x86: set_highmem_pages_init() cleanup
[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 #else
471 static inline void permanent_kmaps_init(pgd_t *pgd_base)
472 {
473 }
474 #endif /* CONFIG_HIGHMEM */
475
476 void __init native_pagetable_setup_start(pgd_t *base)
477 {
478         unsigned long pfn, va;
479         pgd_t *pgd;
480         pud_t *pud;
481         pmd_t *pmd;
482         pte_t *pte;
483
484         /*
485          * Remove any mappings which extend past the end of physical
486          * memory from the boot time page table:
487          */
488         for (pfn = max_low_pfn + 1; pfn < 1<<(32-PAGE_SHIFT); pfn++) {
489                 va = PAGE_OFFSET + (pfn<<PAGE_SHIFT);
490                 pgd = base + pgd_index(va);
491                 if (!pgd_present(*pgd))
492                         break;
493
494                 pud = pud_offset(pgd, va);
495                 pmd = pmd_offset(pud, va);
496                 if (!pmd_present(*pmd))
497                         break;
498
499                 pte = pte_offset_kernel(pmd, va);
500                 if (!pte_present(*pte))
501                         break;
502
503                 pte_clear(NULL, va, pte);
504         }
505         paravirt_alloc_pmd(&init_mm, __pa(base) >> PAGE_SHIFT);
506 }
507
508 void __init native_pagetable_setup_done(pgd_t *base)
509 {
510 }
511
512 /*
513  * Build a proper pagetable for the kernel mappings.  Up until this
514  * point, we've been running on some set of pagetables constructed by
515  * the boot process.
516  *
517  * If we're booting on native hardware, this will be a pagetable
518  * constructed in arch/x86/kernel/head_32.S.  The root of the
519  * pagetable will be swapper_pg_dir.
520  *
521  * If we're booting paravirtualized under a hypervisor, then there are
522  * more options: we may already be running PAE, and the pagetable may
523  * or may not be based in swapper_pg_dir.  In any case,
524  * paravirt_pagetable_setup_start() will set up swapper_pg_dir
525  * appropriately for the rest of the initialization to work.
526  *
527  * In general, pagetable_init() assumes that the pagetable may already
528  * be partially populated, and so it avoids stomping on any existing
529  * mappings.
530  */
531 static void __init early_ioremap_page_table_range_init(pgd_t *pgd_base)
532 {
533         unsigned long vaddr, end;
534
535         /*
536          * Fixed mappings, only the page table structure has to be
537          * created - mappings will be set by set_fixmap():
538          */
539         vaddr = __fix_to_virt(__end_of_fixed_addresses - 1) & PMD_MASK;
540         end = (FIXADDR_TOP + PMD_SIZE - 1) & PMD_MASK;
541         page_table_range_init(vaddr, end, pgd_base);
542         early_ioremap_reset();
543 }
544
545 static void __init pagetable_init(void)
546 {
547         pgd_t *pgd_base = swapper_pg_dir;
548
549         permanent_kmaps_init(pgd_base);
550 }
551
552 #ifdef CONFIG_ACPI_SLEEP
553 /*
554  * ACPI suspend needs this for resume, because things like the intel-agp
555  * driver might have split up a kernel 4MB mapping.
556  */
557 char swsusp_pg_dir[PAGE_SIZE]
558         __attribute__ ((aligned(PAGE_SIZE)));
559
560 static inline void save_pg_dir(void)
561 {
562         memcpy(swsusp_pg_dir, swapper_pg_dir, PAGE_SIZE);
563 }
564 #else /* !CONFIG_ACPI_SLEEP */
565 static inline void save_pg_dir(void)
566 {
567 }
568 #endif /* !CONFIG_ACPI_SLEEP */
569
570 void zap_low_mappings(void)
571 {
572         int i;
573
574         /*
575          * Zap initial low-memory mappings.
576          *
577          * Note that "pgd_clear()" doesn't do it for
578          * us, because pgd_clear() is a no-op on i386.
