2 * sparse memory mappings.
5 #include <linux/mmzone.h>
6 #include <linux/bootmem.h>
7 #include <linux/highmem.h>
8 #include <linux/module.h>
9 #include <linux/spinlock.h>
10 #include <linux/vmalloc.h>
13 #include <asm/pgalloc.h>
14 #include <asm/pgtable.h>
17 * Permanent SPARSEMEM data:
19 * 1) mem_section - memory sections, mem_map's for valid memory
21 #ifdef CONFIG_SPARSEMEM_EXTREME
22 struct mem_section *mem_section[NR_SECTION_ROOTS]
23 ____cacheline_internodealigned_in_smp;
25 struct mem_section mem_section[NR_SECTION_ROOTS][SECTIONS_PER_ROOT]
26 ____cacheline_internodealigned_in_smp;
28 EXPORT_SYMBOL(mem_section);
30 #ifdef NODE_NOT_IN_PAGE_FLAGS
32 * If we did not store the node number in the page then we have to
33 * do a lookup in the section_to_node_table in order to find which
34 * node the page belongs to.
36 #if MAX_NUMNODES <= 256
37 static u8 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned;
39 static u16 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned;
42 int page_to_nid(struct page *page)
44 return section_to_node_table[page_to_section(page)];
46 EXPORT_SYMBOL(page_to_nid);
48 static void set_section_nid(unsigned long section_nr, int nid)
50 section_to_node_table[section_nr] = nid;
52 #else /* !NODE_NOT_IN_PAGE_FLAGS */
53 static inline void set_section_nid(unsigned long section_nr, int nid)
58 #ifdef CONFIG_SPARSEMEM_EXTREME
59 static struct mem_section noinline __init_refok *sparse_index_alloc(int nid)
61 struct mem_section *section = NULL;
62 unsigned long array_size = SECTIONS_PER_ROOT *
63 sizeof(struct mem_section);
65 if (slab_is_available())
66 section = kmalloc_node(array_size, GFP_KERNEL, nid);
68 section = alloc_bootmem_node(NODE_DATA(nid), array_size);
71 memset(section, 0, array_size);
76 static int __meminit sparse_index_init(unsigned long section_nr, int nid)
78 static DEFINE_SPINLOCK(index_init_lock);
79 unsigned long root = SECTION_NR_TO_ROOT(section_nr);
80 struct mem_section *section;
83 if (mem_section[root])
86 section = sparse_index_alloc(nid);
90 * This lock keeps two different sections from
91 * reallocating for the same index
93 spin_lock(&index_init_lock);
95 if (mem_section[root]) {
100 mem_section[root] = section;
102 spin_unlock(&index_init_lock);
105 #else /* !SPARSEMEM_EXTREME */
106 static inline int sparse_index_init(unsigned long section_nr, int nid)
113 * Although written for the SPARSEMEM_EXTREME case, this happens
114 * to also work for the flat array case because
115 * NR_SECTION_ROOTS==NR_MEM_SECTIONS.
117 int __section_nr(struct mem_section* ms)
119 unsigned long root_nr;
120 struct mem_section* root;
122 for (root_nr = 0; root_nr < NR_SECTION_ROOTS; root_nr++) {
123 root = __nr_to_section(root_nr * SECTIONS_PER_ROOT);
127 if ((ms >= root) && (ms < (root + SECTIONS_PER_ROOT)))
131 return (root_nr * SECTIONS_PER_ROOT) + (ms - root);
135 * During early boot, before section_mem_map is used for an actual
136 * mem_map, we use section_mem_map to store the section's NUMA
137 * node. This keeps us from having to use another data structure. The
138 * node information is cleared just before we store the real mem_map.
