458 struct page *map = sparse_buffer_alloc(size);
459 phys_addr_t addr = __pa(MAX_DMA_ADDRESS);
460
461 if (map)
462 return map;
463
464 map = memblock_alloc_try_nid_raw(size, size, addr,
465 MEMBLOCK_ALLOC_ACCESSIBLE, nid);
466 if (!map)
467 panic("%s: Failed to allocate %lu bytes align=0x%lx nid=%d from=%pa\n",
468 __func__, size, PAGE_SIZE, nid, &addr);
469
470 return map;
471}
472#endif /* !CONFIG_SPARSEMEM_VMEMMAP */
473
474static void *sparsemap_buf __meminitdata;
475static void *sparsemap_buf_end __meminitdata;
476
477static inline void __meminit sparse_buffer_free(unsigned long size)
478{
479 WARN_ON(!sparsemap_buf || size == 0);
480 memblock_free_early(__pa(sparsemap_buf), size);
481}
482
483static void __init sparse_buffer_init(unsigned long size, int nid)
484{
485 phys_addr_t addr = __pa(MAX_DMA_ADDRESS);
486 WARN_ON(sparsemap_buf); /* forgot to call sparse_buffer_fini()? */
487 /*
488 * Pre-allocated buffer is mainly used by __populate_section_memmap
489 * and we want it to be properly aligned to the section size - this is
490 * especially the case for VMEMMAP which maps memmap to PMDs
491 */
492 sparsemap_buf = memblock_alloc_exact_nid_raw(size, section_map_size(),
493 addr, MEMBLOCK_ALLOC_ACCESSIBLE, nid);
494 sparsemap_buf_end = sparsemap_buf + size;
495}
496
497static void __init sparse_buffer_fini(void)
498{
499 unsigned long size = sparsemap_buf_end - sparsemap_buf;
500
501 if (sparsemap_buf && size > 0)
502 sparse_buffer_free(size);
503 sparsemap_buf = NULL;
504}
505
506void * __meminit sparse_buffer_alloc(unsigned long size)
507{
508 void *ptr = NULL;
509
510 if (sparsemap_buf) {
511 ptr = (void *) roundup((unsigned long)sparsemap_buf, size);
512 if (ptr + size > sparsemap_buf_end)
513 ptr = NULL;
514 else {
515 /* Free redundant aligned space */
516 if ((unsigned long)(ptr - sparsemap_buf) > 0)
517 sparse_buffer_free((unsigned long)(ptr - sparsemap_buf));
518 sparsemap_buf = ptr + size;
519 }
520 }
521 return ptr;
522}
523
524void __weak __meminit vmemmap_populate_print_last(void)
525{
526}
527
528/*
529 * Initialize sparse on a specific node. The node spans [pnum_begin, pnum_end)
530 * And number of present sections in this node is map_count.
531 */
532static void __init sparse_init_nid(int nid, unsigned long pnum_begin,
533 unsigned long pnum_end,
534 unsigned long map_count)
535{
536 struct mem_section_usage *usage;
537 unsigned long pnum;
538 struct page *map;
539
540 usage = sparse_early_usemaps_alloc_pgdat_section(NODE_DATA(nid),
541 mem_section_usage_size() * map_count);
542 if (!usage) {
543 pr_err("%s: node[%d] usemap allocation failed", __func__, nid);
544 goto failed;
545 }
546 sparse_buffer_init(map_count * section_map_size(), nid);
547 for_each_present_section_nr(pnum_begin, pnum) {
548 unsigned long pfn = section_nr_to_pfn(pnum);
549
550 if (pnum >= pnum_end)
551 break;
552
553 map = __populate_section_memmap(pfn, PAGES_PER_SECTION,
554 nid, NULL);
555 if (!map) {
556 pr_err("%s: node[%d] memory map backing failed. Some memory will not be available.",
557 __func__, nid);
558 pnum_begin = pnum;
559 sparse_buffer_fini();
560 goto failed;
561 }
562 check_usemap_section_nr(nid, usage);
563 sparse_init_one_section(__nr_to_section(pnum), pnum, map, usage,
564 SECTION_IS_EARLY);
565 usage = (void *) usage + mem_section_usage_size();
566 }
567 sparse_buffer_fini();
568 return;
569failed:
570 /* We failed to allocate, mark all the following pnums as not present */
571 for_each_present_section_nr(pnum_begin, pnum) {
572 struct mem_section *ms;
573
574 if (pnum >= pnum_end)
575 break;
576 ms = __nr_to_section(pnum);
577 ms->section_mem_map = 0;
578 }
579}
580
581/*
582 * Allocate the accumulated non-linear sections, allocate a mem_map
583 * for each and record the physical to section mapping.
