1b5e0385f4195ef28501b58d37fdd3cf24564d27
[linux-2.6-microblaze.git] / mm / sparse.c
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
26 struct mem_section **mem_section;
27 #else
28 struct mem_section mem_section[NR_SECTION_ROOTS][SECTIONS_PER_ROOT]
29         ____cacheline_internodealigned_in_smp;
30 #endif
31 EXPORT_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
40 static u8 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned;
41 #else
42 static u16 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned;
43 #endif
44
45 int page_to_nid(const struct page *page)
46 {
47         return section_to_node_table[page_to_section(page)];
48 }
49 EXPORT_SYMBOL(page_to_nid);
50
51 static 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 */
56 static inline void set_section_nid(unsigned long section_nr, int nid)
57 {
58 }
59 #endif
60
61 #ifdef CONFIG_SPARSEMEM_EXTREME
62 static 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
81 static 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 */
105 static inline int sparse_index_init(unsigned long section_nr, int nid)
106 {
107         return 0;
108 }
109 #endif
110
111 #ifdef CONFIG_SPARSEMEM_EXTREME
112 unsigned 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
131 unsigned 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  */
143 static inline unsigned long sparse_encode_early_nid(int nid)
144 {
145         return (nid << SECTION_NID_SHIFT);
146 }
147
148 static 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 */
154 void __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  */
188 unsigned long __highest_present_section_nr;
189 static 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
205 static inline unsigned long first_present_section_nr(void)
206 {
207         return next_present_section_nr(-1);
208 }
209
210 #ifdef CONFIG_SPARSEMEM_VMEMMAP
211 static 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
220 void __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
246 void __init subsection_map_init(unsigned long pfn, unsigned long nr_pages)
247 {
248 }
249 #endif
250
251 /* Record a memory area against a node. */
252 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(). This is a
289  * convenience function that is useful for a number of arches
290  * to mark all of the systems memory as present during initialization.
291  */
292 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  */
308 static 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  */
320 struct 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
327 static 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
337 static unsigned long usemap_size(void)
338 {
339         return BITS_TO_LONGS(SECTION_BLOCKFLAGS_BITS) * sizeof(unsigned long);
340 }
341
342 size_t mem_section_usage_size(void)
343 {
344         return sizeof(struct mem_section_usage) + usemap_size();
345 }
346
347 #ifdef CONFIG_MEMORY_HOTREMOVE
348 static struct mem_section_usage * __init
349 sparse_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);
368 again:
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
377 static 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
420 static struct mem_section_usage * __init
421 sparse_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
427 static 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
434 static unsigned long __init section_map_size(void)
435 {
436         return ALIGN(sizeof(struct page) * PAGES_PER_SECTION, PMD_SIZE);
437 }
438
439 #else
440 static unsigned long __init section_map_size(void)
441 {
442         return PAGE_ALIGN(sizeof(struct page) * PAGES_PER_SECTION);
443 }
444
445 struct 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
465 static void *sparsemap_buf __meminitdata;
466 static void *sparsemap_buf_end __meminitdata;
467
468 static 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
474 static 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
488 static 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
497 void * __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
515 void __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  */
523 static 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;
559 failed:
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  */
575 void __init sparse_init(void)
576 {
577         unsigned long pnum_begin = first_present_section_nr();
578         int nid_begin = sparse_early_nid(__nr_to_section(pnum_begin));
579         unsigned long pnum_end, map_count = 1;
580
581         /* Setup pageblock_order for HUGETLB_PAGE_SIZE_VARIABLE */
582         set_pageblock_order();
583
584         for_each_present_section_nr(pnum_begin + 1, pnum_end) {
585                 int nid = sparse_early_nid(__nr_to_section(pnum_end));
586
587                 if (nid == nid_begin) {
588                         map_count++;
589                         continue;
590                 }
591                 /* Init node with sections in range [pnum_begin, pnum_end) */
592                 sparse_init_nid(nid_begin, pnum_begin, pnum_end, map_count);
593                 nid_begin = nid;
594                 pnum_begin = pnum_end;
595                 map_count = 1;
596         }
597         /* cover the last node */
598         sparse_init_nid(nid_begin, pnum_begin, pnum_end, map_count);
599         vmemmap_populate_print_last();
600 }
601
602 #ifdef CONFIG_MEMORY_HOTPLUG
603
604 /* Mark all memory sections within the pfn range as online */
605 void online_mem_sections(unsigned long start_pfn, unsigned long end_pfn)
606 {
607         unsigned long pfn;
608
609         for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
