GNU Linux-libre 4.19.211-gnu1
[releases.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 <= __highest_present_section_nr));
194
195         return -1;
196 }
197 #define for_each_present_section_nr(start, section_nr)          \
198         for (section_nr = next_present_section_nr(start-1);     \
199              ((section_nr != -1) &&                             \
200               (section_nr <= __highest_present_section_nr));    \
201              section_nr = next_present_section_nr(section_nr))
202
203 static inline unsigned long first_present_section_nr(void)
204 {
205         return next_present_section_nr(-1);
206 }
207
208 /* Record a memory area against a node. */
209 void __init memory_present(int nid, unsigned long start, unsigned long end)
210 {
211         unsigned long pfn;
212
213 #ifdef CONFIG_SPARSEMEM_EXTREME
214         if (unlikely(!mem_section)) {
215                 unsigned long size, align;
216
217                 size = sizeof(struct mem_section*) * NR_SECTION_ROOTS;
218                 align = 1 << (INTERNODE_CACHE_SHIFT);
219                 mem_section = memblock_virt_alloc(size, align);
220         }
221 #endif
222
223         start &= PAGE_SECTION_MASK;
224         mminit_validate_memmodel_limits(&start, &end);
225         for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) {
226                 unsigned long section = pfn_to_section_nr(pfn);
227                 struct mem_section *ms;
228
229                 sparse_index_init(section, nid);
230                 set_section_nid(section, nid);
231
232                 ms = __nr_to_section(section);
233                 if (!ms->section_mem_map) {
234                         ms->section_mem_map = sparse_encode_early_nid(nid) |
235                                                         SECTION_IS_ONLINE;
236                         section_mark_present(ms);
237                 }
238         }
239 }
240
241 /*
242  * Subtle, we encode the real pfn into the mem_map such that
243  * the identity pfn - section_mem_map will return the actual
244  * physical page frame number.
245  */
246 static unsigned long sparse_encode_mem_map(struct page *mem_map, unsigned long pnum)
247 {
248         unsigned long coded_mem_map =
249                 (unsigned long)(mem_map - (section_nr_to_pfn(pnum)));
250         BUILD_BUG_ON(SECTION_MAP_LAST_BIT > (1UL<<PFN_SECTION_SHIFT));
251         BUG_ON(coded_mem_map & ~SECTION_MAP_MASK);
252         return coded_mem_map;
253 }
254
255 /*
256  * Decode mem_map from the coded memmap
257  */
258 struct page *sparse_decode_mem_map(unsigned long coded_mem_map, unsigned long pnum)
259 {
260         /* mask off the extra low bits of information */
261         coded_mem_map &= SECTION_MAP_MASK;
262         return ((struct page *)coded_mem_map) + section_nr_to_pfn(pnum);
263 }
264
265 static void __meminit sparse_init_one_section(struct mem_section *ms,
266                 unsigned long pnum, struct page *mem_map,
267                 unsigned long *pageblock_bitmap)
268 {
269         ms->section_mem_map &= ~SECTION_MAP_MASK;
270         ms->section_mem_map |= sparse_encode_mem_map(mem_map, pnum) |
271                                                         SECTION_HAS_MEM_MAP;
272         ms->pageblock_flags = pageblock_bitmap;
273 }
274
275 unsigned long usemap_size(void)
276 {
277         return BITS_TO_LONGS(SECTION_BLOCKFLAGS_BITS) * sizeof(unsigned long);
278 }
279
280 #ifdef CONFIG_MEMORY_HOTPLUG
281 static unsigned long *__kmalloc_section_usemap(void)
282 {
283         return kmalloc(usemap_size(), GFP_KERNEL);
284 }
285 #endif /* CONFIG_MEMORY_HOTPLUG */
286
287 #ifdef CONFIG_MEMORY_HOTREMOVE
288 static unsigned long * __init
289 sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
290                                          unsigned long size)
291 {
292         unsigned long goal, limit;
293         unsigned long *p;
294         int nid;
295         /*
296          * A page may contain usemaps for other sections preventing the
297          * page being freed and making a section unremovable while
298          * other sections referencing the usemap remain active. Similarly,
299          * a pgdat can prevent a section being removed. If section A
300          * contains a pgdat and section B contains the usemap, both
301          * sections become inter-dependent. This allocates usemaps
302          * from the same section as the pgdat where possible to avoid
303          * this problem.
