Merge branch 'linus' into x86/pat2
[linux-2.6.git] / 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/initrd.h>
22 #include <linux/pagemap.h>
23 #include <linux/bootmem.h>
24 #include <linux/proc_fs.h>
25 #include <linux/pci.h>
26 #include <linux/pfn.h>
27 #include <linux/poison.h>
28 #include <linux/dma-mapping.h>
29 #include <linux/module.h>
30 #include <linux/memory_hotplug.h>
31 #include <linux/nmi.h>
32
33 #include <asm/processor.h>
34 #include <asm/system.h>
35 #include <asm/uaccess.h>
36 #include <asm/pgtable.h>
37 #include <asm/pgalloc.h>
38 #include <asm/dma.h>
39 #include <asm/fixmap.h>
40 #include <asm/e820.h>
41 #include <asm/apic.h>
42 #include <asm/tlb.h>
43 #include <asm/mmu_context.h>
44 #include <asm/proto.h>
45 #include <asm/smp.h>
46 #include <asm/sections.h>
47 #include <asm/kdebug.h>
48 #include <asm/numa.h>
49 #include <asm/cacheflush.h>
50
51 /*
52  * end_pfn only includes RAM, while max_pfn_mapped includes all e820 entries.
53  * The direct mapping extends to max_pfn_mapped, so that we can directly access
54  * apertures, ACPI and other tables without having to play with fixmaps.
55  */
56 unsigned long max_low_pfn_mapped;
57 unsigned long max_pfn_mapped;
58
59 static unsigned long dma_reserve __initdata;
60
61 DEFINE_PER_CPU(struct mmu_gather, mmu_gathers);
62
63 int direct_gbpages
64 #ifdef CONFIG_DIRECT_GBPAGES
65                                 = 1
66 #endif
67 ;
68
69 static int __init parse_direct_gbpages_off(char *arg)
70 {
71         direct_gbpages = 0;
72         return 0;
73 }
74 early_param("nogbpages", parse_direct_gbpages_off);
75
76 static int __init parse_direct_gbpages_on(char *arg)
77 {
78         direct_gbpages = 1;
79         return 0;
80 }
81 early_param("gbpages", parse_direct_gbpages_on);
82
83 /*
84  * NOTE: pagetable_init alloc all the fixmap pagetables contiguous on the
85  * physical space so we can cache the place of the first one and move
86  * around without checking the pgd every time.
87  */
88
89 int after_bootmem;
90
91 /*
92  * NOTE: This function is marked __ref because it calls __init function
93  * (alloc_bootmem_pages). It's safe to do it ONLY when after_bootmem == 0.
94  */
95 static __ref void *spp_getpage(void)
96 {
97         void *ptr;
98
99         if (after_bootmem)
100                 ptr = (void *) get_zeroed_page(GFP_ATOMIC);
101         else
102                 ptr = alloc_bootmem_pages(PAGE_SIZE);
103
104         if (!ptr || ((unsigned long)ptr & ~PAGE_MASK)) {
105                 panic("set_pte_phys: cannot allocate page data %s\n",
106                         after_bootmem ? "after bootmem" : "");
107         }
108
109         pr_debug("spp_getpage %p\n", ptr);
110
111         return ptr;
112 }
113
114 void
115 set_pte_vaddr_pud(pud_t *pud_page, unsigned long vaddr, pte_t new_pte)
116 {
117         pud_t *pud;
118         pmd_t *pmd;
119         pte_t *pte;
120
121         pud = pud_page + pud_index(vaddr);
122         if (pud_none(*pud)) {
123                 pmd = (pmd_t *) spp_getpage();
124                 pud_populate(&init_mm, pud, pmd);
125                 if (pmd != pmd_offset(pud, 0)) {
126                         printk(KERN_ERR "PAGETABLE BUG #01! %p <-> %p\n",
127                                 pmd, pmd_offset(pud, 0));
128                         return;
129                 }
130         }
131         pmd = pmd_offset(pud, vaddr);
132         if (pmd_none(*pmd)) {
133                 pte = (pte_t *) spp_getpage();
134                 pmd_populate_kernel(&init_mm, pmd, pte);
135                 if (pte != pte_offset_kernel(pmd, 0)) {
136                         printk(KERN_ERR "PAGETABLE BUG #02!\n");
137                         return;
138                 }
139         }
140
141         pte = pte_offset_kernel(pmd, vaddr);
142         if (!pte_none(*pte) && pte_val(new_pte) &&
143             pte_val(*pte) != (pte_val(new_pte) & __supported_pte_mask))
144                 pte_ERROR(*pte);
145         set_pte(pte, new_pte);
146
147         /*
148          * It's enough to flush this one mapping.
149          * (PGE mappings get flushed as well)
150          */
151         __flush_tlb_one(vaddr);
152 }
153
154 void
155 set_pte_vaddr(unsigned long vaddr, pte_t pteval)
156 {
157         pgd_t *pgd;
158         pud_t *pud_page;
159
160         pr_debug("set_pte_vaddr %lx to %lx\n", vaddr, native_pte_val(pteval));
161
162         pgd = pgd_offset_k(vaddr);
163         if (pgd_none(*pgd)) {
164                 printk(KERN_ERR
165                         "PGD FIXMAP MISSING, it should be setup in head.S!\n");
166                 return;
167         }
168         pud_page = (pud_t*)pgd_page_vaddr(*pgd);
169         set_pte_vaddr_pud(pud_page, vaddr, pteval);
170 }
171
172 /*
173  * Create large page table mappings for a range of physical addresses.
