x86: fix bootmem cross node for 32bit numa
[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 pte_t *kmap_pte;
358 pgprot_t kmap_prot;
359
360 static inline pte_t *kmap_get_fixmap_pte(unsigned long vaddr)
361 {
362         return pte_offset_kernel(pmd_offset(pud_offset(pgd_offset_k(vaddr),
363                         vaddr), vaddr), vaddr);
364 }
365
366 static void __init kmap_init(void)
367 {
368         unsigned long kmap_vstart;
369
370         /*
371          * Cache the first kmap pte:
372          */
373         kmap_vstart = __fix_to_virt(FIX_KMAP_BEGIN);
374         kmap_pte = kmap_get_fixmap_pte(kmap_vstart);
375
376         kmap_prot = PAGE_KERNEL;
377 }
378
379 #ifdef CONFIG_HIGHMEM
380 static void __init permanent_kmaps_init(pgd_t *pgd_base)
381 {
382         unsigned long vaddr;
383         pgd_t *pgd;
384         pud_t *pud;
385         pmd_t *pmd;
386         pte_t *pte;
387
388         vaddr = PKMAP_BASE;
389         page_table_range_init(vaddr, vaddr + PAGE_SIZE*LAST_PKMAP, pgd_base);
390
391         pgd = swapper_pg_dir + pgd_index(vaddr);
392         pud = pud_offset(pgd, vaddr);
393         pmd = pmd_offset(pud, vaddr);
394         pte = pte_offset_kernel(pmd, vaddr);
395         pkmap_page_table = pte;
396 }
397
398 static void __init add_one_highpage_init(struct page *page, int pfn)
399 {
400         ClearPageReserved(page);
401         init_page_count(page);
402         __free_page(page);
403         totalhigh_pages++;
404 }
405
406 struct add_highpages_data {
407         unsigned long start_pfn;
408         unsigned long end_pfn;
409 };
410
411 static int __init add_highpages_work_fn(unsigned long start_pfn,
412                                          unsigned long end_pfn, void *datax)
413 {
414         int node_pfn;
415         struct page *page;
416         unsigned long final_start_pfn, final_end_pfn;
417         struct add_highpages_data *data;
418
419         data = (struct add_highpages_data *)datax;
420
421         final_start_pfn = max(start_pfn, data->start_pfn);
422         final_end_pfn = min(end_pfn, data->end_pfn);
423         if (final_start_pfn >= final_end_pfn)
424                 return 0;
425
426         for (node_pfn = final_start_pfn; node_pfn < final_end_pfn;
427              node_pfn++) {
428                 if (!pfn_valid(node_pfn))
429                         continue;
430                 page = pfn_to_page(node_pfn);
431                 add_one_highpage_init(page, node_pfn);
432         }
433
434         return 0;
435
436 }
437
438 void __init add_highpages_with_active_regions(int nid, unsigned long start_pfn,
439                                               unsigned long end_pfn)
440 {
441         struct add_highpages_data data;
442
443         data.start_pfn = start_pfn;
444         data.end_pfn = end_pfn;
445
446         work_with_active_regions(nid, add_highpages_work_fn, &data);
447 }
448
449 #else
450 static inline void permanent_kmaps_init(pgd_t *pgd_base)
451 {
452 }
453 #endif /* CONFIG_HIGHMEM */
454
455 void __init native_pagetable_setup_start(pgd_t *base)
456 {
457         unsigned long pfn, va;
458         pgd_t *pgd;
459         pud_t *pud;
460         pmd_t *pmd;
461         pte_t *pte;
462
463         /*
464          * Remove any mappings which extend past the end of physical
465          * memory from the boot time page table:
466          */
467         for (pfn = max_low_pfn + 1; pfn < 1<<(32-PAGE_SHIFT); pfn++) {
468                 va = PAGE_OFFSET + (pfn<<PAGE_SHIFT);
469                 pgd = base + pgd_index(va);
470                 if (!pgd_present(*pgd))
471                         break;
472
473                 pud = pud_offset(pgd, va);
474                 pmd = pmd_offset(pud, va);
475                 if (!pmd_present(*pmd))
476                         break;
477
478                 pte = pte_offset_kernel(pmd, va);
479                 if (!pte_present(*pte))
480                         break;
481
482                 pte_clear(NULL, va, pte);
483         }
484         paravirt_alloc_pmd(&init_mm, __pa(base) >> PAGE_SHIFT);
485 }
486
487 void __init native_pagetable_setup_done(pgd_t *base)
488 {
489 }
490
491 /*
492  * Build a proper pagetable for the kernel mappings.  Up until this
493  * point, we've been running on some set of pagetables constructed by
494  * the boot process.
