3 #include <asm/pgalloc.h>
4 #include <asm/pgtable.h>
6 #include <asm/fixmap.h>
9 #define PGALLOC_GFP (GFP_KERNEL_ACCOUNT | __GFP_NOTRACK | __GFP_ZERO)
12 #define PGALLOC_USER_GFP __GFP_HIGHMEM
14 #define PGALLOC_USER_GFP 0
17 gfp_t __userpte_alloc_gfp
= PGALLOC_GFP
| PGALLOC_USER_GFP
;
19 pte_t
*pte_alloc_one_kernel(struct mm_struct
*mm
, unsigned long address
)
21 return (pte_t
*)__get_free_page(PGALLOC_GFP
& ~__GFP_ACCOUNT
);
24 pgtable_t
pte_alloc_one(struct mm_struct
*mm
, unsigned long address
)
28 pte
= alloc_pages(__userpte_alloc_gfp
, 0);
31 if (!pgtable_page_ctor(pte
)) {
38 static int __init
setup_userpte(char *arg
)
44 * "userpte=nohigh" disables allocation of user pagetables in
47 if (strcmp(arg
, "nohigh") == 0)
48 __userpte_alloc_gfp
&= ~__GFP_HIGHMEM
;
53 early_param("userpte", setup_userpte
);
55 void ___pte_free_tlb(struct mmu_gather
*tlb
, struct page
*pte
)
57 pgtable_page_dtor(pte
);
58 paravirt_release_pte(page_to_pfn(pte
));
59 tlb_remove_page(tlb
, pte
);
62 #if CONFIG_PGTABLE_LEVELS > 2
63 void ___pmd_free_tlb(struct mmu_gather
*tlb
, pmd_t
*pmd
)
65 struct page
*page
= virt_to_page(pmd
);
66 paravirt_release_pmd(__pa(pmd
) >> PAGE_SHIFT
);
68 * NOTE! For PAE, any changes to the top page-directory-pointer-table
69 * entries need a full cr3 reload to flush.
72 tlb
->need_flush_all
= 1;
74 pgtable_pmd_page_dtor(page
);
75 tlb_remove_page(tlb
, page
);
78 #if CONFIG_PGTABLE_LEVELS > 3
79 void ___pud_free_tlb(struct mmu_gather
*tlb
, pud_t
*pud
)
81 paravirt_release_pud(__pa(pud
) >> PAGE_SHIFT
);
82 tlb_remove_page(tlb
, virt_to_page(pud
));
84 #endif /* CONFIG_PGTABLE_LEVELS > 3 */
85 #endif /* CONFIG_PGTABLE_LEVELS > 2 */
87 static inline void pgd_list_add(pgd_t
*pgd
)
89 struct page
*page
= virt_to_page(pgd
);
91 list_add(&page
->lru
, &pgd_list
);
94 static inline void pgd_list_del(pgd_t
*pgd
)
96 struct page
*page
= virt_to_page(pgd
);
101 #define UNSHARED_PTRS_PER_PGD \
102 (SHARED_KERNEL_PMD ? KERNEL_PGD_BOUNDARY : PTRS_PER_PGD)
105 static void pgd_set_mm(pgd_t
*pgd
, struct mm_struct
*mm
)
107 BUILD_BUG_ON(sizeof(virt_to_page(pgd
)->index
) < sizeof(mm
));
108 virt_to_page(pgd
)->index
= (pgoff_t
)mm
;
111 struct mm_struct
*pgd_page_get_mm(struct page
*page
)
113 return (struct mm_struct
*)page
->index
;
116 static void pgd_ctor(struct mm_struct
*mm
, pgd_t
*pgd
)
118 /* If the pgd points to a shared pagetable level (either the
119 ptes in non-PAE, or shared PMD in PAE), then just copy the
120 references from swapper_pg_dir. */
121 if (CONFIG_PGTABLE_LEVELS
== 2 ||
122 (CONFIG_PGTABLE_LEVELS
== 3 && SHARED_KERNEL_PMD
) ||
123 CONFIG_PGTABLE_LEVELS
== 4) {
124 clone_pgd_range(pgd
+ KERNEL_PGD_BOUNDARY
,
125 swapper_pg_dir
+ KERNEL_PGD_BOUNDARY
,
129 /* list required to sync kernel mapping updates */
130 if (!SHARED_KERNEL_PMD
) {
136 static void pgd_dtor(pgd_t
*pgd
)
138 if (SHARED_KERNEL_PMD
)
141 spin_lock(&pgd_lock
);
143 spin_unlock(&pgd_lock
);
147 * List of all pgd's needed for non-PAE so it can invalidate entries
148 * in both cached and uncached pgd's; not needed for PAE since the
149 * kernel pmd is shared. If PAE were not to share the pmd a similar
150 * tactic would be needed. This is essentially codepath-based locking
151 * against pageattr.c; it is the unique case in which a valid change
152 * of kernel pagetables can't be lazily synchronized by vmalloc faults.
