Loading...
1/*
2 * sparse memory mappings.
3 */
4#include <linux/mm.h>
5#include <linux/slab.h>
6#include <linux/mmzone.h>
7#include <linux/bootmem.h>
8#include <linux/highmem.h>
9#include <linux/module.h>
10#include <linux/spinlock.h>
11#include <linux/vmalloc.h>
12#include "internal.h"
13#include <asm/dma.h>
14#include <asm/pgalloc.h>
15#include <asm/pgtable.h>
16
17/*
18 * Permanent SPARSEMEM data:
19 *
20 * 1) mem_section - memory sections, mem_map's for valid memory
21 */
22#ifdef CONFIG_SPARSEMEM_EXTREME
23struct mem_section *mem_section[NR_SECTION_ROOTS]
24 ____cacheline_internodealigned_in_smp;
25#else
26struct mem_section mem_section[NR_SECTION_ROOTS][SECTIONS_PER_ROOT]
27 ____cacheline_internodealigned_in_smp;
28#endif
29EXPORT_SYMBOL(mem_section);
30
31#ifdef NODE_NOT_IN_PAGE_FLAGS
32/*
33 * If we did not store the node number in the page then we have to
34 * do a lookup in the section_to_node_table in order to find which
35 * node the page belongs to.
36 */
37#if MAX_NUMNODES <= 256
38static u8 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned;
39#else
40static u16 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned;
41#endif
42
43int page_to_nid(const struct page *page)
44{
45 return section_to_node_table[page_to_section(page)];
46}
47EXPORT_SYMBOL(page_to_nid);
48
49static void set_section_nid(unsigned long section_nr, int nid)
50{
51 section_to_node_table[section_nr] = nid;
52}
53#else /* !NODE_NOT_IN_PAGE_FLAGS */
54static inline void set_section_nid(unsigned long section_nr, int nid)
55{
56}
57#endif
58
59#ifdef CONFIG_SPARSEMEM_EXTREME
60static struct mem_section noinline __init_refok *sparse_index_alloc(int nid)
61{
62 struct mem_section *section = NULL;
63 unsigned long array_size = SECTIONS_PER_ROOT *
64 sizeof(struct mem_section);
65
66 if (slab_is_available()) {
67 if (node_state(nid, N_HIGH_MEMORY))
68 section = kmalloc_node(array_size, GFP_KERNEL, nid);
69 else
70 section = kmalloc(array_size, GFP_KERNEL);
71 } else
72 section = alloc_bootmem_node(NODE_DATA(nid), array_size);
73
74 if (section)
75 memset(section, 0, array_size);
76
77 return section;
78}
79
80static int __meminit sparse_index_init(unsigned long section_nr, int nid)
81{
82 static DEFINE_SPINLOCK(index_init_lock);
83 unsigned long root = SECTION_NR_TO_ROOT(section_nr);
84 struct mem_section *section;
85 int ret = 0;
86
87 if (mem_section[root])
88 return -EEXIST;
89
90 section = sparse_index_alloc(nid);
91 if (!section)
92 return -ENOMEM;
93 /*
94 * This lock keeps two different sections from
95 * reallocating for the same index
96 */
97 spin_lock(&index_init_lock);
98
99 if (mem_section[root]) {
100 ret = -EEXIST;
101 goto out;
102 }
103
104 mem_section[root] = section;
105out:
106 spin_unlock(&index_init_lock);
107 return ret;
108}
109#else /* !SPARSEMEM_EXTREME */
110static inline int sparse_index_init(unsigned long section_nr, int nid)
111{
112 return 0;
113}
114#endif
115
116/*
117 * Although written for the SPARSEMEM_EXTREME case, this happens
118 * to also work for the flat array case because
119 * NR_SECTION_ROOTS==NR_MEM_SECTIONS.
120 */
121int __section_nr(struct mem_section* ms)
122{
123 unsigned long root_nr;
124 struct mem_section* root;
125
126 for (root_nr = 0; root_nr < NR_SECTION_ROOTS; root_nr++) {
127 root = __nr_to_section(root_nr * SECTIONS_PER_ROOT);
128 if (!root)
129 continue;
130
131 if ((ms >= root) && (ms < (root + SECTIONS_PER_ROOT)))
132 break;
133 }
134
135 return (root_nr * SECTIONS_PER_ROOT) + (ms - root);
136}
137
138/*
139 * During early boot, before section_mem_map is used for an actual
140 * mem_map, we use section_mem_map to store the section's NUMA
141 * node. This keeps us from having to use another data structure. The
142 * node information is cleared just before we store the real mem_map.
143 */
144static inline unsigned long sparse_encode_early_nid(int nid)
145{
146 return (nid << SECTION_NID_SHIFT);
147}
148
149static inline int sparse_early_nid(struct mem_section *section)
150{
151 return (section->section_mem_map >> SECTION_NID_SHIFT);
152}
153
154/* Validate the physical addressing limitations of the model */
155void __meminit mminit_validate_memmodel_limits(unsigned long *start_pfn,
156 unsigned long *end_pfn)
157{
158 unsigned long max_sparsemem_pfn = 1UL << (MAX_PHYSMEM_BITS-PAGE_SHIFT);
159
160 /*
161 * Sanity checks - do not allow an architecture to pass
162 * in larger pfns than the maximum scope of sparsemem:
163 */
164 if (*start_pfn > max_sparsemem_pfn) {
165 mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
166 "Start of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
167 *start_pfn, *end_pfn, max_sparsemem_pfn);
168 WARN_ON_ONCE(1);
169 *start_pfn = max_sparsemem_pfn;
170 *end_pfn = max_sparsemem_pfn;
171 } else if (*end_pfn > max_sparsemem_pfn) {
172 mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
173 "End of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
174 *start_pfn, *end_pfn, max_sparsemem_pfn);
175 WARN_ON_ONCE(1);
176 *end_pfn = max_sparsemem_pfn;
177 }
178}
179
180/* Record a memory area against a node. */
181void __init memory_present(int nid, unsigned long start, unsigned long end)
182{
183 unsigned long pfn;
184
185 start &= PAGE_SECTION_MASK;
186 mminit_validate_memmodel_limits(&start, &end);
187 for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) {
188 unsigned long section = pfn_to_section_nr(pfn);
189 struct mem_section *ms;
190
191 sparse_index_init(section, nid);
192 set_section_nid(section, nid);
193
194 ms = __nr_to_section(section);
195 if (!ms->section_mem_map)
196 ms->section_mem_map = sparse_encode_early_nid(nid) |
197 SECTION_MARKED_PRESENT;
198 }
199}
200
201/*
202 * Only used by the i386 NUMA architecures, but relatively
203 * generic code.
204 */
205unsigned long __init node_memmap_size_bytes(int nid, unsigned long start_pfn,
206 unsigned long end_pfn)
207{
208 unsigned long pfn;
209 unsigned long nr_pages = 0;
210
211 mminit_validate_memmodel_limits(&start_pfn, &end_pfn);
212 for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
213 if (nid != early_pfn_to_nid(pfn))
214 continue;
215
216 if (pfn_present(pfn))
217 nr_pages += PAGES_PER_SECTION;
218 }
219
220 return nr_pages * sizeof(struct page);
221}
222
223/*
224 * Subtle, we encode the real pfn into the mem_map such that
225 * the identity pfn - section_mem_map will return the actual
226 * physical page frame number.
