3 #include <asm/pgalloc.h>
4 #include <asm/pgtable.h>
6 #include <asm/fixmap.h>
8 #define PGALLOC_GFP GFP_KERNEL | __GFP_NOTRACK | __GFP_REPEAT | __GFP_ZERO
11 #define PGALLOC_USER_GFP __GFP_HIGHMEM
13 #define PGALLOC_USER_GFP 0
16 gfp_t __userpte_alloc_gfp = PGALLOC_GFP | PGALLOC_USER_GFP;
18 pte_t *pte_alloc_one_kernel(struct mm_struct *mm, unsigned long address)
20 return (pte_t *)__get_free_page(PGALLOC_GFP);
23 pgtable_t pte_alloc_one(struct mm_struct *mm, unsigned long address)
27 pte = alloc_pages(__userpte_alloc_gfp, 0);
30 if (!pgtable_page_ctor(pte)) {
37 static int __init setup_userpte(char *arg)
43 * "userpte=nohigh" disables allocation of user pagetables in
46 if (strcmp(arg, "nohigh") == 0)
47 __userpte_alloc_gfp &= ~__GFP_HIGHMEM;
52 early_param("userpte", setup_userpte);
54 void ___pte_free_tlb(struct mmu_gather *tlb, struct page *pte)
56 pgtable_page_dtor(pte);
57 paravirt_release_pte(page_to_pfn(pte));
58 tlb_remove_page(tlb, pte);
61 #if CONFIG_PGTABLE_LEVELS > 2
62 void ___pmd_free_tlb(struct mmu_gather *tlb, pmd_t *pmd)
64 struct page *page = virt_to_page(pmd);
65 paravirt_release_pmd(__pa(pmd) >> PAGE_SHIFT);
67 * NOTE! For PAE, any changes to the top page-directory-pointer-table
68 * entries need a full cr3 reload to flush.
71 tlb->need_flush_all = 1;
73 pgtable_pmd_page_dtor(page);
74 tlb_remove_page(tlb, page);
77 #if CONFIG_PGTABLE_LEVELS > 3
78 void ___pud_free_tlb(struct mmu_gather *tlb, pud_t *pud)
80 paravirt_release_pud(__pa(pud) >> PAGE_SHIFT);
81 tlb_remove_page(tlb, virt_to_page(pud));
83 #endif /* CONFIG_PGTABLE_LEVELS > 3 */
84 #endif /* CONFIG_PGTABLE_LEVELS > 2 */
86 static inline void pgd_list_add(pgd_t *pgd)
88 struct page *page = virt_to_page(pgd);
90 list_add(&page->lru, &pgd_list);
93 static inline void pgd_list_del(pgd_t *pgd)
95 struct page *page = virt_to_page(pgd);
100 #define UNSHARED_PTRS_PER_PGD \
101 (SHARED_KERNEL_PMD ? KERNEL_PGD_BOUNDARY : PTRS_PER_PGD)
104 static void pgd_set_mm(pgd_t *pgd, struct mm_struct *mm)
106 BUILD_BUG_ON(sizeof(virt_to_page(pgd)->index) < sizeof(mm));
107 virt_to_page(pgd)->index = (pgoff_t)mm;
110 struct mm_struct *pgd_page_get_mm(struct page *page)
112 return (struct mm_struct *)page->index;
115 static void pgd_ctor(struct mm_struct *mm, pgd_t *pgd)
117 /* If the pgd points to a shared pagetable level (either the
118 ptes in non-PAE, or shared PMD in PAE), then just copy the
119 references from swapper_pg_dir. */
120 if (CONFIG_PGTABLE_LEVELS == 2 ||
121 (CONFIG_PGTABLE_LEVELS == 3 && SHARED_KERNEL_PMD) ||
122 CONFIG_PGTABLE_LEVELS == 4) {
123 clone_pgd_range(pgd + KERNEL_PGD_BOUNDARY,
124 swapper_pg_dir + KERNEL_PGD_BOUNDARY,
128 /* list required to sync kernel mapping updates */
129 if (!SHARED_KERNEL_PMD) {
135 static void pgd_dtor(pgd_t *pgd)
137 if (SHARED_KERNEL_PMD)
140 spin_lock(&pgd_lock);
142 spin_unlock(&pgd_lock);
146 * List of all pgd's needed for non-PAE so it can invalidate entries
147 * in both cached and uncached pgd's; not needed for PAE since the
148 * kernel pmd is shared. If PAE were not to share the pmd a similar
149 * tactic would be needed. This is essentially codepath-based locking
150 * against pageattr.c; it is the unique case in which a valid change
151 * of kernel pagetables can't be lazily synchronized by vmalloc faults.
