3 * Copyright (C) 1995 Linus Torvalds
5 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
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>
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/pfn.h>
25 #include <linux/poison.h>
26 #include <linux/bootmem.h>
27 #include <linux/slab.h>
28 #include <linux/proc_fs.h>
29 #include <linux/memory_hotplug.h>
30 #include <linux/initrd.h>
31 #include <linux/cpumask.h>
34 #include <asm/processor.h>
35 #include <asm/system.h>
36 #include <asm/uaccess.h>
37 #include <asm/pgtable.h>
39 #include <asm/fixmap.h>
44 #include <asm/tlbflush.h>
45 #include <asm/pgalloc.h>
46 #include <asm/sections.h>
47 #include <asm/paravirt.h>
48 #include <asm/setup.h>
49 #include <asm/cacheflush.h>
51 unsigned int __VMALLOC_RESERVE = 128 << 20;
53 unsigned long max_pfn_mapped;
55 DEFINE_PER_CPU(struct mmu_gather, mmu_gathers);
56 unsigned long highstart_pfn, highend_pfn;
58 static noinline int do_test_wp_bit(void);
61 * Creates a middle page table and puts a pointer to it in the
62 * given global directory entry. This only returns the gd entry
63 * in non-PAE compilation mode, since the middle layer is folded.
65 static pmd_t * __init one_md_table_init(pgd_t *pgd)
71 if (!(pgd_val(*pgd) & _PAGE_PRESENT)) {
72 pmd_table = (pmd_t *) alloc_bootmem_low_pages(PAGE_SIZE);
74 paravirt_alloc_pmd(&init_mm, __pa(pmd_table) >> PAGE_SHIFT);
75 set_pgd(pgd, __pgd(__pa(pmd_table) | _PAGE_PRESENT));
76 pud = pud_offset(pgd, 0);
77 BUG_ON(pmd_table != pmd_offset(pud, 0));
80 pud = pud_offset(pgd, 0);
81 pmd_table = pmd_offset(pud, 0);
87 * Create a page table and place a pointer to it in a middle page
90 static pte_t * __init one_page_table_init(pmd_t *pmd)
92 if (!(pmd_val(*pmd) & _PAGE_PRESENT)) {
93 pte_t *page_table = NULL;
95 #ifdef CONFIG_DEBUG_PAGEALLOC
96 page_table = (pte_t *) alloc_bootmem_pages(PAGE_SIZE);
100 (pte_t *)alloc_bootmem_low_pages(PAGE_SIZE);
103 paravirt_alloc_pte(&init_mm, __pa(page_table) >> PAGE_SHIFT);
104 set_pmd(pmd, __pmd(__pa(page_table) | _PAGE_TABLE));
105 BUG_ON(page_table != pte_offset_kernel(pmd, 0));
108 return pte_offset_kernel(pmd, 0);
112 * This function initializes a certain range of kernel virtual memory
113 * with new bootmem page tables, everywhere page tables are missing in
116 * NOTE: The pagetables are allocated contiguous on the physical space
117 * so we can cache the place of the first one and move around without
118 * checking the pgd every time.