153 * vmalloc faults work because attached pagetables are never freed.
157 #ifdef CONFIG_X86_PAE
159 * In PAE mode, we need to do a cr3 reload (=tlb flush) when
160 * updating the top-level pagetable entries to guarantee the
161 * processor notices the update. Since this is expensive, and
162 * all 4 top-level entries are used almost immediately in a
163 * new process's life, we just pre-populate them here.
165 * Also, if we're in a paravirt environment where the kernel pmd is
166 * not shared between pagetables (!SHARED_KERNEL_PMDS), we allocate
167 * and initialize the kernel pmds here.
169 #define PREALLOCATED_PMDS UNSHARED_PTRS_PER_PGD
171 void pud_populate(struct mm_struct
*mm
, pud_t
*pudp
, pmd_t
*pmd
)
173 paravirt_alloc_pmd(mm
, __pa(pmd
) >> PAGE_SHIFT
);
175 /* Note: almost everything apart from _PAGE_PRESENT is
176 reserved at the pmd (PDPT) level. */
177 set_pud(pudp
, __pud(__pa(pmd
) | _PAGE_PRESENT
));
180 * According to Intel App note "TLBs, Paging-Structure Caches,
181 * and Their Invalidation", April 2007, document 317080-001,
182 * section 8.1: in PAE mode we explicitly have to flush the
183 * TLB via cr3 if the top-level pgd is changed...
187 #else /* !CONFIG_X86_PAE */
189 /* No need to prepopulate any pagetable entries in non-PAE modes. */
190 #define PREALLOCATED_PMDS 0
192 #endif /* CONFIG_X86_PAE */
194 static void free_pmds(struct mm_struct
*mm
, pmd_t
*pmds
[])
198 for(i
= 0; i
< PREALLOCATED_PMDS
; i
++)
200 pgtable_pmd_page_dtor(virt_to_page(pmds
[i
]));
201 free_page((unsigned long)pmds
[i
]);
206 static int preallocate_pmds(struct mm_struct
*mm
, pmd_t
*pmds
[])
210 gfp_t gfp
= PGALLOC_GFP
;
213 gfp
&= ~__GFP_ACCOUNT
;
215 for(i
= 0; i
< PREALLOCATED_PMDS
; i
++) {
216 pmd_t
*pmd
= (pmd_t
*)__get_free_page(gfp
);
219 if (pmd
&& !pgtable_pmd_page_ctor(virt_to_page(pmd
))) {
220 free_page((unsigned long)pmd
);
238 * Mop up any pmd pages which may still be attached to the pgd.
