include/asm-ia64/tlb.h

   1 #ifndef _ASM_IA64_TLB_H
   2 #define _ASM_IA64_TLB_H
   3 /*
   4  * Based on <asm-generic/tlb.h>.
   5  *
   6  * Copyright (C) 2002-2003 Hewlett-Packard Co
   7  *      David Mosberger-Tang <davidm@hpl.hp.com>
   8  */
   9 /*
  10  * Removing a translation from a page table (including TLB-shootdown) is a four-step
  11  * procedure:
  12  *
  13  *      (1) Flush (virtual) caches --- ensures virtual memory is coherent with kernel memory
  14  *          (this is a no-op on ia64).
  15  *      (2) Clear the relevant portions of the page-table
  16  *      (3) Flush the TLBs --- ensures that stale content is gone from CPU TLBs
  17  *      (4) Release the pages that were freed up in step (2).
  18  *
  19  * Note that the ordering of these steps is crucial to avoid races on MP machines.
  20  *
  21  * The Linux kernel defines several platform-specific hooks for TLB-shootdown.  When
  22  * unmapping a portion of the virtual address space, these hooks are called according to
  23  * the following template:
  24  *
  25  *      tlb <- tlb_gather_mmu(mm, full_mm_flush);       // start unmap for address space MM
  26  *      {
  27  *        for each vma that needs a shootdown do {
  28  *          tlb_start_vma(tlb, vma);
  29  *            for each page-table-entry PTE that needs to be removed do {
  30  *              tlb_remove_tlb_entry(tlb, pte, address);
  31  *              if (pte refers to a normal page) {
  32  *                tlb_remove_page(tlb, page);
  33  *              }
  34  *            }
  35  *          tlb_end_vma(tlb, vma);
  36  *        }
  37  *      }
  38  *      tlb_finish_mmu(tlb, start, end);        // finish unmap for address space MM
  39  */
  40 #include <linux/mm.h>
  41 #include <linux/pagemap.h>
  42 #include <linux/swap.h>
  43
  44 #include <asm/pgalloc.h>
  45 #include <asm/processor.h>
  46 #include <asm/tlbflush.h>
  47 #include <asm/machvec.h>
  48
  49 #ifdef CONFIG_SMP
  50 # define FREE_PTE_NR            2048
  51 # define tlb_fast_mode(tlb)     ((tlb)->nr == ~0U)
  52 #else
  53 # define FREE_PTE_NR            0
  54 # define tlb_fast_mode(tlb)     (1)
  55 #endif
  56
  57 struct mmu_gather {
  58         struct mm_struct        *mm;
  59         unsigned int            nr;             /* == ~0U => fast mode */
  60         unsigned char           fullmm;         /* non-zero means full mm flush */
  61         unsigned char           need_flush;     /* really unmapped some PTEs? */
  62         unsigned long           start_addr;
  63         unsigned long           end_addr;
  64         struct page             *pages[FREE_PTE_NR];
  65 };
  66
  67 /* Users of the generic TLB shootdown code must declare this storage space. */
  68 DECLARE_PER_CPU(struct mmu_gather, mmu_gathers);
  69
  70 /*
  71  * Flush the TLB for address range START to END and, if not in fast mode, release the
  72  * freed pages that where gathered up to this point.
  73  */
  74 static inline void
  75 ia64_tlb_flush_mmu (struct mmu_gather *tlb, unsigned long start, unsigned long end)
  76 {
  77         unsigned int nr;
  78
  79         if (!tlb->need_flush)
  80                 return;
  81         tlb->need_flush = 0;
  82
  83         if (tlb->fullmm) {
  84                 /*
  85                  * Tearing down the entire address space.  This happens both as a result
  86                  * of exit() and execve().  The latter case necessitates the call to
  87                  * flush_tlb_mm() here.
  88                  */
  89                 flush_tlb_mm(tlb->mm);
  90         } else if (unlikely (end - start >= 1024*1024*1024*1024UL
  91                              || REGION_NUMBER(start) != REGION_NUMBER(end - 1)))
  92         {
  93                 /*
  94                  * If we flush more than a tera-byte or across regions, we're probably
  95                  * better off just flushing the entire TLB(s).  This should be very rare
  96                  * and is not worth optimizing for.
  97                  */
  98                 flush_tlb_all();
  99         } else {
 100                 /*
 101                  * XXX fix me: flush_tlb_range() should take an mm pointer instead of a
 102                  * vma pointer.
 103                  */
 104                 struct vm_area_struct vma;
 105
 106                 vma.vm_mm = tlb->mm;
 107                 /* flush the address range from the tlb: */
 108                 flush_tlb_range(&vma, start, end);
 109                 /* now flush the virt. page-table area mapping the address range: */
 110                 flush_tlb_range(&vma, ia64_thash(start), ia64_thash(end));
 111         }
 112
 113         /* lastly, release the freed pages */
 114         nr = tlb->nr;
 115         if (!tlb_fast_mode(tlb)) {
 116                 unsigned long i;
 117                 tlb->nr = 0;
 118                 tlb->start_addr = ~0UL;
 119                 for (i = 0; i < nr; ++i)
 120                         free_page_and_swap_cache(tlb->pages[i]);
 121         }
 122 }
 123
 124 /*
 125  * Return a pointer to an initialized struct mmu_gather.
