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>
14 #include <linux/mmzone.h>
15 #include <linux/module.h>
16 #include <linux/personality.h>
17 #include <linux/reboot.h>
18 #include <linux/slab.h>
19 #include <linux/swap.h>
20 #include <linux/proc_fs.h>
21 #include <linux/bitops.h>
22 #include <linux/kexec.h>
24 #include <asm/a.out.h>
28 #include <asm/machvec.h>
30 #include <asm/patch.h>
31 #include <asm/pgalloc.h>
33 #include <asm/sections.h>
34 #include <asm/system.h>
36 #include <asm/uaccess.h>
37 #include <asm/unistd.h>
40 DEFINE_PER_CPU(struct mmu_gather, mmu_gathers);
42 DEFINE_PER_CPU(unsigned long *, __pgtable_quicklist);
43 DEFINE_PER_CPU(long, __pgtable_quicklist_size);
45 extern void ia64_tlb_init (void);
47 unsigned long MAX_DMA_ADDRESS = PAGE_OFFSET + 0x100000000UL;
49 #ifdef CONFIG_VIRTUAL_MEM_MAP
50 unsigned long vmalloc_end = VMALLOC_END_INIT;
51 EXPORT_SYMBOL(vmalloc_end);
52 struct page *vmem_map;
53 EXPORT_SYMBOL(vmem_map);
56 struct page *zero_page_memmap_ptr; /* map entry for zero page */
57 EXPORT_SYMBOL(zero_page_memmap_ptr);
59 #define MIN_PGT_PAGES 25UL
60 #define MAX_PGT_FREES_PER_PASS 16L
61 #define PGT_FRACTION_OF_NODE_MEM 16
66 u64 node_free_pages, max_pgt_pages;
69 node_free_pages = nr_free_pages();
71 node_free_pages = nr_free_pages_pgdat(NODE_DATA(numa_node_id()));
73 max_pgt_pages = node_free_pages / PGT_FRACTION_OF_NODE_MEM;
74 max_pgt_pages = max(max_pgt_pages, MIN_PGT_PAGES);
79 min_pages_to_free(void)
83 pages_to_free = pgtable_quicklist_size - max_pgt_pages();
84 pages_to_free = min(pages_to_free, MAX_PGT_FREES_PER_PASS);
93 if (unlikely(pgtable_quicklist_size <= MIN_PGT_PAGES))
97 while (unlikely((pages_to_free = min_pages_to_free()) > 0)) {
98 while (pages_to_free--) {
99 free_page((unsigned long)pgtable_quicklist_alloc());
108 lazy_mmu_prot_update (pte_t pte)
115 return; /* not an executable page... */
117 page = pte_page(pte);
118 addr = (unsigned long) page_address(page);
120 if (test_bit(PG_arch_1, &page->flags))
121 return; /* i-cache is already coherent with d-cache */
123 if (PageCompound(page)) {
124 order = (unsigned long) (page[1].lru.prev);
125 flush_icache_range(addr, addr + (1UL << order << PAGE_SHIFT));
128 flush_icache_range(addr, addr + PAGE_SIZE);
129 set_bit(PG_arch_1, &page->flags); /* mark page as clean */
133 * Since DMA is i-cache coherent, any (complete) pages that were written via
134 * DMA can be marked as "clean" so that lazy_mmu_prot_update() doesn't have to
135 * flush them when they get mapped into an executable vm-area.
138 dma_mark_clean(void *addr, size_t size)
140 unsigned long pg_addr, end;
142 pg_addr = PAGE_ALIGN((unsigned long) addr);
143 end = (unsigned long) addr + size;
144 while (pg_addr + PAGE_SIZE <= end) {
145 struct page *page = virt_to_page(pg_addr);
146 set_bit(PG_arch_1, &page->flags);
147 pg_addr += PAGE_SIZE;
152 ia64_set_rbs_bot (void)
154 unsigned long stack_size = current->signal->rlim[RLIMIT_STACK].rlim_max & -16;
156 if (stack_size > MAX_USER_STACK_SIZE)
157 stack_size = MAX_USER_STACK_SIZE;
158 current->thread.rbs_bot = STACK_TOP - stack_size;
162 * This performs some platform-dependent address space initialization.
163 * On IA-64, we want to setup the VM area for the register backing
164 * store (which grows upwards) and install the gateway page which is
165 * used for signal trampolines, etc.
168 ia64_init_addr_space (void)
170 struct vm_area_struct *vma;
175 * If we're out of memory and kmem_cache_alloc() returns NULL, we simply ignore
176 * the problem. When the process attempts to write to the register backing store
177 * for the first time, it will get a SEGFAULT in this case.
