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/pfn.h>
26 #include <linux/poison.h>
27 #include <linux/dma-mapping.h>
28 #include <linux/module.h>
29 #include <linux/memory_hotplug.h>
30 #include <linux/nmi.h>
32 #include <asm/processor.h>
33 #include <asm/system.h>
34 #include <asm/uaccess.h>
35 #include <asm/pgtable.h>
36 #include <asm/pgalloc.h>
38 #include <asm/fixmap.h>
42 #include <asm/mmu_context.h>
43 #include <asm/proto.h>
45 #include <asm/sections.h>
46 #include <asm/kdebug.h>
48 #include <asm/cacheflush.h>
50 const struct dma_mapping_ops *dma_ops;
51 EXPORT_SYMBOL(dma_ops);
53 static unsigned long dma_reserve __initdata;
55 DEFINE_PER_CPU(struct mmu_gather, mmu_gathers);
57 int direct_gbpages __meminitdata
58 #ifdef CONFIG_DIRECT_GBPAGES
63 static int __init parse_direct_gbpages_off(char *arg)
68 early_param("nogbpages", parse_direct_gbpages_off);
70 static int __init parse_direct_gbpages_on(char *arg)
75 early_param("gbpages", parse_direct_gbpages_on);
78 * NOTE: pagetable_init alloc all the fixmap pagetables contiguous on the
79 * physical space so we can cache the place of the first one and move
80 * around without checking the pgd every time.
85 long i, total = 0, reserved = 0;
86 long shared = 0, cached = 0;
90 printk(KERN_INFO "Mem-info:\n");
92 printk(KERN_INFO "Free swap: %6ldkB\n",
93 nr_swap_pages << (PAGE_SHIFT-10));
95 for_each_online_pgdat(pgdat) {
96 for (i = 0; i < pgdat->node_spanned_pages; ++i) {
98 * This loop can take a while with 256 GB and
99 * 4k pages so defer the NMI watchdog:
101 if (unlikely(i % MAX_ORDER_NR_PAGES == 0))
102 touch_nmi_watchdog();
104 if (!pfn_valid(pgdat->node_start_pfn + i))
107 page = pfn_to_page(pgdat->node_start_pfn + i);
109 if (PageReserved(page))
111 else if (PageSwapCache(page))
113 else if (page_count(page))
114 shared += page_count(page) - 1;
117 printk(KERN_INFO "%lu pages of RAM\n", total);
118 printk(KERN_INFO "%lu reserved pages\n", reserved);
119 printk(KERN_INFO "%lu pages shared\n", shared);
120 printk(KERN_INFO "%lu pages swap cached\n", cached);
125 static __init void *spp_getpage(void)
130 ptr = (void *) get_zeroed_page(GFP_ATOMIC);
132 ptr = alloc_bootmem_pages(PAGE_SIZE);
134 if (!ptr || ((unsigned long)ptr & ~PAGE_MASK)) {
135 panic("set_pte_phys: cannot allocate page data %s\n",
136 after_bootmem ? "after bootmem" : "");
139 pr_debug("spp_getpage %p\n", ptr);
145 set_pte_phys(unsigned long vaddr, unsigned long phys, pgprot_t prot)
152 pr_debug("set_pte_phys %lx to %lx\n", vaddr, phys);
154 pgd = pgd_offset_k(vaddr);
155 if (pgd_none(*pgd)) {
157 "PGD FIXMAP MISSING, it should be setup in head.S!\n");
160 pud = pud_offset(pgd, vaddr);
161 if (pud_none(*pud)) {
162 pmd = (pmd_t *) spp_getpage();
163 set_pud(pud, __pud(__pa(pmd) | _KERNPG_TABLE | _PAGE_USER));
164 if (pmd != pmd_offset(pud, 0)) {
165 printk(KERN_ERR "PAGETABLE BUG #01! %p <-> %p\n",
166 pmd, pmd_offset(pud, 0));
170 pmd = pmd_offset(pud, vaddr);
171 if (pmd_none(*pmd)) {
172 pte = (pte_t *) spp_getpage();
173 set_pmd(pmd, __pmd(__pa(pte) | _KERNPG_TABLE | _PAGE_USER));
174 if (pte != pte_offset_kernel(pmd, 0)) {
175 printk(KERN_ERR "PAGETABLE BUG #02!\n");
179 new_pte = pfn_pte(phys >> PAGE_SHIFT, prot);
181 pte = pte_offset_kernel(pmd, vaddr);
182 if (!pte_none(*pte) &&
183 pte_val(*pte) != (pte_val(new_pte) & __supported_pte_mask))
185 set_pte(pte, new_pte);
188 * It's enough to flush this one mapping.
