2 * linux/arch/i386/mm/init.c
4 * Copyright (C) 1995 Linus Torvalds
6 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
9 #include <linux/config.h>
10 #include <linux/module.h>
11 #include <linux/signal.h>
12 #include <linux/sched.h>
13 #include <linux/kernel.h>
14 #include <linux/errno.h>
15 #include <linux/string.h>
16 #include <linux/types.h>
17 #include <linux/ptrace.h>
18 #include <linux/mman.h>
20 #include <linux/hugetlb.h>
21 #include <linux/swap.h>
22 #include <linux/smp.h>
23 #include <linux/init.h>
24 #include <linux/highmem.h>
25 #include <linux/pagemap.h>
26 #include <linux/bootmem.h>
27 #include <linux/slab.h>
28 #include <linux/proc_fs.h>
29 #include <linux/efi.h>
30 #include <linux/memory_hotplug.h>
31 #include <linux/initrd.h>
33 #include <asm/processor.h>
34 #include <asm/system.h>
35 #include <asm/uaccess.h>
36 #include <asm/pgtable.h>
38 #include <asm/fixmap.h>
42 #include <asm/tlbflush.h>
43 #include <asm/sections.h>
45 unsigned int __VMALLOC_RESERVE = 128 << 20;
47 DEFINE_PER_CPU(struct mmu_gather, mmu_gathers);
48 unsigned long highstart_pfn, highend_pfn;
50 static int noinline do_test_wp_bit(void);
53 * Creates a middle page table and puts a pointer to it in the
54 * given global directory entry. This only returns the gd entry
55 * in non-PAE compilation mode, since the middle layer is folded.
57 static pmd_t * __init one_md_table_init(pgd_t *pgd)
63 pmd_table = (pmd_t *) alloc_bootmem_low_pages(PAGE_SIZE);
64 set_pgd(pgd, __pgd(__pa(pmd_table) | _PAGE_PRESENT));
65 pud = pud_offset(pgd, 0);
66 if (pmd_table != pmd_offset(pud, 0))
69 pud = pud_offset(pgd, 0);
70 pmd_table = pmd_offset(pud, 0);
77 * Create a page table and place a pointer to it in a middle page
80 static pte_t * __init one_page_table_init(pmd_t *pmd)
83 pte_t *page_table = (pte_t *) alloc_bootmem_low_pages(PAGE_SIZE);
84 set_pmd(pmd, __pmd(__pa(page_table) | _PAGE_TABLE));
85 if (page_table != pte_offset_kernel(pmd, 0))
91 return pte_offset_kernel(pmd, 0);
95 * This function initializes a certain range of kernel virtual memory
96 * with new bootmem page tables, everywhere page tables are missing in
101 * NOTE: The pagetables are allocated contiguous on the physical space
102 * so we can cache the place of the first one and move around without
103 * checking the pgd every time.
105 static void __init page_table_range_init (unsigned long start, unsigned long end, pgd_t *pgd_base)
110 int pgd_idx, pmd_idx;
114 pgd_idx = pgd_index(vaddr);
115 pmd_idx = pmd_index(vaddr);
116 pgd = pgd_base + pgd_idx;
118 for ( ; (pgd_idx < PTRS_PER_PGD) && (vaddr != end); pgd++, pgd_idx++) {
120 one_md_table_init(pgd);
121 pud = pud_offset(pgd, vaddr);
122 pmd = pmd_offset(pud, vaddr);
123 for (; (pmd_idx < PTRS_PER_PMD) && (vaddr != end); pmd++, pmd_idx++) {
125 one_page_table_init(pmd);
133 static inline int is_kernel_text(unsigned long addr)
135 if (addr >= PAGE_OFFSET && addr <= (unsigned long)__init_end)
141 * This maps the physical memory to kernel virtual address space, a total
142 * of max_low_pfn pages, by creating page tables starting from address
145 static void __init kernel_physical_mapping_init(pgd_t *pgd_base)
151 int pgd_idx, pmd_idx, pte_ofs;
153 pgd_idx = pgd_index(PAGE_OFFSET);
154 pgd = pgd_base + pgd_idx;
157 for (; pgd_idx < PTRS_PER_PGD; pgd++, pgd_idx++) {
158 pmd = one_md_table_init(pgd);
159 if (pfn >= max_low_pfn)
161 for (pmd_idx = 0; pmd_idx < PTRS_PER_PMD && pfn < max_low_pfn; pmd++, pmd_idx++) {
162 unsigned int address = pfn * PAGE_SIZE + PAGE_OFFSET;
164 /* Map with big pages if possible, otherwise create normal page tables. */
166 unsigned int address2 = (pfn + PTRS_PER_PTE - 1) * PAGE_SIZE + PAGE_OFFSET + PAGE_SIZE-1;
168 if (is_kernel_text(address) || is_kernel_text(address2))
169 set_pmd(pmd, pfn_pmd(pfn, PAGE_KERNEL_LARGE_EXEC));
171 set_pmd(pmd, pfn_pmd(pfn, PAGE_KERNEL_LARGE));
174 pte = one_page_table_init(pmd);
176 for (pte_ofs = 0; pte_ofs < PTRS_PER_PTE && pfn < max_low_pfn; pte++, pfn++, pte_ofs++) {
177 if (is_kernel_text(address))
178 set_pte(pte, pfn_pte(pfn, PAGE_KERNEL_EXEC));
180 set_pte(pte, pfn_pte(pfn, PAGE_KERNEL));
187 static inline int page_kills_ppro(unsigned long pagenr)
189 if (pagenr >= 0x70000 && pagenr <= 0x7003F)
194 extern int is_available_memory(efi_memory_desc_t *);
196 int page_is_ram(unsigned long pagenr)
199 unsigned long addr, end;
202 efi_memory_desc_t *md;
205 for (p = memmap.map; p < memmap.map_end; p += memmap.desc_size) {
207 if (!is_available_memory(md))
209 addr = (md->phys_addr+PAGE_SIZE-1) >> PAGE_SHIFT;
210 end = (md->phys_addr + (md->num_pages << EFI_PAGE_SHIFT)) >> PAGE_SHIFT;
212 if ((pagenr >= addr) && (pagenr < end))
218 for (i = 0; i < e820.nr_map; i++) {
220 if (e820.map[i].type != E820_RAM) /* not usable memory */
223 * !!!FIXME!!! Some BIOSen report areas as RAM that
224 * are not. Notably the 640->1Mb area. We need a sanity
227 addr = (e820.map[i].addr+PAGE_SIZE-1) >> PAGE_SHIFT;
228 end = (e820.map[i].addr+e820.map[i].size) >> PAGE_SHIFT;
229 if ((pagenr >= addr) && (pagenr < end))
236 * devmem_is_allowed() checks to see if /dev/mem access to a certain address is
237 * valid. The argument is a physical page number.
240 * On x86, access has to be given to the first megabyte of ram because that area
241 * contains bios code and data regions used by X and dosemu and similar apps.
242 * Access has to be given to non-kernel-ram areas as well, these contain the PCI
243 * mmio resources as well as potential bios/acpi data regions.
245 int devmem_is_allowed(unsigned long pagenr)
249 if (!page_is_ram(pagenr))
254 EXPORT_SYMBOL_GPL(page_is_ram);
256 #ifdef CONFIG_HIGHMEM
260 #define kmap_get_fixmap_pte(vaddr) \
261 pte_offset_kernel(pmd_offset(pud_offset(pgd_offset_k(vaddr), vaddr), (vaddr)), (vaddr))
263 static void __init kmap_init(void)
265 unsigned long kmap_vstart;
267 /* cache the first kmap pte */
268 kmap_vstart = __fix_to_virt(FIX_KMAP_BEGIN);
269 kmap_pte = kmap_get_fixmap_pte(kmap_vstart);
271 kmap_prot = PAGE_KERNEL;
274 static void __init permanent_kmaps_init(pgd_t *pgd_base)
283 page_table_range_init(vaddr, vaddr + PAGE_SIZE*LAST_PKMAP, pgd_base);
285 pgd = swapper_pg_dir + pgd_index(vaddr);
286 pud = pud_offset(pgd, vaddr);
287 pmd = pmd_offset(pud, vaddr);
288 pte = pte_offset_kernel(pmd, vaddr);
289 pkmap_page_table = pte;
292 static void __meminit free_new_highpage(struct page *page)
294 init_page_count(page);
299 void __init add_one_highpage_init(struct page *page, int pfn, int bad_ppro)
301 if (page_is_ram(pfn) && !(bad_ppro && page_kills_ppro(pfn))) {
302 ClearPageReserved(page);
303 free_new_highpage(page);
305 SetPageReserved(page);
308 static int add_one_highpage_hotplug(struct page *page, unsigned long pfn)
310 free_new_highpage(page);
312 #ifdef CONFIG_FLATMEM
313 max_mapnr = max(pfn, max_mapnr);
320 * Not currently handling the NUMA case.
