arch/powerpc/mm/pgtable.c

   1 /*
   2  * This file contains common routines for dealing with free of page tables
   3  * Along with common page table handling code
   4  *
   5  *  Derived from arch/powerpc/mm/tlb_64.c:
   6  *    Copyright (C) 1995-1996 Gary Thomas (gdt@linuxppc.org)
   7  *
   8  *  Modifications by Paul Mackerras (PowerMac) (paulus@cs.anu.edu.au)
   9  *  and Cort Dougan (PReP) (cort@cs.nmt.edu)
  10  *    Copyright (C) 1996 Paul Mackerras
  11  *
  12  *  Derived from "arch/i386/mm/init.c"
  13  *    Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
  14  *
  15  *  Dave Engebretsen <engebret@us.ibm.com>
  16  *      Rework for PPC64 port.
  17  *
  18  *  This program is free software; you can redistribute it and/or
  19  *  modify it under the terms of the GNU General Public License
  20  *  as published by the Free Software Foundation; either version
  21  *  2 of the License, or (at your option) any later version.
  22  */
  23
  24 #include <linux/kernel.h>
  25 #include <linux/mm.h>
  26 #include <linux/init.h>
  27 #include <linux/percpu.h>
  28 #include <linux/hardirq.h>
  29 #include <asm/pgalloc.h>
  30 #include <asm/tlbflush.h>
  31 #include <asm/tlb.h>
  32
  33 #include "mmu_decl.h"
  34
  35 DEFINE_PER_CPU(struct mmu_gather, mmu_gathers);
  36
  37 #ifdef CONFIG_SMP
  38
  39 /*
  40  * Handle batching of page table freeing on SMP. Page tables are
  41  * queued up and send to be freed later by RCU in order to avoid
  42  * freeing a page table page that is being walked without locks
  43  */
  44
  45 static DEFINE_PER_CPU(struct pte_freelist_batch *, pte_freelist_cur);
  46 static unsigned long pte_freelist_forced_free;
  47
  48 struct pte_freelist_batch
  49 {
  50         struct rcu_head rcu;
  51         unsigned int    index;
  52         pgtable_free_t  tables[0];
  53 };
  54
  55 #define PTE_FREELIST_SIZE \
  56         ((PAGE_SIZE - sizeof(struct pte_freelist_batch)) \
  57           / sizeof(pgtable_free_t))
  58
  59 static void pte_free_smp_sync(void *arg)
  60 {
  61         /* Do nothing, just ensure we sync with all CPUs */
  62 }
  63
  64 /* This is only called when we are critically out of memory
  65  * (and fail to get a page in pte_free_tlb).
  66  */
  67 static void pgtable_free_now(pgtable_free_t pgf)
  68 {
  69         pte_freelist_forced_free++;
  70
  71         smp_call_function(pte_free_smp_sync, NULL, 1);
  72
  73         pgtable_free(pgf);
  74 }
  75
  76 static void pte_free_rcu_callback(struct rcu_head *head)
  77 {
  78         struct pte_freelist_batch *batch =
  79                 container_of(head, struct pte_freelist_batch, rcu);
  80         unsigned int i;
  81
  82         for (i = 0; i < batch->index; i++)
  83                 pgtable_free(batch->tables[i]);
  84
  85         free_page((unsigned long)batch);
  86 }
  87
  88 static void pte_free_submit(struct pte_freelist_batch *batch)
  89 {
  90         INIT_RCU_HEAD(&batch->rcu);
  91         call_rcu(&batch->rcu, pte_free_rcu_callback);
  92 }
  93
  94 void pgtable_free_tlb(struct mmu_gather *tlb, pgtable_free_t pgf)
  95 {
  96         /* This is safe since tlb_gather_mmu has disabled preemption */
  97         struct pte_freelist_batch **batchp = &__get_cpu_var(pte_freelist_cur);
  98
  99         if (atomic_read(&tlb->mm->mm_users) < 2 ||
 100             cpumask_equal(mm_cpumask(tlb->mm), cpumask_of(smp_processor_id()))){
 101                 pgtable_free(pgf);
 102                 return;
 103         }
 104
 105         if (*batchp == NULL) {
 106                 *batchp = (struct pte_freelist_batch *)__get_free_page(GFP_ATOMIC);
 107                 if (*batchp == NULL) {
 108                         pgtable_free_now(pgf);
 109                         return;
 110                 }
 111                 (*batchp)->index = 0;
 112         }
 113         (*batchp)->tables[(*batchp)->index++] = pgf;
 114         if ((*batchp)->index == PTE_FREELIST_SIZE) {
 115                 pte_free_submit(*batchp);
 116                 *batchp = NULL;
 117         }
 118 }
 119
 120 void pte_free_finish(void)
 121 {
 122         /* This is safe since tlb_gather_mmu has disabled preemption */
 123         struct pte_freelist_batch **batchp = &__get_cpu_var(pte_freelist_cur);
 124
 125         if (*batchp == NULL)
 126                 return;
 127         pte_free_submit(*batchp);