239 * Normally they will be freed by munmap/exit_mmap, but any pmd we
240 * preallocate which never got a corresponding vma will need to be
243 static void pgd_mop_up_pmds(struct mm_struct
*mm
, pgd_t
*pgdp
)
247 for(i
= 0; i
< PREALLOCATED_PMDS
; i
++) {
250 if (pgd_val(pgd
) != 0) {
251 pmd_t
*pmd
= (pmd_t
*)pgd_page_vaddr(pgd
);
253 pgdp
[i
] = native_make_pgd(0);
255 paravirt_release_pmd(pgd_val(pgd
) >> PAGE_SHIFT
);
262 static void pgd_prepopulate_pmd(struct mm_struct
*mm
, pgd_t
*pgd
, pmd_t
*pmds
[])
268 if (PREALLOCATED_PMDS
== 0) /* Work around gcc-3.4.x bug */
271 p4d
= p4d_offset(pgd
, 0);
272 pud
= pud_offset(p4d
, 0);
274 for (i
= 0; i
< PREALLOCATED_PMDS
; i
++, pud
++) {
275 pmd_t
*pmd
= pmds
[i
];
277 if (i
>= KERNEL_PGD_BOUNDARY
)
278 memcpy(pmd
, (pmd_t
*)pgd_page_vaddr(swapper_pg_dir
[i
]),
279 sizeof(pmd_t
) * PTRS_PER_PMD
);
281 pud_populate(mm
, pud
, pmd
);
286 * Xen paravirt assumes pgd table should be in one page. 64 bit kernel also
287 * assumes that pgd should be in one page.
289 * But kernel with PAE paging that is not running as a Xen domain
290 * only needs to allocate 32 bytes for pgd instead of one page.
292 #ifdef CONFIG_X86_PAE
294 #include <linux/slab.h>
296 #define PGD_SIZE (PTRS_PER_PGD * sizeof(pgd_t))
299 static struct kmem_cache
*pgd_cache
;
301 static int __init
pgd_cache_init(void)
304 * When PAE kernel is running as a Xen domain, it does not use
305 * shared kernel pmd. And this requires a whole page for pgd.
307 if (!SHARED_KERNEL_PMD
)
311 * when PAE kernel is not running as a Xen domain, it uses
312 * shared kernel pmd. Shared kernel pmd does not require a whole
313 * page for pgd. We are able to just allocate a 32-byte for pgd.
314 * During boot time, we create a 32-byte slab for pgd table allocation.
316 pgd_cache
= kmem_cache_create("pgd_cache", PGD_SIZE
, PGD_ALIGN
,
323 core_initcall(pgd_cache_init
);
325 static inline pgd_t
*_pgd_alloc(void)
328 * If no SHARED_KERNEL_PMD, PAE kernel is running as a Xen domain.
329 * We allocate one page for pgd.
331 if (!SHARED_KERNEL_PMD
)
332 return (pgd_t
*)__get_free_page(PGALLOC_GFP
);
335 * Now PAE kernel is not running as a Xen domain. We can allocate
336 * a 32-byte slab for pgd to save memory space.
338 return kmem_cache_alloc(pgd_cache
, PGALLOC_GFP
);
341 static inline void _pgd_free(pgd_t
*pgd
)
343 if (!SHARED_KERNEL_PMD
)
344 free_page((unsigned long)pgd
);
346 kmem_cache_free(pgd_cache
, pgd
);
349 static inline pgd_t
*_pgd_alloc(void)
351 return (pgd_t
*)__get_free_page(PGALLOC_GFP
);
354 static inline void _pgd_free(pgd_t
*pgd
)
356 free_page((unsigned long)pgd
);
358 #endif /* CONFIG_X86_PAE */
360 pgd_t
*pgd_alloc(struct mm_struct
*mm
)
363 pmd_t
*pmds
[PREALLOCATED_PMDS
];
372 if (preallocate_pmds(mm
, pmds
) != 0)
375 if (paravirt_pgd_alloc(mm
) != 0)
379 * Make sure that pre-populating the pmds is atomic with
380 * respect to anything walking the pgd_list, so that they
381 * never see a partially populated pgd.
383 spin_lock(&pgd_lock
);
386 pgd_prepopulate_pmd(mm
, pgd
, pmds
);
388 spin_unlock(&pgd_lock
);
400 void pgd_free(struct mm_struct
*mm
, pgd_t
*pgd
)
402 pgd_mop_up_pmds(mm
, pgd
);
404 paravirt_pgd_free(mm
, pgd
);
409 * Used to set accessed or dirty bits in the page table entries
410 * on other architectures. On x86, the accessed and dirty bits
411 * are tracked by hardware. However, do_wp_page calls this function
412 * to also make the pte writeable at the same time the dirty bit is
413 * set. In that case we do actually need to write the PTE.
