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>
27 #include <asm/machvec.h>
29 #include <asm/patch.h>
30 #include <asm/pgalloc.h>
32 #include <asm/sections.h>
33 #include <asm/system.h>
35 #include <asm/uaccess.h>
36 #include <asm/unistd.h>
38 #include <asm/paravirt.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 struct page
*zero_page_memmap_ptr
; /* map entry for zero page */
54 EXPORT_SYMBOL(zero_page_memmap_ptr
);
57 __ia64_sync_icache_dcache (pte_t pte
)
63 addr
= (unsigned long) page_address(page
);
65 if (test_bit(PG_arch_1
, &page
->flags
))
66 return; /* i-cache is already coherent with d-cache */
68 flush_icache_range(addr
, addr
+ (PAGE_SIZE
<< compound_order(page
)));
69 set_bit(PG_arch_1
, &page
->flags
); /* mark page as clean */
73 * Since DMA is i-cache coherent, any (complete) pages that were written via
74 * DMA can be marked as "clean" so that lazy_mmu_prot_update() doesn't have to
75 * flush them when they get mapped into an executable vm-area.
78 dma_mark_clean(void *addr
, size_t size
)
80 unsigned long pg_addr
, end
;
82 pg_addr
= PAGE_ALIGN((unsigned long) addr
);
83 end
= (unsigned long) addr
+ size
;
84 while (pg_addr
+ PAGE_SIZE
<= end
) {
85 struct page
*page
= virt_to_page(pg_addr
);
86 set_bit(PG_arch_1
, &page
->flags
);
92 ia64_set_rbs_bot (void)
94 unsigned long stack_size
= rlimit_max(RLIMIT_STACK
) & -16;
96 if (stack_size
> MAX_USER_STACK_SIZE
)
97 stack_size
= MAX_USER_STACK_SIZE
;
98 current
->thread
.rbs_bot
= PAGE_ALIGN(current
->mm
->start_stack
- stack_size
);
102 * This performs some platform-dependent address space initialization.
103 * On IA-64, we want to setup the VM area for the register backing
104 * store (which grows upwards) and install the gateway page which is
105 * used for signal trampolines, etc.
108 ia64_init_addr_space (void)
110 struct vm_area_struct
*vma
;
115 * If we're out of memory and kmem_cache_alloc() returns NULL, we simply ignore
116 * the problem. When the process attempts to write to the register backing store
117 * for the first time, it will get a SEGFAULT in this case.
119 vma
= kmem_cache_zalloc(vm_area_cachep
, GFP_KERNEL
);
121 vma
->vm_mm
= current
->mm
;
122 vma
->vm_start
= current
->thread
.rbs_bot
& PAGE_MASK
;
123 vma
->vm_end
= vma
->vm_start
+ PAGE_SIZE
;
124 vma
->vm_flags
= VM_DATA_DEFAULT_FLAGS
|VM_GROWSUP
|VM_ACCOUNT
;
125 vma
->vm_page_prot
= vm_get_page_prot(vma
->vm_flags
);
126 down_write(¤t
->mm
->mmap_sem
);
127 if (insert_vm_struct(current
->mm
, vma
)) {
128 up_write(¤t
->mm
->mmap_sem
);
129 kmem_cache_free(vm_area_cachep
, vma
);
132 up_write(¤t
->mm
->mmap_sem
);
135 /* map NaT-page at address zero to speed up speculative dereferencing of NULL: */
136 if (!(current
->personality
& MMAP_PAGE_ZERO
)) {
137 vma
= kmem_cache_zalloc(vm_area_cachep
, GFP_KERNEL
);
139 vma
->vm_mm
= current
->mm
;
140 vma
->vm_end
= PAGE_SIZE
;
141 vma
->vm_page_prot
= __pgprot(pgprot_val(PAGE_READONLY
) | _PAGE_MA_NAT
);
142 vma
->vm_flags
= VM_READ
| VM_MAYREAD
| VM_IO
| VM_RESERVED
;
143 down_write(¤t
->mm
->mmap_sem
);
144 if (insert_vm_struct(current
->mm
, vma
)) {
145 up_write(¤t
->mm
->mmap_sem
);
146 kmem_cache_free(vm_area_cachep
, vma
);
149 up_write(¤t
->mm
->mmap_sem
);
157 unsigned long addr
, eaddr
;
159 addr
= (unsigned long) ia64_imva(__init_begin
);
160 eaddr
= (unsigned long) ia64_imva(__init_end
);
161 while (addr
< eaddr
) {
162 ClearPageReserved(virt_to_page(addr
));
163 init_page_count(virt_to_page(addr
));
168 printk(KERN_INFO
"Freeing unused kernel memory: %ldkB freed\n",
169 (__init_end
- __init_begin
) >> 10);
173 free_initrd_mem (unsigned long start
, unsigned long end
)
177 * EFI uses 4KB pages while the kernel can use 4KB or bigger.
