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>
18 * Permanent SPARSEMEM data:
20 * 1) mem_section - memory sections, mem_map's for valid memory
22 #ifdef CONFIG_SPARSEMEM_EXTREME
23 struct mem_section *mem_section[NR_SECTION_ROOTS]
24 ____cacheline_internodealigned_in_smp;
26 struct mem_section mem_section[NR_SECTION_ROOTS][SECTIONS_PER_ROOT]
27 ____cacheline_internodealigned_in_smp;
29 EXPORT_SYMBOL(mem_section);
31 #ifdef NODE_NOT_IN_PAGE_FLAGS
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.
37 #if MAX_NUMNODES <= 256
38 static u8 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned;
40 static u16 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned;
43 int page_to_nid(struct page *page)
45 return section_to_node_table[page_to_section(page)];
47 EXPORT_SYMBOL(page_to_nid);
49 static void set_section_nid(unsigned long section_nr, int nid)
51 section_to_node_table[section_nr] = nid;
53 #else /* !NODE_NOT_IN_PAGE_FLAGS */
54 static inline void set_section_nid(unsigned long section_nr, int nid)
59 #ifdef CONFIG_SPARSEMEM_EXTREME
60 static struct mem_section noinline __init_refok *sparse_index_alloc(int nid)
62 struct mem_section *section = NULL;
63 unsigned long array_size = SECTIONS_PER_ROOT *
64 sizeof(struct mem_section);
66 if (slab_is_available())
67 section = kmalloc_node(array_size, GFP_KERNEL, nid);
69 section = alloc_bootmem_node(NODE_DATA(nid), array_size);
72 memset(section, 0, array_size);
77 static int __meminit sparse_index_init(unsigned long section_nr, int nid)
79 static DEFINE_SPINLOCK(index_init_lock);
80 unsigned long root = SECTION_NR_TO_ROOT(section_nr);
81 struct mem_section *section;
84 if (mem_section[root])
87 section = sparse_index_alloc(nid);
91 * This lock keeps two different sections from
92 * reallocating for the same index
94 spin_lock(&index_init_lock);
96 if (mem_section[root]) {
101 mem_section[root] = section;
103 spin_unlock(&index_init_lock);
106 #else /* !SPARSEMEM_EXTREME */
107 static inline int sparse_index_init(unsigned long section_nr, int nid)
114 * Although written for the SPARSEMEM_EXTREME case, this happens
115 * to also work for the flat array case because
116 * NR_SECTION_ROOTS==NR_MEM_SECTIONS.
118 int __section_nr(struct mem_section* ms)
120 unsigned long root_nr;
121 struct mem_section* root;
123 for (root_nr = 0; root_nr < NR_SECTION_ROOTS; root_nr++) {
124 root = __nr_to_section(root_nr * SECTIONS_PER_ROOT);
128 if ((ms >= root) && (ms < (root + SECTIONS_PER_ROOT)))
132 return (root_nr * SECTIONS_PER_ROOT) + (ms - root);
136 * During early boot, before section_mem_map is used for an actual
137 * mem_map, we use section_mem_map to store the section's NUMA
138 * node. This keeps us from having to use another data structure. The
139 * node information is cleared just before we store the real mem_map.
