1 // SPDX-License-Identifier: GPL-2.0
3 * sparse memory mappings.
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>
17 #include <asm/pgalloc.h>
18 #include <asm/pgtable.h>
21 * Permanent SPARSEMEM data:
23 * 1) mem_section - memory sections, mem_map's for valid memory
25 #ifdef CONFIG_SPARSEMEM_EXTREME
26 struct mem_section **mem_section;
28 struct mem_section mem_section[NR_SECTION_ROOTS][SECTIONS_PER_ROOT]
29 ____cacheline_internodealigned_in_smp;
31 EXPORT_SYMBOL(mem_section);
33 #ifdef NODE_NOT_IN_PAGE_FLAGS
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.
39 #if MAX_NUMNODES <= 256
40 static u8 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned;
42 static u16 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned;
45 int page_to_nid(const struct page *page)
47 return section_to_node_table[page_to_section(page)];
49 EXPORT_SYMBOL(page_to_nid);
51 static void set_section_nid(unsigned long section_nr, int nid)
53 section_to_node_table[section_nr] = nid;
55 #else /* !NODE_NOT_IN_PAGE_FLAGS */
56 static inline void set_section_nid(unsigned long section_nr, int nid)
61 #ifdef CONFIG_SPARSEMEM_EXTREME
62 static noinline struct mem_section __ref *sparse_index_alloc(int nid)
64 struct mem_section *section = NULL;
65 unsigned long array_size = SECTIONS_PER_ROOT *
66 sizeof(struct mem_section);
68 if (slab_is_available())
69 section = kzalloc_node(array_size, GFP_KERNEL, nid);
71 section = memblock_alloc_node(array_size, SMP_CACHE_BYTES,
77 static int __meminit sparse_index_init(unsigned long section_nr, int nid)
79 unsigned long root = SECTION_NR_TO_ROOT(section_nr);
80 struct mem_section *section;
82 if (mem_section[root])
85 section = sparse_index_alloc(nid);
89 mem_section[root] = section;
93 #else /* !SPARSEMEM_EXTREME */
94 static inline int sparse_index_init(unsigned long section_nr, int nid)
100 #ifdef CONFIG_SPARSEMEM_EXTREME
101 int __section_nr(struct mem_section* ms)
103 unsigned long root_nr;
104 struct mem_section *root = NULL;
106 for (root_nr = 0; root_nr < NR_SECTION_ROOTS; root_nr++) {
107 root = __nr_to_section(root_nr * SECTIONS_PER_ROOT);
111 if ((ms >= root) && (ms < (root + SECTIONS_PER_ROOT)))
117 return (root_nr * SECTIONS_PER_ROOT) + (ms - root);
120 int __section_nr(struct mem_section* ms)
122 return (int)(ms - mem_section[0]);
127 * During early boot, before section_mem_map is used for an actual
128 * mem_map, we use section_mem_map to store the section's NUMA
129 * node. This keeps us from having to use another data structure. The
130 * node information is cleared just before we store the real mem_map.
132 static inline unsigned long sparse_encode_early_nid(int nid)
134 return (nid << SECTION_NID_SHIFT);
137 static inline int sparse_early_nid(struct mem_section *section)
139 return (section->section_mem_map >> SECTION_NID_SHIFT);
142 /* Validate the physical addressing limitations of the model */
143 void __meminit mminit_validate_memmodel_limits(unsigned long *start_pfn,
144 unsigned long *end_pfn)
146 unsigned long max_sparsemem_pfn = 1UL << (MAX_PHYSMEM_BITS-PAGE_SHIFT);
149 * Sanity checks - do not allow an architecture to pass
150 * in larger pfns than the maximum scope of sparsemem:
152 if (*start_pfn > max_sparsemem_pfn) {
153 mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
154 "Start of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
155 *start_pfn, *end_pfn, max_sparsemem_pfn);
157 *start_pfn = max_sparsemem_pfn;
158 *end_pfn = max_sparsemem_pfn;
159 } else if (*end_pfn > max_sparsemem_pfn) {
160 mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
161 "End of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
162 *start_pfn, *end_pfn, max_sparsemem_pfn);
164 *end_pfn = max_sparsemem_pfn;
169 * There are a number of times that we loop over NR_MEM_SECTIONS,
170 * looking for section_present() on each. But, when we have very
171 * large physical address spaces, NR_MEM_SECTIONS can also be
172 * very large which makes the loops quite long.
