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);
58 * NOTE: pagetable_init alloc all the fixmap pagetables contiguous on the
59 * physical space so we can cache the place of the first one and move
60 * around without checking the pgd every time.
65 long i, total = 0, reserved = 0;
66 long shared = 0, cached = 0;
70 printk(KERN_INFO "Mem-info:\n");
72 printk(KERN_INFO "Free swap: %6ldkB\n",
73 nr_swap_pages << (PAGE_SHIFT-10));
75 for_each_online_pgdat(pgdat) {
76 for (i = 0; i < pgdat->node_spanned_pages; ++i) {
78 * This loop can take a while with 256 GB and
79 * 4k pages so defer the NMI watchdog:
81 if (unlikely(i % MAX_ORDER_NR_PAGES == 0))
84 if (!pfn_valid(pgdat->node_start_pfn + i))
87 page = pfn_to_page(pgdat->node_start_pfn + i);
89 if (PageReserved(page))
91 else if (PageSwapCache(page))
93 else if (page_count(page))
94 shared += page_count(page) - 1;
97 printk(KERN_INFO "%lu pages of RAM\n", total);
98 printk(KERN_INFO "%lu reserved pages\n", reserved);
99 printk(KERN_INFO "%lu pages shared\n", shared);
100 printk(KERN_INFO "%lu pages swap cached\n", cached);
105 static __init void *spp_getpage(void)
110 ptr = (void *) get_zeroed_page(GFP_ATOMIC);
112 ptr = alloc_bootmem_pages(PAGE_SIZE);
114 if (!ptr || ((unsigned long)ptr & ~PAGE_MASK)) {
115 panic("set_pte_phys: cannot allocate page data %s\n",
116 after_bootmem ? "after bootmem" : "");
119 pr_debug("spp_getpage %p\n", ptr);
125 set_pte_phys(unsigned long vaddr, unsigned long phys, pgprot_t prot)
132 pr_debug("set_pte_phys %lx to %lx\n", vaddr, phys);
134 pgd = pgd_offset_k(vaddr);
135 if (pgd_none(*pgd)) {
137 "PGD FIXMAP MISSING, it should be setup in head.S!\n");
140 pud = pud_offset(pgd, vaddr);
141 if (pud_none(*pud)) {
142 pmd = (pmd_t *) spp_getpage();
143 set_pud(pud, __pud(__pa(pmd) | _KERNPG_TABLE | _PAGE_USER));
144 if (pmd != pmd_offset(pud, 0)) {
145 printk(KERN_ERR "PAGETABLE BUG #01! %p <-> %p\n",
146 pmd, pmd_offset(pud, 0));
150 pmd = pmd_offset(pud, vaddr);
151 if (pmd_none(*pmd)) {
152 pte = (pte_t *) spp_getpage();
153 set_pmd(pmd, __pmd(__pa(pte) | _KERNPG_TABLE | _PAGE_USER));
154 if (pte != pte_offset_kernel(pmd, 0)) {
155 printk(KERN_ERR "PAGETABLE BUG #02!\n");
159 new_pte = pfn_pte(phys >> PAGE_SHIFT, prot);
161 pte = pte_offset_kernel(pmd, vaddr);
162 if (!pte_none(*pte) &&
163 pte_val(*pte) != (pte_val(new_pte) & __supported_pte_mask))
165 set_pte(pte, new_pte);
168 * It's enough to flush this one mapping.
169 * (PGE mappings get flushed as well)
171 __flush_tlb_one(vaddr);
175 * The head.S code sets up the kernel high mapping from:
176 * __START_KERNEL_map to __START_KERNEL_map + KERNEL_TEXT_SIZE
178 * phys_addr holds the negative offset to the kernel, which is added
179 * to the compile time generated pmds. This results in invalid pmds up
180 * to the point where we hit the physaddr 0 mapping.