584 */
585void __init sparse_init(void)
586{
587 unsigned long pnum_end, pnum_begin, map_count = 1;
588 int nid_begin;
589
590 memblocks_present();
591
592 pnum_begin = first_present_section_nr();
593 nid_begin = sparse_early_nid(__nr_to_section(pnum_begin));
594
595 /* Setup pageblock_order for HUGETLB_PAGE_SIZE_VARIABLE */
596 set_pageblock_order();
597
598 for_each_present_section_nr(pnum_begin + 1, pnum_end) {
599 int nid = sparse_early_nid(__nr_to_section(pnum_end));
600
601 if (nid == nid_begin) {
602 map_count++;
603 continue;
604 }
605 /* Init node with sections in range [pnum_begin, pnum_end) */
606 sparse_init_nid(nid_begin, pnum_begin, pnum_end, map_count);
607 nid_begin = nid;
608 pnum_begin = pnum_end;
609 map_count = 1;
610 }
611 /* cover the last node */
612 sparse_init_nid(nid_begin, pnum_begin, pnum_end, map_count);
613 vmemmap_populate_print_last();
614}
615
616#ifdef CONFIG_MEMORY_HOTPLUG
617
618/* Mark all memory sections within the pfn range as online */
619void online_mem_sections(unsigned long start_pfn, unsigned long end_pfn)
620{
621 unsigned long pfn;
622
623 for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
624 unsigned long section_nr = pfn_to_section_nr(pfn);
625 struct mem_section *ms;
626
627 /* onlining code should never touch invalid ranges */
628 if (WARN_ON(!valid_section_nr(section_nr)))
629 continue;
630
631 ms = __nr_to_section(section_nr);
632 ms->section_mem_map |= SECTION_IS_ONLINE;
633 }
634}
635
636/* Mark all memory sections within the pfn range as offline */
637void offline_mem_sections(unsigned long start_pfn, unsigned long end_pfn)
638{
639 unsigned long pfn;
640
641 for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
642 unsigned long section_nr = pfn_to_section_nr(pfn);
643 struct mem_section *ms;
644
645 /*
646 * TODO this needs some double checking. Offlining code makes
647 * sure to check pfn_valid but those checks might be just bogus
648 */
649 if (WARN_ON(!valid_section_nr(section_nr)))
650 continue;
651
652 ms = __nr_to_section(section_nr);
653 ms->section_mem_map &= ~SECTION_IS_ONLINE;
654 }
655}
656
657#ifdef CONFIG_SPARSEMEM_VMEMMAP
658static struct page * __meminit populate_section_memmap(unsigned long pfn,
659 unsigned long nr_pages, int nid, struct vmem_altmap *altmap)
660{
661 return __populate_section_memmap(pfn, nr_pages, nid, altmap);
662}
663
664static void depopulate_section_memmap(unsigned long pfn, unsigned long nr_pages,
665 struct vmem_altmap *altmap)
666{
667 unsigned long start = (unsigned long) pfn_to_page(pfn);
668 unsigned long end = start + nr_pages * sizeof(struct page);
669
670 vmemmap_free(start, end, altmap);
671}
672static void free_map_bootmem(struct page *memmap)
673{
674 unsigned long start = (unsigned long)memmap;
675 unsigned long end = (unsigned long)(memmap + PAGES_PER_SECTION);
676
677 vmemmap_free(start, end, NULL);
678}
679
680static int clear_subsection_map(unsigned long pfn, unsigned long nr_pages)
681{
682 DECLARE_BITMAP(map, SUBSECTIONS_PER_SECTION) = { 0 };
683 DECLARE_BITMAP(tmp, SUBSECTIONS_PER_SECTION) = { 0 };
684 struct mem_section *ms = __pfn_to_section(pfn);
685 unsigned long *subsection_map = ms->usage
686 ? &ms->usage->subsection_map[0] : NULL;
687
688 subsection_mask_set(map, pfn, nr_pages);
689 if (subsection_map)
690 bitmap_and(tmp, map, subsection_map, SUBSECTIONS_PER_SECTION);
691
692 if (WARN(!subsection_map || !bitmap_equal(tmp, map, SUBSECTIONS_PER_SECTION),
693 "section already deactivated (%#lx + %ld)\n",
694 pfn, nr_pages))
695 return -EINVAL;
696
697 bitmap_xor(subsection_map, map, subsection_map, SUBSECTIONS_PER_SECTION);
698 return 0;
699}
700
701static bool is_subsection_map_empty(struct mem_section *ms)
702{
703 return bitmap_empty(&ms->usage->subsection_map[0],
704 SUBSECTIONS_PER_SECTION);
705}
706
707static int fill_subsection_map(unsigned long pfn, unsigned long nr_pages)
708{
709 struct mem_section *ms = __pfn_to_section(pfn);
710 DECLARE_BITMAP(map, SUBSECTIONS_PER_SECTION) = { 0 };
711 unsigned long *subsection_map;
712 int rc = 0;
713
