Merge branch 'x86/core' into core/percpu
[linux-2.6] / arch / x86 / mm / init_32.c
1 /*
2  *
3  *  Copyright (C) 1995  Linus Torvalds
4  *
5  *  Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
6  */
7
8 #include <linux/module.h>
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/hugetlb.h>
19 #include <linux/swap.h>
20 #include <linux/smp.h>
21 #include <linux/init.h>
22 #include <linux/highmem.h>
23 #include <linux/pagemap.h>
24 #include <linux/pci.h>
25 #include <linux/pfn.h>
26 #include <linux/poison.h>
27 #include <linux/bootmem.h>
28 #include <linux/slab.h>
29 #include <linux/proc_fs.h>
30 #include <linux/memory_hotplug.h>
31 #include <linux/initrd.h>
32 #include <linux/cpumask.h>
33
34 #include <asm/asm.h>
35 #include <asm/bios_ebda.h>
36 #include <asm/processor.h>
37 #include <asm/system.h>
38 #include <asm/uaccess.h>
39 #include <asm/pgtable.h>
40 #include <asm/dma.h>
41 #include <asm/fixmap.h>
42 #include <asm/e820.h>
43 #include <asm/apic.h>
44 #include <asm/bugs.h>
45 #include <asm/tlb.h>
46 #include <asm/tlbflush.h>
47 #include <asm/pgalloc.h>
48 #include <asm/sections.h>
49 #include <asm/paravirt.h>
50 #include <asm/setup.h>
51 #include <asm/cacheflush.h>
52
53 unsigned long max_low_pfn_mapped;
54 unsigned long max_pfn_mapped;
55
56 DEFINE_PER_CPU(struct mmu_gather, mmu_gathers);
57 unsigned long highstart_pfn, highend_pfn;
58
59 static noinline int do_test_wp_bit(void);
60
61
62 static unsigned long __initdata table_start;
63 static unsigned long __meminitdata table_end;
64 static unsigned long __meminitdata table_top;
65
66 static int __initdata after_init_bootmem;
67
68 static __init void *alloc_low_page(void)
69 {
70         unsigned long pfn = table_end++;
71         void *adr;
72
73         if (pfn >= table_top)
74                 panic("alloc_low_page: ran out of memory");
75
76         adr = __va(pfn * PAGE_SIZE);
77         memset(adr, 0, PAGE_SIZE);
78         return adr;
79 }
80
81 /*
82  * Creates a middle page table and puts a pointer to it in the
83  * given global directory entry. This only returns the gd entry
84  * in non-PAE compilation mode, since the middle layer is folded.
85  */
86 static pmd_t * __init one_md_table_init(pgd_t *pgd)
87 {
88         pud_t *pud;
89         pmd_t *pmd_table;
90
91 #ifdef CONFIG_X86_PAE
92         if (!(pgd_val(*pgd) & _PAGE_PRESENT)) {
93                 if (after_init_bootmem)
94                         pmd_table = (pmd_t *)alloc_bootmem_low_pages(PAGE_SIZE);
95                 else
96                         pmd_table = (pmd_t *)alloc_low_page();
97                 paravirt_alloc_pmd(&init_mm, __pa(pmd_table) >> PAGE_SHIFT);
98                 set_pgd(pgd, __pgd(__pa(pmd_table) | _PAGE_PRESENT));
99                 pud = pud_offset(pgd, 0);
100                 BUG_ON(pmd_table != pmd_offset(pud, 0));
101
102                 return pmd_table;
103         }
104 #endif
105         pud = pud_offset(pgd, 0);
106         pmd_table = pmd_offset(pud, 0);
107
108         return pmd_table;
109 }
110
111 /*
112  * Create a page table and place a pointer to it in a middle page
113  * directory entry:
114  */
115 static pte_t * __init one_page_table_init(pmd_t *pmd)
116 {
117         if (!(pmd_val(*pmd) & _PAGE_PRESENT)) {
118                 pte_t *page_table = NULL;
119
120                 if (after_init_bootmem) {
121 #ifdef CONFIG_DEBUG_PAGEALLOC
122                         page_table = (pte_t *) alloc_bootmem_pages(PAGE_SIZE);
123 #endif
124                         if (!page_table)
125                                 page_table =
126                                 (pte_t *)alloc_bootmem_low_pages(PAGE_SIZE);
127                 } else
128                         page_table = (pte_t *)alloc_low_page();
129
130                 paravirt_alloc_pte(&init_mm, __pa(page_table) >> PAGE_SHIFT);
131                 set_pmd(pmd, __pmd(__pa(page_table) | _PAGE_TABLE));
132                 BUG_ON(page_table != pte_offset_kernel(pmd, 0));
133         }
134
135         return pte_offset_kernel(pmd, 0);
136 }
137
138 pmd_t * __init populate_extra_pmd(unsigned long vaddr)
139 {
140         int pgd_idx = pgd_index(vaddr);
141         int pmd_idx = pmd_index(vaddr);
142
143         return one_md_table_init(swapper_pg_dir + pgd_idx) + pmd_idx;
144 }
145
146 pte_t * __init populate_extra_pte(unsigned long vaddr)
147 {
148         int pte_idx = pte_index(vaddr);
149         pmd_t *pmd;
150
151         pmd = populate_extra_pmd(vaddr);
152         return one_page_table_init(pmd) + pte_idx;
153 }
154
155 static pte_t *__init page_table_kmap_check(pte_t *pte, pmd_t *pmd,
156                                            unsigned long vaddr, pte_t *lastpte)
157 {
158 #ifdef CONFIG_HIGHMEM
159         /*
160          * Something (early fixmap) may already have put a pte
161          * page here, which causes the page table allocation
162          * to become nonlinear. Attempt to fix it, and if it
163          * is still nonlinear then we have to bug.