182 * We limit the mappings to the region from _text to _end. _end is
183 * rounded up to the 2MB boundary. This catches the invalid pmds as
184 * well, as they are located before _text:
186 void __init cleanup_highmap(void)
188 unsigned long vaddr = __START_KERNEL_map;
189 unsigned long end = round_up((unsigned long)_end, PMD_SIZE) - 1;
190 pmd_t *pmd = level2_kernel_pgt;
191 pmd_t *last_pmd = pmd + PTRS_PER_PMD;
193 for (; pmd < last_pmd; pmd++, vaddr += PMD_SIZE) {
194 if (!pmd_present(*pmd))
196 if (vaddr < (unsigned long) _text || vaddr > end)
197 set_pmd(pmd, __pmd(0));
201 /* NOTE: this is meant to be run only at boot */
203 __set_fixmap(enum fixed_addresses idx, unsigned long phys, pgprot_t prot)
205 unsigned long address = __fix_to_virt(idx);
207 if (idx >= __end_of_fixed_addresses) {
208 printk(KERN_ERR "Invalid __set_fixmap\n");
211 set_pte_phys(address, phys, prot);
214 static unsigned long __initdata table_start;
215 static unsigned long __meminitdata table_end;
217 static __meminit void *alloc_low_page(unsigned long *phys)
219 unsigned long pfn = table_end++;
223 adr = (void *)get_zeroed_page(GFP_ATOMIC);
230 panic("alloc_low_page: ran out of memory");
232 adr = early_ioremap(pfn * PAGE_SIZE, PAGE_SIZE);
233 memset(adr, 0, PAGE_SIZE);
234 *phys = pfn * PAGE_SIZE;
238 static __meminit void unmap_low_page(void *adr)
243 early_iounmap(adr, PAGE_SIZE);
246 /* Must run before zap_low_mappings */
247 __meminit void *early_ioremap(unsigned long addr, unsigned long size)
249 pmd_t *pmd, *last_pmd;
253 pmds = ((addr & ~PMD_MASK) + size + ~PMD_MASK) / PMD_SIZE;
254 vaddr = __START_KERNEL_map;
255 pmd = level2_kernel_pgt;
256 last_pmd = level2_kernel_pgt + PTRS_PER_PMD - 1;
258 for (; pmd <= last_pmd; pmd++, vaddr += PMD_SIZE) {
259 for (i = 0; i < pmds; i++) {
260 if (pmd_present(pmd[i]))
261 goto continue_outer_loop;
263 vaddr += addr & ~PMD_MASK;
266 for (i = 0; i < pmds; i++, addr += PMD_SIZE)
267 set_pmd(pmd+i, __pmd(addr | __PAGE_KERNEL_LARGE_EXEC));
270 return (void *)vaddr;
274 printk(KERN_ERR "early_ioremap(0x%lx, %lu) failed\n", addr, size);
280 * To avoid virtual aliases later:
282 __meminit void early_iounmap(void *addr, unsigned long size)
288 vaddr = (unsigned long)addr;
289 pmds = ((vaddr & ~PMD_MASK) + size + ~PMD_MASK) / PMD_SIZE;
290 pmd = level2_kernel_pgt + pmd_index(vaddr);
292 for (i = 0; i < pmds; i++)
298 static void __meminit
299 phys_pmd_init(pmd_t *pmd_page, unsigned long address, unsigned long end)
301 int i = pmd_index(address);
303 for (; i < PTRS_PER_PMD; i++, address += PMD_SIZE) {
304 pmd_t *pmd = pmd_page + pmd_index(address);
306 if (address >= end) {
307 if (!after_bootmem) {
308 for (; i < PTRS_PER_PMD; i++, pmd++)
309 set_pmd(pmd, __pmd(0));
317 set_pte((pte_t *)pmd,
318 pfn_pte(address >> PAGE_SHIFT, PAGE_KERNEL_LARGE));
322 static void __meminit
323 phys_pmd_update(pud_t *pud, unsigned long address, unsigned long end)
325 pmd_t *pmd = pmd_offset(pud, 0);
326 spin_lock(&init_mm.page_table_lock);
327 phys_pmd_init(pmd, address, end);
328 spin_unlock(&init_mm.page_table_lock);
332 static void __meminit
333 phys_pud_init(pud_t *pud_page, unsigned long addr, unsigned long end)
335 int i = pud_index(addr);
337 for (; i < PTRS_PER_PUD; i++, addr = (addr & PUD_MASK) + PUD_SIZE) {
338 unsigned long pmd_phys;
339 pud_t *pud = pud_page + pud_index(addr);
345 if (!after_bootmem &&
346 !e820_any_mapped(addr, addr+PUD_SIZE, 0)) {
347 set_pud(pud, __pud(0));
352 phys_pmd_update(pud, addr, end);
356 pmd = alloc_low_page(&pmd_phys);
358 spin_lock(&init_mm.page_table_lock);
359 set_pud(pud, __pud(pmd_phys | _KERNPG_TABLE));
360 phys_pmd_init(pmd, addr, end);
361 spin_unlock(&init_mm.page_table_lock);
368 static void __init find_early_table_space(unsigned long end)
370 unsigned long puds, pmds, tables, start;
372 puds = (end + PUD_SIZE - 1) >> PUD_SHIFT;
373 pmds = (end + PMD_SIZE - 1) >> PMD_SHIFT;
374 tables = round_up(puds * sizeof(pud_t), PAGE_SIZE) +
375 round_up(pmds * sizeof(pmd_t), PAGE_SIZE);
378 * RED-PEN putting page tables only on node 0 could
379 * cause a hotspot and fill up ZONE_DMA. The page tables
380 * need roughly 0.5KB per GB.