174  */
175 static void __init __init_extra_mapping(unsigned long phys, unsigned long size,
176                                                 pgprot_t prot)
177 {
178         pgd_t *pgd;
179         pud_t *pud;
180         pmd_t *pmd;
181
182         BUG_ON((phys & ~PMD_MASK) || (size & ~PMD_MASK));
183         for (; size; phys += PMD_SIZE, size -= PMD_SIZE) {
184                 pgd = pgd_offset_k((unsigned long)__va(phys));
185                 if (pgd_none(*pgd)) {
186                         pud = (pud_t *) spp_getpage();
187                         set_pgd(pgd, __pgd(__pa(pud) | _KERNPG_TABLE |
188                                                 _PAGE_USER));
189                 }
190                 pud = pud_offset(pgd, (unsigned long)__va(phys));
191                 if (pud_none(*pud)) {
192                         pmd = (pmd_t *) spp_getpage();
193                         set_pud(pud, __pud(__pa(pmd) | _KERNPG_TABLE |
194                                                 _PAGE_USER));
195                 }
196                 pmd = pmd_offset(pud, phys);
197                 BUG_ON(!pmd_none(*pmd));
198                 set_pmd(pmd, __pmd(phys | pgprot_val(prot)));
199         }
200 }
201
202 void __init init_extra_mapping_wb(unsigned long phys, unsigned long size)
203 {
204         __init_extra_mapping(phys, size, PAGE_KERNEL_LARGE);
205 }
206
207 void __init init_extra_mapping_uc(unsigned long phys, unsigned long size)
208 {
209         __init_extra_mapping(phys, size, PAGE_KERNEL_LARGE_NOCACHE);
210 }
211
212 /*
213  * The head.S code sets up the kernel high mapping:
214  *
215  *   from __START_KERNEL_map to __START_KERNEL_map + size (== _end-_text)
216  *
217  * phys_addr holds the negative offset to the kernel, which is added
218  * to the compile time generated pmds. This results in invalid pmds up
219  * to the point where we hit the physaddr 0 mapping.
220  *
221  * We limit the mappings to the region from _text to _end.  _end is
222  * rounded up to the 2MB boundary. This catches the invalid pmds as
223  * well, as they are located before _text:
224  */
225 void __init cleanup_highmap(void)
226 {
227         unsigned long vaddr = __START_KERNEL_map;
228         unsigned long end = roundup((unsigned long)_end, PMD_SIZE) - 1;
229         pmd_t *pmd = level2_kernel_pgt;
230         pmd_t *last_pmd = pmd + PTRS_PER_PMD;
231
232         for (; pmd < last_pmd; pmd++, vaddr += PMD_SIZE) {
233                 if (pmd_none(*pmd))
234                         continue;
235                 if (vaddr < (unsigned long) _text || vaddr > end)
236                         set_pmd(pmd, __pmd(0));
237         }
238 }
239
240 static unsigned long __initdata table_start;
241 static unsigned long __meminitdata table_end;
242 static unsigned long __meminitdata table_top;
243
244 static __ref void *alloc_low_page(unsigned long *phys)
245 {
246         unsigned long pfn = table_end++;
247         void *adr;
248
249         if (after_bootmem) {
250                 adr = (void *)get_zeroed_page(GFP_ATOMIC);
251                 *phys = __pa(adr);
252
253                 return adr;
254         }
255
256         if (pfn >= table_top)
257                 panic("alloc_low_page: ran out of memory");
258
259         adr = early_ioremap(pfn * PAGE_SIZE, PAGE_SIZE);
260         memset(adr, 0, PAGE_SIZE);
261         *phys  = pfn * PAGE_SIZE;
262         return adr;
263 }
264
265 static __ref void unmap_low_page(void *adr)
266 {
267         if (after_bootmem)
268                 return;
269
270         early_iounmap(adr, PAGE_SIZE);
271 }
272
273 static unsigned long __meminit
274 phys_pte_init(pte_t *pte_page, unsigned long addr, unsigned long end,
275               pgprot_t prot)
276 {
277         unsigned pages = 0;
278         unsigned long last_map_addr = end;
279         int i;
280
281         pte_t *pte = pte_page + pte_index(addr);
282
283         for(i = pte_index(addr); i < PTRS_PER_PTE; i++, addr += PAGE_SIZE, pte++) {
284
285                 if (addr >= end) {
286                         if (!after_bootmem) {
287                                 for(; i < PTRS_PER_PTE; i++, pte++)
288                                         set_pte(pte, __pte(0));
289                         }
290                         break;
291                 }
292
293                 /*
294                  * We will re-use the existing mapping.
295                  * Xen for example has some special requirements, like mapping
296                  * pagetable pages as RO. So assume someone who pre-setup
297                  * these mappings are more intelligent.