164          */
165         int pmd_idx_kmap_begin = fix_to_virt(FIX_KMAP_END) >> PMD_SHIFT;
166         int pmd_idx_kmap_end = fix_to_virt(FIX_KMAP_BEGIN) >> PMD_SHIFT;
167
168         if (pmd_idx_kmap_begin != pmd_idx_kmap_end
169             && (vaddr >> PMD_SHIFT) >= pmd_idx_kmap_begin
170             && (vaddr >> PMD_SHIFT) <= pmd_idx_kmap_end
171             && ((__pa(pte) >> PAGE_SHIFT) < table_start
172                 || (__pa(pte) >> PAGE_SHIFT) >= table_end)) {
173                 pte_t *newpte;
174                 int i;
175
176                 BUG_ON(after_init_bootmem);
177                 newpte = alloc_low_page();
178                 for (i = 0; i < PTRS_PER_PTE; i++)
179                         set_pte(newpte + i, pte[i]);
180
181                 paravirt_alloc_pte(&init_mm, __pa(newpte) >> PAGE_SHIFT);
182                 set_pmd(pmd, __pmd(__pa(newpte)|_PAGE_TABLE));
183                 BUG_ON(newpte != pte_offset_kernel(pmd, 0));
184                 __flush_tlb_all();
185
186                 paravirt_release_pte(__pa(pte) >> PAGE_SHIFT);
187                 pte = newpte;
188         }
189         BUG_ON(vaddr < fix_to_virt(FIX_KMAP_BEGIN - 1)
190                && vaddr > fix_to_virt(FIX_KMAP_END)
191                && lastpte && lastpte + PTRS_PER_PTE != pte);
192 #endif
193         return pte;
194 }
195
196 /*
197  * This function initializes a certain range of kernel virtual memory
198  * with new bootmem page tables, everywhere page tables are missing in
199  * the given range.
200  *
201  * NOTE: The pagetables are allocated contiguous on the physical space
202  * so we can cache the place of the first one and move around without
203  * checking the pgd every time.
204  */
205 static void __init
206 page_table_range_init(unsigned long start, unsigned long end, pgd_t *pgd_base)
207 {
208         int pgd_idx, pmd_idx;
209         unsigned long vaddr;
210         pgd_t *pgd;
211         pmd_t *pmd;
212         pte_t *pte = NULL;
213
214         vaddr = start;
215         pgd_idx = pgd_index(vaddr);
216         pmd_idx = pmd_index(vaddr);
217         pgd = pgd_base + pgd_idx;
218
219         for ( ; (pgd_idx < PTRS_PER_PGD) && (vaddr != end); pgd++, pgd_idx++) {
220                 pmd = one_md_table_init(pgd);
221                 pmd = pmd + pmd_index(vaddr);
222                 for (; (pmd_idx < PTRS_PER_PMD) && (vaddr != end);
223                                                         pmd++, pmd_idx++) {
224                         pte = page_table_kmap_check(one_page_table_init(pmd),
225                                                     pmd, vaddr, pte);
226
227                         vaddr += PMD_SIZE;
228                 }
229                 pmd_idx = 0;
230         }
231 }
232
233 static inline int is_kernel_text(unsigned long addr)
234 {
235         if (addr >= PAGE_OFFSET && addr <= (unsigned long)__init_end)
236                 return 1;
237         return 0;
238 }
239
240 /*
241  * This maps the physical memory to kernel virtual address space, a total
242  * of max_low_pfn pages, by creating page tables starting from address
243  * PAGE_OFFSET:
244  */
245 static void __init kernel_physical_mapping_init(pgd_t *pgd_base,
246                                                 unsigned long start_pfn,
247                                                 unsigned long end_pfn,
248                                                 int use_pse)
249 {
250         int pgd_idx, pmd_idx, pte_ofs;
251         unsigned long pfn;
252         pgd_t *pgd;
253         pmd_t *pmd;
254         pte_t *pte;
255         unsigned pages_2m, pages_4k;
256         int mapping_iter;
257
258         /*
259          * First iteration will setup identity mapping using large/small pages
260          * based on use_pse, with other attributes same as set by
261          * the early code in head_32.S
262          *
263          * Second iteration will setup the appropriate attributes (NX, GLOBAL..)
264          * as desired for the kernel identity mapping.
265          *
266          * This two pass mechanism conforms to the TLB app note which says:
267          *
268          *     "Software should not write to a paging-structure entry in a way
269          *      that would change, for any linear address, both the page size
270          *      and either the page frame or attributes."
271          */
272         mapping_iter = 1;
273
274         if (!cpu_has_pse)
275                 use_pse = 0;
276
277 repeat:
278         pages_2m = pages_4k = 0;
279         pfn = start_pfn;
280         pgd_idx = pgd_index((pfn<<PAGE_SHIFT) + PAGE_OFFSET);
281         pgd = pgd_base + pgd_idx;
282         for (; pgd_idx < PTRS_PER_PGD; pgd++, pgd_idx++) {
283                 pmd = one_md_table_init(pgd);
284
285                 if (pfn >= end_pfn)
286                         continue;
287 #ifdef CONFIG_X86_PAE
288                 pmd_idx = pmd_index((pfn<<PAGE_SHIFT) + PAGE_OFFSET);
289                 pmd += pmd_idx;
290 #else
291                 pmd_idx = 0;
292 #endif
293                 for (; pmd_idx < PTRS_PER_PMD && pfn < end_pfn;
294                      pmd++, pmd_idx++) {
295                         unsigned int addr = pfn * PAGE_SIZE + PAGE_OFFSET;
296
297                         /*
298                          * Map with big pages if possible, otherwise
299                          * create normal page tables:
300                          */
301                         if (use_pse) {
302                                 unsigned int addr2;
303                                 pgprot_t prot = PAGE_KERNEL_LARGE;
304                                 /*
305                                  * first pass will use the same initial
306                                  * identity mapping attribute + _PAGE_PSE.