140 static inline unsigned long sparse_encode_early_nid(int nid)
142 return (nid << SECTION_NID_SHIFT);
145 static inline int sparse_early_nid(struct mem_section *section)
147 return (section->section_mem_map >> SECTION_NID_SHIFT);
150 /* Validate the physical addressing limitations of the model */
151 void __meminit mminit_validate_memmodel_limits(unsigned long *start_pfn,
152 unsigned long *end_pfn)
154 unsigned long max_sparsemem_pfn = 1UL << (MAX_PHYSMEM_BITS-PAGE_SHIFT);
157 * Sanity checks - do not allow an architecture to pass
158 * in larger pfns than the maximum scope of sparsemem:
160 if (*start_pfn > max_sparsemem_pfn) {
161 mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
162 "Start of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
163 *start_pfn, *end_pfn, max_sparsemem_pfn);
165 *start_pfn = max_sparsemem_pfn;
166 *end_pfn = max_sparsemem_pfn;
169 if (*end_pfn > max_sparsemem_pfn) {
170 mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
171 "End of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
172 *start_pfn, *end_pfn, max_sparsemem_pfn);
174 *end_pfn = max_sparsemem_pfn;
178 /* Record a memory area against a node. */
179 void __init memory_present(int nid, unsigned long start, unsigned long end)
183 start &= PAGE_SECTION_MASK;
184 mminit_validate_memmodel_limits(&start, &end);
185 for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) {
186 unsigned long section = pfn_to_section_nr(pfn);
187 struct mem_section *ms;
189 sparse_index_init(section, nid);
190 set_section_nid(section, nid);
192 ms = __nr_to_section(section);
193 if (!ms->section_mem_map)
194 ms->section_mem_map = sparse_encode_early_nid(nid) |
195 SECTION_MARKED_PRESENT;
200 * Only used by the i386 NUMA architecures, but relatively
203 unsigned long __init node_memmap_size_bytes(int nid, unsigned long start_pfn,
204 unsigned long end_pfn)
207 unsigned long nr_pages = 0;
209 mminit_validate_memmodel_limits(&start_pfn, &end_pfn);
210 for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
211 if (nid != early_pfn_to_nid(pfn))
214 if (pfn_present(pfn))
215 nr_pages += PAGES_PER_SECTION;
218 return nr_pages * sizeof(struct page);
222 * Subtle, we encode the real pfn into the mem_map such that
223 * the identity pfn - section_mem_map will return the actual
224 * physical page frame number.
226 static unsigned long sparse_encode_mem_map(struct page *mem_map, unsigned long pnum)
228 return (unsigned long)(mem_map - (section_nr_to_pfn(pnum)));
232 * Decode mem_map from the coded memmap
234 struct page *sparse_decode_mem_map(unsigned long coded_mem_map, unsigned long pnum)
236 /* mask off the extra low bits of information */
237 coded_mem_map &= SECTION_MAP_MASK;
238 return ((struct page *)coded_mem_map) + section_nr_to_pfn(pnum);
241 static int __meminit sparse_init_one_section(struct mem_section *ms,
242 unsigned long pnum, struct page *mem_map,
243 unsigned long *pageblock_bitmap)
245 if (!present_section(ms))
248 ms->section_mem_map &= ~SECTION_MAP_MASK;
249 ms->section_mem_map |= sparse_encode_mem_map(mem_map, pnum) |
251 ms->pageblock_flags = pageblock_bitmap;
256 unsigned long usemap_size(void)
258 unsigned long size_bytes;
259 size_bytes = roundup(SECTION_BLOCKFLAGS_BITS, 8) / 8;
260 size_bytes = roundup(size_bytes, sizeof(unsigned long));
264 #ifdef CONFIG_MEMORY_HOTPLUG
265 static unsigned long *__kmalloc_section_usemap(void)
267 return kmalloc(usemap_size(), GFP_KERNEL);
269 #endif /* CONFIG_MEMORY_HOTPLUG */
271 #ifdef CONFIG_MEMORY_HOTREMOVE
272 static unsigned long * __init
273 sparse_early_usemap_alloc_pgdat_section(struct pglist_data *pgdat)
275 unsigned long section_nr;
278 * A page may contain usemaps for other sections preventing the
279 * page being freed and making a section unremovable while
280 * other sections referencing the usemap retmain active. Similarly,
281 * a pgdat can prevent a section being removed. If section A
282 * contains a pgdat and section B contains the usemap, both
283 * sections become inter-dependent. This allocates usemaps
284 * from the same section as the pgdat where possible to avoid
287 section_nr = pfn_to_section_nr(__pa(pgdat) >> PAGE_SHIFT);
288 return alloc_bootmem_section(usemap_size(), section_nr);
291 static void __init check_usemap_section_nr(int nid, unsigned long *usemap)
293 unsigned long usemap_snr, pgdat_snr;
294 static unsigned long old_usemap_snr = NR_MEM_SECTIONS;
295 static unsigned long old_pgdat_snr = NR_MEM_SECTIONS;
296 struct pglist_data *pgdat = NODE_DATA(nid);
299 usemap_snr = pfn_to_section_nr(__pa(usemap) >> PAGE_SHIFT);
300 pgdat_snr = pfn_to_section_nr(__pa(pgdat) >> PAGE_SHIFT);
301 if (usemap_snr == pgdat_snr)
304 if (old_usemap_snr == usemap_snr && old_pgdat_snr == pgdat_snr)
305 /* skip redundant message */
308 old_usemap_snr = usemap_snr;
309 old_pgdat_snr = pgdat_snr;
311 usemap_nid = sparse_early_nid(__nr_to_section(usemap_snr));
312 if (usemap_nid != nid) {
314 "node %d must be removed before remove section %ld\n",
319 * There is a circular dependency.
