arch/sparc/mm/tlb.c

   1 /* arch/sparc64/mm/tlb.c
   2  *
   3  * Copyright (C) 2004 David S. Miller <davem@redhat.com>
   4  */
   5
   6 #include <linux/kernel.h>
   7 #include <linux/percpu.h>
   8 #include <linux/mm.h>
   9 #include <linux/swap.h>
  10 #include <linux/preempt.h>
  11
  12 #include <asm/pgtable.h>
  13 #include <asm/pgalloc.h>
  14 #include <asm/tlbflush.h>
  15 #include <asm/cacheflush.h>
  16 #include <asm/mmu_context.h>
  17 #include <asm/tlb.h>
  18
  19 /* Heavily inspired by the ppc64 code.  */
  20
  21 static DEFINE_PER_CPU(struct tlb_batch, tlb_batch);
  22
  23 void flush_tlb_pending(void)
  24 {
  25         struct tlb_batch *tb = &get_cpu_var(tlb_batch);
  26         struct mm_struct *mm = tb->mm;
  27
  28         if (!tb->tlb_nr)
  29                 goto out;
  30
  31         flush_tsb_user(tb);
  32
  33         if (CTX_VALID(mm->context)) {
  34                 if (tb->tlb_nr == 1) {
  35                         global_flush_tlb_page(mm, tb->vaddrs[0]);
  36                 } else {
  37 #ifdef CONFIG_SMP
  38                         smp_flush_tlb_pending(tb->mm, tb->tlb_nr,
  39                                               &tb->vaddrs[0]);
  40 #else
  41                         __flush_tlb_pending(CTX_HWBITS(tb->mm->context),
  42                                             tb->tlb_nr, &tb->vaddrs[0]);
  43 #endif
  44                 }
  45         }
  46
  47         tb->tlb_nr = 0;
  48
  49 out:
  50         put_cpu_var(tlb_batch);
  51 }
  52
  53 void arch_enter_lazy_mmu_mode(void)
  54 {
  55         struct tlb_batch *tb = this_cpu_ptr(&tlb_batch);
  56
  57         tb->active = 1;
  58 }
  59
  60 void arch_leave_lazy_mmu_mode(void)
  61 {
  62         struct tlb_batch *tb = this_cpu_ptr(&tlb_batch);
  63
  64         if (tb->tlb_nr)
  65                 flush_tlb_pending();
  66         tb->active = 0;
  67 }
  68
  69 static void tlb_batch_add_one(struct mm_struct *mm, unsigned long vaddr,
  70                               bool exec, bool huge)
  71 {
  72         struct tlb_batch *tb = &get_cpu_var(tlb_batch);
  73         unsigned long nr;
  74
  75         vaddr &= PAGE_MASK;
  76         if (exec)
  77                 vaddr |= 0x1UL;
  78
  79         nr = tb->tlb_nr;
  80
  81         if (unlikely(nr != 0 && mm != tb->mm)) {
  82                 flush_tlb_pending();
  83                 nr = 0;
  84         }
  85
  86         if (!tb->active) {
  87                 flush_tsb_user_page(mm, vaddr, huge);
  88                 global_flush_tlb_page(mm, vaddr);
  89                 goto out;
  90         }
  91
  92         if (nr == 0) {
  93                 tb->mm = mm;
  94                 tb->huge = huge;
  95         }
  96
  97         if (tb->huge != huge) {
  98                 flush_tlb_pending();
  99                 tb->huge = huge;
 100                 nr = 0;
 101         }
 102
 103         tb->vaddrs[nr] = vaddr;
 104         tb->tlb_nr = ++nr;
 105         if (nr >= TLB_BATCH_NR)
 106                 flush_tlb_pending();
 107
 108 out:
 109         put_cpu_var(tlb_batch);
 110 }
 111
 112 void tlb_batch_add(struct mm_struct *mm, unsigned long vaddr,
 113                    pte_t *ptep, pte_t orig, int fullmm)
 114 {
 115         bool huge = is_hugetlb_pte(orig);
 116
 117         if (tlb_type != hypervisor &&
 118             pte_dirty(orig)) {
 119                 unsigned long paddr, pfn = pte_pfn(orig);
 120                 struct address_space *mapping;
 121                 struct page *page;
 122
 123                 if (!pfn_valid(pfn))
 124                         goto no_cache_flush;
 125
 126                 page = pfn_to_page(pfn);
 127                 if (PageReserved(page))
 128                         goto no_cache_flush;
 129
 130                 /* A real file page? */
 131                 mapping = page_mapping(page);
 132                 if (!mapping)
 133                         goto no_cache_flush;
 134
 135                 paddr = (unsigned long) page_address(page);
 136                 if ((paddr ^ vaddr) & (1 << 13))
 137                         flush_dcache_page_all(mm, page);
 138         }
 139
 140 no_cache_flush:
 141         if (!fullmm)
 142                 tlb_batch_add_one(mm, vaddr, pte_exec(orig), huge);
 143 }
 144
 145 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 146 static void tlb_batch_pmd_scan(struct mm_struct *mm, unsigned long vaddr,
 147                                pmd_t pmd)
 148 {
 149         unsigned long end;
 150         pte_t *pte;
 151
 152         pte = pte_offset_map(&pmd, vaddr);
 153         end = vaddr + HPAGE_SIZE;
 154         while (vaddr < end) {
 155                 if (pte_val(*pte) & _PAGE_VALID) {
 156                         bool exec = pte_exec(*pte);
 157
 158                         tlb_batch_add_one(mm, vaddr, exec, false);
 159                 }
 160                 pte++;
 161                 vaddr += PAGE_SIZE;
 162         }
 163         pte_unmap(pte);
 164 }
 165
 166
 167 static void __set_pmd_acct(struct mm_struct *mm, unsigned long addr,
 168                            pmd_t orig, pmd_t pmd)
 169 {
 170         if (mm == &init_mm)
 171                 return;
 172
 173         if ((pmd_val(pmd) ^ pmd_val(orig)) & _PAGE_PMD_HUGE) {
 174                 /*
 175                  * Note that this routine only sets pmds for THP pages.