579          */
580         for (i = 0; i < KERNEL_PGD_BOUNDARY; i++) {
581 #ifdef CONFIG_X86_PAE
582                 set_pgd(swapper_pg_dir+i, __pgd(1 + __pa(empty_zero_page)));
583 #else
584                 set_pgd(swapper_pg_dir+i, __pgd(0));
585 #endif
586         }
587         flush_tlb_all();
588 }
589
590 int nx_enabled;
591
592 pteval_t __supported_pte_mask __read_mostly = ~(_PAGE_NX | _PAGE_GLOBAL | _PAGE_IOMAP);
593 EXPORT_SYMBOL_GPL(__supported_pte_mask);
594
595 #ifdef CONFIG_X86_PAE
596
597 static int disable_nx __initdata;
598
599 /*
600  * noexec = on|off
601  *
602  * Control non executable mappings.
603  *
604  * on      Enable
605  * off     Disable
606  */
607 static int __init noexec_setup(char *str)
608 {
609         if (!str || !strcmp(str, "on")) {
610                 if (cpu_has_nx) {
611                         __supported_pte_mask |= _PAGE_NX;
612                         disable_nx = 0;
613                 }
614         } else {
615                 if (!strcmp(str, "off")) {
616                         disable_nx = 1;
617                         __supported_pte_mask &= ~_PAGE_NX;
618                 } else {
619                         return -EINVAL;
620                 }
621         }
622
623         return 0;
624 }
625 early_param("noexec", noexec_setup);
626
627 static void __init set_nx(void)
628 {
629         unsigned int v[4], l, h;
630
631         if (cpu_has_pae && (cpuid_eax(0x80000000) > 0x80000001)) {
632                 cpuid(0x80000001, &v[0], &v[1], &v[2], &v[3]);
633
634                 if ((v[3] & (1 << 20)) && !disable_nx) {
635                         rdmsr(MSR_EFER, l, h);
636                         l |= EFER_NX;
637                         wrmsr(MSR_EFER, l, h);
638                         nx_enabled = 1;
639                         __supported_pte_mask |= _PAGE_NX;
640                 }
641         }
642 }
643 #endif
644
645 /* user-defined highmem size */
646 static unsigned int highmem_pages = -1;
647
648 /*
649  * highmem=size forces highmem to be exactly 'size' bytes.
650  * This works even on boxes that have no highmem otherwise.
651  * This also works to reduce highmem size on bigger boxes.
652  */
653 static int __init parse_highmem(char *arg)
654 {
655         if (!arg)
656                 return -EINVAL;
657
658         highmem_pages = memparse(arg, &arg) >> PAGE_SHIFT;
659         return 0;
660 }
661 early_param("highmem", parse_highmem);
662
663 #define MSG_HIGHMEM_TOO_BIG \
664         "highmem size (%luMB) is bigger than pages available (%luMB)!\n"
665
666 #define MSG_LOWMEM_TOO_SMALL \
667         "highmem size (%luMB) results in <64MB lowmem, ignoring it!\n"
668 /*
669  * All of RAM fits into lowmem - but if user wants highmem
670  * artificially via the highmem=x boot parameter then create
671  * it:
672  */
673 void __init lowmem_pfn_init(void)
674 {
675         /* max_low_pfn is 0, we already have early_res support */
676         max_low_pfn = max_pfn;
677
678         if (highmem_pages == -1)
679                 highmem_pages = 0;
680 #ifdef CONFIG_HIGHMEM
681         if (highmem_pages >= max_pfn) {
682                 printk(KERN_ERR MSG_HIGHMEM_TOO_BIG,
683                         pages_to_mb(highmem_pages), pages_to_mb(max_pfn));
684                 highmem_pages = 0;
685         }
686         if (highmem_pages) {