320 * Some platforms allow un-removable section because they will just
321 * gather other removable sections for dynamic partitioning.
322 * Just notify un-removable section's number here.
324 printk(KERN_INFO "Section %ld and %ld (node %d)", usemap_snr,
327 " have a circular dependency on usemap and pgdat allocations\n");
330 static unsigned long * __init
331 sparse_early_usemap_alloc_pgdat_section(struct pglist_data *pgdat)
336 static void __init check_usemap_section_nr(int nid, unsigned long *usemap)
339 #endif /* CONFIG_MEMORY_HOTREMOVE */
341 static unsigned long *__init sparse_early_usemap_alloc(unsigned long pnum)
343 unsigned long *usemap;
344 struct mem_section *ms = __nr_to_section(pnum);
345 int nid = sparse_early_nid(ms);
347 usemap = sparse_early_usemap_alloc_pgdat_section(NODE_DATA(nid));
351 usemap = alloc_bootmem_node(NODE_DATA(nid), usemap_size());
353 check_usemap_section_nr(nid, usemap);
357 /* Stupid: suppress gcc warning for SPARSEMEM && !NUMA */
360 printk(KERN_WARNING "%s: allocation failed\n", __func__);
364 #ifndef CONFIG_SPARSEMEM_VMEMMAP
365 struct page __init *sparse_mem_map_populate(unsigned long pnum, int nid)
369 map = alloc_remap(nid, sizeof(struct page) * PAGES_PER_SECTION);
373 map = alloc_bootmem_pages_node(NODE_DATA(nid),
374 PAGE_ALIGN(sizeof(struct page) * PAGES_PER_SECTION));
377 #endif /* !CONFIG_SPARSEMEM_VMEMMAP */
379 static struct page __init *sparse_early_mem_map_alloc(unsigned long pnum)
382 struct mem_section *ms = __nr_to_section(pnum);
383 int nid = sparse_early_nid(ms);
385 map = sparse_mem_map_populate(pnum, nid);
389 printk(KERN_ERR "%s: sparsemem memory map backing failed "
390 "some memory will not be available.\n", __func__);
391 ms->section_mem_map = 0;
395 void __attribute__((weak)) __meminit vmemmap_populate_print_last(void)
399 * Allocate the accumulated non-linear sections, allocate a mem_map
400 * for each and record the physical to section mapping.
402 void __init sparse_init(void)
406 unsigned long *usemap;
407 unsigned long **usemap_map;
411 * map is using big page (aka 2M in x86 64 bit)
412 * usemap is less one page (aka 24 bytes)
413 * so alloc 2M (with 2M align) and 24 bytes in turn will
414 * make next 2M slip to one more 2M later.
415 * then in big system, the memory will have a lot of holes...
416 * here try to allocate 2M pages continously.
418 * powerpc need to call sparse_init_one_section right after each
419 * sparse_early_mem_map_alloc, so allocate usemap_map at first.
