2 * Initialize MMU support.
4 * Copyright (C) 1998-2003 Hewlett-Packard Co
5 * David Mosberger-Tang <davidm@hpl.hp.com>
7 #include <linux/kernel.h>
8 #include <linux/init.h>
10 #include <linux/bootmem.h>
11 #include <linux/efi.h>
12 #include <linux/elf.h>
13 #include <linux/memblock.h>
15 #include <linux/mmzone.h>
16 #include <linux/module.h>
17 #include <linux/personality.h>
18 #include <linux/reboot.h>
19 #include <linux/slab.h>
20 #include <linux/swap.h>
21 #include <linux/proc_fs.h>
22 #include <linux/bitops.h>
23 #include <linux/kexec.h>
27 #include <asm/machvec.h>
29 #include <asm/patch.h>
30 #include <asm/pgalloc.h>
32 #include <asm/sections.h>
34 #include <asm/uaccess.h>
35 #include <asm/unistd.h>
38 extern void ia64_tlb_init (void);
40 unsigned long MAX_DMA_ADDRESS = PAGE_OFFSET + 0x100000000UL;
42 #ifdef CONFIG_VIRTUAL_MEM_MAP
43 unsigned long VMALLOC_END = VMALLOC_END_INIT;
44 EXPORT_SYMBOL(VMALLOC_END);
45 struct page *vmem_map;
46 EXPORT_SYMBOL(vmem_map);
49 struct page *zero_page_memmap_ptr; /* map entry for zero page */
50 EXPORT_SYMBOL(zero_page_memmap_ptr);
53 __ia64_sync_icache_dcache (pte_t pte)
59 addr = (unsigned long) page_address(page);
61 if (test_bit(PG_arch_1, &page->flags))
62 return; /* i-cache is already coherent with d-cache */
64 flush_icache_range(addr, addr + (PAGE_SIZE << compound_order(page)));
65 set_bit(PG_arch_1, &page->flags); /* mark page as clean */
69 * Since DMA is i-cache coherent, any (complete) pages that were written via
70 * DMA can be marked as "clean" so that lazy_mmu_prot_update() doesn't have to
71 * flush them when they get mapped into an executable vm-area.
74 dma_mark_clean(void *addr, size_t size)
76 unsigned long pg_addr, end;
78 pg_addr = PAGE_ALIGN((unsigned long) addr);
79 end = (unsigned long) addr + size;
80 while (pg_addr + PAGE_SIZE <= end) {
81 struct page *page = virt_to_page(pg_addr);
82 set_bit(PG_arch_1, &page->flags);
88 ia64_set_rbs_bot (void)
90 unsigned long stack_size = rlimit_max(RLIMIT_STACK) & -16;
92 if (stack_size > MAX_USER_STACK_SIZE)
93 stack_size = MAX_USER_STACK_SIZE;
94 current->thread.rbs_bot = PAGE_ALIGN(current->mm->start_stack - stack_size);
98 * This performs some platform-dependent address space initialization.
99 * On IA-64, we want to setup the VM area for the register backing
100 * store (which grows upwards) and install the gateway page which is
101 * used for signal trampolines, etc.
104 ia64_init_addr_space (void)
106 struct vm_area_struct *vma;
111 * If we're out of memory and kmem_cache_alloc() returns NULL, we simply ignore
112 * the problem. When the process attempts to write to the register backing store
113 * for the first time, it will get a SEGFAULT in this case.
115 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
117 INIT_LIST_HEAD(&vma->anon_vma_chain);
118 vma->vm_mm = current->mm;
119 vma->vm_start = current->thread.rbs_bot & PAGE_MASK;
120 vma->vm_end = vma->vm_start + PAGE_SIZE;
121 vma->vm_flags = VM_DATA_DEFAULT_FLAGS|VM_GROWSUP|VM_ACCOUNT;
122 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
123 down_write(¤t->mm->mmap_sem);
124 if (insert_vm_struct(current->mm, vma)) {
125 up_write(¤t->mm->mmap_sem);
126 kmem_cache_free(vm_area_cachep, vma);
129 up_write(¤t->mm->mmap_sem);
132 /* map NaT-page at address zero to speed up speculative dereferencing of NULL: */
133 if (!(current->personality & MMAP_PAGE_ZERO)) {
134 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
136 INIT_LIST_HEAD(&vma->anon_vma_chain);
137 vma->vm_mm = current->mm;
138 vma->vm_end = PAGE_SIZE;
139 vma->vm_page_prot = __pgprot(pgprot_val(PAGE_READONLY) | _PAGE_MA_NAT);
140 vma->vm_flags = VM_READ | VM_MAYREAD | VM_IO |
141 VM_DONTEXPAND | VM_DONTDUMP;
142 down_write(¤t->mm->mmap_sem);
143 if (insert_vm_struct(current->mm, vma)) {
144 up_write(¤t->mm->mmap_sem);
145 kmem_cache_free(vm_area_cachep, vma);
148 up_write(¤t->mm->mmap_sem);
156 free_reserved_area(ia64_imva(__init_begin), ia64_imva(__init_end),
157 -1, "unused kernel");
161 free_initrd_mem (unsigned long start, unsigned long end)
164 * EFI uses 4KB pages while the kernel can use 4KB or bigger.
