Merge git://git.kernel.org/pub/scm/linux/kernel/git/gregkh/driver-2.6
[linux-2.6] / arch / x86 / mm / init_64.c
1 /*
2  *  linux/arch/x86_64/mm/init.c
3  *
4  *  Copyright (C) 1995  Linus Torvalds
5  *  Copyright (C) 2000  Pavel Machek <pavel@suse.cz>
6  *  Copyright (C) 2002,2003 Andi Kleen <ak@suse.de>
7  */
8
9 #include <linux/signal.h>
10 #include <linux/sched.h>
11 #include <linux/kernel.h>
12 #include <linux/errno.h>
13 #include <linux/string.h>
14 #include <linux/types.h>
15 #include <linux/ptrace.h>
16 #include <linux/mman.h>
17 #include <linux/mm.h>
18 #include <linux/swap.h>
19 #include <linux/smp.h>
20 #include <linux/init.h>
21 #include <linux/pagemap.h>
22 #include <linux/bootmem.h>
23 #include <linux/proc_fs.h>
24 #include <linux/pci.h>
25 #include <linux/pfn.h>
26 #include <linux/poison.h>
27 #include <linux/dma-mapping.h>
28 #include <linux/module.h>
29 #include <linux/memory_hotplug.h>
30 #include <linux/nmi.h>
31
32 #include <asm/processor.h>
33 #include <asm/system.h>
34 #include <asm/uaccess.h>
35 #include <asm/pgtable.h>
36 #include <asm/pgalloc.h>
37 #include <asm/dma.h>
38 #include <asm/fixmap.h>
39 #include <asm/e820.h>
40 #include <asm/apic.h>
41 #include <asm/tlb.h>
42 #include <asm/mmu_context.h>
43 #include <asm/proto.h>
44 #include <asm/smp.h>
45 #include <asm/sections.h>
46 #include <asm/kdebug.h>
47 #include <asm/numa.h>
48
49 const struct dma_mapping_ops *dma_ops;
50 EXPORT_SYMBOL(dma_ops);
51
52 static unsigned long dma_reserve __initdata;
53
54 DEFINE_PER_CPU(struct mmu_gather, mmu_gathers);
55
56 /*
57  * NOTE: pagetable_init alloc all the fixmap pagetables contiguous on the
58  * physical space so we can cache the place of the first one and move
59  * around without checking the pgd every time.
60  */
61
62 void show_mem(void)
63 {
64         long i, total = 0, reserved = 0;
65         long shared = 0, cached = 0;
66         struct page *page;
67         pg_data_t *pgdat;
68
69         printk(KERN_INFO "Mem-info:\n");
70         show_free_areas();
71         printk(KERN_INFO "Free swap:       %6ldkB\n",
72                 nr_swap_pages << (PAGE_SHIFT-10));
73
74         for_each_online_pgdat(pgdat) {
75                 for (i = 0; i < pgdat->node_spanned_pages; ++i) {
76                         /*
77                          * This loop can take a while with 256 GB and
78                          * 4k pages so defer the NMI watchdog:
79                          */
80                         if (unlikely(i % MAX_ORDER_NR_PAGES == 0))
81                                 touch_nmi_watchdog();
82
83                         if (!pfn_valid(pgdat->node_start_pfn + i))
84                                 continue;
85
86                         page = pfn_to_page(pgdat->node_start_pfn + i);
87                         total++;
88                         if (PageReserved(page))
89                                 reserved++;
90                         else if (PageSwapCache(page))
91                                 cached++;
92                         else if (page_count(page))
93                                 shared += page_count(page) - 1;
94                 }
95         }
96         printk(KERN_INFO "%lu pages of RAM\n",          total);
97         printk(KERN_INFO "%lu reserved pages\n",        reserved);
98         printk(KERN_INFO "%lu pages shared\n",          shared);
99         printk(KERN_INFO "%lu pages swap cached\n",     cached);
100 }
101
102 int after_bootmem;
103
104 static __init void *spp_getpage(void)
105 {
106         void *ptr;
107
108         if (after_bootmem)
109                 ptr = (void *) get_zeroed_page(GFP_ATOMIC);
110         else
111                 ptr = alloc_bootmem_pages(PAGE_SIZE);
112
113         if (!ptr || ((unsigned long)ptr & ~PAGE_MASK)) {
114                 panic("set_pte_phys: cannot allocate page data %s\n",
115                         after_bootmem ? "after bootmem" : "");
116         }
117
118         pr_debug("spp_getpage %p\n", ptr);
119