307                                  */
308                                 pgprot_t init_prot =
309                                         __pgprot(PTE_IDENT_ATTR |
310                                                  _PAGE_PSE);
311
312                                 addr2 = (pfn + PTRS_PER_PTE-1) * PAGE_SIZE +
313                                         PAGE_OFFSET + PAGE_SIZE-1;
314
315                                 if (is_kernel_text(addr) ||
316                                     is_kernel_text(addr2))
317                                         prot = PAGE_KERNEL_LARGE_EXEC;
318
319                                 pages_2m++;
320                                 if (mapping_iter == 1)
321                                         set_pmd(pmd, pfn_pmd(pfn, init_prot));
322                                 else
323                                         set_pmd(pmd, pfn_pmd(pfn, prot));
324
325                                 pfn += PTRS_PER_PTE;
326                                 continue;
327                         }
328                         pte = one_page_table_init(pmd);
329
330                         pte_ofs = pte_index((pfn<<PAGE_SHIFT) + PAGE_OFFSET);
331                         pte += pte_ofs;
332                         for (; pte_ofs < PTRS_PER_PTE && pfn < end_pfn;
333                              pte++, pfn++, pte_ofs++, addr += PAGE_SIZE) {
334                                 pgprot_t prot = PAGE_KERNEL;
335                                 /*
336                                  * first pass will use the same initial
337                                  * identity mapping attribute.
338                                  */
339                                 pgprot_t init_prot = __pgprot(PTE_IDENT_ATTR);
340
341                                 if (is_kernel_text(addr))
342                                         prot = PAGE_KERNEL_EXEC;
343
344                                 pages_4k++;
345                                 if (mapping_iter == 1)
346                                         set_pte(pte, pfn_pte(pfn, init_prot));
347                                 else
348                                         set_pte(pte, pfn_pte(pfn, prot));
349                         }
350                 }
351         }
352         if (mapping_iter == 1) {
353                 /*
354                  * update direct mapping page count only in the first
355                  * iteration.
356                  */
357                 update_page_count(PG_LEVEL_2M, pages_2m);
358                 update_page_count(PG_LEVEL_4K, pages_4k);
359
360                 /*
361                  * local global flush tlb, which will flush the previous
362                  * mappings present in both small and large page TLB's.
363                  */
364                 __flush_tlb_all();
365
366                 /*
367                  * Second iteration will set the actual desired PTE attributes.
368                  */
369                 mapping_iter = 2;
370                 goto repeat;
371         }
372 }
373
374 /*
375  * devmem_is_allowed() checks to see if /dev/mem access to a certain address
376  * is valid. The argument is a physical page number.
377  *
378  *
379  * On x86, access has to be given to the first megabyte of ram because that area
380  * contains bios code and data regions used by X and dosemu and similar apps.
381  * Access has to be given to non-kernel-ram areas as well, these contain the PCI
382  * mmio resources as well as potential bios/acpi data regions.
383  */
384 int devmem_is_allowed(unsigned long pagenr)
385 {
386         if (pagenr <= 256)
387                 return 1;
388         if (iomem_is_exclusive(pagenr << PAGE_SHIFT))
389                 return 0;
390         if (!page_is_ram(pagenr))
391                 return 1;
392         return 0;
393 }
394
395 pte_t *kmap_pte;
396 pgprot_t kmap_prot;
397
398 static inline pte_t *kmap_get_fixmap_pte(unsigned long vaddr)
399 {
400         return pte_offset_kernel(pmd_offset(pud_offset(pgd_offset_k(vaddr),
401                         vaddr), vaddr), vaddr);
402 }
403
404 static void __init kmap_init(void)
405 {
406         unsigned long kmap_vstart;
407
408         /*
409          * Cache the first kmap pte:
410          */
411         kmap_vstart = __fix_to_virt(FIX_KMAP_BEGIN);
412         kmap_pte = kmap_get_fixmap_pte(kmap_vstart);
413
414         kmap_prot = PAGE_KERNEL;
415 }
416
417 #ifdef CONFIG_HIGHMEM
418 static void __init permanent_kmaps_init(pgd_t *pgd_base)
419 {
420         unsigned long vaddr;
421         pgd_t *pgd;
422         pud_t *pud;
423         pmd_t *pmd;
424         pte_t *pte;
425
426         vaddr = PKMAP_BASE;
427         page_table_range_init(vaddr, vaddr + PAGE_SIZE*LAST_PKMAP, pgd_base);
428
429         pgd = swapper_pg_dir + pgd_index(vaddr);
430         pud = pud_offset(pgd, vaddr);
431         pmd = pmd_offset(pud, vaddr);
432         pte = pte_offset_kernel(pmd, vaddr);
433         pkmap_page_table = pte;
434 }
435
436 static void __init add_one_highpage_init(struct page *page, int pfn)
437 {
438         ClearPageReserved(page);
439         init_page_count(page);
440         __free_page(page);
441         totalhigh_pages++;
442 }
443
444 struct add_highpages_data {
445         unsigned long start_pfn;
446         unsigned long end_pfn;
447 };
448
449 static int __init add_highpages_work_fn(unsigned long start_pfn,
450                                          unsigned long end_pfn, void *datax)
451 {
452         int node_pfn;
453         struct page *page;
454         unsigned long final_start_pfn, final_end_pfn;
455         struct add_highpages_data *data;
456
457         data = (struct add_highpages_data *)datax;
458
459         final_start_pfn = max(start_pfn, data->start_pfn);
460         final_end_pfn = min(end_pfn, data->end_pfn);
461         if (final_start_pfn >= final_end_pfn)
462                 return 0;
463
464         for (node_pfn = final_start_pfn; node_pfn < final_end_pfn;
465              node_pfn++) {
466                 if (!pfn_valid(node_pfn))
467                         continue;
468                 page = pfn_to_page(node_pfn);
469                 add_one_highpage_init(page, node_pfn);
470         }
471
472         return 0;
473
474 }
475
476 void __init add_highpages_with_active_regions(int nid, unsigned long start_pfn,
477                                               unsigned long end_pfn)
478 {
479         struct add_highpages_data data;
480
481         data.start_pfn = start_pfn;
482         data.end_pfn = end_pfn;
483
484         work_with_active_regions(nid, add_highpages_work_fn, &data);
485 }
486
487 #else
488 static inline void permanent_kmaps_init(pgd_t *pgd_base)
489 {
490 }
491 #endif /* CONFIG_HIGHMEM */
492
493 void __init native_pagetable_setup_start(pgd_t *base)
494 {
495         unsigned long pfn, va;
496         pgd_t *pgd;
497         pud_t *pud;
498         pmd_t *pmd;
499         pte_t *pte;
500
501         /*
502          * Remove any mappings which extend past the end of physical
503          * memory from the boot time page table:
504          */
505         for (pfn = max_low_pfn + 1; pfn < 1<<(32-PAGE_SHIFT); pfn++) {
506                 va = PAGE_OFFSET + (pfn<<PAGE_SHIFT);
507                 pgd = base + pgd_index(va);
508                 if (!pgd_present(*pgd))
509                         break;
510
511                 pud = pud_offset(pgd, va);
512                 pmd = pmd_offset(pud, va);
513                 if (!pmd_present(*pmd))
514                         break;
515
516                 pte = pte_offset_kernel(pmd, va);
517                 if (!pte_present(*pte))
518                         break;
519
520                 pte_clear(NULL, va, pte);
521         }
522         paravirt_alloc_pmd(&init_mm, __pa(base) >> PAGE_SHIFT);
523 }
524
525 void __init native_pagetable_setup_done(pgd_t *base)
526 {
527 }
528
529 /*
530  * Build a proper pagetable for the kernel mappings.  Up until this
531  * point, we've been running on some set of pagetables constructed by
532  * the boot process.