714 subsection_mask_set(map, pfn, nr_pages);
715
716 subsection_map = &ms->usage->subsection_map[0];
717
718 if (bitmap_empty(map, SUBSECTIONS_PER_SECTION))
719 rc = -EINVAL;
720 else if (bitmap_intersects(map, subsection_map, SUBSECTIONS_PER_SECTION))
721 rc = -EEXIST;
722 else
723 bitmap_or(subsection_map, map, subsection_map,
724 SUBSECTIONS_PER_SECTION);
725
726 return rc;
727}
728#else
729struct page * __meminit populate_section_memmap(unsigned long pfn,
730 unsigned long nr_pages, int nid, struct vmem_altmap *altmap)
731{
732 return kvmalloc_node(array_size(sizeof(struct page),
733 PAGES_PER_SECTION), GFP_KERNEL, nid);
734}
735
736static void depopulate_section_memmap(unsigned long pfn, unsigned long nr_pages,
737 struct vmem_altmap *altmap)
738{
739 kvfree(pfn_to_page(pfn));
740}
741
742static void free_map_bootmem(struct page *memmap)
743{
744 unsigned long maps_section_nr, removing_section_nr, i;
745 unsigned long magic, nr_pages;
746 struct page *page = virt_to_page(memmap);
747
748 nr_pages = PAGE_ALIGN(PAGES_PER_SECTION * sizeof(struct page))
749 >> PAGE_SHIFT;
750
751 for (i = 0; i < nr_pages; i++, page++) {
752 magic = (unsigned long) page->freelist;
753
754 BUG_ON(magic == NODE_INFO);
755
756 maps_section_nr = pfn_to_section_nr(page_to_pfn(page));
757 removing_section_nr = page_private(page);
758
759 /*
760 * When this function is called, the removing section is
761 * logical offlined state. This means all pages are isolated
762 * from page allocator. If removing section's memmap is placed
763 * on the same section, it must not be freed.
764 * If it is freed, page allocator may allocate it which will
765 * be removed physically soon.
766 */
767 if (maps_section_nr != removing_section_nr)
768 put_page_bootmem(page);
769 }
770}
771
772static int clear_subsection_map(unsigned long pfn, unsigned long nr_pages)
773{
774 return 0;
775}
776
777static bool is_subsection_map_empty(struct mem_section *ms)
778{
779 return true;
780}
781
782static int fill_subsection_map(unsigned long pfn, unsigned long nr_pages)
783{
784 return 0;
785}
786#endif /* CONFIG_SPARSEMEM_VMEMMAP */
787
788/*
789 * To deactivate a memory region, there are 3 cases to handle across
790 * two configurations (SPARSEMEM_VMEMMAP={y,n}):
791 *
792 * 1. deactivation of a partial hot-added section (only possible in
793 * the SPARSEMEM_VMEMMAP=y case).
794 * a) section was present at memory init.
795 * b) section was hot-added post memory init.
796 * 2. deactivation of a complete hot-added section.
797 * 3. deactivation of a complete section from memory init.
798 *
799 * For 1, when subsection_map does not empty we will not be freeing the
800 * usage map, but still need to free the vmemmap range.
801 *
802 * For 2 and 3, the SPARSEMEM_VMEMMAP={y,n} cases are unified
803 */
804static void section_deactivate(unsigned long pfn, unsigned long nr_pages,
805 struct vmem_altmap *altmap)
806{
807 struct mem_section *ms = __pfn_to_section(pfn);
808 bool section_is_early = early_section(ms);
809 struct page *memmap = NULL;
810 bool empty;
811
812 if (clear_subsection_map(pfn, nr_pages))
813 return;
814
815 empty = is_subsection_map_empty(ms);
816 if (empty) {
817 unsigned long section_nr = pfn_to_section_nr(pfn);
818
819 /*
820 * When removing an early section, the usage map is kept (as the
821 * usage maps of other sections fall into the same page). It
822 * will be re-used when re-adding the section - which is then no
823 * longer an early section. If the usage map is PageReserved, it
824 * was allocated during boot.
825 */
826 if (!PageReserved(virt_to_page(ms->usage))) {
827 kfree(ms->usage);
828 ms->usage = NULL;
829 }
830 memmap = sparse_decode_mem_map(ms->section_mem_map, section_nr);
831 /*
832 * Mark the section invalid so that valid_section()
833 * return false. This prevents code from dereferencing
834 * ms->usage array.
835 */
836 ms->section_mem_map &= ~SECTION_HAS_MEM_MAP;
837 }
838
839 /*
840 * The memmap of early sections is always fully populated. See
841 * section_activate() and pfn_valid() .