120         return ptr;
121 }
122
123 static __init void
124 set_pte_phys(unsigned long vaddr, unsigned long phys, pgprot_t prot)
125 {
126         pgd_t *pgd;
127         pud_t *pud;
128         pmd_t *pmd;
129         pte_t *pte, new_pte;
130
131         pr_debug("set_pte_phys %lx to %lx\n", vaddr, phys);
132
133         pgd = pgd_offset_k(vaddr);
134         if (pgd_none(*pgd)) {
135                 printk(KERN_ERR
136                         "PGD FIXMAP MISSING, it should be setup in head.S!\n");
137                 return;
138         }
139         pud = pud_offset(pgd, vaddr);
140         if (pud_none(*pud)) {
141                 pmd = (pmd_t *) spp_getpage();
142                 set_pud(pud, __pud(__pa(pmd) | _KERNPG_TABLE | _PAGE_USER));
143                 if (pmd != pmd_offset(pud, 0)) {
144                         printk(KERN_ERR "PAGETABLE BUG #01! %p <-> %p\n",
145                                 pmd, pmd_offset(pud, 0));
146                         return;
147                 }
148         }
149         pmd = pmd_offset(pud, vaddr);
150         if (pmd_none(*pmd)) {
151                 pte = (pte_t *) spp_getpage();
152                 set_pmd(pmd, __pmd(__pa(pte) | _KERNPG_TABLE | _PAGE_USER));
153                 if (pte != pte_offset_kernel(pmd, 0)) {
154                         printk(KERN_ERR "PAGETABLE BUG #02!\n");
155                         return;
156                 }
157         }
158         new_pte = pfn_pte(phys >> PAGE_SHIFT, prot);
159
160         pte = pte_offset_kernel(pmd, vaddr);
161         if (!pte_none(*pte) &&
162             pte_val(*pte) != (pte_val(new_pte) & __supported_pte_mask))
163                 pte_ERROR(*pte);
164         set_pte(pte, new_pte);
165
166         /*
167          * It's enough to flush this one mapping.
168          * (PGE mappings get flushed as well)
169          */
170         __flush_tlb_one(vaddr);
171 }
172
173 /* NOTE: this is meant to be run only at boot */
174 void __init
175 __set_fixmap(enum fixed_addresses idx, unsigned long phys, pgprot_t prot)
176 {
177         unsigned long address = __fix_to_virt(idx);
178
179         if (idx >= __end_of_fixed_addresses) {
180                 printk(KERN_ERR "Invalid __set_fixmap\n");
181                 return;
182         }
183         set_pte_phys(address, phys, prot);
184 }
185
186 static unsigned long __initdata table_start;
187 static unsigned long __meminitdata table_end;
188
189 static __meminit void *alloc_low_page(unsigned long *phys)
190 {
191         unsigned long pfn = table_end++;
192         void *adr;
193
194         if (after_bootmem) {
195                 adr = (void *)get_zeroed_page(GFP_ATOMIC);
196                 *phys = __pa(adr);
197
198                 return adr;
199         }
200
201         if (pfn >= end_pfn)
202                 panic("alloc_low_page: ran out of memory");
203
204         adr = early_ioremap(pfn * PAGE_SIZE, PAGE_SIZE);
205         memset(adr, 0, PAGE_SIZE);
206         *phys  = pfn * PAGE_SIZE;
207         return adr;
208 }
209
210 static __meminit void unmap_low_page(void *adr)
211 {
212         if (after_bootmem)
213                 return;
214
215         early_iounmap(adr, PAGE_SIZE);
216 }
217
218 /* Must run before zap_low_mappings */
219 __meminit void *early_ioremap(unsigned long addr, unsigned long size)
220 {
221         pmd_t *pmd, *last_pmd;
222         unsigned long vaddr;
223         int i, pmds;
224
225         pmds = ((addr & ~PMD_MASK) + size + ~PMD_MASK) / PMD_SIZE;
226         vaddr = __START_KERNEL_map;
227         pmd = level2_kernel_pgt;
228         last_pmd = level2_kernel_pgt + PTRS_PER_PMD - 1;
229
230         for (; pmd <= last_pmd; pmd++, vaddr += PMD_SIZE) {
231                 for (i = 0; i < pmds; i++) {
232                         if (pmd_present(pmd[i]))
233                                 goto continue_outer_loop;
234                 }
235                 vaddr += addr & ~PMD_MASK;
236                 addr &= PMD_MASK;
237