227 */
228static unsigned long sparse_encode_mem_map(struct page *mem_map, unsigned long pnum)
229{
230 return (unsigned long)(mem_map - (section_nr_to_pfn(pnum)));
231}
232
233/*
234 * Decode mem_map from the coded memmap
235 */
236struct page *sparse_decode_mem_map(unsigned long coded_mem_map, unsigned long pnum)
237{
238 /* mask off the extra low bits of information */
239 coded_mem_map &= SECTION_MAP_MASK;
240 return ((struct page *)coded_mem_map) + section_nr_to_pfn(pnum);
241}
242
243static int __meminit sparse_init_one_section(struct mem_section *ms,
244 unsigned long pnum, struct page *mem_map,
245 unsigned long *pageblock_bitmap)
246{
247 if (!present_section(ms))
248 return -EINVAL;
249
250 ms->section_mem_map &= ~SECTION_MAP_MASK;
251 ms->section_mem_map |= sparse_encode_mem_map(mem_map, pnum) |
252 SECTION_HAS_MEM_MAP;
253 ms->pageblock_flags = pageblock_bitmap;
254
255 return 1;
256}
257
258unsigned long usemap_size(void)
259{
260 unsigned long size_bytes;
261 size_bytes = roundup(SECTION_BLOCKFLAGS_BITS, 8) / 8;
262 size_bytes = roundup(size_bytes, sizeof(unsigned long));
263 return size_bytes;
264}
265
266#ifdef CONFIG_MEMORY_HOTPLUG
267static unsigned long *__kmalloc_section_usemap(void)
268{
269 return kmalloc(usemap_size(), GFP_KERNEL);
270}
271#endif /* CONFIG_MEMORY_HOTPLUG */
272
273#ifdef CONFIG_MEMORY_HOTREMOVE
274static unsigned long * __init
275sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
276 unsigned long count)
277{
278 unsigned long section_nr;
279
280 /*
281 * A page may contain usemaps for other sections preventing the
282 * page being freed and making a section unremovable while
283 * other sections referencing the usemap retmain active. Similarly,
284 * a pgdat can prevent a section being removed. If section A
285 * contains a pgdat and section B contains the usemap, both
286 * sections become inter-dependent. This allocates usemaps
287 * from the same section as the pgdat where possible to avoid
288 * this problem.
289 */
290 section_nr = pfn_to_section_nr(__pa(pgdat) >> PAGE_SHIFT);
291 return alloc_bootmem_section(usemap_size() * count, section_nr);
292}
293
294static void __init check_usemap_section_nr(int nid, unsigned long *usemap)
295{
296 unsigned long usemap_snr, pgdat_snr;
297 static unsigned long old_usemap_snr = NR_MEM_SECTIONS;
298 static unsigned long old_pgdat_snr = NR_MEM_SECTIONS;
299 struct pglist_data *pgdat = NODE_DATA(nid);
300 int usemap_nid;
301
302 usemap_snr = pfn_to_section_nr(__pa(usemap) >> PAGE_SHIFT);
303 pgdat_snr = pfn_to_section_nr(__pa(pgdat) >> PAGE_SHIFT);
304 if (usemap_snr == pgdat_snr)
305 return;
306
307 if (old_usemap_snr == usemap_snr && old_pgdat_snr == pgdat_snr)
308 /* skip redundant message */
309 return;
310
311 old_usemap_snr = usemap_snr;
312 old_pgdat_snr = pgdat_snr;
313
314 usemap_nid = sparse_early_nid(__nr_to_section(usemap_snr));
315 if (usemap_nid != nid) {
316 printk(KERN_INFO
317 "node %d must be removed before remove section %ld\n",
318 nid, usemap_snr);
319 return;
320 }
321 /*
322 * There is a circular dependency.
323 * Some platforms allow un-removable section because they will just
324 * gather other removable sections for dynamic partitioning.
325 * Just notify un-removable section's number here.
326 */
327 printk(KERN_INFO "Section %ld and %ld (node %d)", usemap_snr,
328 pgdat_snr, nid);
329 printk(KERN_CONT
330 " have a circular dependency on usemap and pgdat allocations\n");
331}
332#else
333static unsigned long * __init
334sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
335 unsigned long count)
336{
337 return NULL;
338}
339
340static void __init check_usemap_section_nr(int nid, unsigned long *usemap)
341{
342}
343#endif /* CONFIG_MEMORY_HOTREMOVE */
344
345static void __init sparse_early_usemaps_alloc_node(unsigned long**usemap_map,
346 unsigned long pnum_begin,
347 unsigned long pnum_end,
348 unsigned long usemap_count, int nodeid)
349{
350 void *usemap;
351 unsigned long pnum;
352 int size = usemap_size();
353
354 usemap = sparse_early_usemaps_alloc_pgdat_section(NODE_DATA(nodeid),
355 usemap_count);
356 if (usemap) {
357 for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
358 if (!present_section_nr(pnum))
359 continue;
360 usemap_map[pnum] = usemap;
361 usemap += size;
362 }
363 return;
364 }
365
366 usemap = alloc_bootmem_node(NODE_DATA(nodeid), size * usemap_count);
367 if (usemap) {
368 for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
369 if (!present_section_nr(pnum))
370 continue;
371 usemap_map[pnum] = usemap;
372 usemap += size;
373 check_usemap_section_nr(nodeid, usemap_map[pnum]);
374 }
375 return;
376 }
377
378 printk(KERN_WARNING "%s: allocation failed\n", __func__);
379}
380
381#ifndef CONFIG_SPARSEMEM_VMEMMAP
382struct page __init *sparse_mem_map_populate(unsigned long pnum, int nid)
383{
384 struct page *map;
385 unsigned long size;
386
387 map = alloc_remap(nid, sizeof(struct page) * PAGES_PER_SECTION);
388 if (map)
389 return map;
390
391 size = PAGE_ALIGN(sizeof(struct page) * PAGES_PER_SECTION);
392 map = __alloc_bootmem_node_high(NODE_DATA(nid), size,
393 PAGE_SIZE, __pa(MAX_DMA_ADDRESS));
394 return map;
395}
396void __init sparse_mem_maps_populate_node(struct page **map_map,
397 unsigned long pnum_begin,
398 unsigned long pnum_end,
399 unsigned long map_count, int nodeid)
400{
401 void *map;
402 unsigned long pnum;
403 unsigned long size = sizeof(struct page) * PAGES_PER_SECTION;
404
405 map = alloc_remap(nodeid, size * map_count);
406 if (map) {
407 for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
408 if (!present_section_nr(pnum))
409 continue;
410 map_map[pnum] = map;
411 map += size;
412 }
413 return;
414 }
415
416 size = PAGE_ALIGN(size);
417 map = __alloc_bootmem_node_high(NODE_DATA(nodeid), size * map_count,
418 PAGE_SIZE, __pa(MAX_DMA_ADDRESS));
419 if (map) {
420 for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
421 if (!present_section_nr(pnum))
422 continue;
423 map_map[pnum] = map;
424 map += size;
425 }
426 return;
427 }
428
429 /* fallback */
430 for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
431 struct mem_section *ms;
432
433 if (!present_section_nr(pnum))
434 continue;
435 map_map[pnum] = sparse_mem_map_populate(pnum, nodeid);
436 if (map_map[pnum])
437 continue;
438 ms = __nr_to_section(pnum);
439 printk(KERN_ERR "%s: sparsemem memory map backing failed "
440 "some memory will not be available.\n", __func__);
441 ms->section_mem_map = 0;
442 }
443}
444#endif /* !CONFIG_SPARSEMEM_VMEMMAP */
445
446#ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
447static void __init sparse_early_mem_maps_alloc_node(struct page **map_map,
448 unsigned long pnum_begin,
449 unsigned long pnum_end,
450 unsigned long map_count, int nodeid)
451{
452 sparse_mem_maps_populate_node(map_map, pnum_begin, pnum_end,
453 map_count, nodeid);
454}
455#else
456static struct page __init *sparse_early_mem_map_alloc(unsigned long pnum)
457{
458 struct page *map;
459 struct mem_section *ms = __nr_to_section(pnum);
460 int nid = sparse_early_nid(ms);
461
462 map = sparse_mem_map_populate(pnum, nid);
463 if (map)
464 return map;
465
466 printk(KERN_ERR "%s: sparsemem memory map backing failed "
467 "some memory will not be available.\n", __func__);
468 ms->section_mem_map = 0;
469 return NULL;
470}
471#endif
472
473void __attribute__((weak)) __meminit vmemmap_populate_print_last(void)
474{
475}
476
477/*
478 * Allocate the accumulated non-linear sections, allocate a mem_map
479 * for each and record the physical to section mapping.
