2 * Copyright 2002 Andi Kleen, SuSE Labs.
3 * Thanks to Ben LaHaise for precious feedback.
5 #include <linux/highmem.h>
6 #include <linux/bootmem.h>
7 #include <linux/module.h>
8 #include <linux/sched.h>
9 #include <linux/slab.h>
13 #include <asm/processor.h>
14 #include <asm/tlbflush.h>
15 #include <asm/sections.h>
16 #include <asm/uaccess.h>
17 #include <asm/pgalloc.h>
20 * The current flushing context - we pass it instead of 5 arguments:
31 within(unsigned long addr
, unsigned long start
, unsigned long end
)
33 return addr
>= start
&& addr
< end
;
41 * clflush_cache_range - flush a cache range with clflush
42 * @addr: virtual start address
43 * @size: number of bytes to flush
45 * clflush is an unordered instruction which needs fencing with mfence
46 * to avoid ordering issues.
48 void clflush_cache_range(void *vaddr
, unsigned int size
)
50 void *vend
= vaddr
+ size
- 1;
54 for (; vaddr
< vend
; vaddr
+= boot_cpu_data
.x86_clflush_size
)
57 * Flush any possible final partial cacheline:
64 static void __cpa_flush_all(void *arg
)
66 unsigned long cache
= (unsigned long)arg
;
69 * Flush all to work around Errata in early athlons regarding
70 * large page flushing.
74 if (cache
&& boot_cpu_data
.x86_model
>= 4)
78 static void cpa_flush_all(unsigned long cache
)
80 BUG_ON(irqs_disabled());
82 on_each_cpu(__cpa_flush_all
, (void *) cache
, 1, 1);
85 static void __cpa_flush_range(void *arg
)
88 * We could optimize that further and do individual per page
89 * tlb invalidates for a low number of pages. Caveat: we must
90 * flush the high aliases on 64bit as well.
95 static void cpa_flush_range(unsigned long start
, int numpages
, int cache
)
97 unsigned int i
, level
;
100 BUG_ON(irqs_disabled());
101 WARN_ON(PAGE_ALIGN(start
) != start
);
103 on_each_cpu(__cpa_flush_range
, NULL
, 1, 1);
109 * We only need to flush on one CPU,
110 * clflush is a MESI-coherent instruction that
111 * will cause all other CPUs to flush the same
114 for (i
= 0, addr
= start
; i
< numpages
; i
++, addr
+= PAGE_SIZE
) {
115 pte_t
*pte
= lookup_address(addr
, &level
);
118 * Only flush present addresses:
120 if (pte
&& (pte_val(*pte
) & _PAGE_PRESENT
))
121 clflush_cache_range((void *) addr
, PAGE_SIZE
);
125 #define HIGH_MAP_START __START_KERNEL_map
126 #define HIGH_MAP_END (__START_KERNEL_map + KERNEL_TEXT_SIZE)
130 * Converts a virtual address to a X86-64 highmap address
132 static unsigned long virt_to_highmap(void *address
)
135 return __pa((unsigned long)address
) + HIGH_MAP_START
- phys_base
;
137 return (unsigned long)address
;
142 * Certain areas of memory on x86 require very specific protection flags,
143 * for example the BIOS area or kernel text. Callers don't always get this
144 * right (again, ioremap() on BIOS memory is not uncommon) so this function
145 * checks and fixes these known static required protection bits.
147 static inline pgprot_t
static_protections(pgprot_t prot
, unsigned long address
)
149 pgprot_t forbidden
= __pgprot(0);
152 * The BIOS area between 640k and 1Mb needs to be executable for
153 * PCI BIOS based config access (CONFIG_PCI_GOBIOS) support.