304          */
305         goal = __pa(pgdat) & (PAGE_SECTION_MASK << PAGE_SHIFT);
306         limit = goal + (1UL << PA_SECTION_SHIFT);
307         nid = early_pfn_to_nid(goal >> PAGE_SHIFT);
308 again:
309         p = memblock_virt_alloc_try_nid_nopanic(size,
310                                                 SMP_CACHE_BYTES, goal, limit,
311                                                 nid);
312         if (!p && limit) {
313                 limit = 0;
314                 goto again;
315         }
316         return p;
317 }
318
319 static void __init check_usemap_section_nr(int nid, unsigned long *usemap)
320 {
321         unsigned long usemap_snr, pgdat_snr;
322         static unsigned long old_usemap_snr;
323         static unsigned long old_pgdat_snr;
324         struct pglist_data *pgdat = NODE_DATA(nid);
325         int usemap_nid;
326
327         /* First call */
328         if (!old_usemap_snr) {
329                 old_usemap_snr = NR_MEM_SECTIONS;
330                 old_pgdat_snr = NR_MEM_SECTIONS;
331         }
332
333         usemap_snr = pfn_to_section_nr(__pa(usemap) >> PAGE_SHIFT);
334         pgdat_snr = pfn_to_section_nr(__pa(pgdat) >> PAGE_SHIFT);
335         if (usemap_snr == pgdat_snr)
336                 return;
337
338         if (old_usemap_snr == usemap_snr && old_pgdat_snr == pgdat_snr)
339                 /* skip redundant message */
340                 return;
341
342         old_usemap_snr = usemap_snr;
343         old_pgdat_snr = pgdat_snr;
344
345         usemap_nid = sparse_early_nid(__nr_to_section(usemap_snr));
346         if (usemap_nid != nid) {
347                 pr_info("node %d must be removed before remove section %ld\n",
348                         nid, usemap_snr);
349                 return;
350         }
351         /*
352          * There is a circular dependency.
353          * Some platforms allow un-removable section because they will just
354          * gather other removable sections for dynamic partitioning.
355          * Just notify un-removable section's number here.
356          */
357         pr_info("Section %ld and %ld (node %d) have a circular dependency on usemap and pgdat allocations\n",
358                 usemap_snr, pgdat_snr, nid);
359 }
360 #else
361 static unsigned long * __init
362 sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
363                                          unsigned long size)
364 {
365         return memblock_virt_alloc_node_nopanic(size, pgdat->node_id);
366 }
367
368 static void __init check_usemap_section_nr(int nid, unsigned long *usemap)
369 {
370 }
371 #endif /* CONFIG_MEMORY_HOTREMOVE */
372
373 #ifdef CONFIG_SPARSEMEM_VMEMMAP
374 static unsigned long __init section_map_size(void)
375 {
376         return ALIGN(sizeof(struct page) * PAGES_PER_SECTION, PMD_SIZE);
377 }
378
379 #else
380 static unsigned long __init section_map_size(void)
381 {
382         return PAGE_ALIGN(sizeof(struct page) * PAGES_PER_SECTION);
383 }
384
385 struct page __init *sparse_mem_map_populate(unsigned long pnum, int nid,
386                 struct vmem_altmap *altmap)
387 {
388         unsigned long size = section_map_size();
389         struct page *map = sparse_buffer_alloc(size);
390
391         if (map)
392                 return map;
393
394         map = memblock_virt_alloc_try_nid(size,
395                                           PAGE_SIZE, __pa(MAX_DMA_ADDRESS),
396                                           BOOTMEM_ALLOC_ACCESSIBLE, nid);
397         return map;
398 }
399 #endif /* !CONFIG_SPARSEMEM_VMEMMAP */
400
401 static void *sparsemap_buf __meminitdata;
402 static void *sparsemap_buf_end __meminitdata;
403
404 static void __init sparse_buffer_init(unsigned long size, int nid)
405 {
406         WARN_ON(sparsemap_buf); /* forgot to call sparse_buffer_fini()? */
407         sparsemap_buf =
408                 memblock_virt_alloc_try_nid_raw(size, PAGE_SIZE,
409                                                 __pa(MAX_DMA_ADDRESS),
410                                                 BOOTMEM_ALLOC_ACCESSIBLE, nid);
411         sparsemap_buf_end = sparsemap_buf + size;
412 }
413
414 static void __init sparse_buffer_fini(void)
415 {
416         unsigned long size = sparsemap_buf_end - sparsemap_buf;
417
418         if (sparsemap_buf && size > 0)
419                 memblock_free_early(__pa(sparsemap_buf), size);
420         sparsemap_buf = NULL;
421 }
422
423 void * __meminit sparse_buffer_alloc(unsigned long size)
424 {
425         void *ptr = NULL;
426
427         if (sparsemap_buf) {
428                 ptr = PTR_ALIGN(sparsemap_buf, size);
429                 if (ptr + size > sparsemap_buf_end)
430                         ptr = NULL;
431                 else
432                         sparsemap_buf = ptr + size;
433         }
434         return ptr;
435 }
436
437 void __weak __meminit vmemmap_populate_print_last(void)
438 {
439 }
440
441 /*
442  * Initialize sparse on a specific node. The node spans [pnum_begin, pnum_end)
443  * And number of present sections in this node is map_count.