533  *
534  * If we're booting on native hardware, this will be a pagetable
535  * constructed in arch/x86/kernel/head_32.S.  The root of the
536  * pagetable will be swapper_pg_dir.
537  *
538  * If we're booting paravirtualized under a hypervisor, then there are
539  * more options: we may already be running PAE, and the pagetable may
540  * or may not be based in swapper_pg_dir.  In any case,
541  * paravirt_pagetable_setup_start() will set up swapper_pg_dir
542  * appropriately for the rest of the initialization to work.
543  *
544  * In general, pagetable_init() assumes that the pagetable may already
545  * be partially populated, and so it avoids stomping on any existing
546  * mappings.
547  */
548 static void __init early_ioremap_page_table_range_init(pgd_t *pgd_base)
549 {
550         unsigned long vaddr, end;
551
552         /*
553          * Fixed mappings, only the page table structure has to be
554          * created - mappings will be set by set_fixmap():
555          */
556         vaddr = __fix_to_virt(__end_of_fixed_addresses - 1) & PMD_MASK;
557         end = (FIXADDR_TOP + PMD_SIZE - 1) & PMD_MASK;
558         page_table_range_init(vaddr, end, pgd_base);
559         early_ioremap_reset();
560 }
561
562 static void __init pagetable_init(void)
563 {
564         pgd_t *pgd_base = swapper_pg_dir;
565
566         permanent_kmaps_init(pgd_base);
567 }
568
569 #ifdef CONFIG_ACPI_SLEEP
570 /*
571  * ACPI suspend needs this for resume, because things like the intel-agp
572  * driver might have split up a kernel 4MB mapping.
573  */
574 char swsusp_pg_dir[PAGE_SIZE]
575         __attribute__ ((aligned(PAGE_SIZE)));
576
577 static inline void save_pg_dir(void)
578 {
579         memcpy(swsusp_pg_dir, swapper_pg_dir, PAGE_SIZE);
580 }
581 #else /* !CONFIG_ACPI_SLEEP */
582 static inline void save_pg_dir(void)
583 {
584 }
585 #endif /* !CONFIG_ACPI_SLEEP */
586
587 void zap_low_mappings(void)
588 {
589         int i;
590
591         /*
592          * Zap initial low-memory mappings.
593          *
594          * Note that "pgd_clear()" doesn't do it for
595          * us, because pgd_clear() is a no-op on i386.
596          */
597         for (i = 0; i < KERNEL_PGD_BOUNDARY; i++) {
598 #ifdef CONFIG_X86_PAE
599                 set_pgd(swapper_pg_dir+i, __pgd(1 + __pa(empty_zero_page)));
600 #else
601                 set_pgd(swapper_pg_dir+i, __pgd(0));
602 #endif
603         }
604         flush_tlb_all();
605 }
606
607 int nx_enabled;
608
609 pteval_t __supported_pte_mask __read_mostly = ~(_PAGE_NX | _PAGE_GLOBAL | _PAGE_IOMAP);
610 EXPORT_SYMBOL_GPL(__supported_pte_mask);
611
612 #ifdef CONFIG_X86_PAE
613
614 static int disable_nx __initdata;
615
616 /*
617  * noexec = on|off
618  *
619  * Control non executable mappings.
620  *
621  * on      Enable
622  * off     Disable
623  */
624 static int __init noexec_setup(char *str)
625 {
626         if (!str || !strcmp(str, "on")) {
627                 if (cpu_has_nx) {
628                         __supported_pte_mask |= _PAGE_NX;
629                         disable_nx = 0;
630                 }
631         } else {
632                 if (!strcmp(str, "off")) {
633                         disable_nx = 1;
634                         __supported_pte_mask &= ~_PAGE_NX;
635                 } else {
636                         return -EINVAL;
637                 }
638         }
639
640         return 0;
641 }
642 early_param("noexec", noexec_setup);
643
644 static void __init set_nx(void)
645 {
646         unsigned int v[4], l, h;
647
648         if (cpu_has_pae && (cpuid_eax(0x80000000) > 0x80000001)) {
649                 cpuid(0x80000001, &v[0], &v[1], &v[2], &v[3]);
650
651                 if ((v[3] & (1 << 20)) && !disable_nx) {
652                         rdmsr(MSR_EFER, l, h);
653                         l |= EFER_NX;
654                         wrmsr(MSR_EFER, l, h);
655                         nx_enabled = 1;
656                         __supported_pte_mask |= _PAGE_NX;
657                 }
658         }
659 }
660 #endif
661
662 /* user-defined highmem size */
663 static unsigned int highmem_pages = -1;
664
665 /*
666  * highmem=size forces highmem to be exactly 'size' bytes.
667  * This works even on boxes that have no highmem otherwise.
668  * This also works to reduce highmem size on bigger boxes.