189 * (PGE mappings get flushed as well)
191 __flush_tlb_one(vaddr);
195 * The head.S code sets up the kernel high mapping:
197 * from __START_KERNEL_map to __START_KERNEL_map + size (== _end-_text)
199 * phys_addr holds the negative offset to the kernel, which is added
200 * to the compile time generated pmds. This results in invalid pmds up
201 * to the point where we hit the physaddr 0 mapping.
203 * We limit the mappings to the region from _text to _end. _end is
204 * rounded up to the 2MB boundary. This catches the invalid pmds as
205 * well, as they are located before _text:
207 void __init cleanup_highmap(void)
209 unsigned long vaddr = __START_KERNEL_map;
210 unsigned long end = round_up((unsigned long)_end, PMD_SIZE) - 1;
211 pmd_t *pmd = level2_kernel_pgt;
212 pmd_t *last_pmd = pmd + PTRS_PER_PMD;
214 for (; pmd < last_pmd; pmd++, vaddr += PMD_SIZE) {
215 if (!pmd_present(*pmd))
217 if (vaddr < (unsigned long) _text || vaddr > end)
218 set_pmd(pmd, __pmd(0));
222 /* NOTE: this is meant to be run only at boot */
224 __set_fixmap(enum fixed_addresses idx, unsigned long phys, pgprot_t prot)
226 unsigned long address = __fix_to_virt(idx);
228 if (idx >= __end_of_fixed_addresses) {
229 printk(KERN_ERR "Invalid __set_fixmap\n");
232 set_pte_phys(address, phys, prot);
235 static unsigned long __initdata table_start;
236 static unsigned long __meminitdata table_end;
238 static __meminit void *alloc_low_page(unsigned long *phys)
240 unsigned long pfn = table_end++;
244 adr = (void *)get_zeroed_page(GFP_ATOMIC);
251 panic("alloc_low_page: ran out of memory");
253 adr = early_ioremap(pfn * PAGE_SIZE, PAGE_SIZE);
254 memset(adr, 0, PAGE_SIZE);
255 *phys = pfn * PAGE_SIZE;
259 static __meminit void unmap_low_page(void *adr)
264 early_iounmap(adr, PAGE_SIZE);
267 /* Must run before zap_low_mappings */
268 __meminit void *early_ioremap(unsigned long addr, unsigned long size)
270 pmd_t *pmd, *last_pmd;
274 pmds = ((addr & ~PMD_MASK) + size + ~PMD_MASK) / PMD_SIZE;
275 vaddr = __START_KERNEL_map;
276 pmd = level2_kernel_pgt;
277 last_pmd = level2_kernel_pgt + PTRS_PER_PMD - 1;
279 for (; pmd <= last_pmd; pmd++, vaddr += PMD_SIZE) {
280 for (i = 0; i < pmds; i++) {
281 if (pmd_present(pmd[i]))
282 goto continue_outer_loop;
284 vaddr += addr & ~PMD_MASK;
287 for (i = 0; i < pmds; i++, addr += PMD_SIZE)
288 set_pmd(pmd+i, __pmd(addr | __PAGE_KERNEL_LARGE_EXEC));
291 return (void *)vaddr;
295 printk(KERN_ERR "early_ioremap(0x%lx, %lu) failed\n", addr, size);
301 * To avoid virtual aliases later:
303 __meminit void early_iounmap(void *addr, unsigned long size)
309 vaddr = (unsigned long)addr;
310 pmds = ((vaddr & ~PMD_MASK) + size + ~PMD_MASK) / PMD_SIZE;
311 pmd = level2_kernel_pgt + pmd_index(vaddr);
313 for (i = 0; i < pmds; i++)
319 static void __meminit
320 phys_pmd_init(pmd_t *pmd_page, unsigned long address, unsigned long end)
322 int i = pmd_index(address);
324 for (; i < PTRS_PER_PMD; i++, address += PMD_SIZE) {
325 pmd_t *pmd = pmd_page + pmd_index(address);
327 if (address >= end) {
328 if (!after_bootmem) {
329 for (; i < PTRS_PER_PMD; i++, pmd++)
330 set_pmd(pmd, __pmd(0));
338 set_pte((pte_t *)pmd,
339 pfn_pte(address >> PAGE_SHIFT, PAGE_KERNEL_LARGE));
343 static void __meminit
344 phys_pmd_update(pud_t *pud, unsigned long address, unsigned long end)
346 pmd_t *pmd = pmd_offset(pud, 0);
347 spin_lock(&init_mm.page_table_lock);
348 phys_pmd_init(pmd, address, end);
349 spin_unlock(&init_mm.page_table_lock);
353 static void __meminit
