Merge branch 'for-4.16/nfit' into libnvdimm-for-next
[sfrench/cifs-2.6.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/bootmem.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
15 #include "internal.h"
16 #include <asm/dma.h>
17 #include <asm/pgalloc.h>
18 #include <asm/pgtable.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_virt_alloc_node(array_size, nid);
72
73         return section;
74 }
75
76 static int __meminit sparse_index_init(unsigned long section_nr, int nid)
77 {
78         unsigned long root = SECTION_NR_TO_ROOT(section_nr);
79         struct mem_section *section;
80
81         if (mem_section[root])
82                 return -EEXIST;
83
84         section = sparse_index_alloc(nid);
85         if (!section)
86                 return -ENOMEM;
87
88         mem_section[root] = section;
89
90         return 0;
91 }
92 #else /* !SPARSEMEM_EXTREME */
93 static inline int sparse_index_init(unsigned long section_nr, int nid)
94 {
95         return 0;
96 }
97 #endif
98
99 #ifdef CONFIG_SPARSEMEM_EXTREME
100 int __section_nr(struct mem_section* ms)
101 {
102         unsigned long root_nr;
103         struct mem_section *root = NULL;
104
105         for (root_nr = 0; root_nr < NR_SECTION_ROOTS; root_nr++) {
106                 root = __nr_to_section(root_nr * SECTIONS_PER_ROOT);
107                 if (!root)
108                         continue;
109
110                 if ((ms >= root) && (ms < (root + SECTIONS_PER_ROOT)))
111                      break;
112         }
113
114         VM_BUG_ON(!root);
115
116         return (root_nr * SECTIONS_PER_ROOT) + (ms - root);
117 }
118 #else
119 int __section_nr(struct mem_section* ms)
120 {
121         return (int)(ms - mem_section[0]);
122 }
123 #endif
124
125 /*
126  * During early boot, before section_mem_map is used for an actual
127  * mem_map, we use section_mem_map to store the section's NUMA
128  * node.  This keeps us from having to use another data structure.  The
129  * node information is cleared just before we store the real mem_map.
130  */
131 static inline unsigned long sparse_encode_early_nid(int nid)
132 {
133         return (nid << SECTION_NID_SHIFT);
134 }
135
136 static inline int sparse_early_nid(struct mem_section *section)
137 {
138         return (section->section_mem_map >> SECTION_NID_SHIFT);
139 }
140
141 /* Validate the physical addressing limitations of the model */
142 void __meminit mminit_validate_memmodel_limits(unsigned long *start_pfn,
143                                                 unsigned long *end_pfn)
144 {
145         unsigned long max_sparsemem_pfn = 1UL << (MAX_PHYSMEM_BITS-PAGE_SHIFT);
146
147         /*
148          * Sanity checks - do not allow an architecture to pass
149          * in larger pfns than the maximum scope of sparsemem:
150          */
151         if (*start_pfn > max_sparsemem_pfn) {
152                 mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
153                         "Start of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
154                         *start_pfn, *end_pfn, max_sparsemem_pfn);
155                 WARN_ON_ONCE(1);
156                 *start_pfn = max_sparsemem_pfn;
157                 *end_pfn = max_sparsemem_pfn;
158         } else if (*end_pfn > max_sparsemem_pfn) {
159                 mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
160                         "End of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
161                         *start_pfn, *end_pfn, max_sparsemem_pfn);
162                 WARN_ON_ONCE(1);
163                 *end_pfn = max_sparsemem_pfn;
164         }
165 }
166
167 /*
168  * There are a number of times that we loop over NR_MEM_SECTIONS,
169  * looking for section_present() on each.  But, when we have very
170  * large physical address spaces, NR_MEM_SECTIONS can also be
171  * very large which makes the loops quite long.
172  *
173  * Keeping track of this gives us an easy way to break out of
174  * those loops early.