669  */
670 static int __init parse_highmem(char *arg)
671 {
672         if (!arg)
673                 return -EINVAL;
674
675         highmem_pages = memparse(arg, &arg) >> PAGE_SHIFT;
676         return 0;
677 }
678 early_param("highmem", parse_highmem);
679
680 #define MSG_HIGHMEM_TOO_BIG \
681         "highmem size (%luMB) is bigger than pages available (%luMB)!\n"
682
683 #define MSG_LOWMEM_TOO_SMALL \
684         "highmem size (%luMB) results in <64MB lowmem, ignoring it!\n"
685 /*
686  * All of RAM fits into lowmem - but if user wants highmem
687  * artificially via the highmem=x boot parameter then create
688  * it:
689  */
690 void __init lowmem_pfn_init(void)
691 {
692         /* max_low_pfn is 0, we already have early_res support */
693         max_low_pfn = max_pfn;
694
695         if (highmem_pages == -1)
696                 highmem_pages = 0;
697 #ifdef CONFIG_HIGHMEM
698         if (highmem_pages >= max_pfn) {
699                 printk(KERN_ERR MSG_HIGHMEM_TOO_BIG,
700                         pages_to_mb(highmem_pages), pages_to_mb(max_pfn));
701                 highmem_pages = 0;
702         }
703         if (highmem_pages) {
704                 if (max_low_pfn - highmem_pages < 64*1024*1024/PAGE_SIZE) {
705                         printk(KERN_ERR MSG_LOWMEM_TOO_SMALL,
706                                 pages_to_mb(highmem_pages));
707                         highmem_pages = 0;
708                 }
709                 max_low_pfn -= highmem_pages;
710         }
711 #else
712         if (highmem_pages)
713                 printk(KERN_ERR "ignoring highmem size on non-highmem kernel!\n");
714 #endif
715 }
716
717 #define MSG_HIGHMEM_TOO_SMALL \
718         "only %luMB highmem pages available, ignoring highmem size of %luMB!\n"
719
720 #define MSG_HIGHMEM_TRIMMED \
721         "Warning: only 4GB will be used. Use a HIGHMEM64G enabled kernel!\n"
722 /*
723  * We have more RAM than fits into lowmem - we try to put it into
724  * highmem, also taking the highmem=x boot parameter into account:
725  */
726 void __init highmem_pfn_init(void)
727 {
728         max_low_pfn = MAXMEM_PFN;
729
730         if (highmem_pages == -1)
731                 highmem_pages = max_pfn - MAXMEM_PFN;
732
733         if (highmem_pages + MAXMEM_PFN < max_pfn)
734                 max_pfn = MAXMEM_PFN + highmem_pages;
735
736         if (highmem_pages + MAXMEM_PFN > max_pfn) {
737                 printk(KERN_WARNING MSG_HIGHMEM_TOO_SMALL,
738                         pages_to_mb(max_pfn - MAXMEM_PFN),
739                         pages_to_mb(highmem_pages));
740                 highmem_pages = 0;
741         }
742 #ifndef CONFIG_HIGHMEM
743         /* Maximum memory usable is what is directly addressable */
744         printk(KERN_WARNING "Warning only %ldMB will be used.\n", MAXMEM>>20);
745         if (max_pfn > MAX_NONPAE_PFN)
746                 printk(KERN_WARNING "Use a HIGHMEM64G enabled kernel.\n");
747         else
748                 printk(KERN_WARNING "Use a HIGHMEM enabled kernel.\n");
749         max_pfn = MAXMEM_PFN;
750 #else /* !CONFIG_HIGHMEM */
751 #ifndef CONFIG_HIGHMEM64G
752         if (max_pfn > MAX_NONPAE_PFN) {
753                 max_pfn = MAX_NONPAE_PFN;
754                 printk(KERN_WARNING MSG_HIGHMEM_TRIMMED);
755         }
756 #endif /* !CONFIG_HIGHMEM64G */
757 #endif /* !CONFIG_HIGHMEM */
758 }
759
760 /*
761  * Determine low and high memory ranges:
762  */
763 void __init find_low_pfn_range(void)
764 {
765         /* it could update max_pfn */
766
767         if (max_pfn <= MAXMEM_PFN)
768                 lowmem_pfn_init();
769         else
770                 highmem_pfn_init();
771 }
772
773 #ifndef CONFIG_NEED_MULTIPLE_NODES
774 void __init initmem_init(unsigned long start_pfn,
775                                   unsigned long end_pfn)
776 {
777 #ifdef CONFIG_HIGHMEM
778         highstart_pfn = highend_pfn = max_pfn;
779         if (max_pfn > max_low_pfn)
780                 highstart_pfn = max_low_pfn;
781         memory_present(0, 0, highend_pfn);
782         e820_register_active_regions(0, 0, highend_pfn);
783         printk(KERN_NOTICE "%ldMB HIGHMEM available.\n",
784                 pages_to_mb(highend_pfn - highstart_pfn));
785         num_physpages = highend_pfn;
786         high_memory = (void *) __va(highstart_pfn * PAGE_SIZE - 1) + 1;
787 #else
788         memory_present(0, 0, max_low_pfn);
789         e820_register_active_regions(0, 0, max_low_pfn);
790         num_physpages = max_low_pfn;
791         high_memory = (void *) __va(max_low_pfn * PAGE_SIZE - 1) + 1;
792 #endif
793 #ifdef CONFIG_FLATMEM
794         max_mapnr = num_physpages;
795 #endif
796         printk(KERN_NOTICE "%ldMB LOWMEM available.\n",
797                         pages_to_mb(max_low_pfn));
798
799         setup_bootmem_allocator();
800 }
801 #endif /* !CONFIG_NEED_MULTIPLE_NODES */
802
803 static void __init zone_sizes_init(void)
804 {
805         unsigned long max_zone_pfns[MAX_NR_ZONES];
806         memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
807         max_zone_pfns[ZONE_DMA] =