121 page_table_range_init(unsigned long start, unsigned long end, pgd_t *pgd_base)
123 int pgd_idx, pmd_idx;
129 pgd_idx = pgd_index(vaddr);
130 pmd_idx = pmd_index(vaddr);
131 pgd = pgd_base + pgd_idx;
133 for ( ; (pgd_idx < PTRS_PER_PGD) && (vaddr != end); pgd++, pgd_idx++) {
134 pmd = one_md_table_init(pgd);
135 pmd = pmd + pmd_index(vaddr);
136 for (; (pmd_idx < PTRS_PER_PMD) && (vaddr != end);
138 one_page_table_init(pmd);
146 static inline int is_kernel_text(unsigned long addr)
148 if (addr >= PAGE_OFFSET && addr <= (unsigned long)__init_end)
154 * This maps the physical memory to kernel virtual address space, a total
155 * of max_low_pfn pages, by creating page tables starting from address
158 static void __init kernel_physical_mapping_init(pgd_t *pgd_base)
160 int pgd_idx, pmd_idx, pte_ofs;
166 pgd_idx = pgd_index(PAGE_OFFSET);
167 pgd = pgd_base + pgd_idx;
170 for (; pgd_idx < PTRS_PER_PGD; pgd++, pgd_idx++) {
171 pmd = one_md_table_init(pgd);
172 if (pfn >= max_low_pfn)
176 pmd_idx < PTRS_PER_PMD && pfn < max_low_pfn;
178 unsigned int addr = pfn * PAGE_SIZE + PAGE_OFFSET;
181 * Map with big pages if possible, otherwise
182 * create normal page tables:
184 * Don't use a large page for the first 2/4MB of memory
185 * because there are often fixed size MTRRs in there
186 * and overlapping MTRRs into large pages can cause
189 if (cpu_has_pse && !(pgd_idx == 0 && pmd_idx == 0)) {
191 pgprot_t prot = PAGE_KERNEL_LARGE;
193 addr2 = (pfn + PTRS_PER_PTE-1) * PAGE_SIZE +
194 PAGE_OFFSET + PAGE_SIZE-1;
196 if (is_kernel_text(addr) ||
197 is_kernel_text(addr2))
198 prot = PAGE_KERNEL_LARGE_EXEC;
200 set_pmd(pmd, pfn_pmd(pfn, prot));
203 max_pfn_mapped = pfn;
206 pte = one_page_table_init(pmd);
209 pte_ofs < PTRS_PER_PTE && pfn < max_low_pfn;
210 pte++, pfn++, pte_ofs++, addr += PAGE_SIZE) {
211 pgprot_t prot = PAGE_KERNEL;
213 if (is_kernel_text(addr))
214 prot = PAGE_KERNEL_EXEC;
216 set_pte(pte, pfn_pte(pfn, prot));
218 max_pfn_mapped = pfn;
223 static inline int page_kills_ppro(unsigned long pagenr)
225 if (pagenr >= 0x70000 && pagenr <= 0x7003F)
231 * devmem_is_allowed() checks to see if /dev/mem access to a certain address
232 * is valid. The argument is a physical page number.
235 * On x86, access has to be given to the first megabyte of ram because that area
236 * contains bios code and data regions used by X and dosemu and similar apps.
237 * Access has to be given to non-kernel-ram areas as well, these contain the PCI
238 * mmio resources as well as potential bios/acpi data regions.
240 int devmem_is_allowed(unsigned long pagenr)
244 if (!page_is_ram(pagenr))
249 #ifdef CONFIG_HIGHMEM
253 static inline pte_t *kmap_get_fixmap_pte(unsigned long vaddr)
255 return pte_offset_kernel(pmd_offset(pud_offset(pgd_offset_k(vaddr),
256 vaddr), vaddr), vaddr);
259 static void __init kmap_init(void)
261 unsigned long kmap_vstart;
264 * Cache the first kmap pte:
266 kmap_vstart = __fix_to_virt(FIX_KMAP_BEGIN);
267 kmap_pte = kmap_get_fixmap_pte(kmap_vstart);
269 kmap_prot = PAGE_KERNEL;
272 static void __init permanent_kmaps_init(pgd_t *pgd_base)
281 page_table_range_init(vaddr, vaddr + PAGE_SIZE*LAST_PKMAP, pgd_base);
283 pgd = swapper_pg_dir + pgd_index(vaddr);
284 pud = pud_offset(pgd, vaddr);
285 pmd = pmd_offset(pud, vaddr);
286 pte = pte_offset_kernel(pmd, vaddr);
287 pkmap_page_table = pte;
290 void __init add_one_highpage_init(struct page *page, int pfn, int bad_ppro)
292 if (page_is_ram(pfn) && !(bad_ppro && page_kills_ppro(pfn))) {
293 ClearPageReserved(page);
294 init_page_count(page);
298 SetPageReserved(page);
302 static void __init set_highmem_pages_init(int bad_ppro)
306 for (pfn = highstart_pfn; pfn < highend_pfn; pfn++) {
308 * Holes under sparsemem might not have no mem_map[]:
311 add_one_highpage_init(pfn_to_page(pfn), pfn, bad_ppro);
313 totalram_pages += totalhigh_pages;
315 #endif /* !CONFIG_NUMA */
318 # define kmap_init() do { } while (0)
319 # define permanent_kmaps_init(pgd_base) do { } while (0)
320 # define set_highmem_pages_init(bad_ppro) do { } while (0)
321 #endif /* CONFIG_HIGHMEM */
323 pteval_t __PAGE_KERNEL = _PAGE_KERNEL;
324 EXPORT_SYMBOL(__PAGE_KERNEL);
326 pteval_t __PAGE_KERNEL_EXEC = _PAGE_KERNEL_EXEC;
328 void __init native_pagetable_setup_start(pgd_t *base)
330 unsigned long pfn, va;
337 * Remove any mappings which extend past the end of physical
338 * memory from the boot time page table:
340 for (pfn = max_low_pfn + 1; pfn < 1<<(32-PAGE_SHIFT); pfn++) {
341 va = PAGE_OFFSET + (pfn<<PAGE_SHIFT);
342 pgd = base + pgd_index(va);
343 if (!pgd_present(*pgd))
346 pud = pud_offset(pgd, va);
347 pmd = pmd_offset(pud, va);
348 if (!pmd_present(*pmd))
351 pte = pte_offset_kernel(pmd, va);
352 if (!pte_present(*pte))
355 pte_clear(NULL, va, pte);
357 paravirt_alloc_pmd(&init_mm, __pa(base) >> PAGE_SHIFT);
360 void __init native_pagetable_setup_done(pgd_t *base)
365 * Build a proper pagetable for the kernel mappings. Up until this
366 * point, we've been running on some set of pagetables constructed by
369 * If we're booting on native hardware, this will be a pagetable
370 * constructed in arch/x86/kernel/head_32.S. The root of the
371 * pagetable will be swapper_pg_dir.
373 * If we're booting paravirtualized under a hypervisor, then there are
374 * more options: we may already be running PAE, and the pagetable may
375 * or may not be based in swapper_pg_dir. In any case,
376 * paravirt_pagetable_setup_start() will set up swapper_pg_dir
377 * appropriately for the rest of the initialization to work.
379 * In general, pagetable_init() assumes that the pagetable may already
380 * be partially populated, and so it avoids stomping on any existing
383 static void __init pagetable_init(void)
385 pgd_t *pgd_base = swapper_pg_dir;
386 unsigned long vaddr, end;
388 paravirt_pagetable_setup_start(pgd_base);
390 /* Enable PSE if available */
392 set_in_cr4(X86_CR4_PSE);
394 /* Enable PGE if available */
396 set_in_cr4(X86_CR4_PGE);
397 __PAGE_KERNEL |= _PAGE_GLOBAL;
398 __PAGE_KERNEL_EXEC |= _PAGE_GLOBAL;
401 kernel_physical_mapping_init(pgd_base);
405 * Fixed mappings, only the page table structure has to be
406 * created - mappings will be set by set_fixmap():
408 early_ioremap_clear();
409 vaddr = __fix_to_virt(__end_of_fixed_addresses - 1) & PMD_MASK;
410 end = (FIXADDR_TOP + PMD_SIZE - 1) & PMD_MASK;
411 page_table_range_init(vaddr, end, pgd_base);
412 early_ioremap_reset();
414 permanent_kmaps_init(pgd_base);
416 paravirt_pagetable_setup_done(pgd_base);
419 #ifdef CONFIG_ACPI_SLEEP
421 * ACPI suspend needs this for resume, because things like the intel-agp
422 * driver might have split up a kernel 4MB mapping.