610                 unsigned long section_nr = pfn_to_section_nr(pfn);
611                 struct mem_section *ms;
612
613                 /* onlining code should never touch invalid ranges */
614                 if (WARN_ON(!valid_section_nr(section_nr)))
615                         continue;
616
617                 ms = __nr_to_section(section_nr);
618                 ms->section_mem_map |= SECTION_IS_ONLINE;
619         }
620 }
621
622 #ifdef CONFIG_MEMORY_HOTREMOVE
623 /* Mark all memory sections within the pfn range as offline */
624 void offline_mem_sections(unsigned long start_pfn, unsigned long end_pfn)
625 {
626         unsigned long pfn;
627
628         for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
629                 unsigned long section_nr = pfn_to_section_nr(pfn);
630                 struct mem_section *ms;
631
632                 /*
633                  * TODO this needs some double checking. Offlining code makes
634                  * sure to check pfn_valid but those checks might be just bogus
635                  */
636                 if (WARN_ON(!valid_section_nr(section_nr)))
637                         continue;
638
639                 ms = __nr_to_section(section_nr);
640                 ms->section_mem_map &= ~SECTION_IS_ONLINE;
641         }
642 }
643 #endif
644
645 #ifdef CONFIG_SPARSEMEM_VMEMMAP
646 static struct page * __meminit populate_section_memmap(unsigned long pfn,
647                 unsigned long nr_pages, int nid, struct vmem_altmap *altmap)
648 {
649         return __populate_section_memmap(pfn, nr_pages, nid, altmap);
650 }
651
652 static void depopulate_section_memmap(unsigned long pfn, unsigned long nr_pages,
653                 struct vmem_altmap *altmap)
654 {
655         unsigned long start = (unsigned long) pfn_to_page(pfn);
656         unsigned long end = start + nr_pages * sizeof(struct page);
657
658         vmemmap_free(start, end, altmap);
659 }
660 static void free_map_bootmem(struct page *memmap)
661 {
662         unsigned long start = (unsigned long)memmap;
663         unsigned long end = (unsigned long)(memmap + PAGES_PER_SECTION);
664
665         vmemmap_free(start, end, NULL);
666 }
667
668 static int clear_subsection_map(unsigned long pfn, unsigned long nr_pages)
669 {
670         DECLARE_BITMAP(map, SUBSECTIONS_PER_SECTION) = { 0 };
671         DECLARE_BITMAP(tmp, SUBSECTIONS_PER_SECTION) = { 0 };
672         struct mem_section *ms = __pfn_to_section(pfn);
673         unsigned long *subsection_map = ms->usage
674                 ? &ms->usage->subsection_map[0] : NULL;
675
676         subsection_mask_set(map, pfn, nr_pages);
677         if (subsection_map)
678                 bitmap_and(tmp, map, subsection_map, SUBSECTIONS_PER_SECTION);
679
680         if (WARN(!subsection_map || !bitmap_equal(tmp, map, SUBSECTIONS_PER_SECTION),
681                                 "section already deactivated (%#lx + %ld)\n",
682                                 pfn, nr_pages))
683                 return -EINVAL;
684
685         bitmap_xor(subsection_map, map, subsection_map, SUBSECTIONS_PER_SECTION);
686         return 0;
687 }
688
689 static bool is_subsection_map_empty(struct mem_section *ms)
690 {
691         return bitmap_empty(&ms->usage->subsection_map[0],
692                             SUBSECTIONS_PER_SECTION);
693 }
694
695 static int fill_subsection_map(unsigned long pfn, unsigned long nr_pages)
696 {
697         struct mem_section *ms = __pfn_to_section(pfn);
698         DECLARE_BITMAP(map, SUBSECTIONS_PER_SECTION) = { 0 };
699         unsigned long *subsection_map;
700         int rc = 0;
701
702         subsection_mask_set(map, pfn, nr_pages);
703
704         subsection_map = &ms->usage->subsection_map[0];
705
706         if (bitmap_empty(map, SUBSECTIONS_PER_SECTION))
707                 rc = -EINVAL;
708         else if (bitmap_intersects(map, subsection_map, SUBSECTIONS_PER_SECTION))
709                 rc = -EEXIST;
710         else
711                 bitmap_or(subsection_map, map, subsection_map,
712                                 SUBSECTIONS_PER_SECTION);
713
714         return rc;
715 }
716 #else
717 struct page * __meminit populate_section_memmap(unsigned long pfn,
718                 unsigned long nr_pages, int nid, struct vmem_altmap *altmap)
719 {
720         return kvmalloc_node(array_size(sizeof(struct page),
721                                         PAGES_PER_SECTION), GFP_KERNEL, nid);
722 }
723
724 static void depopulate_section_memmap(unsigned long pfn, unsigned long nr_pages,
725                 struct vmem_altmap *altmap)
726 {
727         kvfree(pfn_to_page(pfn));
728 }
729
730 static void free_map_bootmem(struct page *memmap)
731 {
732         unsigned long maps_section_nr, removing_section_nr, i;
733         unsigned long magic, nr_pages;
734         struct page *page = virt_to_page(memmap);
735
736         nr_pages = PAGE_ALIGN(PAGES_PER_SECTION * sizeof(struct page))
737                 >> PAGE_SHIFT;
738
739         for (i = 0; i < nr_pages; i++, page++) {
740                 magic = (unsigned long) page->freelist;
741
742                 BUG_ON(magic == NODE_INFO);
743
744                 maps_section_nr = pfn_to_section_nr(page_to_pfn(page));
745                 removing_section_nr = page_private(page);
746
747                 /*
748                  * When this function is called, the removing section is
749                  * logical offlined state. This means all pages are isolated
750                  * from page allocator. If removing section's memmap is placed
751                  * on the same section, it must not be freed.