808                 virt_to_phys((char *)MAX_DMA_ADDRESS) >> PAGE_SHIFT;
809         max_zone_pfns[ZONE_NORMAL] = max_low_pfn;
810 #ifdef CONFIG_HIGHMEM
811         max_zone_pfns[ZONE_HIGHMEM] = highend_pfn;
812 #endif
813
814         free_area_init_nodes(max_zone_pfns);
815 }
816
817 void __init setup_bootmem_allocator(void)
818 {
819         int i;
820         unsigned long bootmap_size, bootmap;
821         /*
822          * Initialize the boot-time allocator (with low memory only):
823          */
824         bootmap_size = bootmem_bootmap_pages(max_low_pfn)<<PAGE_SHIFT;
825         bootmap = find_e820_area(min_low_pfn<<PAGE_SHIFT,
826                                  max_pfn_mapped<<PAGE_SHIFT, bootmap_size,
827                                  PAGE_SIZE);
828         if (bootmap == -1L)
829                 panic("Cannot find bootmem map of size %ld\n", bootmap_size);
830         reserve_early(bootmap, bootmap + bootmap_size, "BOOTMAP");
831
832         /* don't touch min_low_pfn */
833         bootmap_size = init_bootmem_node(NODE_DATA(0), bootmap >> PAGE_SHIFT,
834                                          min_low_pfn, max_low_pfn);
835         printk(KERN_INFO "  mapped low ram: 0 - %08lx\n",
836                  max_pfn_mapped<<PAGE_SHIFT);
837         printk(KERN_INFO "  low ram: %08lx - %08lx\n",
838                  min_low_pfn<<PAGE_SHIFT, max_low_pfn<<PAGE_SHIFT);
839         printk(KERN_INFO "  bootmap %08lx - %08lx\n",
840                  bootmap, bootmap + bootmap_size);
841         for_each_online_node(i)
842                 free_bootmem_with_active_regions(i, max_low_pfn);
843         early_res_to_bootmem(0, max_low_pfn<<PAGE_SHIFT);
844
845         after_init_bootmem = 1;
846 }
847
848 static void __init find_early_table_space(unsigned long end, int use_pse)
849 {
850         unsigned long puds, pmds, ptes, tables, start;
851
852         puds = (end + PUD_SIZE - 1) >> PUD_SHIFT;
853         tables = roundup(puds * sizeof(pud_t), PAGE_SIZE);
854
855         pmds = (end + PMD_SIZE - 1) >> PMD_SHIFT;
856         tables += roundup(pmds * sizeof(pmd_t), PAGE_SIZE);
857
858         if (use_pse) {
859                 unsigned long extra;
860
861                 extra = end - ((end>>PMD_SHIFT) << PMD_SHIFT);
862                 extra += PMD_SIZE;
863                 ptes = (extra + PAGE_SIZE - 1) >> PAGE_SHIFT;
864         } else
865                 ptes = (end + PAGE_SIZE - 1) >> PAGE_SHIFT;
866
867         tables += roundup(ptes * sizeof(pte_t), PAGE_SIZE);
868
869         /* for fixmap */
870         tables += roundup(__end_of_fixed_addresses * sizeof(pte_t), PAGE_SIZE);
871
872         /*
873          * RED-PEN putting page tables only on node 0 could
874          * cause a hotspot and fill up ZONE_DMA. The page tables
875          * need roughly 0.5KB per GB.
876          */
877         start = 0x7000;
878         table_start = find_e820_area(start, max_pfn_mapped<<PAGE_SHIFT,
879                                         tables, PAGE_SIZE);
880         if (table_start == -1UL)
881                 panic("Cannot find space for the kernel page tables");
882
883         table_start >>= PAGE_SHIFT;
884         table_end = table_start;
885         table_top = table_start + (tables>>PAGE_SHIFT);
886
887         printk(KERN_DEBUG "kernel direct mapping tables up to %lx @ %lx-%lx\n",
888                 end, table_start << PAGE_SHIFT,
889                 (table_start << PAGE_SHIFT) + tables);
890 }
891
892 unsigned long __init_refok init_memory_mapping(unsigned long start,
893                                                 unsigned long end)
894 {
895         pgd_t *pgd_base = swapper_pg_dir;
896         unsigned long start_pfn, end_pfn;
897         unsigned long big_page_start;
898 #ifdef CONFIG_DEBUG_PAGEALLOC
899         /*
900          * For CONFIG_DEBUG_PAGEALLOC, identity mapping will use small pages.
901          * This will simplify cpa(), which otherwise needs to support splitting
902          * large pages into small in interrupt context, etc.
903          */
904         int use_pse = 0;
905 #else
906         int use_pse = cpu_has_pse;
907 #endif
908
909         /*
910          * Find space for the kernel direct mapping tables.
911          */
912         if (!after_init_bootmem)
913                 find_early_table_space(end, use_pse);
914
915 #ifdef CONFIG_X86_PAE
916         set_nx();
917         if (nx_enabled)
918                 printk(KERN_INFO "NX (Execute Disable) protection: active\n");
919 #endif
920
921         /* Enable PSE if available */
922         if (cpu_has_pse)
923                 set_in_cr4(X86_CR4_PSE);
924
925         /* Enable PGE if available */
926         if (cpu_has_pge) {
927                 set_in_cr4(X86_CR4_PGE);
928                 __supported_pte_mask |= _PAGE_GLOBAL;
929         }
930
931         /*
932          * Don't use a large page for the first 2/4MB of memory
933          * because there are often fixed size MTRRs in there
934          * and overlapping MTRRs into large pages can cause
935          * slowdowns.