298                  */
299                 if (pte_val(*pte))
300                         continue;
301
302                 if (0)
303                         printk("   pte=%p addr=%lx pte=%016lx\n",
304                                pte, addr, pfn_pte(addr >> PAGE_SHIFT, PAGE_KERNEL).pte);
305                 pages++;
306                 set_pte(pte, pfn_pte(addr >> PAGE_SHIFT, prot));
307                 last_map_addr = (addr & PAGE_MASK) + PAGE_SIZE;
308         }
309
310         update_page_count(PG_LEVEL_4K, pages);
311
312         return last_map_addr;
313 }
314
315 static unsigned long __meminit
316 phys_pte_update(pmd_t *pmd, unsigned long address, unsigned long end,
317                 pgprot_t prot)
318 {
319         pte_t *pte = (pte_t *)pmd_page_vaddr(*pmd);
320
321         return phys_pte_init(pte, address, end, prot);
322 }
323
324 static unsigned long __meminit
325 phys_pmd_init(pmd_t *pmd_page, unsigned long address, unsigned long end,
326               unsigned long page_size_mask, pgprot_t prot)
327 {
328         unsigned long pages = 0;
329         unsigned long last_map_addr = end;
330
331         int i = pmd_index(address);
332
333         for (; i < PTRS_PER_PMD; i++, address += PMD_SIZE) {
334                 unsigned long pte_phys;
335                 pmd_t *pmd = pmd_page + pmd_index(address);
336                 pte_t *pte;
337                 pgprot_t new_prot = prot;
338
339                 if (address >= end) {
340                         if (!after_bootmem) {
341                                 for (; i < PTRS_PER_PMD; i++, pmd++)
342                                         set_pmd(pmd, __pmd(0));
343                         }
344                         break;
345                 }
346
347                 if (pmd_val(*pmd)) {
348                         if (!pmd_large(*pmd)) {
349                                 spin_lock(&init_mm.page_table_lock);
350                                 last_map_addr = phys_pte_update(pmd, address,
351                                                                 end, prot);
352                                 spin_unlock(&init_mm.page_table_lock);
353                                 continue;
354                         }
355                         /*
356                          * If we are ok with PG_LEVEL_2M mapping, then we will
357                          * use the existing mapping,
358                          *
359                          * Otherwise, we will split the large page mapping but
360                          * use the same existing protection bits except for
361                          * large page, so that we don't violate Intel's TLB
362                          * Application note (317080) which says, while changing
363                          * the page sizes, new and old translations should
364                          * not differ with respect to page frame and
365                          * attributes.
366                          */
367                         if (page_size_mask & (1 << PG_LEVEL_2M))
368                                 continue;
369                         new_prot = pte_pgprot(pte_clrhuge(*(pte_t *)pmd));
370                 }
371
372                 if (page_size_mask & (1<<PG_LEVEL_2M)) {
373                         pages++;
374                         spin_lock(&init_mm.page_table_lock);
375                         set_pte((pte_t *)pmd,
376                                 pfn_pte(address >> PAGE_SHIFT,
377                                         __pgprot(pgprot_val(prot) | _PAGE_PSE)));
378                         spin_unlock(&init_mm.page_table_lock);
379                         last_map_addr = (address & PMD_MASK) + PMD_SIZE;
380                         continue;
381                 }
382
383                 pte = alloc_low_page(&pte_phys);
384                 last_map_addr = phys_pte_init(pte, address, end, new_prot);
385                 unmap_low_page(pte);
386
387                 spin_lock(&init_mm.page_table_lock);
388                 pmd_populate_kernel(&init_mm, pmd, __va(pte_phys));
389                 spin_unlock(&init_mm.page_table_lock);
390         }
391         update_page_count(PG_LEVEL_2M, pages);
392         return last_map_addr;
393 }
394
395 static unsigned long __meminit
396 phys_pmd_update(pud_t *pud, unsigned long address, unsigned long end,
397                 unsigned long page_size_mask, pgprot_t prot)
398 {
399         pmd_t *pmd = pmd_offset(pud, 0);
400         unsigned long last_map_addr;
401
402         last_map_addr = phys_pmd_init(pmd, address, end, page_size_mask, prot);
403         __flush_tlb_all();
404         return last_map_addr;
405 }
406
407 static unsigned long __meminit
408 phys_pud_init(pud_t *pud_page, unsigned long addr, unsigned long end,
409                          unsigned long page_size_mask)
410 {
411         unsigned long pages = 0;
412         unsigned long last_map_addr = end;
413         int i = pud_index(addr);
414
415         for (; i < PTRS_PER_PUD; i++, addr = (addr & PUD_MASK) + PUD_SIZE) {
416                 unsigned long pmd_phys;
417                 pud_t *pud = pud_page + pud_index(addr);
418                 pmd_t *pmd;
419                 pgprot_t prot = PAGE_KERNEL;
420
421                 if (addr >= end)
422                         break;
423
424                 if (!after_bootmem &&
425                                 !e820_any_mapped(addr, addr+PUD_SIZE, 0)) {
426                         set_pud(pud, __pud(0));
427                         continue;
428                 }
429
430                 if (pud_val(*pud)) {
431                         if (!pud_large(*pud)) {
432                                 last_map_addr = phys_pmd_update(pud, addr, end,
433                                                          page_size_mask, prot);
434                                 continue;
435                         }
436                         /*
437                          * If we are ok with PG_LEVEL_1G mapping, then we will
438                          * use the existing mapping.
439                          *
440                          * Otherwise, we will split the gbpage mapping but use
441                          * the same existing protection  bits except for large
442                          * page, so that we don't violate Intel's TLB
443                          * Application note (317080) which says, while changing
444                          * the page sizes, new and old translations should
445                          * not differ with respect to page frame and
446                          * attributes.
