2 * Initialize MMU support.
4 * Copyright (C) 1998-2003 Hewlett-Packard Co
5 * David Mosberger-Tang <davidm@hpl.hp.com>
7 #include <linux/config.h>
8 #include <linux/kernel.h>
9 #include <linux/init.h>
11 #include <linux/bootmem.h>
12 #include <linux/efi.h>
13 #include <linux/elf.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>
23 #include <asm/a.out.h>
24 #include <asm/bitops.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 extern void ia64_tlb_init (void);
44 unsigned long MAX_DMA_ADDRESS = PAGE_OFFSET + 0x100000000UL;
46 #ifdef CONFIG_VIRTUAL_MEM_MAP
47 unsigned long vmalloc_end = VMALLOC_END_INIT;
48 EXPORT_SYMBOL(vmalloc_end);
49 struct page *vmem_map;
50 EXPORT_SYMBOL(vmem_map);
53 static int pgt_cache_water[2] = { 25, 50 };
55 struct page *zero_page_memmap_ptr; /* map entry for zero page */
56 EXPORT_SYMBOL(zero_page_memmap_ptr);
59 check_pgt_cache (void)
63 low = pgt_cache_water[0];
64 high = pgt_cache_water[1];
67 if (pgtable_cache_size > (u64) high) {
70 free_page((unsigned long)pgd_alloc_one_fast(NULL));
72 free_page((unsigned long)pmd_alloc_one_fast(NULL, 0));
73 } while (pgtable_cache_size > (u64) low);
79 update_mmu_cache (struct vm_area_struct *vma, unsigned long vaddr, pte_t pte)
85 return; /* not an executable page... */
88 /* don't use VADDR: it may not be mapped on this CPU (or may have just been flushed): */
89 addr = (unsigned long) page_address(page);
91 if (test_bit(PG_arch_1, &page->flags))
92 return; /* i-cache is already coherent with d-cache */
94 flush_icache_range(addr, addr + PAGE_SIZE);
95 set_bit(PG_arch_1, &page->flags); /* mark page as clean */
99 ia64_set_rbs_bot (void)
101 unsigned long stack_size = current->rlim[RLIMIT_STACK].rlim_max & -16;
103 if (stack_size > MAX_USER_STACK_SIZE)
104 stack_size = MAX_USER_STACK_SIZE;
105 current->thread.rbs_bot = STACK_TOP - stack_size;
109 * This performs some platform-dependent address space initialization.
110 * On IA-64, we want to setup the VM area for the register backing
111 * store (which grows upwards) and install the gateway page which is
112 * used for signal trampolines, etc.
115 ia64_init_addr_space (void)
117 struct vm_area_struct *vma;
122 * If we're out of memory and kmem_cache_alloc() returns NULL, we simply ignore
123 * the problem. When the process attempts to write to the register backing store
124 * for the first time, it will get a SEGFAULT in this case.
126 vma = kmem_cache_alloc(vm_area_cachep, SLAB_KERNEL);
128 memset(vma, 0, sizeof(*vma));
129 vma->vm_mm = current->mm;
130 vma->vm_start = current->thread.rbs_bot & PAGE_MASK;
131 vma->vm_end = vma->vm_start + PAGE_SIZE;
132 vma->vm_page_prot = protection_map[VM_DATA_DEFAULT_FLAGS & 0x7];
133 vma->vm_flags = VM_DATA_DEFAULT_FLAGS | VM_GROWSUP;
134 insert_vm_struct(current->mm, vma);
137 /* map NaT-page at address zero to speed up speculative dereferencing of NULL: */
138 if (!(current->personality & MMAP_PAGE_ZERO)) {
139 vma = kmem_cache_alloc(vm_area_cachep, SLAB_KERNEL);
141 memset(vma, 0, sizeof(*vma));
142 vma->vm_mm = current->mm;
143 vma->vm_end = PAGE_SIZE;
144 vma->vm_page_prot = __pgprot(pgprot_val(PAGE_READONLY) | _PAGE_MA_NAT);
145 vma->vm_flags = VM_READ | VM_MAYREAD | VM_IO | VM_RESERVED;
146 insert_vm_struct(current->mm, vma);
154 unsigned long addr, eaddr;
156 addr = (unsigned long) ia64_imva(__init_begin);
157 eaddr = (unsigned long) ia64_imva(__init_end);
158 while (addr < eaddr) {
159 ClearPageReserved(virt_to_page(addr));
160 set_page_count(virt_to_page(addr), 1);
165 printk(KERN_INFO "Freeing unused kernel memory: %ldkB freed\n",
166 (__init_end - __init_begin) >> 10);
170 free_initrd_mem (unsigned long start, unsigned long end)
174 * EFI uses 4KB pages while the kernel can use 4KB or bigger.
