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