687                 if (max_low_pfn - highmem_pages < 64*1024*1024/PAGE_SIZE) {
688                         printk(KERN_ERR MSG_LOWMEM_TOO_SMALL,
689                                 pages_to_mb(highmem_pages));
690                         highmem_pages = 0;
691                 }
692                 max_low_pfn -= highmem_pages;
693         }
694 #else
695         if (highmem_pages)
696                 printk(KERN_ERR "ignoring highmem size on non-highmem kernel!\n");
697 #endif
698 }
699
700 #define MSG_HIGHMEM_TOO_SMALL \
701         "only %luMB highmem pages available, ignoring highmem size of %luMB!\n"
702
703 #define MSG_HIGHMEM_TRIMMED \
704         "Warning: only 4GB will be used. Use a HIGHMEM64G enabled kernel!\n"
705 /*
706  * We have more RAM than fits into lowmem - we try to put it into
707  * highmem, also taking the highmem=x boot parameter into account:
708  */
709 void __init highmem_pfn_init(void)
710 {
711         max_low_pfn = MAXMEM_PFN;
712
713         if (highmem_pages == -1)
714                 highmem_pages = max_pfn - MAXMEM_PFN;
715
716         if (highmem_pages + MAXMEM_PFN < max_pfn)
717                 max_pfn = MAXMEM_PFN + highmem_pages;
718
719         if (highmem_pages + MAXMEM_PFN > max_pfn) {
720                 printk(KERN_WARNING MSG_HIGHMEM_TOO_SMALL,
721                         pages_to_mb(max_pfn - MAXMEM_PFN),
722                         pages_to_mb(highmem_pages));
723                 highmem_pages = 0;
724         }
725 #ifndef CONFIG_HIGHMEM
726         /* Maximum memory usable is what is directly addressable */
727         printk(KERN_WARNING "Warning only %ldMB will be used.\n", MAXMEM>>20);
728         if (max_pfn > MAX_NONPAE_PFN)
729                 printk(KERN_WARNING "Use a HIGHMEM64G enabled kernel.\n");
730         else
731                 printk(KERN_WARNING "Use a HIGHMEM enabled kernel.\n");
732         max_pfn = MAXMEM_PFN;
733 #else /* !CONFIG_HIGHMEM */
734 #ifndef CONFIG_HIGHMEM64G
735         if (max_pfn > MAX_NONPAE_PFN) {
736                 max_pfn = MAX_NONPAE_PFN;
737                 printk(KERN_WARNING MSG_HIGHMEM_TRIMMED);
738         }
739 #endif /* !CONFIG_HIGHMEM64G */
740 #endif /* !CONFIG_HIGHMEM */
741 }
742
743 /*
744  * Determine low and high memory ranges:
745  */
746 void __init find_low_pfn_range(void)
747 {
748         /* it could update max_pfn */
749
750         if (max_pfn <= MAXMEM_PFN)
751                 lowmem_pfn_init();
752         else
753                 highmem_pfn_init();
754 }
755
756 #ifndef CONFIG_NEED_MULTIPLE_NODES
757 void __init initmem_init(unsigned long start_pfn,
758                                   unsigned long end_pfn)
759 {
760 #ifdef CONFIG_HIGHMEM
761         highstart_pfn = highend_pfn = max_pfn;
762         if (max_pfn > max_low_pfn)
763                 highstart_pfn = max_low_pfn;
764         memory_present(0, 0, highend_pfn);
765         e820_register_active_regions(0, 0, highend_pfn);
766         printk(KERN_NOTICE "%ldMB HIGHMEM available.\n",
767                 pages_to_mb(highend_pfn - highstart_pfn));
768         num_physpages = highend_pfn;
769         high_memory = (void *) __va(highstart_pfn * PAGE_SIZE - 1) + 1;
770 #else
771         memory_present(0, 0, max_low_pfn);
772         e820_register_active_regions(0, 0, max_low_pfn);
773         num_physpages = max_low_pfn;
774         high_memory = (void *) __va(max_low_pfn * PAGE_SIZE - 1) + 1;