175 * Thus EFI and the kernel may have different page sizes. It is
176 * therefore possible to have the initrd share the same page as
177 * the end of the kernel (given current setup).
179 * To avoid freeing/using the wrong page (kernel sized) we:
180 * - align up the beginning of initrd
181 * - align down the end of initrd
184 * |=============| a000
190 * |=============| 8000
193 * |/////////////| 7000
196 * |=============| 6000
199 * K=kernel using 8KB pages
201 * In this example, we must free page 8000 ONLY. So we must align up
202 * initrd_start and keep initrd_end as is.
204 start = PAGE_ALIGN(start);
205 end = end & PAGE_MASK;
208 printk(KERN_INFO "Freeing initrd memory: %ldkB freed\n", (end - start) >> 10);
210 for (; start < end; start += PAGE_SIZE) {
211 if (!virt_addr_valid(start))
213 page = virt_to_page(start);
214 ClearPageReserved(page);
215 set_page_count(page, 1);
222 * This installs a clean page in the kernel's page table.
225 put_kernel_page (struct page *page, unsigned long address, pgprot_t pgprot)
231 if (!PageReserved(page))
232 printk(KERN_ERR "put_kernel_page: page at 0x%p not in reserved memory\n",
235 pgd = pgd_offset_k(address); /* note: this is NOT pgd_offset()! */
237 spin_lock(&init_mm.page_table_lock);
239 pmd = pmd_alloc(&init_mm, pgd, address);
242 pte = pte_alloc_map(&init_mm, pmd, address);
245 if (!pte_none(*pte)) {
249 set_pte(pte, mk_pte(page, pgprot));
252 out: spin_unlock(&init_mm.page_table_lock);
253 /* no need for flush_tlb */
263 * Map the gate page twice: once read-only to export the ELF headers etc. and once
264 * execute-only page to enable privilege-promotion via "epc":
266 page = virt_to_page(ia64_imva(__start_gate_section));
267 put_kernel_page(page, GATE_ADDR, PAGE_READONLY);
268 #ifdef HAVE_BUGGY_SEGREL
269 page = virt_to_page(ia64_imva(__start_gate_section + PAGE_SIZE));
270 put_kernel_page(page, GATE_ADDR + PAGE_SIZE, PAGE_GATE);
272 put_kernel_page(page, GATE_ADDR + PERCPU_PAGE_SIZE, PAGE_GATE);
278 ia64_mmu_init (void *my_cpu_data)
280 unsigned long psr, pta, impl_va_bits;
281 extern void __devinit tlb_init (void);
284 #ifdef CONFIG_DISABLE_VHPT
285 # define VHPT_ENABLE_BIT 0
287 # define VHPT_ENABLE_BIT 1
290 /* Pin mapping for percpu area into TLB */
291 psr = ia64_clear_ic();
292 ia64_itr(0x2, IA64_TR_PERCPU_DATA, PERCPU_ADDR,
293 pte_val(pfn_pte(__pa(my_cpu_data) >> PAGE_SHIFT, PAGE_KERNEL)),
300 * Check if the virtually mapped linear page table (VMLPT) overlaps with a mapped
301 * address space. The IA-64 architecture guarantees that at least 50 bits of
302 * virtual address space are implemented but if we pick a large enough page size
303 * (e.g., 64KB), the mapped address space is big enough that it will overlap with
304 * VMLPT. I assume that once we run on machines big enough to warrant 64KB pages,
305 * IMPL_VA_MSB will be significantly bigger, so this is unlikely to become a
306 * problem in practice. Alternatively, we could truncate the top of the mapped
307 * address space to not permit mappings that would overlap with the VMLPT.
311 # define mapped_space_bits (3*(PAGE_SHIFT - pte_bits) + PAGE_SHIFT)
313 * The virtual page table has to cover the entire implemented address space within
314 * a region even though not all of this space may be mappable. The reason for
315 * this is that the Access bit and Dirty bit fault handlers perform
316 * non-speculative accesses to the virtual page table, so the address range of the
317 * virtual page table itself needs to be covered by virtual page table.