152 * vmalloc faults work because attached pagetables are never freed.
156 #ifdef CONFIG_X86_PAE
158 * In PAE mode, we need to do a cr3 reload (=tlb flush) when
159 * updating the top-level pagetable entries to guarantee the
160 * processor notices the update. Since this is expensive, and
161 * all 4 top-level entries are used almost immediately in a
162 * new process's life, we just pre-populate them here.
164 * Also, if we're in a paravirt environment where the kernel pmd is
165 * not shared between pagetables (!SHARED_KERNEL_PMDS), we allocate
166 * and initialize the kernel pmds here.
168 #define PREALLOCATED_PMDS UNSHARED_PTRS_PER_PGD
170 void pud_populate(struct mm_struct *mm, pud_t *pudp, pmd_t *pmd)
172 paravirt_alloc_pmd(mm, __pa(pmd) >> PAGE_SHIFT);
174 /* Note: almost everything apart from _PAGE_PRESENT is
175 reserved at the pmd (PDPT) level. */
176 set_pud(pudp, __pud(__pa(pmd) | _PAGE_PRESENT));
179 * According to Intel App note "TLBs, Paging-Structure Caches,
180 * and Their Invalidation", April 2007, document 317080-001,
181 * section 8.1: in PAE mode we explicitly have to flush the
182 * TLB via cr3 if the top-level pgd is changed...
186 #else /* !CONFIG_X86_PAE */
188 /* No need to prepopulate any pagetable entries in non-PAE modes. */
189 #define PREALLOCATED_PMDS 0
191 #endif /* CONFIG_X86_PAE */
193 static void free_pmds(struct mm_struct *mm, pmd_t *pmds[])
197 for(i = 0; i < PREALLOCATED_PMDS; i++)
199 pgtable_pmd_page_dtor(virt_to_page(pmds[i]));
200 free_page((unsigned long)pmds[i]);
205 static int preallocate_pmds(struct mm_struct *mm, pmd_t *pmds[])
210 for(i = 0; i < PREALLOCATED_PMDS; i++) {
211 pmd_t *pmd = (pmd_t *)__get_free_page(PGALLOC_GFP);
214 if (pmd && !pgtable_pmd_page_ctor(virt_to_page(pmd))) {
215 free_page((unsigned long)pmd);
233 * Mop up any pmd pages which may still be attached to the pgd.
234 * Normally they will be freed by munmap/exit_mmap, but any pmd we
235 * preallocate which never got a corresponding vma will need to be
238 static void pgd_mop_up_pmds(struct mm_struct *mm, pgd_t *pgdp)
242 for(i = 0; i < PREALLOCATED_PMDS; i++) {
245 if (pgd_val(pgd) != 0) {
246 pmd_t *pmd = (pmd_t *)pgd_page_vaddr(pgd);
248 pgdp[i] = native_make_pgd(0);
250 paravirt_release_pmd(pgd_val(pgd) >> PAGE_SHIFT);
257 static void pgd_prepopulate_pmd(struct mm_struct *mm, pgd_t *pgd, pmd_t *pmds[])
262 if (PREALLOCATED_PMDS == 0) /* Work around gcc-3.4.x bug */
265 pud = pud_offset(pgd, 0);
267 for (i = 0; i < PREALLOCATED_PMDS; i++, pud++) {
268 pmd_t *pmd = pmds[i];
270 if (i >= KERNEL_PGD_BOUNDARY)
271 memcpy(pmd, (pmd_t *)pgd_page_vaddr(swapper_pg_dir[i]),
272 sizeof(pmd_t) * PTRS_PER_PMD);
274 pud_populate(mm, pud, pmd);
279 * Xen paravirt assumes pgd table should be in one page. 64 bit kernel also
280 * assumes that pgd should be in one page.
