cd8d6732613834bfdfcbca2df6b9604e20393757
[linux-2.6.git] / 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 static pte_t *__init page_table_kmap_check(pte_t *pte, pmd_t *pmd,
139                                            unsigned long vaddr, pte_t *lastpte)
140 {
141 #ifdef CONFIG_HIGHMEM
142         /*
143          * Something (early fixmap) may already have put a pte
144          * page here, which causes the page table allocation
145          * to become nonlinear. Attempt to fix it, and if it
146          * is still nonlinear then we have to bug.
147          */
148         int pmd_idx_kmap_begin = fix_to_virt(FIX_KMAP_END) >> PMD_SHIFT;
149         int pmd_idx_kmap_end = fix_to_virt(FIX_KMAP_BEGIN) >> PMD_SHIFT;
150
151         if (pmd_idx_kmap_begin != pmd_idx_kmap_end
152             && (vaddr >> PMD_SHIFT) >= pmd_idx_kmap_begin
153             && (vaddr >> PMD_SHIFT) <= pmd_idx_kmap_end
154             && ((__pa(pte) >> PAGE_SHIFT) < table_start
155                 || (__pa(pte) >> PAGE_SHIFT) >= table_end)) {
156                 pte_t *newpte;
157                 int i;
158
159                 BUG_ON(after_init_bootmem);
160                 newpte = alloc_low_page();
161                 for (i = 0; i < PTRS_PER_PTE; i++)
162                         set_pte(newpte + i, pte[i]);
163
164                 paravirt_alloc_pte(&init_mm, __pa(newpte) >> PAGE_SHIFT);
165                 set_pmd(pmd, __pmd(__pa(newpte)|_PAGE_TABLE));
166                 BUG_ON(newpte != pte_offset_kernel(pmd, 0));
167                 __flush_tlb_all();
168
169                 paravirt_release_pte(__pa(pte) >> PAGE_SHIFT);
170                 pte = newpte;
171         }
172         BUG_ON(vaddr < fix_to_virt(FIX_KMAP_BEGIN - 1)
173                && vaddr > fix_to_virt(FIX_KMAP_END)
174                && lastpte && lastpte + PTRS_PER_PTE != pte);
175 #endif
176         return pte;
177 }
178
179 /*
180  * This function initializes a certain range of kernel virtual memory
181  * with new bootmem page tables, everywhere page tables are missing in
182  * the given range.
183  *
184  * NOTE: The pagetables are allocated contiguous on the physical space
185  * so we can cache the place of the first one and move around without
186  * checking the pgd every time.
187  */
188 static void __init
189 page_table_range_init(unsigned long start, unsigned long end, pgd_t *pgd_base)
190 {
191         int pgd_idx, pmd_idx;
192         unsigned long vaddr;
193         pgd_t *pgd;
194         pmd_t *pmd;
195         pte_t *pte = NULL;
196
197         vaddr = start;
198         pgd_idx = pgd_index(vaddr);
199         pmd_idx = pmd_index(vaddr);
200         pgd = pgd_base + pgd_idx;
201
202         for ( ; (pgd_idx < PTRS_PER_PGD) && (vaddr != end); pgd++, pgd_idx++) {
203                 pmd = one_md_table_init(pgd);
204                 pmd = pmd + pmd_index(vaddr);
205                 for (; (pmd_idx < PTRS_PER_PMD) && (vaddr != end);
206                                                         pmd++, pmd_idx++) {
207                         pte = page_table_kmap_check(one_page_table_init(pmd),
208                                                     pmd, vaddr, pte);
209
210                         vaddr += PMD_SIZE;
211                 }
212                 pmd_idx = 0;
213         }
214 }
215
216 static inline int is_kernel_text(unsigned long addr)
217 {
218         if (addr >= PAGE_OFFSET && addr <= (unsigned long)__init_end)
219                 return 1;
220         return 0;
221 }
222
223 /*
224  * This maps the physical memory to kernel virtual address space, a total
225  * of max_low_pfn pages, by creating page tables starting from address
226  * PAGE_OFFSET:
227  */
228 static void __init kernel_physical_mapping_init(pgd_t *pgd_base,
229                                                 unsigned long start_pfn,
230                                                 unsigned long end_pfn,
231                                                 int use_pse)
232 {
233         int pgd_idx, pmd_idx, pte_ofs;
234         unsigned long pfn;
235         pgd_t *pgd;
236         pmd_t *pmd;
237         pte_t *pte;
238         unsigned pages_2m, pages_4k;
239         int mapping_iter;
240
241         /*
242          * First iteration will setup identity mapping using large/small pages
243          * based on use_pse, with other attributes same as set by
244          * the early code in head_32.S
245          *
246          * Second iteration will setup the appropriate attributes (NX, GLOBAL..)
247          * as desired for the kernel identity mapping.
248          *
249          * This two pass mechanism conforms to the TLB app note which says:
250          *
251          *     "Software should not write to a paging-structure entry in a way
252          *      that would change, for any linear address, both the page size
253          *      and either the page frame or attributes."