775 #endif
776 #ifdef CONFIG_FLATMEM
777         max_mapnr = num_physpages;
778 #endif
779         printk(KERN_NOTICE "%ldMB LOWMEM available.\n",
780                         pages_to_mb(max_low_pfn));
781
782         setup_bootmem_allocator();
783 }
784 #endif /* !CONFIG_NEED_MULTIPLE_NODES */
785
786 static void __init zone_sizes_init(void)
787 {
788         unsigned long max_zone_pfns[MAX_NR_ZONES];
789         memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
790         max_zone_pfns[ZONE_DMA] =
791                 virt_to_phys((char *)MAX_DMA_ADDRESS) >> PAGE_SHIFT;
792         max_zone_pfns[ZONE_NORMAL] = max_low_pfn;
793 #ifdef CONFIG_HIGHMEM
794         max_zone_pfns[ZONE_HIGHMEM] = highend_pfn;
795 #endif
796
797         free_area_init_nodes(max_zone_pfns);
798 }
799
800 void __init setup_bootmem_allocator(void)
801 {
802         int i;
803         unsigned long bootmap_size, bootmap;
804         /*
805          * Initialize the boot-time allocator (with low memory only):
806          */
807         bootmap_size = bootmem_bootmap_pages(max_low_pfn)<<PAGE_SHIFT;
808         bootmap = find_e820_area(min_low_pfn<<PAGE_SHIFT,
809                                  max_pfn_mapped<<PAGE_SHIFT, bootmap_size,
810                                  PAGE_SIZE);
811         if (bootmap == -1L)
812                 panic("Cannot find bootmem map of size %ld\n", bootmap_size);
813         reserve_early(bootmap, bootmap + bootmap_size, "BOOTMAP");
814
815         /* don't touch min_low_pfn */
816         bootmap_size = init_bootmem_node(NODE_DATA(0), bootmap >> PAGE_SHIFT,
817                                          min_low_pfn, max_low_pfn);
818         printk(KERN_INFO "  mapped low ram: 0 - %08lx\n",
819                  max_pfn_mapped<<PAGE_SHIFT);
820         printk(KERN_INFO "  low ram: %08lx - %08lx\n",
821                  min_low_pfn<<PAGE_SHIFT, max_low_pfn<<PAGE_SHIFT);
822         printk(KERN_INFO "  bootmap %08lx - %08lx\n",
823                  bootmap, bootmap + bootmap_size);
824         for_each_online_node(i)
825                 free_bootmem_with_active_regions(i, max_low_pfn);
826         early_res_to_bootmem(0, max_low_pfn<<PAGE_SHIFT);
827
828         after_init_bootmem = 1;
829 }
830
831 static void __init find_early_table_space(unsigned long end, int use_pse)
832 {
833         unsigned long puds, pmds, ptes, tables, start;
834
835         puds = (end + PUD_SIZE - 1) >> PUD_SHIFT;
836         tables = roundup(puds * sizeof(pud_t), PAGE_SIZE);
837
838         pmds = (end + PMD_SIZE - 1) >> PMD_SHIFT;
839         tables += roundup(pmds * sizeof(pmd_t), PAGE_SIZE);
840
841         if (use_pse) {
842                 unsigned long extra;
843
844                 extra = end - ((end>>PMD_SHIFT) << PMD_SHIFT);
845                 extra += PMD_SIZE;
846                 ptes = (extra + PAGE_SIZE - 1) >> PAGE_SHIFT;
847         } else
848                 ptes = (end + PAGE_SIZE - 1) >> PAGE_SHIFT;
849
850         tables += roundup(ptes * sizeof(pte_t), PAGE_SIZE);
851
852         /* for fixmap */
853         tables += roundup(__end_of_fixed_addresses * sizeof(pte_t), PAGE_SIZE);
854
855         /*
856          * RED-PEN putting page tables only on node 0 could
857          * cause a hotspot and fill up ZONE_DMA. The page tables
858          * need roughly 0.5KB per GB.