155 if (within(__pa(address
), BIOS_BEGIN
, BIOS_END
))
156 pgprot_val(forbidden
) |= _PAGE_NX
;
159 * The kernel text needs to be executable for obvious reasons
160 * Does not cover __inittext since that is gone later on
162 if (within(address
, (unsigned long)_text
, (unsigned long)_etext
))
163 pgprot_val(forbidden
) |= _PAGE_NX
;
165 * Do the same for the x86-64 high kernel mapping
167 if (within(address
, virt_to_highmap(_text
), virt_to_highmap(_etext
)))
168 pgprot_val(forbidden
) |= _PAGE_NX
;
171 #ifdef CONFIG_DEBUG_RODATA
172 /* The .rodata section needs to be read-only */
173 if (within(address
, (unsigned long)__start_rodata
,
174 (unsigned long)__end_rodata
))
175 pgprot_val(forbidden
) |= _PAGE_RW
;
177 * Do the same for the x86-64 high kernel mapping
179 if (within(address
, virt_to_highmap(__start_rodata
),
180 virt_to_highmap(__end_rodata
)))
181 pgprot_val(forbidden
) |= _PAGE_RW
;
184 prot
= __pgprot(pgprot_val(prot
) & ~pgprot_val(forbidden
));
190 * Lookup the page table entry for a virtual address. Return a pointer
191 * to the entry and the level of the mapping.
193 * Note: We return pud and pmd either when the entry is marked large
194 * or when the present bit is not set. Otherwise we would return a
195 * pointer to a nonexisting mapping.
197 pte_t
*lookup_address(unsigned long address
, int *level
)
199 pgd_t
*pgd
= pgd_offset_k(address
);
203 *level
= PG_LEVEL_NONE
;
208 pud
= pud_offset(pgd
, address
);
212 *level
= PG_LEVEL_1G
;
213 if (pud_large(*pud
) || !pud_present(*pud
))
216 pmd
= pmd_offset(pud
, address
);
220 *level
= PG_LEVEL_2M
;
221 if (pmd_large(*pmd
) || !pmd_present(*pmd
))
224 *level
= PG_LEVEL_4K
;
226 return pte_offset_kernel(pmd
, address
);
230 * Set the new pmd in all the pgds we know about:
232 static void __set_pmd_pte(pte_t
*kpte
, unsigned long address
, pte_t pte
)
235 set_pte_atomic(kpte
, pte
);
237 if (!SHARED_KERNEL_PMD
) {
240 list_for_each_entry(page
, &pgd_list
, lru
) {
245 pgd
= (pgd_t
*)page_address(page
) + pgd_index(address
);
246 pud
= pud_offset(pgd
, address
);
247 pmd
= pmd_offset(pud
, address
);
248 set_pte_atomic((pte_t
*)pmd
, pte
);
255 try_preserve_large_page(pte_t
*kpte
, unsigned long address
,
256 struct cpa_data
*cpa
)
258 unsigned long nextpage_addr
, numpages
, pmask
, psize
, flags
;
259 pte_t new_pte
, old_pte
, *tmp
;
260 pgprot_t old_prot
, new_prot
;
261 int level
, do_split
= 1;
263 spin_lock_irqsave(&pgd_lock
, flags
);
265 * Check for races, another CPU might have split this page
268 tmp
= lookup_address(address
, &level
);
274 psize
= PMD_PAGE_SIZE
;
275 pmask
= PMD_PAGE_MASK
;
279 psize
= PMD_PAGE_SIZE
;
280 pmask
= PMD_PAGE_MASK
;
289 * Calculate the number of pages, which fit into this large
290 * page starting at address:
292 nextpage_addr
= (address
+ psize
) & pmask
;
293 numpages
= (nextpage_addr
- address
) >> PAGE_SHIFT
;
294 if (numpages
< cpa
->numpages
)
295 cpa
->numpages
= numpages
;
298 * We are safe now. Check whether the new pgprot is the same:
301 old_prot
= new_prot
= pte_pgprot(old_pte
);
303 pgprot_val(new_prot
) &= ~pgprot_val(cpa
->mask_clr
);
304 pgprot_val(new_prot
) |= pgprot_val(cpa
->mask_set
);
305 new_prot
= static_protections(new_prot
, address
);
308 * If there are no changes, return. maxpages has been updated
311 if (pgprot_val(new_prot
) == pgprot_val(old_prot
)) {
317 * We need to change the attributes. Check, whether we can
318 * change the large page in one go. We request a split, when
319 * the address is not aligned and the number of pages is
320 * smaller than the number of pages in the large page. Note
321 * that we limited the number of possible pages already to
322 * the number of pages in the large page.
324 if (address
== (nextpage_addr
- psize
) && cpa
->numpages
== numpages
) {
326 * The address is aligned and the number of pages
327 * covers the full page.