254          */
255         mapping_iter = 1;
256
257         if (!cpu_has_pse)
258                 use_pse = 0;
259
260 repeat:
261         pages_2m = pages_4k = 0;
262         pfn = start_pfn;
263         pgd_idx = pgd_index((pfn<<PAGE_SHIFT) + PAGE_OFFSET);
264         pgd = pgd_base + pgd_idx;
265         for (; pgd_idx < PTRS_PER_PGD; pgd++, pgd_idx++) {
266                 pmd = one_md_table_init(pgd);
267
268                 if (pfn >= end_pfn)
269                         continue;
270 #ifdef CONFIG_X86_PAE
271                 pmd_idx = pmd_index((pfn<<PAGE_SHIFT) + PAGE_OFFSET);
272                 pmd += pmd_idx;
273 #else
274                 pmd_idx = 0;
275 #endif
276                 for (; pmd_idx < PTRS_PER_PMD && pfn < end_pfn;
277                      pmd++, pmd_idx++) {
278                         unsigned int addr = pfn * PAGE_SIZE + PAGE_OFFSET;
279
280                         /*
281                          * Map with big pages if possible, otherwise
282                          * create normal page tables:
283                          */
284                         if (use_pse) {
285                                 unsigned int addr2;
286                                 pgprot_t prot = PAGE_KERNEL_LARGE;
287                                 /*
288                                  * first pass will use the same initial
289                                  * identity mapping attribute + _PAGE_PSE.
290                                  */
291                                 pgprot_t init_prot =
292                                         __pgprot(PTE_IDENT_ATTR |
293                                                  _PAGE_PSE);
294
295                                 addr2 = (pfn + PTRS_PER_PTE-1) * PAGE_SIZE +
296                                         PAGE_OFFSET + PAGE_SIZE-1;
297
298                                 if (is_kernel_text(addr) ||
299                                     is_kernel_text(addr2))
300                                         prot = PAGE_KERNEL_LARGE_EXEC;
301
302                                 pages_2m++;
303                                 if (mapping_iter == 1)
304                                         set_pmd(pmd, pfn_pmd(pfn, init_prot));
305                                 else
306                                         set_pmd(pmd, pfn_pmd(pfn, prot));
307
308                                 pfn += PTRS_PER_PTE;
309                                 continue;
310                         }
311                         pte = one_page_table_init(pmd);
312
313                         pte_ofs = pte_index((pfn<<PAGE_SHIFT) + PAGE_OFFSET);
314                         pte += pte_ofs;
315                         for (; pte_ofs < PTRS_PER_PTE && pfn < end_pfn;
316                              pte++, pfn++, pte_ofs++, addr += PAGE_SIZE) {
317                                 pgprot_t prot = PAGE_KERNEL;
318                                 /*
319                                  * first pass will use the same initial
320                                  * identity mapping attribute.
321                                  */
322                                 pgprot_t init_prot = __pgprot(PTE_IDENT_ATTR);
323
324                                 if (is_kernel_text(addr))
325                                         prot = PAGE_KERNEL_EXEC;
326
327                                 pages_4k++;
328                                 if (mapping_iter == 1)
329                                         set_pte(pte, pfn_pte(pfn, init_prot));
330                                 else
331                                         set_pte(pte, pfn_pte(pfn, prot));
332                         }
333                 }
334         }
335         if (mapping_iter == 1) {
336                 /*
337                  * update direct mapping page count only in the first
338                  * iteration.
339                  */
340                 update_page_count(PG_LEVEL_2M, pages_2m);
341                 update_page_count(PG_LEVEL_4K, pages_4k);
342
343                 /*
344                  * local global flush tlb, which will flush the previous
345                  * mappings present in both small and large page TLB's.
346                  */
347                 __flush_tlb_all();
348
349                 /*
350                  * Second iteration will set the actual desired PTE attributes.
351                  */
352                 mapping_iter = 2;
353                 goto repeat;
354         }
355 }
356
357 /*
358  * devmem_is_allowed() checks to see if /dev/mem access to a certain address
359  * is valid. The argument is a physical page number.
360  *
361  *
362  * On x86, access has to be given to the first megabyte of ram because that area
363  * contains bios code and data regions used by X and dosemu and similar apps.
364  * Access has to be given to non-kernel-ram areas as well, these contain the PCI
365  * mmio resources as well as potential bios/acpi data regions.
366  */
367 int devmem_is_allowed(unsigned long pagenr)
368 {
369         if (pagenr <= 256)
370                 return 1;
371         if (iomem_is_exclusive(pagenr << PAGE_SHIFT))
372                 return 0;
373         if (!page_is_ram(pagenr))
374                 return 1;
375         return 0;
376 }
377
378 pte_t *kmap_pte;
379 pgprot_t kmap_prot;
380
381 static inline pte_t *kmap_get_fixmap_pte(unsigned long vaddr)
382 {
383         return pte_offset_kernel(pmd_offset(pud_offset(pgd_offset_k(vaddr),
384                         vaddr), vaddr), vaddr);
385 }
386
387 static void __init kmap_init(void)
388 {
389         unsigned long kmap_vstart;
390
391         /*
392          * Cache the first kmap pte:
393          */