444  */
445 static void __init sparse_init_nid(int nid, unsigned long pnum_begin,
446                                    unsigned long pnum_end,
447                                    unsigned long map_count)
448 {
449         unsigned long pnum, usemap_longs, *usemap;
450         struct page *map;
451
452         usemap_longs = BITS_TO_LONGS(SECTION_BLOCKFLAGS_BITS);
453         usemap = sparse_early_usemaps_alloc_pgdat_section(NODE_DATA(nid),
454                                                           usemap_size() *
455                                                           map_count);
456         if (!usemap) {
457                 pr_err("%s: node[%d] usemap allocation failed", __func__, nid);
458                 goto failed;
459         }
460         sparse_buffer_init(map_count * section_map_size(), nid);
461         for_each_present_section_nr(pnum_begin, pnum) {
462                 if (pnum >= pnum_end)
463                         break;
464
465                 map = sparse_mem_map_populate(pnum, nid, NULL);
466                 if (!map) {
467                         pr_err("%s: node[%d] memory map backing failed. Some memory will not be available.",
468                                __func__, nid);
469                         pnum_begin = pnum;
470                         sparse_buffer_fini();
471                         goto failed;
472                 }
473                 check_usemap_section_nr(nid, usemap);
474                 sparse_init_one_section(__nr_to_section(pnum), pnum, map, usemap);
475                 usemap += usemap_longs;
476         }
477         sparse_buffer_fini();
478         return;
479 failed:
480         /* We failed to allocate, mark all the following pnums as not present */
481         for_each_present_section_nr(pnum_begin, pnum) {
482                 struct mem_section *ms;
483
484                 if (pnum >= pnum_end)
485                         break;
486                 ms = __nr_to_section(pnum);
487                 ms->section_mem_map = 0;
488         }
489 }
490
491 /*
492  * Allocate the accumulated non-linear sections, allocate a mem_map
493  * for each and record the physical to section mapping.