859          */
860         start = 0x7000;
861         table_start = find_e820_area(start, max_pfn_mapped<<PAGE_SHIFT,
862                                         tables, PAGE_SIZE);
863         if (table_start == -1UL)
864                 panic("Cannot find space for the kernel page tables");
865
866         table_start >>= PAGE_SHIFT;
867         table_end = table_start;
868         table_top = table_start + (tables>>PAGE_SHIFT);
869
870         printk(KERN_DEBUG "kernel direct mapping tables up to %lx @ %lx-%lx\n",
871                 end, table_start << PAGE_SHIFT,
872                 (table_start << PAGE_SHIFT) + tables);
873 }
874
875 unsigned long __init_refok init_memory_mapping(unsigned long start,
876                                                 unsigned long end)
877 {
878         pgd_t *pgd_base = swapper_pg_dir;
879         unsigned long start_pfn, end_pfn;
880         unsigned long big_page_start;
881 #ifdef CONFIG_DEBUG_PAGEALLOC
882         /*
883          * For CONFIG_DEBUG_PAGEALLOC, identity mapping will use small pages.
884          * This will simplify cpa(), which otherwise needs to support splitting
885          * large pages into small in interrupt context, etc.
886          */
887         int use_pse = 0;
888 #else
889         int use_pse = cpu_has_pse;
890 #endif
891
892         /*
893          * Find space for the kernel direct mapping tables.
894          */
895         if (!after_init_bootmem)
896                 find_early_table_space(end, use_pse);
897
898 #ifdef CONFIG_X86_PAE
899         set_nx();
900         if (nx_enabled)
901                 printk(KERN_INFO "NX (Execute Disable) protection: active\n");
902 #endif
903
904         /* Enable PSE if available */
905         if (cpu_has_pse)
906                 set_in_cr4(X86_CR4_PSE);
907
908         /* Enable PGE if available */
909         if (cpu_has_pge) {
910                 set_in_cr4(X86_CR4_PGE);
911                 __supported_pte_mask |= _PAGE_GLOBAL;
912         }
913
914         /*
915          * Don't use a large page for the first 2/4MB of memory
916          * because there are often fixed size MTRRs in there
917          * and overlapping MTRRs into large pages can cause
918          * slowdowns.
919          */
920         big_page_start = PMD_SIZE;
921
922         if (start < big_page_start) {
923                 start_pfn = start >> PAGE_SHIFT;
924                 end_pfn = min(big_page_start>>PAGE_SHIFT, end>>PAGE_SHIFT);
925         } else {
926                 /* head is not big page alignment ? */
927                 start_pfn = start >> PAGE_SHIFT;
928                 end_pfn = ((start + (PMD_SIZE - 1))>>PMD_SHIFT)
929                                  << (PMD_SHIFT - PAGE_SHIFT);
930         }
931         if (start_pfn < end_pfn)
932                 kernel_physical_mapping_init(pgd_base, start_pfn, end_pfn, 0);
933
934         /* big page range */
935         start_pfn = ((start + (PMD_SIZE - 1))>>PMD_SHIFT)
936                          << (PMD_SHIFT - PAGE_SHIFT);
937         if (start_pfn < (big_page_start >> PAGE_SHIFT))
938                 start_pfn =  big_page_start >> PAGE_SHIFT;
939         end_pfn = (end>>PMD_SHIFT) << (PMD_SHIFT - PAGE_SHIFT);
940         if (start_pfn < end_pfn)
941                 kernel_physical_mapping_init(pgd_base, start_pfn, end_pfn,
942                                              use_pse);
943
944         /* tail is not big page alignment ? */
945         start_pfn = end_pfn;
946         if (start_pfn > (big_page_start>>PAGE_SHIFT)) {
947                 end_pfn = end >> PAGE_SHIFT;
948                 if (start_pfn < end_pfn)
949                         kernel_physical_mapping_init(pgd_base, start_pfn,
950                                                          end_pfn, 0);
951         }
952
953         early_ioremap_page_table_range_init(pgd_base);
954
955         load_cr3(swapper_pg_dir);
956
957         __flush_tlb_all();
958
959         if (!after_init_bootmem)
960                 reserve_early(table_start << PAGE_SHIFT,
961                                  table_end << PAGE_SHIFT, "PGTABLE");
962
963         if (!after_init_bootmem)
964                 early_memtest(start, end);
965
966         return end >> PAGE_SHIFT;
967 }
968
969
970 /*
971  * paging_init() sets up the page tables - note that the first 8MB are
972  * already mapped by head.S.
