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>
22 #include <linux/bitops.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)
114 return; /* not an executable page... */
116 page = pte_page(pte);
117 addr = (unsigned long) page_address(page);
119 if (test_bit(PG_arch_1, &page->flags))
120 return; /* i-cache is already coherent with d-cache */
122 flush_icache_range(addr, addr + PAGE_SIZE);
123 set_bit(PG_arch_1, &page->flags); /* mark page as clean */
127 ia64_set_rbs_bot (void)
129 unsigned long stack_size = current->signal->rlim[RLIMIT_STACK].rlim_max & -16;
131 if (stack_size > MAX_USER_STACK_SIZE)
132 stack_size = MAX_USER_STACK_SIZE;
133 current->thread.rbs_bot = STACK_TOP - stack_size;
137 * This performs some platform-dependent address space initialization.
138 * On IA-64, we want to setup the VM area for the register backing
139 * store (which grows upwards) and install the gateway page which is
140 * used for signal trampolines, etc.
143 ia64_init_addr_space (void)
145 struct vm_area_struct *vma;
150 * If we're out of memory and kmem_cache_alloc() returns NULL, we simply ignore
151 * the problem. When the process attempts to write to the register backing store
152 * for the first time, it will get a SEGFAULT in this case.
154 vma = kmem_cache_alloc(vm_area_cachep, SLAB_KERNEL);
156 memset(vma, 0, sizeof(*vma));
157 vma->vm_mm = current->mm;
158 vma->vm_start = current->thread.rbs_bot & PAGE_MASK;
159 vma->vm_end = vma->vm_start + PAGE_SIZE;
160 vma->vm_page_prot = protection_map[VM_DATA_DEFAULT_FLAGS & 0x7];
161 vma->vm_flags = VM_DATA_DEFAULT_FLAGS|VM_GROWSUP|VM_ACCOUNT;
162 down_write(¤t->mm->mmap_sem);
163 if (insert_vm_struct(current->mm, vma)) {
164 up_write(¤t->mm->mmap_sem);
165 kmem_cache_free(vm_area_cachep, vma);
168 up_write(¤t->mm->mmap_sem);
171 /* map NaT-page at address zero to speed up speculative dereferencing of NULL: */
172 if (!(current->personality & MMAP_PAGE_ZERO)) {
173 vma = kmem_cache_alloc(vm_area_cachep, SLAB_KERNEL);
175 memset(vma, 0, sizeof(*vma));
176 vma->vm_mm = current->mm;
177 vma->vm_end = PAGE_SIZE;
178 vma->vm_page_prot = __pgprot(pgprot_val(PAGE_READONLY) | _PAGE_MA_NAT);
179 vma->vm_flags = VM_READ | VM_MAYREAD | VM_IO | VM_RESERVED;
180 down_write(¤t->mm->mmap_sem);
181 if (insert_vm_struct(current->mm, vma)) {
182 up_write(¤t->mm->mmap_sem);
183 kmem_cache_free(vm_area_cachep, vma);
186 up_write(¤t->mm->mmap_sem);
194 unsigned long addr, eaddr;
196 addr = (unsigned long) ia64_imva(__init_begin);
197 eaddr = (unsigned long) ia64_imva(__init_end);
198 while (addr < eaddr) {
199 ClearPageReserved(virt_to_page(addr));
200 init_page_count(virt_to_page(addr));
205 printk(KERN_INFO "Freeing unused kernel memory: %ldkB freed\n",
206 (__init_end - __init_begin) >> 10);
210 free_initrd_mem (unsigned long start, unsigned long end)
214 * EFI uses 4KB pages while the kernel can use 4KB or bigger.
215 * Thus EFI and the kernel may have different page sizes. It is
216 * therefore possible to have the initrd share the same page as
217 * the end of the kernel (given current setup).
219 * To avoid freeing/using the wrong page (kernel sized) we:
220 * - align up the beginning of initrd
221 * - align down the end of initrd
224 * |=============| a000
230 * |=============| 8000
233 * |/////////////| 7000
236 * |=============| 6000
239 * K=kernel using 8KB pages
241 * In this example, we must free page 8000 ONLY. So we must align up
242 * initrd_start and keep initrd_end as is.