394         kmap_vstart = __fix_to_virt(FIX_KMAP_BEGIN);
395         kmap_pte = kmap_get_fixmap_pte(kmap_vstart);
396
397         kmap_prot = PAGE_KERNEL;
398 }
399
400 #ifdef CONFIG_HIGHMEM
401 static void __init permanent_kmaps_init(pgd_t *pgd_base)
402 {
403         unsigned long vaddr;
404         pgd_t *pgd;
405         pud_t *pud;
406         pmd_t *pmd;
407         pte_t *pte;
408
409         vaddr = PKMAP_BASE;
410         page_table_range_init(vaddr, vaddr + PAGE_SIZE*LAST_PKMAP, pgd_base);
411
412         pgd = swapper_pg_dir + pgd_index(vaddr);
413         pud = pud_offset(pgd, vaddr);
414         pmd = pmd_offset(pud, vaddr);
415         pte = pte_offset_kernel(pmd, vaddr);
416         pkmap_page_table = pte;
417 }
418
419 static void __init add_one_highpage_init(struct page *page, int pfn)
420 {
421         ClearPageReserved(page);
422         init_page_count(page);
423         __free_page(page);
424         totalhigh_pages++;
425 }
426
427 struct add_highpages_data {
428         unsigned long start_pfn;
429         unsigned long end_pfn;
430 };
431
432 static int __init add_highpages_work_fn(unsigned long start_pfn,
433                                          unsigned long end_pfn, void *datax)
434 {
435         int node_pfn;
436         struct page *page;
437         unsigned long final_start_pfn, final_end_pfn;
438         struct add_highpages_data *data;
439
440         data = (struct add_highpages_data *)datax;
441
442         final_start_pfn = max(start_pfn, data->start_pfn);
443         final_end_pfn = min(end_pfn, data->end_pfn);
444         if (final_start_pfn >= final_end_pfn)
445                 return 0;
446
447         for (node_pfn = final_start_pfn; node_pfn < final_end_pfn;
448              node_pfn++) {
449                 if (!pfn_valid(node_pfn))
450                         continue;
451                 page = pfn_to_page(node_pfn);
452                 add_one_highpage_init(page, node_pfn);
453         }
454
455         return 0;
456
457 }
458
459 void __init add_highpages_with_active_regions(int nid, unsigned long start_pfn,
460                                               unsigned long end_pfn)
461 {
462         struct add_highpages_data data;
463
464         data.start_pfn = start_pfn;
465         data.end_pfn = end_pfn;
466
467         work_with_active_regions(nid, add_highpages_work_fn, &data);
468 }
469
470 #ifndef CONFIG_NUMA
471 static void __init set_highmem_pages_init(void)
472 {
473         add_highpages_with_active_regions(0, highstart_pfn, highend_pfn);
474
475         totalram_pages += totalhigh_pages;
476 }
477 #endif /* !CONFIG_NUMA */
478
479 #else
480 static inline void permanent_kmaps_init(pgd_t *pgd_base)
481 {
482 }
483 static inline void set_highmem_pages_init(void)
484 {
485 }
486 #endif /* CONFIG_HIGHMEM */
487
488 void __init native_pagetable_setup_start(pgd_t *base)
489 {
490         unsigned long pfn, va;
491         pgd_t *pgd;
492         pud_t *pud;
493         pmd_t *pmd;
494         pte_t *pte;
495
496         /*
497          * Remove any mappings which extend past the end of physical
498          * memory from the boot time page table:
499          */
500         for (pfn = max_low_pfn + 1; pfn < 1<<(32-PAGE_SHIFT); pfn++) {
501                 va = PAGE_OFFSET + (pfn<<PAGE_SHIFT);
502                 pgd = base + pgd_index(va);
503                 if (!pgd_present(*pgd))
504                         break;
505
506                 pud = pud_offset(pgd, va);
507                 pmd = pmd_offset(pud, va);
508                 if (!pmd_present(*pmd))
509                         break;
510
511                 pte = pte_offset_kernel(pmd, va);
512                 if (!pte_present(*pte))
513                         break;
514
515                 pte_clear(NULL, va, pte);
516         }
517         paravirt_alloc_pmd(&init_mm, __pa(base) >> PAGE_SHIFT);
518 }
519
520 void __init native_pagetable_setup_done(pgd_t *base)
521 {
522 }
523
524 /*
525  * Build a proper pagetable for the kernel mappings.  Up until this
526  * point, we've been running on some set of pagetables constructed by
527  * the boot process.
528  *
529  * If we're booting on native hardware, this will be a pagetable
530  * constructed in arch/x86/kernel/head_32.S.  The root of the
531  * pagetable will be swapper_pg_dir.
532  *
533  * If we're booting paravirtualized under a hypervisor, then there are
534  * more options: we may already be running PAE, and the pagetable may
535  * or may not be based in swapper_pg_dir.  In any case,
536  * paravirt_pagetable_setup_start() will set up swapper_pg_dir
537  * appropriately for the rest of the initialization to work.
538  *
539  * In general, pagetable_init() assumes that the pagetable may already
540  * be partially populated, and so it avoids stomping on any existing
541  * mappings.
542  */
543 static void __init early_ioremap_page_table_range_init(pgd_t *pgd_base)
544 {
545         unsigned long vaddr, end;
546
547         /*
548          * Fixed mappings, only the page table structure has to be
549          * created - mappings will be set by set_fixmap():
550          */
551         vaddr = __fix_to_virt(__end_of_fixed_addresses - 1) & PMD_MASK;
552         end = (FIXADDR_TOP + PMD_SIZE - 1) & PMD_MASK;
553         page_table_range_init(vaddr, end, pgd_base);
554         early_ioremap_reset();
555 }
556
557 static void __init pagetable_init(void)
558 {
559         pgd_t *pgd_base = swapper_pg_dir;
560
561         permanent_kmaps_init(pgd_base);
562 }
563
564 #ifdef CONFIG_ACPI_SLEEP
565 /*
566  * ACPI suspend needs this for resume, because things like the intel-agp
567  * driver might have split up a kernel 4MB mapping.