973  *
974  * This routines also unmaps the page at virtual kernel address 0, so
975  * that we can trap those pesky NULL-reference errors in the kernel.
976  */
977 void __init paging_init(void)
978 {
979         pagetable_init();
980
981         __flush_tlb_all();
982
983         kmap_init();
984
985         /*
986          * NOTE: at this point the bootmem allocator is fully available.
987          */
988         sparse_init();
989         zone_sizes_init();
990 }
991
992 /*
993  * Test if the WP bit works in supervisor mode. It isn't supported on 386's
994  * and also on some strange 486's. All 586+'s are OK. This used to involve
995  * black magic jumps to work around some nasty CPU bugs, but fortunately the
996  * switch to using exceptions got rid of all that.
997  */
998 static void __init test_wp_bit(void)
999 {
1000         printk(KERN_INFO
1001   "Checking if this processor honours the WP bit even in supervisor mode...");
1002
1003         /* Any page-aligned address will do, the test is non-destructive */
1004         __set_fixmap(FIX_WP_TEST, __pa(&swapper_pg_dir), PAGE_READONLY);
1005         boot_cpu_data.wp_works_ok = do_test_wp_bit();
1006         clear_fixmap(FIX_WP_TEST);
1007
1008         if (!boot_cpu_data.wp_works_ok) {
1009                 printk(KERN_CONT "No.\n");
1010 #ifdef CONFIG_X86_WP_WORKS_OK
1011                 panic(
1012   "This kernel doesn't support CPU's with broken WP. Recompile it for a 386!");
1013 #endif
1014         } else {
1015                 printk(KERN_CONT "Ok.\n");
1016         }
1017 }
1018
1019 static struct kcore_list kcore_mem, kcore_vmalloc;
1020
1021 void __init mem_init(void)
1022 {
1023         int codesize, reservedpages, datasize, initsize;
1024         int tmp;
1025
1026         pci_iommu_alloc();
1027
1028 #ifdef CONFIG_FLATMEM
1029         BUG_ON(!mem_map);
1030 #endif
1031         /* this will put all low memory onto the freelists */
1032         totalram_pages += free_all_bootmem();
1033
1034         reservedpages = 0;
1035         for (tmp = 0; tmp < max_low_pfn; tmp++)
1036                 /*
1037                  * Only count reserved RAM pages:
1038                  */
1039                 if (page_is_ram(tmp) && PageReserved(pfn_to_page(tmp)))
1040                         reservedpages++;
1041
1042         set_highmem_pages_init();
1043
1044         codesize =  (unsigned long) &_etext - (unsigned long) &_text;
1045         datasize =  (unsigned long) &_edata - (unsigned long) &_etext;
1046         initsize =  (unsigned long) &__init_end - (unsigned long) &__init_begin;
1047
1048         kclist_add(&kcore_mem, __va(0), max_low_pfn << PAGE_SHIFT);
1049         kclist_add(&kcore_vmalloc, (void *)VMALLOC_START,
1050                    VMALLOC_END-VMALLOC_START);
1051
1052         printk(KERN_INFO "Memory: %luk/%luk available (%dk kernel code, "
1053                         "%dk reserved, %dk data, %dk init, %ldk highmem)\n",
1054                 (unsigned long) nr_free_pages() << (PAGE_SHIFT-10),
1055                 num_physpages << (PAGE_SHIFT-10),
1056                 codesize >> 10,
1057                 reservedpages << (PAGE_SHIFT-10),
1058                 datasize >> 10,
1059                 initsize >> 10,
1060                 (unsigned long) (totalhigh_pages << (PAGE_SHIFT-10))
1061                );