321 * Assuming single node and all memory that
322 * has been added dynamically that would be
323 * onlined here is in HIGHMEM
325 void online_page(struct page *page)
327 ClearPageReserved(page);
328 add_one_highpage_hotplug(page, page_to_pfn(page));
333 extern void set_highmem_pages_init(int);
335 static void __init set_highmem_pages_init(int bad_ppro)
338 for (pfn = highstart_pfn; pfn < highend_pfn; pfn++)
339 add_one_highpage_init(pfn_to_page(pfn), pfn, bad_ppro);
340 totalram_pages += totalhigh_pages;
342 #endif /* CONFIG_FLATMEM */
345 #define kmap_init() do { } while (0)
346 #define permanent_kmaps_init(pgd_base) do { } while (0)
347 #define set_highmem_pages_init(bad_ppro) do { } while (0)
348 #endif /* CONFIG_HIGHMEM */
350 unsigned long long __PAGE_KERNEL = _PAGE_KERNEL;
351 EXPORT_SYMBOL(__PAGE_KERNEL);
352 unsigned long long __PAGE_KERNEL_EXEC = _PAGE_KERNEL_EXEC;
355 extern void __init remap_numa_kva(void);
357 #define remap_numa_kva() do {} while (0)
360 static void __init pagetable_init (void)
363 pgd_t *pgd_base = swapper_pg_dir;
365 #ifdef CONFIG_X86_PAE
367 /* Init entries of the first-level page table to the zero page */
368 for (i = 0; i < PTRS_PER_PGD; i++)
369 set_pgd(pgd_base + i, __pgd(__pa(empty_zero_page) | _PAGE_PRESENT));
372 /* Enable PSE if available */
374 set_in_cr4(X86_CR4_PSE);
377 /* Enable PGE if available */
379 set_in_cr4(X86_CR4_PGE);
380 __PAGE_KERNEL |= _PAGE_GLOBAL;
381 __PAGE_KERNEL_EXEC |= _PAGE_GLOBAL;
384 kernel_physical_mapping_init(pgd_base);
388 * Fixed mappings, only the page table structure has to be
389 * created - mappings will be set by set_fixmap():
391 vaddr = __fix_to_virt(__end_of_fixed_addresses - 1) & PMD_MASK;
392 page_table_range_init(vaddr, 0, pgd_base);
394 permanent_kmaps_init(pgd_base);
396 #ifdef CONFIG_X86_PAE
398 * Add low memory identity-mappings - SMP needs it when
399 * starting up on an AP from real-mode. In the non-PAE
400 * case we already have these mappings through head.S.
401 * All user-space mappings are explicitly cleared after
404 set_pgd(&pgd_base[0], pgd_base[USER_PTRS_PER_PGD]);
408 #ifdef CONFIG_SOFTWARE_SUSPEND
410 * Swap suspend & friends need this for resume because things like the intel-agp
411 * driver might have split up a kernel 4MB mapping.
413 char __nosavedata swsusp_pg_dir[PAGE_SIZE]
414 __attribute__ ((aligned (PAGE_SIZE)));
416 static inline void save_pg_dir(void)
418 memcpy(swsusp_pg_dir, swapper_pg_dir, PAGE_SIZE);
421 static inline void save_pg_dir(void)
426 void zap_low_mappings (void)
433 * Zap initial low-memory mappings.
435 * Note that "pgd_clear()" doesn't do it for
436 * us, because pgd_clear() is a no-op on i386.
438 for (i = 0; i < USER_PTRS_PER_PGD; i++)
439 #ifdef CONFIG_X86_PAE
440 set_pgd(swapper_pg_dir+i, __pgd(1 + __pa(empty_zero_page)));
442 set_pgd(swapper_pg_dir+i, __pgd(0));
447 static int disable_nx __initdata = 0;
448 u64 __supported_pte_mask __read_mostly = ~_PAGE_NX;
453 * Control non executable mappings.
456 * off Disable (disables exec-shield too)
458 void __init noexec_setup(const char *str)
460 if (!strncmp(str, "on",2) && cpu_has_nx) {
461 __supported_pte_mask |= _PAGE_NX;
463 } else if (!strncmp(str,"off",3)) {
465 __supported_pte_mask &= ~_PAGE_NX;
471 #ifdef CONFIG_X86_PAE
473 static void __init set_nx(void)
475 unsigned int v[4], l, h;
477 if (cpu_has_pae && (cpuid_eax(0x80000000) > 0x80000001)) {
478 cpuid(0x80000001, &v[0], &v[1], &v[2], &v[3]);
479 if ((v[3] & (1 << 20)) && !disable_nx) {
480 rdmsr(MSR_EFER, l, h);
482 wrmsr(MSR_EFER, l, h);
484 __supported_pte_mask |= _PAGE_NX;
490 * Enables/disables executability of a given kernel page and
491 * returns the previous setting.