936          */
937         big_page_start = PMD_SIZE;
938
939         if (start < big_page_start) {
940                 start_pfn = start >> PAGE_SHIFT;
941                 end_pfn = min(big_page_start>>PAGE_SHIFT, end>>PAGE_SHIFT);
942         } else {
943                 /* head is not big page alignment ? */
944                 start_pfn = start >> PAGE_SHIFT;
945                 end_pfn = ((start + (PMD_SIZE - 1))>>PMD_SHIFT)
946                                  << (PMD_SHIFT - PAGE_SHIFT);
947         }
948         if (start_pfn < end_pfn)
949                 kernel_physical_mapping_init(pgd_base, start_pfn, end_pfn, 0);
950
951         /* big page range */
952         start_pfn = ((start + (PMD_SIZE - 1))>>PMD_SHIFT)
953                          << (PMD_SHIFT - PAGE_SHIFT);
954         if (start_pfn < (big_page_start >> PAGE_SHIFT))
955                 start_pfn =  big_page_start >> PAGE_SHIFT;
956         end_pfn = (end>>PMD_SHIFT) << (PMD_SHIFT - PAGE_SHIFT);
957         if (start_pfn < end_pfn)
958                 kernel_physical_mapping_init(pgd_base, start_pfn, end_pfn,
959                                              use_pse);
960
961         /* tail is not big page alignment ? */
962         start_pfn = end_pfn;
963         if (start_pfn > (big_page_start>>PAGE_SHIFT)) {
964                 end_pfn = end >> PAGE_SHIFT;
965                 if (start_pfn < end_pfn)
966                         kernel_physical_mapping_init(pgd_base, start_pfn,
967                                                          end_pfn, 0);
968         }
969
970         early_ioremap_page_table_range_init(pgd_base);
971
972         load_cr3(swapper_pg_dir);
973
974         __flush_tlb_all();
975
976         if (!after_init_bootmem)
977                 reserve_early(table_start << PAGE_SHIFT,
978                                  table_end << PAGE_SHIFT, "PGTABLE");
979
980         if (!after_init_bootmem)
981                 early_memtest(start, end);
982
983         return end >> PAGE_SHIFT;
984 }
985
986
987 /*
988  * paging_init() sets up the page tables - note that the first 8MB are
989  * already mapped by head.S.
990  *
991  * This routines also unmaps the page at virtual kernel address 0, so
992  * that we can trap those pesky NULL-reference errors in the kernel.
993  */
994 void __init paging_init(void)
995 {
996         pagetable_init();
997
998         __flush_tlb_all();
999
1000         kmap_init();
1001
1002         /*
1003          * NOTE: at this point the bootmem allocator is fully available.
1004          */
1005         sparse_init();
1006         zone_sizes_init();
1007 }
1008
1009 /*
1010  * Test if the WP bit works in supervisor mode. It isn't supported on 386's
1011  * and also on some strange 486's. All 586+'s are OK. This used to involve
1012  * black magic jumps to work around some nasty CPU bugs, but fortunately the
1013  * switch to using exceptions got rid of all that.
1014  */
1015 static void __init test_wp_bit(void)
1016 {
1017         printk(KERN_INFO
1018   "Checking if this processor honours the WP bit even in supervisor mode...");
1019
1020         /* Any page-aligned address will do, the test is non-destructive */
1021         __set_fixmap(FIX_WP_TEST, __pa(&swapper_pg_dir), PAGE_READONLY);
1022         boot_cpu_data.wp_works_ok = do_test_wp_bit();
1023         clear_fixmap(FIX_WP_TEST);
1024
1025         if (!boot_cpu_data.wp_works_ok) {
1026                 printk(KERN_CONT "No.\n");
1027 #ifdef CONFIG_X86_WP_WORKS_OK
1028                 panic(
1029   "This kernel doesn't support CPU's with broken WP. Recompile it for a 386!");
1030 #endif
1031         } else {
1032                 printk(KERN_CONT "Ok.\n");
1033         }
1034 }
1035
1036 static struct kcore_list kcore_mem, kcore_vmalloc;
1037
1038 void __init mem_init(void)
1039 {
1040         int codesize, reservedpages, datasize, initsize;
1041         int tmp;
1042
1043         pci_iommu_alloc();
1044
1045 #ifdef CONFIG_FLATMEM
1046         BUG_ON(!mem_map);
1047 #endif
1048         /* this will put all low memory onto the freelists */
1049         totalram_pages += free_all_bootmem();
1050
1051         reservedpages = 0;
1052         for (tmp = 0; tmp < max_low_pfn; tmp++)
1053                 /*
1054                  * Only count reserved RAM pages:
1055                  */
1056                 if (page_is_ram(tmp) && PageReserved(pfn_to_page(tmp)))
1057                         reservedpages++;
1058
1059         set_highmem_pages_init();
1060
1061         codesize =  (unsigned long) &_etext - (unsigned long) &_text;
1062         datasize =  (unsigned long) &_edata - (unsigned long) &_etext;
1063         initsize =  (unsigned long) &__init_end - (unsigned long) &__init_begin;
1064
1065         kclist_add(&kcore_mem, __va(0), max_low_pfn << PAGE_SHIFT);
1066         kclist_add(&kcore_vmalloc, (void *)VMALLOC_START,
1067                    VMALLOC_END-VMALLOC_START);
1068
1069         printk(KERN_INFO "Memory: %luk/%luk available (%dk kernel code, "
1070                         "%dk reserved, %dk data, %dk init, %ldk highmem)\n",
1071                 (unsigned long) nr_free_pages() << (PAGE_SHIFT-10),
1072                 num_physpages << (PAGE_SHIFT-10),
1073                 codesize >> 10,
1074                 reservedpages << (PAGE_SHIFT-10),
1075                 datasize >> 10,
1076                 initsize >> 10,
1077                 (unsigned long) (totalhigh_pages << (PAGE_SHIFT-10))
1078                );
1079
1080         printk(KERN_INFO "virtual kernel memory layout:\n"
1081                 "    fixmap  : 0x%08lx - 0x%08lx   (%4ld kB)\n"
1082 #ifdef CONFIG_HIGHMEM
1083                 "    pkmap   : 0x%08lx - 0x%08lx   (%4ld kB)\n"
1084 #endif
1085                 "    vmalloc : 0x%08lx - 0x%08lx   (%4ld MB)\n"
1086                 "    lowmem  : 0x%08lx - 0x%08lx   (%4ld MB)\n"
1087                 "      .init : 0x%08lx - 0x%08lx   (%4ld kB)\n"
1088                 "      .data : 0x%08lx - 0x%08lx   (%4ld kB)\n"
1089                 "      .text : 0x%08lx - 0x%08lx   (%4ld kB)\n",
1090                 FIXADDR_START, FIXADDR_TOP,
1091                 (FIXADDR_TOP - FIXADDR_START) >> 10,
1092
1093 #ifdef CONFIG_HIGHMEM
1094                 PKMAP_BASE, PKMAP_BASE+LAST_PKMAP*PAGE_SIZE,
1095                 (LAST_PKMAP*PAGE_SIZE) >> 10,
1096 #endif
1097
1098                 VMALLOC_START, VMALLOC_END,
1099                 (VMALLOC_END - VMALLOC_START) >> 20,
1100
1101                 (unsigned long)__va(0), (unsigned long)high_memory,
1102                 ((unsigned long)high_memory - (unsigned long)__va(0)) >> 20,
1103
1104                 (unsigned long)&__init_begin, (unsigned long)&__init_end,
1105                 ((unsigned long)&__init_end -
1106                  (unsigned long)&__init_begin) >> 10,
1107
1108                 (unsigned long)&_etext, (unsigned long)&_edata,
1109                 ((unsigned long)&_edata - (unsigned long)&_etext) >> 10,
1110
1111                 (unsigned long)&_text, (unsigned long)&_etext,
1112                 ((unsigned long)&_etext - (unsigned long)&_text) >> 10);
1113
1114         /*
1115          * Check boundaries twice: Some fundamental inconsistencies can
1116          * be detected at build time already.