480 */
481void __init sparse_init(void)
482{
483 unsigned long pnum;
484 struct page *map;
485 unsigned long *usemap;
486 unsigned long **usemap_map;
487 int size;
488 int nodeid_begin = 0;
489 unsigned long pnum_begin = 0;
490 unsigned long usemap_count;
491#ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
492 unsigned long map_count;
493 int size2;
494 struct page **map_map;
495#endif
496
497 /*
498 * map is using big page (aka 2M in x86 64 bit)
499 * usemap is less one page (aka 24 bytes)
500 * so alloc 2M (with 2M align) and 24 bytes in turn will
501 * make next 2M slip to one more 2M later.
502 * then in big system, the memory will have a lot of holes...
503 * here try to allocate 2M pages continuously.
504 *
505 * powerpc need to call sparse_init_one_section right after each
506 * sparse_early_mem_map_alloc, so allocate usemap_map at first.
507 */
508 size = sizeof(unsigned long *) * NR_MEM_SECTIONS;
509 usemap_map = alloc_bootmem(size);
510 if (!usemap_map)
511 panic("can not allocate usemap_map\n");
512
513 for (pnum = 0; pnum < NR_MEM_SECTIONS; pnum++) {
514 struct mem_section *ms;
515
516 if (!present_section_nr(pnum))
517 continue;
518 ms = __nr_to_section(pnum);
519 nodeid_begin = sparse_early_nid(ms);
520 pnum_begin = pnum;
521 break;
522 }
523 usemap_count = 1;
524 for (pnum = pnum_begin + 1; pnum < NR_MEM_SECTIONS; pnum++) {
525 struct mem_section *ms;
526 int nodeid;
527
528 if (!present_section_nr(pnum))
529 continue;
530 ms = __nr_to_section(pnum);
531 nodeid = sparse_early_nid(ms);
532 if (nodeid == nodeid_begin) {
533 usemap_count++;
534 continue;
535 }
536 /* ok, we need to take cake of from pnum_begin to pnum - 1*/
537 sparse_early_usemaps_alloc_node(usemap_map, pnum_begin, pnum,
538 usemap_count, nodeid_begin);
539 /* new start, update count etc*/
540 nodeid_begin = nodeid;
541 pnum_begin = pnum;
542 usemap_count = 1;
543 }
544 /* ok, last chunk */
545 sparse_early_usemaps_alloc_node(usemap_map, pnum_begin, NR_MEM_SECTIONS,
546 usemap_count, nodeid_begin);
547
548#ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
549 size2 = sizeof(struct page *) * NR_MEM_SECTIONS;
550 map_map = alloc_bootmem(size2);
551 if (!map_map)
552 panic("can not allocate map_map\n");
553
554 for (pnum = 0; pnum < NR_MEM_SECTIONS; pnum++) {
555 struct mem_section *ms;
556
557 if (!present_section_nr(pnum))
558 continue;
559 ms = __nr_to_section(pnum);
560 nodeid_begin = sparse_early_nid(ms);
561 pnum_begin = pnum;
562 break;
563 }
564 map_count = 1;
565 for (pnum = pnum_begin + 1; pnum < NR_MEM_SECTIONS; pnum++) {
566 struct mem_section *ms;
567 int nodeid;
568
569 if (!present_section_nr(pnum))
570 continue;
571 ms = __nr_to_section(pnum);
572 nodeid = sparse_early_nid(ms);
573 if (nodeid == nodeid_begin) {
574 map_count++;
575 continue;
576 }
577 /* ok, we need to take cake of from pnum_begin to pnum - 1*/
578 sparse_early_mem_maps_alloc_node(map_map, pnum_begin, pnum,
579 map_count, nodeid_begin);
580 /* new start, update count etc*/
581 nodeid_begin = nodeid;
582 pnum_begin = pnum;
583 map_count = 1;
584 }
585 /* ok, last chunk */
586 sparse_early_mem_maps_alloc_node(map_map, pnum_begin, NR_MEM_SECTIONS,
587 map_count, nodeid_begin);
588#endif
589
590 for (pnum = 0; pnum < NR_MEM_SECTIONS; pnum++) {
591 if (!present_section_nr(pnum))
592 continue;
593
594 usemap = usemap_map[pnum];
595 if (!usemap)
596 continue;
597
598#ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
599 map = map_map[pnum];
600#else
601 map = sparse_early_mem_map_alloc(pnum);
602#endif
603 if (!map)
604 continue;
605
606 sparse_init_one_section(__nr_to_section(pnum), pnum, map,
607 usemap);
608 }
609
610 vmemmap_populate_print_last();
611
612#ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
613 free_bootmem(__pa(map_map), size2);
614#endif
615 free_bootmem(__pa(usemap_map), size);
616}
617
618#ifdef CONFIG_MEMORY_HOTPLUG
619#ifdef CONFIG_SPARSEMEM_VMEMMAP
620static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid,
621 unsigned long nr_pages)
622{
623 /* This will make the necessary allocations eventually. */
624 return sparse_mem_map_populate(pnum, nid);
625}
626static void __kfree_section_memmap(struct page *memmap, unsigned long nr_pages)
627{
628 return; /* XXX: Not implemented yet */
629}
630static void free_map_bootmem(struct page *page, unsigned long nr_pages)
631{
632}
633#else
634static struct page *__kmalloc_section_memmap(unsigned long nr_pages)
635{
636 struct page *page, *ret;
637 unsigned long memmap_size = sizeof(struct page) * nr_pages;
638
639 page = alloc_pages(GFP_KERNEL|__GFP_NOWARN, get_order(memmap_size));
640 if (page)
641 goto got_map_page;
642
643 ret = vmalloc(memmap_size);
644 if (ret)
645 goto got_map_ptr;
646
647 return NULL;
648got_map_page:
649 ret = (struct page *)pfn_to_kaddr(page_to_pfn(page));
650got_map_ptr:
651 memset(ret, 0, memmap_size);
652
653 return ret;
654}
655
656static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid,
657 unsigned long nr_pages)
658{
659 return __kmalloc_section_memmap(nr_pages);
660}
661
662static void __kfree_section_memmap(struct page *memmap, unsigned long nr_pages)
663{
664 if (is_vmalloc_addr(memmap))
665 vfree(memmap);
666 else
667 free_pages((unsigned long)memmap,
668 get_order(sizeof(struct page) * nr_pages));
669}
670
671static void free_map_bootmem(struct page *page, unsigned long nr_pages)
672{
673 unsigned long maps_section_nr, removing_section_nr, i;
674 unsigned long magic;
675
676 for (i = 0; i < nr_pages; i++, page++) {
677 magic = (unsigned long) page->lru.next;
678
679 BUG_ON(magic == NODE_INFO);
680
681 maps_section_nr = pfn_to_section_nr(page_to_pfn(page));
682 removing_section_nr = page->private;
683
684 /*
685 * When this function is called, the removing section is
686 * logical offlined state. This means all pages are isolated
687 * from page allocator. If removing section's memmap is placed
688 * on the same section, it must not be freed.