141 static inline unsigned long sparse_encode_early_nid(int nid)
143 return (nid << SECTION_NID_SHIFT);
146 static inline int sparse_early_nid(struct mem_section *section)
148 return (section->section_mem_map >> SECTION_NID_SHIFT);
151 /* Validate the physical addressing limitations of the model */
152 void __meminit mminit_validate_memmodel_limits(unsigned long *start_pfn,
153 unsigned long *end_pfn)
155 unsigned long max_sparsemem_pfn = 1UL << (MAX_PHYSMEM_BITS-PAGE_SHIFT);
158 * Sanity checks - do not allow an architecture to pass
159 * in larger pfns than the maximum scope of sparsemem:
161 if (*start_pfn > max_sparsemem_pfn) {
162 mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
163 "Start of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
164 *start_pfn, *end_pfn, max_sparsemem_pfn);
166 *start_pfn = max_sparsemem_pfn;
167 *end_pfn = max_sparsemem_pfn;
170 if (*end_pfn > max_sparsemem_pfn) {
171 mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
172 "End of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
173 *start_pfn, *end_pfn, max_sparsemem_pfn);
175 *end_pfn = max_sparsemem_pfn;
179 /* Record a memory area against a node. */
180 void __init memory_present(int nid, unsigned long start, unsigned long end)
184 start &= PAGE_SECTION_MASK;
185 mminit_validate_memmodel_limits(&start, &end);
186 for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) {
187 unsigned long section = pfn_to_section_nr(pfn);
188 struct mem_section *ms;
190 sparse_index_init(section, nid);
191 set_section_nid(section, nid);
193 ms = __nr_to_section(section);
194 if (!ms->section_mem_map)
195 ms->section_mem_map = sparse_encode_early_nid(nid) |
196 SECTION_MARKED_PRESENT;
201 * Only used by the i386 NUMA architecures, but relatively
204 unsigned long __init node_memmap_size_bytes(int nid, unsigned long start_pfn,
205 unsigned long end_pfn)
208 unsigned long nr_pages = 0;
210 mminit_validate_memmodel_limits(&start_pfn, &end_pfn);
211 for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
212 if (nid != early_pfn_to_nid(pfn))
215 if (pfn_present(pfn))
216 nr_pages += PAGES_PER_SECTION;
219 return nr_pages * sizeof(struct page);
223 * Subtle, we encode the real pfn into the mem_map such that
224 * the identity pfn - section_mem_map will return the actual
225 * physical page frame number.
227 static unsigned long sparse_encode_mem_map(struct page *mem_map, unsigned long pnum)
229 return (unsigned long)(mem_map - (section_nr_to_pfn(pnum)));
233 * Decode mem_map from the coded memmap
235 struct page *sparse_decode_mem_map(unsigned long coded_mem_map, unsigned long pnum)
237 /* mask off the extra low bits of information */
238 coded_mem_map &= SECTION_MAP_MASK;
239 return ((struct page *)coded_mem_map) + section_nr_to_pfn(pnum);
242 static int __meminit sparse_init_one_section(struct mem_section *ms,
243 unsigned long pnum, struct page *mem_map,
244 unsigned long *pageblock_bitmap)
246 if (!present_section(ms))
249 ms->section_mem_map &= ~SECTION_MAP_MASK;
250 ms->section_mem_map |= sparse_encode_mem_map(mem_map, pnum) |
252 ms->pageblock_flags = pageblock_bitmap;
257 unsigned long usemap_size(void)
259 unsigned long size_bytes;
260 size_bytes = roundup(SECTION_BLOCKFLAGS_BITS, 8) / 8;
261 size_bytes = roundup(size_bytes, sizeof(unsigned long));
265 #ifdef CONFIG_MEMORY_HOTPLUG
266 static unsigned long *__kmalloc_section_usemap(void)
268 return kmalloc(usemap_size(), GFP_KERNEL);
270 #endif /* CONFIG_MEMORY_HOTPLUG */
272 #ifdef CONFIG_MEMORY_HOTREMOVE
273 static unsigned long * __init
274 sparse_early_usemap_alloc_pgdat_section(struct pglist_data *pgdat)
276 unsigned long section_nr;
279 * A page may contain usemaps for other sections preventing the
280 * page being freed and making a section unremovable while
281 * other sections referencing the usemap retmain active. Similarly,
282 * a pgdat can prevent a section being removed. If section A
283 * contains a pgdat and section B contains the usemap, both
284 * sections become inter-dependent. This allocates usemaps
285 * from the same section as the pgdat where possible to avoid
288 section_nr = pfn_to_section_nr(__pa(pgdat) >> PAGE_SHIFT);
289 return alloc_bootmem_section(usemap_size(), section_nr);
292 static void __init check_usemap_section_nr(int nid, unsigned long *usemap)
294 unsigned long usemap_snr, pgdat_snr;
295 static unsigned long old_usemap_snr = NR_MEM_SECTIONS;
296 static unsigned long old_pgdat_snr = NR_MEM_SECTIONS;
297 struct pglist_data *pgdat = NODE_DATA(nid);
300 usemap_snr = pfn_to_section_nr(__pa(usemap) >> PAGE_SHIFT);
301 pgdat_snr = pfn_to_section_nr(__pa(pgdat) >> PAGE_SHIFT);
302 if (usemap_snr == pgdat_snr)
305 if (old_usemap_snr == usemap_snr && old_pgdat_snr == pgdat_snr)
306 /* skip redundant message */
309 old_usemap_snr = usemap_snr;
310 old_pgdat_snr = pgdat_snr;
312 usemap_nid = sparse_early_nid(__nr_to_section(usemap_snr));
313 if (usemap_nid != nid) {
315 "node %d must be removed before remove section %ld\n",
320 * There is a circular dependency.
