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,
74 panic("%s: Failed to allocate %lu bytes nid=%d\n",
75 __func__, array_size, nid);
81 static int __meminit sparse_index_init(unsigned long section_nr, int nid)
83 unsigned long root = SECTION_NR_TO_ROOT(section_nr);
84 struct mem_section *section;
86 if (mem_section[root])
89 section = sparse_index_alloc(nid);
93 mem_section[root] = section;
97 #else /* !SPARSEMEM_EXTREME */
98 static inline int sparse_index_init(unsigned long section_nr, int nid)
104 #ifdef CONFIG_SPARSEMEM_EXTREME
105 int __section_nr(struct mem_section* ms)
107 unsigned long root_nr;
108 struct mem_section *root = NULL;
110 for (root_nr = 0; root_nr < NR_SECTION_ROOTS; root_nr++) {
111 root = __nr_to_section(root_nr * SECTIONS_PER_ROOT);
115 if ((ms >= root) && (ms < (root + SECTIONS_PER_ROOT)))
121 return (root_nr * SECTIONS_PER_ROOT) + (ms - root);
124 int __section_nr(struct mem_section* ms)
126 return (int)(ms - mem_section[0]);
131 * During early boot, before section_mem_map is used for an actual
132 * mem_map, we use section_mem_map to store the section's NUMA
133 * node. This keeps us from having to use another data structure. The
134 * node information is cleared just before we store the real mem_map.
136 static inline unsigned long sparse_encode_early_nid(int nid)
138 return (nid << SECTION_NID_SHIFT);
141 static inline int sparse_early_nid(struct mem_section *section)
143 return (section->section_mem_map >> SECTION_NID_SHIFT);
146 /* Validate the physical addressing limitations of the model */
147 void __meminit mminit_validate_memmodel_limits(unsigned long *start_pfn,
148 unsigned long *end_pfn)
150 unsigned long max_sparsemem_pfn = 1UL << (MAX_PHYSMEM_BITS-PAGE_SHIFT);
153 * Sanity checks - do not allow an architecture to pass
154 * in larger pfns than the maximum scope of sparsemem:
156 if (*start_pfn > max_sparsemem_pfn) {
157 mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
158 "Start of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
159 *start_pfn, *end_pfn, max_sparsemem_pfn);
161 *start_pfn = max_sparsemem_pfn;
162 *end_pfn = max_sparsemem_pfn;
163 } else if (*end_pfn > max_sparsemem_pfn) {
164 mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
165 "End of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
166 *start_pfn, *end_pfn, max_sparsemem_pfn);
168 *end_pfn = max_sparsemem_pfn;
173 * There are a number of times that we loop over NR_MEM_SECTIONS,
174 * looking for section_present() on each. But, when we have very
175 * large physical address spaces, NR_MEM_SECTIONS can also be
176 * very large which makes the loops quite long.
178 * Keeping track of this gives us an easy way to break out of
181 int __highest_present_section_nr;
182 static void section_mark_present(struct mem_section *ms)
184 int section_nr = __section_nr(ms);
186 if (section_nr > __highest_present_section_nr)
187 __highest_present_section_nr = section_nr;
189 ms->section_mem_map |= SECTION_MARKED_PRESENT;
192 static inline int next_present_section_nr(int section_nr)
196 if (present_section_nr(section_nr))
198 } while ((section_nr <= __highest_present_section_nr));
202 #define for_each_present_section_nr(start, section_nr) \
203 for (section_nr = next_present_section_nr(start-1); \
204 ((section_nr != -1) && \
205 (section_nr <= __highest_present_section_nr)); \
206 section_nr = next_present_section_nr(section_nr))
208 static inline unsigned long first_present_section_nr(void)
210 return next_present_section_nr(-1);
213 /* Record a memory area against a node. */
214 void __init memory_present(int nid, unsigned long start, unsigned long end)
218 #ifdef CONFIG_SPARSEMEM_EXTREME
219 if (unlikely(!mem_section)) {
220 unsigned long size, align;
222 size = sizeof(struct mem_section*) * NR_SECTION_ROOTS;
223 align = 1 << (INTERNODE_CACHE_SHIFT);
224 mem_section = memblock_alloc(size, align);
226 panic("%s: Failed to allocate %lu bytes align=0x%lx\n",
227 __func__, size, align);
231 start &= PAGE_SECTION_MASK;
232 mminit_validate_memmodel_limits(&start, &end);
233 for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) {
234 unsigned long section = pfn_to_section_nr(pfn);
235 struct mem_section *ms;
237 sparse_index_init(section, nid);
238 set_section_nid(section, nid);
240 ms = __nr_to_section(section);
241 if (!ms->section_mem_map) {
242 ms->section_mem_map = sparse_encode_early_nid(nid) |
244 section_mark_present(ms);
250 * Mark all memblocks as present using memory_present(). This is a
251 * convienence function that is useful for a number of arches
252 * to mark all of the systems memory as present during initialization.
