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 ia64_set_rbs_bot (void)
135 unsigned long stack_size
= current
->signal
->rlim
[RLIMIT_STACK
].rlim_max
& -16;
137 if (stack_size
> MAX_USER_STACK_SIZE
)
138 stack_size
= MAX_USER_STACK_SIZE
;
139 current
->thread
.rbs_bot
= STACK_TOP
- stack_size
;
143 * This performs some platform-dependent address space initialization.
144 * On IA-64, we want to setup the VM area for the register backing
145 * store (which grows upwards) and install the gateway page which is
146 * used for signal trampolines, etc.
149 ia64_init_addr_space (void)
151 struct vm_area_struct
*vma
;
156 * If we're out of memory and kmem_cache_alloc() returns NULL, we simply ignore
157 * the problem. When the process attempts to write to the register backing store
158 * for the first time, it will get a SEGFAULT in this case.
160 vma
= kmem_cache_alloc(vm_area_cachep
, GFP_KERNEL
);
162 memset(vma
, 0, sizeof(*vma
));
163 vma
->vm_mm
= current
->mm
;
164 vma
->vm_start
= current
->thread
.rbs_bot
& PAGE_MASK
;
165 vma
->vm_end
= vma
->vm_start
+ PAGE_SIZE
;
166 vma
->vm_page_prot
= protection_map
[VM_DATA_DEFAULT_FLAGS
& 0x7];
167 vma
->vm_flags
= VM_DATA_DEFAULT_FLAGS
|VM_GROWSUP
|VM_ACCOUNT
;
168 down_write(¤t
->mm
->mmap_sem
);
169 if (insert_vm_struct(current
->mm
, vma
)) {
170 up_write(¤t
->mm
->mmap_sem
);
171 kmem_cache_free(vm_area_cachep
, vma
);
174 up_write(¤t
->mm
->mmap_sem
);
177 /* map NaT-page at address zero to speed up speculative dereferencing of NULL: */
178 if (!(current
->personality
& MMAP_PAGE_ZERO
)) {
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_end
= PAGE_SIZE
;
184 vma
->vm_page_prot
= __pgprot(pgprot_val(PAGE_READONLY
) | _PAGE_MA_NAT
);
185 vma
->vm_flags
= VM_READ
| VM_MAYREAD
| VM_IO
| VM_RESERVED
;
186 down_write(¤t
->mm
->mmap_sem
);
187 if (insert_vm_struct(current
->mm
, vma
)) {
188 up_write(¤t
->mm
->mmap_sem
);
189 kmem_cache_free(vm_area_cachep
, vma
);
192 up_write(¤t
->mm
->mmap_sem
);
200 unsigned long addr
, eaddr
;
202 addr
= (unsigned long) ia64_imva(__init_begin
);
203 eaddr
= (unsigned long) ia64_imva(__init_end
);
204 while (addr
< eaddr
) {
205 ClearPageReserved(virt_to_page(addr
));
206 init_page_count(virt_to_page(addr
));
211 printk(KERN_INFO
"Freeing unused kernel memory: %ldkB freed\n",
212 (__init_end
- __init_begin
) >> 10);
216 free_initrd_mem (unsigned long start
, unsigned long end
)
220 * EFI uses 4KB pages while the kernel can use 4KB or bigger.
221 * Thus EFI and the kernel may have different page sizes. It is
222 * therefore possible to have the initrd share the same page as
223 * the end of the kernel (given current setup).
225 * To avoid freeing/using the wrong page (kernel sized) we:
226 * - align up the beginning of initrd
227 * - align down the end of initrd
230 * |=============| a000
236 * |=============| 8000
239 * |/////////////| 7000
242 * |=============| 6000
245 * K=kernel using 8KB pages
247 * In this example, we must free page 8000 ONLY. So we must align up
248 * initrd_start and keep initrd_end as is.