319 # define vmlpt_bits (impl_va_bits - PAGE_SHIFT + pte_bits)
320 # define POW2(n) (1ULL << (n))
322 impl_va_bits = ffz(~(local_cpu_data->unimpl_va_mask | (7UL << 61)));
324 if (impl_va_bits < 51 || impl_va_bits > 61)
325 panic("CPU has bogus IMPL_VA_MSB value of %lu!\n", impl_va_bits - 1);
327 /* place the VMLPT at the end of each page-table mapped region: */
328 pta = POW2(61) - POW2(vmlpt_bits);
330 if (POW2(mapped_space_bits) >= pta)
331 panic("mm/init: overlap between virtually mapped linear page table and "
332 "mapped kernel space!");
334 * Set the (virtually mapped linear) page table address. Bit
335 * 8 selects between the short and long format, bits 2-7 the
336 * size of the table, and bit 0 whether the VHPT walker is
339 ia64_set_pta(pta | (0 << 8) | (vmlpt_bits << 2) | VHPT_ENABLE_BIT);
343 #ifdef CONFIG_HUGETLB_PAGE
344 ia64_set_rr(HPAGE_REGION_BASE, HPAGE_SHIFT << 2);
348 cpu = smp_processor_id();
350 /* mca handler uses cr.lid as key to pick the right entry */
351 ia64_mca_tlb_list[cpu].cr_lid = ia64_getreg(_IA64_REG_CR_LID);
353 /* insert this percpu data information into our list for MCA recovery purposes */
354 ia64_mca_tlb_list[cpu].percpu_paddr = pte_val(mk_pte_phys(__pa(my_cpu_data), PAGE_KERNEL));
355 /* Also save per-cpu tlb flush recipe for use in physical mode mca handler */
356 ia64_mca_tlb_list[cpu].ptce_base = local_cpu_data->ptce_base;
357 ia64_mca_tlb_list[cpu].ptce_count[0] = local_cpu_data->ptce_count[0];
358 ia64_mca_tlb_list[cpu].ptce_count[1] = local_cpu_data->ptce_count[1];
359 ia64_mca_tlb_list[cpu].ptce_stride[0] = local_cpu_data->ptce_stride[0];
360 ia64_mca_tlb_list[cpu].ptce_stride[1] = local_cpu_data->ptce_stride[1];
363 #ifdef CONFIG_VIRTUAL_MEM_MAP
366 create_mem_map_page_table (u64 start, u64 end, void *arg)
368 unsigned long address, start_page, end_page;
369 struct page *map_start, *map_end;
375 map_start = vmem_map + (__pa(start) >> PAGE_SHIFT);
376 map_end = vmem_map + (__pa(end) >> PAGE_SHIFT);
378 start_page = (unsigned long) map_start & PAGE_MASK;
379 end_page = PAGE_ALIGN((unsigned long) map_end);
380 node = paddr_to_nid(__pa(start));
382 for (address = start_page; address < end_page; address += PAGE_SIZE) {
383 pgd = pgd_offset_k(address);
385 pgd_populate(&init_mm, pgd, alloc_bootmem_pages_node(NODE_DATA(node), PAGE_SIZE));
386 pmd = pmd_offset(pgd, address);
389 pmd_populate_kernel(&init_mm, pmd, alloc_bootmem_pages_node(NODE_DATA(node), PAGE_SIZE));
390 pte = pte_offset_kernel(pmd, address);
393 set_pte(pte, pfn_pte(__pa(alloc_bootmem_pages_node(NODE_DATA(node), PAGE_SIZE)) >> PAGE_SHIFT,
399 struct memmap_init_callback_data {
407 virtual_memmap_init (u64 start, u64 end, void *arg)
409 struct memmap_init_callback_data *args;
410 struct page *map_start, *map_end;
412 args = (struct memmap_init_callback_data *) arg;
414 map_start = vmem_map + (__pa(start) >> PAGE_SHIFT);
415 map_end = vmem_map + (__pa(end) >> PAGE_SHIFT);
417 if (map_start < args->start)
418 map_start = args->start;
419 if (map_end > args->end)
423 * We have to initialize "out of bounds" struct page elements that fit completely
424 * on the same pages that were allocated for the "in bounds" elements because they
425 * may be referenced later (and found to be "reserved").