329 new_pte
= pfn_pte(pte_pfn(old_pte
), canon_pgprot(new_prot
));
330 __set_pmd_pte(kpte
, address
, new_pte
);
336 spin_unlock_irqrestore(&pgd_lock
, flags
);
341 static int split_large_page(pte_t
*kpte
, unsigned long address
)
343 unsigned long flags
, pfn
, pfninc
= 1;
344 gfp_t gfp_flags
= GFP_KERNEL
;
345 unsigned int i
, level
;
350 #ifdef CONFIG_DEBUG_PAGEALLOC
351 gfp_flags
= GFP_ATOMIC
| __GFP_NOWARN
;
353 base
= alloc_pages(gfp_flags
, 0);
357 spin_lock_irqsave(&pgd_lock
, flags
);
359 * Check for races, another CPU might have split this page
362 tmp
= lookup_address(address
, &level
);
366 pbase
= (pte_t
*)page_address(base
);
368 paravirt_alloc_pt(&init_mm
, page_to_pfn(base
));
370 ref_prot
= pte_pgprot(pte_clrhuge(*kpte
));
373 if (level
== PG_LEVEL_1G
) {
374 pfninc
= PMD_PAGE_SIZE
>> PAGE_SHIFT
;
375 pgprot_val(ref_prot
) |= _PAGE_PSE
;
380 * Get the target pfn from the original entry:
382 pfn
= pte_pfn(*kpte
);
383 for (i
= 0; i
< PTRS_PER_PTE
; i
++, pfn
+= pfninc
)
384 set_pte(&pbase
[i
], pfn_pte(pfn
, ref_prot
));
387 * Install the new, split up pagetable. Important details here:
389 * On Intel the NX bit of all levels must be cleared to make a
390 * page executable. See section 4.13.2 of Intel 64 and IA-32
391 * Architectures Software Developer's Manual).
393 * Mark the entry present. The current mapping might be
394 * set to not present, which we preserved above.
396 ref_prot
= pte_pgprot(pte_mkexec(pte_clrhuge(*kpte
)));
397 pgprot_val(ref_prot
) |= _PAGE_PRESENT
;
398 __set_pmd_pte(kpte
, address
, mk_pte(base
, ref_prot
));
402 spin_unlock_irqrestore(&pgd_lock
, flags
);
405 __free_pages(base
, 0);
410 static int __change_page_attr(unsigned long address
, struct cpa_data
*cpa
)
412 int level
, do_split
, err
;
413 struct page
*kpte_page
;
417 kpte
= lookup_address(address
, &level
);
421 kpte_page
= virt_to_page(kpte
);
422 BUG_ON(PageLRU(kpte_page
));
423 BUG_ON(PageCompound(kpte_page
));
425 if (level
== PG_LEVEL_4K
) {
426 pte_t new_pte
, old_pte
= *kpte
;
427 pgprot_t new_prot
= pte_pgprot(old_pte
);
429 if(!pte_val(old_pte
)) {
430 printk(KERN_WARNING
"CPA: called for zero pte. "
431 "vaddr = %lx cpa->vaddr = %lx\n", address
,
437 pgprot_val(new_prot
) &= ~pgprot_val(cpa
->mask_clr
);
438 pgprot_val(new_prot
) |= pgprot_val(cpa
->mask_set
);
440 new_prot
= static_protections(new_prot
, address
);
443 * We need to keep the pfn from the existing PTE,
444 * after all we're only going to change it's attributes
445 * not the memory it points to
447 new_pte
= pfn_pte(pte_pfn(old_pte
), canon_pgprot(new_prot
));
450 * Do we really change anything ?
452 if (pte_val(old_pte
) != pte_val(new_pte
)) {
453 set_pte_atomic(kpte
, new_pte
);
461 * Check, whether we can keep the large page intact
462 * and just change the pte:
464 do_split
= try_preserve_large_page(kpte
, address
, cpa
);
466 * When the range fits into the existing large page,
467 * return. cp->numpages and cpa->tlbflush have been updated in
474 * We have to split the large page:
476 err
= split_large_page(kpte
, address
);
486 * change_page_attr_addr - Change page table attributes in linear mapping
487 * @address: Virtual address in linear mapping.