174 * Keeping track of this gives us an easy way to break out of
177 int __highest_present_section_nr;
178 static void section_mark_present(struct mem_section *ms)
180 int section_nr = __section_nr(ms);
182 if (section_nr > __highest_present_section_nr)
183 __highest_present_section_nr = section_nr;
185 ms->section_mem_map |= SECTION_MARKED_PRESENT;
188 static inline int next_present_section_nr(int section_nr)
192 if (present_section_nr(section_nr))
194 } while ((section_nr <= __highest_present_section_nr));
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 <= __highest_present_section_nr)); \
202 section_nr = next_present_section_nr(section_nr))
204 static inline unsigned long first_present_section_nr(void)
206 return next_present_section_nr(-1);
209 /* Record a memory area against a node. */
210 void __init memory_present(int nid, unsigned long start, unsigned long end)
214 #ifdef CONFIG_SPARSEMEM_EXTREME
215 if (unlikely(!mem_section)) {
216 unsigned long size, align;
218 size = sizeof(struct mem_section*) * NR_SECTION_ROOTS;
219 align = 1 << (INTERNODE_CACHE_SHIFT);
220 mem_section = memblock_alloc(size, align);
224 start &= PAGE_SECTION_MASK;
225 mminit_validate_memmodel_limits(&start, &end);
226 for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) {
227 unsigned long section = pfn_to_section_nr(pfn);
228 struct mem_section *ms;
230 sparse_index_init(section, nid);
231 set_section_nid(section, nid);
233 ms = __nr_to_section(section);
234 if (!ms->section_mem_map) {
235 ms->section_mem_map = sparse_encode_early_nid(nid) |
237 section_mark_present(ms);
243 * Subtle, we encode the real pfn into the mem_map such that
244 * the identity pfn - section_mem_map will return the actual
245 * physical page frame number.
247 static unsigned long sparse_encode_mem_map(struct page *mem_map, unsigned long pnum)
249 unsigned long coded_mem_map =
250 (unsigned long)(mem_map - (section_nr_to_pfn(pnum)));
251 BUILD_BUG_ON(SECTION_MAP_LAST_BIT > (1UL<<PFN_SECTION_SHIFT));
252 BUG_ON(coded_mem_map & ~SECTION_MAP_MASK);
253 return coded_mem_map;
257 * Decode mem_map from the coded memmap
259 struct page *sparse_decode_mem_map(unsigned long coded_mem_map, unsigned long pnum)
261 /* mask off the extra low bits of information */
262 coded_mem_map &= SECTION_MAP_MASK;
263 return ((struct page *)coded_mem_map) + section_nr_to_pfn(pnum);
266 static void __meminit sparse_init_one_section(struct mem_section *ms,
267 unsigned long pnum, struct page *mem_map,
268 unsigned long *pageblock_bitmap)
270 ms->section_mem_map &= ~SECTION_MAP_MASK;
271 ms->section_mem_map |= sparse_encode_mem_map(mem_map, pnum) |
273 ms->pageblock_flags = pageblock_bitmap;
276 unsigned long usemap_size(void)
278 return BITS_TO_LONGS(SECTION_BLOCKFLAGS_BITS) * sizeof(unsigned long);
281 #ifdef CONFIG_MEMORY_HOTPLUG
282 static unsigned long *__kmalloc_section_usemap(void)
284 return kmalloc(usemap_size(), GFP_KERNEL);
286 #endif /* CONFIG_MEMORY_HOTPLUG */
288 #ifdef CONFIG_MEMORY_HOTREMOVE
289 static unsigned long * __init
290 sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
293 unsigned long goal, limit;
297 * A page may contain usemaps for other sections preventing the
298 * page being freed and making a section unremovable while
299 * other sections referencing the usemap remain active. Similarly,
300 * a pgdat can prevent a section being removed. If section A
301 * contains a pgdat and section B contains the usemap, both
302 * sections become inter-dependent. This allocates usemaps
303 * from the same section as the pgdat where possible to avoid
306 goal = __pa(pgdat) & (PAGE_SECTION_MASK << PAGE_SHIFT);
307 limit = goal + (1UL << PA_SECTION_SHIFT);
308 nid = early_pfn_to_nid(goal >> PAGE_SHIFT);
310 p = memblock_alloc_try_nid_nopanic(size,
311 SMP_CACHE_BYTES, goal, limit,
320 static void __init check_usemap_section_nr(int nid, unsigned long *usemap)
322 unsigned long usemap_snr, pgdat_snr;
323 static unsigned long old_usemap_snr;
324 static unsigned long old_pgdat_snr;
325 struct pglist_data *pgdat = NODE_DATA(nid);
329 if (!old_usemap_snr) {
330 old_usemap_snr = NR_MEM_SECTIONS;
331 old_pgdat_snr = NR_MEM_SECTIONS;
334 usemap_snr = pfn_to_section_nr(__pa(usemap) >> PAGE_SHIFT);
335 pgdat_snr = pfn_to_section_nr(__pa(pgdat) >> PAGE_SHIFT);
336 if (usemap_snr == pgdat_snr)
339 if (old_usemap_snr == usemap_snr && old_pgdat_snr == pgdat_snr)
340 /* skip redundant message */
343 old_usemap_snr = usemap_snr;
344 old_pgdat_snr = pgdat_snr;
346 usemap_nid = sparse_early_nid(__nr_to_section(usemap_snr));
347 if (usemap_nid != nid) {
348 pr_info("node %d must be removed before remove section %ld\n",
353 * There is a circular dependency.