175  */
176 int __highest_present_section_nr;
177 static void section_mark_present(struct mem_section *ms)
178 {
179         int section_nr = __section_nr(ms);
180
181         if (section_nr > __highest_present_section_nr)
182                 __highest_present_section_nr = section_nr;
183
184         ms->section_mem_map |= SECTION_MARKED_PRESENT;
185 }
186
187 static inline int next_present_section_nr(int section_nr)
188 {
189         do {
190                 section_nr++;
191                 if (present_section_nr(section_nr))
192                         return section_nr;
193         } while ((section_nr < NR_MEM_SECTIONS) &&
194                  (section_nr <= __highest_present_section_nr));
195
196         return -1;
197 }
198 #define for_each_present_section_nr(start, section_nr)          \
199         for (section_nr = next_present_section_nr(start-1);     \
200              ((section_nr >= 0) &&                              \
201               (section_nr < NR_MEM_SECTIONS) &&                 \
202               (section_nr <= __highest_present_section_nr));    \
203              section_nr = next_present_section_nr(section_nr))
204
205 /* Record a memory area against a node. */
206 void __init memory_present(int nid, unsigned long start, unsigned long end)
207 {
208         unsigned long pfn;
209
210 #ifdef CONFIG_SPARSEMEM_EXTREME
211         if (unlikely(!mem_section)) {
212                 unsigned long size, align;
213
214                 size = sizeof(struct mem_section*) * NR_SECTION_ROOTS;
215                 align = 1 << (INTERNODE_CACHE_SHIFT);
216                 mem_section = memblock_virt_alloc(size, align);
217         }
218 #endif
219
220         start &= PAGE_SECTION_MASK;
221         mminit_validate_memmodel_limits(&start, &end);
222         for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) {
223                 unsigned long section = pfn_to_section_nr(pfn);
224                 struct mem_section *ms;
225
226                 sparse_index_init(section, nid);
227                 set_section_nid(section, nid);
228
229                 ms = __nr_to_section(section);
230                 if (!ms->section_mem_map) {
231                         ms->section_mem_map = sparse_encode_early_nid(nid) |
232                                                         SECTION_IS_ONLINE;
233                         section_mark_present(ms);
234                 }
235         }
236 }
237
238 /*
239  * Only used by the i386 NUMA architecures, but relatively
240  * generic code.
241  */
242 unsigned long __init node_memmap_size_bytes(int nid, unsigned long start_pfn,
243                                                      unsigned long end_pfn)
244 {
245         unsigned long pfn;
246         unsigned long nr_pages = 0;
247
248         mminit_validate_memmodel_limits(&start_pfn, &end_pfn);
249         for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
250                 if (nid != early_pfn_to_nid(pfn))
251                         continue;
252
253                 if (pfn_present(pfn))
254                         nr_pages += PAGES_PER_SECTION;
255         }
256
257         return nr_pages * sizeof(struct page);
258 }
259
260 /*
261  * Subtle, we encode the real pfn into the mem_map such that
262  * the identity pfn - section_mem_map will return the actual
263  * physical page frame number.
264  */
265 static unsigned long sparse_encode_mem_map(struct page *mem_map, unsigned long pnum)
266 {
267         return (unsigned long)(mem_map - (section_nr_to_pfn(pnum)));
268 }
269
270 /*
271  * Decode mem_map from the coded memmap
272  */
273 struct page *sparse_decode_mem_map(unsigned long coded_mem_map, unsigned long pnum)
274 {
275         /* mask off the extra low bits of information */
276         coded_mem_map &= SECTION_MAP_MASK;
277         return ((struct page *)coded_mem_map) + section_nr_to_pfn(pnum);
278 }
279
280 static int __meminit sparse_init_one_section(struct mem_section *ms,
281                 unsigned long pnum, struct page *mem_map,
282                 unsigned long *pageblock_bitmap)
283 {
284         if (!present_section(ms))
285                 return -EINVAL;
286
287         ms->section_mem_map &= ~SECTION_MAP_MASK;
288         ms->section_mem_map |= sparse_encode_mem_map(mem_map, pnum) |
289                                                         SECTION_HAS_MEM_MAP;
290         ms->pageblock_flags = pageblock_bitmap;
291
292         return 1;
293 }
294
295 unsigned long usemap_size(void)
296 {
297         return BITS_TO_LONGS(SECTION_BLOCKFLAGS_BITS) * sizeof(unsigned long);
298 }
299
300 #ifdef CONFIG_MEMORY_HOTPLUG
301 static unsigned long *__kmalloc_section_usemap(void)
302 {
303         return kmalloc(usemap_size(), GFP_KERNEL);
304 }
305 #endif /* CONFIG_MEMORY_HOTPLUG */
306
307 #ifdef CONFIG_MEMORY_HOTREMOVE
308 static unsigned long * __init
309 sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
310                                          unsigned long size)
311 {
312         unsigned long goal, limit;
313         unsigned long *p;
314         int nid;
315         /*
316          * A page may contain usemaps for other sections preventing the
317          * page being freed and making a section unremovable while
318          * other sections referencing the usemap remain active. Similarly,
319          * a pgdat can prevent a section being removed. If section A
320          * contains a pgdat and section B contains the usemap, both
321          * sections become inter-dependent. This allocates usemaps
322          * from the same section as the pgdat where possible to avoid
323          * this problem.