424 char swsusp_pg_dir[PAGE_SIZE]
425 __attribute__ ((aligned(PAGE_SIZE)));
427 static inline void save_pg_dir(void)
429 memcpy(swsusp_pg_dir, swapper_pg_dir, PAGE_SIZE);
431 #else /* !CONFIG_ACPI_SLEEP */
432 static inline void save_pg_dir(void)
435 #endif /* !CONFIG_ACPI_SLEEP */
437 void zap_low_mappings(void)
444 * Zap initial low-memory mappings.
446 * Note that "pgd_clear()" doesn't do it for
447 * us, because pgd_clear() is a no-op on i386.
449 for (i = 0; i < KERNEL_PGD_BOUNDARY; i++) {
450 #ifdef CONFIG_X86_PAE
451 set_pgd(swapper_pg_dir+i, __pgd(1 + __pa(empty_zero_page)));
453 set_pgd(swapper_pg_dir+i, __pgd(0));
461 pteval_t __supported_pte_mask __read_mostly = ~_PAGE_NX;
462 EXPORT_SYMBOL_GPL(__supported_pte_mask);
464 #ifdef CONFIG_X86_PAE
466 static int disable_nx __initdata;
471 * Control non executable mappings.
476 static int __init noexec_setup(char *str)
478 if (!str || !strcmp(str, "on")) {
480 __supported_pte_mask |= _PAGE_NX;
484 if (!strcmp(str, "off")) {
486 __supported_pte_mask &= ~_PAGE_NX;
494 early_param("noexec", noexec_setup);
496 static void __init set_nx(void)
498 unsigned int v[4], l, h;
500 if (cpu_has_pae && (cpuid_eax(0x80000000) > 0x80000001)) {
501 cpuid(0x80000001, &v[0], &v[1], &v[2], &v[3]);
503 if ((v[3] & (1 << 20)) && !disable_nx) {
504 rdmsr(MSR_EFER, l, h);
506 wrmsr(MSR_EFER, l, h);
508 __supported_pte_mask |= _PAGE_NX;
515 * paging_init() sets up the page tables - note that the first 8MB are
516 * already mapped by head.S.
518 * This routines also unmaps the page at virtual kernel address 0, so
519 * that we can trap those pesky NULL-reference errors in the kernel.
521 void __init paging_init(void)
523 #ifdef CONFIG_X86_PAE
526 printk(KERN_INFO "NX (Execute Disable) protection: active\n");
530 load_cr3(swapper_pg_dir);
538 * Test if the WP bit works in supervisor mode. It isn't supported on 386's
539 * and also on some strange 486's. All 586+'s are OK. This used to involve
540 * black magic jumps to work around some nasty CPU bugs, but fortunately the
541 * switch to using exceptions got rid of all that.
543 static void __init test_wp_bit(void)
546 "Checking if this processor honours the WP bit even in supervisor mode...");
548 /* Any page-aligned address will do, the test is non-destructive */
549 __set_fixmap(FIX_WP_TEST, __pa(&swapper_pg_dir), PAGE_READONLY);
550 boot_cpu_data.wp_works_ok = do_test_wp_bit();
551 clear_fixmap(FIX_WP_TEST);
553 if (!boot_cpu_data.wp_works_ok) {
554 printk(KERN_CONT "No.\n");
555 #ifdef CONFIG_X86_WP_WORKS_OK
557 "This kernel doesn't support CPU's with broken WP. Recompile it for a 386!");
560 printk(KERN_CONT "Ok.\n");
564 static struct kcore_list kcore_mem, kcore_vmalloc;
566 void __init mem_init(void)
568 int codesize, reservedpages, datasize, initsize;
571 #ifdef CONFIG_FLATMEM
574 bad_ppro = ppro_with_ram_bug();
576 #ifdef CONFIG_HIGHMEM
577 /* check that fixmap and pkmap do not overlap */