354 * Some platforms allow un-removable section because they will just
355 * gather other removable sections for dynamic partitioning.
356 * Just notify un-removable section's number here.
358 pr_info("Section %ld and %ld (node %d) have a circular dependency on usemap and pgdat allocations\n",
359 usemap_snr, pgdat_snr, nid);
362 static unsigned long * __init
363 sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
366 return memblock_alloc_node_nopanic(size, pgdat->node_id);
369 static void __init check_usemap_section_nr(int nid, unsigned long *usemap)
372 #endif /* CONFIG_MEMORY_HOTREMOVE */
374 #ifdef CONFIG_SPARSEMEM_VMEMMAP
375 static unsigned long __init section_map_size(void)
377 return ALIGN(sizeof(struct page) * PAGES_PER_SECTION, PMD_SIZE);
381 static unsigned long __init section_map_size(void)
383 return PAGE_ALIGN(sizeof(struct page) * PAGES_PER_SECTION);
386 struct page __init *sparse_mem_map_populate(unsigned long pnum, int nid,
387 struct vmem_altmap *altmap)
389 unsigned long size = section_map_size();
390 struct page *map = sparse_buffer_alloc(size);
395 map = memblock_alloc_try_nid(size,
396 PAGE_SIZE, __pa(MAX_DMA_ADDRESS),
397 MEMBLOCK_ALLOC_ACCESSIBLE, nid);
400 #endif /* !CONFIG_SPARSEMEM_VMEMMAP */
402 static void *sparsemap_buf __meminitdata;
403 static void *sparsemap_buf_end __meminitdata;
405 static void __init sparse_buffer_init(unsigned long size, int nid)
407 WARN_ON(sparsemap_buf); /* forgot to call sparse_buffer_fini()? */
409 memblock_alloc_try_nid_raw(size, PAGE_SIZE,
410 __pa(MAX_DMA_ADDRESS),
411 MEMBLOCK_ALLOC_ACCESSIBLE, nid);
412 sparsemap_buf_end = sparsemap_buf + size;
415 static void __init sparse_buffer_fini(void)
417 unsigned long size = sparsemap_buf_end - sparsemap_buf;
419 if (sparsemap_buf && size > 0)
420 memblock_free_early(__pa(sparsemap_buf), size);
421 sparsemap_buf = NULL;
424 void * __meminit sparse_buffer_alloc(unsigned long size)
429 ptr = PTR_ALIGN(sparsemap_buf, size);
430 if (ptr + size > sparsemap_buf_end)
433 sparsemap_buf = ptr + size;
438 void __weak __meminit vmemmap_populate_print_last(void)
443 * Initialize sparse on a specific node. The node spans [pnum_begin, pnum_end)
444 * And number of present sections in this node is map_count.