689 * If it is freed, page allocator may allocate it which will
690 * be removed physically soon.
691 */
692 if (maps_section_nr != removing_section_nr)
693 put_page_bootmem(page);
694 }
695}
696#endif /* CONFIG_SPARSEMEM_VMEMMAP */
697
698static void free_section_usemap(struct page *memmap, unsigned long *usemap)
699{
700 struct page *usemap_page;
701 unsigned long nr_pages;
702
703 if (!usemap)
704 return;
705
706 usemap_page = virt_to_page(usemap);
707 /*
708 * Check to see if allocation came from hot-plug-add
709 */
710 if (PageSlab(usemap_page)) {
711 kfree(usemap);
712 if (memmap)
713 __kfree_section_memmap(memmap, PAGES_PER_SECTION);
714 return;
715 }
716
717 /*
718 * The usemap came from bootmem. This is packed with other usemaps
719 * on the section which has pgdat at boot time. Just keep it as is now.
720 */
721
722 if (memmap) {
723 struct page *memmap_page;
724 memmap_page = virt_to_page(memmap);
725
726 nr_pages = PAGE_ALIGN(PAGES_PER_SECTION * sizeof(struct page))
727 >> PAGE_SHIFT;
728
729 free_map_bootmem(memmap_page, nr_pages);
730 }
731}
732
733/*
734 * returns the number of sections whose mem_maps were properly
735 * set. If this is <=0, then that means that the passed-in
736 * map was not consumed and must be freed.
737 */
738int __meminit sparse_add_one_section(struct zone *zone, unsigned long start_pfn,
739 int nr_pages)
740{
741 unsigned long section_nr = pfn_to_section_nr(start_pfn);
742 struct pglist_data *pgdat = zone->zone_pgdat;
743 struct mem_section *ms;
744 struct page *memmap;
745 unsigned long *usemap;
746 unsigned long flags;
747 int ret;
748
749 /*
750 * no locking for this, because it does its own
751 * plus, it does a kmalloc
752 */
753 ret = sparse_index_init(section_nr, pgdat->node_id);
754 if (ret < 0 && ret != -EEXIST)
755 return ret;
756 memmap = kmalloc_section_memmap(section_nr, pgdat->node_id, nr_pages);
757 if (!memmap)
758 return -ENOMEM;
759 usemap = __kmalloc_section_usemap();
760 if (!usemap) {
761 __kfree_section_memmap(memmap, nr_pages);
762 return -ENOMEM;
763 }
764
765 pgdat_resize_lock(pgdat, &flags);
766
767 ms = __pfn_to_section(start_pfn);
768 if (ms->section_mem_map & SECTION_MARKED_PRESENT) {
769 ret = -EEXIST;
770 goto out;
771 }
772
773 ms->section_mem_map |= SECTION_MARKED_PRESENT;
774
775 ret = sparse_init_one_section(ms, section_nr, memmap, usemap);
776
777out:
778 pgdat_resize_unlock(pgdat, &flags);
779 if (ret <= 0) {
780 kfree(usemap);
781 __kfree_section_memmap(memmap, nr_pages);
782 }
783 return ret;
784}
785
786void sparse_remove_one_section(struct zone *zone, struct mem_section *ms)
787{
788 struct page *memmap = NULL;
789 unsigned long *usemap = NULL;
790
791 if (ms->section_mem_map) {
792 usemap = ms->pageblock_flags;
793 memmap = sparse_decode_mem_map(ms->section_mem_map,
794 __section_nr(ms));
795 ms->section_mem_map = 0;
796 ms->pageblock_flags = NULL;
797 }
798
799 free_section_usemap(memmap, usemap);
800}
801#endif
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#include <linux/bootmem_info.h>
17#include <linux/vmstat.h>
18#include "internal.h"
19#include <asm/dma.h>
20
21/*
22 * Permanent SPARSEMEM data:
23 *
24 * 1) mem_section - memory sections, mem_map's for valid memory
25 */
26#ifdef CONFIG_SPARSEMEM_EXTREME
27struct mem_section **mem_section;
28#else
29struct mem_section mem_section[NR_SECTION_ROOTS][SECTIONS_PER_ROOT]
30 ____cacheline_internodealigned_in_smp;
31#endif
32EXPORT_SYMBOL(mem_section);
33
34#ifdef NODE_NOT_IN_PAGE_FLAGS
35/*
36 * If we did not store the node number in the page then we have to
37 * do a lookup in the section_to_node_table in order to find which
38 * node the page belongs to.
39 */
40#if MAX_NUMNODES <= 256
41static u8 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned;
42#else
43static u16 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned;
44#endif
45
46int page_to_nid(const struct page *page)
47{
48 return section_to_node_table[page_to_section(page)];
49}
50EXPORT_SYMBOL(page_to_nid);
51
52static void set_section_nid(unsigned long section_nr, int nid)
53{
54 section_to_node_table[section_nr] = nid;
55}
56#else /* !NODE_NOT_IN_PAGE_FLAGS */
57static inline void set_section_nid(unsigned long section_nr, int nid)
58{
59}
60#endif
61
62#ifdef CONFIG_SPARSEMEM_EXTREME
63static noinline struct mem_section __ref *sparse_index_alloc(int nid)
64{
65 struct mem_section *section = NULL;
66 unsigned long array_size = SECTIONS_PER_ROOT *
67 sizeof(struct mem_section);
68
69 if (slab_is_available()) {
70 section = kzalloc_node(array_size, GFP_KERNEL, nid);
71 } else {
72 section = memblock_alloc_node(array_size, SMP_CACHE_BYTES,
73 nid);
74 if (!section)
75 panic("%s: Failed to allocate %lu bytes nid=%d\n",
76 __func__, array_size, nid);
77 }
78
79 return section;
80}
81
82static int __meminit sparse_index_init(unsigned long section_nr, int nid)
83{
84 unsigned long root = SECTION_NR_TO_ROOT(section_nr);
85 struct mem_section *section;
86
87 /*
88 * An existing section is possible in the sub-section hotplug
89 * case. First hot-add instantiates, follow-on hot-add reuses
90 * the existing section.
91 *
92 * The mem_hotplug_lock resolves the apparent race below.
93 */
94 if (mem_section[root])
95 return 0;
96
97 section = sparse_index_alloc(nid);
98 if (!section)
99 return -ENOMEM;
100
101 mem_section[root] = section;
102
103 return 0;
104}
105#else /* !SPARSEMEM_EXTREME */
106static inline int sparse_index_init(unsigned long section_nr, int nid)
107{
108 return 0;
109}
110#endif
111
112/*
113 * During early boot, before section_mem_map is used for an actual
114 * mem_map, we use section_mem_map to store the section's NUMA
115 * node. This keeps us from having to use another data structure. The
116 * node information is cleared just before we store the real mem_map.