179 vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
181 memset(vma, 0, sizeof(*vma));
182 vma->vm_mm = current->mm;
183 vma->vm_start = current->thread.rbs_bot & PAGE_MASK;
184 vma->vm_end = vma->vm_start + PAGE_SIZE;
185 vma->vm_page_prot = protection_map[VM_DATA_DEFAULT_FLAGS & 0x7];
186 vma->vm_flags = VM_DATA_DEFAULT_FLAGS|VM_GROWSUP|VM_ACCOUNT;
187 down_write(¤t->mm->mmap_sem);
188 if (insert_vm_struct(current->mm, vma)) {
189 up_write(¤t->mm->mmap_sem);
190 kmem_cache_free(vm_area_cachep, vma);
193 up_write(¤t->mm->mmap_sem);
196 /* map NaT-page at address zero to speed up speculative dereferencing of NULL: */
197 if (!(current->personality & MMAP_PAGE_ZERO)) {
198 vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
200 memset(vma, 0, sizeof(*vma));
201 vma->vm_mm = current->mm;
202 vma->vm_end = PAGE_SIZE;
203 vma->vm_page_prot = __pgprot(pgprot_val(PAGE_READONLY) | _PAGE_MA_NAT);
204 vma->vm_flags = VM_READ | VM_MAYREAD | VM_IO | VM_RESERVED;
205 down_write(¤t->mm->mmap_sem);
206 if (insert_vm_struct(current->mm, vma)) {
207 up_write(¤t->mm->mmap_sem);
208 kmem_cache_free(vm_area_cachep, vma);
211 up_write(¤t->mm->mmap_sem);
219 unsigned long addr, eaddr;
221 addr = (unsigned long) ia64_imva(__init_begin);
222 eaddr = (unsigned long) ia64_imva(__init_end);
223 while (addr < eaddr) {
224 ClearPageReserved(virt_to_page(addr));
225 init_page_count(virt_to_page(addr));
230 printk(KERN_INFO "Freeing unused kernel memory: %ldkB freed\n",
231 (__init_end - __init_begin) >> 10);
235 free_initrd_mem (unsigned long start, unsigned long end)
239 * EFI uses 4KB pages while the kernel can use 4KB or bigger.
240 * Thus EFI and the kernel may have different page sizes. It is
241 * therefore possible to have the initrd share the same page as
242 * the end of the kernel (given current setup).
244 * To avoid freeing/using the wrong page (kernel sized) we:
245 * - align up the beginning of initrd
246 * - align down the end of initrd
249 * |=============| a000
255 * |=============| 8000
258 * |/////////////| 7000
261 * |=============| 6000
264 * K=kernel using 8KB pages
266 * In this example, we must free page 8000 ONLY. So we must align up
267 * initrd_start and keep initrd_end as is.
269 start = PAGE_ALIGN(start);
270 end = end & PAGE_MASK;
273 printk(KERN_INFO "Freeing initrd memory: %ldkB freed\n", (end - start) >> 10);
275 for (; start < end; start += PAGE_SIZE) {
276 if (!virt_addr_valid(start))
278 page = virt_to_page(start);
279 ClearPageReserved(page);
280 init_page_count(page);
287 * This installs a clean page in the kernel's page table.