447                          */
448                         if (page_size_mask & (1 << PG_LEVEL_1G))
449                                 continue;
450                         prot = pte_pgprot(pte_clrhuge(*(pte_t *)pud));
451                 }
452
453                 if (page_size_mask & (1<<PG_LEVEL_1G)) {
454                         pages++;
455                         spin_lock(&init_mm.page_table_lock);
456                         set_pte((pte_t *)pud,
457                                 pfn_pte(addr >> PAGE_SHIFT, PAGE_KERNEL_LARGE));
458                         spin_unlock(&init_mm.page_table_lock);
459                         last_map_addr = (addr & PUD_MASK) + PUD_SIZE;
460                         continue;
461                 }
462
463                 pmd = alloc_low_page(&pmd_phys);
464                 last_map_addr = phys_pmd_init(pmd, addr, end, page_size_mask,
465                                               prot);
466                 unmap_low_page(pmd);
467
468                 spin_lock(&init_mm.page_table_lock);
469                 pud_populate(&init_mm, pud, __va(pmd_phys));
470                 spin_unlock(&init_mm.page_table_lock);
471         }
472         __flush_tlb_all();
473
474         update_page_count(PG_LEVEL_1G, pages);
475
476         return last_map_addr;
477 }
478
479 static unsigned long __meminit
480 phys_pud_update(pgd_t *pgd, unsigned long addr, unsigned long end,
481                  unsigned long page_size_mask)
482 {
483         pud_t *pud;
484
485         pud = (pud_t *)pgd_page_vaddr(*pgd);
486
487         return phys_pud_init(pud, addr, end, page_size_mask);
488 }
489
490 static void __init find_early_table_space(unsigned long end, int use_pse,
491                                           int use_gbpages)
492 {
493         unsigned long puds, pmds, ptes, tables, start;
494
495         puds = (end + PUD_SIZE - 1) >> PUD_SHIFT;
496         tables = roundup(puds * sizeof(pud_t), PAGE_SIZE);
497         if (use_gbpages) {
498                 unsigned long extra;
499                 extra = end - ((end>>PUD_SHIFT) << PUD_SHIFT);
500                 pmds = (extra + PMD_SIZE - 1) >> PMD_SHIFT;
501         } else
502                 pmds = (end + PMD_SIZE - 1) >> PMD_SHIFT;
503         tables += roundup(pmds * sizeof(pmd_t), PAGE_SIZE);
504
505         if (use_pse) {
506                 unsigned long extra;
507                 extra = end - ((end>>PMD_SHIFT) << PMD_SHIFT);
508                 ptes = (extra + PAGE_SIZE - 1) >> PAGE_SHIFT;
509         } else
510                 ptes = (end + PAGE_SIZE - 1) >> PAGE_SHIFT;
511         tables += roundup(ptes * sizeof(pte_t), PAGE_SIZE);
512
513         /*
514          * RED-PEN putting page tables only on node 0 could
515          * cause a hotspot and fill up ZONE_DMA. The page tables
516          * need roughly 0.5KB per GB.
517          */
518         start = 0x8000;
519         table_start = find_e820_area(start, end, tables, PAGE_SIZE);
520         if (table_start == -1UL)
521                 panic("Cannot find space for the kernel page tables");
522
523         table_start >>= PAGE_SHIFT;
524         table_end = table_start;
525         table_top = table_start + (tables >> PAGE_SHIFT);
526
527         printk(KERN_DEBUG "kernel direct mapping tables up to %lx @ %lx-%lx\n",
528                 end, table_start << PAGE_SHIFT, table_top << PAGE_SHIFT);
529 }
530
531 static void __init init_gbpages(void)
532 {
533         if (direct_gbpages && cpu_has_gbpages)
534                 printk(KERN_INFO "Using GB pages for direct mapping\n");
535         else
536                 direct_gbpages = 0;
537 }
538
539 static unsigned long __init kernel_physical_mapping_init(unsigned long start,
540                                                 unsigned long end,
541                                                 unsigned long page_size_mask)
542 {
543
544         unsigned long next, last_map_addr = end;
545
546         start = (unsigned long)__va(start);
547         end = (unsigned long)__va(end);
548
549         for (; start < end; start = next) {
550                 pgd_t *pgd = pgd_offset_k(start);
551                 unsigned long pud_phys;
552                 pud_t *pud;
553
554                 next = (start + PGDIR_SIZE) & PGDIR_MASK;
555                 if (next > end)
556                         next = end;
557
558                 if (pgd_val(*pgd)) {
559                         last_map_addr = phys_pud_update(pgd, __pa(start),
560                                                  __pa(end), page_size_mask);
561                         continue;
562                 }
563
564                 pud = alloc_low_page(&pud_phys);
565                 last_map_addr = phys_pud_init(pud, __pa(start), __pa(next),
566                                                  page_size_mask);
567                 unmap_low_page(pud);
568
569                 spin_lock(&init_mm.page_table_lock);
570                 pgd_populate(&init_mm, pgd, __va(pud_phys));
571                 spin_unlock(&init_mm.page_table_lock);
572         }
573         __flush_tlb_all();
574
575         return last_map_addr;
576 }
577
578 struct map_range {
579         unsigned long start;
580         unsigned long end;
581         unsigned page_size_mask;
582 };
583
584 #define NR_RANGE_MR 5
585
586 static int save_mr(struct map_range *mr, int nr_range,
587                    unsigned long start_pfn, unsigned long end_pfn,
588                    unsigned long page_size_mask)
589 {
590
591         if (start_pfn < end_pfn) {
592                 if (nr_range >= NR_RANGE_MR)
593                         panic("run out of range for init_memory_mapping\n");
594                 mr[nr_range].start = start_pfn<<PAGE_SHIFT;
595                 mr[nr_range].end   = end_pfn<<PAGE_SHIFT;
596                 mr[nr_range].page_size_mask = page_size_mask;
597                 nr_range++;
598         }
599
600         return nr_range;
601 }
602
603 /*
604  * Setup the direct mapping of the physical memory at PAGE_OFFSET.