321 * Some platforms allow un-removable section because they will just
322 * gather other removable sections for dynamic partitioning.
323 * Just notify un-removable section's number here.
325 printk(KERN_INFO "Section %ld and %ld (node %d)", usemap_snr,
328 " have a circular dependency on usemap and pgdat allocations\n");
331 static unsigned long * __init
332 sparse_early_usemap_alloc_pgdat_section(struct pglist_data *pgdat)
337 static void __init check_usemap_section_nr(int nid, unsigned long *usemap)
340 #endif /* CONFIG_MEMORY_HOTREMOVE */
342 static unsigned long *__init sparse_early_usemap_alloc(unsigned long pnum)
344 unsigned long *usemap;
345 struct mem_section *ms = __nr_to_section(pnum);
346 int nid = sparse_early_nid(ms);
348 usemap = sparse_early_usemap_alloc_pgdat_section(NODE_DATA(nid));
352 usemap = alloc_bootmem_node(NODE_DATA(nid), usemap_size());
354 check_usemap_section_nr(nid, usemap);
358 /* Stupid: suppress gcc warning for SPARSEMEM && !NUMA */
361 printk(KERN_WARNING "%s: allocation failed\n", __func__);
365 #ifndef CONFIG_SPARSEMEM_VMEMMAP
366 struct page __init *sparse_mem_map_populate(unsigned long pnum, int nid)
370 map = alloc_remap(nid, sizeof(struct page) * PAGES_PER_SECTION);
374 map = alloc_bootmem_pages_node(NODE_DATA(nid),
375 PAGE_ALIGN(sizeof(struct page) * PAGES_PER_SECTION));
378 #endif /* !CONFIG_SPARSEMEM_VMEMMAP */
380 struct page __init *sparse_early_mem_map_alloc(unsigned long pnum)
383 struct mem_section *ms = __nr_to_section(pnum);
384 int nid = sparse_early_nid(ms);
386 map = sparse_mem_map_populate(pnum, nid);
390 printk(KERN_ERR "%s: sparsemem memory map backing failed "
391 "some memory will not be available.\n", __func__);
392 ms->section_mem_map = 0;
396 void __attribute__((weak)) __meminit vmemmap_populate_print_last(void)
400 * Allocate the accumulated non-linear sections, allocate a mem_map
401 * for each and record the physical to section mapping.
403 void __init sparse_init(void)
407 unsigned long *usemap;
408 unsigned long **usemap_map;
412 * map is using big page (aka 2M in x86 64 bit)
413 * usemap is less one page (aka 24 bytes)
414 * so alloc 2M (with 2M align) and 24 bytes in turn will
415 * make next 2M slip to one more 2M later.
416 * then in big system, the memory will have a lot of holes...
417 * here try to allocate 2M pages continously.
419 * powerpc need to call sparse_init_one_section right after each
420 * sparse_early_mem_map_alloc, so allocate usemap_map at first.