324          */
325         goal = __pa(pgdat) & (PAGE_SECTION_MASK << PAGE_SHIFT);
326         limit = goal + (1UL << PA_SECTION_SHIFT);
327         nid = early_pfn_to_nid(goal >> PAGE_SHIFT);
328 again:
329         p = memblock_virt_alloc_try_nid_nopanic(size,
330                                                 SMP_CACHE_BYTES, goal, limit,
331                                                 nid);
332         if (!p && limit) {
333                 limit = 0;
334                 goto again;
335         }
336         return p;
337 }
338
339 static void __init check_usemap_section_nr(int nid, unsigned long *usemap)
340 {
341         unsigned long usemap_snr, pgdat_snr;
342         static unsigned long old_usemap_snr;
343         static unsigned long old_pgdat_snr;
344         struct pglist_data *pgdat = NODE_DATA(nid);
345         int usemap_nid;
346
347         /* First call */
348         if (!old_usemap_snr) {
349                 old_usemap_snr = NR_MEM_SECTIONS;
350                 old_pgdat_snr = NR_MEM_SECTIONS;
351         }
352
353         usemap_snr = pfn_to_section_nr(__pa(usemap) >> PAGE_SHIFT);
354         pgdat_snr = pfn_to_section_nr(__pa(pgdat) >> PAGE_SHIFT);
355         if (usemap_snr == pgdat_snr)
356                 return;
357
358         if (old_usemap_snr == usemap_snr && old_pgdat_snr == pgdat_snr)
359                 /* skip redundant message */
360                 return;
361
362         old_usemap_snr = usemap_snr;
363         old_pgdat_snr = pgdat_snr;
364
365         usemap_nid = sparse_early_nid(__nr_to_section(usemap_snr));
366         if (usemap_nid != nid) {
367                 pr_info("node %d must be removed before remove section %ld\n",
368                         nid, usemap_snr);
369                 return;
370         }
371         /*
372          * There is a circular dependency.
373          * Some platforms allow un-removable section because they will just
374          * gather other removable sections for dynamic partitioning.
375          * Just notify un-removable section's number here.
376          */
377         pr_info("Section %ld and %ld (node %d) have a circular dependency on usemap and pgdat allocations\n",
378                 usemap_snr, pgdat_snr, nid);
379 }
380 #else
381 static unsigned long * __init
382 sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
383                                          unsigned long size)
384 {
385         return memblock_virt_alloc_node_nopanic(size, pgdat->node_id);
386 }
387
388 static void __init check_usemap_section_nr(int nid, unsigned long *usemap)
389 {
390 }
391 #endif /* CONFIG_MEMORY_HOTREMOVE */
392
393 static void __init sparse_early_usemaps_alloc_node(void *data,
394                                  unsigned long pnum_begin,
395                                  unsigned long pnum_end,
396                                  unsigned long usemap_count, int nodeid)
397 {
398         void *usemap;
399         unsigned long pnum;
400         unsigned long **usemap_map = (unsigned long **)data;
401         int size = usemap_size();
402
403         usemap = sparse_early_usemaps_alloc_pgdat_section(NODE_DATA(nodeid),
404                                                           size * usemap_count);
405         if (!usemap) {
406                 pr_warn("%s: allocation failed\n", __func__);
407                 return;
408         }
409
410         for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
411                 if (!present_section_nr(pnum))
412                         continue;
413                 usemap_map[pnum] = usemap;
414                 usemap += size;
415                 check_usemap_section_nr(nodeid, usemap_map[pnum]);
416         }
417 }
418
419 #ifndef CONFIG_SPARSEMEM_VMEMMAP
420 struct page __init *sparse_mem_map_populate(unsigned long pnum, int nid,
421                 struct vmem_altmap *altmap)
422 {
423         struct page *map;
424         unsigned long size;
425
426         map = alloc_remap(nid, sizeof(struct page) * PAGES_PER_SECTION);
427         if (map)
428                 return map;
429
430         size = PAGE_ALIGN(sizeof(struct page) * PAGES_PER_SECTION);
431         map = memblock_virt_alloc_try_nid(size,
432                                           PAGE_SIZE, __pa(MAX_DMA_ADDRESS),
433                                           BOOTMEM_ALLOC_ACCESSIBLE, nid);
434         return map;
435 }
436 void __init sparse_mem_maps_populate_node(struct page **map_map,
437                                           unsigned long pnum_begin,
438                                           unsigned long pnum_end,
439                                           unsigned long map_count, int nodeid)
440 {
441         void *map;
442         unsigned long pnum;
443         unsigned long size = sizeof(struct page) * PAGES_PER_SECTION;
444