605  * This runs before bootmem is initialized and gets pages directly from
606  * the physical memory. To access them they are temporarily mapped.
607  */
608 unsigned long __init_refok init_memory_mapping(unsigned long start,
609                                                unsigned long end)
610 {
611         unsigned long last_map_addr = 0;
612         unsigned long page_size_mask = 0;
613         unsigned long start_pfn, end_pfn;
614
615         struct map_range mr[NR_RANGE_MR];
616         int nr_range, i;
617         int use_pse, use_gbpages;
618
619         printk(KERN_INFO "init_memory_mapping\n");
620
621         /*
622          * Find space for the kernel direct mapping tables.
623          *
624          * Later we should allocate these tables in the local node of the
625          * memory mapped. Unfortunately this is done currently before the
626          * nodes are discovered.
627          */
628         if (!after_bootmem)
629                 init_gbpages();
630
631 #ifdef CONFIG_DEBUG_PAGEALLOC
632         /*
633          * For CONFIG_DEBUG_PAGEALLOC, identity mapping will use small pages.
634          * This will simplify cpa(), which otherwise needs to support splitting
635          * large pages into small in interrupt context, etc.
636          */
637         use_pse = use_gbpages = 0;
638 #else
639         use_pse = cpu_has_pse;
640         use_gbpages = direct_gbpages;
641 #endif
642
643         if (use_gbpages)
644                 page_size_mask |= 1 << PG_LEVEL_1G;
645         if (use_pse)
646                 page_size_mask |= 1 << PG_LEVEL_2M;
647
648         memset(mr, 0, sizeof(mr));
649         nr_range = 0;
650
651         /* head if not big page alignment ?*/
652         start_pfn = start >> PAGE_SHIFT;
653         end_pfn = ((start + (PMD_SIZE - 1)) >> PMD_SHIFT)
654                         << (PMD_SHIFT - PAGE_SHIFT);
655         nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
656
657         /* big page (2M) range*/
658         start_pfn = ((start + (PMD_SIZE - 1))>>PMD_SHIFT)
659                          << (PMD_SHIFT - PAGE_SHIFT);
660         end_pfn = ((start + (PUD_SIZE - 1))>>PUD_SHIFT)
661                          << (PUD_SHIFT - PAGE_SHIFT);
662         if (end_pfn > ((end>>PUD_SHIFT)<<(PUD_SHIFT - PAGE_SHIFT)))
663                 end_pfn = ((end>>PUD_SHIFT)<<(PUD_SHIFT - PAGE_SHIFT));
664         nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
665                         page_size_mask & (1<<PG_LEVEL_2M));
666
667         /* big page (1G) range */
668         start_pfn = end_pfn;
669         end_pfn = (end>>PUD_SHIFT) << (PUD_SHIFT - PAGE_SHIFT);
670         nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
671                                 page_size_mask &
672                                  ((1<<PG_LEVEL_2M)|(1<<PG_LEVEL_1G)));
673
674         /* tail is not big page (1G) alignment */
675         start_pfn = end_pfn;
676         end_pfn = (end>>PMD_SHIFT) << (PMD_SHIFT - PAGE_SHIFT);
677         nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
678                         page_size_mask & (1<<PG_LEVEL_2M));
679
680         /* tail is not big page (2M) alignment */
681         start_pfn = end_pfn;
682         end_pfn = end>>PAGE_SHIFT;
683         nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
684
685         /* try to merge same page size and continuous */
686         for (i = 0; nr_range > 1 && i < nr_range - 1; i++) {
687                 unsigned long old_start;
688                 if (mr[i].end != mr[i+1].start ||
689                     mr[i].page_size_mask != mr[i+1].page_size_mask)
690                         continue;
691                 /* move it */
692                 old_start = mr[i].start;
693                 memmove(&mr[i], &mr[i+1],
694                          (nr_range - 1 - i) * sizeof (struct map_range));
695                 mr[i].start = old_start;
696                 nr_range--;
697         }
698
699         for (i = 0; i < nr_range; i++)
700                 printk(KERN_DEBUG " %010lx - %010lx page %s\n",
701                                 mr[i].start, mr[i].end,
702                         (mr[i].page_size_mask & (1<<PG_LEVEL_1G))?"1G":(
703                          (mr[i].page_size_mask & (1<<PG_LEVEL_2M))?"2M":"4k"));
704
705         if (!after_bootmem)
706                 find_early_table_space(end, use_pse, use_gbpages);
707
708         for (i = 0; i < nr_range; i++)
709                 last_map_addr = kernel_physical_mapping_init(
710                                         mr[i].start, mr[i].end,
711                                         mr[i].page_size_mask);
712
713         if (!after_bootmem)
714                 mmu_cr4_features = read_cr4();
715         __flush_tlb_all();
716
717         if (!after_bootmem && table_end > table_start)
718                 reserve_early(table_start << PAGE_SHIFT,
719                                  table_end << PAGE_SHIFT, "PGTABLE");
720
721         printk(KERN_INFO "last_map_addr: %lx end: %lx\n",
722                          last_map_addr, end);
723
724         if (!after_bootmem)
725                 early_memtest(start, end);
726
727         return last_map_addr >> PAGE_SHIFT;
728 }
729