842 */
843 if (!section_is_early)
844 depopulate_section_memmap(pfn, nr_pages, altmap);
845 else if (memmap)
846 free_map_bootmem(memmap);
847
848 if (empty)
849 ms->section_mem_map = (unsigned long)NULL;
850}
851
852static struct page * __meminit section_activate(int nid, unsigned long pfn,
853 unsigned long nr_pages, struct vmem_altmap *altmap)
854{
855 struct mem_section *ms = __pfn_to_section(pfn);
856 struct mem_section_usage *usage = NULL;
857 struct page *memmap;
858 int rc = 0;
859
860 if (!ms->usage) {
861 usage = kzalloc(mem_section_usage_size(), GFP_KERNEL);
862 if (!usage)
863 return ERR_PTR(-ENOMEM);
864 ms->usage = usage;
865 }
866
867 rc = fill_subsection_map(pfn, nr_pages);
868 if (rc) {
869 if (usage)
870 ms->usage = NULL;
871 kfree(usage);
872 return ERR_PTR(rc);
873 }
874
875 /*
876 * The early init code does not consider partially populated
877 * initial sections, it simply assumes that memory will never be
878 * referenced. If we hot-add memory into such a section then we
879 * do not need to populate the memmap and can simply reuse what
880 * is already there.
881 */
882 if (nr_pages < PAGES_PER_SECTION && early_section(ms))
883 return pfn_to_page(pfn);
884
885 memmap = populate_section_memmap(pfn, nr_pages, nid, altmap);
886 if (!memmap) {
887 section_deactivate(pfn, nr_pages, altmap);
888 return ERR_PTR(-ENOMEM);
889 }
890
891 return memmap;
892}
893
894/**
895 * sparse_add_section - add a memory section, or populate an existing one
896 * @nid: The node to add section on
897 * @start_pfn: start pfn of the memory range
898 * @nr_pages: number of pfns to add in the section
899 * @altmap: device page map
900 *
901 * This is only intended for hotplug.
902 *
903 * Note that only VMEMMAP supports sub-section aligned hotplug,
904 * the proper alignment and size are gated by check_pfn_span().
905 *
906 *
907 * Return:
908 * * 0 - On success.
909 * * -EEXIST - Section has been present.
910 * * -ENOMEM - Out of memory.
911 */
912int __meminit sparse_add_section(int nid, unsigned long start_pfn,
913 unsigned long nr_pages, struct vmem_altmap *altmap)
914{
915 unsigned long section_nr = pfn_to_section_nr(start_pfn);
916 struct mem_section *ms;
917 struct page *memmap;
918 int ret;
919
920 ret = sparse_index_init(section_nr, nid);
921 if (ret < 0)
922 return ret;
923
924 memmap = section_activate(nid, start_pfn, nr_pages, altmap);
925 if (IS_ERR(memmap))
926 return PTR_ERR(memmap);
927
928 /*
929 * Poison uninitialized struct pages in order to catch invalid flags
930 * combinations.
931 */
932 page_init_poison(memmap, sizeof(struct page) * nr_pages);
933
934 ms = __nr_to_section(section_nr);
935 set_section_nid(section_nr, nid);
936 section_mark_present(ms);
937
938 /* Align memmap to section boundary in the subsection case */
939 if (section_nr_to_pfn(section_nr) != start_pfn)
940 memmap = pfn_to_page(section_nr_to_pfn(section_nr));
941 sparse_init_one_section(ms, section_nr, memmap, ms->usage, 0);
942
943 return 0;
944}
945
946#ifdef CONFIG_MEMORY_FAILURE
947static void clear_hwpoisoned_pages(struct page *memmap, int nr_pages)
948{
949 int i;
950
951 /*
952 * A further optimization is to have per section refcounted
953 * num_poisoned_pages. But that would need more space per memmap, so
954 * for now just do a quick global check to speed up this routine in the
955 * absence of bad pages.
956 */
957 if (atomic_long_read(&num_poisoned_pages) == 0)
958 return;
959
960 for (i = 0; i < nr_pages; i++) {
961 if (PageHWPoison(&memmap[i])) {
962 num_poisoned_pages_dec();
963 ClearPageHWPoison(&memmap[i]);
964 }
965 }
966}
967#else
968static inline void clear_hwpoisoned_pages(struct page *memmap, int nr_pages)
969{
970}
971#endif
972
973void sparse_remove_section(struct mem_section *ms, unsigned long pfn,
974 unsigned long nr_pages, unsigned long map_offset,
975 struct vmem_altmap *altmap)
976{
977 clear_hwpoisoned_pages(pfn_to_page(pfn) + map_offset,
978 nr_pages - map_offset);
979 section_deactivate(pfn, nr_pages, altmap);
980}
981#endif /* CONFIG_MEMORY_HOTPLUG */