488 * @prot: New page table attribute (PAGE_*)
490 * Change page attributes of a page in the direct mapping. This is a variant
491 * of change_page_attr() that also works on memory holes that do not have
492 * mem_map entry (pfn_valid() is false).
494 * See change_page_attr() documentation for more details.
496 * Modules and drivers should use the set_memory_* APIs instead.
498 static int change_page_attr_addr(struct cpa_data
*cpa
)
501 unsigned long address
= cpa
->vaddr
;
504 unsigned long phys_addr
= __pa(address
);
507 * If we are inside the high mapped kernel range, then we
508 * fixup the low mapping first. __va() returns the virtual
509 * address in the linear mapping:
511 if (within(address
, HIGH_MAP_START
, HIGH_MAP_END
))
512 address
= (unsigned long) __va(phys_addr
);
515 err
= __change_page_attr(address
, cpa
);
521 * If the physical address is inside the kernel map, we need
522 * to touch the high mapped kernel as well:
524 if (within(phys_addr
, 0, KERNEL_TEXT_SIZE
)) {
526 * Calc the high mapping address. See __phys_addr()
527 * for the non obvious details.
529 * Note that NX and other required permissions are
530 * checked in static_protections().
532 address
= phys_addr
+ HIGH_MAP_START
- phys_base
;
535 * Our high aliases are imprecise, because we check
536 * everything between 0 and KERNEL_TEXT_SIZE, so do
537 * not propagate lookup failures back to users:
539 __change_page_attr(address
, cpa
);
545 static int __change_page_attr_set_clr(struct cpa_data
*cpa
)
547 int ret
, numpages
= cpa
->numpages
;
551 * Store the remaining nr of pages for the large page
552 * preservation check.
554 cpa
->numpages
= numpages
;
555 ret
= change_page_attr_addr(cpa
);
560 * Adjust the number of pages with the result of the
561 * CPA operation. Either a large page has been
562 * preserved or a single page update happened.
564 BUG_ON(cpa
->numpages
> numpages
);
565 numpages
-= cpa
->numpages
;
566 cpa
->vaddr
+= cpa
->numpages
* PAGE_SIZE
;
571 static inline int cache_attr(pgprot_t attr
)
573 return pgprot_val(attr
) &
574 (_PAGE_PAT
| _PAGE_PAT_LARGE
| _PAGE_PWT
| _PAGE_PCD
);
577 static int change_page_attr_set_clr(unsigned long addr
, int numpages
,
578 pgprot_t mask_set
, pgprot_t mask_clr
)
584 * Check, if we are requested to change a not supported
587 mask_set
= canon_pgprot(mask_set
);
588 mask_clr
= canon_pgprot(mask_clr
);
589 if (!pgprot_val(mask_set
) && !pgprot_val(mask_clr
))
593 cpa
.numpages
= numpages
;
594 cpa
.mask_set
= mask_set
;
595 cpa
.mask_clr
= mask_clr
;
598 ret
= __change_page_attr_set_clr(&cpa
);
601 * Check whether we really changed something:
607 * No need to flush, when we did not set any of the caching
610 cache
= cache_attr(mask_set
);
613 * On success we use clflush, when the CPU supports it to
614 * avoid the wbindv. If the CPU does not support it and in the
615 * error case we fall back to cpa_flush_all (which uses
618 if (!ret
&& cpu_has_clflush
)
619 cpa_flush_range(addr
, numpages
, cache
);
621 cpa_flush_all(cache
);
626 static inline int change_page_attr_set(unsigned long addr
, int numpages
,
629 return change_page_attr_set_clr(addr
, numpages
, mask
, __pgprot(0));
632 static inline int change_page_attr_clear(unsigned long addr
, int numpages
,
635 return change_page_attr_set_clr(addr
, numpages
, __pgprot(0), mask
);
638 int set_memory_uc(unsigned long addr
, int numpages
)
640 return change_page_attr_set(addr
, numpages
,
641 __pgprot(_PAGE_PCD