354 phys_pud_init(pud_t *pud_page, unsigned long addr, unsigned long end)
356 int i = pud_index(addr);
358 for (; i < PTRS_PER_PUD; i++, addr = (addr & PUD_MASK) + PUD_SIZE) {
359 unsigned long pmd_phys;
360 pud_t *pud = pud_page + pud_index(addr);
366 if (!after_bootmem &&
367 !e820_any_mapped(addr, addr+PUD_SIZE, 0)) {
368 set_pud(pud, __pud(0));
373 if (!pud_large(*pud))
374 phys_pmd_update(pud, addr, end);
378 if (direct_gbpages) {
379 set_pte((pte_t *)pud,
380 pfn_pte(addr >> PAGE_SHIFT, PAGE_KERNEL_LARGE));
384 pmd = alloc_low_page(&pmd_phys);
386 spin_lock(&init_mm.page_table_lock);
387 set_pud(pud, __pud(pmd_phys | _KERNPG_TABLE));
388 phys_pmd_init(pmd, addr, end);
389 spin_unlock(&init_mm.page_table_lock);
396 static void __init find_early_table_space(unsigned long end)
398 unsigned long puds, pmds, tables, start;
400 puds = (end + PUD_SIZE - 1) >> PUD_SHIFT;
401 tables = round_up(puds * sizeof(pud_t), PAGE_SIZE);
402 if (!direct_gbpages) {
403 pmds = (end + PMD_SIZE - 1) >> PMD_SHIFT;
404 tables += round_up(pmds * sizeof(pmd_t), PAGE_SIZE);
408 * RED-PEN putting page tables only on node 0 could
409 * cause a hotspot and fill up ZONE_DMA. The page tables
410 * need roughly 0.5KB per GB.
413 table_start = find_e820_area(start, end, tables, PAGE_SIZE);
414 if (table_start == -1UL)
415 panic("Cannot find space for the kernel page tables");
417 table_start >>= PAGE_SHIFT;
418 table_end = table_start;
420 early_printk("kernel direct mapping tables up to %lx @ %lx-%lx\n",
421 end, table_start << PAGE_SHIFT,
422 (table_start << PAGE_SHIFT) + tables);
425 static void __init init_gbpages(void)
427 if (direct_gbpages && cpu_has_gbpages)
428 printk(KERN_INFO "Using GB pages for direct mapping\n");
434 * Setup the direct mapping of the physical memory at PAGE_OFFSET.
435 * This runs before bootmem is initialized and gets pages directly from
436 * the physical memory. To access them they are temporarily mapped.
438 void __init_refok init_memory_mapping(unsigned long start, unsigned long end)
442 pr_debug("init_memory_mapping\n");
445 * Find space for the kernel direct mapping tables.
447 * Later we should allocate these tables in the local node of the
448 * memory mapped. Unfortunately this is done currently before the
449 * nodes are discovered.
451 if (!after_bootmem) {
453 find_early_table_space(end);
456 start = (unsigned long)__va(start);
457 end = (unsigned long)__va(end);
459 for (; start < end; start = next) {
460 pgd_t *pgd = pgd_offset_k(start);
461 unsigned long pud_phys;
465 pud = pud_offset(pgd, start & PGDIR_MASK);
467 pud = alloc_low_page(&pud_phys);
469 next = start + PGDIR_SIZE;
472 phys_pud_init(pud, __pa(start), __pa(next));
474 set_pgd(pgd_offset_k(start), mk_kernel_pgd(pud_phys));
479 mmu_cr4_features = read_cr4();
483 reserve_early(table_start << PAGE_SHIFT,
484 table_end << PAGE_SHIFT, "PGTABLE");
488 void __init paging_init(void)
490 unsigned long max_zone_pfns[MAX_NR_ZONES];
492 memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
493 max_zone_pfns[ZONE_DMA] = MAX_DMA_PFN;
494 max_zone_pfns[ZONE_DMA32] = MAX_DMA32_PFN;
495 max_zone_pfns[ZONE_NORMAL] = end_pfn;
497 memory_present(0, 0, end_pfn);
499 free_area_init_nodes(max_zone_pfns);
504 * Memory hotplug specific functions
506 void online_page(struct page *page)
508 ClearPageReserved(page);
509 init_page_count(page);
515 #ifdef CONFIG_MEMORY_HOTPLUG
517 * Memory is added always to NORMAL zone. This means you will never get
518 * additional DMA/DMA32 memory.
520 int arch_add_memory(int nid, u64 start, u64 size)
522 struct pglist_data *pgdat = NODE_DATA(nid);
523 struct zone *zone = pgdat->node_zones + ZONE_NORMAL;
524 unsigned long start_pfn = start >> PAGE_SHIFT;
525 unsigned long nr_pages = size >> PAGE_SHIFT;
528 init_memory_mapping(start, start + size-1);
530 ret = __add_pages(zone, start_pfn, nr_pages);
535 EXPORT_SYMBOL_GPL(arch_add_memory);
537 #if !defined(CONFIG_ACPI_NUMA) && defined(CONFIG_NUMA)
538 int memory_add_physaddr_to_nid(u64 start)
542 EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid);
545 #endif /* CONFIG_MEMORY_HOTPLUG */
547 static struct kcore_list kcore_mem, kcore_vmalloc, kcore_kernel,
548 kcore_modules, kcore_vsyscall;
550 void __init mem_init(void)
552 long codesize, reservedpages, datasize, initsize;
556 /* clear_bss() already clear the empty_zero_page */
560 /* this will put all low memory onto the freelists */
562 totalram_pages = numa_free_all_bootmem();
564 totalram_pages = free_all_bootmem();
566 reservedpages = end_pfn - totalram_pages -
567 absent_pages_in_range(0, end_pfn);
570 codesize = (unsigned long) &_etext - (unsigned long) &_text;
571 datasize = (unsigned long) &_edata - (unsigned long) &_etext;
572 initsize = (unsigned long) &__init_end - (unsigned long) &__init_begin;
574 /* Register memory areas for /proc/kcore */
575 kclist_add(&kcore_mem, __va(0), max_low_pfn << PAGE_SHIFT);
576 kclist_add(&kcore_vmalloc, (void *)VMALLOC_START,
577 VMALLOC_END-VMALLOC_START);
578 kclist_add(&kcore_kernel, &_stext, _end - _stext);
579 kclist_add(&kcore_modules, (void *)MODULES_VADDR, MODULES_LEN);
580 kclist_add(&kcore_vsyscall, (void *)VSYSCALL_START,
581 VSYSCALL_END - VSYSCALL_START);
583 printk(KERN_INFO "Memory: %luk/%luk available (%ldk kernel code, "
584 "%ldk reserved, %ldk data, %ldk init)\n",
585 (unsigned long) nr_free_pages() << (PAGE_SHIFT-10),
586 end_pfn << (PAGE_SHIFT-10),
588 reservedpages << (PAGE_SHIFT-10),
595 void free_init_pages(char *what, unsigned long begin, unsigned long end)
597 unsigned long addr = begin;
603 * If debugging page accesses then do not free this memory but
604 * mark them not present - any buggy init-section access will
605 * create a kernel page fault:
607 #ifdef CONFIG_DEBUG_PAGEALLOC
608 printk(KERN_INFO "debug: unmapping init memory %08lx..%08lx\n",
609 begin, PAGE_ALIGN(end));
610 set_memory_np(begin, (end - begin) >> PAGE_SHIFT);
612 printk(KERN_INFO "Freeing %s: %luk freed\n", what, (end - begin) >> 10);
614 for (; addr < end; addr += PAGE_SIZE) {
615 ClearPageReserved(virt_to_page(addr));
616 init_page_count(virt_to_page(addr));
617 memset((void *)(addr & ~(PAGE_SIZE-1)),
618 POISON_FREE_INITMEM, PAGE_SIZE);
625 void free_initmem(void)
627 free_init_pages("unused kernel memory",
628 (unsigned long)(&__init_begin),
629 (unsigned long)(&__init_end));
632 #ifdef CONFIG_DEBUG_RODATA
633 const int rodata_test_data = 0xC3;
634 EXPORT_SYMBOL_GPL(rodata_test_data);
636 void mark_rodata_ro(void)
638 unsigned long start = PFN_ALIGN(_stext), end = PFN_ALIGN(__end_rodata);
640 printk(KERN_INFO "Write protecting the kernel read-only data: %luk\n",
641 (end - start) >> 10);
642 set_memory_ro(start, (end - start) >> PAGE_SHIFT);
645 * The rodata section (but not the kernel text!) should also be
648 start = ((unsigned long)__start_rodata + PAGE_SIZE - 1) & PAGE_MASK;
649 set_memory_nx(start, (end - start) >> PAGE_SHIFT);
653 #ifdef CONFIG_CPA_DEBUG
654 printk(KERN_INFO "Testing CPA: undo %lx-%lx\n", start, end);
655 set_memory_rw(start, (end-start) >> PAGE_SHIFT);
657 printk(KERN_INFO "Testing CPA: again\n");
658 set_memory_ro(start, (end-start) >> PAGE_SHIFT);
664 #ifdef CONFIG_BLK_DEV_INITRD
665 void free_initrd_mem(unsigned long start, unsigned long end)
667 free_init_pages("initrd memory", start, end);
671 void __init reserve_bootmem_generic(unsigned long phys, unsigned len)
674 int nid = phys_to_nid(phys);
676 unsigned long pfn = phys >> PAGE_SHIFT;
678 if (pfn >= end_pfn) {
680 * This can happen with kdump kernels when accessing
683 if (pfn < end_pfn_map)
686 printk(KERN_ERR "reserve_bootmem: illegal reserve %lx %u\n",
691 /* Should check here against the e820 map to avoid double free */
693 reserve_bootmem_node(NODE_DATA(nid), phys, len, BOOTMEM_DEFAULT);
695 reserve_bootmem(phys, len, BOOTMEM_DEFAULT);
697 if (phys+len <= MAX_DMA_PFN*PAGE_SIZE) {
698 dma_reserve += len / PAGE_SIZE;
699 set_dma_reserve(dma_reserve);
703 int kern_addr_valid(unsigned long addr)
705 unsigned long above = ((long)addr) >> __VIRTUAL_MASK_SHIFT;
711 if (above != 0 && above != -1UL)
714 pgd = pgd_offset_k(addr);
718 pud = pud_offset(pgd, addr);
722 pmd = pmd_offset(pud, addr);
727 return pfn_valid(pmd_pfn(*pmd));
729 pte = pte_offset_kernel(pmd, addr);
733 return pfn_valid(pte_pfn(*pte));
737 * A pseudo VMA to allow ptrace access for the vsyscall page. This only
738 * covers the 64bit vsyscall page now. 32bit has a real VMA now and does
739 * not need special handling anymore:
741 static struct vm_area_struct gate_vma = {
742 .vm_start = VSYSCALL_START,
743 .vm_end = VSYSCALL_START + (VSYSCALL_MAPPED_PAGES * PAGE_SIZE),
744 .vm_page_prot = PAGE_READONLY_EXEC,
745 .vm_flags = VM_READ | VM_EXEC
748 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
750 #ifdef CONFIG_IA32_EMULATION
751 if (test_tsk_thread_flag(tsk, TIF_IA32))
757 int in_gate_area(struct task_struct *task, unsigned long addr)
759 struct vm_area_struct *vma = get_gate_vma(task);
764 return (addr >= vma->vm_start) && (addr < vma->vm_end);
768 * Use this when you have no reliable task/vma, typically from interrupt
769 * context. It is less reliable than using the task's vma and may give
772 int in_gate_area_no_task(unsigned long addr)
774 return (addr >= VSYSCALL_START) && (addr < VSYSCALL_END);
777 const char *arch_vma_name(struct vm_area_struct *vma)
779 if (vma->vm_mm && vma->vm_start == (long)vma->vm_mm->context.vdso)
781 if (vma == &gate_vma)
786 #ifdef CONFIG_SPARSEMEM_VMEMMAP
788 * Initialise the sparsemem vmemmap using huge-pages at the PMD level.
791 vmemmap_populate(struct page *start_page, unsigned long size, int node)
793 unsigned long addr = (unsigned long)start_page;
794 unsigned long end = (unsigned long)(start_page + size);
800 for (; addr < end; addr = next) {
801 next = pmd_addr_end(addr, end);
803 pgd = vmemmap_pgd_populate(addr, node);
807 pud = vmemmap_pud_populate(pgd, addr, node);
811 pmd = pmd_offset(pud, addr);
812 if (pmd_none(*pmd)) {
816 p = vmemmap_alloc_block(PMD_SIZE, node);
820 entry = pfn_pte(__pa(p) >> PAGE_SHIFT,
822 set_pmd(pmd, __pmd(pte_val(entry)));
824 printk(KERN_DEBUG " [%lx-%lx] PMD ->%p on node %d\n",
825 addr, addr + PMD_SIZE - 1, p, node);
827 vmemmap_verify((pte_t *)pmd, node, addr, next);