494  */
495 void __init sparse_init(void)
496 {
497         unsigned long pnum_begin = first_present_section_nr();
498         int nid_begin = sparse_early_nid(__nr_to_section(pnum_begin));
499         unsigned long pnum_end, map_count = 1;
500
501         /* Setup pageblock_order for HUGETLB_PAGE_SIZE_VARIABLE */
502         set_pageblock_order();
503
504         for_each_present_section_nr(pnum_begin + 1, pnum_end) {
505                 int nid = sparse_early_nid(__nr_to_section(pnum_end));
506
507                 if (nid == nid_begin) {
508                         map_count++;
509                         continue;
510                 }
511                 /* Init node with sections in range [pnum_begin, pnum_end) */
512                 sparse_init_nid(nid_begin, pnum_begin, pnum_end, map_count);
513                 nid_begin = nid;
514                 pnum_begin = pnum_end;
515                 map_count = 1;
516         }
517         /* cover the last node */
518         sparse_init_nid(nid_begin, pnum_begin, pnum_end, map_count);
519         vmemmap_populate_print_last();
520 }
521
522 #ifdef CONFIG_MEMORY_HOTPLUG
523
524 /* Mark all memory sections within the pfn range as online */
525 void online_mem_sections(unsigned long start_pfn, unsigned long end_pfn)
526 {
527         unsigned long pfn;
528
529         for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
530                 unsigned long section_nr = pfn_to_section_nr(pfn);
531                 struct mem_section *ms;
532
533                 /* onlining code should never touch invalid ranges */
534                 if (WARN_ON(!valid_section_nr(section_nr)))
535                         continue;
536
537                 ms = __nr_to_section(section_nr);
538                 ms->section_mem_map |= SECTION_IS_ONLINE;
539         }
540 }
541
542 #ifdef CONFIG_MEMORY_HOTREMOVE
543 /* Mark all memory sections within the pfn range as online */
544 void offline_mem_sections(unsigned long start_pfn, unsigned long end_pfn)
545 {
546         unsigned long pfn;
547
548         for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
549                 unsigned long section_nr = pfn_to_section_nr(pfn);
550                 struct mem_section *ms;
551
552                 /*
553                  * TODO this needs some double checking. Offlining code makes
554                  * sure to check pfn_valid but those checks might be just bogus
555                  */
556                 if (WARN_ON(!valid_section_nr(section_nr)))
557                         continue;
558
559                 ms = __nr_to_section(section_nr);
560                 ms->section_mem_map &= ~SECTION_IS_ONLINE;
561         }
562 }
563 #endif
564
565 #ifdef CONFIG_SPARSEMEM_VMEMMAP
566 static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid,
567                 struct vmem_altmap *altmap)
568 {
569         /* This will make the necessary allocations eventually. */
570         return sparse_mem_map_populate(pnum, nid, altmap);
571 }
572 static void __kfree_section_memmap(struct page *memmap,
573                 struct vmem_altmap *altmap)
574 {
575         unsigned long start = (unsigned long)memmap;
576         unsigned long end = (unsigned long)(memmap + PAGES_PER_SECTION);
577
578         vmemmap_free(start, end, altmap);
579 }
580 static void free_map_bootmem(struct page *memmap)
581 {
582         unsigned long start = (unsigned long)memmap;
583         unsigned long end = (unsigned long)(memmap + PAGES_PER_SECTION);
584
585         vmemmap_free(start, end, NULL);
586 }
587 #else
588 static struct page *__kmalloc_section_memmap(void)
589 {
590         struct page *page, *ret;
591         unsigned long memmap_size = sizeof(struct page) * PAGES_PER_SECTION;
592
593         page = alloc_pages(GFP_KERNEL|__GFP_NOWARN, get_order(memmap_size));
594         if (page)
595                 goto got_map_page;
596
597         ret = vmalloc(memmap_size);
598         if (ret)
599                 goto got_map_ptr;
600
601         return NULL;
602 got_map_page:
603         ret = (struct page *)pfn_to_kaddr(page_to_pfn(page));
604 got_map_ptr:
605
606         return ret;
607 }
608
609 static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid,
610                 struct vmem_altmap *altmap)
611 {
612         return __kmalloc_section_memmap();
613 }
614
615 static void __kfree_section_memmap(struct page *memmap,
616                 struct vmem_altmap *altmap)
617 {
618         if (is_vmalloc_addr(memmap))
619                 vfree(memmap);
620         else
621                 free_pages((unsigned long)memmap,
622                            get_order(sizeof(struct page) * PAGES_PER_SECTION));
623 }
624
625 static void free_map_bootmem(struct page *memmap)
626 {
627         unsigned long maps_section_nr, removing_section_nr, i;
628         unsigned long magic, nr_pages;
629         struct page *page = virt_to_page(memmap);
630
631         nr_pages = PAGE_ALIGN(PAGES_PER_SECTION * sizeof(struct page))
632                 >> PAGE_SHIFT;
633
634         for (i = 0; i < nr_pages; i++, page++) {
635                 magic = (unsigned long) page->freelist;
636
637                 BUG_ON(magic == NODE_INFO);
638
639                 maps_section_nr = pfn_to_section_nr(page_to_pfn(page));
640                 removing_section_nr = page_private(page);
641
642                 /*
643                  * When this function is called, the removing section is
644                  * logical offlined state. This means all pages are isolated
645                  * from page allocator. If removing section's memmap is placed
646                  * on the same section, it must not be freed.
647                  * If it is freed, page allocator may allocate it which will
648                  * be removed physically soon.