244 start = PAGE_ALIGN(start);
245 end = end & PAGE_MASK;
248 printk(KERN_INFO "Freeing initrd memory: %ldkB freed\n", (end - start) >> 10);
250 for (; start < end; start += PAGE_SIZE) {
251 if (!virt_addr_valid(start))
253 page = virt_to_page(start);
254 ClearPageReserved(page);
255 init_page_count(page);
262 * This installs a clean page in the kernel's page table.
264 static struct page * __init
265 put_kernel_page (struct page *page, unsigned long address, pgprot_t pgprot)
272 if (!PageReserved(page))
273 printk(KERN_ERR "put_kernel_page: page at 0x%p not in reserved memory\n",
276 pgd = pgd_offset_k(address); /* note: this is NOT pgd_offset()! */
279 pud = pud_alloc(&init_mm, pgd, address);
282 pmd = pmd_alloc(&init_mm, pud, address);
285 pte = pte_alloc_kernel(pmd, address);
290 set_pte(pte, mk_pte(page, pgprot));
293 /* no need for flush_tlb */
303 * Map the gate page twice: once read-only to export the ELF
304 * headers etc. and once execute-only page to enable
305 * privilege-promotion via "epc":
307 page = virt_to_page(ia64_imva(__start_gate_section));
308 put_kernel_page(page, GATE_ADDR, PAGE_READONLY);
309 #ifdef HAVE_BUGGY_SEGREL
310 page = virt_to_page(ia64_imva(__start_gate_section + PAGE_SIZE));
311 put_kernel_page(page, GATE_ADDR + PAGE_SIZE, PAGE_GATE);
313 put_kernel_page(page, GATE_ADDR + PERCPU_PAGE_SIZE, PAGE_GATE);
314 /* Fill in the holes (if any) with read-only zero pages: */
318 for (addr = GATE_ADDR + PAGE_SIZE;
319 addr < GATE_ADDR + PERCPU_PAGE_SIZE;
322 put_kernel_page(ZERO_PAGE(0), addr,
324 put_kernel_page(ZERO_PAGE(0), addr + PERCPU_PAGE_SIZE,
333 ia64_mmu_init (void *my_cpu_data)
335 unsigned long psr, pta, impl_va_bits;
336 extern void __devinit tlb_init (void);
338 #ifdef CONFIG_DISABLE_VHPT
339 # define VHPT_ENABLE_BIT 0
341 # define VHPT_ENABLE_BIT 1
344 /* Pin mapping for percpu area into TLB */
345 psr = ia64_clear_ic();
346 ia64_itr(0x2, IA64_TR_PERCPU_DATA, PERCPU_ADDR,
347 pte_val(pfn_pte(__pa(my_cpu_data) >> PAGE_SHIFT, PAGE_KERNEL)),
354 * Check if the virtually mapped linear page table (VMLPT) overlaps with a mapped
355 * address space. The IA-64 architecture guarantees that at least 50 bits of
356 * virtual address space are implemented but if we pick a large enough page size
357 * (e.g., 64KB), the mapped address space is big enough that it will overlap with
358 * VMLPT. I assume that once we run on machines big enough to warrant 64KB pages,
359 * IMPL_VA_MSB will be significantly bigger, so this is unlikely to become a
360 * problem in practice. Alternatively, we could truncate the top of the mapped
361 * address space to not permit mappings that would overlap with the VMLPT.
365 # define mapped_space_bits (3*(PAGE_SHIFT - pte_bits) + PAGE_SHIFT)
367 * The virtual page table has to cover the entire implemented address space within
368 * a region even though not all of this space may be mappable. The reason for
369 * this is that the Access bit and Dirty bit fault handlers perform
370 * non-speculative accesses to the virtual page table, so the address range of the
371 * virtual page table itself needs to be covered by virtual page table.
373 # define vmlpt_bits (impl_va_bits - PAGE_SHIFT + pte_bits)
374 # define POW2(n) (1ULL << (n))
376 impl_va_bits = ffz(~(local_cpu_data->unimpl_va_mask | (7UL << 61)));
378 if (impl_va_bits < 51 || impl_va_bits > 61)
379 panic("CPU has bogus IMPL_VA_MSB value of %lu!\n", impl_va_bits - 1);
381 * mapped_space_bits - PAGE_SHIFT is the total number of ptes we need,
382 * which must fit into "vmlpt_bits - pte_bits" slots. Second half of
383 * the test makes sure that our mapped space doesn't overlap the
384 * unimplemented hole in the middle of the region.