427 map_start -= ((unsigned long) map_start & (PAGE_SIZE - 1)) / sizeof(struct page);
428 map_end += ((PAGE_ALIGN((unsigned long) map_end) - (unsigned long) map_end)
429 / sizeof(struct page));
431 if (map_start < map_end)
432 memmap_init_zone((unsigned long)(map_end - map_start),
433 args->nid, args->zone, page_to_pfn(map_start));
438 memmap_init (unsigned long size, int nid, unsigned long zone,
439 unsigned long start_pfn)
442 memmap_init_zone(size, nid, zone, start_pfn);
445 struct memmap_init_callback_data args;
447 start = pfn_to_page(start_pfn);
449 args.end = start + size;
453 efi_memmap_walk(virtual_memmap_init, &args);
458 ia64_pfn_valid (unsigned long pfn)
461 struct page *pg = pfn_to_page(pfn);
463 return (__get_user(byte, (char __user *) pg) == 0)
464 && ((((u64)pg & PAGE_MASK) == (((u64)(pg + 1) - 1) & PAGE_MASK))
465 || (__get_user(byte, (char __user *) (pg + 1) - 1) == 0));
467 EXPORT_SYMBOL(ia64_pfn_valid);
470 find_largest_hole (u64 start, u64 end, void *arg)
474 static u64 last_end = PAGE_OFFSET;
476 /* NOTE: this algorithm assumes efi memmap table is ordered */
478 if (*max_gap < (start - last_end))
479 *max_gap = start - last_end;
483 #endif /* CONFIG_VIRTUAL_MEM_MAP */
486 count_reserved_pages (u64 start, u64 end, void *arg)
488 unsigned long num_reserved = 0;
489 unsigned long *count = arg;
491 for (; start < end; start += PAGE_SIZE)
492 if (PageReserved(virt_to_page(start)))
494 *count += num_reserved;
499 * Boot command-line option "nolwsys" can be used to disable the use of any light-weight
500 * system call handler. When this option is in effect, all fsyscalls will end up bubbling
501 * down into the kernel and calling the normal (heavy-weight) syscall handler. This is
502 * useful for performance testing, but conceivably could also come in handy for debugging
509 nolwsys_setup (char *s)
515 __setup("nolwsys", nolwsys_setup);
520 long reserved_pages, codesize, datasize, initsize;
521 unsigned long num_pgt_pages;
524 static struct kcore_list kcore_mem, kcore_vmem, kcore_kernel;
528 * This needs to be called _after_ the command line has been parsed but _before_
529 * any drivers that may need the PCI DMA interface are initialized or bootmem has
535 #ifndef CONFIG_DISCONTIGMEM
538 max_mapnr = max_low_pfn;
541 high_memory = __va(max_low_pfn * PAGE_SIZE);
543 kclist_add(&kcore_mem, __va(0), max_low_pfn * PAGE_SIZE);
544 kclist_add(&kcore_vmem, (void *)VMALLOC_START, VMALLOC_END-VMALLOC_START);
545 kclist_add(&kcore_kernel, _stext, _end - _stext);
547 for_each_pgdat(pgdat)
548 totalram_pages += free_all_bootmem_node(pgdat);
551 efi_memmap_walk(count_reserved_pages, &reserved_pages);
553 codesize = (unsigned long) _etext - (unsigned long) _stext;
554 datasize = (unsigned long) _edata - (unsigned long) _etext;
555 initsize = (unsigned long) __init_end - (unsigned long) __init_begin;
557 printk(KERN_INFO "Memory: %luk/%luk available (%luk code, %luk reserved, "
558 "%luk data, %luk init)\n", (unsigned long) nr_free_pages() << (PAGE_SHIFT - 10),
559 num_physpages << (PAGE_SHIFT - 10), codesize >> 10,
560 reserved_pages << (PAGE_SHIFT - 10), datasize >> 10, initsize >> 10);
563 * Allow for enough (cached) page table pages so that we can map the entire memory
564 * at least once. Each task also needs a couple of page tables pages, so add in a
565 * fudge factor for that (don't use "threads-max" here; that would be wrong!).
566 * Don't allow the cache to be more than 10% of total memory, though.
568 # define NUM_TASKS 500 /* typical number of tasks */
569 num_pgt_pages = nr_free_pages() / PTRS_PER_PGD + NUM_TASKS;
570 if (num_pgt_pages > nr_free_pages() / 10)
571 num_pgt_pages = nr_free_pages() / 10;
572 if (num_pgt_pages > (u64) pgt_cache_water[1])
573 pgt_cache_water[1] = num_pgt_pages;
576 * For fsyscall entrpoints with no light-weight handler, use the ordinary
577 * (heavy-weight) handler, but mark it by setting bit 0, so the fsyscall entry
578 * code can tell them apart.
580 for (i = 0; i < NR_syscalls; ++i) {
581 extern unsigned long fsyscall_table[NR_syscalls];
582 extern unsigned long sys_call_table[NR_syscalls];
584 if (!fsyscall_table[i] || nolwsys)
585 fsyscall_table[i] = sys_call_table[i] | 1;
589 #ifdef CONFIG_IA32_SUPPORT