445         map = alloc_remap(nodeid, size * map_count);
446         if (map) {
447                 for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
448                         if (!present_section_nr(pnum))
449                                 continue;
450                         map_map[pnum] = map;
451                         map += size;
452                 }
453                 return;
454         }
455
456         size = PAGE_ALIGN(size);
457         map = memblock_virt_alloc_try_nid_raw(size * map_count,
458                                               PAGE_SIZE, __pa(MAX_DMA_ADDRESS),
459                                               BOOTMEM_ALLOC_ACCESSIBLE, nodeid);
460         if (map) {
461                 for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
462                         if (!present_section_nr(pnum))
463                                 continue;
464                         map_map[pnum] = map;
465                         map += size;
466                 }
467                 return;
468         }
469
470         /* fallback */
471         for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
472                 struct mem_section *ms;
473
474                 if (!present_section_nr(pnum))
475                         continue;
476                 map_map[pnum] = sparse_mem_map_populate(pnum, nodeid, NULL);
477                 if (map_map[pnum])
478                         continue;
479                 ms = __nr_to_section(pnum);
480                 pr_err("%s: sparsemem memory map backing failed some memory will not be available\n",
481                        __func__);
482                 ms->section_mem_map = 0;
483         }
484 }
485 #endif /* !CONFIG_SPARSEMEM_VMEMMAP */
486
487 #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
488 static void __init sparse_early_mem_maps_alloc_node(void *data,
489                                  unsigned long pnum_begin,
490                                  unsigned long pnum_end,
491                                  unsigned long map_count, int nodeid)
492 {
493         struct page **map_map = (struct page **)data;
494         sparse_mem_maps_populate_node(map_map, pnum_begin, pnum_end,
495                                          map_count, nodeid);
496 }
497 #else
498 static struct page __init *sparse_early_mem_map_alloc(unsigned long pnum)
499 {
500         struct page *map;
501         struct mem_section *ms = __nr_to_section(pnum);
502         int nid = sparse_early_nid(ms);
503
504         map = sparse_mem_map_populate(pnum, nid, NULL);
505         if (map)
506                 return map;
507
508         pr_err("%s: sparsemem memory map backing failed some memory will not be available\n",
509                __func__);
510         ms->section_mem_map = 0;
511         return NULL;
512 }
513 #endif
514
515 void __weak __meminit vmemmap_populate_print_last(void)
516 {
517 }
518
519 /**
520  *  alloc_usemap_and_memmap - memory alloction for pageblock flags and vmemmap
521  *  @map: usemap_map for pageblock flags or mmap_map for vmemmap
522  */
523 static void __init alloc_usemap_and_memmap(void (*alloc_func)
524                                         (void *, unsigned long, unsigned long,
525                                         unsigned long, int), void *data)
526 {
527         unsigned long pnum;
528         unsigned long map_count;
529         int nodeid_begin = 0;
530         unsigned long pnum_begin = 0;
531
532         for_each_present_section_nr(0, pnum) {
533                 struct mem_section *ms;
534
535                 ms = __nr_to_section(pnum);
536                 nodeid_begin = sparse_early_nid(ms);
537                 pnum_begin = pnum;
538                 break;
539         }
540         map_count = 1;
541         for_each_present_section_nr(pnum_begin + 1, pnum) {
542                 struct mem_section *ms;
543                 int nodeid;
544
545                 ms = __nr_to_section(pnum);
546                 nodeid = sparse_early_nid(ms);
547                 if (nodeid == nodeid_begin) {
548                         map_count++;
549                         continue;
550                 }
551                 /* ok, we need to take cake of from pnum_begin to pnum - 1*/
552                 alloc_func(data, pnum_begin, pnum,
553                                                 map_count, nodeid_begin);
554                 /* new start, update count etc*/
555                 nodeid_begin = nodeid;
556                 pnum_begin = pnum;
557                 map_count = 1;
558         }
559         /* ok, last chunk */
560         alloc_func(data, pnum_begin, NR_MEM_SECTIONS,
561                                                 map_count, nodeid_begin);
562 }
563
564 /*
565  * Allocate the accumulated non-linear sections, allocate a mem_map
566  * for each and record the physical to section mapping.