568  */
569 char swsusp_pg_dir[PAGE_SIZE]
570         __attribute__ ((aligned(PAGE_SIZE)));
571
572 static inline void save_pg_dir(void)
573 {
574         memcpy(swsusp_pg_dir, swapper_pg_dir, PAGE_SIZE);
575 }
576 #else /* !CONFIG_ACPI_SLEEP */
577 static inline void save_pg_dir(void)
578 {
579 }
580 #endif /* !CONFIG_ACPI_SLEEP */
581
582 void zap_low_mappings(void)
583 {
584         int i;
585
586         /*
587          * Zap initial low-memory mappings.
588          *
589          * Note that "pgd_clear()" doesn't do it for
590          * us, because pgd_clear() is a no-op on i386.
591          */
592         for (i = 0; i < KERNEL_PGD_BOUNDARY; i++) {
593 #ifdef CONFIG_X86_PAE
594                 set_pgd(swapper_pg_dir+i, __pgd(1 + __pa(empty_zero_page)));
595 #else
596                 set_pgd(swapper_pg_dir+i, __pgd(0));
597 #endif
598         }
599         flush_tlb_all();
600 }
601
602 int nx_enabled;
603
604 pteval_t __supported_pte_mask __read_mostly = ~(_PAGE_NX | _PAGE_GLOBAL | _PAGE_IOMAP);
605 EXPORT_SYMBOL_GPL(__supported_pte_mask);
606
607 #ifdef CONFIG_X86_PAE
608
609 static int disable_nx __initdata;
610
611 /*
612  * noexec = on|off
613  *
614  * Control non executable mappings.
615  *
616  * on      Enable
617  * off     Disable
618  */
619 static int __init noexec_setup(char *str)
620 {
621         if (!str || !strcmp(str, "on")) {
622                 if (cpu_has_nx) {
623                         __supported_pte_mask |= _PAGE_NX;
624                         disable_nx = 0;
625                 }
626         } else {
627                 if (!strcmp(str, "off")) {
628                         disable_nx = 1;
629                         __supported_pte_mask &= ~_PAGE_NX;
630                 } else {
631                         return -EINVAL;
632                 }
633         }
634
635         return 0;
636 }
637 early_param("noexec", noexec_setup);
638
639 static void __init set_nx(void)
640 {
641         unsigned int v[4], l, h;
642
643         if (cpu_has_pae && (cpuid_eax(0x80000000) > 0x80000001)) {
644                 cpuid(0x80000001, &v[0], &v[1], &v[2], &v[3]);
645
646                 if ((v[3] & (1 << 20)) && !disable_nx) {
647                         rdmsr(MSR_EFER, l, h);
648                         l |= EFER_NX;
649                         wrmsr(MSR_EFER, l, h);
650                         nx_enabled = 1;
651                         __supported_pte_mask |= _PAGE_NX;
652                 }
653         }
654 }
655 #endif
656
657 /* user-defined highmem size */
658 static unsigned int highmem_pages = -1;
659
660 /*
661  * highmem=size forces highmem to be exactly 'size' bytes.
662  * This works even on boxes that have no highmem otherwise.
663  * This also works to reduce highmem size on bigger boxes.
664  */
665 static int __init parse_highmem(char *arg)
666 {
667         if (!arg)
668                 return -EINVAL;
669
670         highmem_pages = memparse(arg, &arg) >> PAGE_SHIFT;
671         return 0;
672 }
673 early_param("highmem", parse_highmem);
674
675 #define MSG_HIGHMEM_TOO_BIG \
676         "highmem size (%luMB) is bigger than pages available (%luMB)!\n"
677
678 #define MSG_LOWMEM_TOO_SMALL \
679         "highmem size (%luMB) results in <64MB lowmem, ignoring it!\n"
680 /*
681  * All of RAM fits into lowmem - but if user wants highmem
682  * artificially via the highmem=x boot parameter then create
683  * it:
684  */
685 void __init lowmem_pfn_init(void)
686 {
687         /* max_low_pfn is 0, we already have early_res support */
688         max_low_pfn = max_pfn;
689
690         if (highmem_pages == -1)
691                 highmem_pages = 0;
692 #ifdef CONFIG_HIGHMEM
693         if (highmem_pages >= max_pfn) {
694                 printk(KERN_ERR MSG_HIGHMEM_TOO_BIG,
695                         pages_to_mb(highmem_pages), pages_to_mb(max_pfn));
696                 highmem_pages = 0;
697         }
698         if (highmem_pages) {
699                 if (max_low_pfn - highmem_pages < 64*1024*1024/PAGE_SIZE) {
700                         printk(KERN_ERR MSG_LOWMEM_TOO_SMALL,
701                                 pages_to_mb(highmem_pages));
702                         highmem_pages = 0;
703                 }
704                 max_low_pfn -= highmem_pages;
705         }
706 #else
707         if (highmem_pages)
708                 printk(KERN_ERR "ignoring highmem size on non-highmem kernel!\n");
709 #endif
710 }
711
712 #define MSG_HIGHMEM_TOO_SMALL \
713         "only %luMB highmem pages available, ignoring highmem size of %luMB!\n"
714
715 #define MSG_HIGHMEM_TRIMMED \
716         "Warning: only 4GB will be used. Use a HIGHMEM64G enabled kernel!\n"
717 /*
718  * We have more RAM than fits into lowmem - we try to put it into
719  * highmem, also taking the highmem=x boot parameter into account:
720  */
721 void __init highmem_pfn_init(void)
722 {
723         max_low_pfn = MAXMEM_PFN;
724
725         if (highmem_pages == -1)
726                 highmem_pages = max_pfn - MAXMEM_PFN;
727
728         if (highmem_pages + MAXMEM_PFN < max_pfn)
729                 max_pfn = MAXMEM_PFN + highmem_pages;
730
731         if (highmem_pages + MAXMEM_PFN > max_pfn) {
732                 printk(KERN_WARNING MSG_HIGHMEM_TOO_SMALL,
733                         pages_to_mb(max_pfn - MAXMEM_PFN),
734                         pages_to_mb(highmem_pages));
