arch/arm/include/asm/tlb.h

   1 /*
   2  *  arch/arm/include/asm/tlb.h
   3  *
   4  *  Copyright (C) 2002 Russell King
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 as
   8  * published by the Free Software Foundation.
   9  *
  10  *  Experimentation shows that on a StrongARM, it appears to be faster
  11  *  to use the "invalidate whole tlb" rather than "invalidate single
  12  *  tlb" for this.
  13  *
  14  *  This appears true for both the process fork+exit case, as well as
  15  *  the munmap-large-area case.
  16  */
  17 #ifndef __ASMARM_TLB_H
  18 #define __ASMARM_TLB_H
  19
  20 #include <asm/cacheflush.h>
  21
  22 #ifndef CONFIG_MMU
  23
  24 #include <linux/pagemap.h>
  25
  26 #define tlb_flush(tlb)  ((void) tlb)
  27
  28 #include <asm-generic/tlb.h>
  29
  30 #else /* !CONFIG_MMU */
  31
  32 #include <linux/swap.h>
  33 #include <asm/pgalloc.h>
  34 #include <asm/tlbflush.h>
  35
  36 /*
  37  * We need to delay page freeing for SMP as other CPUs can access pages
  38  * which have been removed but not yet had their TLB entries invalidated.
  39  * Also, as ARMv7 speculative prefetch can drag new entries into the TLB,
  40  * we need to apply this same delaying tactic to ensure correct operation.
  41  */
  42 #if defined(CONFIG_SMP) || defined(CONFIG_CPU_32v7)
  43 #define tlb_fast_mode(tlb)      0
  44 #else
  45 #define tlb_fast_mode(tlb)      1
  46 #endif
  47
  48 #define MMU_GATHER_BUNDLE       8
  49
  50 /*
  51  * TLB handling.  This allows us to remove pages from the page
  52  * tables, and efficiently handle the TLB issues.
  53  */
  54 struct mmu_gather {
  55         struct mm_struct        *mm;
  56         unsigned int            fullmm;
  57         struct vm_area_struct   *vma;
  58         unsigned long           range_start;
  59         unsigned long           range_end;
  60         unsigned int            nr;
  61         unsigned int            max;
  62         struct page             **pages;
  63         struct page             *local[MMU_GATHER_BUNDLE];
  64 };
  65
  66 DECLARE_PER_CPU(struct mmu_gather, mmu_gathers);
  67
  68 /*
  69  * This is unnecessarily complex.  There's three ways the TLB shootdown
  70  * code is used:
  71  *  1. Unmapping a range of vmas.  See zap_page_range(), unmap_region().
  72  *     tlb->fullmm = 0, and tlb_start_vma/tlb_end_vma will be called.
  73  *     tlb->vma will be non-NULL.
  74  *  2. Unmapping all vmas.  See exit_mmap().
  75  *     tlb->fullmm = 1, and tlb_start_vma/tlb_end_vma will be called.
  76  *     tlb->vma will be non-NULL.  Additionally, page tables will be freed.
  77  *  3. Unmapping argument pages.  See shift_arg_pages().
  78  *     tlb->fullmm = 0, but tlb_start_vma/tlb_end_vma will not be called.
  79  *     tlb->vma will be NULL.
  80  */
  81 static inline void tlb_flush(struct mmu_gather *tlb)
  82 {
  83         if (tlb->fullmm || !tlb->vma)
  84                 flush_tlb_mm(tlb->mm);
  85         else if (tlb->range_end > 0) {
  86                 flush_tlb_range(tlb->vma, tlb->range_start, tlb->range_end);
  87                 tlb->range_start = TASK_SIZE;
  88                 tlb->range_end = 0;
  89         }
  90 }
  91
  92 static inline void tlb_add_flush(struct mmu_gather *tlb, unsigned long addr)
  93 {
  94         if (!tlb->fullmm) {
  95                 if (addr < tlb->range_start)
  96                         tlb->range_start = addr;
  97                 if (addr + PAGE_SIZE > tlb->range_end)
  98                         tlb->range_end = addr + PAGE_SIZE;
  99         }
 100 }
 101
 102 static inline void __tlb_alloc_page(struct mmu_gather *tlb)
 103 {
 104         unsigned long addr = __get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
 105
 106         if (addr) {
 107                 tlb->pages = (void *)addr;
 108                 tlb->max = PAGE_SIZE / sizeof(struct page *);
 109         }
 110 }
 111
 112 static inline void tlb_flush_mmu(struct mmu_gather *tlb)
 113 {
 114         tlb_flush(tlb);
 115         if (!tlb_fast_mode(tlb)) {
 116                 free_pages_and_swap_cache(tlb->pages, tlb->nr);
 117                 tlb->nr = 0;
 118                 if (tlb->pages == tlb->local)
 119                         __tlb_alloc_page(tlb);
 120         }
 121 }
 122
 123 static inline void
 124 tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, unsigned int fullmm)
 125 {
 126         tlb->mm = mm;
 127         tlb->fullmm = fullmm;
 128         tlb->vma = NULL;
 129         tlb->max = ARRAY_SIZE(tlb->local);
 130         tlb->pages = tlb->local;
 131         tlb->nr = 0;
 132         __tlb_alloc_page(tlb);
 133 }
 134
 135 static inline void
 136 tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
 137 {
 138         tlb_flush_mmu(tlb);
 139
 140         /* keep the page table cache within bounds */
 141         check_pgt_cache();
 142
 143         if (tlb->pages != tlb->local)
 144                 free_pages((unsigned long)tlb->pages, 0);
 145 }
 146
 147 /*
 148  * Memorize the range for the TLB flush.