178 * Thus EFI and the kernel may have different page sizes. It is
179 * therefore possible to have the initrd share the same page as
180 * the end of the kernel (given current setup).
182 * To avoid freeing/using the wrong page (kernel sized) we:
183 * - align up the beginning of initrd
184 * - align down the end of initrd
187 * |=============| a000
193 * |=============| 8000
196 * |/////////////| 7000
199 * |=============| 6000
202 * K=kernel using 8KB pages
204 * In this example, we must free page 8000 ONLY. So we must align up
205 * initrd_start and keep initrd_end as is.
207 start
= PAGE_ALIGN(start
);
208 end
= end
& PAGE_MASK
;
211 printk(KERN_INFO
"Freeing initrd memory: %ldkB freed\n", (end
- start
) >> 10);
213 for (; start
< end
; start
+= PAGE_SIZE
) {
214 if (!virt_addr_valid(start
))
216 page
= virt_to_page(start
);
217 ClearPageReserved(page
);
218 init_page_count(page
);
225 * This installs a clean page in the kernel's page table.
227 static struct page
* __init
228 put_kernel_page (struct page
*page
, unsigned long address
, pgprot_t pgprot
)
235 if (!PageReserved(page
))
236 printk(KERN_ERR
"put_kernel_page: page at 0x%p not in reserved memory\n",
239 pgd
= pgd_offset_k(address
); /* note: this is NOT pgd_offset()! */
242 pud
= pud_alloc(&init_mm
, pgd
, address
);
245 pmd
= pmd_alloc(&init_mm
, pud
, address
);
248 pte
= pte_alloc_kernel(pmd
, address
);
253 set_pte(pte
, mk_pte(page
, pgprot
));
256 /* no need for flush_tlb */
267 * Map the gate page twice: once read-only to export the ELF
268 * headers etc. and once execute-only page to enable
269 * privilege-promotion via "epc":
271 gate_section
= paravirt_get_gate_section();
272 page
= virt_to_page(ia64_imva(gate_section
));
273 put_kernel_page(page
, GATE_ADDR
, PAGE_READONLY
);
274 #ifdef HAVE_BUGGY_SEGREL
275 page
= virt_to_page(ia64_imva(gate_section
+ PAGE_SIZE
));
276 put_kernel_page(page
, GATE_ADDR
+ PAGE_SIZE
, PAGE_GATE
);
278 put_kernel_page(page
, GATE_ADDR
+ PERCPU_PAGE_SIZE
, PAGE_GATE
);
279 /* Fill in the holes (if any) with read-only zero pages: */
283 for (addr
= GATE_ADDR
+ PAGE_SIZE
;
284 addr
< GATE_ADDR
+ PERCPU_PAGE_SIZE
;
287 put_kernel_page(ZERO_PAGE(0), addr
,
289 put_kernel_page(ZERO_PAGE(0), addr
+ PERCPU_PAGE_SIZE
,
298 ia64_mmu_init (void *my_cpu_data
)
300 unsigned long pta
, impl_va_bits
;
301 extern void __devinit
tlb_init (void);
303 #ifdef CONFIG_DISABLE_VHPT
304 # define VHPT_ENABLE_BIT 0
306 # define VHPT_ENABLE_BIT 1
310 * Check if the virtually mapped linear page table (VMLPT) overlaps with a mapped
311 * address space. The IA-64 architecture guarantees that at least 50 bits of
312 * virtual address space are implemented but if we pick a large enough page size
313 * (e.g., 64KB), the mapped address space is big enough that it will overlap with
314 * VMLPT. I assume that once we run on machines big enough to warrant 64KB pages,
315 * IMPL_VA_MSB will be significantly bigger, so this is unlikely to become a
316 * problem in practice. Alternatively, we could truncate the top of the mapped
317 * address space to not permit mappings that would overlap with the VMLPT.
321 # define mapped_space_bits (3*(PAGE_SHIFT - pte_bits) + PAGE_SHIFT)
323 * The virtual page table has to cover the entire implemented address space within
324 * a region even though not all of this space may be mappable. The reason for
325 * this is that the Access bit and Dirty bit fault handlers perform
326 * non-speculative accesses to the virtual page table, so the address range of the
327 * virtual page table itself needs to be covered by virtual page table.