 176                  * Hugetlb pages are handled elsewhere.  We need to check
 177                  * for huge zero page.  Huge zero pages are like hugetlb
 178                  * pages in that there is no RSS, but there is the need
 179                  * for TSB entries.  So, huge zero page counts go into
 180                  * hugetlb_pte_count.
 181                  */
 182                 if (pmd_val(pmd) & _PAGE_PMD_HUGE) {
 183                         if (is_huge_zero_page(pmd_page(pmd)))
 184                                 mm->context.hugetlb_pte_count++;
 185                         else
 186                                 mm->context.thp_pte_count++;
 187                 } else {
 188                         if (is_huge_zero_page(pmd_page(orig)))
 189                                 mm->context.hugetlb_pte_count--;
 190                         else
 191                                 mm->context.thp_pte_count--;
 192                 }
 193
 194                 /* Do not try to allocate the TSB hash table if we
 195                  * don't have one already.  We have various locks held
 196                  * and thus we'll end up doing a GFP_KERNEL allocation
 197                  * in an atomic context.
 198                  *
 199                  * Instead, we let the first TLB miss on a hugepage
 200                  * take care of this.
 201                  */
 202         }
 203
 204         if (!pmd_none(orig)) {
 205                 addr &= HPAGE_MASK;
 206                 if (pmd_trans_huge(orig)) {
 207                         pte_t orig_pte = __pte(pmd_val(orig));
 208                         bool exec = pte_exec(orig_pte);
 209
 210                         tlb_batch_add_one(mm, addr, exec, true);
 211                         tlb_batch_add_one(mm, addr + REAL_HPAGE_SIZE, exec,
 212                                         true);
 213                 } else {
 214                         tlb_batch_pmd_scan(mm, addr, orig);
 215                 }
 216         }
 217 }
 218
 219 void set_pmd_at(struct mm_struct *mm, unsigned long addr,
 220                 pmd_t *pmdp, pmd_t pmd)
 221 {
 222         pmd_t orig = *pmdp;
 223
 224         *pmdp = pmd;
 225         __set_pmd_acct(mm, addr, orig, pmd);
 226 }
 227
 228 static inline pmd_t pmdp_establish(struct vm_area_struct *vma,
 229                 unsigned long address, pmd_t *pmdp, pmd_t pmd)
 230 {
 231         pmd_t old;
 232
 233         do {
 234                 old = *pmdp;
 235         } while (cmpxchg64(&pmdp->pmd, old.pmd, pmd.pmd) != old.pmd);
 236         __set_pmd_acct(vma->vm_mm, address, old, pmd);
 237
 238         return old;
 239 }
 240
 241 /*
 242  * This routine is only called when splitting a THP
 243  */
 244 pmd_t pmdp_invalidate(struct vm_area_struct *vma, unsigned long address,
 245                      pmd_t *pmdp)
 246 {
 247         pmd_t old, entry;
 248
 249         entry = __pmd(pmd_val(*pmdp) & ~_PAGE_VALID);
 250         old = pmdp_establish(vma, address, pmdp, entry);
 251         flush_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
 252
 253         /*
 254          * set_pmd_at() will not be called in a way to decrement
 255          * thp_pte_count when splitting a THP, so do it now.
 256          * Sanity check pmd before doing the actual decrement.
 257          */
 258         if ((pmd_val(entry) & _PAGE_PMD_HUGE) &&
 259             !is_huge_zero_page(pmd_page(entry)))
 260                 (vma->vm_mm)->context.thp_pte_count--;
 261
 262         return old;
 263 }
 264
 265 void pgtable_trans_huge_deposit(struct mm_struct *mm, pmd_t *pmdp,
 266                                 pgtable_t pgtable)
 267 {
 268         struct list_head *lh = (struct list_head *) pgtable;
 269
 270         assert_spin_locked(&mm->page_table_lock);
 271
 272         /* FIFO */
 273         if (!pmd_huge_pte(mm, pmdp))
 274                 INIT_LIST_HEAD(lh);
 275         else
 276                 list_add(lh, (struct list_head *) pmd_huge_pte(mm, pmdp));
 277         pmd_huge_pte(mm, pmdp) = pgtable;
 278 }
 279
 280 pgtable_t pgtable_trans_huge_withdraw(struct mm_struct *mm, pmd_t *pmdp)
 281 {
 282         struct list_head *lh;
 283         pgtable_t pgtable;
 284
 285         assert_spin_locked(&mm->page_table_lock);
 286
 287         /* FIFO */
 288         pgtable = pmd_huge_pte(mm, pmdp);
 289         lh = (struct list_head *) pgtable;
 290         if (list_empty(lh))
 291                 pmd_huge_pte(mm, pmdp) = NULL;
 292         else {
 293                 pmd_huge_pte(mm, pmdp) = (pgtable_t) lh->next;
 294                 list_del(lh);
 295         }
 296         pte_val(pgtable[0]) = 0;
 297         pte_val(pgtable[1]) = 0;
 298
 299         return pgtable;
 300 }
 301 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */