2 * linux/arch/x86_64/mm/init.c
4 * Copyright (C) 1995 Linus Torvalds
5 * Copyright (C) 2000 Pavel Machek <pavel@suse.cz>
6 * Copyright (C) 2002,2003 Andi Kleen <ak@suse.de>
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/swap.h>
19 #include <linux/smp.h>
20 #include <linux/init.h>
21 #include <linux/pagemap.h>
22 #include <linux/bootmem.h>
23 #include <linux/proc_fs.h>
24 #include <linux/pci.h>
25 #include <linux/poison.h>
26 #include <linux/dma-mapping.h>
27 #include <linux/module.h>
28 #include <linux/memory_hotplug.h>
30 #include <asm/processor.h>
31 #include <asm/system.h>
32 #include <asm/uaccess.h>
33 #include <asm/pgtable.h>
34 #include <asm/pgalloc.h>
36 #include <asm/fixmap.h>
40 #include <asm/mmu_context.h>
41 #include <asm/proto.h>
43 #include <asm/sections.h>
49 struct dma_mapping_ops* dma_ops;
50 EXPORT_SYMBOL(dma_ops);
52 static unsigned long dma_reserve __initdata;
54 DEFINE_PER_CPU(struct mmu_gather, mmu_gathers);
57 * NOTE: pagetable_init alloc all the fixmap pagetables contiguous on the
58 * physical space so we can cache the place of the first one and move
59 * around without checking the pgd every time.
64 long i, total = 0, reserved = 0;
65 long shared = 0, cached = 0;
69 printk(KERN_INFO "Mem-info:\n");
71 printk(KERN_INFO "Free swap: %6ldkB\n", nr_swap_pages<<(PAGE_SHIFT-10));
73 for_each_online_pgdat(pgdat) {
74 for (i = 0; i < pgdat->node_spanned_pages; ++i) {
75 page = pfn_to_page(pgdat->node_start_pfn + i);
77 if (PageReserved(page))
79 else if (PageSwapCache(page))
81 else if (page_count(page))
82 shared += page_count(page) - 1;
85 printk(KERN_INFO "%lu pages of RAM\n", total);
86 printk(KERN_INFO "%lu reserved pages\n",reserved);
87 printk(KERN_INFO "%lu pages shared\n",shared);
88 printk(KERN_INFO "%lu pages swap cached\n",cached);
93 static __init void *spp_getpage(void)
97 ptr = (void *) get_zeroed_page(GFP_ATOMIC);
99 ptr = alloc_bootmem_pages(PAGE_SIZE);
100 if (!ptr || ((unsigned long)ptr & ~PAGE_MASK))
101 panic("set_pte_phys: cannot allocate page data %s\n", after_bootmem?"after bootmem":"");
103 Dprintk("spp_getpage %p\n", ptr);
107 static __init void set_pte_phys(unsigned long vaddr,
108 unsigned long phys, pgprot_t prot)
115 Dprintk("set_pte_phys %lx to %lx\n", vaddr, phys);
117 pgd = pgd_offset_k(vaddr);
118 if (pgd_none(*pgd)) {
119 printk("PGD FIXMAP MISSING, it should be setup in head.S!\n");
122 pud = pud_offset(pgd, vaddr);
123 if (pud_none(*pud)) {
124 pmd = (pmd_t *) spp_getpage();
125 set_pud(pud, __pud(__pa(pmd) | _KERNPG_TABLE | _PAGE_USER));
126 if (pmd != pmd_offset(pud, 0)) {
127 printk("PAGETABLE BUG #01! %p <-> %p\n", pmd, pmd_offset(pud,0));
131 pmd = pmd_offset(pud, vaddr);
132 if (pmd_none(*pmd)) {
133 pte = (pte_t *) spp_getpage();
134 set_pmd(pmd, __pmd(__pa(pte) | _KERNPG_TABLE | _PAGE_USER));
135 if (pte != pte_offset_kernel(pmd, 0)) {
136 printk("PAGETABLE BUG #02!\n");
140 new_pte = pfn_pte(phys >> PAGE_SHIFT, prot);
142 pte = pte_offset_kernel(pmd, vaddr);
143 if (!pte_none(*pte) &&
144 pte_val(*pte) != (pte_val(new_pte) & __supported_pte_mask))
146 set_pte(pte, new_pte);
149 * It's enough to flush this one mapping.
150 * (PGE mappings get flushed as well)
152 __flush_tlb_one(vaddr);
155 /* NOTE: this is meant to be run only at boot */
157 __set_fixmap (enum fixed_addresses idx, unsigned long phys, pgprot_t prot)
159 unsigned long address = __fix_to_virt(idx);
161 if (idx >= __end_of_fixed_addresses) {
162 printk("Invalid __set_fixmap\n");
165 set_pte_phys(address, phys, prot);
168 unsigned long __initdata table_start, table_end;
170 extern pmd_t temp_boot_pmds[];
172 static struct temp_map {
176 } temp_mappings[] __initdata = {
177 { &temp_boot_pmds[0], (void *)(40UL * 1024 * 1024) },
178 { &temp_boot_pmds[1], (void *)(42UL * 1024 * 1024) },
182 static __meminit void *alloc_low_page(int *index, unsigned long *phys)
186 unsigned long pfn = table_end++, paddr;
190 adr = (void *)get_zeroed_page(GFP_ATOMIC);
196 panic("alloc_low_page: ran out of memory");
197 for (i = 0; temp_mappings[i].allocated; i++) {
198 if (!temp_mappings[i].pmd)
199 panic("alloc_low_page: ran out of temp mappings");
201 ti = &temp_mappings[i];
202 paddr = (pfn << PAGE_SHIFT) & PMD_MASK;
203 set_pmd(ti->pmd, __pmd(paddr | _KERNPG_TABLE | _PAGE_PSE));
206 adr = ti->address + ((pfn << PAGE_SHIFT) & ~PMD_MASK);
207 memset(adr, 0, PAGE_SIZE);
209 *phys = pfn * PAGE_SIZE;
213 static __meminit void unmap_low_page(int i)
220 ti = &temp_mappings[i];
221 set_pmd(ti->pmd, __pmd(0));
225 /* Must run before zap_low_mappings */
226 __init void *early_ioremap(unsigned long addr, unsigned long size)
228 unsigned long map = round_down(addr, LARGE_PAGE_SIZE);
230 /* actually usually some more */
231 if (size >= LARGE_PAGE_SIZE) {
234 set_pmd(temp_mappings[0].pmd, __pmd(map | _KERNPG_TABLE | _PAGE_PSE));
235 map += LARGE_PAGE_SIZE;
236 set_pmd(temp_mappings[1].pmd, __pmd(map | _KERNPG_TABLE | _PAGE_PSE));
238 return temp_mappings[0].address + (addr & (LARGE_PAGE_SIZE-1));
241 /* To avoid virtual aliases later */
242 __init void early_iounmap(void *addr, unsigned long size)
244 if ((void *)round_down((unsigned long)addr, LARGE_PAGE_SIZE) != temp_mappings[0].address)
245 printk("early_iounmap: bad address %p\n", addr);
246 set_pmd(temp_mappings[0].pmd, __pmd(0));
247 set_pmd(temp_mappings[1].pmd, __pmd(0));
251 static void __meminit
252 phys_pmd_init(pmd_t *pmd_page, unsigned long address, unsigned long end)
254 int i = pmd_index(address);
256 for (; i < PTRS_PER_PMD; i++, address += PMD_SIZE) {
258 pmd_t *pmd = pmd_page + pmd_index(address);
260 if (address >= end) {
262 for (; i < PTRS_PER_PMD; i++, pmd++)
263 set_pmd(pmd, __pmd(0));
270 entry = _PAGE_NX|_PAGE_PSE|_KERNPG_TABLE|_PAGE_GLOBAL|address;
271 entry &= __supported_pte_mask;
272 set_pmd(pmd, __pmd(entry));
276 static void __meminit
277 phys_pmd_update(pud_t *pud, unsigned long address, unsigned long end)
279 pmd_t *pmd = pmd_offset(pud,0);
280 spin_lock(&init_mm.page_table_lock);
281 phys_pmd_init(pmd, address, end);
282 spin_unlock(&init_mm.page_table_lock);
286 static void __meminit phys_pud_init(pud_t *pud_page, unsigned long addr, unsigned long end)
288 int i = pud_index(addr);
291 for (; i < PTRS_PER_PUD; i++, addr = (addr & PUD_MASK) + PUD_SIZE ) {
293 unsigned long pmd_phys;
294 pud_t *pud = pud_page + pud_index(addr);
300 if (!after_bootmem && !e820_any_mapped(addr,addr+PUD_SIZE,0)) {
301 set_pud(pud, __pud(0));
306 phys_pmd_update(pud, addr, end);
310 pmd = alloc_low_page(&map, &pmd_phys);
311 spin_lock(&init_mm.page_table_lock);
312 set_pud(pud, __pud(pmd_phys | _KERNPG_TABLE));
313 phys_pmd_init(pmd, addr, end);
314 spin_unlock(&init_mm.page_table_lock);
320 static void __init find_early_table_space(unsigned long end)
322 unsigned long puds, pmds, tables, start;
324 puds = (end + PUD_SIZE - 1) >> PUD_SHIFT;
325 pmds = (end + PMD_SIZE - 1) >> PMD_SHIFT;
326 tables = round_up(puds * sizeof(pud_t), PAGE_SIZE) +
327 round_up(pmds * sizeof(pmd_t), PAGE_SIZE);
329 /* RED-PEN putting page tables only on node 0 could
330 cause a hotspot and fill up ZONE_DMA. The page tables
331 need roughly 0.5KB per GB. */
333 table_start = find_e820_area(start, end, tables);
334 if (table_start == -1UL)
335 panic("Cannot find space for the kernel page tables");
337 table_start >>= PAGE_SHIFT;
338 table_end = table_start;
340 early_printk("kernel direct mapping tables up to %lx @ %lx-%lx\n",
341 end, table_start << PAGE_SHIFT,
342 (table_start << PAGE_SHIFT) + tables);
345 /* Setup the direct mapping of the physical memory at PAGE_OFFSET.
346 This runs before bootmem is initialized and gets pages directly from the
347 physical memory. To access them they are temporarily mapped. */
348 void __meminit init_memory_mapping(unsigned long start, unsigned long end)
352 Dprintk("init_memory_mapping\n");
355 * Find space for the kernel direct mapping tables.
356 * Later we should allocate these tables in the local node of the memory
357 * mapped. Unfortunately this is done currently before the nodes are
361 find_early_table_space(end);
363 start = (unsigned long)__va(start);
364 end = (unsigned long)__va(end);
366 for (; start < end; start = next) {
368 unsigned long pud_phys;
369 pgd_t *pgd = pgd_offset_k(start);
373 pud = pud_offset(pgd, start & PGDIR_MASK);
375 pud = alloc_low_page(&map, &pud_phys);
377 next = start + PGDIR_SIZE;
380 phys_pud_init(pud, __pa(start), __pa(next));
382 set_pgd(pgd_offset_k(start), mk_kernel_pgd(pud_phys));
387 asm volatile("movq %%cr4,%0" : "=r" (mmu_cr4_features));
391 void __cpuinit zap_low_mappings(int cpu)
394 pgd_t *pgd = pgd_offset_k(0UL);
398 * For AP's, zap the low identity mappings by changing the cr3
399 * to init_level4_pgt and doing local flush tlb all
401 asm volatile("movq %0,%%cr3" :: "r" (__pa_symbol(&init_level4_pgt)));
407 void __init paging_init(void)
409 unsigned long max_zone_pfns[MAX_NR_ZONES];
410 memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
411 max_zone_pfns[ZONE_DMA] = MAX_DMA_PFN;
412 max_zone_pfns[ZONE_DMA32] = MAX_DMA32_PFN;
413 max_zone_pfns[ZONE_NORMAL] = end_pfn;
415 memory_present(0, 0, end_pfn);
417 free_area_init_nodes(max_zone_pfns);
421 /* Unmap a kernel mapping if it exists. This is useful to avoid prefetches
422 from the CPU leading to inconsistent cache lines. address and size
423 must be aligned to 2MB boundaries.
424 Does nothing when the mapping doesn't exist. */
425 void __init clear_kernel_mapping(unsigned long address, unsigned long size)
427 unsigned long end = address + size;
429 BUG_ON(address & ~LARGE_PAGE_MASK);
430 BUG_ON(size & ~LARGE_PAGE_MASK);
432 for (; address < end; address += LARGE_PAGE_SIZE) {
433 pgd_t *pgd = pgd_offset_k(address);
438 pud = pud_offset(pgd, address);
441 pmd = pmd_offset(pud, address);
442 if (!pmd || pmd_none(*pmd))
444 if (0 == (pmd_val(*pmd) & _PAGE_PSE)) {
445 /* Could handle this, but it should not happen currently. */
447 "clear_kernel_mapping: mapping has been split. will leak memory\n");
450 set_pmd(pmd, __pmd(0));
456 * Memory hotplug specific functions
458 void online_page(struct page *page)
460 ClearPageReserved(page);
461 init_page_count(page);
467 #ifdef CONFIG_MEMORY_HOTPLUG
469 * Memory is added always to NORMAL zone. This means you will never get
470 * additional DMA/DMA32 memory.
472 int arch_add_memory(int nid, u64 start, u64 size)
474 struct pglist_data *pgdat = NODE_DATA(nid);
475 struct zone *zone = pgdat->node_zones + ZONE_NORMAL;
476 unsigned long start_pfn = start >> PAGE_SHIFT;
477 unsigned long nr_pages = size >> PAGE_SHIFT;
480 init_memory_mapping(start, (start + size -1));
482 ret = __add_pages(zone, start_pfn, nr_pages);
488 printk("%s: Problem encountered in __add_pages!\n", __func__);
491 EXPORT_SYMBOL_GPL(arch_add_memory);
493 int remove_memory(u64 start, u64 size)
497 EXPORT_SYMBOL_GPL(remove_memory);
499 #ifndef CONFIG_ACPI_NUMA
500 int memory_add_physaddr_to_nid(u64 start)
504 EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid);
507 #ifndef CONFIG_ACPI_NUMA
508 int memory_add_physaddr_to_nid(u64 start)
514 #endif /* CONFIG_MEMORY_HOTPLUG */
516 #ifdef CONFIG_MEMORY_HOTPLUG_RESERVE
518 * Memory Hotadd without sparsemem. The mem_maps have been allocated in advance,
519 * just online the pages.
521 int __add_pages(struct zone *z, unsigned long start_pfn, unsigned long nr_pages)
525 unsigned long total = 0, mem = 0;
526 for (pfn = start_pfn; pfn < start_pfn + nr_pages; pfn++) {
527 if (pfn_valid(pfn)) {
528 online_page(pfn_to_page(pfn));
535 z->spanned_pages += total;
536 z->present_pages += mem;
537 z->zone_pgdat->node_spanned_pages += total;
538 z->zone_pgdat->node_present_pages += mem;
544 static struct kcore_list kcore_mem, kcore_vmalloc, kcore_kernel, kcore_modules,
547 void __init mem_init(void)
549 long codesize, reservedpages, datasize, initsize;
553 /* clear the zero-page */
554 memset(empty_zero_page, 0, PAGE_SIZE);
558 /* this will put all low memory onto the freelists */
560 totalram_pages = numa_free_all_bootmem();
562 totalram_pages = free_all_bootmem();
564 reservedpages = end_pfn - totalram_pages -
565 absent_pages_in_range(0, end_pfn);
569 codesize = (unsigned long) &_etext - (unsigned long) &_text;
570 datasize = (unsigned long) &_edata - (unsigned long) &_etext;
571 initsize = (unsigned long) &__init_end - (unsigned long) &__init_begin;
573 /* Register memory areas for /proc/kcore */
574 kclist_add(&kcore_mem, __va(0), max_low_pfn << PAGE_SHIFT);
575 kclist_add(&kcore_vmalloc, (void *)VMALLOC_START,
576 VMALLOC_END-VMALLOC_START);
577 kclist_add(&kcore_kernel, &_stext, _end - _stext);
578 kclist_add(&kcore_modules, (void *)MODULES_VADDR, MODULES_LEN);
579 kclist_add(&kcore_vsyscall, (void *)VSYSCALL_START,
580 VSYSCALL_END - VSYSCALL_START);
582 printk("Memory: %luk/%luk available (%ldk kernel code, %ldk reserved, %ldk data, %ldk init)\n",
583 (unsigned long) nr_free_pages() << (PAGE_SHIFT-10),
584 end_pfn << (PAGE_SHIFT-10),
586 reservedpages << (PAGE_SHIFT-10),
592 * Sync boot_level4_pgt mappings with the init_level4_pgt
593 * except for the low identity mappings which are already zapped
594 * in init_level4_pgt. This sync-up is essential for AP's bringup
596 memcpy(boot_level4_pgt+1, init_level4_pgt+1, (PTRS_PER_PGD-1)*sizeof(pgd_t));
600 void free_init_pages(char *what, unsigned long begin, unsigned long end)
607 printk(KERN_INFO "Freeing %s: %ldk freed\n", what, (end - begin) >> 10);
608 for (addr = begin; addr < end; addr += PAGE_SIZE) {
609 ClearPageReserved(virt_to_page(addr));
610 init_page_count(virt_to_page(addr));
611 memset((void *)(addr & ~(PAGE_SIZE-1)),
612 POISON_FREE_INITMEM, PAGE_SIZE);
618 void free_initmem(void)
620 memset(__initdata_begin, POISON_FREE_INITDATA,
621 __initdata_end - __initdata_begin);
622 free_init_pages("unused kernel memory",
623 (unsigned long)(&__init_begin),
624 (unsigned long)(&__init_end));
627 #ifdef CONFIG_DEBUG_RODATA
629 void mark_rodata_ro(void)
631 unsigned long addr = (unsigned long)__start_rodata;
633 for (; addr < (unsigned long)__end_rodata; addr += PAGE_SIZE)
634 change_page_attr_addr(addr, 1, PAGE_KERNEL_RO);
636 printk ("Write protecting the kernel read-only data: %luk\n",
637 (__end_rodata - __start_rodata) >> 10);
640 * change_page_attr_addr() requires a global_flush_tlb() call after it.