289 static struct page * __init
290 put_kernel_page (struct page *page, unsigned long address, pgprot_t pgprot)
297 if (!PageReserved(page))
298 printk(KERN_ERR "put_kernel_page: page at 0x%p not in reserved memory\n",
301 pgd = pgd_offset_k(address); /* note: this is NOT pgd_offset()! */
304 pud = pud_alloc(&init_mm, pgd, address);
307 pmd = pmd_alloc(&init_mm, pud, address);
310 pte = pte_alloc_kernel(pmd, address);
315 set_pte(pte, mk_pte(page, pgprot));
318 /* no need for flush_tlb */
328 * Map the gate page twice: once read-only to export the ELF
329 * headers etc. and once execute-only page to enable
330 * privilege-promotion via "epc":
332 page = virt_to_page(ia64_imva(__start_gate_section));
333 put_kernel_page(page, GATE_ADDR, PAGE_READONLY);
334 #ifdef HAVE_BUGGY_SEGREL
335 page = virt_to_page(ia64_imva(__start_gate_section + PAGE_SIZE));
336 put_kernel_page(page, GATE_ADDR + PAGE_SIZE, PAGE_GATE);
338 put_kernel_page(page, GATE_ADDR + PERCPU_PAGE_SIZE, PAGE_GATE);
339 /* Fill in the holes (if any) with read-only zero pages: */
343 for (addr = GATE_ADDR + PAGE_SIZE;
344 addr < GATE_ADDR + PERCPU_PAGE_SIZE;
347 put_kernel_page(ZERO_PAGE(0), addr,
349 put_kernel_page(ZERO_PAGE(0), addr + PERCPU_PAGE_SIZE,
358 ia64_mmu_init (void *my_cpu_data)
360 unsigned long psr, pta, impl_va_bits;
361 extern void __devinit tlb_init (void);
363 #ifdef CONFIG_DISABLE_VHPT
364 # define VHPT_ENABLE_BIT 0
366 # define VHPT_ENABLE_BIT 1
369 /* Pin mapping for percpu area into TLB */
370 psr = ia64_clear_ic();
371 ia64_itr(0x2, IA64_TR_PERCPU_DATA, PERCPU_ADDR,
372 pte_val(pfn_pte(__pa(my_cpu_data) >> PAGE_SHIFT, PAGE_KERNEL)),
379 * Check if the virtually mapped linear page table (VMLPT) overlaps with a mapped
380 * address space. The IA-64 architecture guarantees that at least 50 bits of
381 * virtual address space are implemented but if we pick a large enough page size
382 * (e.g., 64KB), the mapped address space is big enough that it will overlap with
383 * VMLPT. I assume that once we run on machines big enough to warrant 64KB pages,
384 * IMPL_VA_MSB will be significantly bigger, so this is unlikely to become a
385 * problem in practice. Alternatively, we could truncate the top of the mapped
386 * address space to not permit mappings that would overlap with the VMLPT.
390 # define mapped_space_bits (3*(PAGE_SHIFT - pte_bits) + PAGE_SHIFT)
392 * The virtual page table has to cover the entire implemented address space within
393 * a region even though not all of this space may be mappable. The reason for
394 * this is that the Access bit and Dirty bit fault handlers perform
395 * non-speculative accesses to the virtual page table, so the address range of the
396 * virtual page table itself needs to be covered by virtual page table.
398 # define vmlpt_bits (impl_va_bits - PAGE_SHIFT + pte_bits)
399 # define POW2(n) (1ULL << (n))
401 impl_va_bits = ffz(~(local_cpu_data->unimpl_va_mask | (7UL << 61)));
403 if (impl_va_bits < 51 || impl_va_bits > 61)
404 panic("CPU has bogus IMPL_VA_MSB value of %lu!\n", impl_va_bits - 1);
406 * mapped_space_bits - PAGE_SHIFT is the total number of ptes we need,
407 * which must fit into "vmlpt_bits - pte_bits" slots. Second half of
408 * the test makes sure that our mapped space doesn't overlap the
409 * unimplemented hole in the middle of the region.
411 if ((mapped_space_bits - PAGE_SHIFT > vmlpt_bits - pte_bits) ||
412 (mapped_space_bits > impl_va_bits - 1))
413 panic("Cannot build a big enough virtual-linear page table"
414 " to cover mapped address space.\n"
415 " Try using a smaller page size.\n");
418 /* place the VMLPT at the end of each page-table mapped region: */
419 pta = POW2(61) - POW2(vmlpt_bits);
422 * Set the (virtually mapped linear) page table address. Bit
423 * 8 selects between the short and long format, bits 2-7 the
424 * size of the table, and bit 0 whether the VHPT walker is