495  *
496  * If we're booting on native hardware, this will be a pagetable
497  * constructed in arch/x86/kernel/head_32.S.  The root of the
498  * pagetable will be swapper_pg_dir.
499  *
500  * If we're booting paravirtualized under a hypervisor, then there are
501  * more options: we may already be running PAE, and the pagetable may
502  * or may not be based in swapper_pg_dir.  In any case,
503  * paravirt_pagetable_setup_start() will set up swapper_pg_dir
504  * appropriately for the rest of the initialization to work.
505  *
506  * In general, pagetable_init() assumes that the pagetable may already
507  * be partially populated, and so it avoids stomping on any existing
508  * mappings.
509  */
510 static void __init early_ioremap_page_table_range_init(pgd_t *pgd_base)
511 {
512         unsigned long vaddr, end;
513
514         /*
515          * Fixed mappings, only the page table structure has to be
516          * created - mappings will be set by set_fixmap():
517          */
518         vaddr = __fix_to_virt(__end_of_fixed_addresses - 1) & PMD_MASK;
519         end = (FIXADDR_TOP + PMD_SIZE - 1) & PMD_MASK;
520         page_table_range_init(vaddr, end, pgd_base);
521         early_ioremap_reset();
522 }
523
524 static void __init pagetable_init(void)
525 {
526         pgd_t *pgd_base = swapper_pg_dir;
527
528         permanent_kmaps_init(pgd_base);
529 }
530
531 #ifdef CONFIG_ACPI_SLEEP
532 /*
533  * ACPI suspend needs this for resume, because things like the intel-agp
534  * driver might have split up a kernel 4MB mapping.
535  */
536 char swsusp_pg_dir[PAGE_SIZE]
537         __attribute__ ((aligned(PAGE_SIZE)));
538
539 static inline void save_pg_dir(void)
540 {
541         memcpy(swsusp_pg_dir, swapper_pg_dir, PAGE_SIZE);
542 }
543 #else /* !CONFIG_ACPI_SLEEP */
544 static inline void save_pg_dir(void)
545 {
546 }
547 #endif /* !CONFIG_ACPI_SLEEP */
548
549 void zap_low_mappings(void)
550 {
551         int i;
552
553         /*
554          * Zap initial low-memory mappings.
555          *
556          * Note that "pgd_clear()" doesn't do it for
557          * us, because pgd_clear() is a no-op on i386.
558          */
559         for (i = 0; i < KERNEL_PGD_BOUNDARY; i++) {
560 #ifdef CONFIG_X86_PAE
561                 set_pgd(swapper_pg_dir+i, __pgd(1 + __pa(empty_zero_page)));
562 #else
563                 set_pgd(swapper_pg_dir+i, __pgd(0));
564 #endif
565         }
566         flush_tlb_all();
567 }
568
569 int nx_enabled;
570
571 pteval_t __supported_pte_mask __read_mostly = ~(_PAGE_NX | _PAGE_GLOBAL | _PAGE_IOMAP);
572 EXPORT_SYMBOL_GPL(__supported_pte_mask);
573
574 #ifdef CONFIG_X86_PAE
575
576 static int disable_nx __initdata;
577
578 /*
579  * noexec = on|off
580  *
581  * Control non executable mappings.
582  *
583  * on      Enable
584  * off     Disable
585  */
586 static int __init noexec_setup(char *str)
587 {
588         if (!str || !strcmp(str, "on")) {
589                 if (cpu_has_nx) {
590                         __supported_pte_mask |= _PAGE_NX;
591                         disable_nx = 0;
592                 }
593         } else {
594                 if (!strcmp(str, "off")) {
595                         disable_nx = 1;
596                         __supported_pte_mask &= ~_PAGE_NX;
597                 } else {
598                         return -EINVAL;
599                 }
600         }
601
602         return 0;
603 }
604 early_param("noexec", noexec_setup);
605
606 static void __init set_nx(void)
607 {
608         unsigned int v[4], l, h;
609
610         if (cpu_has_pae && (cpuid_eax(0x80000000) > 0x80000001)) {
611                 cpuid(0x80000001, &v[0], &v[1], &v[2], &v[3]);
612
613                 if ((v[3] & (1 << 20)) && !disable_nx) {
614                         rdmsr(MSR_EFER, l, h);
615                         l |= EFER_NX;
616                         wrmsr(MSR_EFER, l, h);
617                         nx_enabled = 1;
618                         __supported_pte_mask |= _PAGE_NX;
619                 }
620         }
621 }
622 #endif
623
624 /* user-defined highmem size */
625 static unsigned int highmem_pages = -1;
626
627 /*
628  * highmem=size forces highmem to be exactly 'size' bytes.