165 * Thus EFI and the kernel may have different page sizes. It is
166 * therefore possible to have the initrd share the same page as
167 * the end of the kernel (given current setup).
169 * To avoid freeing/using the wrong page (kernel sized) we:
170 * - align up the beginning of initrd
171 * - align down the end of initrd
174 * |=============| a000
180 * |=============| 8000
183 * |/////////////| 7000
186 * |=============| 6000
189 * K=kernel using 8KB pages
191 * In this example, we must free page 8000 ONLY. So we must align up
192 * initrd_start and keep initrd_end as is.
194 start = PAGE_ALIGN(start);
195 end = end & PAGE_MASK;
198 printk(KERN_INFO "Freeing initrd memory: %ldkB freed\n", (end - start) >> 10);
200 for (; start < end; start += PAGE_SIZE) {
201 if (!virt_addr_valid(start))
203 free_reserved_page(virt_to_page(start));
208 * This installs a clean page in the kernel's page table.
210 static struct page * __init
211 put_kernel_page (struct page *page, unsigned long address, pgprot_t pgprot)
218 if (!PageReserved(page))
219 printk(KERN_ERR "put_kernel_page: page at 0x%p not in reserved memory\n",
222 pgd = pgd_offset_k(address); /* note: this is NOT pgd_offset()! */
225 pud = pud_alloc(&init_mm, pgd, address);
228 pmd = pmd_alloc(&init_mm, pud, address);
231 pte = pte_alloc_kernel(pmd, address);
236 set_pte(pte, mk_pte(page, pgprot));
239 /* no need for flush_tlb */
249 * Map the gate page twice: once read-only to export the ELF
250 * headers etc. and once execute-only page to enable
251 * privilege-promotion via "epc":
253 page = virt_to_page(ia64_imva(__start_gate_section));
254 put_kernel_page(page, GATE_ADDR, PAGE_READONLY);
255 #ifdef HAVE_BUGGY_SEGREL
256 page = virt_to_page(ia64_imva(__start_gate_section + PAGE_SIZE));
257 put_kernel_page(page, GATE_ADDR + PAGE_SIZE, PAGE_GATE);
259 put_kernel_page(page, GATE_ADDR + PERCPU_PAGE_SIZE, PAGE_GATE);
260 /* Fill in the holes (if any) with read-only zero pages: */
264 for (addr = GATE_ADDR + PAGE_SIZE;
265 addr < GATE_ADDR + PERCPU_PAGE_SIZE;
268 put_kernel_page(ZERO_PAGE(0), addr,
270 put_kernel_page(ZERO_PAGE(0), addr + PERCPU_PAGE_SIZE,
278 static struct vm_area_struct gate_vma;
280 static int __init gate_vma_init(void)
282 gate_vma.vm_mm = NULL;
283 gate_vma.vm_start = FIXADDR_USER_START;
284 gate_vma.vm_end = FIXADDR_USER_END;
285 gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
286 gate_vma.vm_page_prot = __P101;
290 __initcall(gate_vma_init);
292 struct vm_area_struct *get_gate_vma(struct mm_struct *mm)
297 int in_gate_area_no_mm(unsigned long addr)
299 if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
304 int in_gate_area(struct mm_struct *mm, unsigned long addr)
306 return in_gate_area_no_mm(addr);
309 void ia64_mmu_init(void *my_cpu_data)
311 unsigned long pta, impl_va_bits;
312 extern void tlb_init(void);
314 #ifdef CONFIG_DISABLE_VHPT
315 # define VHPT_ENABLE_BIT 0
317 # define VHPT_ENABLE_BIT 1
321 * Check if the virtually mapped linear page table (VMLPT) overlaps with a mapped
322 * address space. The IA-64 architecture guarantees that at least 50 bits of
323 * virtual address space are implemented but if we pick a large enough page size
324 * (e.g., 64KB), the mapped address space is big enough that it will overlap with
325 * VMLPT. I assume that once we run on machines big enough to warrant 64KB pages,
326 * IMPL_VA_MSB will be significantly bigger, so this is unlikely to become a
327 * problem in practice. Alternatively, we could truncate the top of the mapped
328 * address space to not permit mappings that would overlap with the VMLPT.