383 table_start = find_e820_area(start, end, tables, PAGE_SIZE);
384 if (table_start == -1UL)
385 panic("Cannot find space for the kernel page tables");
387 table_start >>= PAGE_SHIFT;
388 table_end = table_start;
390 early_printk("kernel direct mapping tables up to %lx @ %lx-%lx\n",
391 end, table_start << PAGE_SHIFT,
392 (table_start << PAGE_SHIFT) + tables);
396 * Setup the direct mapping of the physical memory at PAGE_OFFSET.
397 * This runs before bootmem is initialized and gets pages directly from
398 * the physical memory. To access them they are temporarily mapped.
400 void __init_refok init_memory_mapping(unsigned long start, unsigned long end)
404 pr_debug("init_memory_mapping\n");
407 * Find space for the kernel direct mapping tables.
409 * Later we should allocate these tables in the local node of the
410 * memory mapped. Unfortunately this is done currently before the
411 * nodes are discovered.
414 find_early_table_space(end);
416 start = (unsigned long)__va(start);
417 end = (unsigned long)__va(end);
419 for (; start < end; start = next) {
420 pgd_t *pgd = pgd_offset_k(start);
421 unsigned long pud_phys;
425 pud = pud_offset(pgd, start & PGDIR_MASK);
427 pud = alloc_low_page(&pud_phys);
429 next = start + PGDIR_SIZE;
432 phys_pud_init(pud, __pa(start), __pa(next));
434 set_pgd(pgd_offset_k(start), mk_kernel_pgd(pud_phys));
439 mmu_cr4_features = read_cr4();
443 reserve_early(table_start << PAGE_SHIFT,
444 table_end << PAGE_SHIFT, "PGTABLE");
448 void __init paging_init(void)
450 unsigned long max_zone_pfns[MAX_NR_ZONES];
452 memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
453 max_zone_pfns[ZONE_DMA] = MAX_DMA_PFN;
454 max_zone_pfns[ZONE_DMA32] = MAX_DMA32_PFN;
455 max_zone_pfns[ZONE_NORMAL] = end_pfn;
457 memory_present(0, 0, end_pfn);
459 free_area_init_nodes(max_zone_pfns);
464 * Memory hotplug specific functions
466 void online_page(struct page *page)
468 ClearPageReserved(page);
469 init_page_count(page);
475 #ifdef CONFIG_MEMORY_HOTPLUG
477 * Memory is added always to NORMAL zone. This means you will never get
478 * additional DMA/DMA32 memory.