629  * This works even on boxes that have no highmem otherwise.
630  * This also works to reduce highmem size on bigger boxes.
631  */
632 static int __init parse_highmem(char *arg)
633 {
634         if (!arg)
635                 return -EINVAL;
636
637         highmem_pages = memparse(arg, &arg) >> PAGE_SHIFT;
638         return 0;
639 }
640 early_param("highmem", parse_highmem);
641
642 #define MSG_HIGHMEM_TOO_BIG \
643         "highmem size (%luMB) is bigger than pages available (%luMB)!\n"
644
645 #define MSG_LOWMEM_TOO_SMALL \
646         "highmem size (%luMB) results in <64MB lowmem, ignoring it!\n"
647 /*
648  * All of RAM fits into lowmem - but if user wants highmem
649  * artificially via the highmem=x boot parameter then create
650  * it:
651  */
652 void __init lowmem_pfn_init(void)
653 {
654         /* max_low_pfn is 0, we already have early_res support */
655         max_low_pfn = max_pfn;
656
657         if (highmem_pages == -1)
658                 highmem_pages = 0;
659 #ifdef CONFIG_HIGHMEM
660         if (highmem_pages >= max_pfn) {
661                 printk(KERN_ERR MSG_HIGHMEM_TOO_BIG,
662                         pages_to_mb(highmem_pages), pages_to_mb(max_pfn));
663                 highmem_pages = 0;
664         }
665         if (highmem_pages) {
666                 if (max_low_pfn - highmem_pages < 64*1024*1024/PAGE_SIZE) {
667                         printk(KERN_ERR MSG_LOWMEM_TOO_SMALL,
668                                 pages_to_mb(highmem_pages));
669                         highmem_pages = 0;
670                 }
671                 max_low_pfn -= highmem_pages;
672         }
673 #else
674         if (highmem_pages)
675                 printk(KERN_ERR "ignoring highmem size on non-highmem kernel!\n");
676 #endif
677 }
678
679 #define MSG_HIGHMEM_TOO_SMALL \
680         "only %luMB highmem pages available, ignoring highmem size of %luMB!\n"
681
682 #define MSG_HIGHMEM_TRIMMED \
683         "Warning: only 4GB will be used. Use a HIGHMEM64G enabled kernel!\n"
684 /*
685  * We have more RAM than fits into lowmem - we try to put it into
686  * highmem, also taking the highmem=x boot parameter into account:
687  */
688 void __init highmem_pfn_init(void)
689 {
690         max_low_pfn = MAXMEM_PFN;
691
692         if (highmem_pages == -1)
693                 highmem_pages = max_pfn - MAXMEM_PFN;
694
695         if (highmem_pages + MAXMEM_PFN < max_pfn)
696                 max_pfn = MAXMEM_PFN + highmem_pages;
697
698         if (highmem_pages + MAXMEM_PFN > max_pfn) {
699                 printk(KERN_WARNING MSG_HIGHMEM_TOO_SMALL,
700                         pages_to_mb(max_pfn - MAXMEM_PFN),
701                         pages_to_mb(highmem_pages));
702                 highmem_pages = 0;
703         }
704 #ifndef CONFIG_HIGHMEM
705         /* Maximum memory usable is what is directly addressable */
706         printk(KERN_WARNING "Warning only %ldMB will be used.\n", MAXMEM>>20);
707         if (max_pfn > MAX_NONPAE_PFN)
708                 printk(KERN_WARNING "Use a HIGHMEM64G enabled kernel.\n");
709         else
710                 printk(KERN_WARNING "Use a HIGHMEM enabled kernel.\n");
711         max_pfn = MAXMEM_PFN;
712 #else /* !CONFIG_HIGHMEM */
713 #ifndef CONFIG_HIGHMEM64G
714         if (max_pfn > MAX_NONPAE_PFN) {
715                 max_pfn = MAX_NONPAE_PFN;
716                 printk(KERN_WARNING MSG_HIGHMEM_TRIMMED);
717         }
718 #endif /* !CONFIG_HIGHMEM64G */
719 #endif /* !CONFIG_HIGHMEM */
720 }
721
722 /*
723  * Determine low and high memory ranges:
724  */
725 void __init find_low_pfn_range(void)
726 {
727         /* it could update max_pfn */
728
729         if (max_pfn <= MAXMEM_PFN)
730                 lowmem_pfn_init();
731         else
732                 highmem_pfn_init();
733 }
734
735 #ifndef CONFIG_NEED_MULTIPLE_NODES
736 void __init initmem_init(unsigned long start_pfn,