578 if (PKMAP_BASE + LAST_PKMAP*PAGE_SIZE >= FIXADDR_START) {
580 "fixmap and kmap areas overlap - this will crash\n");
581 printk(KERN_ERR "pkstart: %lxh pkend: %lxh fixstart %lxh\n",
582 PKMAP_BASE, PKMAP_BASE + LAST_PKMAP*PAGE_SIZE,
587 /* this will put all low memory onto the freelists */
588 totalram_pages += free_all_bootmem();
591 for (tmp = 0; tmp < max_low_pfn; tmp++)
593 * Only count reserved RAM pages:
595 if (page_is_ram(tmp) && PageReserved(pfn_to_page(tmp)))
598 set_highmem_pages_init(bad_ppro);
600 codesize = (unsigned long) &_etext - (unsigned long) &_text;
601 datasize = (unsigned long) &_edata - (unsigned long) &_etext;
602 initsize = (unsigned long) &__init_end - (unsigned long) &__init_begin;
604 kclist_add(&kcore_mem, __va(0), max_low_pfn << PAGE_SHIFT);
605 kclist_add(&kcore_vmalloc, (void *)VMALLOC_START,
606 VMALLOC_END-VMALLOC_START);
608 printk(KERN_INFO "Memory: %luk/%luk available (%dk kernel code, "
609 "%dk reserved, %dk data, %dk init, %ldk highmem)\n",
610 (unsigned long) nr_free_pages() << (PAGE_SHIFT-10),
611 num_physpages << (PAGE_SHIFT-10),
613 reservedpages << (PAGE_SHIFT-10),
616 (unsigned long) (totalhigh_pages << (PAGE_SHIFT-10))
619 #if 1 /* double-sanity-check paranoia */
620 printk(KERN_INFO "virtual kernel memory layout:\n"
621 " fixmap : 0x%08lx - 0x%08lx (%4ld kB)\n"
622 #ifdef CONFIG_HIGHMEM
623 " pkmap : 0x%08lx - 0x%08lx (%4ld kB)\n"
625 " vmalloc : 0x%08lx - 0x%08lx (%4ld MB)\n"
626 " lowmem : 0x%08lx - 0x%08lx (%4ld MB)\n"
627 " .init : 0x%08lx - 0x%08lx (%4ld kB)\n"
628 " .data : 0x%08lx - 0x%08lx (%4ld kB)\n"
629 " .text : 0x%08lx - 0x%08lx (%4ld kB)\n",
630 FIXADDR_START, FIXADDR_TOP,
631 (FIXADDR_TOP - FIXADDR_START) >> 10,
633 #ifdef CONFIG_HIGHMEM
634 PKMAP_BASE, PKMAP_BASE+LAST_PKMAP*PAGE_SIZE,
635 (LAST_PKMAP*PAGE_SIZE) >> 10,
638 VMALLOC_START, VMALLOC_END,
639 (VMALLOC_END - VMALLOC_START) >> 20,
641 (unsigned long)__va(0), (unsigned long)high_memory,
642 ((unsigned long)high_memory - (unsigned long)__va(0)) >> 20,
644 (unsigned long)&__init_begin, (unsigned long)&__init_end,
645 ((unsigned long)&__init_end -
646 (unsigned long)&__init_begin) >> 10,
648 (unsigned long)&_etext, (unsigned long)&_edata,
649 ((unsigned long)&_edata - (unsigned long)&_etext) >> 10,
651 (unsigned long)&_text, (unsigned long)&_etext,
652 ((unsigned long)&_etext - (unsigned long)&_text) >> 10);
654 #ifdef CONFIG_HIGHMEM
655 BUG_ON(PKMAP_BASE + LAST_PKMAP*PAGE_SIZE > FIXADDR_START);
656 BUG_ON(VMALLOC_END > PKMAP_BASE);
658 BUG_ON(VMALLOC_START > VMALLOC_END);
659 BUG_ON((unsigned long)high_memory > VMALLOC_START);
660 #endif /* double-sanity-check paranoia */
662 if (boot_cpu_data.wp_works_ok < 0)
668 * Subtle. SMP is doing it's boot stuff late (because it has to
669 * fork idle threads) - but it also needs low mappings for the
670 * protected-mode entry to work. We zap these entries only after
671 * the WP-bit has been tested.