117 */
118static inline unsigned long sparse_encode_early_nid(int nid)
119{
120 return ((unsigned long)nid << SECTION_NID_SHIFT);
121}
122
123static inline int sparse_early_nid(struct mem_section *section)
124{
125 return (section->section_mem_map >> SECTION_NID_SHIFT);
126}
127
128/* Validate the physical addressing limitations of the model */
129static void __meminit mminit_validate_memmodel_limits(unsigned long *start_pfn,
130 unsigned long *end_pfn)
131{
132 unsigned long max_sparsemem_pfn = (DIRECT_MAP_PHYSMEM_END + 1) >> PAGE_SHIFT;
133
134 /*
135 * Sanity checks - do not allow an architecture to pass
136 * in larger pfns than the maximum scope of sparsemem:
137 */
138 if (*start_pfn > max_sparsemem_pfn) {
139 mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
140 "Start of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
141 *start_pfn, *end_pfn, max_sparsemem_pfn);
142 WARN_ON_ONCE(1);
143 *start_pfn = max_sparsemem_pfn;
144 *end_pfn = max_sparsemem_pfn;
145 } else if (*end_pfn > max_sparsemem_pfn) {
146 mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
147 "End of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
148 *start_pfn, *end_pfn, max_sparsemem_pfn);
149 WARN_ON_ONCE(1);
150 *end_pfn = max_sparsemem_pfn;
151 }
152}
153
154/*
155 * There are a number of times that we loop over NR_MEM_SECTIONS,
156 * looking for section_present() on each. But, when we have very
157 * large physical address spaces, NR_MEM_SECTIONS can also be
158 * very large which makes the loops quite long.
159 *
160 * Keeping track of this gives us an easy way to break out of
161 * those loops early.
162 */
163unsigned long __highest_present_section_nr;
164static void __section_mark_present(struct mem_section *ms,
165 unsigned long section_nr)
166{
167 if (section_nr > __highest_present_section_nr)
168 __highest_present_section_nr = section_nr;
169
170 ms->section_mem_map |= SECTION_MARKED_PRESENT;
171}
172
173#define for_each_present_section_nr(start, section_nr) \
174 for (section_nr = next_present_section_nr(start-1); \
175 section_nr != -1; \
176 section_nr = next_present_section_nr(section_nr))
177
178static inline unsigned long first_present_section_nr(void)
179{
180 return next_present_section_nr(-1);
181}
182
183#ifdef CONFIG_SPARSEMEM_VMEMMAP
184static void subsection_mask_set(unsigned long *map, unsigned long pfn,
185 unsigned long nr_pages)
186{
187 int idx = subsection_map_index(pfn);
188 int end = subsection_map_index(pfn + nr_pages - 1);
189
190 bitmap_set(map, idx, end - idx + 1);
191}
192
193void __init subsection_map_init(unsigned long pfn, unsigned long nr_pages)
194{
195 int end_sec_nr = pfn_to_section_nr(pfn + nr_pages - 1);
196 unsigned long nr, start_sec_nr = pfn_to_section_nr(pfn);
197
198 for (nr = start_sec_nr; nr <= end_sec_nr; nr++) {
199 struct mem_section *ms;
200 unsigned long pfns;
201
202 pfns = min(nr_pages, PAGES_PER_SECTION
203 - (pfn & ~PAGE_SECTION_MASK));
204 ms = __nr_to_section(nr);
205 subsection_mask_set(ms->usage->subsection_map, pfn, pfns);
206
207 pr_debug("%s: sec: %lu pfns: %lu set(%d, %d)\n", __func__, nr,
208 pfns, subsection_map_index(pfn),
209 subsection_map_index(pfn + pfns - 1));
210
211 pfn += pfns;
212 nr_pages -= pfns;
213 }
214}
215#else
216void __init subsection_map_init(unsigned long pfn, unsigned long nr_pages)
217{
218}
219#endif
220
221/* Record a memory area against a node. */
222static void __init memory_present(int nid, unsigned long start, unsigned long end)
223{
224 unsigned long pfn;
225
226 start &= PAGE_SECTION_MASK;
227 mminit_validate_memmodel_limits(&start, &end);
228 for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) {
229 unsigned long section_nr = pfn_to_section_nr(pfn);
230 struct mem_section *ms;
231
232 sparse_index_init(section_nr, nid);
233 set_section_nid(section_nr, nid);
234
235 ms = __nr_to_section(section_nr);
236 if (!ms->section_mem_map) {
237 ms->section_mem_map = sparse_encode_early_nid(nid) |
238 SECTION_IS_ONLINE;
239 __section_mark_present(ms, section_nr);
240 }
241 }
242}
243
244/*
245 * Mark all memblocks as present using memory_present().
246 * This is a convenience function that is useful to mark all of the systems
247 * memory as present during initialization.
248 */
249static void __init memblocks_present(void)
250{
251 unsigned long start, end;
252 int i, nid;
253
254#ifdef CONFIG_SPARSEMEM_EXTREME
255 if (unlikely(!mem_section)) {
256 unsigned long size, align;
257
258 size = sizeof(struct mem_section *) * NR_SECTION_ROOTS;
259 align = 1 << (INTERNODE_CACHE_SHIFT);
260 mem_section = memblock_alloc(size, align);
261 if (!mem_section)
262 panic("%s: Failed to allocate %lu bytes align=0x%lx\n",
263 __func__, size, align);
264 }
265#endif
266
267 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid)
268 memory_present(nid, start, end);
269}
270
271/*
272 * Subtle, we encode the real pfn into the mem_map such that
273 * the identity pfn - section_mem_map will return the actual
274 * physical page frame number.
275 */
276static unsigned long sparse_encode_mem_map(struct page *mem_map, unsigned long pnum)
277{
278 unsigned long coded_mem_map =
279 (unsigned long)(mem_map - (section_nr_to_pfn(pnum)));
280 BUILD_BUG_ON(SECTION_MAP_LAST_BIT > PFN_SECTION_SHIFT);
281 BUG_ON(coded_mem_map & ~SECTION_MAP_MASK);
282 return coded_mem_map;
283}
284
285#ifdef CONFIG_MEMORY_HOTPLUG
286/*
287 * Decode mem_map from the coded memmap
288 */
289struct page *sparse_decode_mem_map(unsigned long coded_mem_map, unsigned long pnum)
290{
291 /* mask off the extra low bits of information */
292 coded_mem_map &= SECTION_MAP_MASK;
293 return ((struct page *)coded_mem_map) + section_nr_to_pfn(pnum);
294}
295#endif /* CONFIG_MEMORY_HOTPLUG */
296
297static void __meminit sparse_init_one_section(struct mem_section *ms,
298 unsigned long pnum, struct page *mem_map,
299 struct mem_section_usage *usage, unsigned long flags)
300{
301 ms->section_mem_map &= ~SECTION_MAP_MASK;
302 ms->section_mem_map |= sparse_encode_mem_map(mem_map, pnum)
303 | SECTION_HAS_MEM_MAP | flags;
304 ms->usage = usage;
305}
306
307static unsigned long usemap_size(void)
308{
309 return BITS_TO_LONGS(SECTION_BLOCKFLAGS_BITS) * sizeof(unsigned long);
310}
311
312size_t mem_section_usage_size(void)
313{
314 return sizeof(struct mem_section_usage) + usemap_size();
315}
316
317#ifdef CONFIG_MEMORY_HOTREMOVE
318static inline phys_addr_t pgdat_to_phys(struct pglist_data *pgdat)
319{
320#ifndef CONFIG_NUMA
321 VM_BUG_ON(pgdat != &contig_page_data);
322 return __pa_symbol(&contig_page_data);
323#else
324 return __pa(pgdat);
325#endif
326}
327
328static struct mem_section_usage * __init
329sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
330 unsigned long size)
331{
332 struct mem_section_usage *usage;
333 unsigned long goal, limit;
334 int nid;
335 /*
336 * A page may contain usemaps for other sections preventing the
337 * page being freed and making a section unremovable while
338 * other sections referencing the usemap remain active. Similarly,
339 * a pgdat can prevent a section being removed. If section A
340 * contains a pgdat and section B contains the usemap, both
341 * sections become inter-dependent. This allocates usemaps
342 * from the same section as the pgdat where possible to avoid
343 * this problem.