737                                   unsigned long end_pfn)
738 {
739 #ifdef CONFIG_HIGHMEM
740         highstart_pfn = highend_pfn = max_pfn;
741         if (max_pfn > max_low_pfn)
742                 highstart_pfn = max_low_pfn;
743         memory_present(0, 0, highend_pfn);
744         e820_register_active_regions(0, 0, highend_pfn);
745         printk(KERN_NOTICE "%ldMB HIGHMEM available.\n",
746                 pages_to_mb(highend_pfn - highstart_pfn));
747         num_physpages = highend_pfn;
748         high_memory = (void *) __va(highstart_pfn * PAGE_SIZE - 1) + 1;
749 #else
750         memory_present(0, 0, max_low_pfn);
751         e820_register_active_regions(0, 0, max_low_pfn);
752         num_physpages = max_low_pfn;
753         high_memory = (void *) __va(max_low_pfn * PAGE_SIZE - 1) + 1;
754 #endif
755 #ifdef CONFIG_FLATMEM
756         max_mapnr = num_physpages;
757 #endif
758         printk(KERN_NOTICE "%ldMB LOWMEM available.\n",
759                         pages_to_mb(max_low_pfn));
760
761         setup_bootmem_allocator();
762 }
763 #endif /* !CONFIG_NEED_MULTIPLE_NODES */
764
765 static void __init zone_sizes_init(void)
766 {
767         unsigned long max_zone_pfns[MAX_NR_ZONES];
768         memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
769         max_zone_pfns[ZONE_DMA] =
770                 virt_to_phys((char *)MAX_DMA_ADDRESS) >> PAGE_SHIFT;
771         max_zone_pfns[ZONE_NORMAL] = max_low_pfn;
772 #ifdef CONFIG_HIGHMEM
773         max_zone_pfns[ZONE_HIGHMEM] = highend_pfn;
774 #endif
775
776         free_area_init_nodes(max_zone_pfns);
777 }
778
779 #ifdef CONFIG_NEED_MULTIPLE_NODES
780 static unsigned long __init setup_node_bootmem(int nodeid,
781                                  unsigned long start_pfn,
782                                  unsigned long end_pfn,
783                                  unsigned long bootmap)
784 {
785         unsigned long bootmap_size;
786
787         if (start_pfn > max_low_pfn)
788                 return bootmap;
789         if (end_pfn > max_low_pfn)
790                 end_pfn = max_low_pfn;
791
792         /* don't touch min_low_pfn */
793         bootmap_size = init_bootmem_node(NODE_DATA(nodeid),
794                                          bootmap >> PAGE_SHIFT,
795                                          start_pfn, end_pfn);
796         printk(KERN_INFO "  node %d low ram: %08lx - %08lx\n",
797                 nodeid, start_pfn<<PAGE_SHIFT, end_pfn<<PAGE_SHIFT);
798         printk(KERN_INFO "  node %d bootmap %08lx - %08lx\n",
799                  nodeid, bootmap, bootmap + bootmap_size);
800         free_bootmem_with_active_regions(nodeid, end_pfn);
801         early_res_to_bootmem(start_pfn<<PAGE_SHIFT, end_pfn<<PAGE_SHIFT);
802
803         return bootmap + bootmap_size;
804 }
805 #endif
806
807 void __init setup_bootmem_allocator(void)
808 {
809         int nodeid;
810         unsigned long bootmap_size, bootmap;
811         /*
812          * Initialize the boot-time allocator (with low memory only):
813          */
814         bootmap_size = bootmem_bootmap_pages(max_low_pfn)<<PAGE_SHIFT;
815         bootmap = find_e820_area(min_low_pfn<<PAGE_SHIFT,
816                                  max_pfn_mapped<<PAGE_SHIFT, bootmap_size,
817                                  PAGE_SIZE);
818         if (bootmap == -1L)
819                 panic("Cannot find bootmem map of size %ld\n", bootmap_size);
820         reserve_early(bootmap, bootmap + bootmap_size, "BOOTMAP");
821
822         printk(KERN_INFO "  mapped low ram: 0 - %08lx\n",
823                  max_pfn_mapped<<PAGE_SHIFT);
824         printk(KERN_INFO "  low ram: %08lx - %08lx\n",
825                  min_low_pfn<<PAGE_SHIFT, max_low_pfn<<PAGE_SHIFT);
826
827 #ifdef CONFIG_NEED_MULTIPLE_NODES
828         for_each_online_node(nodeid)
829                 bootmap = setup_node_bootmem(nodeid, node_start_pfn[nodeid],