446 static void __init sparse_init_nid(int nid, unsigned long pnum_begin,
447 unsigned long pnum_end,
448 unsigned long map_count)
450 unsigned long pnum, usemap_longs, *usemap;
453 usemap_longs = BITS_TO_LONGS(SECTION_BLOCKFLAGS_BITS);
454 usemap = sparse_early_usemaps_alloc_pgdat_section(NODE_DATA(nid),
458 pr_err("%s: node[%d] usemap allocation failed", __func__, nid);
461 sparse_buffer_init(map_count * section_map_size(), nid);
462 for_each_present_section_nr(pnum_begin, pnum) {
463 if (pnum >= pnum_end)
466 map = sparse_mem_map_populate(pnum, nid, NULL);
468 pr_err("%s: node[%d] memory map backing failed. Some memory will not be available.",
473 check_usemap_section_nr(nid, usemap);
474 sparse_init_one_section(__nr_to_section(pnum), pnum, map, usemap);
475 usemap += usemap_longs;
477 sparse_buffer_fini();
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;
484 if (pnum >= pnum_end)
486 ms = __nr_to_section(pnum);
487 ms->section_mem_map = 0;
492 * Allocate the accumulated non-linear sections, allocate a mem_map
493 * for each and record the physical to section mapping.
495 void __init sparse_init(void)
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;
501 /* Setup pageblock_order for HUGETLB_PAGE_SIZE_VARIABLE */
502 set_pageblock_order();
504 for_each_present_section_nr(pnum_begin + 1, pnum_end) {
505 int nid = sparse_early_nid(__nr_to_section(pnum_end));
507 if (nid == nid_begin) {
511 /* Init node with sections in range [pnum_begin, pnum_end) */
512 sparse_init_nid(nid_begin, pnum_begin, pnum_end, map_count);
514 pnum_begin = pnum_end;
517 /* cover the last node */
518 sparse_init_nid(nid_begin, pnum_begin, pnum_end, map_count);
519 vmemmap_populate_print_last();
522 #ifdef CONFIG_MEMORY_HOTPLUG
524 /* Mark all memory sections within the pfn range as online */
525 void online_mem_sections(unsigned long start_pfn, unsigned long end_pfn)
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;
533 /* onlining code should never touch invalid ranges */
534 if (WARN_ON(!valid_section_nr(section_nr)))
537 ms = __nr_to_section(section_nr);
538 ms->section_mem_map |= SECTION_IS_ONLINE;
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)
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;
553 * TODO this needs some double checking. Offlining code makes
554 * sure to check pfn_valid but those checks might be just bogus
556 if (WARN_ON(!valid_section_nr(section_nr)))
559 ms = __nr_to_section(section_nr);
560 ms->section_mem_map &= ~SECTION_IS_ONLINE;
565 #ifdef CONFIG_SPARSEMEM_VMEMMAP
566 static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid,
567 struct vmem_altmap *altmap)
569 /* This will make the necessary allocations eventually. */
570 return sparse_mem_map_populate(pnum, nid, altmap);
572 static void __kfree_section_memmap(struct page *memmap,
573 struct vmem_altmap *altmap)
575 unsigned long start = (unsigned long)memmap;
576 unsigned long end = (unsigned long)(memmap + PAGES_PER_SECTION);
578 vmemmap_free(start, end, altmap);
580 #ifdef CONFIG_MEMORY_HOTREMOVE
581 static void free_map_bootmem(struct page *memmap)
583 unsigned long start = (unsigned long)memmap;
584 unsigned long end = (unsigned long)(memmap + PAGES_PER_SECTION);
586 vmemmap_free(start, end, NULL);
588 #endif /* CONFIG_MEMORY_HOTREMOVE */
590 static struct page *__kmalloc_section_memmap(void)
592 struct page *page, *ret;
593 unsigned long memmap_size = sizeof(struct page) * PAGES_PER_SECTION;
595 page = alloc_pages(GFP_KERNEL|__GFP_NOWARN, get_order(memmap_size));
599 ret = vmalloc(memmap_size);
605 ret = (struct page *)pfn_to_kaddr(page_to_pfn(page));
611 static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid,
612 struct vmem_altmap *altmap)
614 return __kmalloc_section_memmap();
617 static void __kfree_section_memmap(struct page *memmap,
618 struct vmem_altmap *altmap)
620 if (is_vmalloc_addr(memmap))
623 free_pages((unsigned long)memmap,
624 get_order(sizeof(struct page) * PAGES_PER_SECTION));
627 #ifdef CONFIG_MEMORY_HOTREMOVE
628 static void free_map_bootmem(struct page *memmap)
630 unsigned long maps_section_nr, removing_section_nr, i;
631 unsigned long magic, nr_pages;
632 struct page *page = virt_to_page(memmap);
634 nr_pages = PAGE_ALIGN(PAGES_PER_SECTION * sizeof(struct page))
637 for (i = 0; i < nr_pages; i++, page++) {
638 magic = (unsigned long) page->freelist;
640 BUG_ON(magic == NODE_INFO);
642 maps_section_nr = pfn_to_section_nr(page_to_pfn(page));
643 removing_section_nr = page_private(page);
646 * When this function is called, the removing section is
647 * logical offlined state. This means all pages are isolated
648 * from page allocator. If removing section's memmap is placed
649 * on the same section, it must not be freed.