329 # define vmlpt_bits (impl_va_bits - PAGE_SHIFT + pte_bits)
330 # define POW2(n) (1ULL << (n))
332 impl_va_bits
= ffz(~(local_cpu_data
->unimpl_va_mask
| (7UL << 61)));
334 if (impl_va_bits
< 51 || impl_va_bits
> 61)
335 panic("CPU has bogus IMPL_VA_MSB value of %lu!\n", impl_va_bits
- 1);
337 * mapped_space_bits - PAGE_SHIFT is the total number of ptes we need,
338 * which must fit into "vmlpt_bits - pte_bits" slots. Second half of
339 * the test makes sure that our mapped space doesn't overlap the
340 * unimplemented hole in the middle of the region.
342 if ((mapped_space_bits
- PAGE_SHIFT
> vmlpt_bits
- pte_bits
) ||
343 (mapped_space_bits
> impl_va_bits
- 1))
344 panic("Cannot build a big enough virtual-linear page table"
345 " to cover mapped address space.\n"
346 " Try using a smaller page size.\n");
349 /* place the VMLPT at the end of each page-table mapped region: */
350 pta
= POW2(61) - POW2(vmlpt_bits
);
353 * Set the (virtually mapped linear) page table address. Bit
354 * 8 selects between the short and long format, bits 2-7 the
355 * size of the table, and bit 0 whether the VHPT walker is
358 ia64_set_pta(pta
| (0 << 8) | (vmlpt_bits
<< 2) | VHPT_ENABLE_BIT
);
362 #ifdef CONFIG_HUGETLB_PAGE
363 ia64_set_rr(HPAGE_REGION_BASE
, HPAGE_SHIFT
<< 2);
368 #ifdef CONFIG_VIRTUAL_MEM_MAP
369 int vmemmap_find_next_valid_pfn(int node
, int i
)
371 unsigned long end_address
, hole_next_pfn
;
372 unsigned long stop_address
;
373 pg_data_t
*pgdat
= NODE_DATA(node
);
375 end_address
= (unsigned long) &vmem_map
[pgdat
->node_start_pfn
+ i
];
376 end_address
= PAGE_ALIGN(end_address
);
378 stop_address
= (unsigned long) &vmem_map
[
379 pgdat
->node_start_pfn
+ pgdat
->node_spanned_pages
];
387 pgd
= pgd_offset_k(end_address
);
388 if (pgd_none(*pgd
)) {
389 end_address
+= PGDIR_SIZE
;
393 pud
= pud_offset(pgd
, end_address
);
394 if (pud_none(*pud
)) {
395 end_address
+= PUD_SIZE
;
399 pmd
= pmd_offset(pud
, end_address
);
400 if (pmd_none(*pmd
)) {
401 end_address
+= PMD_SIZE
;
405 pte
= pte_offset_kernel(pmd
, end_address
);
407 if (pte_none(*pte
)) {
408 end_address
+= PAGE_SIZE
;
410 if ((end_address
< stop_address
) &&
411 (end_address
!= ALIGN(end_address
, 1UL << PMD_SHIFT
)))
415 /* Found next valid vmem_map page */
417 } while (end_address
< stop_address
);
419 end_address
= min(end_address
, stop_address
);
420 end_address
= end_address
- (unsigned long) vmem_map
+ sizeof(struct page
) - 1;
421 hole_next_pfn
= end_address
/ sizeof(struct page
);
422 return hole_next_pfn
- pgdat
->node_start_pfn
;
425 int __init
create_mem_map_page_table(u64 start
, u64 end
, void *arg
)
427 unsigned long address
, start_page
, end_page
;
428 struct page
*map_start
, *map_end
;
435 map_start
= vmem_map
+ (__pa(start
) >> PAGE_SHIFT
);
436 map_end
= vmem_map
+ (__pa(end
) >> PAGE_SHIFT
);
438 start_page
= (unsigned long) map_start
& PAGE_MASK
;
439 end_page
= PAGE_ALIGN((unsigned long) map_end
);
440 node
= paddr_to_nid(__pa(start
));
442 for (address
= start_page
; address
< end_page
; address
+= PAGE_SIZE
) {
443 pgd
= pgd_offset_k(address
);
445 pgd_populate(&init_mm
, pgd
, alloc_bootmem_pages_node(NODE_DATA(node
), PAGE_SIZE
));
446 pud
= pud_offset(pgd
, address
);
449 pud_populate(&init_mm
, pud
, alloc_bootmem_pages_node(NODE_DATA(node