480 int arch_add_memory(int nid, u64 start, u64 size)
482 struct pglist_data *pgdat = NODE_DATA(nid);
483 struct zone *zone = pgdat->node_zones + ZONE_NORMAL;
484 unsigned long start_pfn = start >> PAGE_SHIFT;
485 unsigned long nr_pages = size >> PAGE_SHIFT;
488 init_memory_mapping(start, start + size-1);
490 ret = __add_pages(zone, start_pfn, nr_pages);
495 EXPORT_SYMBOL_GPL(arch_add_memory);
497 #if !defined(CONFIG_ACPI_NUMA) && defined(CONFIG_NUMA)
498 int memory_add_physaddr_to_nid(u64 start)
502 EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid);
505 #endif /* CONFIG_MEMORY_HOTPLUG */
507 static struct kcore_list kcore_mem, kcore_vmalloc, kcore_kernel,
508 kcore_modules, kcore_vsyscall;
510 void __init mem_init(void)
512 long codesize, reservedpages, datasize, initsize;
516 /* clear_bss() already clear the empty_zero_page */
518 /* temporary debugging - double check it's true: */
522 for (i = 0; i < 1024; i++)
523 WARN_ON_ONCE(empty_zero_page[i]);
528 /* this will put all low memory onto the freelists */
530 totalram_pages = numa_free_all_bootmem();
532 totalram_pages = free_all_bootmem();
534 reservedpages = end_pfn - totalram_pages -
535 absent_pages_in_range(0, end_pfn);
538 codesize = (unsigned long) &_etext - (unsigned long) &_text;
539 datasize = (unsigned long) &_edata - (unsigned long) &_etext;
540 initsize = (unsigned long) &__init_end - (unsigned long) &__init_begin;
542 /* Register memory areas for /proc/kcore */
543 kclist_add(&kcore_mem, __va(0), max_low_pfn << PAGE_SHIFT);
544 kclist_add(&kcore_vmalloc, (void *)VMALLOC_START,
545 VMALLOC_END-VMALLOC_START);
546 kclist_add(&kcore_kernel, &_stext, _end - _stext);
547 kclist_add(&kcore_modules, (void *)MODULES_VADDR, MODULES_LEN);
548 kclist_add(&kcore_vsyscall, (void *)VSYSCALL_START,
549 VSYSCALL_END - VSYSCALL_START);
551 printk(KERN_INFO "Memory: %luk/%luk available (%ldk kernel code, "
552 "%ldk reserved, %ldk data, %ldk init)\n",
553 (unsigned long) nr_free_pages() << (PAGE_SHIFT-10),
554 end_pfn << (PAGE_SHIFT-10),
556 reservedpages << (PAGE_SHIFT-10),
563 void free_init_pages(char *what, unsigned long begin, unsigned long end)
565 unsigned long addr = begin;
571 * If debugging page accesses then do not free this memory but
572 * mark them not present - any buggy init-section access will
573 * create a kernel page fault:
575 #ifdef CONFIG_DEBUG_PAGEALLOC
576 printk(KERN_INFO "debug: unmapping init memory %08lx..%08lx\n",
577 begin, PAGE_ALIGN(end));
578 set_memory_np(begin, (end - begin) >> PAGE_SHIFT);
580 printk(KERN_INFO "Freeing %s: %luk freed\n", what, (end - begin) >> 10);
582 for (; addr < end; addr += PAGE_SIZE) {
583 ClearPageReserved(virt_to_page(addr));
584 init_page_count(virt_to_page(addr));
585 memset((void *)(addr & ~(PAGE_SIZE-1)),
586 POISON_FREE_INITMEM, PAGE_SIZE);
593 void free_initmem(void)
595 free_init_pages("unused kernel memory",
596 (unsigned long)(&__init_begin),
597 (unsigned long)(&__init_end));
600 #ifdef CONFIG_DEBUG_RODATA
601 const int rodata_test_data = 0xC3;
602 EXPORT_SYMBOL_GPL(rodata_test_data);
604 void mark_rodata_ro(void)
606 unsigned long start = (unsigned long)_stext, end;
608 #ifdef CONFIG_HOTPLUG_CPU
609 /* It must still be possible to apply SMP alternatives. */
610 if (num_possible_cpus() > 1)
611 start = (unsigned long)_etext;
614 #ifdef CONFIG_KPROBES
615 start = (unsigned long)__start_rodata;
618 end = (unsigned long)__end_rodata;
619 start = (start + PAGE_SIZE - 1) & PAGE_MASK;
625 printk(KERN_INFO "Write protecting the kernel read-only data: %luk\n",
626 (end - start) >> 10);