427 ia64_set_pta(pta | (0 << 8) | (vmlpt_bits << 2) | VHPT_ENABLE_BIT);
431 #ifdef CONFIG_HUGETLB_PAGE
432 ia64_set_rr(HPAGE_REGION_BASE, HPAGE_SHIFT << 2);
437 #ifdef CONFIG_VIRTUAL_MEM_MAP
438 int vmemmap_find_next_valid_pfn(int node, int i)
440 unsigned long end_address, hole_next_pfn;
441 unsigned long stop_address;
442 pg_data_t *pgdat = NODE_DATA(node);
444 end_address = (unsigned long) &vmem_map[pgdat->node_start_pfn + i];
445 end_address = PAGE_ALIGN(end_address);
447 stop_address = (unsigned long) &vmem_map[
448 pgdat->node_start_pfn + pgdat->node_spanned_pages];
456 pgd = pgd_offset_k(end_address);
457 if (pgd_none(*pgd)) {
458 end_address += PGDIR_SIZE;
462 pud = pud_offset(pgd, end_address);
463 if (pud_none(*pud)) {
464 end_address += PUD_SIZE;
468 pmd = pmd_offset(pud, end_address);
469 if (pmd_none(*pmd)) {
470 end_address += PMD_SIZE;
474 pte = pte_offset_kernel(pmd, end_address);
476 if (pte_none(*pte)) {
477 end_address += PAGE_SIZE;
479 if ((end_address < stop_address) &&
480 (end_address != ALIGN(end_address, 1UL << PMD_SHIFT)))
484 /* Found next valid vmem_map page */
486 } while (end_address < stop_address);
488 end_address = min(end_address, stop_address);
489 end_address = end_address - (unsigned long) vmem_map + sizeof(struct page) - 1;
490 hole_next_pfn = end_address / sizeof(struct page);
491 return hole_next_pfn - pgdat->node_start_pfn;
495 create_mem_map_page_table (u64 start, u64 end, void *arg)
497 unsigned long address, start_page, end_page;
498 struct page *map_start, *map_end;
505 map_start = vmem_map + (__pa(start) >> PAGE_SHIFT);
506 map_end = vmem_map + (__pa(end) >> PAGE_SHIFT);
508 start_page = (unsigned long) map_start & PAGE_MASK;
509 end_page = PAGE_ALIGN((unsigned long) map_end);
510 node = paddr_to_nid(__pa(start));
512 for (address = start_page; address < end_page; address += PAGE_SIZE) {
513 pgd = pgd_offset_k(address);
515 pgd_populate(&init_mm, pgd, alloc_bootmem_pages_node(NODE_DATA(node), PAGE_SIZE));
516 pud = pud_offset(pgd, address);
519 pud_populate(&init_mm, pud, alloc_bootmem_pages_node(NODE_DATA(node), PAGE_SIZE));
520 pmd = pmd_offset(pud, address);
523 pmd_populate_kernel(&init_mm, pmd, alloc_bootmem_pages_node(NODE_DATA(node), PAGE_SIZE));
524 pte = pte_offset_kernel(pmd, address);
527 set_pte(pte, pfn_pte(__pa(alloc_bootmem_pages_node(NODE_DATA(node), PAGE_SIZE)) >> PAGE_SHIFT,
533 struct memmap_init_callback_data {
541 virtual_memmap_init (u64 start, u64 end, void *arg)
543 struct memmap_init_callback_data *args;
544 struct page *map_start, *map_end;
546 args = (struct memmap_init_callback_data *) arg;
547 map_start = vmem_map + (__pa(start) >> PAGE_SHIFT);
548 map_end = vmem_map + (__pa(end) >> PAGE_SHIFT);
550 if (map_start < args->start)
551 map_start = args->start;
552 if (map_end > args->end)
556 * We have to initialize "out of bounds" struct page elements that fit completely
557 * on the same pages that were allocated for the "in bounds" elements because they
558 * may be referenced later (and found to be "reserved").
560 map_start -= ((unsigned long) map_start & (PAGE_SIZE - 1)) / sizeof(struct page);
561 map_end += ((PAGE_ALIGN((unsigned long) map_end) - (unsigned long) map_end)
562 / sizeof(struct page));
564 if (map_start < map_end)
565 memmap_init_zone((unsigned long)(map_end - map_start),
566 args->nid, args->zone, page_to_pfn(map_start),
572 memmap_init (unsigned long size, int nid, unsigned long zone,
573 unsigned long start_pfn)
576 memmap_init_zone(size, nid, zone, start_pfn, MEMMAP_EARLY);
579 struct memmap_init_callback_data args;
581 start = pfn_to_page(start_pfn);
583 args.end = start + size;
587 efi_memmap_walk(virtual_memmap_init, &args);
592 ia64_pfn_valid (unsigned long pfn)
595 struct page *pg = pfn_to_page(pfn);
597 return (__get_user(byte, (char __user *) pg) == 0)
598 && ((((u64)pg & PAGE_MASK) == (((u64)(pg + 1) - 1) & PAGE_MASK))