735                 highmem_pages = 0;
736         }
737 #ifndef CONFIG_HIGHMEM
738         /* Maximum memory usable is what is directly addressable */
739         printk(KERN_WARNING "Warning only %ldMB will be used.\n", MAXMEM>>20);
740         if (max_pfn > MAX_NONPAE_PFN)
741                 printk(KERN_WARNING "Use a HIGHMEM64G enabled kernel.\n");
742         else
743                 printk(KERN_WARNING "Use a HIGHMEM enabled kernel.\n");
744         max_pfn = MAXMEM_PFN;
745 #else /* !CONFIG_HIGHMEM */
746 #ifndef CONFIG_HIGHMEM64G
747         if (max_pfn > MAX_NONPAE_PFN) {
748                 max_pfn = MAX_NONPAE_PFN;
749                 printk(KERN_WARNING MSG_HIGHMEM_TRIMMED);
750         }
751 #endif /* !CONFIG_HIGHMEM64G */
752 #endif /* !CONFIG_HIGHMEM */
753 }
754
755 /*
756  * Determine low and high memory ranges:
757  */
758 void __init find_low_pfn_range(void)
759 {
760         /* it could update max_pfn */
761
762         if (max_pfn <= MAXMEM_PFN)
763                 lowmem_pfn_init();
764         else
765                 highmem_pfn_init();
766 }
767
768 #ifndef CONFIG_NEED_MULTIPLE_NODES
769 void __init initmem_init(unsigned long start_pfn,
770                                   unsigned long end_pfn)
771 {
772 #ifdef CONFIG_HIGHMEM
773         highstart_pfn = highend_pfn = max_pfn;
774         if (max_pfn > max_low_pfn)
775                 highstart_pfn = max_low_pfn;
776         memory_present(0, 0, highend_pfn);
777         e820_register_active_regions(0, 0, highend_pfn);
778         printk(KERN_NOTICE "%ldMB HIGHMEM available.\n",
779                 pages_to_mb(highend_pfn - highstart_pfn));
780         num_physpages = highend_pfn;
781         high_memory = (void *) __va(highstart_pfn * PAGE_SIZE - 1) + 1;
782 #else
783         memory_present(0, 0, max_low_pfn);
784         e820_register_active_regions(0, 0, max_low_pfn);
785         num_physpages = max_low_pfn;
786         high_memory = (void *) __va(max_low_pfn * PAGE_SIZE - 1) + 1;
787 #endif
788 #ifdef CONFIG_FLATMEM
789         max_mapnr = num_physpages;
790 #endif
791         printk(KERN_NOTICE "%ldMB LOWMEM available.\n",
792                         pages_to_mb(max_low_pfn));
793
794         setup_bootmem_allocator();
795 }
796 #endif /* !CONFIG_NEED_MULTIPLE_NODES */
797
798 static void __init zone_sizes_init(void)
799 {
800         unsigned long max_zone_pfns[MAX_NR_ZONES];
801         memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
802         max_zone_pfns[ZONE_DMA] =
803                 virt_to_phys((char *)MAX_DMA_ADDRESS) >> PAGE_SHIFT;
804         max_zone_pfns[ZONE_NORMAL] = max_low_pfn;
805 #ifdef CONFIG_HIGHMEM
806         max_zone_pfns[ZONE_HIGHMEM] = highend_pfn;
807 #endif
808
809         free_area_init_nodes(max_zone_pfns);
810 }
811
812 void __init setup_bootmem_allocator(void)
813 {
814         int i;
815         unsigned long bootmap_size, bootmap;
816         /*
817          * Initialize the boot-time allocator (with low memory only):
818          */
819         bootmap_size = bootmem_bootmap_pages(max_low_pfn)<<PAGE_SHIFT;
820         bootmap = find_e820_area(min_low_pfn<<PAGE_SHIFT,
821                                  max_pfn_mapped<<PAGE_SHIFT, bootmap_size,
822                                  PAGE_SIZE);
823         if (bootmap == -1L)
824                 panic("Cannot find bootmem map of size %ld\n", bootmap_size);
825         reserve_early(bootmap, bootmap + bootmap_size, "BOOTMAP");
826
827         /* don't touch min_low_pfn */
828         bootmap_size = init_bootmem_node(NODE_DATA(0), bootmap >> PAGE_SHIFT,
829                                          min_low_pfn, max_low_pfn);
830         printk(KERN_INFO "  mapped low ram: 0 - %08lx\n",
831                  max_pfn_mapped<<PAGE_SHIFT);
832         printk(KERN_INFO "  low ram: %08lx - %08lx\n",
833                  min_low_pfn<<PAGE_SHIFT, max_low_pfn<<PAGE_SHIFT);
834         printk(KERN_INFO "  bootmap %08lx - %08lx\n",
835                  bootmap, bootmap + bootmap_size);
836         for_each_online_node(i)
837                 free_bootmem_with_active_regions(i, max_low_pfn);
838         early_res_to_bootmem(0, max_low_pfn<<PAGE_SHIFT);
839
840         after_init_bootmem = 1;
841 }
842
843 static void __init find_early_table_space(unsigned long end, int use_pse)
844 {
845         unsigned long puds, pmds, ptes, tables, start;
846
847         puds = (end + PUD_SIZE - 1) >> PUD_SHIFT;
848         tables = roundup(puds * sizeof(pud_t), PAGE_SIZE);
849
850         pmds = (end + PMD_SIZE - 1) >> PMD_SHIFT;
851         tables += roundup(pmds * sizeof(pmd_t), PAGE_SIZE);
852
853         if (use_pse) {
854                 unsigned long extra;
855
856                 extra = end - ((end>>PMD_SHIFT) << PMD_SHIFT);
857                 extra += PMD_SIZE;
858                 ptes = (extra + PAGE_SIZE - 1) >> PAGE_SHIFT;
859         } else
860                 ptes = (end + PAGE_SIZE - 1) >> PAGE_SHIFT;
861
862         tables += roundup(ptes * sizeof(pte_t), PAGE_SIZE);
863
864         /* for fixmap */
865         tables += roundup(__end_of_fixed_addresses * sizeof(pte_t), PAGE_SIZE);
866
867         /*
868          * RED-PEN putting page tables only on node 0 could
869          * cause a hotspot and fill up ZONE_DMA. The page tables
870          * need roughly 0.5KB per GB.