1062
1063         printk(KERN_INFO "virtual kernel memory layout:\n"
1064                 "    fixmap  : 0x%08lx - 0x%08lx   (%4ld kB)\n"
1065 #ifdef CONFIG_HIGHMEM
1066                 "    pkmap   : 0x%08lx - 0x%08lx   (%4ld kB)\n"
1067 #endif
1068                 "    vmalloc : 0x%08lx - 0x%08lx   (%4ld MB)\n"
1069                 "    lowmem  : 0x%08lx - 0x%08lx   (%4ld MB)\n"
1070                 "      .init : 0x%08lx - 0x%08lx   (%4ld kB)\n"
1071                 "      .data : 0x%08lx - 0x%08lx   (%4ld kB)\n"
1072                 "      .text : 0x%08lx - 0x%08lx   (%4ld kB)\n",
1073                 FIXADDR_START, FIXADDR_TOP,
1074                 (FIXADDR_TOP - FIXADDR_START) >> 10,
1075
1076 #ifdef CONFIG_HIGHMEM
1077                 PKMAP_BASE, PKMAP_BASE+LAST_PKMAP*PAGE_SIZE,
1078                 (LAST_PKMAP*PAGE_SIZE) >> 10,
1079 #endif
1080
1081                 VMALLOC_START, VMALLOC_END,
1082                 (VMALLOC_END - VMALLOC_START) >> 20,
1083
1084                 (unsigned long)__va(0), (unsigned long)high_memory,
1085                 ((unsigned long)high_memory - (unsigned long)__va(0)) >> 20,
1086
1087                 (unsigned long)&__init_begin, (unsigned long)&__init_end,
1088                 ((unsigned long)&__init_end -
1089                  (unsigned long)&__init_begin) >> 10,
1090
1091                 (unsigned long)&_etext, (unsigned long)&_edata,
1092                 ((unsigned long)&_edata - (unsigned long)&_etext) >> 10,
1093
1094                 (unsigned long)&_text, (unsigned long)&_etext,
1095                 ((unsigned long)&_etext - (unsigned long)&_text) >> 10);
1096
1097         /*
1098          * Check boundaries twice: Some fundamental inconsistencies can
1099          * be detected at build time already.
1100          */
1101 #define __FIXADDR_TOP (-PAGE_SIZE)
1102 #ifdef CONFIG_HIGHMEM
1103         BUILD_BUG_ON(PKMAP_BASE + LAST_PKMAP*PAGE_SIZE  > FIXADDR_START);
1104         BUILD_BUG_ON(VMALLOC_END                        > PKMAP_BASE);
1105 #endif
1106 #define high_memory (-128UL << 20)
1107         BUILD_BUG_ON(VMALLOC_START                      >= VMALLOC_END);
1108 #undef high_memory
1109 #undef __FIXADDR_TOP
1110
1111 #ifdef CONFIG_HIGHMEM
1112         BUG_ON(PKMAP_BASE + LAST_PKMAP*PAGE_SIZE        > FIXADDR_START);
1113         BUG_ON(VMALLOC_END                              > PKMAP_BASE);
1114 #endif
1115         BUG_ON(VMALLOC_START                            >= VMALLOC_END);
1116         BUG_ON((unsigned long)high_memory               > VMALLOC_START);
1117
1118         if (boot_cpu_data.wp_works_ok < 0)
1119                 test_wp_bit();
1120
1121         save_pg_dir();
1122         zap_low_mappings();
1123 }
1124
1125 #ifdef CONFIG_MEMORY_HOTPLUG
1126 int arch_add_memory(int nid, u64 start, u64 size)
1127 {
1128         struct pglist_data *pgdata = NODE_DATA(nid);
1129         struct zone *zone = pgdata->node_zones + ZONE_HIGHMEM;
1130         unsigned long start_pfn = start >> PAGE_SHIFT;
1131         unsigned long nr_pages = size >> PAGE_SHIFT;
1132
1133         return __add_pages(nid, zone, start_pfn, nr_pages);
1134 }
1135 #endif
1136
1137 /*
1138  * This function cannot be __init, since exceptions don't work in that
1139  * section.  Put this after the callers, so that it cannot be inlined.