627 set_memory_ro(start, (end - start) >> PAGE_SHIFT);
630 * The rodata section (but not the kernel text!) should also be
633 start = ((unsigned long)__start_rodata + PAGE_SIZE - 1) & PAGE_MASK;
634 set_memory_nx(start, (end - start) >> PAGE_SHIFT);
638 #ifdef CONFIG_CPA_DEBUG
639 printk(KERN_INFO "Testing CPA: undo %lx-%lx\n", start, end);
640 set_memory_rw(start, (end-start) >> PAGE_SHIFT);
642 printk(KERN_INFO "Testing CPA: again\n");
643 set_memory_ro(start, (end-start) >> PAGE_SHIFT);
648 #ifdef CONFIG_BLK_DEV_INITRD
649 void free_initrd_mem(unsigned long start, unsigned long end)
651 free_init_pages("initrd memory", start, end);
655 void __init reserve_bootmem_generic(unsigned long phys, unsigned len)
658 int nid = phys_to_nid(phys);
660 unsigned long pfn = phys >> PAGE_SHIFT;
662 if (pfn >= end_pfn) {
664 * This can happen with kdump kernels when accessing
667 if (pfn < end_pfn_map)
670 printk(KERN_ERR "reserve_bootmem: illegal reserve %lx %u\n",
675 /* Should check here against the e820 map to avoid double free */
677 reserve_bootmem_node(NODE_DATA(nid), phys, len, BOOTMEM_DEFAULT);
679 reserve_bootmem(phys, len, BOOTMEM_DEFAULT);
681 if (phys+len <= MAX_DMA_PFN*PAGE_SIZE) {
682 dma_reserve += len / PAGE_SIZE;
683 set_dma_reserve(dma_reserve);
687 int kern_addr_valid(unsigned long addr)
689 unsigned long above = ((long)addr) >> __VIRTUAL_MASK_SHIFT;
695 if (above != 0 && above != -1UL)
698 pgd = pgd_offset_k(addr);
702 pud = pud_offset(pgd, addr);
706 pmd = pmd_offset(pud, addr);
711 return pfn_valid(pmd_pfn(*pmd));
713 pte = pte_offset_kernel(pmd, addr);
717 return pfn_valid(pte_pfn(*pte));
721 * A pseudo VMA to allow ptrace access for the vsyscall page. This only
722 * covers the 64bit vsyscall page now. 32bit has a real VMA now and does
723 * not need special handling anymore:
725 static struct vm_area_struct gate_vma = {
726 .vm_start = VSYSCALL_START,
727 .vm_end = VSYSCALL_START + (VSYSCALL_MAPPED_PAGES * PAGE_SIZE),
728 .vm_page_prot = PAGE_READONLY_EXEC,
729 .vm_flags = VM_READ | VM_EXEC
732 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
734 #ifdef CONFIG_IA32_EMULATION
735 if (test_tsk_thread_flag(tsk, TIF_IA32))
741 int in_gate_area(struct task_struct *task, unsigned long addr)
743 struct vm_area_struct *vma = get_gate_vma(task);
748 return (addr >= vma->vm_start) && (addr < vma->vm_end);
752 * Use this when you have no reliable task/vma, typically from interrupt
753 * context. It is less reliable than using the task's vma and may give
756 int in_gate_area_no_task(unsigned long addr)
758 return (addr >= VSYSCALL_START) && (addr < VSYSCALL_END);
761 const char *arch_vma_name(struct vm_area_struct *vma)
763 if (vma->vm_mm && vma->vm_start == (long)vma->vm_mm->context.vdso)
765 if (vma == &gate_vma)
770 #ifdef CONFIG_SPARSEMEM_VMEMMAP
772 * Initialise the sparsemem vmemmap using huge-pages at the PMD level.
775 vmemmap_populate(struct page *start_page, unsigned long size, int node)
777 unsigned long addr = (unsigned long)start_page;
778 unsigned long end = (unsigned long)(start_page + size);
784 for (; addr < end; addr = next) {
785 next = pmd_addr_end(addr, end);
787 pgd = vmemmap_pgd_populate(addr, node);
791 pud = vmemmap_pud_populate(pgd, addr, node);
795 pmd = pmd_offset(pud, addr);
796 if (pmd_none(*pmd)) {
800 p = vmemmap_alloc_block(PMD_SIZE, node);
804 entry = pfn_pte(__pa(p) >> PAGE_SHIFT,
806 set_pmd(pmd, __pmd(pte_val(entry)));
808 printk(KERN_DEBUG " [%lx-%lx] PMD ->%p on node %d\n",
809 addr, addr + PMD_SIZE - 1, p, node);
811 vmemmap_verify((pte_t *)pmd, node, addr, next);