567  */
568 void __init sparse_init(void)
569 {
570         unsigned long pnum;
571         struct page *map;
572         unsigned long *usemap;
573         unsigned long **usemap_map;
574         int size;
575 #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
576         int size2;
577         struct page **map_map;
578 #endif
579
580         /* see include/linux/mmzone.h 'struct mem_section' definition */
581         BUILD_BUG_ON(!is_power_of_2(sizeof(struct mem_section)));
582
583         /* Setup pageblock_order for HUGETLB_PAGE_SIZE_VARIABLE */
584         set_pageblock_order();
585
586         /*
587          * map is using big page (aka 2M in x86 64 bit)
588          * usemap is less one page (aka 24 bytes)
589          * so alloc 2M (with 2M align) and 24 bytes in turn will
590          * make next 2M slip to one more 2M later.
591          * then in big system, the memory will have a lot of holes...
592          * here try to allocate 2M pages continuously.
593          *
594          * powerpc need to call sparse_init_one_section right after each
595          * sparse_early_mem_map_alloc, so allocate usemap_map at first.
596          */
597         size = sizeof(unsigned long *) * NR_MEM_SECTIONS;
598         usemap_map = memblock_virt_alloc(size, 0);
599         if (!usemap_map)
600                 panic("can not allocate usemap_map\n");
601         alloc_usemap_and_memmap(sparse_early_usemaps_alloc_node,
602                                                         (void *)usemap_map);
603
604 #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
605         size2 = sizeof(struct page *) * NR_MEM_SECTIONS;
606         map_map = memblock_virt_alloc(size2, 0);
607         if (!map_map)
608                 panic("can not allocate map_map\n");
609         alloc_usemap_and_memmap(sparse_early_mem_maps_alloc_node,
610                                                         (void *)map_map);
611 #endif
612
613         for_each_present_section_nr(0, pnum) {
614                 usemap = usemap_map[pnum];
615                 if (!usemap)
616                         continue;
617
618 #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
619                 map = map_map[pnum];
620 #else
621                 map = sparse_early_mem_map_alloc(pnum);
622 #endif
623                 if (!map)
624                         continue;
625
626                 sparse_init_one_section(__nr_to_section(pnum), pnum, map,
627                                                                 usemap);
628         }
629
630         vmemmap_populate_print_last();
631
632 #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
633         memblock_free_early(__pa(map_map), size2);
634 #endif
635         memblock_free_early(__pa(usemap_map), size);
636 }
637
638 #ifdef CONFIG_MEMORY_HOTPLUG
639
640 /* Mark all memory sections within the pfn range as online */
641 void online_mem_sections(unsigned long start_pfn, unsigned long end_pfn)
642 {
643         unsigned long pfn;
644
645         for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
646                 unsigned long section_nr = pfn_to_section_nr(pfn);
647                 struct mem_section *ms;
648
649                 /* onlining code should never touch invalid ranges */
650                 if (WARN_ON(!valid_section_nr(section_nr)))
651                         continue;
652
653                 ms = __nr_to_section(section_nr);
654                 ms->section_mem_map |= SECTION_IS_ONLINE;
655         }
656 }
657
658 #ifdef CONFIG_MEMORY_HOTREMOVE
659 /* Mark all memory sections within the pfn range as online */
660 void offline_mem_sections(unsigned long start_pfn, unsigned long end_pfn)
661 {
662         unsigned long pfn;
663
664         for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
665                 unsigned long section_nr = pfn_to_section_nr(start_pfn);
666                 struct mem_section *ms;
667
668                 /*
669                  * TODO this needs some double checking. Offlining code makes