282 * But kernel with PAE paging that is not running as a Xen domain
283 * only needs to allocate 32 bytes for pgd instead of one page.
285 #ifdef CONFIG_X86_PAE
287 #include <linux/slab.h>
289 #define PGD_SIZE (PTRS_PER_PGD * sizeof(pgd_t))
292 static struct kmem_cache *pgd_cache;
294 static int __init pgd_cache_init(void)
297 * When PAE kernel is running as a Xen domain, it does not use
298 * shared kernel pmd. And this requires a whole page for pgd.
300 if (!SHARED_KERNEL_PMD)
304 * when PAE kernel is not running as a Xen domain, it uses
305 * shared kernel pmd. Shared kernel pmd does not require a whole
306 * page for pgd. We are able to just allocate a 32-byte for pgd.
307 * During boot time, we create a 32-byte slab for pgd table allocation.
309 pgd_cache = kmem_cache_create("pgd_cache", PGD_SIZE, PGD_ALIGN,
316 core_initcall(pgd_cache_init);
318 static inline pgd_t *_pgd_alloc(void)
321 * If no SHARED_KERNEL_PMD, PAE kernel is running as a Xen domain.
322 * We allocate one page for pgd.
324 if (!SHARED_KERNEL_PMD)
325 return (pgd_t *)__get_free_page(PGALLOC_GFP);
328 * Now PAE kernel is not running as a Xen domain. We can allocate
329 * a 32-byte slab for pgd to save memory space.
331 return kmem_cache_alloc(pgd_cache, PGALLOC_GFP);
334 static inline void _pgd_free(pgd_t *pgd)
336 if (!SHARED_KERNEL_PMD)
337 free_page((unsigned long)pgd);
339 kmem_cache_free(pgd_cache, pgd);
342 static inline pgd_t *_pgd_alloc(void)
344 return (pgd_t *)__get_free_page(PGALLOC_GFP);
347 static inline void _pgd_free(pgd_t *pgd)
349 free_page((unsigned long)pgd);
351 #endif /* CONFIG_X86_PAE */
353 pgd_t *pgd_alloc(struct mm_struct *mm)
356 pmd_t *pmds[PREALLOCATED_PMDS];
365 if (preallocate_pmds(mm, pmds) != 0)
368 if (paravirt_pgd_alloc(mm) != 0)
372 * Make sure that pre-populating the pmds is atomic with
373 * respect to anything walking the pgd_list, so that they
374 * never see a partially populated pgd.
376 spin_lock(&pgd_lock);
379 pgd_prepopulate_pmd(mm, pgd, pmds);
381 spin_unlock(&pgd_lock);
393 void pgd_free(struct mm_struct *mm, pgd_t *pgd)
395 pgd_mop_up_pmds(mm, pgd);
397 paravirt_pgd_free(mm, pgd);
402 * Used to set accessed or dirty bits in the page table entries
403 * on other architectures. On x86, the accessed and dirty bits
404 * are tracked by hardware. However, do_wp_page calls this function
405 * to also make the pte writeable at the same time the dirty bit is
406 * set. In that case we do actually need to write the PTE.