386 if ((mapped_space_bits - PAGE_SHIFT > vmlpt_bits - pte_bits) ||
387 (mapped_space_bits > impl_va_bits - 1))
388 panic("Cannot build a big enough virtual-linear page table"
389 " to cover mapped address space.\n"
390 " Try using a smaller page size.\n");
393 /* place the VMLPT at the end of each page-table mapped region: */
394 pta = POW2(61) - POW2(vmlpt_bits);
397 * Set the (virtually mapped linear) page table address. Bit
398 * 8 selects between the short and long format, bits 2-7 the
399 * size of the table, and bit 0 whether the VHPT walker is
402 ia64_set_pta(pta | (0 << 8) | (vmlpt_bits << 2) | VHPT_ENABLE_BIT);
406 #ifdef CONFIG_HUGETLB_PAGE
407 ia64_set_rr(HPAGE_REGION_BASE, HPAGE_SHIFT << 2);
412 #ifdef CONFIG_VIRTUAL_MEM_MAP
415 create_mem_map_page_table (u64 start, u64 end, void *arg)
417 unsigned long address, start_page, end_page;
418 struct page *map_start, *map_end;
425 map_start = vmem_map + (__pa(start) >> PAGE_SHIFT);
426 map_end = vmem_map + (__pa(end) >> PAGE_SHIFT);
428 start_page = (unsigned long) map_start & PAGE_MASK;
429 end_page = PAGE_ALIGN((unsigned long) map_end);
430 node = paddr_to_nid(__pa(start));
432 for (address = start_page; address < end_page; address += PAGE_SIZE) {
433 pgd = pgd_offset_k(address);
435 pgd_populate(&init_mm, pgd, alloc_bootmem_pages_node(NODE_DATA(node), PAGE_SIZE));
436 pud = pud_offset(pgd, address);
439 pud_populate(&init_mm, pud, alloc_bootmem_pages_node(NODE_DATA(node), PAGE_SIZE));
440 pmd = pmd_offset(pud, address);
443 pmd_populate_kernel(&init_mm, pmd, alloc_bootmem_pages_node(NODE_DATA(node), PAGE_SIZE));
444 pte = pte_offset_kernel(pmd, address);
447 set_pte(pte, pfn_pte(__pa(alloc_bootmem_pages_node(NODE_DATA(node), PAGE_SIZE)) >> PAGE_SHIFT,
453 struct memmap_init_callback_data {
461 virtual_memmap_init (u64 start, u64 end, void *arg)
463 struct memmap_init_callback_data *args;
464 struct page *map_start, *map_end;
466 args = (struct memmap_init_callback_data *) arg;
467 map_start = vmem_map + (__pa(start) >> PAGE_SHIFT);
468 map_end = vmem_map + (__pa(end) >> PAGE_SHIFT);
470 if (map_start < args->start)
471 map_start = args->start;
472 if (map_end > args->end)
476 * We have to initialize "out of bounds" struct page elements that fit completely
477 * on the same pages that were allocated for the "in bounds" elements because they
478 * may be referenced later (and found to be "reserved").