| _PAGE_PWT
));
643 EXPORT_SYMBOL(set_memory_uc
);
645 int set_memory_wb(unsigned long addr
, int numpages
)
647 return change_page_attr_clear(addr
, numpages
,
648 __pgprot(_PAGE_PCD
| _PAGE_PWT
));
650 EXPORT_SYMBOL(set_memory_wb
);
652 int set_memory_x(unsigned long addr
, int numpages
)
654 return change_page_attr_clear(addr
, numpages
, __pgprot(_PAGE_NX
));
656 EXPORT_SYMBOL(set_memory_x
);
658 int set_memory_nx(unsigned long addr
, int numpages
)
660 return change_page_attr_set(addr
, numpages
, __pgprot(_PAGE_NX
));
662 EXPORT_SYMBOL(set_memory_nx
);
664 int set_memory_ro(unsigned long addr
, int numpages
)
666 return change_page_attr_clear(addr
, numpages
, __pgprot(_PAGE_RW
));
669 int set_memory_rw(unsigned long addr
, int numpages
)
671 return change_page_attr_set(addr
, numpages
, __pgprot(_PAGE_RW
));
674 int set_memory_np(unsigned long addr
, int numpages
)
676 return change_page_attr_clear(addr
, numpages
, __pgprot(_PAGE_PRESENT
));
679 int set_pages_uc(struct page
*page
, int numpages
)
681 unsigned long addr
= (unsigned long)page_address(page
);
683 return set_memory_uc(addr
, numpages
);
685 EXPORT_SYMBOL(set_pages_uc
);
687 int set_pages_wb(struct page
*page
, int numpages
)
689 unsigned long addr
= (unsigned long)page_address(page
);
691 return set_memory_wb(addr
, numpages
);
693 EXPORT_SYMBOL(set_pages_wb
);
695 int set_pages_x(struct page
*page
, int numpages
)
697 unsigned long addr
= (unsigned long)page_address(page
);
699 return set_memory_x(addr
, numpages
);
701 EXPORT_SYMBOL(set_pages_x
);
703 int set_pages_nx(struct page
*page
, int numpages
)
705 unsigned long addr
= (unsigned long)page_address(page
);
707 return set_memory_nx(addr
, numpages
);
709 EXPORT_SYMBOL(set_pages_nx
);
711 int set_pages_ro(struct page
*page
, int numpages
)
713 unsigned long addr
= (unsigned long)page_address(page
);
715 return set_memory_ro(addr
, numpages
);
718 int set_pages_rw(struct page
*page
, int numpages
)
720 unsigned long addr
= (unsigned long)page_address(page
);
722 return set_memory_rw(addr
, numpages
);
725 #ifdef CONFIG_DEBUG_PAGEALLOC
727 static int __set_pages_p(struct page
*page
, int numpages
)
729 struct cpa_data cpa
= { .vaddr
= (unsigned long) page_address(page
),
730 .numpages
= numpages
,
731 .mask_set
= __pgprot(_PAGE_PRESENT
| _PAGE_RW
),
732 .mask_clr
= __pgprot(0)};
734 return __change_page_attr_set_clr(&cpa
);
737 static int __set_pages_np(struct page
*page
, int numpages
)
739 struct cpa_data cpa
= { .vaddr
= (unsigned long) page_address(page
),
740 .numpages
= numpages
,
741 .mask_set
= __pgprot(0),
742 .mask_clr
= __pgprot(_PAGE_PRESENT
| _PAGE_RW
)};
744 return __change_page_attr_set_clr(&cpa
);
747 void kernel_map_pages(struct page
*page
, int numpages
, int enable
)
749 if (PageHighMem(page
))
752 debug_check_no_locks_freed(page_address(page
),
753 numpages
* PAGE_SIZE
);
757 * If page allocator is not up yet then do not call c_p_a():
759 if (!debug_pagealloc_enabled
)
763 * The return value is ignored - the calls cannot fail,
764 * large pages are disabled at boot time:
767 __set_pages_p(page
, numpages
);
769 __set_pages_np(page
, numpages
);
772 * We should perform an IPI and flush all tlbs,
773 * but that can deadlock->flush only current cpu:
780 * The testcases use internal knowledge of the implementation that shouldn't
781 * be exposed to the rest of the kernel. Include these directly here.
783 #ifdef CONFIG_CPA_DEBUG
784 #include "pageattr-test.c"