250 start
= PAGE_ALIGN(start
);
251 end
= end
& PAGE_MASK
;
254 printk(KERN_INFO
"Freeing initrd memory: %ldkB freed\n", (end
- start
) >> 10);
256 for (; start
< end
; start
+= PAGE_SIZE
) {
257 if (!virt_addr_valid(start
))
259 page
= virt_to_page(start
);
260 ClearPageReserved(page
);
261 init_page_count(page
);
268 * This installs a clean page in the kernel's page table.
270 static struct page
* __init
271 put_kernel_page (struct page
*page
, unsigned long address
, pgprot_t pgprot
)
278 if (!PageReserved(page
))
279 printk(KERN_ERR
"put_kernel_page: page at 0x%p not in reserved memory\n",
282 pgd
= pgd_offset_k(address
); /* note: this is NOT pgd_offset()! */
285 pud
= pud_alloc(&init_mm
, pgd
, address
);
288 pmd
= pmd_alloc(&init_mm
, pud
, address
);
291 pte
= pte_alloc_kernel(pmd
, address
);
296 set_pte(pte
, mk_pte(page
, pgprot
));
299 /* no need for flush_tlb */
309 * Map the gate page twice: once read-only to export the ELF
310 * headers etc. and once execute-only page to enable
311 * privilege-promotion via "epc":
313 page
= virt_to_page(ia64_imva(__start_gate_section
));
314 put_kernel_page(page
, GATE_ADDR
, PAGE_READONLY
);
315 #ifdef HAVE_BUGGY_SEGREL
316 page
= virt_to_page(ia64_imva(__start_gate_section
+ PAGE_SIZE
));
317 put_kernel_page(page
, GATE_ADDR
+ PAGE_SIZE
, PAGE_GATE
);
319 put_kernel_page(page
, GATE_ADDR
+ PERCPU_PAGE_SIZE
, PAGE_GATE
);
320 /* Fill in the holes (if any) with read-only zero pages: */
324 for (addr
= GATE_ADDR
+ PAGE_SIZE
;
325 addr
< GATE_ADDR
+ PERCPU_PAGE_SIZE
;
328 put_kernel_page(ZERO_PAGE(0), addr
,
330 put_kernel_page(ZERO_PAGE(0), addr
+ PERCPU_PAGE_SIZE
,
339 ia64_mmu_init (void *my_cpu_data
)
341 unsigned long psr
, pta
, impl_va_bits
;
342 extern void __devinit
tlb_init (void);
344 #ifdef CONFIG_DISABLE_VHPT
345 # define VHPT_ENABLE_BIT 0
347 # define VHPT_ENABLE_BIT 1
350 /* Pin mapping for percpu area into TLB */
351 psr
= ia64_clear_ic();
352 ia64_itr(0x2, IA64_TR_PERCPU_DATA
, PERCPU_ADDR
,
353 pte_val(pfn_pte(__pa(my_cpu_data
) >> PAGE_SHIFT
, PAGE_KERNEL
)),
360 * Check if the virtually mapped linear page table (VMLPT) overlaps with a mapped
361 * address space. The IA-64 architecture guarantees that at least 50 bits of
362 * virtual address space are implemented but if we pick a large enough page size
363 * (e.g., 64KB), the mapped address space is big enough that it will overlap with
364 * VMLPT. I assume that once we run on machines big enough to warrant 64KB pages,
365 * IMPL_VA_MSB will be significantly bigger, so this is unlikely to become a
366 * problem in practice. Alternatively, we could truncate the top of the mapped
367 * address space to not permit mappings that would overlap with the VMLPT.
371 # define mapped_space_bits (3*(PAGE_SHIFT - pte_bits) + PAGE_SHIFT)
373 * The virtual page table has to cover the entire implemented address space within
374 * a region even though not all of this space may be mappable. The reason for
375 * this is that the Access bit and Dirty bit fault handlers perform
376 * non-speculative accesses to the virtual page table, so the address range of the
377 * virtual page table itself needs to be covered by virtual page table.