 149  */
 150 static inline void
 151 tlb_remove_tlb_entry(struct mmu_gather *tlb, pte_t *ptep, unsigned long addr)
 152 {
 153         tlb_add_flush(tlb, addr);
 154 }
 155
 156 /*
 157  * In the case of tlb vma handling, we can optimise these away in the
 158  * case where we're doing a full MM flush.  When we're doing a munmap,
 159  * the vmas are adjusted to only cover the region to be torn down.
 160  */
 161 static inline void
 162 tlb_start_vma(struct mmu_gather *tlb, struct vm_area_struct *vma)
 163 {
 164         if (!tlb->fullmm) {
 165                 flush_cache_range(vma, vma->vm_start, vma->vm_end);
 166                 tlb->vma = vma;
 167                 tlb->range_start = TASK_SIZE;
 168                 tlb->range_end = 0;
 169         }
 170 }
 171
 172 static inline void
 173 tlb_end_vma(struct mmu_gather *tlb, struct vm_area_struct *vma)
 174 {
 175         if (!tlb->fullmm)
 176                 tlb_flush(tlb);
 177 }
 178
 179 static inline int __tlb_remove_page(struct mmu_gather *tlb, struct page *page)
 180 {
 181         if (tlb_fast_mode(tlb)) {
 182                 free_page_and_swap_cache(page);
 183                 return 1; /* avoid calling tlb_flush_mmu */
 184         }
 185
 186         tlb->pages[tlb->nr++] = page;
 187         VM_BUG_ON(tlb->nr > tlb->max);
 188         return tlb->max - tlb->nr;
 189 }
 190
 191 static inline void tlb_remove_page(struct mmu_gather *tlb, struct page *page)
 192 {
 193         if (!__tlb_remove_page(tlb, page))
 194                 tlb_flush_mmu(tlb);
 195 }
 196
 197 static inline void __pte_free_tlb(struct mmu_gather *tlb, pgtable_t pte,
 198         unsigned long addr)
 199 {
 200         pgtable_page_dtor(pte);
 201
 202 #ifdef CONFIG_ARM_LPAE
 203         tlb_add_flush(tlb, addr);
 204 #else
 205         /*
 206          * With the classic ARM MMU, a pte page has two corresponding pmd
 207          * entries, each covering 1MB.
 208          */
 209         addr &= PMD_MASK;
 210         tlb_add_flush(tlb, addr + SZ_1M - PAGE_SIZE);
 211         tlb_add_flush(tlb, addr + SZ_1M);
 212 #endif
 213
 214         tlb_remove_page(tlb, pte);
 215 }
 216
 217 static inline void __pmd_free_tlb(struct mmu_gather *tlb, pmd_t *pmdp,
 218                                   unsigned long addr)
 219 {
 220 #ifdef CONFIG_ARM_LPAE
 221         tlb_add_flush(tlb, addr);
 222         tlb_remove_page(tlb, virt_to_page(pmdp));
 223 #endif
 224 }
 225
 226 #define pte_free_tlb(tlb, ptep, addr)   __pte_free_tlb(tlb, ptep, addr)
 227 #define pmd_free_tlb(tlb, pmdp, addr)   __pmd_free_tlb(tlb, pmdp, addr)
 228 #define pud_free_tlb(tlb, pudp, addr)   pud_free((tlb)->mm, pudp)
 229
 230 #define tlb_migrate_finish(mm)          do { } while (0)
 231
 232 #endif /* CONFIG_MMU */
 233 #endif