332 # define mapped_space_bits (3*(PAGE_SHIFT - pte_bits) + PAGE_SHIFT)
334 * The virtual page table has to cover the entire implemented address space within
335 * a region even though not all of this space may be mappable. The reason for
336 * this is that the Access bit and Dirty bit fault handlers perform
337 * non-speculative accesses to the virtual page table, so the address range of the
338 * virtual page table itself needs to be covered by virtual page table.
340 # define vmlpt_bits (impl_va_bits - PAGE_SHIFT + pte_bits)
341 # define POW2(n) (1ULL << (n))
343 impl_va_bits = ffz(~(local_cpu_data->unimpl_va_mask | (7UL << 61)));
345 if (impl_va_bits < 51 || impl_va_bits > 61)
346 panic("CPU has bogus IMPL_VA_MSB value of %lu!\n", impl_va_bits - 1);
348 * mapped_space_bits - PAGE_SHIFT is the total number of ptes we need,
349 * which must fit into "vmlpt_bits - pte_bits" slots. Second half of
350 * the test makes sure that our mapped space doesn't overlap the
351 * unimplemented hole in the middle of the region.
353 if ((mapped_space_bits - PAGE_SHIFT > vmlpt_bits - pte_bits) ||
354 (mapped_space_bits > impl_va_bits - 1))
355 panic("Cannot build a big enough virtual-linear page table"
356 " to cover mapped address space.\n"
357 " Try using a smaller page size.\n");
360 /* place the VMLPT at the end of each page-table mapped region: */
361 pta = POW2(61) - POW2(vmlpt_bits);
364 * Set the (virtually mapped linear) page table address. Bit
365 * 8 selects between the short and long format, bits 2-7 the
366 * size of the table, and bit 0 whether the VHPT walker is
369 ia64_set_pta(pta | (0 << 8) | (vmlpt_bits << 2) | VHPT_ENABLE_BIT);
373 #ifdef CONFIG_HUGETLB_PAGE
374 ia64_set_rr(HPAGE_REGION_BASE, HPAGE_SHIFT << 2);
379 #ifdef CONFIG_VIRTUAL_MEM_MAP
380 int vmemmap_find_next_valid_pfn(int node, int i)
382 unsigned long end_address, hole_next_pfn;
383 unsigned long stop_address;
384 pg_data_t *pgdat = NODE_DATA(node);
386 end_address = (unsigned long) &vmem_map[pgdat->node_start_pfn + i];
387 end_address = PAGE_ALIGN(end_address);
388 stop_address = (unsigned long) &vmem_map[pgdat_end_pfn(pgdat)];
396 pgd = pgd_offset_k(end_address);
397 if (pgd_none(*pgd)) {
398 end_address += PGDIR_SIZE;
402 pud = pud_offset(pgd, end_address);
403 if (pud_none(*pud)) {
404 end_address += PUD_SIZE;
408 pmd = pmd_offset(pud, end_address);
409 if (pmd_none(*pmd)) {
410 end_address += PMD_SIZE;
414 pte = pte_offset_kernel(pmd, end_address);
416 if (pte_none(*pte)) {
417 end_address += PAGE_SIZE;
419 if ((end_address < stop_address) &&
420 (end_address != ALIGN(end_address, 1UL << PMD_SHIFT)))
424 /* Found next valid vmem_map page */
426 } while (end_address < stop_address);
428 end_address = min(end_address, stop_address);
429 end_address = end_address - (unsigned long) vmem_map + sizeof(struct page) - 1;
430 hole_next_pfn = end_address / sizeof(struct page);
431 return hole_next_pfn - pgdat->node_start_pfn;
434 int __init create_mem_map_page_table(u64 start, u64 end, void *arg)
436 unsigned long address, start_page, end_page;
437 struct page *map_start, *map_end;
444 map_start = vmem_map + (__pa(start) >> PAGE_SHIFT);
445 map_end = vmem_map + (__pa(end) >> PAGE_SHIFT);
447 start_page = (unsigned long) map_start & PAGE_MASK;
448 end_page = PAGE_ALIGN((unsigned long) map_end);
449 node = paddr_to_nid(__pa(start));