379 # define vmlpt_bits (impl_va_bits - PAGE_SHIFT + pte_bits)
380 # define POW2(n) (1ULL << (n))
382 impl_va_bits
= ffz(~(local_cpu_data
->unimpl_va_mask
| (7UL << 61)));
384 if (impl_va_bits
< 51 || impl_va_bits
> 61)
385 panic("CPU has bogus IMPL_VA_MSB value of %lu!\n", impl_va_bits
- 1);
387 * mapped_space_bits - PAGE_SHIFT is the total number of ptes we need,
388 * which must fit into "vmlpt_bits - pte_bits" slots. Second half of
389 * the test makes sure that our mapped space doesn't overlap the
390 * unimplemented hole in the middle of the region.
392 if ((mapped_space_bits
- PAGE_SHIFT
> vmlpt_bits
- pte_bits
) ||
393 (mapped_space_bits
> impl_va_bits
- 1))
394 panic("Cannot build a big enough virtual-linear page table"
395 " to cover mapped address space.\n"
396 " Try using a smaller page size.\n");
399 /* place the VMLPT at the end of each page-table mapped region: */
400 pta
= POW2(61) - POW2(vmlpt_bits
);
403 * Set the (virtually mapped linear) page table address. Bit
404 * 8 selects between the short and long format, bits 2-7 the
405 * size of the table, and bit 0 whether the VHPT walker is
408 ia64_set_pta(pta
| (0 << 8) | (vmlpt_bits
<< 2) | VHPT_ENABLE_BIT
);
412 #ifdef CONFIG_HUGETLB_PAGE
413 ia64_set_rr(HPAGE_REGION_BASE
, HPAGE_SHIFT
<< 2);
418 #ifdef CONFIG_VIRTUAL_MEM_MAP
419 int vmemmap_find_next_valid_pfn(int node
, int i
)
421 unsigned long end_address
, hole_next_pfn
;
422 unsigned long stop_address
;
423 pg_data_t
*pgdat
= NODE_DATA(node
);
425 end_address
= (unsigned long) &vmem_map
[pgdat
->node_start_pfn
+ i
];
426 end_address
= PAGE_ALIGN(end_address
);
428 stop_address
= (unsigned long) &vmem_map
[
429 pgdat
->node_start_pfn
+ pgdat
->node_spanned_pages
];
437 pgd
= pgd_offset_k(end_address
);
438 if (pgd_none(*pgd
)) {
439 end_address
+= PGDIR_SIZE
;
443 pud
= pud_offset(pgd
, end_address
);
444 if (pud_none(*pud
)) {
445 end_address
+= PUD_SIZE
;
449 pmd
= pmd_offset(pud
, end_address
);
450 if (pmd_none(*pmd
)) {
451 end_address
+= PMD_SIZE
;
455 pte
= pte_offset_kernel(pmd
, end_address
);
457 if (pte_none(*pte
)) {
458 end_address
+= PAGE_SIZE
;
460 if ((end_address
< stop_address
) &&
461 (end_address
!= ALIGN(end_address
, 1UL << PMD_SHIFT
)))
465 /* Found next valid vmem_map page */
467 } while (end_address
< stop_address
);
469 end_address
= min(end_address
, stop_address
);
470 end_address
= end_address
- (unsigned long) vmem_map
+ sizeof(struct page
) - 1;
471 hole_next_pfn
= end_address
/ sizeof(struct page
);
472 return hole_next_pfn
- pgdat
->node_start_pfn
;
476 create_mem_map_page_table (u64 start
, u64 end
, void *arg
)
478 unsigned long address
, start_page
, end_page
;
479 struct page
*map_start
, *map_end
;
486 map_start
= vmem_map
+ (__pa(start
) >> PAGE_SHIFT
);
487 map_end
= vmem_map
+ (__pa(end
) >> PAGE_SHIFT
);
489 start_page
= (unsigned long) map_start
& PAGE_MASK
;
490 end_page
= PAGE_ALIGN((unsigned long) map_end
);
491 node
= paddr_to_nid(__pa(start
));
493 for (address
= start_page
; address
< end_page
; address
+= PAGE_SIZE