254 void __init memblocks_present(void)
256 struct memblock_region *reg;
258 for_each_memblock(memory, reg) {
259 memory_present(memblock_get_region_node(reg),
260 memblock_region_memory_base_pfn(reg),
261 memblock_region_memory_end_pfn(reg));
266 * Subtle, we encode the real pfn into the mem_map such that
267 * the identity pfn - section_mem_map will return the actual
268 * physical page frame number.
270 static unsigned long sparse_encode_mem_map(struct page *mem_map, unsigned long pnum)
272 unsigned long coded_mem_map =
273 (unsigned long)(mem_map - (section_nr_to_pfn(pnum)));
274 BUILD_BUG_ON(SECTION_MAP_LAST_BIT > (1UL<<PFN_SECTION_SHIFT));
275 BUG_ON(coded_mem_map & ~SECTION_MAP_MASK);
276 return coded_mem_map;
280 * Decode mem_map from the coded memmap
282 struct page *sparse_decode_mem_map(unsigned long coded_mem_map, unsigned long pnum)
284 /* mask off the extra low bits of information */
285 coded_mem_map &= SECTION_MAP_MASK;
286 return ((struct page *)coded_mem_map) + section_nr_to_pfn(pnum);
289 static void __meminit sparse_init_one_section(struct mem_section *ms,
290 unsigned long pnum, struct page *mem_map,
291 unsigned long *pageblock_bitmap)
293 ms->section_mem_map &= ~SECTION_MAP_MASK;
294 ms->section_mem_map |= sparse_encode_mem_map(mem_map, pnum) |
296 ms->pageblock_flags = pageblock_bitmap;
299 unsigned long usemap_size(void)
301 return BITS_TO_LONGS(SECTION_BLOCKFLAGS_BITS) * sizeof(unsigned long);
304 #ifdef CONFIG_MEMORY_HOTPLUG
305 static unsigned long *__kmalloc_section_usemap(void)
307 return kmalloc(usemap_size(), GFP_KERNEL);
309 #endif /* CONFIG_MEMORY_HOTPLUG */
311 #ifdef CONFIG_MEMORY_HOTREMOVE
312 static unsigned long * __init
313 sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
316 unsigned long goal, limit;
320 * A page may contain usemaps for other sections preventing the
321 * page being freed and making a section unremovable while
322 * other sections referencing the usemap remain active. Similarly,
323 * a pgdat can prevent a section being removed. If section A
324 * contains a pgdat and section B contains the usemap, both
325 * sections become inter-dependent. This allocates usemaps
326 * from the same section as the pgdat where possible to avoid
329 goal = __pa(pgdat) & (PAGE_SECTION_MASK << PAGE_SHIFT);
330 limit = goal + (1UL << PA_SECTION_SHIFT);
331 nid = early_pfn_to_nid(goal >> PAGE_SHIFT);
333 p = memblock_alloc_try_nid_nopanic(size,
334 SMP_CACHE_BYTES, goal, limit,
343 static void __init check_usemap_section_nr(int nid, unsigned long *usemap)
345 unsigned long usemap_snr, pgdat_snr;
346 static unsigned long old_usemap_snr;
347 static unsigned long old_pgdat_snr;
348 struct pglist_data *pgdat = NODE_DATA(nid);
352 if (!old_usemap_snr) {
353 old_usemap_snr = NR_MEM_SECTIONS;
354 old_pgdat_snr = NR_MEM_SECTIONS;
357 usemap_snr = pfn_to_section_nr(__pa(usemap) >> PAGE_SHIFT);
358 pgdat_snr = pfn_to_section_nr(__pa(pgdat) >> PAGE_SHIFT);
359 if (usemap_snr == pgdat_snr)
362 if (old_usemap_snr == usemap_snr && old_pgdat_snr == pgdat_snr)
363 /* skip redundant message */
366 old_usemap_snr = usemap_snr;
367 old_pgdat_snr = pgdat_snr;
369 usemap_nid = sparse_early_nid(__nr_to_section(usemap_snr));
370 if (usemap_nid != nid) {
371 pr_info("node %d must be removed before remove section %ld\n",
376 * There is a circular dependency.