1140  */
1141 static noinline int do_test_wp_bit(void)
1142 {
1143         char tmp_reg;
1144         int flag;
1145
1146         __asm__ __volatile__(
1147                 "       movb %0, %1     \n"
1148                 "1:     movb %1, %0     \n"
1149                 "       xorl %2, %2     \n"
1150                 "2:                     \n"
1151                 _ASM_EXTABLE(1b,2b)
1152                 :"=m" (*(char *)fix_to_virt(FIX_WP_TEST)),
1153                  "=q" (tmp_reg),
1154                  "=r" (flag)
1155                 :"2" (1)
1156                 :"memory");
1157
1158         return flag;
1159 }
1160
1161 #ifdef CONFIG_DEBUG_RODATA
1162 const int rodata_test_data = 0xC3;
1163 EXPORT_SYMBOL_GPL(rodata_test_data);
1164
1165 void mark_rodata_ro(void)
1166 {
1167         unsigned long start = PFN_ALIGN(_text);
1168         unsigned long size = PFN_ALIGN(_etext) - start;
1169
1170 #ifndef CONFIG_DYNAMIC_FTRACE
1171         /* Dynamic tracing modifies the kernel text section */
1172         set_pages_ro(virt_to_page(start), size >> PAGE_SHIFT);
1173         printk(KERN_INFO "Write protecting the kernel text: %luk\n",
1174                 size >> 10);
1175
1176 #ifdef CONFIG_CPA_DEBUG
1177         printk(KERN_INFO "Testing CPA: Reverting %lx-%lx\n",
1178                 start, start+size);
1179         set_pages_rw(virt_to_page(start), size>>PAGE_SHIFT);
1180
1181         printk(KERN_INFO "Testing CPA: write protecting again\n");
1182         set_pages_ro(virt_to_page(start), size>>PAGE_SHIFT);
1183 #endif
1184 #endif /* CONFIG_DYNAMIC_FTRACE */
1185
1186         start += size;
1187         size = (unsigned long)__end_rodata - start;
1188         set_pages_ro(virt_to_page(start), size >> PAGE_SHIFT);
1189         printk(KERN_INFO "Write protecting the kernel read-only data: %luk\n",
1190                 size >> 10);
1191         rodata_test();
1192
1193 #ifdef CONFIG_CPA_DEBUG
1194         printk(KERN_INFO "Testing CPA: undo %lx-%lx\n", start, start + size);
1195         set_pages_rw(virt_to_page(start), size >> PAGE_SHIFT);
1196
1197         printk(KERN_INFO "Testing CPA: write protecting again\n");
1198         set_pages_ro(virt_to_page(start), size >> PAGE_SHIFT);
1199 #endif
1200 }
1201 #endif
1202
1203 #ifdef CONFIG_BLK_DEV_INITRD
1204 void free_initrd_mem(unsigned long start, unsigned long end)
1205 {
1206         free_init_pages("initrd memory", start, end);
1207 }
1208 #endif
1209
1210 int __init reserve_bootmem_generic(unsigned long phys, unsigned long len,
1211                                    int flags)
1212 {
1213         return reserve_bootmem(phys, len, flags);
1214 }