730 #ifndef CONFIG_NUMA
731 void __init initmem_init(unsigned long start_pfn, unsigned long end_pfn)
732 {
733         unsigned long bootmap_size, bootmap;
734
735         bootmap_size = bootmem_bootmap_pages(end_pfn)<<PAGE_SHIFT;
736         bootmap = find_e820_area(0, end_pfn<<PAGE_SHIFT, bootmap_size,
737                                  PAGE_SIZE);
738         if (bootmap == -1L)
739                 panic("Cannot find bootmem map of size %ld\n", bootmap_size);
740         /* don't touch min_low_pfn */
741         bootmap_size = init_bootmem_node(NODE_DATA(0), bootmap >> PAGE_SHIFT,
742                                          0, end_pfn);
743         e820_register_active_regions(0, start_pfn, end_pfn);
744         free_bootmem_with_active_regions(0, end_pfn);
745         early_res_to_bootmem(0, end_pfn<<PAGE_SHIFT);
746         reserve_bootmem(bootmap, bootmap_size, BOOTMEM_DEFAULT);
747 }
748
749 void __init paging_init(void)
750 {
751         unsigned long max_zone_pfns[MAX_NR_ZONES];
752
753         memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
754         max_zone_pfns[ZONE_DMA] = MAX_DMA_PFN;
755         max_zone_pfns[ZONE_DMA32] = MAX_DMA32_PFN;
756         max_zone_pfns[ZONE_NORMAL] = max_pfn;
757
758         memory_present(0, 0, max_pfn);
759         sparse_init();
760         free_area_init_nodes(max_zone_pfns);
761 }
762 #endif
763
764 /*
765  * Memory hotplug specific functions
766  */
767 #ifdef CONFIG_MEMORY_HOTPLUG
768 /*
769  * Memory is added always to NORMAL zone. This means you will never get
770  * additional DMA/DMA32 memory.
771  */
772 int arch_add_memory(int nid, u64 start, u64 size)
773 {
774         struct pglist_data *pgdat = NODE_DATA(nid);
775         struct zone *zone = pgdat->node_zones + ZONE_NORMAL;
776         unsigned long last_mapped_pfn, start_pfn = start >> PAGE_SHIFT;
777         unsigned long nr_pages = size >> PAGE_SHIFT;
778         int ret;
779
780         last_mapped_pfn = init_memory_mapping(start, start + size-1);
781         if (last_mapped_pfn > max_pfn_mapped)
782                 max_pfn_mapped = last_mapped_pfn;
783
784         ret = __add_pages(zone, start_pfn, nr_pages);
785         WARN_ON(1);
786
787         return ret;
788 }
789 EXPORT_SYMBOL_GPL(arch_add_memory);
790
791 #if !defined(CONFIG_ACPI_NUMA) && defined(CONFIG_NUMA)
792 int memory_add_physaddr_to_nid(u64 start)
793 {
794         return 0;
795 }
796 EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid);
797 #endif
798
799 #endif /* CONFIG_MEMORY_HOTPLUG */
800
801 /*
802  * devmem_is_allowed() checks to see if /dev/mem access to a certain address
803  * is valid. The argument is a physical page number.
804  *
805  *
806  * On x86, access has to be given to the first megabyte of ram because that area
807  * contains bios code and data regions used by X and dosemu and similar apps.
808  * Access has to be given to non-kernel-ram areas as well, these contain the PCI
809  * mmio resources as well as potential bios/acpi data regions.
810  */
811 int devmem_is_allowed(unsigned long pagenr)
812 {
813         if (pagenr <= 256)
814                 return 1;
815         if (!page_is_ram(pagenr))
816                 return 1;
817         return 0;
818 }
819
820
821 static struct kcore_list kcore_mem, kcore_vmalloc, kcore_kernel,
822                          kcore_modules, kcore_vsyscall;
823
824 void __init mem_init(void)
825 {
826         long codesize, reservedpages, datasize, initsize;
827
828         pci_iommu_alloc();
829
830         /* clear_bss() already clear the empty_zero_page */
831
832         reservedpages = 0;
833
834         /* this will put all low memory onto the freelists */
835 #ifdef CONFIG_NUMA
836         totalram_pages = numa_free_all_bootmem();
837 #else
838         totalram_pages = free_all_bootmem();
839 #endif
840         reservedpages = max_pfn - totalram_pages -
841                                         absent_pages_in_range(0, max_pfn);
842         after_bootmem = 1;
843
844         codesize =  (unsigned long) &_etext - (unsigned long) &_text;
845         datasize =  (unsigned long) &_edata - (unsigned long) &_etext;
846         initsize =  (unsigned long) &__init_end - (unsigned long) &__init_begin;
847
848         /* Register memory areas for /proc/kcore */
849         kclist_add(&kcore_mem, __va(0), max_low_pfn << PAGE_SHIFT);
850         kclist_add(&kcore_vmalloc, (void *)VMALLOC_START,
851                    VMALLOC_END-VMALLOC_START);
852         kclist_add(&kcore_kernel, &_stext, _end - _stext);
853         kclist_add(&kcore_modules, (void *)MODULES_VADDR, MODULES_LEN);
854         kclist_add(&kcore_vsyscall, (void *)VSYSCALL_START,
855                                  VSYSCALL_END - VSYSCALL_START);
856
857         printk(KERN_INFO "Memory: %luk/%luk available (%ldk kernel code, "
858                                 "%ldk reserved, %ldk data, %ldk init)\n",
859                 (unsigned long) nr_free_pages() << (PAGE_SHIFT-10),
860                 max_pfn << (PAGE_SHIFT-10),
861                 codesize >> 10,
862                 reservedpages << (PAGE_SHIFT-10),
863                 datasize >> 10,
864                 initsize >> 10);
865 }
866
867 void free_init_pages(char *what, unsigned long begin, unsigned long end)