377 * Some platforms allow un-removable section because they will just
378 * gather other removable sections for dynamic partitioning.
379 * Just notify un-removable section's number here.
381 pr_info("Section %ld and %ld (node %d) have a circular dependency on usemap and pgdat allocations\n",
382 usemap_snr, pgdat_snr, nid);
385 static unsigned long * __init
386 sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
389 return memblock_alloc_node_nopanic(size, pgdat->node_id);
392 static void __init check_usemap_section_nr(int nid, unsigned long *usemap)
395 #endif /* CONFIG_MEMORY_HOTREMOVE */
397 #ifdef CONFIG_SPARSEMEM_VMEMMAP
398 static unsigned long __init section_map_size(void)
400 return ALIGN(sizeof(struct page) * PAGES_PER_SECTION, PMD_SIZE);
404 static unsigned long __init section_map_size(void)
406 return PAGE_ALIGN(sizeof(struct page) * PAGES_PER_SECTION);
409 struct page __init *sparse_mem_map_populate(unsigned long pnum, int nid,
410 struct vmem_altmap *altmap)
412 unsigned long size = section_map_size();
413 struct page *map = sparse_buffer_alloc(size);
414 phys_addr_t addr = __pa(MAX_DMA_ADDRESS);
419 map = memblock_alloc_try_nid(size,
421 MEMBLOCK_ALLOC_ACCESSIBLE, nid);
423 panic("%s: Failed to allocate %lu bytes align=0x%lx nid=%d from=%pa\n",
424 __func__, size, PAGE_SIZE, nid, &addr);
428 #endif /* !CONFIG_SPARSEMEM_VMEMMAP */
430 static void *sparsemap_buf __meminitdata;
431 static void *sparsemap_buf_end __meminitdata;
433 static void __init sparse_buffer_init(unsigned long size, int nid)
435 phys_addr_t addr = __pa(MAX_DMA_ADDRESS);
436 WARN_ON(sparsemap_buf); /* forgot to call sparse_buffer_fini()? */
438 memblock_alloc_try_nid_raw(size, PAGE_SIZE,
440 MEMBLOCK_ALLOC_ACCESSIBLE, nid);
441 sparsemap_buf_end = sparsemap_buf + size;
444 static void __init sparse_buffer_fini(void)
446 unsigned long size = sparsemap_buf_end - sparsemap_buf;
448 if (sparsemap_buf && size > 0)
449 memblock_free_early(__pa(sparsemap_buf), size);
450 sparsemap_buf = NULL;
453 void * __meminit sparse_buffer_alloc(unsigned long size)
458 ptr = PTR_ALIGN(sparsemap_buf, size);
459 if (ptr + size > sparsemap_buf_end)
462 sparsemap_buf = ptr + size;
467 void __weak __meminit vmemmap_populate_print_last(void)
472 * Initialize sparse on a specific node. The node spans [pnum_begin, pnum_end)
473 * And number of present sections in this node is map_count.
475 static void __init sparse_init_nid(int nid, unsigned long pnum_begin,
476 unsigned long pnum_end,
477 unsigned long map_count)
479 unsigned long pnum, usemap_longs, *usemap;
482 usemap_longs = BITS_TO_LONGS(SECTION_BLOCKFLAGS_BITS);
483 usemap = sparse_early_usemaps_alloc_pgdat_section(NODE_DATA(nid),
487 pr_err("%s: node[%d] usemap allocation failed", __func__, nid);
490 sparse_buffer_init(map_count * section_map_size(), nid);
491 for_each_present_section_nr(pnum_begin, pnum) {
492 if (pnum >= pnum_end)
495 map = sparse_mem_map_populate(pnum, nid, NULL);
497 pr_err("%s: node[%d] memory map backing failed. Some memory will not be available.",
502 check_usemap_section_nr(nid, usemap);
503 sparse_init_one_section(__nr_to_section(pnum), pnum, map, usemap);
504 usemap += usemap_longs;
506 sparse_buffer_fini();
509 /* We failed to allocate, mark all the following pnums as not present */
510 for_each_present_section_nr(pnum_begin, pnum) {
511 struct mem_section *ms;
513 if (pnum >= pnum_end)
515 ms = __nr_to_section(pnum);
516 ms->section_mem_map = 0;
521 * Allocate the accumulated non-linear sections, allocate a mem_map
522 * for each and record the physical to section mapping.