871          */
872         start = 0x7000;
873         table_start = find_e820_area(start, max_pfn_mapped<<PAGE_SHIFT,
874                                         tables, PAGE_SIZE);
875         if (table_start == -1UL)
876                 panic("Cannot find space for the kernel page tables");
877
878         table_start >>= PAGE_SHIFT;
879         table_end = table_start;
880         table_top = table_start + (tables>>PAGE_SHIFT);
881
882         printk(KERN_DEBUG "kernel direct mapping tables up to %lx @ %lx-%lx\n",
883                 end, table_start << PAGE_SHIFT,
884                 (table_start << PAGE_SHIFT) + tables);
885 }
886
887 unsigned long __init_refok init_memory_mapping(unsigned long start,
888                                                 unsigned long end)
889 {
890         pgd_t *pgd_base = swapper_pg_dir;
891         unsigned long start_pfn, end_pfn;
892         unsigned long big_page_start;
893 #ifdef CONFIG_DEBUG_PAGEALLOC
894         /*
895          * For CONFIG_DEBUG_PAGEALLOC, identity mapping will use small pages.
896          * This will simplify cpa(), which otherwise needs to support splitting
897          * large pages into small in interrupt context, etc.
898          */
899         int use_pse = 0;
900 #else
901         int use_pse = cpu_has_pse;
902 #endif
903
904         /*
905          * Find space for the kernel direct mapping tables.
906          */
907         if (!after_init_bootmem)
908                 find_early_table_space(end, use_pse);
909
910 #ifdef CONFIG_X86_PAE
911         set_nx();
912         if (nx_enabled)
913                 printk(KERN_INFO "NX (Execute Disable) protection: active\n");
914 #endif
915
916         /* Enable PSE if available */
917         if (cpu_has_pse)
918                 set_in_cr4(X86_CR4_PSE);
919
920         /* Enable PGE if available */
921         if (cpu_has_pge) {
922                 set_in_cr4(X86_CR4_PGE);
923                 __supported_pte_mask |= _PAGE_GLOBAL;
924         }
925
926         /*
927          * Don't use a large page for the first 2/4MB of memory
928          * because there are often fixed size MTRRs in there
929          * and overlapping MTRRs into large pages can cause
930          * slowdowns.
931          */
932         big_page_start = PMD_SIZE;
933
934         if (start < big_page_start) {
935                 start_pfn = start >> PAGE_SHIFT;
936                 end_pfn = min(big_page_start>>PAGE_SHIFT, end>>PAGE_SHIFT);
937         } else {
938                 /* head is not big page alignment ? */
939                 start_pfn = start >> PAGE_SHIFT;
940                 end_pfn = ((start + (PMD_SIZE - 1))>>PMD_SHIFT)
941                                  << (PMD_SHIFT - PAGE_SHIFT);
942         }
943         if (start_pfn < end_pfn)
944                 kernel_physical_mapping_init(pgd_base, start_pfn, end_pfn, 0);
945
946         /* big page range */
947         start_pfn = ((start + (PMD_SIZE - 1))>>PMD_SHIFT)
948                          << (PMD_SHIFT - PAGE_SHIFT);
949         if (start_pfn < (big_page_start >> PAGE_SHIFT))
950                 start_pfn =  big_page_start >> PAGE_SHIFT;
951         end_pfn = (end>>PMD_SHIFT) << (PMD_SHIFT - PAGE_SHIFT);
952         if (start_pfn < end_pfn)
953                 kernel_physical_mapping_init(pgd_base, start_pfn, end_pfn,
954                                              use_pse);
955
956         /* tail is not big page alignment ? */
957         start_pfn = end_pfn;
958         if (start_pfn > (big_page_start>>PAGE_SHIFT)) {
959                 end_pfn = end >> PAGE_SHIFT;
960                 if (start_pfn < end_pfn)
961                         kernel_physical_mapping_init(pgd_base, start_pfn,
962                                                          end_pfn, 0);
963         }
964
965         early_ioremap_page_table_range_init(pgd_base);
966
967         load_cr3(swapper_pg_dir);
968
969         __flush_tlb_all();
970
971         if (!after_init_bootmem)
972                 reserve_early(table_start << PAGE_SHIFT,
973                                  table_end << PAGE_SHIFT, "PGTABLE");
974
975         if (!after_init_bootmem)
976                 early_memtest(start, end);
977
978         return end >> PAGE_SHIFT;
979 }
980
981
982 /*
983  * paging_init() sets up the page tables - note that the first 8MB are
984  * already mapped by head.S.