421 size = sizeof(unsigned long *) * NR_MEM_SECTIONS;
422 usemap_map = alloc_bootmem(size);
424 panic("can not allocate usemap_map\n");
426 for (pnum = 0; pnum < NR_MEM_SECTIONS; pnum++) {
427 if (!present_section_nr(pnum))
429 usemap_map[pnum] = sparse_early_usemap_alloc(pnum);
432 for (pnum = 0; pnum < NR_MEM_SECTIONS; pnum++) {
433 if (!present_section_nr(pnum))
436 usemap = usemap_map[pnum];
440 map = sparse_early_mem_map_alloc(pnum);
444 sparse_init_one_section(__nr_to_section(pnum), pnum, map,
448 vmemmap_populate_print_last();
450 free_bootmem(__pa(usemap_map), size);
453 #ifdef CONFIG_MEMORY_HOTPLUG
454 #ifdef CONFIG_SPARSEMEM_VMEMMAP
455 static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid,
456 unsigned long nr_pages)
458 /* This will make the necessary allocations eventually. */
459 return sparse_mem_map_populate(pnum, nid);
461 static void __kfree_section_memmap(struct page *memmap, unsigned long nr_pages)
463 return; /* XXX: Not implemented yet */
465 static void free_map_bootmem(struct page *page, unsigned long nr_pages)
469 static struct page *__kmalloc_section_memmap(unsigned long nr_pages)
471 struct page *page, *ret;
472 unsigned long memmap_size = sizeof(struct page) * nr_pages;
474 page = alloc_pages(GFP_KERNEL|__GFP_NOWARN, get_order(memmap_size));
478 ret = vmalloc(memmap_size);
484 ret = (struct page *)pfn_to_kaddr(page_to_pfn(page));
486 memset(ret, 0, memmap_size);
491 static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid,
492 unsigned long nr_pages)
494 return __kmalloc_section_memmap(nr_pages);
497 static void __kfree_section_memmap(struct page *memmap, unsigned long nr_pages)
499 if (is_vmalloc_addr(memmap))
502 free_pages((unsigned long)memmap,
503 get_order(sizeof(struct page) * nr_pages));
506 static void free_map_bootmem(struct page *page, unsigned long nr_pages)
508 unsigned long maps_section_nr, removing_section_nr, i;
511 for (i = 0; i < nr_pages; i++, page++) {
512 magic = atomic_read(&page->_mapcount);
514 BUG_ON(magic == NODE_INFO);
516 maps_section_nr = pfn_to_section_nr(page_to_pfn(page));
517 removing_section_nr = page->private;
520 * When this function is called, the removing section is
521 * logical offlined state. This means all pages are isolated
522 * from page allocator. If removing section's memmap is placed
523 * on the same section, it must not be freed.
524 * If it is freed, page allocator may allocate it which will
525 * be removed physically soon.
527 if (maps_section_nr != removing_section_nr)
528 put_page_bootmem(page);
531 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
533 static void free_section_usemap(struct page *memmap, unsigned long *usemap)
535 struct page *usemap_page;
536 unsigned long nr_pages;
541 usemap_page = virt_to_page(usemap);
543 * Check to see if allocation came from hot-plug-add
545 if (PageSlab(usemap_page)) {
548 __kfree_section_memmap(memmap, PAGES_PER_SECTION);
553 * The usemap came from bootmem. This is packed with other usemaps
554 * on the section which has pgdat at boot time. Just keep it as is now.
558 struct page *memmap_page;
559 memmap_page = virt_to_page(memmap);
561 nr_pages = PAGE_ALIGN(PAGES_PER_SECTION * sizeof(struct page))
564 free_map_bootmem(memmap_page, nr_pages);
569 * returns the number of sections whose mem_maps were properly
570 * set. If this is <=0, then that means that the passed-in
571 * map was not consumed and must be freed.
573 int sparse_add_one_section(struct zone *zone, unsigned long start_pfn,
576 unsigned long section_nr = pfn_to_section_nr(start_pfn);
577 struct pglist_data *pgdat = zone->zone_pgdat;
578 struct mem_section *ms;
580 unsigned long *usemap;
585 * no locking for this, because it does its own
586 * plus, it does a kmalloc
588 ret = sparse_index_init(section_nr, pgdat->node_id);
589 if (ret < 0 && ret != -EEXIST)
591 memmap = kmalloc_section_memmap(section_nr, pgdat->node_id, nr_pages);
594 usemap = __kmalloc_section_usemap();
596 __kfree_section_memmap(memmap, nr_pages);
600 pgdat_resize_lock(pgdat, &flags);
602 ms = __pfn_to_section(start_pfn);
603 if (ms->section_mem_map & SECTION_MARKED_PRESENT) {
608 ms->section_mem_map |= SECTION_MARKED_PRESENT;
610 ret = sparse_init_one_section(ms, section_nr, memmap, usemap);
613 pgdat_resize_unlock(pgdat, &flags);
616 __kfree_section_memmap(memmap, nr_pages);
621 void sparse_remove_one_section(struct zone *zone, struct mem_section *ms)
623 struct page *memmap = NULL;
624 unsigned long *usemap = NULL;
626 if (ms->section_mem_map) {
627 usemap = ms->pageblock_flags;
628 memmap = sparse_decode_mem_map(ms->section_mem_map,
630 ms->section_mem_map = 0;
631 ms->pageblock_flags = NULL;
634 free_section_usemap(memmap, usemap);