649                  */
650                 if (maps_section_nr != removing_section_nr)
651                         put_page_bootmem(page);
652         }
653 }
654 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
655
656 /*
657  * returns the number of sections whose mem_maps were properly
658  * set.  If this is <=0, then that means that the passed-in
659  * map was not consumed and must be freed.
660  */
661 int __meminit sparse_add_one_section(int nid, unsigned long start_pfn,
662                                      struct vmem_altmap *altmap)
663 {
664         unsigned long section_nr = pfn_to_section_nr(start_pfn);
665         struct mem_section *ms;
666         struct page *memmap;
667         unsigned long *usemap;
668         int ret;
669
670         /*
671          * no locking for this, because it does its own
672          * plus, it does a kmalloc
673          */
674         ret = sparse_index_init(section_nr, nid);
675         if (ret < 0 && ret != -EEXIST)
676                 return ret;
677         ret = 0;
678         memmap = kmalloc_section_memmap(section_nr, nid, altmap);
679         if (!memmap)
680                 return -ENOMEM;
681         usemap = __kmalloc_section_usemap();
682         if (!usemap) {
683                 __kfree_section_memmap(memmap, altmap);
684                 return -ENOMEM;
685         }
686
687         ms = __pfn_to_section(start_pfn);
688         if (ms->section_mem_map & SECTION_MARKED_PRESENT) {
689                 ret = -EEXIST;
690                 goto out;
691         }
692
693 #ifdef CONFIG_DEBUG_VM
694         /*
695          * Poison uninitialized struct pages in order to catch invalid flags
696          * combinations.
697          */
698         memset(memmap, PAGE_POISON_PATTERN, sizeof(struct page) * PAGES_PER_SECTION);
699 #endif
700
701         section_mark_present(ms);
702         sparse_init_one_section(ms, section_nr, memmap, usemap);
703
704 out:
705         if (ret < 0) {
706                 kfree(usemap);
707                 __kfree_section_memmap(memmap, altmap);
708         }
709         return ret;
710 }
711
712 #ifdef CONFIG_MEMORY_FAILURE
713 static void clear_hwpoisoned_pages(struct page *memmap, int nr_pages)
714 {
715         int i;
716
717         if (!memmap)
718                 return;
719
720         for (i = 0; i < nr_pages; i++) {
721                 if (PageHWPoison(&memmap[i])) {
722                         atomic_long_sub(1, &num_poisoned_pages);
723                         ClearPageHWPoison(&memmap[i]);
724                 }
725         }
726 }
727 #else
728 static inline void clear_hwpoisoned_pages(struct page *memmap, int nr_pages)
729 {
730 }
731 #endif
732
733 static void free_section_usemap(struct page *memmap, unsigned long *usemap,
734                 struct vmem_altmap *altmap)
735 {
736         struct page *usemap_page;
737
738         if (!usemap)
739                 return;
740
741         usemap_page = virt_to_page(usemap);
742         /*
743          * Check to see if allocation came from hot-plug-add
744          */
745         if (PageSlab(usemap_page) || PageCompound(usemap_page)) {
746                 kfree(usemap);
747                 if (memmap)
748                         __kfree_section_memmap(memmap, altmap);
749                 return;
750         }
751
752         /*
753          * The usemap came from bootmem. This is packed with other usemaps
754          * on the section which has pgdat at boot time. Just keep it as is now.
755          */
756
757         if (memmap)
758                 free_map_bootmem(memmap);
759 }
760
761 void sparse_remove_one_section(struct mem_section *ms, unsigned long map_offset,
762                                struct vmem_altmap *altmap)
763 {
764         struct page *memmap = NULL;
765         unsigned long *usemap = NULL;
766
767         if (ms->section_mem_map) {
768                 usemap = ms->pageblock_flags;
769                 memmap = sparse_decode_mem_map(ms->section_mem_map,
770                                                 __section_nr(ms));
771                 ms->section_mem_map = 0;
772                 ms->pageblock_flags = NULL;
773         }
774
775         clear_hwpoisoned_pages(memmap + map_offset,
776                         PAGES_PER_SECTION - map_offset);
777         free_section_usemap(memmap, usemap, altmap);
778 }
779 #endif /* CONFIG_MEMORY_HOTPLUG */