678 #ifdef CONFIG_MEMORY_HOTPLUG
679 int arch_add_memory(int nid, u64 start, u64 size)
681 struct pglist_data *pgdata = NODE_DATA(nid);
682 struct zone *zone = pgdata->node_zones + ZONE_HIGHMEM;
683 unsigned long start_pfn = start >> PAGE_SHIFT;
684 unsigned long nr_pages = size >> PAGE_SHIFT;
686 return __add_pages(zone, start_pfn, nr_pages);
691 * This function cannot be __init, since exceptions don't work in that
692 * section. Put this after the callers, so that it cannot be inlined.
694 static noinline int do_test_wp_bit(void)
699 __asm__ __volatile__(
705 :"=m" (*(char *)fix_to_virt(FIX_WP_TEST)),
714 #ifdef CONFIG_DEBUG_RODATA
715 const int rodata_test_data = 0xC3;
716 EXPORT_SYMBOL_GPL(rodata_test_data);
718 void mark_rodata_ro(void)
720 unsigned long start = PFN_ALIGN(_text);
721 unsigned long size = PFN_ALIGN(_etext) - start;
723 set_pages_ro(virt_to_page(start), size >> PAGE_SHIFT);
724 printk(KERN_INFO "Write protecting the kernel text: %luk\n",
727 #ifdef CONFIG_CPA_DEBUG
728 printk(KERN_INFO "Testing CPA: Reverting %lx-%lx\n",
730 set_pages_rw(virt_to_page(start), size>>PAGE_SHIFT);
732 printk(KERN_INFO "Testing CPA: write protecting again\n");
733 set_pages_ro(virt_to_page(start), size>>PAGE_SHIFT);
736 size = (unsigned long)__end_rodata - start;
737 set_pages_ro(virt_to_page(start), size >> PAGE_SHIFT);
738 printk(KERN_INFO "Write protecting the kernel read-only data: %luk\n",
742 #ifdef CONFIG_CPA_DEBUG
743 printk(KERN_INFO "Testing CPA: undo %lx-%lx\n", start, start + size);
744 set_pages_rw(virt_to_page(start), size >> PAGE_SHIFT);
746 printk(KERN_INFO "Testing CPA: write protecting again\n");
747 set_pages_ro(virt_to_page(start), size >> PAGE_SHIFT);
752 void free_init_pages(char *what, unsigned long begin, unsigned long end)
754 #ifdef CONFIG_DEBUG_PAGEALLOC
756 * If debugging page accesses then do not free this memory but
757 * mark them not present - any buggy init-section access will
758 * create a kernel page fault:
760 printk(KERN_INFO "debug: unmapping init memory %08lx..%08lx\n",
761 begin, PAGE_ALIGN(end));
762 set_memory_np(begin, (end - begin) >> PAGE_SHIFT);
767 * We just marked the kernel text read only above, now that
768 * we are going to free part of that, we need to make that
771 set_memory_rw(begin, (end - begin) >> PAGE_SHIFT);
773 for (addr = begin; addr < end; addr += PAGE_SIZE) {
774 ClearPageReserved(virt_to_page(addr));
775 init_page_count(virt_to_page(addr));
776 memset((void *)addr, POISON_FREE_INITMEM, PAGE_SIZE);
780 printk(KERN_INFO "Freeing %s: %luk freed\n", what, (end - begin) >> 10);
784 void free_initmem(void)
786 free_init_pages("unused kernel memory",
787 (unsigned long)(&__init_begin),
788 (unsigned long)(&__init_end));
791 #ifdef CONFIG_BLK_DEV_INITRD
792 void free_initrd_mem(unsigned long start, unsigned long end)
794 free_init_pages("initrd memory", start, end);