752                  * If it is freed, page allocator may allocate it which will
753                  * be removed physically soon.
754                  */
755                 if (maps_section_nr != removing_section_nr)
756                         put_page_bootmem(page);
757         }
758 }
759
760 static int clear_subsection_map(unsigned long pfn, unsigned long nr_pages)
761 {
762         return 0;
763 }
764
765 static bool is_subsection_map_empty(struct mem_section *ms)
766 {
767         return true;
768 }
769
770 static int fill_subsection_map(unsigned long pfn, unsigned long nr_pages)
771 {
772         return 0;
773 }
774 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
775
776 /*
777  * To deactivate a memory region, there are 3 cases to handle across
778  * two configurations (SPARSEMEM_VMEMMAP={y,n}):
779  *
780  * 1. deactivation of a partial hot-added section (only possible in
781  *    the SPARSEMEM_VMEMMAP=y case).
782  *      a) section was present at memory init.
783  *      b) section was hot-added post memory init.
784  * 2. deactivation of a complete hot-added section.
785  * 3. deactivation of a complete section from memory init.
786  *
787  * For 1, when subsection_map does not empty we will not be freeing the
788  * usage map, but still need to free the vmemmap range.
789  *
790  * For 2 and 3, the SPARSEMEM_VMEMMAP={y,n} cases are unified
791  */
792 static void section_deactivate(unsigned long pfn, unsigned long nr_pages,
793                 struct vmem_altmap *altmap)
794 {
795         struct mem_section *ms = __pfn_to_section(pfn);
796         bool section_is_early = early_section(ms);
797         struct page *memmap = NULL;
798         bool empty;
799
800         if (clear_subsection_map(pfn, nr_pages))
801                 return;
802
803         empty = is_subsection_map_empty(ms);
804         if (empty) {
805                 unsigned long section_nr = pfn_to_section_nr(pfn);
806
807                 /*
808                  * When removing an early section, the usage map is kept (as the
809                  * usage maps of other sections fall into the same page). It
810                  * will be re-used when re-adding the section - which is then no
811                  * longer an early section. If the usage map is PageReserved, it
812                  * was allocated during boot.
813                  */
814                 if (!PageReserved(virt_to_page(ms->usage))) {
815                         kfree(ms->usage);
816                         ms->usage = NULL;
817                 }
818                 memmap = sparse_decode_mem_map(ms->section_mem_map, section_nr);
819                 /*
820                  * Mark the section invalid so that valid_section()
821                  * return false. This prevents code from dereferencing
822                  * ms->usage array.
823                  */
824                 ms->section_mem_map &= ~SECTION_HAS_MEM_MAP;
825         }
826
827         /*
828          * The memmap of early sections is always fully populated. See
829          * section_activate() and pfn_valid() .