344 */
345 goal = pgdat_to_phys(pgdat) & (PAGE_SECTION_MASK << PAGE_SHIFT);
346 limit = goal + (1UL << PA_SECTION_SHIFT);
347 nid = early_pfn_to_nid(goal >> PAGE_SHIFT);
348again:
349 usage = memblock_alloc_try_nid(size, SMP_CACHE_BYTES, goal, limit, nid);
350 if (!usage && limit) {
351 limit = MEMBLOCK_ALLOC_ACCESSIBLE;
352 goto again;
353 }
354 return usage;
355}
356
357static void __init check_usemap_section_nr(int nid,
358 struct mem_section_usage *usage)
359{
360 unsigned long usemap_snr, pgdat_snr;
361 static unsigned long old_usemap_snr;
362 static unsigned long old_pgdat_snr;
363 struct pglist_data *pgdat = NODE_DATA(nid);
364 int usemap_nid;
365
366 /* First call */
367 if (!old_usemap_snr) {
368 old_usemap_snr = NR_MEM_SECTIONS;
369 old_pgdat_snr = NR_MEM_SECTIONS;
370 }
371
372 usemap_snr = pfn_to_section_nr(__pa(usage) >> PAGE_SHIFT);
373 pgdat_snr = pfn_to_section_nr(pgdat_to_phys(pgdat) >> PAGE_SHIFT);
374 if (usemap_snr == pgdat_snr)
375 return;
376
377 if (old_usemap_snr == usemap_snr && old_pgdat_snr == pgdat_snr)
378 /* skip redundant message */
379 return;
380
381 old_usemap_snr = usemap_snr;
382 old_pgdat_snr = pgdat_snr;
383
384 usemap_nid = sparse_early_nid(__nr_to_section(usemap_snr));
385 if (usemap_nid != nid) {
386 pr_info("node %d must be removed before remove section %ld\n",
387 nid, usemap_snr);
388 return;
389 }
390 /*
391 * There is a circular dependency.
392 * Some platforms allow un-removable section because they will just
393 * gather other removable sections for dynamic partitioning.
394 * Just notify un-removable section's number here.
395 */
396 pr_info("Section %ld and %ld (node %d) have a circular dependency on usemap and pgdat allocations\n",
397 usemap_snr, pgdat_snr, nid);
398}
399#else
400static struct mem_section_usage * __init
401sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
402 unsigned long size)
403{
404 return memblock_alloc_node(size, SMP_CACHE_BYTES, pgdat->node_id);
405}
406
407static void __init check_usemap_section_nr(int nid,
408 struct mem_section_usage *usage)
409{
410}
411#endif /* CONFIG_MEMORY_HOTREMOVE */
412
413#ifdef CONFIG_SPARSEMEM_VMEMMAP
414static unsigned long __init section_map_size(void)
415{
416 return ALIGN(sizeof(struct page) * PAGES_PER_SECTION, PMD_SIZE);
417}
418
419#else
420static unsigned long __init section_map_size(void)
421{
422 return PAGE_ALIGN(sizeof(struct page) * PAGES_PER_SECTION);
423}
424
425struct page __init *__populate_section_memmap(unsigned long pfn,
426 unsigned long nr_pages, int nid, struct vmem_altmap *altmap,
427 struct dev_pagemap *pgmap)
428{
429 unsigned long size = section_map_size();
430 struct page *map = sparse_buffer_alloc(size);
431 phys_addr_t addr = __pa(MAX_DMA_ADDRESS);
432
433 if (map)
434 return map;
435
436 map = memmap_alloc(size, size, addr, nid, false);
437 if (!map)
438 panic("%s: Failed to allocate %lu bytes align=0x%lx nid=%d from=%pa\n",
439 __func__, size, PAGE_SIZE, nid, &addr);
440
441 return map;
442}
443#endif /* !CONFIG_SPARSEMEM_VMEMMAP */
444
445static void *sparsemap_buf __meminitdata;
446static void *sparsemap_buf_end __meminitdata;
447
448static inline void __meminit sparse_buffer_free(unsigned long size)
449{
450 WARN_ON(!sparsemap_buf || size == 0);
451 memblock_free(sparsemap_buf, size);
452}
453
454static void __init sparse_buffer_init(unsigned long size, int nid)
455{
456 phys_addr_t addr = __pa(MAX_DMA_ADDRESS);
457 WARN_ON(sparsemap_buf); /* forgot to call sparse_buffer_fini()? */
458 /*
459 * Pre-allocated buffer is mainly used by __populate_section_memmap
460 * and we want it to be properly aligned to the section size - this is
461 * especially the case for VMEMMAP which maps memmap to PMDs
462 */
463 sparsemap_buf = memmap_alloc(size, section_map_size(), addr, nid, true);
464 sparsemap_buf_end = sparsemap_buf + size;
465#ifndef CONFIG_SPARSEMEM_VMEMMAP
466 memmap_boot_pages_add(DIV_ROUND_UP(size, PAGE_SIZE));
467#endif
468}
469
470static void __init sparse_buffer_fini(void)
471{
472 unsigned long size = sparsemap_buf_end - sparsemap_buf;
473
474 if (sparsemap_buf && size > 0)
475 sparse_buffer_free(size);
476 sparsemap_buf = NULL;
477}
478
479void * __meminit sparse_buffer_alloc(unsigned long size)
480{
481 void *ptr = NULL;
482
483 if (sparsemap_buf) {
484 ptr = (void *) roundup((unsigned long)sparsemap_buf, size);
485 if (ptr + size > sparsemap_buf_end)
486 ptr = NULL;
487 else {
488 /* Free redundant aligned space */
489 if ((unsigned long)(ptr - sparsemap_buf) > 0)
490 sparse_buffer_free((unsigned long)(ptr - sparsemap_buf));
491 sparsemap_buf = ptr + size;
492 }
493 }
494 return ptr;
495}
496
497void __weak __meminit vmemmap_populate_print_last(void)
498{
499}
500
501/*
502 * Initialize sparse on a specific node. The node spans [pnum_begin, pnum_end)
503 * And number of present sections in this node is map_count.