480 map_start -= ((unsigned long) map_start & (PAGE_SIZE - 1)) / sizeof(struct page);
481 map_end += ((PAGE_ALIGN((unsigned long) map_end) - (unsigned long) map_end)
482 / sizeof(struct page));
484 if (map_start < map_end)
485 memmap_init_zone((unsigned long)(map_end - map_start),
486 args->nid, args->zone, page_to_pfn(map_start));
491 memmap_init (unsigned long size, int nid, unsigned long zone,
492 unsigned long start_pfn)
495 memmap_init_zone(size, nid, zone, start_pfn);
498 struct memmap_init_callback_data args;
500 start = pfn_to_page(start_pfn);
502 args.end = start + size;
506 efi_memmap_walk(virtual_memmap_init, &args);
511 ia64_pfn_valid (unsigned long pfn)
514 struct page *pg = pfn_to_page(pfn);
516 return (__get_user(byte, (char __user *) pg) == 0)
517 && ((((u64)pg & PAGE_MASK) == (((u64)(pg + 1) - 1) & PAGE_MASK))
518 || (__get_user(byte, (char __user *) (pg + 1) - 1) == 0));
520 EXPORT_SYMBOL(ia64_pfn_valid);
523 find_largest_hole (u64 start, u64 end, void *arg)
527 static u64 last_end = PAGE_OFFSET;
529 /* NOTE: this algorithm assumes efi memmap table is ordered */
531 if (*max_gap < (start - last_end))
532 *max_gap = start - last_end;
536 #endif /* CONFIG_VIRTUAL_MEM_MAP */
539 count_reserved_pages (u64 start, u64 end, void *arg)
541 unsigned long num_reserved = 0;
542 unsigned long *count = arg;
544 for (; start < end; start += PAGE_SIZE)
545 if (PageReserved(virt_to_page(start)))
547 *count += num_reserved;
552 * Boot command-line option "nolwsys" can be used to disable the use of any light-weight
553 * system call handler. When this option is in effect, all fsyscalls will end up bubbling
554 * down into the kernel and calling the normal (heavy-weight) syscall handler. This is
555 * useful for performance testing, but conceivably could also come in handy for debugging
559 static int nolwsys __initdata;
562 nolwsys_setup (char *s)
568 __setup("nolwsys", nolwsys_setup);
573 long reserved_pages, codesize, datasize, initsize;
576 static struct kcore_list kcore_mem, kcore_vmem, kcore_kernel;
578 BUG_ON(PTRS_PER_PGD * sizeof(pgd_t) != PAGE_SIZE);
579 BUG_ON(PTRS_PER_PMD * sizeof(pmd_t) != PAGE_SIZE);
580 BUG_ON(PTRS_PER_PTE * sizeof(pte_t) != PAGE_SIZE);
584 * This needs to be called _after_ the command line has been parsed but _before_
585 * any drivers that may need the PCI DMA interface are initialized or bootmem has
591 #ifdef CONFIG_FLATMEM
594 max_mapnr = max_low_pfn;
597 high_memory = __va(max_low_pfn * PAGE_SIZE);
599 kclist_add(&kcore_mem, __va(0), max_low_pfn * PAGE_SIZE);
600 kclist_add(&kcore_vmem, (void *)VMALLOC_START, VMALLOC_END-VMALLOC_START);
601 kclist_add(&kcore_kernel, _stext, _end - _stext);
603 for_each_pgdat(pgdat)
604 if (pgdat->bdata->node_bootmem_map)
605 totalram_pages += free_all_bootmem_node(pgdat);
608 efi_memmap_walk(count_reserved_pages, &reserved_pages);
610 codesize = (unsigned long) _etext - (unsigned long) _stext;
611 datasize = (unsigned long) _edata - (unsigned long) _etext;
612 initsize = (unsigned long) __init_end - (unsigned long) __init_begin;
614 printk(KERN_INFO "Memory: %luk/%luk available (%luk code, %luk reserved, "
615 "%luk data, %luk init)\n", (unsigned long) nr_free_pages() << (PAGE_SHIFT - 10),
616 num_physpages << (PAGE_SHIFT - 10), codesize >> 10,
617 reserved_pages << (PAGE_SHIFT - 10), datasize >> 10, initsize >> 10);
621 * For fsyscall entrpoints with no light-weight handler, use the ordinary
622 * (heavy-weight) handler, but mark it by setting bit 0, so the fsyscall entry
623 * code can tell them apart.
625 for (i = 0; i < NR_syscalls; ++i) {
626 extern unsigned long fsyscall_table[NR_syscalls];
627 extern unsigned long sys_call_table[NR_syscalls];
629 if (!fsyscall_table[i] || nolwsys)
630 fsyscall_table[i] = sys_call_table[i] | 1;
634 #ifdef CONFIG_IA32_SUPPORT
639 #ifdef CONFIG_MEMORY_HOTPLUG
640 void online_page(struct page *page)
642 ClearPageReserved(page);
643 init_page_count(page);
649 int add_memory(u64 start, u64 size)
653 unsigned long start_pfn = start >> PAGE_SHIFT;
654 unsigned long nr_pages = size >> PAGE_SHIFT;
657 pgdat = NODE_DATA(0);
659 zone = pgdat->node_zones + ZONE_NORMAL;
660 ret = __add_pages(zone, start_pfn, nr_pages);
663 printk("%s: Problem encountered in __add_pages() as ret=%d\n",
669 int remove_memory(u64 start, u64 size)