) {
494 pgd
= pgd_offset_k(address
);
496 pgd_populate(&init_mm
, pgd
, alloc_bootmem_pages_node(NODE_DATA(node
), PAGE_SIZE
));
497 pud
= pud_offset(pgd
, address
);
500 pud_populate(&init_mm
, pud
, alloc_bootmem_pages_node(NODE_DATA(node
), PAGE_SIZE
));
501 pmd
= pmd_offset(pud
, address
);
504 pmd_populate_kernel(&init_mm
, pmd
, alloc_bootmem_pages_node(NODE_DATA(node
), PAGE_SIZE
));
505 pte
= pte_offset_kernel(pmd
, address
);
508 set_pte(pte
, pfn_pte(__pa(alloc_bootmem_pages_node(NODE_DATA(node
), PAGE_SIZE
)) >> PAGE_SHIFT
,
514 struct memmap_init_callback_data
{
522 virtual_memmap_init (u64 start
, u64 end
, void *arg
)
524 struct memmap_init_callback_data
*args
;
525 struct page
*map_start
, *map_end
;
527 args
= (struct memmap_init_callback_data
*) arg
;
528 map_start
= vmem_map
+ (__pa(start
) >> PAGE_SHIFT
);
529 map_end
= vmem_map
+ (__pa(end
) >> PAGE_SHIFT
);
531 if (map_start
< args
->start
)
532 map_start
= args
->start
;
533 if (map_end
> args
->end
)
537 * We have to initialize "out of bounds" struct page elements that fit completely
538 * on the same pages that were allocated for the "in bounds" elements because they
539 * may be referenced later (and found to be "reserved").
541 map_start
-= ((unsigned long) map_start
& (PAGE_SIZE
- 1)) / sizeof(struct page
);
542 map_end
+= ((PAGE_ALIGN((unsigned long) map_end
) - (unsigned long) map_end
)
543 / sizeof(struct page
));
545 if (map_start
< map_end
)
546 memmap_init_zone((unsigned long)(map_end
- map_start
),
547 args
->nid
, args
->zone
, page_to_pfn(map_start
),
553 memmap_init (unsigned long size
, int nid
, unsigned long zone
,
554 unsigned long start_pfn
)
557 memmap_init_zone(size
, nid
, zone
, start_pfn
, MEMMAP_EARLY
);
560 struct memmap_init_callback_data args
;
562 start
= pfn_to_page(start_pfn
);
564 args
.end
= start
+ size
;
568 efi_memmap_walk(virtual_memmap_init
, &args
);
573 ia64_pfn_valid (unsigned long pfn
)
576 struct page
*pg
= pfn_to_page(pfn
);
578 return (__get_user(byte
, (char __user
*) pg
) == 0)
579 && ((((u64
)pg
& PAGE_MASK
) == (((u64
)(pg
+ 1) - 1) & PAGE_MASK
))
580 || (__get_user(byte
, (char __user
*) (pg
+ 1) - 1) == 0));
582 EXPORT_SYMBOL(ia64_pfn_valid
);
585 find_largest_hole (u64 start
, u64 end
, void *arg
)
589 static u64 last_end
= PAGE_OFFSET
;
591 /* NOTE: this algorithm assumes efi memmap table is ordered */
593 if (*max_gap
< (start
- last_end
))
594 *max_gap
= start
- last_end
;
599 #endif /* CONFIG_VIRTUAL_MEM_MAP */
602 register_active_ranges(u64 start
, u64 end
, void *arg
)
604 int nid
= paddr_to_nid(__pa(start
));
609 if (start
> crashk_res
.start
&& start
< crashk_res
.end
)
610 start
= crashk_res
.end
;
611 if (end
> crashk_res
.start
&& end
< crashk_res
.end
)
612 end
= crashk_res
.start
;
616 add_active_range(nid
, __pa(start
) >> PAGE_SHIFT
,
617 __pa(end
) >> PAGE_SHIFT
);
622 count_reserved_pages (u64 start
, u64 end
, void *arg
)
624 unsigned long num_reserved
= 0;
625 unsigned long *count
= arg
;
627 for (; start
< end
; start
+= PAGE_SIZE
)