238                 for (i = 0; i < pmds; i++, addr += PMD_SIZE)
239                         set_pmd(pmd+i, __pmd(addr | __PAGE_KERNEL_LARGE_EXEC));
240                 __flush_tlb_all();
241
242                 return (void *)vaddr;
243 continue_outer_loop:
244                 ;
245         }
246         printk(KERN_ERR "early_ioremap(0x%lx, %lu) failed\n", addr, size);
247
248         return NULL;
249 }
250
251 /*
252  * To avoid virtual aliases later:
253  */
254 __meminit void early_iounmap(void *addr, unsigned long size)
255 {
256         unsigned long vaddr;
257         pmd_t *pmd;
258         int i, pmds;
259
260         vaddr = (unsigned long)addr;
261         pmds = ((vaddr & ~PMD_MASK) + size + ~PMD_MASK) / PMD_SIZE;
262         pmd = level2_kernel_pgt + pmd_index(vaddr);
263
264         for (i = 0; i < pmds; i++)
265                 pmd_clear(pmd + i);
266
267         __flush_tlb_all();
268 }
269
270 static void __meminit
271 phys_pmd_init(pmd_t *pmd_page, unsigned long address, unsigned long end)
272 {
273         int i = pmd_index(address);
274
275         for (; i < PTRS_PER_PMD; i++, address += PMD_SIZE) {
276                 unsigned long entry;
277                 pmd_t *pmd = pmd_page + pmd_index(address);
278
279                 if (address >= end) {
280                         if (!after_bootmem) {
281                                 for (; i < PTRS_PER_PMD; i++, pmd++)
282                                         set_pmd(pmd, __pmd(0));
283                         }
284                         break;
285                 }
286
287                 if (pmd_val(*pmd))
288                         continue;
289
290                 entry = __PAGE_KERNEL_LARGE|_PAGE_GLOBAL|address;
291                 entry &= __supported_pte_mask;
292                 set_pmd(pmd, __pmd(entry));
293         }
294 }
295
296 static void __meminit
297 phys_pmd_update(pud_t *pud, unsigned long address, unsigned long end)
298 {
299         pmd_t *pmd = pmd_offset(pud, 0);
300         spin_lock(&init_mm.page_table_lock);
301         phys_pmd_init(pmd, address, end);
302         spin_unlock(&init_mm.page_table_lock);
303         __flush_tlb_all();
304 }
305
306 static void __meminit
307 phys_pud_init(pud_t *pud_page, unsigned long addr, unsigned long end)
308 {
309         int i = pud_index(addr);
310
311         for (; i < PTRS_PER_PUD; i++, addr = (addr & PUD_MASK) + PUD_SIZE) {
312                 unsigned long pmd_phys;
313                 pud_t *pud = pud_page + pud_index(addr);
314                 pmd_t *pmd;
315
316                 if (addr >= end)
317                         break;
318
319                 if (!after_bootmem &&
320                                 !e820_any_mapped(addr, addr+PUD_SIZE, 0)) {
321                         set_pud(pud, __pud(0));
322                         continue;
323                 }
324
325                 if (pud_val(*pud)) {
326                         phys_pmd_update(pud, addr, end);
327                         continue;
328                 }
329
330                 pmd = alloc_low_page(&pmd_phys);
331
332                 spin_lock(&init_mm.page_table_lock);
333                 set_pud(pud, __pud(pmd_phys | _KERNPG_TABLE));
334                 phys_pmd_init(pmd, addr, end);
335                 spin_unlock(&init_mm.page_table_lock);
336
337                 unmap_low_page(pmd);
338         }
339         __flush_tlb_all();
340 }
341
342 static void __init find_early_table_space(unsigned long end)
343 {
344         unsigned long puds, pmds, tables, start;
345
346         puds = (end + PUD_SIZE - 1) >> PUD_SHIFT;
347         pmds = (end + PMD_SIZE - 1) >> PMD_SHIFT;
348         tables = round_up(puds * sizeof(pud_t), PAGE_SIZE) +
349                  round_up(pmds * sizeof(pmd_t), PAGE_SIZE);
350
351         /*
352          * RED-PEN putting page tables only on node 0 could
353          * cause a hotspot and fill up ZONE_DMA. The page tables
354          * need roughly 0.5KB per GB.