415 int ptep_set_access_flags(struct vm_area_struct
*vma
,
416 unsigned long address
, pte_t
*ptep
,
417 pte_t entry
, int dirty
)
419 int changed
= !pte_same(*ptep
, entry
);
421 if (changed
&& dirty
) {
423 pte_update(vma
->vm_mm
, address
, ptep
);
429 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
430 int pmdp_set_access_flags(struct vm_area_struct
*vma
,
431 unsigned long address
, pmd_t
*pmdp
,
432 pmd_t entry
, int dirty
)
434 int changed
= !pmd_same(*pmdp
, entry
);
436 VM_BUG_ON(address
& ~HPAGE_PMD_MASK
);
438 if (changed
&& dirty
) {
441 * We had a write-protection fault here and changed the pmd
442 * to to more permissive. No need to flush the TLB for that,
443 * #PF is architecturally guaranteed to do that and in the
444 * worst-case we'll generate a spurious fault.
451 int pudp_set_access_flags(struct vm_area_struct
*vma
, unsigned long address
,
452 pud_t
*pudp
, pud_t entry
, int dirty
)
454 int changed
= !pud_same(*pudp
, entry
);
456 VM_BUG_ON(address
& ~HPAGE_PUD_MASK
);
458 if (changed
&& dirty
) {
461 * We had a write-protection fault here and changed the pud
462 * to to more permissive. No need to flush the TLB for that,
463 * #PF is architecturally guaranteed to do that and in the
464 * worst-case we'll generate a spurious fault.
472 int ptep_test_and_clear_young(struct vm_area_struct
*vma
,
473 unsigned long addr
, pte_t
*ptep
)
477 if (pte_young(*ptep
))
478 ret
= test_and_clear_bit(_PAGE_BIT_ACCESSED
,
479 (unsigned long *) &ptep
->pte
);
482 pte_update(vma
->vm_mm
, addr
, ptep
);
487 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
488 int pmdp_test_and_clear_young(struct vm_area_struct
*vma
,
489 unsigned long addr
, pmd_t
*pmdp
)
493 if (pmd_young(*pmdp
))
494 ret
= test_and_clear_bit(_PAGE_BIT_ACCESSED
,
495 (unsigned long *)pmdp
);
499 int pudp_test_and_clear_young(struct vm_area_struct
*vma
,
500 unsigned long addr
, pud_t
*pudp
)
504 if (pud_young(*pudp
))
505 ret
= test_and_clear_bit(_PAGE_BIT_ACCESSED
,
506 (unsigned long *)pudp
);
512 int ptep_clear_flush_young(struct vm_area_struct
*vma
,
513 unsigned long address
, pte_t
*ptep
)
516 * On x86 CPUs, clearing the accessed bit without a TLB flush
517 * doesn't cause data corruption. [ It could cause incorrect
518 * page aging and the (mistaken) reclaim of hot pages, but the
519 * chance of that should be relatively low. ]
521 * So as a performance optimization don't flush the TLB when
522 * clearing the accessed bit, it will eventually be flushed by
523 * a context switch or a VM operation anyway. [ In the rare
524 * event of it not getting flushed for a long time the delay
525 * shouldn't really matter because there's no real memory
526 * pressure for swapout to react to. ]
528 return ptep_test_and_clear_young(vma
, address
, ptep
);
531 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
532 int pmdp_clear_flush_young(struct vm_area_struct
*vma
,
533 unsigned long address
, pmd_t
*pmdp
)
537 VM_BUG_ON(address
& ~HPAGE_PMD_MASK
);
539 young
= pmdp_test_and_clear_young(vma
, address
, pmdp
);
541 flush_tlb_range(vma
, address
, address
+ HPAGE_PMD_SIZE
);
548 * reserve_top_address - reserves a hole in the top of kernel address space
549 * @reserve - size of hole to reserve
551 * Can be used to relocate the fixmap area and poke a hole in the top
552 * of kernel address space to make room for a hypervisor.