504 */
505static void __init sparse_init_nid(int nid, unsigned long pnum_begin,
506 unsigned long pnum_end,
507 unsigned long map_count)
508{
509 struct mem_section_usage *usage;
510 unsigned long pnum;
511 struct page *map;
512
513 usage = sparse_early_usemaps_alloc_pgdat_section(NODE_DATA(nid),
514 mem_section_usage_size() * map_count);
515 if (!usage) {
516 pr_err("%s: node[%d] usemap allocation failed", __func__, nid);
517 goto failed;
518 }
519 sparse_buffer_init(map_count * section_map_size(), nid);
520 for_each_present_section_nr(pnum_begin, pnum) {
521 unsigned long pfn = section_nr_to_pfn(pnum);
522
523 if (pnum >= pnum_end)
524 break;
525
526 map = __populate_section_memmap(pfn, PAGES_PER_SECTION,
527 nid, NULL, NULL);
528 if (!map) {
529 pr_err("%s: node[%d] memory map backing failed. Some memory will not be available.",
530 __func__, nid);
531 pnum_begin = pnum;
532 sparse_buffer_fini();
533 goto failed;
534 }
535 check_usemap_section_nr(nid, usage);
536 sparse_init_one_section(__nr_to_section(pnum), pnum, map, usage,
537 SECTION_IS_EARLY);
538 usage = (void *) usage + mem_section_usage_size();
539 }
540 sparse_buffer_fini();
541 return;
542failed:
543 /* We failed to allocate, mark all the following pnums as not present */
544 for_each_present_section_nr(pnum_begin, pnum) {
545 struct mem_section *ms;
546
547 if (pnum >= pnum_end)
548 break;
549 ms = __nr_to_section(pnum);
550 ms->section_mem_map = 0;
551 }
552}
553
554/*
555 * Allocate the accumulated non-linear sections, allocate a mem_map
556 * for each and record the physical to section mapping.
557 */
558void __init sparse_init(void)
559{
560 unsigned long pnum_end, pnum_begin, map_count = 1;
561 int nid_begin;
562
563 /* see include/linux/mmzone.h 'struct mem_section' definition */
564 BUILD_BUG_ON(!is_power_of_2(sizeof(struct mem_section)));
565 memblocks_present();
566
567 pnum_begin = first_present_section_nr();
568 nid_begin = sparse_early_nid(__nr_to_section(pnum_begin));
569
570 /* Setup pageblock_order for HUGETLB_PAGE_SIZE_VARIABLE */
571 set_pageblock_order();
572
573 for_each_present_section_nr(pnum_begin + 1, pnum_end) {
574 int nid = sparse_early_nid(__nr_to_section(pnum_end));
575
576 if (nid == nid_begin) {
577 map_count++;
578 continue;
579 }
580 /* Init node with sections in range [pnum_begin, pnum_end) */
581 sparse_init_nid(nid_begin, pnum_begin, pnum_end, map_count);
582 nid_begin = nid;
583 pnum_begin = pnum_end;
584 map_count = 1;
585 }
586 /* cover the last node */
587 sparse_init_nid(nid_begin, pnum_begin, pnum_end, map_count);
588 vmemmap_populate_print_last();
589}
590
591#ifdef CONFIG_MEMORY_HOTPLUG
592
593/* Mark all memory sections within the pfn range as online */
594void online_mem_sections(unsigned long start_pfn, unsigned long end_pfn)
595{
596 unsigned long pfn;
597
598 for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
599 unsigned long section_nr = pfn_to_section_nr(pfn);
600 struct mem_section *ms;
601
602 /* onlining code should never touch invalid ranges */
603 if (WARN_ON(!valid_section_nr(section_nr)))
604 continue;
605
606 ms = __nr_to_section(section_nr);
607 ms->section_mem_map |= SECTION_IS_ONLINE;
608 }
609}
610
611/* Mark all memory sections within the pfn range as offline */
612void offline_mem_sections(unsigned long start_pfn, unsigned long end_pfn)
613{
614 unsigned long pfn;
615
616 for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
617 unsigned long section_nr = pfn_to_section_nr(pfn);
618 struct mem_section *ms;
619
620 /*
621 * TODO this needs some double checking. Offlining code makes
622 * sure to check pfn_valid but those checks might be just bogus
623 */
624 if (WARN_ON(!valid_section_nr(section_nr)))
625 continue;
626
627 ms = __nr_to_section(section_nr);
628 ms->section_mem_map &= ~SECTION_IS_ONLINE;
629 }
630}
631
632#ifdef CONFIG_SPARSEMEM_VMEMMAP
633static struct page * __meminit populate_section_memmap(unsigned long pfn,
634 unsigned long nr_pages, int nid, struct vmem_altmap *altmap,
635 struct dev_pagemap *pgmap)
636{
637 return __populate_section_memmap(pfn, nr_pages, nid, altmap, pgmap);
638}
639
640static void depopulate_section_memmap(unsigned long pfn, unsigned long nr_pages,
641 struct vmem_altmap *altmap)
642{
643 unsigned long start = (unsigned long) pfn_to_page(pfn);
644 unsigned long end = start + nr_pages * sizeof(struct page);
645
646 memmap_pages_add(-1L * (DIV_ROUND_UP(end - start, PAGE_SIZE)));
647 vmemmap_free(start, end, altmap);
648}
649static void free_map_bootmem(struct page *memmap)
650{
651 unsigned long start = (unsigned long)memmap;
652 unsigned long end = (unsigned long)(memmap + PAGES_PER_SECTION);
653
654 vmemmap_free(start, end, NULL);
655}
656
657static int clear_subsection_map(unsigned long pfn, unsigned long nr_pages)
658{
659 DECLARE_BITMAP(map, SUBSECTIONS_PER_SECTION) = { 0 };
660 DECLARE_BITMAP(tmp, SUBSECTIONS_PER_SECTION) = { 0 };
661 struct mem_section *ms = __pfn_to_section(pfn);
662 unsigned long *subsection_map = ms->usage
663 ? &ms->usage->subsection_map[0] : NULL;
664
665 subsection_mask_set(map, pfn, nr_pages);
666 if (subsection_map)
667 bitmap_and(tmp, map, subsection_map, SUBSECTIONS_PER_SECTION);
668
669 if (WARN(!subsection_map || !bitmap_equal(tmp, map, SUBSECTIONS_PER_SECTION),
670 "section already deactivated (%#lx + %ld)\n",
671 pfn, nr_pages))
672 return -EINVAL;
673
674 bitmap_xor(subsection_map, map, subsection_map, SUBSECTIONS_PER_SECTION);
675 return 0;
676}
677
678static bool is_subsection_map_empty(struct mem_section *ms)
679{
680 return bitmap_empty(&ms->usage->subsection_map[0],
681 SUBSECTIONS_PER_SECTION);
682}
683
684static int fill_subsection_map(unsigned long pfn, unsigned long nr_pages)
685{
686 struct mem_section *ms = __pfn_to_section(pfn);
687 DECLARE_BITMAP(map, SUBSECTIONS_PER_SECTION) = { 0 };
688 unsigned long *subsection_map;
689 int rc = 0;
690
691 subsection_mask_set(map, pfn, nr_pages);
692
693 subsection_map = &ms->usage->subsection_map[0];
694
695 if (bitmap_empty(map, SUBSECTIONS_PER_SECTION))
696 rc = -EINVAL;
697 else if (bitmap_intersects(map, subsection_map, SUBSECTIONS_PER_SECTION))
698 rc = -EEXIST;
699 else
700 bitmap_or(subsection_map, map, subsection_map,
701 SUBSECTIONS_PER_SECTION);
702
703 return rc;
704}
705#else
706static struct page * __meminit populate_section_memmap(unsigned long pfn,
707 unsigned long nr_pages, int nid, struct vmem_altmap *altmap,
708 struct dev_pagemap *pgmap)
709{
710 return kvmalloc_node(array_size(sizeof(struct page),
711 PAGES_PER_SECTION), GFP_KERNEL, nid);
712}
713
714static void depopulate_section_memmap(unsigned long pfn, unsigned long nr_pages,
715 struct vmem_altmap *altmap)
716{
717 kvfree(pfn_to_page(pfn));
718}
719
720static void free_map_bootmem(struct page *memmap)
721{
722 unsigned long maps_section_nr, removing_section_nr, i;
723 unsigned long type, nr_pages;
724 struct page *page = virt_to_page(memmap);
725
726 nr_pages = PAGE_ALIGN(PAGES_PER_SECTION * sizeof(struct page))
727 >> PAGE_SHIFT;
728
729 for (i = 0; i < nr_pages; i++, page++) {
730 type = bootmem_type(page);
731
732 BUG_ON(type == NODE_INFO);
733
734 maps_section_nr = pfn_to_section_nr(page_to_pfn(page));
735 removing_section_nr = bootmem_info(page);
736
737 /*
738 * When this function is called, the removing section is
739 * logical offlined state. This means all pages are isolated
740 * from page allocator. If removing section's memmap is placed
741 * on the same section, it must not be freed.