 126  */
 127 static inline struct mmu_gather *
 128 tlb_gather_mmu (struct mm_struct *mm, unsigned int full_mm_flush)
 129 {
 130         struct mmu_gather *tlb = &get_cpu_var(mmu_gathers);
 131
 132         tlb->mm = mm;
 133         /*
 134          * Use fast mode if only 1 CPU is online.
 135          *
 136          * It would be tempting to turn on fast-mode for full_mm_flush as well.  But this
 137          * doesn't work because of speculative accesses and software prefetching: the page
 138          * table of "mm" may (and usually is) the currently active page table and even
 139          * though the kernel won't do any user-space accesses during the TLB shoot down, a
 140          * compiler might use speculation or lfetch.fault on what happens to be a valid
 141          * user-space address.  This in turn could trigger a TLB miss fault (or a VHPT
 142          * walk) and re-insert a TLB entry we just removed.  Slow mode avoids such
 143          * problems.  (We could make fast-mode work by switching the current task to a
 144          * different "mm" during the shootdown.) --davidm 08/02/2002
 145          */
 146         tlb->nr = (num_online_cpus() == 1) ? ~0U : 0;
 147         tlb->fullmm = full_mm_flush;
 148         tlb->start_addr = ~0UL;
 149         return tlb;
 150 }
 151
 152 /*
 153  * Called at the end of the shootdown operation to free up any resources that were
 154  * collected.
 155  */
 156 static inline void
 157 tlb_finish_mmu (struct mmu_gather *tlb, unsigned long start, unsigned long end)
 158 {
 159         /*
 160          * Note: tlb->nr may be 0 at this point, so we can't rely on tlb->start_addr and
 161          * tlb->end_addr.
 162          */
 163         ia64_tlb_flush_mmu(tlb, start, end);
 164
 165         /* keep the page table cache within bounds */
 166         check_pgt_cache();
 167
 168         put_cpu_var(mmu_gathers);
 169 }
 170
 171 /*
 172  * Logically, this routine frees PAGE.  On MP machines, the actual freeing of the page
 173  * must be delayed until after the TLB has been flushed (see comments at the beginning of
 174  * this file).
 175  */
 176 static inline void
 177 tlb_remove_page (struct mmu_gather *tlb, struct page *page)
 178 {
 179         tlb->need_flush = 1;
 180
 181         if (tlb_fast_mode(tlb)) {
 182                 free_page_and_swap_cache(page);
 183                 return;
 184         }
 185         tlb->pages[tlb->nr++] = page;
 186         if (tlb->nr >= FREE_PTE_NR)
 187                 ia64_tlb_flush_mmu(tlb, tlb->start_addr, tlb->end_addr);
 188 }
 189
 190 /*
 191  * Remove TLB entry for PTE mapped at virtual address ADDRESS.  This is called for any
 192  * PTE, not just those pointing to (normal) physical memory.
 193  */
 194 static inline void
 195 __tlb_remove_tlb_entry (struct mmu_gather *tlb, pte_t *ptep, unsigned long address)
 196 {
 197         if (tlb->start_addr == ~0UL)
 198                 tlb->start_addr = address;
 199         tlb->end_addr = address + PAGE_SIZE;
 200 }
 201
 202 #define tlb_migrate_finish(mm)  platform_tlb_migrate_finish(mm)
 203
 204 #define tlb_start_vma(tlb, vma)                 do { } while (0)
 205 #define tlb_end_vma(tlb, vma)                   do { } while (0)
 206
 207 #define tlb_remove_tlb_entry(tlb, ptep, addr)           \
 208 do {                                                    \
 209         tlb->need_flush = 1;                            \
 210         __tlb_remove_tlb_entry(tlb, ptep, addr);        \
 211 } while (0)
 212
 213 #define pte_free_tlb(tlb, ptep)                         \
 214 do {                                                    \
 215         tlb->need_flush = 1;                            \
 216         __pte_free_tlb(tlb, ptep);                      \
 217 } while (0)
 218
 219 #define pmd_free_tlb(tlb, ptep)                         \
 220 do {                                                    \
 221         tlb->need_flush = 1;                            \
 222         __pmd_free_tlb(tlb, ptep);                      \
 223 } while (0)
 224
 225 #define pud_free_tlb(tlb, pudp)                         \
 226 do {                                                    \
 227         tlb->need_flush = 1;                            \
 228         __pud_free_tlb(tlb, pudp);                      \
 229 } while (0)
 230
 231 #endif /* _ASM_IA64_TLB_H */