985  *
986  * This routines also unmaps the page at virtual kernel address 0, so
987  * that we can trap those pesky NULL-reference errors in the kernel.
988  */
989 void __init paging_init(void)
990 {
991         pagetable_init();
992
993         __flush_tlb_all();
994
995         kmap_init();
996
997         /*
998          * NOTE: at this point the bootmem allocator is fully available.
999          */
1000         sparse_init();
1001         zone_sizes_init();
1002 }
1003
1004 /*
1005  * Test if the WP bit works in supervisor mode. It isn't supported on 386's
1006  * and also on some strange 486's. All 586+'s are OK. This used to involve
1007  * black magic jumps to work around some nasty CPU bugs, but fortunately the
1008  * switch to using exceptions got rid of all that.
1009  */
1010 static void __init test_wp_bit(void)
1011 {
1012         printk(KERN_INFO
1013   "Checking if this processor honours the WP bit even in supervisor mode...");
1014
1015         /* Any page-aligned address will do, the test is non-destructive */
1016         __set_fixmap(FIX_WP_TEST, __pa(&swapper_pg_dir), PAGE_READONLY);
1017         boot_cpu_data.wp_works_ok = do_test_wp_bit();
1018         clear_fixmap(FIX_WP_TEST);
1019
1020         if (!boot_cpu_data.wp_works_ok) {
1021                 printk(KERN_CONT "No.\n");
1022 #ifdef CONFIG_X86_WP_WORKS_OK
1023                 panic(
1024   "This kernel doesn't support CPU's with broken WP. Recompile it for a 386!");
1025 #endif
1026         } else {
1027                 printk(KERN_CONT "Ok.\n");
1028         }
1029 }
1030
1031 static struct kcore_list kcore_mem, kcore_vmalloc;
1032
1033 void __init mem_init(void)
1034 {
1035         int codesize, reservedpages, datasize, initsize;
1036         int tmp;
1037
1038         pci_iommu_alloc();
1039
1040 #ifdef CONFIG_FLATMEM
1041         BUG_ON(!mem_map);
1042 #endif
1043         /* this will put all low memory onto the freelists */
1044         totalram_pages += free_all_bootmem();
1045
1046         reservedpages = 0;
1047         for (tmp = 0; tmp < max_low_pfn; tmp++)
1048                 /*
1049                  * Only count reserved RAM pages:
1050                  */
1051                 if (page_is_ram(tmp) && PageReserved(pfn_to_page(tmp)))
1052                         reservedpages++;
1053
1054         set_highmem_pages_init();
1055
1056         codesize =  (unsigned long) &_etext - (unsigned long) &_text;
1057         datasize =  (unsigned long) &_edata - (unsigned long) &_etext;
1058         initsize =  (unsigned long) &__init_end - (unsigned long) &__init_begin;
1059
1060         kclist_add(&kcore_mem, __va(0), max_low_pfn << PAGE_SHIFT);
1061         kclist_add(&kcore_vmalloc, (void *)VMALLOC_START,
1062                    VMALLOC_END-VMALLOC_START);
1063
1064         printk(KERN_INFO "Memory: %luk/%luk available (%dk kernel code, "
1065                         "%dk reserved, %dk data, %dk init, %ldk highmem)\n",
1066                 (unsigned long) nr_free_pages() << (PAGE_SHIFT-10),
1067                 num_physpages << (PAGE_SHIFT-10),
1068                 codesize >> 10,
1069                 reservedpages << (PAGE_SHIFT-10),
1070                 datasize >> 10,
1071                 initsize >> 10,
1072                 (unsigned long) (totalhigh_pages << (PAGE_SHIFT-10))
1073                );
1074
1075         printk(KERN_INFO "virtual kernel memory layout:\n"
1076                 "    fixmap  : 0x%08lx - 0x%08lx   (%4ld kB)\n"
1077 #ifdef CONFIG_HIGHMEM
1078                 "    pkmap   : 0x%08lx - 0x%08lx   (%4ld kB)\n"
1079 #endif
1080                 "    vmalloc : 0x%08lx - 0x%08lx   (%4ld MB)\n"
1081                 "    lowmem  : 0x%08lx - 0x%08lx   (%4ld MB)\n"
1082                 "      .init : 0x%08lx - 0x%08lx   (%4ld kB)\n"
1083                 "      .data : 0x%08lx - 0x%08lx   (%4ld kB)\n"
1084                 "      .text : 0x%08lx - 0x%08lx   (%4ld kB)\n",
1085                 FIXADDR_START, FIXADDR_TOP,
1086                 (FIXADDR_TOP - FIXADDR_START) >> 10,
1087
1088 #ifdef CONFIG_HIGHMEM
1089                 PKMAP_BASE, PKMAP_BASE+LAST_PKMAP*PAGE_SIZE,
1090                 (LAST_PKMAP*PAGE_SIZE) >> 10,
1091 #endif
1092
1093                 VMALLOC_START, VMALLOC_END,
1094                 (VMALLOC_END - VMALLOC_START) >> 20,
1095
1096                 (unsigned long)__va(0), (unsigned long)high_memory,
1097                 ((unsigned long)high_memory - (unsigned long)__va(0)) >> 20,
1098
1099                 (unsigned long)&__init_begin, (unsigned long)&__init_end,
1100                 ((unsigned long)&__init_end -
1101                  (unsigned long)&__init_begin) >> 10,
1102
1103                 (unsigned long)&_etext, (unsigned long)&_edata,
1104                 ((unsigned long)&_edata - (unsigned long)&_etext) >> 10,
1105
1106                 (unsigned long)&_text, (unsigned long)&_etext,
1107                 ((unsigned long)&_etext - (unsigned long)&_text) >> 10);
1108
1109         /*
1110          * Check boundaries twice: Some fundamental inconsistencies can
1111          * be detected at build time already.