355          */
356         start = 0x8000;
357         table_start = find_e820_area(start, end, tables);
358         if (table_start == -1UL)
359                 panic("Cannot find space for the kernel page tables");
360
361         /*
362          * When you have a lot of RAM like 256GB, early_table will not fit
363          * into 0x8000 range, find_e820_area() will find area after kernel
364          * bss but the table_start is not page aligned, so need to round it
365          * up to avoid overlap with bss:
366          */
367         table_start = round_up(table_start, PAGE_SIZE);
368         table_start >>= PAGE_SHIFT;
369         table_end = table_start;
370
371         early_printk("kernel direct mapping tables up to %lx @ %lx-%lx\n",
372                 end, table_start << PAGE_SHIFT,
373                 (table_start << PAGE_SHIFT) + tables);
374 }
375
376 /*
377  * Setup the direct mapping of the physical memory at PAGE_OFFSET.
378  * This runs before bootmem is initialized and gets pages directly from
379  * the physical memory. To access them they are temporarily mapped.
380  */
381 void __init_refok init_memory_mapping(unsigned long start, unsigned long end)
382 {
383         unsigned long next;
384
385         pr_debug("init_memory_mapping\n");
386
387         /*
388          * Find space for the kernel direct mapping tables.
389          *
390          * Later we should allocate these tables in the local node of the
391          * memory mapped. Unfortunately this is done currently before the
392          * nodes are discovered.
393          */
394         if (!after_bootmem)
395                 find_early_table_space(end);
396
397         start = (unsigned long)__va(start);
398         end = (unsigned long)__va(end);
399
400         for (; start < end; start = next) {
401                 pgd_t *pgd = pgd_offset_k(start);
402                 unsigned long pud_phys;
403                 pud_t *pud;
404
405                 if (after_bootmem)
406                         pud = pud_offset(pgd, start & PGDIR_MASK);
407                 else
408                         pud = alloc_low_page(&pud_phys);
409
410                 next = start + PGDIR_SIZE;
411                 if (next > end)
412                         next = end;
413                 phys_pud_init(pud, __pa(start), __pa(next));
414                 if (!after_bootmem)
415                         set_pgd(pgd_offset_k(start), mk_kernel_pgd(pud_phys));
416                 unmap_low_page(pud);
417         }
418
419         if (!after_bootmem)
420                 mmu_cr4_features = read_cr4();
421         __flush_tlb_all();
422
423         reserve_early(table_start << PAGE_SHIFT, table_end << PAGE_SHIFT);
424 }
425
426 #ifndef CONFIG_NUMA
427 void __init paging_init(void)
428 {
429         unsigned long max_zone_pfns[MAX_NR_ZONES];
430
431         memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
432         max_zone_pfns[ZONE_DMA] = MAX_DMA_PFN;
433         max_zone_pfns[ZONE_DMA32] = MAX_DMA32_PFN;
434         max_zone_pfns[ZONE_NORMAL] = end_pfn;
435
436         memory_present(0, 0, end_pfn);
437         sparse_init();
438         free_area_init_nodes(max_zone_pfns);
439 }
440 #endif
441
442 /*
443  * Unmap a kernel mapping if it exists. This is useful to avoid
444  * prefetches from the CPU leading to inconsistent cache lines.
445  * address and size must be aligned to 2MB boundaries.
446  * Does nothing when the mapping doesn't exist.