599 || (__get_user(byte, (char __user *) (pg + 1) - 1) == 0));
601 EXPORT_SYMBOL(ia64_pfn_valid);
604 find_largest_hole (u64 start, u64 end, void *arg)
608 static u64 last_end = PAGE_OFFSET;
610 /* NOTE: this algorithm assumes efi memmap table is ordered */
612 if (*max_gap < (start - last_end))
613 *max_gap = start - last_end;
618 #endif /* CONFIG_VIRTUAL_MEM_MAP */
621 register_active_ranges(u64 start, u64 end, void *arg)
623 int nid = paddr_to_nid(__pa(start));
628 if (start > crashk_res.start && start < crashk_res.end)
629 start = crashk_res.end;
630 if (end > crashk_res.start && end < crashk_res.end)
631 end = crashk_res.start;
635 add_active_range(nid, __pa(start) >> PAGE_SHIFT,
636 __pa(end) >> PAGE_SHIFT);
641 count_reserved_pages (u64 start, u64 end, void *arg)
643 unsigned long num_reserved = 0;
644 unsigned long *count = arg;
646 for (; start < end; start += PAGE_SIZE)
647 if (PageReserved(virt_to_page(start)))
649 *count += num_reserved;
654 * Boot command-line option "nolwsys" can be used to disable the use of any light-weight
655 * system call handler. When this option is in effect, all fsyscalls will end up bubbling
656 * down into the kernel and calling the normal (heavy-weight) syscall handler. This is
657 * useful for performance testing, but conceivably could also come in handy for debugging
661 static int nolwsys __initdata;
664 nolwsys_setup (char *s)
670 __setup("nolwsys", nolwsys_setup);
675 long reserved_pages, codesize, datasize, initsize;
678 static struct kcore_list kcore_mem, kcore_vmem, kcore_kernel;
680 BUG_ON(PTRS_PER_PGD * sizeof(pgd_t) != PAGE_SIZE);
681 BUG_ON(PTRS_PER_PMD * sizeof(pmd_t) != PAGE_SIZE);
682 BUG_ON(PTRS_PER_PTE * sizeof(pte_t) != PAGE_SIZE);
686 * This needs to be called _after_ the command line has been parsed but _before_
687 * any drivers that may need the PCI DMA interface are initialized or bootmem has
693 #ifdef CONFIG_FLATMEM
696 max_mapnr = max_low_pfn;
699 high_memory = __va(max_low_pfn * PAGE_SIZE);
701 kclist_add(&kcore_mem, __va(0), max_low_pfn * PAGE_SIZE);
702 kclist_add(&kcore_vmem, (void *)VMALLOC_START, VMALLOC_END-VMALLOC_START);
703 kclist_add(&kcore_kernel, _stext, _end - _stext);
705 for_each_online_pgdat(pgdat)
706 if (pgdat->bdata->node_bootmem_map)
707 totalram_pages += free_all_bootmem_node(pgdat);
710 efi_memmap_walk(count_reserved_pages, &reserved_pages);
712 codesize = (unsigned long) _etext - (unsigned long) _stext;
713 datasize = (unsigned long) _edata - (unsigned long) _etext;
714 initsize = (unsigned long) __init_end - (unsigned long) __init_begin;
716 printk(KERN_INFO "Memory: %luk/%luk available (%luk code, %luk reserved, "
717 "%luk data, %luk init)\n", (unsigned long) nr_free_pages() << (PAGE_SHIFT - 10),
718 num_physpages << (PAGE_SHIFT - 10), codesize >> 10,
719 reserved_pages << (PAGE_SHIFT - 10), datasize >> 10, initsize >> 10);
723 * For fsyscall entrpoints with no light-weight handler, use the ordinary
724 * (heavy-weight) handler, but mark it by setting bit 0, so the fsyscall entry
725 * code can tell them apart.
727 for (i = 0; i < NR_syscalls; ++i) {
728 extern unsigned long fsyscall_table[NR_syscalls];
729 extern unsigned long sys_call_table[NR_syscalls];
731 if (!fsyscall_table[i] || nolwsys)
732 fsyscall_table[i] = sys_call_table[i] | 1;
736 #ifdef CONFIG_IA32_SUPPORT
741 #ifdef CONFIG_MEMORY_HOTPLUG
742 void online_page(struct page *page)
744 ClearPageReserved(page);
745 init_page_count(page);
751 int arch_add_memory(int nid, u64 start, u64 size)
755 unsigned long start_pfn = start >> PAGE_SHIFT;
756 unsigned long nr_pages = size >> PAGE_SHIFT;
759 pgdat = NODE_DATA(nid);
761 zone = pgdat->node_zones + ZONE_NORMAL;
762 ret = __add_pages(zone, start_pfn, nr_pages);
765 printk("%s: Problem encountered in __add_pages() as ret=%d\n",
771 int remove_memory(u64 start, u64 size)
775 EXPORT_SYMBOL_GPL(remove_memory);