1112          */
1113 #define __FIXADDR_TOP (-PAGE_SIZE)
1114 #ifdef CONFIG_HIGHMEM
1115         BUILD_BUG_ON(PKMAP_BASE + LAST_PKMAP*PAGE_SIZE  > FIXADDR_START);
1116         BUILD_BUG_ON(VMALLOC_END                        > PKMAP_BASE);
1117 #endif
1118 #define high_memory (-128UL << 20)
1119         BUILD_BUG_ON(VMALLOC_START                      >= VMALLOC_END);
1120 #undef high_memory
1121 #undef __FIXADDR_TOP
1122
1123 #ifdef CONFIG_HIGHMEM
1124         BUG_ON(PKMAP_BASE + LAST_PKMAP*PAGE_SIZE        > FIXADDR_START);
1125         BUG_ON(VMALLOC_END                              > PKMAP_BASE);
1126 #endif
1127         BUG_ON(VMALLOC_START                            >= VMALLOC_END);
1128         BUG_ON((unsigned long)high_memory               > VMALLOC_START);
1129
1130         if (boot_cpu_data.wp_works_ok < 0)
1131                 test_wp_bit();
1132
1133         save_pg_dir();
1134         zap_low_mappings();
1135 }
1136
1137 #ifdef CONFIG_MEMORY_HOTPLUG
1138 int arch_add_memory(int nid, u64 start, u64 size)
1139 {
1140         struct pglist_data *pgdata = NODE_DATA(nid);
1141         struct zone *zone = pgdata->node_zones + ZONE_HIGHMEM;
1142         unsigned long start_pfn = start >> PAGE_SHIFT;
1143         unsigned long nr_pages = size >> PAGE_SHIFT;
1144
1145         return __add_pages(nid, zone, start_pfn, nr_pages);
1146 }
1147 #endif
1148
1149 /*
1150  * This function cannot be __init, since exceptions don't work in that
1151  * section.  Put this after the callers, so that it cannot be inlined.
1152  */
1153 static noinline int do_test_wp_bit(void)
1154 {
1155         char tmp_reg;
1156         int flag;
1157
1158         __asm__ __volatile__(
1159                 "       movb %0, %1     \n"
1160                 "1:     movb %1, %0     \n"
1161                 "       xorl %2, %2     \n"
1162                 "2:                     \n"
1163                 _ASM_EXTABLE(1b,2b)
1164                 :"=m" (*(char *)fix_to_virt(FIX_WP_TEST)),
1165                  "=q" (tmp_reg),
1166                  "=r" (flag)
1167                 :"2" (1)
1168                 :"memory");
1169
1170         return flag;
1171 }
1172
1173 #ifdef CONFIG_DEBUG_RODATA
1174 const int rodata_test_data = 0xC3;
1175 EXPORT_SYMBOL_GPL(rodata_test_data);
1176
1177 void mark_rodata_ro(void)
1178 {
1179         unsigned long start = PFN_ALIGN(_text);
1180         unsigned long size = PFN_ALIGN(_etext) - start;
1181
1182 #ifndef CONFIG_DYNAMIC_FTRACE
1183         /* Dynamic tracing modifies the kernel text section */
1184         set_pages_ro(virt_to_page(start), size >> PAGE_SHIFT);
1185         printk(KERN_INFO "Write protecting the kernel text: %luk\n",
1186                 size >> 10);
1187
1188 #ifdef CONFIG_CPA_DEBUG
1189         printk(KERN_INFO "Testing CPA: Reverting %lx-%lx\n",
1190                 start, start+size);
1191         set_pages_rw(virt_to_page(start), size>>PAGE_SHIFT);
1192
1193         printk(KERN_INFO "Testing CPA: write protecting again\n");
1194         set_pages_ro(virt_to_page(start), size>>PAGE_SHIFT);
1195 #endif
1196 #endif /* CONFIG_DYNAMIC_FTRACE */
1197
1198         start += size;
1199         size = (unsigned long)__end_rodata - start;
1200         set_pages_ro(virt_to_page(start), size >> PAGE_SHIFT);
1201         printk(KERN_INFO "Write protecting the kernel read-only data: %luk\n",
1202                 size >> 10);
1203         rodata_test();
1204
1205 #ifdef CONFIG_CPA_DEBUG
1206         printk(KERN_INFO "Testing CPA: undo %lx-%lx\n", start, start + size);
1207         set_pages_rw(virt_to_page(start), size >> PAGE_SHIFT);
1208
1209         printk(KERN_INFO "Testing CPA: write protecting again\n");
1210         set_pages_ro(virt_to_page(start), size >> PAGE_SHIFT);
1211 #endif
1212 }
1213 #endif
1214
1215 #ifdef CONFIG_BLK_DEV_INITRD
1216 void free_initrd_mem(unsigned long start, unsigned long end)
1217 {
1218         free_init_pages("initrd memory", start, end);
1219 }
1220 #endif
1221
1222 int __init reserve_bootmem_generic(unsigned long phys, unsigned long len,
1223                                    int flags)
1224 {
1225         return reserve_bootmem(phys, len, flags);
1226 }