830                                         node_end_pfn[nodeid], bootmap);
831 #else
832         /* don't touch min_low_pfn */
833         bootmap_size = init_bootmem_node(NODE_DATA(0), bootmap >> PAGE_SHIFT,
834                                          min_low_pfn, max_low_pfn);
835         printk(KERN_INFO "  bootmap %08lx - %08lx\n",
836                  bootmap, bootmap + bootmap_size);
837         free_bootmem_with_active_regions(0, max_low_pfn);
838         early_res_to_bootmem(0, max_low_pfn<<PAGE_SHIFT);
839 #endif
840
841         after_init_bootmem = 1;
842 }
843
844 static void __init find_early_table_space(unsigned long end, int use_pse)
845 {
846         unsigned long puds, pmds, ptes, tables, start;
847
848         puds = (end + PUD_SIZE - 1) >> PUD_SHIFT;
849         tables = roundup(puds * sizeof(pud_t), PAGE_SIZE);
850
851         pmds = (end + PMD_SIZE - 1) >> PMD_SHIFT;
852         tables += roundup(pmds * sizeof(pmd_t), PAGE_SIZE);
853
854         if (use_pse) {
855                 unsigned long extra;
856
857                 extra = end - ((end>>PMD_SHIFT) << PMD_SHIFT);
858                 extra += PMD_SIZE;
859                 ptes = (extra + PAGE_SIZE - 1) >> PAGE_SHIFT;
860         } else
861                 ptes = (end + PAGE_SIZE - 1) >> PAGE_SHIFT;
862
863         tables += roundup(ptes * sizeof(pte_t), PAGE_SIZE);
864
865         /* for fixmap */
866         tables += roundup(__end_of_fixed_addresses * sizeof(pte_t), PAGE_SIZE);
867
868         /*
869          * RED-PEN putting page tables only on node 0 could
870          * cause a hotspot and fill up ZONE_DMA. The page tables
871          * need roughly 0.5KB per GB.
872          */
873         start = 0x7000;
874         table_start = find_e820_area(start, max_pfn_mapped<<PAGE_SHIFT,
875                                         tables, PAGE_SIZE);
876         if (table_start == -1UL)
877                 panic("Cannot find space for the kernel page tables");
878
879         table_start >>= PAGE_SHIFT;
880         table_end = table_start;
881         table_top = table_start + (tables>>PAGE_SHIFT);
882
883         printk(KERN_DEBUG "kernel direct mapping tables up to %lx @ %lx-%lx\n",
884                 end, table_start << PAGE_SHIFT,
885                 (table_start << PAGE_SHIFT) + tables);
886 }
887
888 unsigned long __init_refok init_memory_mapping(unsigned long start,
889                                                 unsigned long end)
890 {
891         pgd_t *pgd_base = swapper_pg_dir;
892         unsigned long start_pfn, end_pfn;
893         unsigned long big_page_start;
894 #ifdef CONFIG_DEBUG_PAGEALLOC
895         /*
896          * For CONFIG_DEBUG_PAGEALLOC, identity mapping will use small pages.
897          * This will simplify cpa(), which otherwise needs to support splitting
898          * large pages into small in interrupt context, etc.
899          */
900         int use_pse = 0;
901 #else
902         int use_pse = cpu_has_pse;
903 #endif
904
905         /*
906          * Find space for the kernel direct mapping tables.
907          */
908         if (!after_init_bootmem)
909                 find_early_table_space(end, use_pse);
910
911 #ifdef CONFIG_X86_PAE
912         set_nx();
913         if (nx_enabled)
914                 printk(KERN_INFO "NX (Execute Disable) protection: active\n");
915 #endif
916
917         /* Enable PSE if available */
918         if (cpu_has_pse)
919                 set_in_cr4(X86_CR4_PSE);
920
921         /* Enable PGE if available */
922         if (cpu_has_pge) {
923                 set_in_cr4(X86_CR4_PGE);
924                 __supported_pte_mask |= _PAGE_GLOBAL;
925         }
926
927         /*
928          * Don't use a large page for the first 2/4MB of memory
929          * because there are often fixed size MTRRs in there
930          * and overlapping MTRRs into large pages can cause
931          * slowdowns.