451 for (address = start_page; address < end_page; address += PAGE_SIZE) {
452 pgd = pgd_offset_k(address);
454 pgd_populate(&init_mm, pgd, alloc_bootmem_pages_node(NODE_DATA(node), PAGE_SIZE));
455 pud = pud_offset(pgd, address);
458 pud_populate(&init_mm, pud, alloc_bootmem_pages_node(NODE_DATA(node), PAGE_SIZE));
459 pmd = pmd_offset(pud, address);
462 pmd_populate_kernel(&init_mm, pmd, alloc_bootmem_pages_node(NODE_DATA(node), PAGE_SIZE));
463 pte = pte_offset_kernel(pmd, address);
466 set_pte(pte, pfn_pte(__pa(alloc_bootmem_pages_node(NODE_DATA(node), PAGE_SIZE)) >> PAGE_SHIFT,
472 struct memmap_init_callback_data {
480 virtual_memmap_init(u64 start, u64 end, void *arg)
482 struct memmap_init_callback_data *args;
483 struct page *map_start, *map_end;
485 args = (struct memmap_init_callback_data *) arg;
486 map_start = vmem_map + (__pa(start) >> PAGE_SHIFT);
487 map_end = vmem_map + (__pa(end) >> PAGE_SHIFT);
489 if (map_start < args->start)
490 map_start = args->start;
491 if (map_end > args->end)
495 * We have to initialize "out of bounds" struct page elements that fit completely
496 * on the same pages that were allocated for the "in bounds" elements because they
497 * may be referenced later (and found to be "reserved").
499 map_start -= ((unsigned long) map_start & (PAGE_SIZE - 1)) / sizeof(struct page);
500 map_end += ((PAGE_ALIGN((unsigned long) map_end) - (unsigned long) map_end)
501 / sizeof(struct page));
503 if (map_start < map_end)
504 memmap_init_zone((unsigned long)(map_end - map_start),
505 args->nid, args->zone, page_to_pfn(map_start),
511 memmap_init (unsigned long size, int nid, unsigned long zone,
512 unsigned long start_pfn)
515 memmap_init_zone(size, nid, zone, start_pfn, MEMMAP_EARLY);
518 struct memmap_init_callback_data args;
520 start = pfn_to_page(start_pfn);
522 args.end = start + size;
526 efi_memmap_walk(virtual_memmap_init, &args);
531 ia64_pfn_valid (unsigned long pfn)
534 struct page *pg = pfn_to_page(pfn);
536 return (__get_user(byte, (char __user *) pg) == 0)
537 && ((((u64)pg & PAGE_MASK) == (((u64)(pg + 1) - 1) & PAGE_MASK))
538 || (__get_user(byte, (char __user *) (pg + 1) - 1) == 0));
540 EXPORT_SYMBOL(ia64_pfn_valid);
542 int __init find_largest_hole(u64 start, u64 end, void *arg)
546 static u64 last_end = PAGE_OFFSET;
548 /* NOTE: this algorithm assumes efi memmap table is ordered */
550 if (*max_gap < (start - last_end))
551 *max_gap = start - last_end;
556 #endif /* CONFIG_VIRTUAL_MEM_MAP */
558 int __init register_active_ranges(u64 start, u64 len, int nid)
560 u64 end = start + len;
563 if (start > crashk_res.start && start < crashk_res.end)
564 start = crashk_res.end;
565 if (end > crashk_res.start && end < crashk_res.end)
566 end = crashk_res.start;
570 memblock_add_node(__pa(start), end - start, nid);
575 find_max_min_low_pfn (u64 start, u64 end, void *arg)
577 unsigned long pfn_start, pfn_end;
578 #ifdef CONFIG_FLATMEM
579 pfn_start = (PAGE_ALIGN(__pa(start))) >> PAGE_SHIFT;
580 pfn_end = (PAGE_ALIGN(__pa(end - 1))) >> PAGE_SHIFT;
582 pfn_start = GRANULEROUNDDOWN(__pa(start)) >> PAGE_SHIFT;
583 pfn_end = GRANULEROUNDUP(__pa(end - 1)) >> PAGE_SHIFT;
585 min_low_pfn = min(min_low_pfn, pfn_start);
586 max_low_pfn = max(max_low_pfn, pfn_end);
591 * Boot command-line option "nolwsys" can be used to disable the use of any light-weight