447  */
448 void __init clear_kernel_mapping(unsigned long address, unsigned long size)
449 {
450         unsigned long end = address + size;
451
452         BUG_ON(address & ~LARGE_PAGE_MASK);
453         BUG_ON(size & ~LARGE_PAGE_MASK);
454
455         for (; address < end; address += LARGE_PAGE_SIZE) {
456                 pgd_t *pgd = pgd_offset_k(address);
457                 pud_t *pud;
458                 pmd_t *pmd;
459
460                 if (pgd_none(*pgd))
461                         continue;
462
463                 pud = pud_offset(pgd, address);
464                 if (pud_none(*pud))
465                         continue;
466
467                 pmd = pmd_offset(pud, address);
468                 if (!pmd || pmd_none(*pmd))
469                         continue;
470
471                 if (!(pmd_val(*pmd) & _PAGE_PSE)) {
472                         /*
473                          * Could handle this, but it should not happen
474                          * currently:
475                          */
476                         printk(KERN_ERR "clear_kernel_mapping: "
477                                 "mapping has been split. will leak memory\n");
478                         pmd_ERROR(*pmd);
479                 }
480                 set_pmd(pmd, __pmd(0));
481         }
482         __flush_tlb_all();
483 }
484
485 /*
486  * Memory hotplug specific functions
487  */
488 void online_page(struct page *page)
489 {
490         ClearPageReserved(page);
491         init_page_count(page);
492         __free_page(page);
493         totalram_pages++;
494         num_physpages++;
495 }
496
497 #ifdef CONFIG_MEMORY_HOTPLUG
498 /*
499  * Memory is added always to NORMAL zone. This means you will never get
500  * additional DMA/DMA32 memory.
501  */
502 int arch_add_memory(int nid, u64 start, u64 size)
503 {
504         struct pglist_data *pgdat = NODE_DATA(nid);
505         struct zone *zone = pgdat->node_zones + ZONE_NORMAL;
506         unsigned long start_pfn = start >> PAGE_SHIFT;
507         unsigned long nr_pages = size >> PAGE_SHIFT;
508         int ret;
509
510         init_memory_mapping(start, start + size-1);
511
512         ret = __add_pages(zone, start_pfn, nr_pages);
513         WARN_ON(1);
514
515         return ret;
516 }
517 EXPORT_SYMBOL_GPL(arch_add_memory);
518
519 #if !defined(CONFIG_ACPI_NUMA) && defined(CONFIG_NUMA)
520 int memory_add_physaddr_to_nid(u64 start)
521 {
522         return 0;
523 }
524 EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid);
525 #endif
526
527 #endif /* CONFIG_MEMORY_HOTPLUG */
528
529 static struct kcore_list kcore_mem, kcore_vmalloc, kcore_kernel,
530                          kcore_modules, kcore_vsyscall;
531
532 void __init mem_init(void)
533 {
534         long codesize, reservedpages, datasize, initsize;
535
536         pci_iommu_alloc();
537
538         /* clear_bss() already clear the empty_zero_page */
539
540         /* temporary debugging - double check it's true: */
541         {
542                 int i;
543
544                 for (i = 0; i < 1024; i++)
545                         WARN_ON_ONCE(empty_zero_page[i]);
546         }
547
548         reservedpages = 0;
549
550         /* this will put all low memory onto the freelists */
551 #ifdef CONFIG_NUMA
552         totalram_pages = numa_free_all_bootmem();
553 #else
554         totalram_pages = free_all_bootmem();
555 #endif
556         reservedpages = end_pfn - totalram_pages -
557                                         absent_pages_in_range(0, end_pfn);
558         after_bootmem = 1;
559
560         codesize =  (unsigned long) &_etext - (unsigned long) &_text;
561         datasize =  (unsigned long) &_edata - (unsigned long) &_etext;
562         initsize =  (unsigned long) &__init_end - (unsigned long) &__init_begin;
563
564         /* Register memory areas for /proc/kcore */
565         kclist_add(&kcore_mem, __va(0), max_low_pfn << PAGE_SHIFT);
566         kclist_add(&kcore_vmalloc, (void *)VMALLOC_START,
567                    VMALLOC_END-VMALLOC_START);
568         kclist_add(&kcore_kernel, &_stext, _end - _stext);
569         kclist_add(&kcore_modules, (void *)MODULES_VADDR, MODULES_LEN);
570         kclist_add(&kcore_vsyscall, (void *)VSYSCALL_START,