422 size = sizeof(unsigned long *) * NR_MEM_SECTIONS;
423 usemap_map = alloc_bootmem(size);
425 panic("can not allocate usemap_map\n");
427 for (pnum = 0; pnum < NR_MEM_SECTIONS; pnum++) {
428 if (!present_section_nr(pnum))
430 usemap_map[pnum] = sparse_early_usemap_alloc(pnum);
433 for (pnum = 0; pnum < NR_MEM_SECTIONS; pnum++) {
434 if (!present_section_nr(pnum))
437 usemap = usemap_map[pnum];
441 map = sparse_early_mem_map_alloc(pnum);
445 sparse_init_one_section(__nr_to_section(pnum), pnum, map,
449 vmemmap_populate_print_last();
451 free_bootmem(__pa(usemap_map), size);
454 #ifdef CONFIG_MEMORY_HOTPLUG
455 #ifdef CONFIG_SPARSEMEM_VMEMMAP
456 static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid,
457 unsigned long nr_pages)
459 /* This will make the necessary allocations eventually. */
460 return sparse_mem_map_populate(pnum, nid);
462 static void __kfree_section_memmap(struct page *memmap, unsigned long nr_pages)
464 return; /* XXX: Not implemented yet */
466 static void free_map_bootmem(struct page *page, unsigned long nr_pages)
470 static struct page *__kmalloc_section_memmap(unsigned long nr_pages)
472 struct page *page, *ret;
473 unsigned long memmap_size = sizeof(struct page) * nr_pages;
475 page = alloc_pages(GFP_KERNEL|__GFP_NOWARN, get_order(memmap_size));
479 ret = vmalloc(memmap_size);
485 ret = (struct page *)pfn_to_kaddr(page_to_pfn(page));
487 memset(ret, 0, memmap_size);
492 static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid,
493 unsigned long nr_pages)
495 return __kmalloc_section_memmap(nr_pages);
498 static void __kfree_section_memmap(struct page *memmap, unsigned long nr_pages)
500 if (is_vmalloc_addr(memmap))
503 free_pages((unsigned long)memmap,
504 get_order(sizeof(struct page) * nr_pages));
507 static void free_map_bootmem(struct page *page, unsigned long nr_pages)
509 unsigned long maps_section_nr, removing_section_nr, i;
512 for (i = 0; i < nr_pages; i++, page++) {
513 magic = atomic_read(&page->_mapcount);
515 BUG_ON(magic == NODE_INFO);
517 maps_section_nr = pfn_to_section_nr(page_to_pfn(page));
518 removing_section_nr = page->private;
521 * When this function is called, the removing section is
522 * logical offlined state. This means all pages are isolated
523 * from page allocator. If removing section's memmap is placed
524 * on the same section, it must not be freed.
525 * If it is freed, page allocator may allocate it which will
526 * be removed physically soon.
528 if (maps_section_nr != removing_section_nr)
529 put_page_bootmem(page);
532 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
534 static void free_section_usemap(struct page *memmap, unsigned long *usemap)
536 struct page *usemap_page;
537 unsigned long nr_pages;
542 usemap_page = virt_to_page(usemap);
544 * Check to see if allocation came from hot-plug-add
546 if (PageSlab(usemap_page)) {
549 __kfree_section_memmap(memmap, PAGES_PER_SECTION);
554 * The usemap came from bootmem. This is packed with other usemaps
555 * on the section which has pgdat at boot time. Just keep it as is now.
559 struct page *memmap_page;
560 memmap_page = virt_to_page(memmap);
562 nr_pages = PAGE_ALIGN(PAGES_PER_SECTION * sizeof(struct page))
565 free_map_bootmem(memmap_page, nr_pages);
570 * returns the number of sections whose mem_maps were properly
571 * set. If this is <=0, then that means that the passed-in
572 * map was not consumed and must be freed.
574 int sparse_add_one_section(struct zone *zone, unsigned long start_pfn,
577 unsigned long section_nr = pfn_to_section_nr(start_pfn);
578 struct pglist_data *pgdat = zone->zone_pgdat;
579 struct mem_section *ms;
581 unsigned long *usemap;
586 * no locking for this, because it does its own
587 * plus, it does a kmalloc
589 ret = sparse_index_init(section_nr, pgdat->node_id);
590 if (ret < 0 && ret != -EEXIST)
592 memmap = kmalloc_section_memmap(section_nr, pgdat->node_id, nr_pages);
595 usemap = __kmalloc_section_usemap();
597 __kfree_section_memmap(memmap, nr_pages);
601 pgdat_resize_lock(pgdat, &flags);
603 ms = __pfn_to_section(start_pfn);
604 if (ms->section_mem_map & SECTION_MARKED_PRESENT) {
609 ms->section_mem_map |= SECTION_MARKED_PRESENT;
611 ret = sparse_init_one_section(ms, section_nr, memmap, usemap);
614 pgdat_resize_unlock(pgdat, &flags);
617 __kfree_section_memmap(memmap, nr_pages);
622 void sparse_remove_one_section(struct zone *zone, struct mem_section *ms)
624 struct page *memmap = NULL;
625 unsigned long *usemap = NULL;
627 if (ms->section_mem_map) {
628 usemap = ms->pageblock_flags;
629 memmap = sparse_decode_mem_map(ms->section_mem_map,
631 ms->section_mem_map = 0;
632 ms->pageblock_flags = NULL;
635 free_section_usemap(memmap, usemap);