650 * If it is freed, page allocator may allocate it which will
651 * be removed physically soon.
653 if (maps_section_nr != removing_section_nr)
654 put_page_bootmem(page);
657 #endif /* CONFIG_MEMORY_HOTREMOVE */
658 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
661 * returns the number of sections whose mem_maps were properly
662 * set. If this is <=0, then that means that the passed-in
663 * map was not consumed and must be freed.
665 int __meminit sparse_add_one_section(struct pglist_data *pgdat,
666 unsigned long start_pfn, struct vmem_altmap *altmap)
668 unsigned long section_nr = pfn_to_section_nr(start_pfn);
669 struct mem_section *ms;
671 unsigned long *usemap;
676 * no locking for this, because it does its own
677 * plus, it does a kmalloc
679 ret = sparse_index_init(section_nr, pgdat->node_id);
680 if (ret < 0 && ret != -EEXIST)
683 memmap = kmalloc_section_memmap(section_nr, pgdat->node_id, altmap);
686 usemap = __kmalloc_section_usemap();
688 __kfree_section_memmap(memmap, altmap);
692 pgdat_resize_lock(pgdat, &flags);
694 ms = __pfn_to_section(start_pfn);
695 if (ms->section_mem_map & SECTION_MARKED_PRESENT) {
701 * Poison uninitialized struct pages in order to catch invalid flags
704 page_init_poison(memmap, sizeof(struct page) * PAGES_PER_SECTION);
706 section_mark_present(ms);
707 sparse_init_one_section(ms, section_nr, memmap, usemap);
710 pgdat_resize_unlock(pgdat, &flags);
713 __kfree_section_memmap(memmap, altmap);
718 #ifdef CONFIG_MEMORY_HOTREMOVE
719 #ifdef CONFIG_MEMORY_FAILURE
720 static void clear_hwpoisoned_pages(struct page *memmap, int nr_pages)
727 for (i = 0; i < nr_pages; i++) {
728 if (PageHWPoison(&memmap[i])) {
729 atomic_long_sub(1, &num_poisoned_pages);
730 ClearPageHWPoison(&memmap[i]);
735 static inline void clear_hwpoisoned_pages(struct page *memmap, int nr_pages)
740 static void free_section_usemap(struct page *memmap, unsigned long *usemap,
741 struct vmem_altmap *altmap)
743 struct page *usemap_page;
748 usemap_page = virt_to_page(usemap);
750 * Check to see if allocation came from hot-plug-add
752 if (PageSlab(usemap_page) || PageCompound(usemap_page)) {
755 __kfree_section_memmap(memmap, altmap);
760 * The usemap came from bootmem. This is packed with other usemaps
761 * on the section which has pgdat at boot time. Just keep it as is now.
765 free_map_bootmem(memmap);
768 void sparse_remove_one_section(struct zone *zone, struct mem_section *ms,
769 unsigned long map_offset, struct vmem_altmap *altmap)
771 struct page *memmap = NULL;
772 unsigned long *usemap = NULL, flags;
773 struct pglist_data *pgdat = zone->zone_pgdat;
775 pgdat_resize_lock(pgdat, &flags);
776 if (ms->section_mem_map) {
777 usemap = ms->pageblock_flags;
778 memmap = sparse_decode_mem_map(ms->section_mem_map,
780 ms->section_mem_map = 0;
781 ms->pageblock_flags = NULL;
783 pgdat_resize_unlock(pgdat, &flags);
785 clear_hwpoisoned_pages(memmap + map_offset,
786 PAGES_PER_SECTION - map_offset);
787 free_section_usemap(memmap, usemap, altmap);
789 #endif /* CONFIG_MEMORY_HOTREMOVE */
790 #endif /* CONFIG_MEMORY_HOTPLUG */