), PAGE_SIZE
));
450 pmd
= pmd_offset(pud
, address
);
453 pmd_populate_kernel(&init_mm
, pmd
, alloc_bootmem_pages_node(NODE_DATA(node
), PAGE_SIZE
));
454 pte
= pte_offset_kernel(pmd
, address
);
457 set_pte(pte
, pfn_pte(__pa(alloc_bootmem_pages_node(NODE_DATA(node
), PAGE_SIZE
)) >> PAGE_SHIFT
,
463 struct memmap_init_callback_data
{
471 virtual_memmap_init(u64 start
, u64 end
, void *arg
)
473 struct memmap_init_callback_data
*args
;
474 struct page
*map_start
, *map_end
;
476 args
= (struct memmap_init_callback_data
*) arg
;
477 map_start
= vmem_map
+ (__pa(start
) >> PAGE_SHIFT
);
478 map_end
= vmem_map
+ (__pa(end
) >> PAGE_SHIFT
);
480 if (map_start
< args
->start
)
481 map_start
= args
->start
;
482 if (map_end
> args
->end
)
486 * We have to initialize "out of bounds" struct page elements that fit completely
487 * on the same pages that were allocated for the "in bounds" elements because they
488 * may be referenced later (and found to be "reserved").
490 map_start
-= ((unsigned long) map_start
& (PAGE_SIZE
- 1)) / sizeof(struct page
);
491 map_end
+= ((PAGE_ALIGN((unsigned long) map_end
) - (unsigned long) map_end
)
492 / sizeof(struct page
));
494 if (map_start
< map_end
)
495 memmap_init_zone((unsigned long)(map_end
- map_start
),
496 args
->nid
, args
->zone
, page_to_pfn(map_start
),
502 memmap_init (unsigned long size
, int nid
, unsigned long zone
,
503 unsigned long start_pfn
)
506 memmap_init_zone(size
, nid
, zone
, start_pfn
, MEMMAP_EARLY
);
509 struct memmap_init_callback_data args
;
511 start
= pfn_to_page(start_pfn
);
513 args
.end
= start
+ size
;
517 efi_memmap_walk(virtual_memmap_init
, &args
);
522 ia64_pfn_valid (unsigned long pfn
)
525 struct page
*pg
= pfn_to_page(pfn
);
527 return (__get_user(byte
, (char __user
*) pg
) == 0)
528 && ((((u64
)pg
& PAGE_MASK
) == (((u64
)(pg
+ 1) - 1) & PAGE_MASK
))
529 || (__get_user(byte
, (char __user
*) (pg
+ 1) - 1) == 0));
531 EXPORT_SYMBOL(ia64_pfn_valid
);
533 int __init
find_largest_hole(u64 start
, u64 end
, void *arg
)
537 static u64 last_end
= PAGE_OFFSET
;
539 /* NOTE: this algorithm assumes efi memmap table is ordered */
541 if (*max_gap
< (start
- last_end
))
542 *max_gap
= start
- last_end
;
547 #endif /* CONFIG_VIRTUAL_MEM_MAP */
549 int __init
register_active_ranges(u64 start
, u64 len
, int nid
)
551 u64 end
= start
+ len
;
554 if (start
> crashk_res
.start
&& start
< crashk_res
.end
)
555 start
= crashk_res
.end
;
556 if (end
> crashk_res
.start
&& end
< crashk_res
.end
)
557 end
= crashk_res
.start
;
561 add_active_range(nid
, __pa(start
) >> PAGE_SHIFT
,
562 __pa(end
) >> PAGE_SHIFT
);
567 count_reserved_pages(u64 start
, u64 end
, void *arg
)
569 unsigned long num_reserved
= 0;
570 unsigned long *count
= arg
;
572 for (; start
< end
; start
+= PAGE_SIZE
)
573 if (PageReserved(virt_to_page(start
)))
575 *count
+= num_reserved
;
580 find_max_min_low_pfn (u64 start
, u64 end
, void *arg
)
582 unsigned long pfn_start
, pfn_end
;
583 #ifdef CONFIG_FLATMEM
584 pfn_start
= (PAGE_ALIGN(__pa(start
))) >> PAGE_SHIFT
;
585 pfn_end
= (PAGE_ALIGN(__pa(end
- 1))) >> PAGE_SHIFT
;
587 pfn_start
= GRANULEROUNDDOWN(__pa(start
)) >> PAGE_SHIFT
;
588 pfn_end
= GRANULEROUNDUP(__pa(end
- 1)) >> PAGE_SHIFT
;
590 min_low_pfn
= min(min_low_pfn
, pfn_start