554 void __init
reserve_top_address(unsigned long reserve
)
557 BUG_ON(fixmaps_set
> 0);
558 __FIXADDR_TOP
= round_down(-reserve
, 1 << PMD_SHIFT
) - PAGE_SIZE
;
559 printk(KERN_INFO
"Reserving virtual address space above 0x%08lx (rounded to 0x%08lx)\n",
560 -reserve
, __FIXADDR_TOP
+ PAGE_SIZE
);
566 void __native_set_fixmap(enum fixed_addresses idx
, pte_t pte
)
568 unsigned long address
= __fix_to_virt(idx
);
570 if (idx
>= __end_of_fixed_addresses
) {
574 set_pte_vaddr(address
, pte
);
578 void native_set_fixmap(enum fixed_addresses idx
, phys_addr_t phys
,
581 __native_set_fixmap(idx
, pfn_pte(phys
>> PAGE_SHIFT
, flags
));
584 #ifdef CONFIG_HAVE_ARCH_HUGE_VMAP
586 * pud_set_huge - setup kernel PUD mapping
588 * MTRRs can override PAT memory types with 4KiB granularity. Therefore, this
589 * function sets up a huge page only if any of the following conditions are met:
591 * - MTRRs are disabled, or
593 * - MTRRs are enabled and the range is completely covered by a single MTRR, or
595 * - MTRRs are enabled and the corresponding MTRR memory type is WB, which
596 * has no effect on the requested PAT memory type.
598 * Callers should try to decrease page size (1GB -> 2MB -> 4K) if the bigger
599 * page mapping attempt fails.
601 * Returns 1 on success and 0 on failure.
603 int pud_set_huge(pud_t
*pud
, phys_addr_t addr
, pgprot_t prot
)
607 mtrr
= mtrr_type_lookup(addr
, addr
+ PUD_SIZE
, &uniform
);
608 if ((mtrr
!= MTRR_TYPE_INVALID
) && (!uniform
) &&
609 (mtrr
!= MTRR_TYPE_WRBACK
))
612 prot
= pgprot_4k_2_large(prot
);
614 set_pte((pte_t
*)pud
, pfn_pte(
615 (u64
)addr
>> PAGE_SHIFT
,
616 __pgprot(pgprot_val(prot
) | _PAGE_PSE
)));
622 * pmd_set_huge - setup kernel PMD mapping
624 * See text over pud_set_huge() above.
626 * Returns 1 on success and 0 on failure.
628 int pmd_set_huge(pmd_t
*pmd
, phys_addr_t addr
, pgprot_t prot
)
632 mtrr
= mtrr_type_lookup(addr
, addr
+ PMD_SIZE
, &uniform
);
633 if ((mtrr
!= MTRR_TYPE_INVALID
) && (!uniform
) &&
634 (mtrr
!= MTRR_TYPE_WRBACK
)) {
635 pr_warn_once("%s: Cannot satisfy [mem %#010llx-%#010llx] with a huge-page mapping due to MTRR override.\n",
636 __func__
, addr
, addr
+ PMD_SIZE
);
640 prot
= pgprot_4k_2_large(prot
);
642 set_pte((pte_t
*)pmd
, pfn_pte(
643 (u64
)addr
>> PAGE_SHIFT
,
644 __pgprot(pgprot_val(prot
) | _PAGE_PSE
)));
650 * pud_clear_huge - clear kernel PUD mapping when it is set
652 * Returns 1 on success and 0 on failure (no PUD map is found).
654 int pud_clear_huge(pud_t
*pud
)
656 if (pud_large(*pud
)) {
665 * pmd_clear_huge - clear kernel PMD mapping when it is set
667 * Returns 1 on success and 0 on failure (no PMD map is found).
669 int pmd_clear_huge(pmd_t
*pmd
)
671 if (pmd_large(*pmd
)) {
678 #endif /* CONFIG_HAVE_ARCH_HUGE_VMAP */