493 int __init set_kernel_exec(unsigned long vaddr, int enable)
501 pte = lookup_address(vaddr);
504 if (!pte_exec_kernel(*pte))
508 pte->pte_high &= ~(1 << (_PAGE_BIT_NX - 32));
510 pte->pte_high |= 1 << (_PAGE_BIT_NX - 32);
519 * paging_init() sets up the page tables - note that the first 8MB are
520 * already mapped by head.S.
522 * This routines also unmaps the page at virtual kernel address 0, so
523 * that we can trap those pesky NULL-reference errors in the kernel.
525 void __init paging_init(void)
527 #ifdef CONFIG_X86_PAE
530 printk("NX (Execute Disable) protection: active\n");
534 printk("Using x86 segment limits to approximate NX protection\n");
538 load_cr3(swapper_pg_dir);
540 #ifdef CONFIG_X86_PAE
542 * We will bail out later - printk doesn't work right now so
543 * the user would just see a hanging kernel.
546 set_in_cr4(X86_CR4_PAE);
554 * Test if the WP bit works in supervisor mode. It isn't supported on 386's
555 * and also on some strange 486's (NexGen etc.). All 586+'s are OK. This
556 * used to involve black magic jumps to work around some nasty CPU bugs,
557 * but fortunately the switch to using exceptions got rid of all that.
560 static void __init test_wp_bit(void)
562 printk("Checking if this processor honours the WP bit even in supervisor mode... ");
564 /* Any page-aligned address will do, the test is non-destructive */
565 __set_fixmap(FIX_WP_TEST, __pa(&swapper_pg_dir), PAGE_READONLY);
566 boot_cpu_data.wp_works_ok = do_test_wp_bit();
567 clear_fixmap(FIX_WP_TEST);
569 if (!boot_cpu_data.wp_works_ok) {
571 #ifdef CONFIG_X86_WP_WORKS_OK
572 panic("This kernel doesn't support CPU's with broken WP. Recompile it for a 386!");
579 static void __init set_max_mapnr_init(void)
581 #ifdef CONFIG_HIGHMEM
582 num_physpages = highend_pfn;
584 num_physpages = max_low_pfn;
586 #ifdef CONFIG_FLATMEM
587 max_mapnr = num_physpages;
591 static struct kcore_list kcore_mem, kcore_vmalloc;
593 void __init mem_init(void)
595 extern int ppro_with_ram_bug(void);
596 int codesize, reservedpages, datasize, initsize;
600 #ifdef CONFIG_FLATMEM
605 bad_ppro = ppro_with_ram_bug();
607 #ifdef CONFIG_HIGHMEM
608 /* check that fixmap and pkmap do not overlap */
609 if (PKMAP_BASE+LAST_PKMAP*PAGE_SIZE >= FIXADDR_START) {
610 printk(KERN_ERR "fixmap and kmap areas overlap - this will crash\n");
611 printk(KERN_ERR "pkstart: %lxh pkend: %lxh fixstart %lxh\n",
612 PKMAP_BASE, PKMAP_BASE+LAST_PKMAP*PAGE_SIZE, FIXADDR_START);
617 set_max_mapnr_init();
619 #ifdef CONFIG_HIGHMEM
620 high_memory = (void *) __va(highstart_pfn * PAGE_SIZE - 1) + 1;
622 high_memory = (void *) __va(max_low_pfn * PAGE_SIZE - 1) + 1;
625 /* this will put all low memory onto the freelists */
626 totalram_pages += free_all_bootmem();
629 for (tmp = 0; tmp < max_low_pfn; tmp++)
631 * Only count reserved RAM pages
633 if (page_is_ram(tmp) && PageReserved(pfn_to_page(tmp)))
636 set_highmem_pages_init(bad_ppro);
638 codesize = (unsigned long) &_etext - (unsigned long) &_text;
639 datasize = (unsigned long) &_edata - (unsigned long) &_etext;
640 initsize = (unsigned long) &__init_end - (unsigned long) &__init_begin;
642 kclist_add(&kcore_mem, __va(0), max_low_pfn << PAGE_SHIFT);
643 kclist_add(&kcore_vmalloc, (void *)VMALLOC_START,
644 VMALLOC_END-VMALLOC_START);
646 printk(KERN_INFO "Memory: %luk/%luk available (%dk kernel code, %dk reserved, %dk data, %dk init, %ldk highmem)\n",
647 (unsigned long) nr_free_pages() << (PAGE_SHIFT-10),
648 num_physpages << (PAGE_SHIFT-10),
650 reservedpages << (PAGE_SHIFT-10),
653 (unsigned long) (totalhigh_pages << (PAGE_SHIFT-10))
656 #ifdef CONFIG_X86_PAE
658 panic("cannot execute a PAE-enabled kernel on a PAE-less CPU!");
660 if (boot_cpu_data.wp_works_ok < 0)
664 * Subtle. SMP is doing it's boot stuff late (because it has to
665 * fork idle threads) - but it also needs low mappings for the
666 * protected-mode entry to work. We zap these entries only after
667 * the WP-bit has been tested.