830          */
831         if (!section_is_early)
832                 depopulate_section_memmap(pfn, nr_pages, altmap);
833         else if (memmap)
834                 free_map_bootmem(memmap);
835
836         if (empty)
837                 ms->section_mem_map = (unsigned long)NULL;
838 }
839
840 static struct page * __meminit section_activate(int nid, unsigned long pfn,
841                 unsigned long nr_pages, struct vmem_altmap *altmap)
842 {
843         struct mem_section *ms = __pfn_to_section(pfn);
844         struct mem_section_usage *usage = NULL;
845         struct page *memmap;
846         int rc = 0;
847
848         if (!ms->usage) {
849                 usage = kzalloc(mem_section_usage_size(), GFP_KERNEL);
850                 if (!usage)
851                         return ERR_PTR(-ENOMEM);
852                 ms->usage = usage;
853         }
854
855         rc = fill_subsection_map(pfn, nr_pages);
856         if (rc) {
857                 if (usage)
858                         ms->usage = NULL;
859                 kfree(usage);
860                 return ERR_PTR(rc);
861         }
862
863         /*
864          * The early init code does not consider partially populated
865          * initial sections, it simply assumes that memory will never be
866          * referenced.  If we hot-add memory into such a section then we
867          * do not need to populate the memmap and can simply reuse what
868          * is already there.
869          */
870         if (nr_pages < PAGES_PER_SECTION && early_section(ms))
871                 return pfn_to_page(pfn);
872
873         memmap = populate_section_memmap(pfn, nr_pages, nid, altmap);
874         if (!memmap) {
875                 section_deactivate(pfn, nr_pages, altmap);
876                 return ERR_PTR(-ENOMEM);
877         }
878
879         return memmap;
880 }
881
882 /**
883  * sparse_add_section - add a memory section, or populate an existing one
884  * @nid: The node to add section on
885  * @start_pfn: start pfn of the memory range
886  * @nr_pages: number of pfns to add in the section
887  * @altmap: device page map
888  *
889  * This is only intended for hotplug.
890  *
891  * Note that only VMEMMAP supports sub-section aligned hotplug,
892  * the proper alignment and size are gated by check_pfn_span().
893  *
894  *
895  * Return:
896  * * 0          - On success.
897  * * -EEXIST    - Section has been present.
898  * * -ENOMEM    - Out of memory.
899  */
900 int __meminit sparse_add_section(int nid, unsigned long start_pfn,
901                 unsigned long nr_pages, struct vmem_altmap *altmap)
902 {
903         unsigned long section_nr = pfn_to_section_nr(start_pfn);
904         struct mem_section *ms;
905         struct page *memmap;
906         int ret;
907
908         ret = sparse_index_init(section_nr, nid);
909         if (ret < 0)
910                 return ret;
911
912         memmap = section_activate(nid, start_pfn, nr_pages, altmap);
913         if (IS_ERR(memmap))
914                 return PTR_ERR(memmap);
915
916         /*
917          * Poison uninitialized struct pages in order to catch invalid flags
918          * combinations.
919          */
920         page_init_poison(memmap, sizeof(struct page) * nr_pages);
921
922         ms = __nr_to_section(section_nr);
923         set_section_nid(section_nr, nid);
924         section_mark_present(ms);
925
926         /* Align memmap to section boundary in the subsection case */
927         if (section_nr_to_pfn(section_nr) != start_pfn)
928                 memmap = pfn_to_page(section_nr_to_pfn(section_nr));
929         sparse_init_one_section(ms, section_nr, memmap, ms->usage, 0);
930
931         return 0;
932 }
933
934 #ifdef CONFIG_MEMORY_FAILURE
935 static void clear_hwpoisoned_pages(struct page *memmap, int nr_pages)
936 {
937         int i;
938
939         /*
940          * A further optimization is to have per section refcounted
941          * num_poisoned_pages.  But that would need more space per memmap, so
942          * for now just do a quick global check to speed up this routine in the
943          * absence of bad pages.
944          */
945         if (atomic_long_read(&num_poisoned_pages) == 0)
946                 return;
947
948         for (i = 0; i < nr_pages; i++) {
949                 if (PageHWPoison(&memmap[i])) {
950                         num_poisoned_pages_dec();
951                         ClearPageHWPoison(&memmap[i]);
952                 }
953         }
954 }
955 #else
956 static inline void clear_hwpoisoned_pages(struct page *memmap, int nr_pages)
957 {
958 }
959 #endif
960
961 void sparse_remove_section(struct mem_section *ms, unsigned long pfn,
962                 unsigned long nr_pages, unsigned long map_offset,
963                 struct vmem_altmap *altmap)
964 {
965         clear_hwpoisoned_pages(pfn_to_page(pfn) + map_offset,
966                         nr_pages - map_offset);
967         section_deactivate(pfn, nr_pages, altmap);
968 }
969 #endif /* CONFIG_MEMORY_HOTPLUG */