524 void __init sparse_init(void)
526 unsigned long pnum_begin = first_present_section_nr();
527 int nid_begin = sparse_early_nid(__nr_to_section(pnum_begin));
528 unsigned long pnum_end, map_count = 1;
530 /* Setup pageblock_order for HUGETLB_PAGE_SIZE_VARIABLE */
531 set_pageblock_order();
533 for_each_present_section_nr(pnum_begin + 1, pnum_end) {
534 int nid = sparse_early_nid(__nr_to_section(pnum_end));
536 if (nid == nid_begin) {
540 /* Init node with sections in range [pnum_begin, pnum_end) */
541 sparse_init_nid(nid_begin, pnum_begin, pnum_end, map_count);
543 pnum_begin = pnum_end;
546 /* cover the last node */
547 sparse_init_nid(nid_begin, pnum_begin, pnum_end, map_count);
548 vmemmap_populate_print_last();
551 #ifdef CONFIG_MEMORY_HOTPLUG
553 /* Mark all memory sections within the pfn range as online */
554 void online_mem_sections(unsigned long start_pfn, unsigned long end_pfn)
558 for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
559 unsigned long section_nr = pfn_to_section_nr(pfn);
560 struct mem_section *ms;
562 /* onlining code should never touch invalid ranges */
563 if (WARN_ON(!valid_section_nr(section_nr)))
566 ms = __nr_to_section(section_nr);
567 ms->section_mem_map |= SECTION_IS_ONLINE;
571 #ifdef CONFIG_MEMORY_HOTREMOVE
572 /* Mark all memory sections within the pfn range as online */
573 void offline_mem_sections(unsigned long start_pfn, unsigned long end_pfn)
577 for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
578 unsigned long section_nr = pfn_to_section_nr(pfn);
579 struct mem_section *ms;
582 * TODO this needs some double checking. Offlining code makes
583 * sure to check pfn_valid but those checks might be just bogus
585 if (WARN_ON(!valid_section_nr(section_nr)))
588 ms = __nr_to_section(section_nr);
589 ms->section_mem_map &= ~SECTION_IS_ONLINE;
594 #ifdef CONFIG_SPARSEMEM_VMEMMAP
595 static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid,
596 struct vmem_altmap *altmap)
598 /* This will make the necessary allocations eventually. */
599 return sparse_mem_map_populate(pnum, nid, altmap);
601 static void __kfree_section_memmap(struct page *memmap,
602 struct vmem_altmap *altmap)
604 unsigned long start = (unsigned long)memmap;
605 unsigned long end = (unsigned long)(memmap + PAGES_PER_SECTION);
607 vmemmap_free(start, end, altmap);
609 #ifdef CONFIG_MEMORY_HOTREMOVE
610 static void free_map_bootmem(struct page *memmap)
612 unsigned long start = (unsigned long)memmap;
613 unsigned long end = (unsigned long)(memmap + PAGES_PER_SECTION);
615 vmemmap_free(start, end, NULL);
617 #endif /* CONFIG_MEMORY_HOTREMOVE */
619 static struct page *__kmalloc_section_memmap(void)
621 struct page *page, *ret;
622 unsigned long memmap_size = sizeof(struct page) * PAGES_PER_SECTION;
624 page = alloc_pages(GFP_KERNEL|__GFP_NOWARN, get_order(memmap_size));
628 ret = vmalloc(memmap_size);
634 ret = (struct page *)pfn_to_kaddr(page_to_pfn(page));
640 static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid,
641 struct vmem_altmap *altmap)
643 return __kmalloc_section_memmap();
646 static void __kfree_section_memmap(struct page *memmap,
647 struct vmem_altmap *altmap)
649 if (is_vmalloc_addr(memmap))
652 free_pages((unsigned long)memmap,
653 get_order(sizeof(struct page) * PAGES_PER_SECTION));
656 #ifdef CONFIG_MEMORY_HOTREMOVE
657 static void free_map_bootmem(struct page *memmap)
659 unsigned long maps_section_nr, removing_section_nr, i;
660 unsigned long magic, nr_pages;
661 struct page *page = virt_to_page(memmap);
663 nr_pages = PAGE_ALIGN(PAGES_PER_SECTION * sizeof(struct page))
666 for (i = 0; i < nr_pages; i++, page++) {
667 magic = (unsigned long) page->freelist;
669 BUG_ON(magic == NODE_INFO);
671 maps_section_nr = pfn_to_section_nr(page_to_pfn(page));
672 removing_section_nr = page_private(page);
675 * When this function is called, the removing section is
676 * logical offlined state. This means all pages are isolated
677 * from page allocator. If removing section's memmap is placed
678 * on the same section, it must not be freed.