);
591 max_low_pfn
= max(max_low_pfn
, pfn_end
);
596 * Boot command-line option "nolwsys" can be used to disable the use of any light-weight
597 * system call handler. When this option is in effect, all fsyscalls will end up bubbling
598 * down into the kernel and calling the normal (heavy-weight) syscall handler. This is
599 * useful for performance testing, but conceivably could also come in handy for debugging
603 static int nolwsys __initdata
;
606 nolwsys_setup (char *s
)
612 __setup("nolwsys", nolwsys_setup
);
617 long reserved_pages
, codesize
, datasize
, initsize
;
621 BUG_ON(PTRS_PER_PGD
* sizeof(pgd_t
) != PAGE_SIZE
);
622 BUG_ON(PTRS_PER_PMD
* sizeof(pmd_t
) != PAGE_SIZE
);
623 BUG_ON(PTRS_PER_PTE
* sizeof(pte_t
) != PAGE_SIZE
);
627 * This needs to be called _after_ the command line has been parsed but _before_
628 * any drivers that may need the PCI DMA interface are initialized or bootmem has
634 #ifdef CONFIG_FLATMEM
636 max_mapnr
= max_low_pfn
;
639 high_memory
= __va(max_low_pfn
* PAGE_SIZE
);
641 for_each_online_pgdat(pgdat
)
642 if (pgdat
->bdata
->node_bootmem_map
)
643 totalram_pages
+= free_all_bootmem_node(pgdat
);
646 efi_memmap_walk(count_reserved_pages
, &reserved_pages
);
648 codesize
= (unsigned long) _etext
- (unsigned long) _stext
;
649 datasize
= (unsigned long) _edata
- (unsigned long) _etext
;
650 initsize
= (unsigned long) __init_end
- (unsigned long) __init_begin
;
652 printk(KERN_INFO
"Memory: %luk/%luk available (%luk code, %luk reserved, "
653 "%luk data, %luk init)\n", nr_free_pages() << (PAGE_SHIFT
- 10),
654 num_physpages
<< (PAGE_SHIFT
- 10), codesize
>> 10,
655 reserved_pages
<< (PAGE_SHIFT
- 10), datasize
>> 10, initsize
>> 10);
659 * For fsyscall entrpoints with no light-weight handler, use the ordinary
660 * (heavy-weight) handler, but mark it by setting bit 0, so the fsyscall entry
661 * code can tell them apart.
663 for (i
= 0; i
< NR_syscalls
; ++i
) {
664 extern unsigned long sys_call_table
[NR_syscalls
];
665 unsigned long *fsyscall_table
= paravirt_get_fsyscall_table();
667 if (!fsyscall_table
[i
] || nolwsys
)
668 fsyscall_table
[i
] = sys_call_table
[i
] | 1;
672 #ifdef CONFIG_IA32_SUPPORT
677 #ifdef CONFIG_MEMORY_HOTPLUG
678 int arch_add_memory(int nid
, u64 start
, u64 size
)
682 unsigned long start_pfn
= start
>> PAGE_SHIFT
;
683 unsigned long nr_pages
= size
>> PAGE_SHIFT
;
686 pgdat
= NODE_DATA(nid
);
688 zone
= pgdat
->node_zones
+ ZONE_NORMAL
;
689 ret
= __add_pages(nid
, zone
, start_pfn
, nr_pages
);
692 printk("%s: Problem encountered in __add_pages() as ret=%d\n",
700 * Even when CONFIG_IA32_SUPPORT is not enabled it is
701 * useful to have the Linux/x86 domain registered to
702 * avoid an attempted module load when emulators call
703 * personality(PER_LINUX32). This saves several milliseconds
706 static struct exec_domain ia32_exec_domain
;
709 per_linux32_init(void)
711 ia32_exec_domain
.name
= "Linux/x86";
712 ia32_exec_domain
.handler
= NULL
;
713 ia32_exec_domain
.pers_low
= PER_LINUX32
;
714 ia32_exec_domain
.pers_high
= PER_LINUX32
;
715 ia32_exec_domain
.signal_map
= default_exec_domain
.signal_map
;
716 ia32_exec_domain
.signal_invmap
= default_exec_domain
.signal_invmap
;
717 register_exec_domain(&ia32_exec_domain
);
722 __initcall(per_linux32_init
);