571                                  VSYSCALL_END - VSYSCALL_START);
572
573         printk(KERN_INFO "Memory: %luk/%luk available (%ldk kernel code, "
574                                 "%ldk reserved, %ldk data, %ldk init)\n",
575                 (unsigned long) nr_free_pages() << (PAGE_SHIFT-10),
576                 end_pfn << (PAGE_SHIFT-10),
577                 codesize >> 10,
578                 reservedpages << (PAGE_SHIFT-10),
579                 datasize >> 10,
580                 initsize >> 10);
581 }
582
583 void free_init_pages(char *what, unsigned long begin, unsigned long end)
584 {
585         unsigned long addr;
586
587         if (begin >= end)
588                 return;
589
590         /*
591          * If debugging page accesses then do not free this memory but
592          * mark them not present - any buggy init-section access will
593          * create a kernel page fault:
594          */
595 #ifdef CONFIG_DEBUG_PAGEALLOC
596         printk(KERN_INFO "debug: unmapping init memory %08lx..%08lx\n",
597                 begin, PAGE_ALIGN(end));
598         set_memory_np(begin, (end - begin) >> PAGE_SHIFT);
599 #else
600         printk(KERN_INFO "Freeing %s: %luk freed\n", what, (end - begin) >> 10);
601
602         for (addr = begin; addr < end; addr += PAGE_SIZE) {
603                 ClearPageReserved(virt_to_page(addr));
604                 init_page_count(virt_to_page(addr));
605                 memset((void *)(addr & ~(PAGE_SIZE-1)),
606                         POISON_FREE_INITMEM, PAGE_SIZE);
607                 free_page(addr);
608                 totalram_pages++;
609         }
610 #endif
611 }
612
613 void free_initmem(void)
614 {
615         free_init_pages("unused kernel memory",
616                         (unsigned long)(&__init_begin),
617                         (unsigned long)(&__init_end));
618 }
619
620 #ifdef CONFIG_DEBUG_RODATA
621 const int rodata_test_data = 0xC3;
622 EXPORT_SYMBOL_GPL(rodata_test_data);
623
624 void mark_rodata_ro(void)
625 {
626         unsigned long start = (unsigned long)_stext, end;
627
628 #ifdef CONFIG_HOTPLUG_CPU
629         /* It must still be possible to apply SMP alternatives. */
630         if (num_possible_cpus() > 1)
631                 start = (unsigned long)_etext;
632 #endif
633
634 #ifdef CONFIG_KPROBES
635         start = (unsigned long)__start_rodata;
636 #endif
637
638         end = (unsigned long)__end_rodata;
639         start = (start + PAGE_SIZE - 1) & PAGE_MASK;
640         end &= PAGE_MASK;
641         if (end <= start)
642                 return;
643
644         set_memory_ro(start, (end - start) >> PAGE_SHIFT);
645
646         printk(KERN_INFO "Write protecting the kernel read-only data: %luk\n",
647                (end - start) >> 10);
648
649         rodata_test();
650
651 #ifdef CONFIG_CPA_DEBUG
652         printk(KERN_INFO "Testing CPA: undo %lx-%lx\n", start, end);
653         set_memory_rw(start, (end-start) >> PAGE_SHIFT);
654
655         printk(KERN_INFO "Testing CPA: again\n");
656         set_memory_ro(start, (end-start) >> PAGE_SHIFT);
657 #endif
658 }
659 #endif
660
661 #ifdef CONFIG_BLK_DEV_INITRD
662 void free_initrd_mem(unsigned long start, unsigned long end)
663 {
664         free_init_pages("initrd memory", start, end);
665 }
666 #endif
667
668 void __init reserve_bootmem_generic(unsigned long phys, unsigned len)
669 {
670 #ifdef CONFIG_NUMA
671         int nid = phys_to_nid(phys);
672 #endif
673         unsigned long pfn = phys >> PAGE_SHIFT;
674
675         if (pfn >= end_pfn) {
676                 /*
677                  * This can happen with kdump kernels when accessing
678                  * firmware tables:
679                  */
680                 if (pfn < end_pfn_map)
681                         return;
682
683                 printk(KERN_ERR "reserve_bootmem: illegal reserve %lx %u\n",
684                                 phys, len);
685                 return;
686         }
687
688         /* Should check here against the e820 map to avoid double free */
689 #ifdef CONFIG_NUMA
690         reserve_bootmem_node(NODE_DATA(nid), phys, len);
691 #else
692         reserve_bootmem(phys, len);
693 #endif
694         if (phys+len <= MAX_DMA_PFN*PAGE_SIZE) {
695                 dma_reserve += len / PAGE_SIZE;