932          */
933         big_page_start = PMD_SIZE;
934
935         if (start < big_page_start) {
936                 start_pfn = start >> PAGE_SHIFT;
937                 end_pfn = min(big_page_start>>PAGE_SHIFT, end>>PAGE_SHIFT);
938         } else {
939                 /* head is not big page alignment ? */
940                 start_pfn = start >> PAGE_SHIFT;
941                 end_pfn = ((start + (PMD_SIZE - 1))>>PMD_SHIFT)
942                                  << (PMD_SHIFT - PAGE_SHIFT);
943         }
944         if (start_pfn < end_pfn)
945                 kernel_physical_mapping_init(pgd_base, start_pfn, end_pfn, 0);
946
947         /* big page range */
948         start_pfn = ((start + (PMD_SIZE - 1))>>PMD_SHIFT)
949                          << (PMD_SHIFT - PAGE_SHIFT);
950         if (start_pfn < (big_page_start >> PAGE_SHIFT))
951                 start_pfn =  big_page_start >> PAGE_SHIFT;
952         end_pfn = (end>>PMD_SHIFT) << (PMD_SHIFT - PAGE_SHIFT);
953         if (start_pfn < end_pfn)
954                 kernel_physical_mapping_init(pgd_base, start_pfn, end_pfn,
955                                              use_pse);
956
957         /* tail is not big page alignment ? */
958         start_pfn = end_pfn;
959         if (start_pfn > (big_page_start>>PAGE_SHIFT)) {
960                 end_pfn = end >> PAGE_SHIFT;
961                 if (start_pfn < end_pfn)
962                         kernel_physical_mapping_init(pgd_base, start_pfn,
963                                                          end_pfn, 0);
964         }
965
966         early_ioremap_page_table_range_init(pgd_base);
967
968         load_cr3(swapper_pg_dir);
969
970         __flush_tlb_all();
971
972         if (!after_init_bootmem)
973                 reserve_early(table_start << PAGE_SHIFT,
974                                  table_end << PAGE_SHIFT, "PGTABLE");
975
976         if (!after_init_bootmem)
977                 early_memtest(start, end);
978
979         return end >> PAGE_SHIFT;
980 }
981
982
983 /*
984  * paging_init() sets up the page tables - note that the first 8MB are
985  * already mapped by head.S.
986  *
987  * This routines also unmaps the page at virtual kernel address 0, so
988  * that we can trap those pesky NULL-reference errors in the kernel.
989  */
990 void __init paging_init(void)
991 {
992         pagetable_init();
993
994         __flush_tlb_all();
995
996         kmap_init();
997
998         /*
999          * NOTE: at this point the bootmem allocator is fully available.
1000          */
1001         sparse_init();
1002         zone_sizes_init();
1003 }
1004
1005 /*
1006  * Test if the WP bit works in supervisor mode. It isn't supported on 386's
1007  * and also on some strange 486's. All 586+'s are OK. This used to involve
1008  * black magic jumps to work around some nasty CPU bugs, but fortunately the
1009  * switch to using exceptions got rid of all that.