675 * this is for the non-NUMA, single node SMP system case.
676 * Specifically, in the case of x86, we will always add
677 * memory to the highmem for now.
679 #ifdef CONFIG_MEMORY_HOTPLUG
680 #ifndef CONFIG_NEED_MULTIPLE_NODES
681 int add_memory(u64 start, u64 size)
683 struct pglist_data *pgdata = &contig_page_data;
684 struct zone *zone = pgdata->node_zones + MAX_NR_ZONES-1;
685 unsigned long start_pfn = start >> PAGE_SHIFT;
686 unsigned long nr_pages = size >> PAGE_SHIFT;
688 return __add_pages(zone, start_pfn, nr_pages);
691 int remove_memory(u64 start, u64 size)
698 kmem_cache_t *pgd_cache;
699 kmem_cache_t *pmd_cache;
701 void __init pgtable_cache_init(void)
703 if (PTRS_PER_PMD > 1) {
704 pmd_cache = kmem_cache_create("pmd",
705 PTRS_PER_PMD*sizeof(pmd_t),
706 PTRS_PER_PMD*sizeof(pmd_t),
711 panic("pgtable_cache_init(): cannot create pmd cache");
713 pgd_cache = kmem_cache_create("pgd",
714 PTRS_PER_PGD*sizeof(pgd_t),
715 PTRS_PER_PGD*sizeof(pgd_t),
718 PTRS_PER_PMD == 1 ? pgd_dtor : NULL);
720 panic("pgtable_cache_init(): Cannot create pgd cache");
724 * This function cannot be __init, since exceptions don't work in that
725 * section. Put this after the callers, so that it cannot be inlined.
727 static int noinline do_test_wp_bit(void)
732 __asm__ __volatile__(
737 ".section __ex_table,\"a\"\n"
741 :"=m" (*(char *)fix_to_virt(FIX_WP_TEST)),
750 #ifdef CONFIG_DEBUG_RODATA
752 extern char __start_rodata, __end_rodata;
753 void mark_rodata_ro(void)
755 unsigned long addr = (unsigned long)&__start_rodata;
757 for (; addr < (unsigned long)&__end_rodata; addr += PAGE_SIZE)
758 change_page_attr(virt_to_page(addr), 1, PAGE_KERNEL_RO);
760 printk ("Write protecting the kernel read-only data: %luk\n",
761 (unsigned long)(&__end_rodata - &__start_rodata) >> 10);
764 * change_page_attr() requires a global_flush_tlb() call after it.
765 * We do this after the printk so that if something went wrong in the
766 * change, the printk gets out at least to give a better debug hint
767 * of who is the culprit.
773 void free_init_pages(char *what, unsigned long begin, unsigned long end)
777 for (addr = begin; addr < end; addr += PAGE_SIZE) {
778 ClearPageReserved(virt_to_page(addr));
779 init_page_count(virt_to_page(addr));
780 memset((void *)addr, 0xcc, PAGE_SIZE);
784 printk(KERN_INFO "Freeing %s: %ldk freed\n", what, (end - begin) >> 10);
787 void free_initmem(void)
789 free_init_pages("unused kernel memory",
790 (unsigned long)(&__init_begin),
791 (unsigned long)(&__init_end));
794 #ifdef CONFIG_BLK_DEV_INITRD
795 void free_initrd_mem(unsigned long start, unsigned long end)
797 free_init_pages("initrd memory", start, end);