408 int ptep_set_access_flags(struct vm_area_struct *vma,
409 unsigned long address, pte_t *ptep,
410 pte_t entry, int dirty)
412 int changed = !pte_same(*ptep, entry);
414 if (changed && dirty) {
416 pte_update_defer(vma->vm_mm, address, ptep);
422 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
423 int pmdp_set_access_flags(struct vm_area_struct *vma,
424 unsigned long address, pmd_t *pmdp,
425 pmd_t entry, int dirty)
427 int changed = !pmd_same(*pmdp, entry);
429 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
431 if (changed && dirty) {
433 pmd_update_defer(vma->vm_mm, address, pmdp);
435 * We had a write-protection fault here and changed the pmd
436 * to to more permissive. No need to flush the TLB for that,
437 * #PF is architecturally guaranteed to do that and in the
438 * worst-case we'll generate a spurious fault.
446 int ptep_test_and_clear_young(struct vm_area_struct *vma,
447 unsigned long addr, pte_t *ptep)
451 if (pte_young(*ptep))
452 ret = test_and_clear_bit(_PAGE_BIT_ACCESSED,
453 (unsigned long *) &ptep->pte);
456 pte_update(vma->vm_mm, addr, ptep);
461 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
462 int pmdp_test_and_clear_young(struct vm_area_struct *vma,
463 unsigned long addr, pmd_t *pmdp)
467 if (pmd_young(*pmdp))
468 ret = test_and_clear_bit(_PAGE_BIT_ACCESSED,
469 (unsigned long *)pmdp);
472 pmd_update(vma->vm_mm, addr, pmdp);
478 int ptep_clear_flush_young(struct vm_area_struct *vma,
479 unsigned long address, pte_t *ptep)
482 * On x86 CPUs, clearing the accessed bit without a TLB flush
483 * doesn't cause data corruption. [ It could cause incorrect
484 * page aging and the (mistaken) reclaim of hot pages, but the
485 * chance of that should be relatively low. ]
487 * So as a performance optimization don't flush the TLB when
488 * clearing the accessed bit, it will eventually be flushed by
489 * a context switch or a VM operation anyway. [ In the rare
490 * event of it not getting flushed for a long time the delay
491 * shouldn't really matter because there's no real memory
492 * pressure for swapout to react to. ]
494 return ptep_test_and_clear_young(vma, address, ptep);
497 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
498 int pmdp_clear_flush_young(struct vm_area_struct *vma,
499 unsigned long address, pmd_t *pmdp)
503 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
505 young = pmdp_test_and_clear_young(vma, address, pmdp);
507 flush_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
512 void pmdp_splitting_flush(struct vm_area_struct *vma,
513 unsigned long address, pmd_t *pmdp)
516 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
517 set = !test_and_set_bit(_PAGE_BIT_SPLITTING,
518 (unsigned long *)pmdp);
520 pmd_update(vma->vm_mm, address, pmdp);
521 /* need tlb flush only to serialize against gup-fast */
522 flush_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
528 * reserve_top_address - reserves a hole in the top of kernel address space
529 * @reserve - size of hole to reserve
531 * Can be used to relocate the fixmap area and poke a hole in the top
532 * of kernel address space to make room for a hypervisor.
534 void __init reserve_top_address(unsigned long reserve)
537 BUG_ON(fixmaps_set > 0);
538 __FIXADDR_TOP = round_down(-reserve, 1 << PMD_SHIFT) - PAGE_SIZE;
539 printk(KERN_INFO "Reserving virtual address space above 0x%08lx (rounded to 0x%08lx)\n",
540 -reserve, __FIXADDR_TOP + PAGE_SIZE);
546 void __native_set_fixmap(enum fixed_addresses idx, pte_t pte)
548 unsigned long address = __fix_to_virt(idx);
550 if (idx >= __end_of_fixed_addresses) {
554 set_pte_vaddr(address, pte);
558 void native_set_fixmap(enum fixed_addresses idx, phys_addr_t phys,
561 __native_set_fixmap(idx, pfn_pte(phys >> PAGE_SHIFT, flags));