742 * If it is freed, page allocator may allocate it which will
743 * be removed physically soon.
744 */
745 if (maps_section_nr != removing_section_nr)
746 put_page_bootmem(page);
747 }
748}
749
750static int clear_subsection_map(unsigned long pfn, unsigned long nr_pages)
751{
752 return 0;
753}
754
755static bool is_subsection_map_empty(struct mem_section *ms)
756{
757 return true;
758}
759
760static int fill_subsection_map(unsigned long pfn, unsigned long nr_pages)
761{
762 return 0;
763}
764#endif /* CONFIG_SPARSEMEM_VMEMMAP */
765
766/*
767 * To deactivate a memory region, there are 3 cases to handle across
768 * two configurations (SPARSEMEM_VMEMMAP={y,n}):
769 *
770 * 1. deactivation of a partial hot-added section (only possible in
771 * the SPARSEMEM_VMEMMAP=y case).
772 * a) section was present at memory init.
773 * b) section was hot-added post memory init.
774 * 2. deactivation of a complete hot-added section.
775 * 3. deactivation of a complete section from memory init.
776 *
777 * For 1, when subsection_map does not empty we will not be freeing the
778 * usage map, but still need to free the vmemmap range.
779 *
780 * For 2 and 3, the SPARSEMEM_VMEMMAP={y,n} cases are unified
781 */
782static void section_deactivate(unsigned long pfn, unsigned long nr_pages,
783 struct vmem_altmap *altmap)
784{
785 struct mem_section *ms = __pfn_to_section(pfn);
786 bool section_is_early = early_section(ms);
787 struct page *memmap = NULL;
788 bool empty;
789
790 if (clear_subsection_map(pfn, nr_pages))
791 return;
792
793 empty = is_subsection_map_empty(ms);
794 if (empty) {
795 unsigned long section_nr = pfn_to_section_nr(pfn);
796
797 /*
798 * Mark the section invalid so that valid_section()
799 * return false. This prevents code from dereferencing
800 * ms->usage array.
801 */
802 ms->section_mem_map &= ~SECTION_HAS_MEM_MAP;
803
804 /*
805 * When removing an early section, the usage map is kept (as the
806 * usage maps of other sections fall into the same page). It
807 * will be re-used when re-adding the section - which is then no
808 * longer an early section. If the usage map is PageReserved, it
809 * was allocated during boot.
810 */
811 if (!PageReserved(virt_to_page(ms->usage))) {
812 kfree_rcu(ms->usage, rcu);
813 WRITE_ONCE(ms->usage, NULL);
814 }
815 memmap = sparse_decode_mem_map(ms->section_mem_map, section_nr);
816 }
817
818 /*
819 * The memmap of early sections is always fully populated. See
820 * section_activate() and pfn_valid() .
821 */
822 if (!section_is_early)
823 depopulate_section_memmap(pfn, nr_pages, altmap);
824 else if (memmap)
825 free_map_bootmem(memmap);
826
827 if (empty)
828 ms->section_mem_map = (unsigned long)NULL;
829}
830
831static struct page * __meminit section_activate(int nid, unsigned long pfn,
832 unsigned long nr_pages, struct vmem_altmap *altmap,
833 struct dev_pagemap *pgmap)
834{
835 struct mem_section *ms = __pfn_to_section(pfn);
836 struct mem_section_usage *usage = NULL;
837 struct page *memmap;
838 int rc;
839
840 if (!ms->usage) {
841 usage = kzalloc(mem_section_usage_size(), GFP_KERNEL);
842 if (!usage)
843 return ERR_PTR(-ENOMEM);
844 ms->usage = usage;
845 }
846
847 rc = fill_subsection_map(pfn, nr_pages);
848 if (rc) {
849 if (usage)
850 ms->usage = NULL;
851 kfree(usage);
852 return ERR_PTR(rc);
853 }
854
855 /*
856 * The early init code does not consider partially populated
857 * initial sections, it simply assumes that memory will never be
858 * referenced. If we hot-add memory into such a section then we
859 * do not need to populate the memmap and can simply reuse what
860 * is already there.
861 */
862 if (nr_pages < PAGES_PER_SECTION && early_section(ms))
863 return pfn_to_page(pfn);
864
865 memmap = populate_section_memmap(pfn, nr_pages, nid, altmap, pgmap);
866 if (!memmap) {
867 section_deactivate(pfn, nr_pages, altmap);
868 return ERR_PTR(-ENOMEM);
869 }
870
871 return memmap;
872}
873
874/**
875 * sparse_add_section - add a memory section, or populate an existing one
876 * @nid: The node to add section on
877 * @start_pfn: start pfn of the memory range
878 * @nr_pages: number of pfns to add in the section
879 * @altmap: alternate pfns to allocate the memmap backing store
880 * @pgmap: alternate compound page geometry for devmap mappings
881 *
882 * This is only intended for hotplug.
883 *
884 * Note that only VMEMMAP supports sub-section aligned hotplug,
885 * the proper alignment and size are gated by check_pfn_span().
886 *
887 *
888 * Return:
889 * * 0 - On success.
890 * * -EEXIST - Section has been present.
891 * * -ENOMEM - Out of memory.
892 */
893int __meminit sparse_add_section(int nid, unsigned long start_pfn,
894 unsigned long nr_pages, struct vmem_altmap *altmap,
895 struct dev_pagemap *pgmap)
896{
897 unsigned long section_nr = pfn_to_section_nr(start_pfn);
898 struct mem_section *ms;
899 struct page *memmap;
900 int ret;
901
902 ret = sparse_index_init(section_nr, nid);
903 if (ret < 0)
904 return ret;
905
906 memmap = section_activate(nid, start_pfn, nr_pages, altmap, pgmap);
907 if (IS_ERR(memmap))
908 return PTR_ERR(memmap);
909
910 /*
911 * Poison uninitialized struct pages in order to catch invalid flags
912 * combinations.
913 */
914 if (!altmap || !altmap->inaccessible)
915 page_init_poison(memmap, sizeof(struct page) * nr_pages);
916
917 ms = __nr_to_section(section_nr);
918 set_section_nid(section_nr, nid);
919 __section_mark_present(ms, section_nr);
920
921 /* Align memmap to section boundary in the subsection case */
922 if (section_nr_to_pfn(section_nr) != start_pfn)
923 memmap = pfn_to_page(section_nr_to_pfn(section_nr));
924 sparse_init_one_section(ms, section_nr, memmap, ms->usage, 0);
925
926 return 0;
927}
928
929void sparse_remove_section(unsigned long pfn, unsigned long nr_pages,
930 struct vmem_altmap *altmap)
931{
932 struct mem_section *ms = __pfn_to_section(pfn);
933
934 if (WARN_ON_ONCE(!valid_section(ms)))
935 return;
936
937 section_deactivate(pfn, nr_pages, altmap);
938}
939#endif /* CONFIG_MEMORY_HOTPLUG */