1010  */
1011 static void __init test_wp_bit(void)
1012 {
1013         printk(KERN_INFO
1014   "Checking if this processor honours the WP bit even in supervisor mode...");
1015
1016         /* Any page-aligned address will do, the test is non-destructive */
1017         __set_fixmap(FIX_WP_TEST, __pa(&swapper_pg_dir), PAGE_READONLY);
1018         boot_cpu_data.wp_works_ok = do_test_wp_bit();
1019         clear_fixmap(FIX_WP_TEST);
1020
1021         if (!boot_cpu_data.wp_works_ok) {
1022                 printk(KERN_CONT "No.\n");
1023 #ifdef CONFIG_X86_WP_WORKS_OK
1024                 panic(
1025   "This kernel doesn't support CPU's with broken WP. Recompile it for a 386!");
1026 #endif
1027         } else {
1028                 printk(KERN_CONT "Ok.\n");
1029         }
1030 }
1031
1032 static struct kcore_list kcore_mem, kcore_vmalloc;
1033
1034 void __init mem_init(void)
1035 {
1036         int codesize, reservedpages, datasize, initsize;
1037         int tmp;
1038
1039         pci_iommu_alloc();
1040
1041 #ifdef CONFIG_FLATMEM
1042         BUG_ON(!mem_map);
1043 #endif
1044         /* this will put all low memory onto the freelists */
1045         totalram_pages += free_all_bootmem();
1046
1047         reservedpages = 0;
1048         for (tmp = 0; tmp < max_low_pfn; tmp++)
1049                 /*
1050                  * Only count reserved RAM pages:
1051                  */
1052                 if (page_is_ram(tmp) && PageReserved(pfn_to_page(tmp)))
1053                         reservedpages++;
1054
1055         set_highmem_pages_init();
1056
1057         codesize =  (unsigned long) &_etext - (unsigned long) &_text;
1058         datasize =  (unsigned long) &_edata - (unsigned long) &_etext;
1059         initsize =  (unsigned long) &__init_end - (unsigned long) &__init_begin;
1060
1061         kclist_add(&kcore_mem, __va(0), max_low_pfn << PAGE_SHIFT);
1062         kclist_add(&kcore_vmalloc, (void *)VMALLOC_START,
1063                    VMALLOC_END-VMALLOC_START);
1064
1065         printk(KERN_INFO "Memory: %luk/%luk available (%dk kernel code, "
1066                         "%dk reserved, %dk data, %dk init, %ldk highmem)\n",
1067                 (unsigned long) nr_free_pages() << (PAGE_SHIFT-10),
1068                 num_physpages << (PAGE_SHIFT-10),
1069                 codesize >> 10,
1070                 reservedpages << (PAGE_SHIFT-10),
1071                 datasize >> 10,
1072                 initsize >> 10,
1073                 (unsigned long) (totalhigh_pages << (PAGE_SHIFT-10))
1074                );
1075
1076         printk(KERN_INFO "virtual kernel memory layout:\n"
1077                 "    fixmap  : 0x%08lx - 0x%08lx   (%4ld kB)\n"
1078 #ifdef CONFIG_HIGHMEM
1079                 "    pkmap   : 0x%08lx - 0x%08lx   (%4ld kB)\n"
1080 #endif
1081                 "    vmalloc : 0x%08lx - 0x%08lx   (%4ld MB)\n"
1082                 "    lowmem  : 0x%08lx - 0x%08lx   (%4ld MB)\n"
1083                 "      .init : 0x%08lx - 0x%08lx   (%4ld kB)\n"
1084                 "      .data : 0x%08lx - 0x%08lx   (%4ld kB)\n"
1085                 "      .text : 0x%08lx - 0x%08lx   (%4ld kB)\n",
1086                 FIXADDR_START, FIXADDR_TOP,
1087                 (FIXADDR_TOP - FIXADDR_START) >> 10,
1088
1089 #ifdef CONFIG_HIGHMEM
1090                 PKMAP_BASE, PKMAP_BASE+LAST_PKMAP*PAGE_SIZE,
1091                 (LAST_PKMAP*PAGE_SIZE) >> 10,
1092 #endif
1093
1094                 VMALLOC_START, VMALLOC_END,
1095                 (VMALLOC_END - VMALLOC_START) >> 20,
1096
1097                 (unsigned long)__va(0), (unsigned long)high_memory,
1098                 ((unsigned long)high_memory - (unsigned long)__va(0)) >> 20,
1099
1100                 (unsigned long)&__init_begin, (unsigned long)&__init_end,
1101                 ((unsigned long)&__init_end -
1102                  (unsigned long)&__init_begin) >> 10,
1103
1104                 (unsigned long)&_etext, (unsigned long)&_edata,
1105                 ((unsigned long)&_edata - (unsigned long)&_etext) >> 10,
1106
1107                 (unsigned long)&_text, (unsigned long)&_etext,
1108                 ((unsigned long)&_etext - (unsigned long)&_text) >> 10);
1109
1110         /*
1111          * Check boundaries twice: Some fundamental inconsistencies can
1112          * be detected at build time already.