670                  * sure to check pfn_valid but those checks might be just bogus
671                  */
672                 if (WARN_ON(!valid_section_nr(section_nr)))
673                         continue;
674
675                 ms = __nr_to_section(section_nr);
676                 ms->section_mem_map &= ~SECTION_IS_ONLINE;
677         }
678 }
679 #endif
680
681 #ifdef CONFIG_SPARSEMEM_VMEMMAP
682 static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid,
683                 struct vmem_altmap *altmap)
684 {
685         /* This will make the necessary allocations eventually. */
686         return sparse_mem_map_populate(pnum, nid, altmap);
687 }
688 static void __kfree_section_memmap(struct page *memmap,
689                 struct vmem_altmap *altmap)
690 {
691         unsigned long start = (unsigned long)memmap;
692         unsigned long end = (unsigned long)(memmap + PAGES_PER_SECTION);
693
694         vmemmap_free(start, end, altmap);
695 }
696 #ifdef CONFIG_MEMORY_HOTREMOVE
697 static void free_map_bootmem(struct page *memmap)
698 {
699         unsigned long start = (unsigned long)memmap;
700         unsigned long end = (unsigned long)(memmap + PAGES_PER_SECTION);
701
702         vmemmap_free(start, end, NULL);
703 }
704 #endif /* CONFIG_MEMORY_HOTREMOVE */
705 #else
706 static struct page *__kmalloc_section_memmap(void)
707 {
708         struct page *page, *ret;
709         unsigned long memmap_size = sizeof(struct page) * PAGES_PER_SECTION;
710
711         page = alloc_pages(GFP_KERNEL|__GFP_NOWARN, get_order(memmap_size));
712         if (page)
713                 goto got_map_page;
714
715         ret = vmalloc(memmap_size);
716         if (ret)
717                 goto got_map_ptr;
718
719         return NULL;
720 got_map_page:
721         ret = (struct page *)pfn_to_kaddr(page_to_pfn(page));
722 got_map_ptr:
723
724         return ret;
725 }
726
727 static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid,
728                 struct vmem_altmap *altmap)
729 {
730         return __kmalloc_section_memmap();
731 }
732
733 static void __kfree_section_memmap(struct page *memmap,
734                 struct vmem_altmap *altmap)
735 {
736         if (is_vmalloc_addr(memmap))
737                 vfree(memmap);
738         else
739                 free_pages((unsigned long)memmap,
740                            get_order(sizeof(struct page) * PAGES_PER_SECTION));
741 }
742
743 #ifdef CONFIG_MEMORY_HOTREMOVE
744 static void free_map_bootmem(struct page *memmap)
745 {
746         unsigned long maps_section_nr, removing_section_nr, i;
747         unsigned long magic, nr_pages;
748         struct page *page = virt_to_page(memmap);
749
750         nr_pages = PAGE_ALIGN(PAGES_PER_SECTION * sizeof(struct page))
751                 >> PAGE_SHIFT;
752
753         for (i = 0; i < nr_pages; i++, page++) {
754                 magic = (unsigned long) page->freelist;
755
756                 BUG_ON(magic == NODE_INFO);
757
758                 maps_section_nr = pfn_to_section_nr(page_to_pfn(page));
759                 removing_section_nr = page_private(page);
760
761                 /*
762                  * When this function is called, the removing section is
763                  * logical offlined state. This means all pages are isolated
764                  * from page allocator. If removing section's memmap is placed
765                  * on the same section, it must not be freed.
766                  * If it is freed, page allocator may allocate it which will
767                  * be removed physically soon.
768                  */
769                 if (maps_section_nr != removing_section_nr)
770                         put_page_bootmem(page);
771         }
772 }
773 #endif /* CONFIG_MEMORY_HOTREMOVE */
774 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
775
776 /*
777  * returns the number of sections whose mem_maps were properly
778  * set.  If this is <=0, then that means that the passed-in
779  * map was not consumed and must be freed.