679 * If it is freed, page allocator may allocate it which will
680 * be removed physically soon.
682 if (maps_section_nr != removing_section_nr)
683 put_page_bootmem(page);
686 #endif /* CONFIG_MEMORY_HOTREMOVE */
687 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
690 * returns the number of sections whose mem_maps were properly
691 * set. If this is <=0, then that means that the passed-in
692 * map was not consumed and must be freed.
694 int __meminit sparse_add_one_section(int nid, unsigned long start_pfn,
695 struct vmem_altmap *altmap)
697 unsigned long section_nr = pfn_to_section_nr(start_pfn);
698 struct mem_section *ms;
700 unsigned long *usemap;
704 * no locking for this, because it does its own
705 * plus, it does a kmalloc
707 ret = sparse_index_init(section_nr, nid);
708 if (ret < 0 && ret != -EEXIST)
711 memmap = kmalloc_section_memmap(section_nr, nid, altmap);
714 usemap = __kmalloc_section_usemap();
716 __kfree_section_memmap(memmap, altmap);
720 ms = __pfn_to_section(start_pfn);
721 if (ms->section_mem_map & SECTION_MARKED_PRESENT) {
727 * Poison uninitialized struct pages in order to catch invalid flags
730 page_init_poison(memmap, sizeof(struct page) * PAGES_PER_SECTION);
732 section_mark_present(ms);
733 sparse_init_one_section(ms, section_nr, memmap, usemap);
738 __kfree_section_memmap(memmap, altmap);
743 #ifdef CONFIG_MEMORY_HOTREMOVE
744 #ifdef CONFIG_MEMORY_FAILURE
745 static void clear_hwpoisoned_pages(struct page *memmap, int nr_pages)
753 * A further optimization is to have per section refcounted
754 * num_poisoned_pages. But that would need more space per memmap, so
755 * for now just do a quick global check to speed up this routine in the
756 * absence of bad pages.
758 if (atomic_long_read(&num_poisoned_pages) == 0)
761 for (i = 0; i < nr_pages; i++) {
762 if (PageHWPoison(&memmap[i])) {
763 atomic_long_sub(1, &num_poisoned_pages);
764 ClearPageHWPoison(&memmap[i]);
769 static inline void clear_hwpoisoned_pages(struct page *memmap, int nr_pages)
774 static void free_section_usemap(struct page *memmap, unsigned long *usemap,
775 struct vmem_altmap *altmap)
777 struct page *usemap_page;
782 usemap_page = virt_to_page(usemap);
784 * Check to see if allocation came from hot-plug-add
786 if (PageSlab(usemap_page) || PageCompound(usemap_page)) {
789 __kfree_section_memmap(memmap, altmap);
794 * The usemap came from bootmem. This is packed with other usemaps
795 * on the section which has pgdat at boot time. Just keep it as is now.
799 free_map_bootmem(memmap);
802 void sparse_remove_one_section(struct zone *zone, struct mem_section *ms,
803 unsigned long map_offset, struct vmem_altmap *altmap)
805 struct page *memmap = NULL;
806 unsigned long *usemap = NULL;
808 if (ms->section_mem_map) {
809 usemap = ms->pageblock_flags;
810 memmap = sparse_decode_mem_map(ms->section_mem_map,
812 ms->section_mem_map = 0;
813 ms->pageblock_flags = NULL;
816 clear_hwpoisoned_pages(memmap + map_offset,
817 PAGES_PER_SECTION - map_offset);
818 free_section_usemap(memmap, usemap, altmap);
820 #endif /* CONFIG_MEMORY_HOTREMOVE */
821 #endif /* CONFIG_MEMORY_HOTPLUG */