868 {
869         unsigned long addr = begin;
870
871         if (addr >= end)
872                 return;
873
874         /*
875          * If debugging page accesses then do not free this memory but
876          * mark them not present - any buggy init-section access will
877          * create a kernel page fault:
878          */
879 #ifdef CONFIG_DEBUG_PAGEALLOC
880         printk(KERN_INFO "debug: unmapping init memory %08lx..%08lx\n",
881                 begin, PAGE_ALIGN(end));
882         set_memory_np(begin, (end - begin) >> PAGE_SHIFT);
883 #else
884         printk(KERN_INFO "Freeing %s: %luk freed\n", what, (end - begin) >> 10);
885
886         for (; addr < end; addr += PAGE_SIZE) {
887                 ClearPageReserved(virt_to_page(addr));
888                 init_page_count(virt_to_page(addr));
889                 memset((void *)(addr & ~(PAGE_SIZE-1)),
890                         POISON_FREE_INITMEM, PAGE_SIZE);
891                 free_page(addr);
892                 totalram_pages++;
893         }
894 #endif
895 }
896
897 void free_initmem(void)
898 {
899         free_init_pages("unused kernel memory",
900                         (unsigned long)(&__init_begin),
901                         (unsigned long)(&__init_end));
902 }
903
904 #ifdef CONFIG_DEBUG_RODATA
905 const int rodata_test_data = 0xC3;
906 EXPORT_SYMBOL_GPL(rodata_test_data);
907
908 void mark_rodata_ro(void)
909 {
910         unsigned long start = PFN_ALIGN(_stext), end = PFN_ALIGN(__end_rodata);
911         unsigned long rodata_start =
912                 ((unsigned long)__start_rodata + PAGE_SIZE - 1) & PAGE_MASK;
913
914 #ifdef CONFIG_DYNAMIC_FTRACE
915         /* Dynamic tracing modifies the kernel text section */
916         start = rodata_start;
917 #endif
918
919         printk(KERN_INFO "Write protecting the kernel read-only data: %luk\n",
920                (end - start) >> 10);
921         set_memory_ro(start, (end - start) >> PAGE_SHIFT);
922
923         /*
924          * The rodata section (but not the kernel text!) should also be
925          * not-executable.
926          */
927         set_memory_nx(rodata_start, (end - rodata_start) >> PAGE_SHIFT);
928
929         rodata_test();
930
931 #ifdef CONFIG_CPA_DEBUG
932         printk(KERN_INFO "Testing CPA: undo %lx-%lx\n", start, end);
933         set_memory_rw(start, (end-start) >> PAGE_SHIFT);
934
935         printk(KERN_INFO "Testing CPA: again\n");
936         set_memory_ro(start, (end-start) >> PAGE_SHIFT);
937 #endif
938 }
939
940 #endif
941
942 #ifdef CONFIG_BLK_DEV_INITRD
943 void free_initrd_mem(unsigned long start, unsigned long end)
944 {
945         free_init_pages("initrd memory", start, end);
946 }
947 #endif
948
949 int __init reserve_bootmem_generic(unsigned long phys, unsigned long len,
950                                    int flags)
951 {
952 #ifdef CONFIG_NUMA
953         int nid, next_nid;
954         int ret;
955 #endif
956         unsigned long pfn = phys >> PAGE_SHIFT;
957
958         if (pfn >= max_pfn) {
959                 /*
960                  * This can happen with kdump kernels when accessing
961                  * firmware tables:
962                  */
963                 if (pfn < max_pfn_mapped)
964                         return -EFAULT;
965
966                 printk(KERN_ERR "reserve_bootmem: illegal reserve %lx %lu\n",
967                                 phys, len);
968                 return -EFAULT;
969         }
970
971         /* Should check here against the e820 map to avoid double free */
972 #ifdef CONFIG_NUMA
973         nid = phys_to_nid(phys);
974         next_nid = phys_to_nid(phys + len - 1);
975         if (nid == next_nid)
976                 ret = reserve_bootmem_node(NODE_DATA(nid), phys, len, flags);
977         else
978                 ret = reserve_bootmem(phys, len, flags);
979
980         if (ret != 0)
981                 return ret;
982
983 #else
984         reserve_bootmem(phys, len, BOOTMEM_DEFAULT);
985 #endif
986
987         if (phys+len <= MAX_DMA_PFN*PAGE_SIZE) {
988                 dma_reserve += len / PAGE_SIZE;
989                 set_dma_reserve(dma_reserve);
990         }
991
992         return 0;
993 }
994
995 int kern_addr_valid(unsigned long addr)
996 {
997         unsigned long above = ((long)addr) >> __VIRTUAL_MASK_SHIFT;
998         pgd_t *pgd;
999         pud_t *pud;
1000         pmd_t *pmd;
1001         pte_t *pte;
1002
1003         if (above != 0 && above != -1UL)
1004                 return 0;
1005
1006         pgd = pgd_offset_k(addr);
1007         if (pgd_none(*pgd))
1008                 return 0;
1009
1010         pud = pud_offset(pgd, addr);
1011         if (pud_none(*pud))
1012                 return 0;
1013
1014         pmd = pmd_offset(pud, addr);
1015         if (pmd_none(*pmd))
1016                 return 0;
1017
1018         if (pmd_large(*pmd))
1019                 return pfn_valid(pmd_pfn(*pmd));
1020
1021         pte = pte_offset_kernel(pmd, addr);
1022         if (pte_none(*pte))
1023                 return 0;
1024
1025         return pfn_valid(pte_pfn(*pte));
1026 }
1027
1028 /*
1029  * A pseudo VMA to allow ptrace access for the vsyscall page.  This only