780  */
781 int __meminit sparse_add_one_section(struct pglist_data *pgdat,
782                 unsigned long start_pfn, struct vmem_altmap *altmap)
783 {
784         unsigned long section_nr = pfn_to_section_nr(start_pfn);
785         struct mem_section *ms;
786         struct page *memmap;
787         unsigned long *usemap;
788         unsigned long flags;
789         int ret;
790
791         /*
792          * no locking for this, because it does its own
793          * plus, it does a kmalloc
794          */
795         ret = sparse_index_init(section_nr, pgdat->node_id);
796         if (ret < 0 && ret != -EEXIST)
797                 return ret;
798         memmap = kmalloc_section_memmap(section_nr, pgdat->node_id, altmap);
799         if (!memmap)
800                 return -ENOMEM;
801         usemap = __kmalloc_section_usemap();
802         if (!usemap) {
803                 __kfree_section_memmap(memmap, altmap);
804                 return -ENOMEM;
805         }
806
807         pgdat_resize_lock(pgdat, &flags);
808
809         ms = __pfn_to_section(start_pfn);
810         if (ms->section_mem_map & SECTION_MARKED_PRESENT) {
811                 ret = -EEXIST;
812                 goto out;
813         }
814
815         memset(memmap, 0, sizeof(struct page) * PAGES_PER_SECTION);
816
817         section_mark_present(ms);
818
819         ret = sparse_init_one_section(ms, section_nr, memmap, usemap);
820
821 out:
822         pgdat_resize_unlock(pgdat, &flags);
823         if (ret <= 0) {
824                 kfree(usemap);
825                 __kfree_section_memmap(memmap, altmap);
826         }
827         return ret;
828 }
829
830 #ifdef CONFIG_MEMORY_HOTREMOVE
831 #ifdef CONFIG_MEMORY_FAILURE
832 static void clear_hwpoisoned_pages(struct page *memmap, int nr_pages)
833 {
834         int i;
835
836         if (!memmap)
837                 return;
838
839         for (i = 0; i < nr_pages; i++) {
840                 if (PageHWPoison(&memmap[i])) {
841                         atomic_long_sub(1, &num_poisoned_pages);
842                         ClearPageHWPoison(&memmap[i]);
843                 }
844         }
845 }
846 #else
847 static inline void clear_hwpoisoned_pages(struct page *memmap, int nr_pages)
848 {
849 }
850 #endif
851
852 static void free_section_usemap(struct page *memmap, unsigned long *usemap,
853                 struct vmem_altmap *altmap)
854 {
855         struct page *usemap_page;
856
857         if (!usemap)
858                 return;
859
860         usemap_page = virt_to_page(usemap);
861         /*
862          * Check to see if allocation came from hot-plug-add
863          */
864         if (PageSlab(usemap_page) || PageCompound(usemap_page)) {
865                 kfree(usemap);
866                 if (memmap)
867                         __kfree_section_memmap(memmap, altmap);
868                 return;
869         }
870
871         /*
872          * The usemap came from bootmem. This is packed with other usemaps
873          * on the section which has pgdat at boot time. Just keep it as is now.
874          */
875
876         if (memmap)
877                 free_map_bootmem(memmap);
878 }
879
880 void sparse_remove_one_section(struct zone *zone, struct mem_section *ms,
881                 unsigned long map_offset, struct vmem_altmap *altmap)
882 {
883         struct page *memmap = NULL;
884         unsigned long *usemap = NULL, flags;
885         struct pglist_data *pgdat = zone->zone_pgdat;
886
887         pgdat_resize_lock(pgdat, &flags);
888         if (ms->section_mem_map) {
889                 usemap = ms->pageblock_flags;
890                 memmap = sparse_decode_mem_map(ms->section_mem_map,
891                                                 __section_nr(ms));
892                 ms->section_mem_map = 0;
893                 ms->pageblock_flags = NULL;
894         }
895         pgdat_resize_unlock(pgdat, &flags);
896
897         clear_hwpoisoned_pages(memmap + map_offset,
898                         PAGES_PER_SECTION - map_offset);
899         free_section_usemap(memmap, usemap, altmap);
900 }
901 #endif /* CONFIG_MEMORY_HOTREMOVE */
902 #endif /* CONFIG_MEMORY_HOTPLUG */