1117          */
1118 #define __FIXADDR_TOP (-PAGE_SIZE)
1119 #ifdef CONFIG_HIGHMEM
1120         BUILD_BUG_ON(PKMAP_BASE + LAST_PKMAP*PAGE_SIZE  > FIXADDR_START);
1121         BUILD_BUG_ON(VMALLOC_END                        > PKMAP_BASE);
1122 #endif
1123 #define high_memory (-128UL << 20)
1124         BUILD_BUG_ON(VMALLOC_START                      >= VMALLOC_END);
1125 #undef high_memory
1126 #undef __FIXADDR_TOP
1127
1128 #ifdef CONFIG_HIGHMEM
1129         BUG_ON(PKMAP_BASE + LAST_PKMAP*PAGE_SIZE        > FIXADDR_START);
1130         BUG_ON(VMALLOC_END                              > PKMAP_BASE);
1131 #endif
1132         BUG_ON(VMALLOC_START                            >= VMALLOC_END);
1133         BUG_ON((unsigned long)high_memory               > VMALLOC_START);
1134
1135         if (boot_cpu_data.wp_works_ok < 0)
1136                 test_wp_bit();
1137
1138         save_pg_dir();
1139         zap_low_mappings();
1140 }
1141
1142 #ifdef CONFIG_MEMORY_HOTPLUG
1143 int arch_add_memory(int nid, u64 start, u64 size)
1144 {
1145         struct pglist_data *pgdata = NODE_DATA(nid);
1146         struct zone *zone = pgdata->node_zones + ZONE_HIGHMEM;
1147         unsigned long start_pfn = start >> PAGE_SHIFT;
1148         unsigned long nr_pages = size >> PAGE_SHIFT;
1149
1150         return __add_pages(nid, zone, start_pfn, nr_pages);
1151 }
1152 #endif
1153
1154 /*
1155  * This function cannot be __init, since exceptions don't work in that
1156  * section.  Put this after the callers, so that it cannot be inlined.
1157  */
1158 static noinline int do_test_wp_bit(void)
1159 {
1160         char tmp_reg;
1161         int flag;
1162
1163         __asm__ __volatile__(
1164                 "       movb %0, %1     \n"
1165                 "1:     movb %1, %0     \n"
1166                 "       xorl %2, %2     \n"
1167                 "2:                     \n"
1168                 _ASM_EXTABLE(1b,2b)
1169                 :"=m" (*(char *)fix_to_virt(FIX_WP_TEST)),
1170                  "=q" (tmp_reg),
1171                  "=r" (flag)
1172                 :"2" (1)
1173                 :"memory");
1174
1175         return flag;
1176 }
1177
1178 #ifdef CONFIG_DEBUG_RODATA
1179 const int rodata_test_data = 0xC3;
1180 EXPORT_SYMBOL_GPL(rodata_test_data);
1181
1182 void mark_rodata_ro(void)
1183 {
1184         unsigned long start = PFN_ALIGN(_text);
1185         unsigned long size = PFN_ALIGN(_etext) - start;
1186
1187 #ifndef CONFIG_DYNAMIC_FTRACE
1188         /* Dynamic tracing modifies the kernel text section */
1189         set_pages_ro(virt_to_page(start), size >> PAGE_SHIFT);
1190         printk(KERN_INFO "Write protecting the kernel text: %luk\n",
1191                 size >> 10);
1192
1193 #ifdef CONFIG_CPA_DEBUG
1194         printk(KERN_INFO "Testing CPA: Reverting %lx-%lx\n",
1195                 start, start+size);
1196         set_pages_rw(virt_to_page(start), size>>PAGE_SHIFT);
1197
1198         printk(KERN_INFO "Testing CPA: write protecting again\n");
1199         set_pages_ro(virt_to_page(start), size>>PAGE_SHIFT);
1200 #endif
1201 #endif /* CONFIG_DYNAMIC_FTRACE */
1202
1203         start += size;
1204         size = (unsigned long)__end_rodata - start;
1205         set_pages_ro(virt_to_page(start), size >> PAGE_SHIFT);
1206         printk(KERN_INFO "Write protecting the kernel read-only data: %luk\n",
1207                 size >> 10);
1208         rodata_test();
1209
1210 #ifdef CONFIG_CPA_DEBUG
1211         printk(KERN_INFO "Testing CPA: undo %lx-%lx\n", start, start + size);
1212         set_pages_rw(virt_to_page(start), size >> PAGE_SHIFT);
1213
1214         printk(KERN_INFO "Testing CPA: write protecting again\n");
1215         set_pages_ro(virt_to_page(start), size >> PAGE_SHIFT);
1216 #endif
1217 }
1218 #endif
1219
1220 #ifdef CONFIG_BLK_DEV_INITRD
1221 void free_initrd_mem(unsigned long start, unsigned long end)
1222 {
1223         free_init_pages("initrd memory", start, end);
1224 }
1225 #endif
1226
1227 int __init reserve_bootmem_generic(unsigned long phys, unsigned long len,
1228                                    int flags)
1229 {
1230         return reserve_bootmem(phys, len, flags);
1231 }