592 * system call handler. When this option is in effect, all fsyscalls will end up bubbling
593 * down into the kernel and calling the normal (heavy-weight) syscall handler. This is
594 * useful for performance testing, but conceivably could also come in handy for debugging
598 static int nolwsys __initdata;
601 nolwsys_setup (char *s)
607 __setup("nolwsys", nolwsys_setup);
614 BUG_ON(PTRS_PER_PGD * sizeof(pgd_t) != PAGE_SIZE);
615 BUG_ON(PTRS_PER_PMD * sizeof(pmd_t) != PAGE_SIZE);
616 BUG_ON(PTRS_PER_PTE * sizeof(pte_t) != PAGE_SIZE);
620 * This needs to be called _after_ the command line has been parsed but _before_
621 * any drivers that may need the PCI DMA interface are initialized or bootmem has
627 #ifdef CONFIG_FLATMEM
631 set_max_mapnr(max_low_pfn);
632 high_memory = __va(max_low_pfn * PAGE_SIZE);
634 mem_init_print_info(NULL);
637 * For fsyscall entrpoints with no light-weight handler, use the ordinary
638 * (heavy-weight) handler, but mark it by setting bit 0, so the fsyscall entry
639 * code can tell them apart.
641 for (i = 0; i < NR_syscalls; ++i) {
642 extern unsigned long fsyscall_table[NR_syscalls];
643 extern unsigned long sys_call_table[NR_syscalls];
645 if (!fsyscall_table[i] || nolwsys)
646 fsyscall_table[i] = sys_call_table[i] | 1;
651 #ifdef CONFIG_MEMORY_HOTPLUG
652 int arch_add_memory(int nid, u64 start, u64 size)
656 unsigned long start_pfn = start >> PAGE_SHIFT;
657 unsigned long nr_pages = size >> PAGE_SHIFT;
660 pgdat = NODE_DATA(nid);
662 zone = pgdat->node_zones +
663 zone_for_memory(nid, start, size, ZONE_NORMAL);
664 ret = __add_pages(nid, zone, start_pfn, nr_pages);
667 printk("%s: Problem encountered in __add_pages() as ret=%d\n",
673 #ifdef CONFIG_MEMORY_HOTREMOVE
674 int arch_remove_memory(u64 start, u64 size)
676 unsigned long start_pfn = start >> PAGE_SHIFT;
677 unsigned long nr_pages = size >> PAGE_SHIFT;
681 zone = page_zone(pfn_to_page(start_pfn));
682 ret = __remove_pages(zone, start_pfn, nr_pages);
684 pr_warn("%s: Problem encountered in __remove_pages() as"
685 " ret=%d\n", __func__, ret);
693 * show_mem - give short summary of memory stats
695 * Shows a simple page count of reserved and used pages in the system.
696 * For discontig machines, it does this on a per-pgdat basis.
698 void show_mem(unsigned int filter)
700 int total_reserved = 0;
701 unsigned long total_present = 0;
704 printk(KERN_INFO "Mem-info:\n");
705 show_free_areas(filter);
706 printk(KERN_INFO "Node memory in pages:\n");
707 for_each_online_pgdat(pgdat) {
708 unsigned long present;
711 int nid = pgdat->node_id;
714 if (skip_free_areas_node(filter, nid))
716 pgdat_resize_lock(pgdat, &flags);
718 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
719 struct zone *zone = &pgdat->node_zones[zoneid];
720 if (!populated_zone(zone))
723 reserved += zone->present_pages - zone->managed_pages;
725 present = pgdat->node_present_pages;
727 pgdat_resize_unlock(pgdat, &flags);
728 total_present += present;
729 total_reserved += reserved;
730 printk(KERN_INFO "Node %4d: RAM: %11ld, rsvd: %8d, ",
731 nid, present, reserved);
733 printk(KERN_INFO "%ld pages of RAM\n", total_present);
734 printk(KERN_INFO "%d reserved pages\n", total_reserved);
735 printk(KERN_INFO "Total of %ld pages in page table cache\n",
736 quicklist_total_size());
737 printk(KERN_INFO "%ld free buffer pages\n", nr_free_buffer_pages());