628 if (PageReserved(virt_to_page(start
)))
630 *count
+= num_reserved
;
635 * Boot command-line option "nolwsys" can be used to disable the use of any light-weight
636 * system call handler. When this option is in effect, all fsyscalls will end up bubbling
637 * down into the kernel and calling the normal (heavy-weight) syscall handler. This is
638 * useful for performance testing, but conceivably could also come in handy for debugging
642 static int nolwsys __initdata
;
645 nolwsys_setup (char *s
)
651 __setup("nolwsys", nolwsys_setup
);
656 long reserved_pages
, codesize
, datasize
, initsize
;
659 static struct kcore_list kcore_mem
, kcore_vmem
, kcore_kernel
;
661 BUG_ON(PTRS_PER_PGD
* sizeof(pgd_t
) != PAGE_SIZE
);
662 BUG_ON(PTRS_PER_PMD
* sizeof(pmd_t
) != PAGE_SIZE
);
663 BUG_ON(PTRS_PER_PTE
* sizeof(pte_t
) != PAGE_SIZE
);
667 * This needs to be called _after_ the command line has been parsed but _before_
668 * any drivers that may need the PCI DMA interface are initialized or bootmem has
674 #ifdef CONFIG_FLATMEM
677 max_mapnr
= max_low_pfn
;
680 high_memory
= __va(max_low_pfn
* PAGE_SIZE
);
682 kclist_add(&kcore_mem
, __va(0), max_low_pfn
* PAGE_SIZE
);
683 kclist_add(&kcore_vmem
, (void *)VMALLOC_START
, VMALLOC_END
-VMALLOC_START
);
684 kclist_add(&kcore_kernel
, _stext
, _end
- _stext
);
686 for_each_online_pgdat(pgdat
)
687 if (pgdat
->bdata
->node_bootmem_map
)
688 totalram_pages
+= free_all_bootmem_node(pgdat
);
691 efi_memmap_walk(count_reserved_pages
, &reserved_pages
);
693 codesize
= (unsigned long) _etext
- (unsigned long) _stext
;
694 datasize
= (unsigned long) _edata
- (unsigned long) _etext
;
695 initsize
= (unsigned long) __init_end
- (unsigned long) __init_begin
;
697 printk(KERN_INFO
"Memory: %luk/%luk available (%luk code, %luk reserved, "
698 "%luk data, %luk init)\n", (unsigned long) nr_free_pages() << (PAGE_SHIFT
- 10),
699 num_physpages
<< (PAGE_SHIFT
- 10), codesize
>> 10,
700 reserved_pages
<< (PAGE_SHIFT
- 10), datasize
>> 10, initsize
>> 10);
704 * For fsyscall entrpoints with no light-weight handler, use the ordinary
705 * (heavy-weight) handler, but mark it by setting bit 0, so the fsyscall entry
706 * code can tell them apart.
708 for (i
= 0; i
< NR_syscalls
; ++i
) {
709 extern unsigned long fsyscall_table
[NR_syscalls
];
710 extern unsigned long sys_call_table
[NR_syscalls
];
712 if (!fsyscall_table
[i
] || nolwsys
)
713 fsyscall_table
[i
] = sys_call_table
[i
] | 1;
717 #ifdef CONFIG_IA32_SUPPORT
722 #ifdef CONFIG_MEMORY_HOTPLUG
723 void online_page(struct page
*page
)
725 ClearPageReserved(page
);
726 init_page_count(page
);
732 int arch_add_memory(int nid
, u64 start
, u64 size
)
736 unsigned long start_pfn
= start
>> PAGE_SHIFT
;
737 unsigned long nr_pages
= size
>> PAGE_SHIFT
;
740 pgdat
= NODE_DATA(nid
);
742 zone
= pgdat
->node_zones
+ ZONE_NORMAL
;
743 ret
= __add_pages(zone
, start_pfn
, nr_pages
);
746 printk("%s: Problem encountered in __add_pages() as ret=%d\n",
752 int remove_memory(u64 start
, u64 size
)
756 EXPORT_SYMBOL_GPL(remove_memory
);