1113          */
1114 #define __FIXADDR_TOP (-PAGE_SIZE)
1115 #ifdef CONFIG_HIGHMEM
1116         BUILD_BUG_ON(PKMAP_BASE + LAST_PKMAP*PAGE_SIZE  > FIXADDR_START);
1117         BUILD_BUG_ON(VMALLOC_END                        > PKMAP_BASE);
1118 #endif
1119 #define high_memory (-128UL << 20)
1120         BUILD_BUG_ON(VMALLOC_START                      >= VMALLOC_END);
1121 #undef high_memory
1122 #undef __FIXADDR_TOP
1123
1124 #ifdef CONFIG_HIGHMEM
1125         BUG_ON(PKMAP_BASE + LAST_PKMAP*PAGE_SIZE        > FIXADDR_START);
1126         BUG_ON(VMALLOC_END                              > PKMAP_BASE);
1127 #endif
1128         BUG_ON(VMALLOC_START                            >= VMALLOC_END);
1129         BUG_ON((unsigned long)high_memory               > VMALLOC_START);
1130
1131         if (boot_cpu_data.wp_works_ok < 0)
1132                 test_wp_bit();
1133
1134         save_pg_dir();
1135         zap_low_mappings();
1136 }
1137
1138 #ifdef CONFIG_MEMORY_HOTPLUG
1139 int arch_add_memory(int nid, u64 start, u64 size)
1140 {
1141         struct pglist_data *pgdata = NODE_DATA(nid);
1142         struct zone *zone = pgdata->node_zones + ZONE_HIGHMEM;
1143         unsigned long start_pfn = start >> PAGE_SHIFT;
1144         unsigned long nr_pages = size >> PAGE_SHIFT;
1145
1146         return __add_pages(nid, zone, start_pfn, nr_pages);
1147 }
1148 #endif
1149
1150 /*
1151  * This function cannot be __init, since exceptions don't work in that
1152  * section.  Put this after the callers, so that it cannot be inlined.
1153  */
1154 static noinline int do_test_wp_bit(void)
1155 {
1156         char tmp_reg;
1157         int flag;
1158
1159         __asm__ __volatile__(
1160                 "       movb %0, %1     \n"
1161                 "1:     movb %1, %0     \n"
1162                 "       xorl %2, %2     \n"
1163                 "2:                     \n"
1164                 _ASM_EXTABLE(1b,2b)
1165                 :"=m" (*(char *)fix_to_virt(FIX_WP_TEST)),
1166                  "=q" (tmp_reg),
1167                  "=r" (flag)
1168                 :"2" (1)
1169                 :"memory");
1170
1171         return flag;
1172 }
1173
1174 #ifdef CONFIG_DEBUG_RODATA
1175 const int rodata_test_data = 0xC3;
1176 EXPORT_SYMBOL_GPL(rodata_test_data);
1177
1178 void mark_rodata_ro(void)
1179 {
1180         unsigned long start = PFN_ALIGN(_text);
1181         unsigned long size = PFN_ALIGN(_etext) - start;
1182
1183 #ifndef CONFIG_DYNAMIC_FTRACE
1184         /* Dynamic tracing modifies the kernel text section */
1185         set_pages_ro(virt_to_page(start), size >> PAGE_SHIFT);
1186         printk(KERN_INFO "Write protecting the kernel text: %luk\n",
1187                 size >> 10);
1188
1189 #ifdef CONFIG_CPA_DEBUG
1190         printk(KERN_INFO "Testing CPA: Reverting %lx-%lx\n",
1191                 start, start+size);
1192         set_pages_rw(virt_to_page(start), size>>PAGE_SHIFT);
1193
1194         printk(KERN_INFO "Testing CPA: write protecting again\n");
1195         set_pages_ro(virt_to_page(start), size>>PAGE_SHIFT);
1196 #endif
1197 #endif /* CONFIG_DYNAMIC_FTRACE */
1198
1199         start += size;
1200         size = (unsigned long)__end_rodata - start;
1201         set_pages_ro(virt_to_page(start), size >> PAGE_SHIFT);
1202         printk(KERN_INFO "Write protecting the kernel read-only data: %luk\n",
1203                 size >> 10);
1204         rodata_test();
1205
1206 #ifdef CONFIG_CPA_DEBUG
1207         printk(KERN_INFO "Testing CPA: undo %lx-%lx\n", start, start + size);
1208         set_pages_rw(virt_to_page(start), size >> PAGE_SHIFT);
1209
1210         printk(KERN_INFO "Testing CPA: write protecting again\n");
1211         set_pages_ro(virt_to_page(start), size >> PAGE_SHIFT);
1212 #endif
1213 }
1214 #endif
1215
1216 #ifdef CONFIG_BLK_DEV_INITRD
1217 void free_initrd_mem(unsigned long start, unsigned long end)
1218 {
1219         free_init_pages("initrd memory", start, end);
1220 }
1221 #endif
1222
1223 int __init reserve_bootmem_generic(unsigned long phys, unsigned long len,
1224                                    int flags)
1225 {
1226         return reserve_bootmem(phys, len, flags);
1227 }