1030  * covers the 64bit vsyscall page now. 32bit has a real VMA now and does
1031  * not need special handling anymore:
1032  */
1033 static struct vm_area_struct gate_vma = {
1034         .vm_start       = VSYSCALL_START,
1035         .vm_end         = VSYSCALL_START + (VSYSCALL_MAPPED_PAGES * PAGE_SIZE),
1036         .vm_page_prot   = PAGE_READONLY_EXEC,
1037         .vm_flags       = VM_READ | VM_EXEC
1038 };
1039
1040 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
1041 {
1042 #ifdef CONFIG_IA32_EMULATION
1043         if (test_tsk_thread_flag(tsk, TIF_IA32))
1044                 return NULL;
1045 #endif
1046         return &gate_vma;
1047 }
1048
1049 int in_gate_area(struct task_struct *task, unsigned long addr)
1050 {
1051         struct vm_area_struct *vma = get_gate_vma(task);
1052
1053         if (!vma)
1054                 return 0;
1055
1056         return (addr >= vma->vm_start) && (addr < vma->vm_end);
1057 }
1058
1059 /*
1060  * Use this when you have no reliable task/vma, typically from interrupt
1061  * context. It is less reliable than using the task's vma and may give
1062  * false positives:
1063  */
1064 int in_gate_area_no_task(unsigned long addr)
1065 {
1066         return (addr >= VSYSCALL_START) && (addr < VSYSCALL_END);
1067 }
1068
1069 const char *arch_vma_name(struct vm_area_struct *vma)
1070 {
1071         if (vma->vm_mm && vma->vm_start == (long)vma->vm_mm->context.vdso)
1072                 return "[vdso]";
1073         if (vma == &gate_vma)
1074                 return "[vsyscall]";
1075         return NULL;
1076 }
1077
1078 #ifdef CONFIG_SPARSEMEM_VMEMMAP
1079 /*
1080  * Initialise the sparsemem vmemmap using huge-pages at the PMD level.
1081  */
1082 static long __meminitdata addr_start, addr_end;
1083 static void __meminitdata *p_start, *p_end;
1084 static int __meminitdata node_start;
1085
1086 int __meminit
1087 vmemmap_populate(struct page *start_page, unsigned long size, int node)
1088 {
1089         unsigned long addr = (unsigned long)start_page;
1090         unsigned long end = (unsigned long)(start_page + size);
1091         unsigned long next;
1092         pgd_t *pgd;
1093         pud_t *pud;
1094         pmd_t *pmd;
1095
1096         for (; addr < end; addr = next) {
1097                 void *p = NULL;
1098
1099                 pgd = vmemmap_pgd_populate(addr, node);
1100                 if (!pgd)
1101                         return -ENOMEM;
1102
1103                 pud = vmemmap_pud_populate(pgd, addr, node);
1104                 if (!pud)
1105                         return -ENOMEM;
1106
1107                 if (!cpu_has_pse) {
1108                         next = (addr + PAGE_SIZE) & PAGE_MASK;
1109                         pmd = vmemmap_pmd_populate(pud, addr, node);
1110
1111                         if (!pmd)
1112                                 return -ENOMEM;
1113
1114                         p = vmemmap_pte_populate(pmd, addr, node);
1115
1116                         if (!p)
1117                                 return -ENOMEM;
1118
1119                         addr_end = addr + PAGE_SIZE;
1120                         p_end = p + PAGE_SIZE;
1121                 } else {
1122                         next = pmd_addr_end(addr, end);
1123
1124                         pmd = pmd_offset(pud, addr);
1125                         if (pmd_none(*pmd)) {
1126                                 pte_t entry;
1127
1128                                 p = vmemmap_alloc_block(PMD_SIZE, node);
1129                                 if (!p)
1130                                         return -ENOMEM;
1131
1132                                 entry = pfn_pte(__pa(p) >> PAGE_SHIFT,
1133                                                 PAGE_KERNEL_LARGE);
1134                                 set_pmd(pmd, __pmd(pte_val(entry)));
1135
1136                                 /* check to see if we have contiguous blocks */
1137                                 if (p_end != p || node_start != node) {
1138                                         if (p_start)
1139                                                 printk(KERN_DEBUG " [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1140                                                        addr_start, addr_end-1, p_start, p_end-1, node_start);
1141                                         addr_start = addr;
1142                                         node_start = node;
1143                                         p_start = p;
1144                                 }
1145
1146                                 addr_end = addr + PMD_SIZE;
1147                                 p_end = p + PMD_SIZE;
1148                         } else
1149                                 vmemmap_verify((pte_t *)pmd, node, addr, next);
1150                 }
1151
1152         }
1153         return 0;
1154 }
1155
1156 void __meminit vmemmap_populate_print_last(void)
1157 {
1158         if (p_start) {
1159                 printk(KERN_DEBUG " [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1160                         addr_start, addr_end-1, p_start, p_end-1, node_start);
1161                 p_start = NULL;
1162                 p_end = NULL;
1163                 node_start = 0;
1164         }
1165 }
1166 #endif