696                 set_dma_reserve(dma_reserve);
697         }
698 }
699
700 int kern_addr_valid(unsigned long addr)
701 {
702         unsigned long above = ((long)addr) >> __VIRTUAL_MASK_SHIFT;
703         pgd_t *pgd;
704         pud_t *pud;
705         pmd_t *pmd;
706         pte_t *pte;
707
708         if (above != 0 && above != -1UL)
709                 return 0;
710
711         pgd = pgd_offset_k(addr);
712         if (pgd_none(*pgd))
713                 return 0;
714
715         pud = pud_offset(pgd, addr);
716         if (pud_none(*pud))
717                 return 0;
718
719         pmd = pmd_offset(pud, addr);
720         if (pmd_none(*pmd))
721                 return 0;
722
723         if (pmd_large(*pmd))
724                 return pfn_valid(pmd_pfn(*pmd));
725
726         pte = pte_offset_kernel(pmd, addr);
727         if (pte_none(*pte))
728                 return 0;
729
730         return pfn_valid(pte_pfn(*pte));
731 }
732
733 /*
734  * A pseudo VMA to allow ptrace access for the vsyscall page.  This only
735  * covers the 64bit vsyscall page now. 32bit has a real VMA now and does
736  * not need special handling anymore:
737  */
738 static struct vm_area_struct gate_vma = {
739         .vm_start       = VSYSCALL_START,
740         .vm_end         = VSYSCALL_START + (VSYSCALL_MAPPED_PAGES * PAGE_SIZE),
741         .vm_page_prot   = PAGE_READONLY_EXEC,
742         .vm_flags       = VM_READ | VM_EXEC
743 };
744
745 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
746 {
747 #ifdef CONFIG_IA32_EMULATION
748         if (test_tsk_thread_flag(tsk, TIF_IA32))
749                 return NULL;
750 #endif
751         return &gate_vma;
752 }
753
754 int in_gate_area(struct task_struct *task, unsigned long addr)
755 {
756         struct vm_area_struct *vma = get_gate_vma(task);
757
758         if (!vma)
759                 return 0;
760
761         return (addr >= vma->vm_start) && (addr < vma->vm_end);
762 }
763
764 /*
765  * Use this when you have no reliable task/vma, typically from interrupt
766  * context. It is less reliable than using the task's vma and may give
767  * false positives:
768  */
769 int in_gate_area_no_task(unsigned long addr)
770 {
771         return (addr >= VSYSCALL_START) && (addr < VSYSCALL_END);
772 }
773
774 const char *arch_vma_name(struct vm_area_struct *vma)
775 {
776         if (vma->vm_mm && vma->vm_start == (long)vma->vm_mm->context.vdso)
777                 return "[vdso]";
778         if (vma == &gate_vma)
779                 return "[vsyscall]";
780         return NULL;
781 }
782
783 #ifdef CONFIG_SPARSEMEM_VMEMMAP
784 /*
785  * Initialise the sparsemem vmemmap using huge-pages at the PMD level.
786  */
787 int __meminit
788 vmemmap_populate(struct page *start_page, unsigned long size, int node)
789 {
790         unsigned long addr = (unsigned long)start_page;
791         unsigned long end = (unsigned long)(start_page + size);
792         unsigned long next;
793         pgd_t *pgd;
794         pud_t *pud;
795         pmd_t *pmd;
796
797         for (; addr < end; addr = next) {
798                 next = pmd_addr_end(addr, end);
799
800                 pgd = vmemmap_pgd_populate(addr, node);
801                 if (!pgd)
802                         return -ENOMEM;
803
804                 pud = vmemmap_pud_populate(pgd, addr, node);
805                 if (!pud)
806                         return -ENOMEM;
807
808                 pmd = pmd_offset(pud, addr);
809                 if (pmd_none(*pmd)) {
810                         pte_t entry;
811                         void *p;
812
813                         p = vmemmap_alloc_block(PMD_SIZE, node);
814                         if (!p)
815                                 return -ENOMEM;
816
817                         entry = pfn_pte(__pa(p) >> PAGE_SHIFT,
818                                                         PAGE_KERNEL_LARGE);
819                         set_pmd(pmd, __pmd(pte_val(entry)));
820
821                         printk(KERN_DEBUG " [%lx-%lx] PMD ->%p on node %d\n",
822                                 addr, addr + PMD_SIZE - 1, p, node);
823                 } else {
824                         vmemmap_verify((pte_t *)pmd, node, addr, next);
825                 }
826         }
827         return 0;
828 }
829 #endif