641 * We do this after the printk so that if something went wrong in the
642 * change, the printk gets out at least to give a better debug hint
643 * of who is the culprit.
649 #ifdef CONFIG_BLK_DEV_INITRD
650 void free_initrd_mem(unsigned long start, unsigned long end)
652 free_init_pages("initrd memory", start, end);
656 void __init reserve_bootmem_generic(unsigned long phys, unsigned len)
659 int nid = phys_to_nid(phys);
661 unsigned long pfn = phys >> PAGE_SHIFT;
662 if (pfn >= end_pfn) {
663 /* This can happen with kdump kernels when accessing firmware
665 if (pfn < end_pfn_map)
667 printk(KERN_ERR "reserve_bootmem: illegal reserve %lx %u\n",
672 /* Should check here against the e820 map to avoid double free */
674 reserve_bootmem_node(NODE_DATA(nid), phys, len);
676 reserve_bootmem(phys, len);
678 if (phys+len <= MAX_DMA_PFN*PAGE_SIZE) {
679 dma_reserve += len / PAGE_SIZE;
680 set_dma_reserve(dma_reserve);
684 int kern_addr_valid(unsigned long addr)
686 unsigned long above = ((long)addr) >> __VIRTUAL_MASK_SHIFT;
692 if (above != 0 && above != -1UL)
695 pgd = pgd_offset_k(addr);
699 pud = pud_offset(pgd, addr);
703 pmd = pmd_offset(pud, addr);
707 return pfn_valid(pmd_pfn(*pmd));
709 pte = pte_offset_kernel(pmd, addr);
712 return pfn_valid(pte_pfn(*pte));
716 #include <linux/sysctl.h>
718 extern int exception_trace, page_fault_trace;
720 static ctl_table debug_table2[] = {
721 { 99, "exception-trace", &exception_trace, sizeof(int), 0644, NULL,
726 static ctl_table debug_root_table2[] = {
727 { .ctl_name = CTL_DEBUG, .procname = "debug", .mode = 0555,
728 .child = debug_table2 },
732 static __init int x8664_sysctl_init(void)
734 register_sysctl_table(debug_root_table2, 1);
737 __initcall(x8664_sysctl_init);
740 /* A pseudo VMAs to allow ptrace access for the vsyscall page. This only
741 covers the 64bit vsyscall page now. 32bit has a real VMA now and does
742 not need special handling anymore. */
744 static struct vm_area_struct gate_vma = {
745 .vm_start = VSYSCALL_START,
746 .vm_end = VSYSCALL_END,
747 .vm_page_prot = PAGE_READONLY
750 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
752 #ifdef CONFIG_IA32_EMULATION
753 if (test_tsk_thread_flag(tsk, TIF_IA32))
759 int in_gate_area(struct task_struct *task, unsigned long addr)
761 struct vm_area_struct *vma = get_gate_vma(task);
764 return (addr >= vma->vm_start) && (addr < vma->vm_end);
767 /* Use this when you have no reliable task/vma, typically from interrupt
768 * context. It is less reliable than using the task's vma and may give
771 int in_gate_area_no_task(unsigned long addr)
773 return (addr >= VSYSCALL_START) && (addr < VSYSCALL_END);