 128         *batchp = NULL;
 129 }
 130
 131 #endif /* CONFIG_SMP */
 132
 133 static inline int is_exec_fault(void)
 134 {
 135         return current->thread.regs && TRAP(current->thread.regs) == 0x400;
 136 }
 137
 138 /* We only try to do i/d cache coherency on stuff that looks like
 139  * reasonably "normal" PTEs. We currently require a PTE to be present
 140  * and we avoid _PAGE_SPECIAL and _PAGE_NO_CACHE. We also only do that
 141  * on userspace PTEs
 142  */
 143 static inline int pte_looks_normal(pte_t pte)
 144 {
 145         return (pte_val(pte) &
 146             (_PAGE_PRESENT | _PAGE_SPECIAL | _PAGE_NO_CACHE | _PAGE_USER)) ==
 147             (_PAGE_PRESENT | _PAGE_USER);
 148 }
 149
 150 struct page * maybe_pte_to_page(pte_t pte)
 151 {
 152         unsigned long pfn = pte_pfn(pte);
 153         struct page *page;
 154
 155         if (unlikely(!pfn_valid(pfn)))
 156                 return NULL;
 157         page = pfn_to_page(pfn);
 158         if (PageReserved(page))
 159                 return NULL;
 160         return page;
 161 }
 162
 163 #if defined(CONFIG_PPC_STD_MMU) || _PAGE_EXEC == 0
 164
 165 /* Server-style MMU handles coherency when hashing if HW exec permission
 166  * is supposed per page (currently 64-bit only). If not, then, we always
 167  * flush the cache for valid PTEs in set_pte. Embedded CPU without HW exec
 168  * support falls into the same category.
 169  */
 170
 171 static pte_t set_pte_filter(pte_t pte, unsigned long addr)
 172 {
 173         pte = __pte(pte_val(pte) & ~_PAGE_HPTEFLAGS);
 174         if (pte_looks_normal(pte) && !(cpu_has_feature(CPU_FTR_COHERENT_ICACHE) ||
 175                                        cpu_has_feature(CPU_FTR_NOEXECUTE))) {
 176                 struct page *pg = maybe_pte_to_page(pte);
 177                 if (!pg)
 178                         return pte;
 179                 if (!test_bit(PG_arch_1, &pg->flags)) {
 180 #ifdef CONFIG_8xx
 181                         /* On 8xx, cache control instructions (particularly
 182                          * "dcbst" from flush_dcache_icache) fault as write
 183                          * operation if there is an unpopulated TLB entry
 184                          * for the address in question. To workaround that,
 185                          * we invalidate the TLB here, thus avoiding dcbst
 186                          * misbehaviour.
 187                          */
 188                         /* 8xx doesn't care about PID, size or ind args */
 189                         _tlbil_va(addr, 0, 0, 0);
 190 #endif /* CONFIG_8xx */
 191                         flush_dcache_icache_page(pg);
 192                         set_bit(PG_arch_1, &pg->flags);
 193                 }
 194         }
 195         return pte;
 196 }
 197
 198 static pte_t set_access_flags_filter(pte_t pte, struct vm_area_struct *vma,
 199                                      int dirty)
 200 {
 201         return pte;
 202 }
 203
 204 #else /* defined(CONFIG_PPC_STD_MMU) || _PAGE_EXEC == 0 */
 205
 206 /* Embedded type MMU with HW exec support. This is a bit more complicated
 207  * as we don't have two bits to spare for _PAGE_EXEC and _PAGE_HWEXEC so
 208  * instead we "filter out" the exec permission for non clean pages.
 209  */
 210 static pte_t set_pte_filter(pte_t pte, unsigned long addr)
 211 {
 212         struct page *pg;
 213
 214         /* No exec permission in the first place, move on */
 215         if (!(pte_val(pte) & _PAGE_EXEC) || !pte_looks_normal(pte))
 216                 return pte;
 217
 218         /* If you set _PAGE_EXEC on weird pages you're on your own */
 219         pg = maybe_pte_to_page(pte);
 220         if (unlikely(!pg))
 221                 return pte;
 222
 223         /* If the page clean, we move on */
 224         if (test_bit(PG_arch_1, &pg->flags))
 225                 return pte;
 226
 227         /* If it's an exec fault, we flush the cache and make it clean */
 228         if (is_exec_fault()) {
 229                 flush_dcache_icache_page(pg);
 230                 set_bit(PG_arch_1, &pg->flags);
 231                 return pte;
 232         }
 233
 234         /* Else, we filter out _PAGE_EXEC */
 235         return __pte(pte_val(pte) & ~_PAGE_EXEC);
 236 }
 237
 238 static pte_t set_access_flags_filter(pte_t pte, struct vm_area_struct *vma,
 239                                      int dirty)
 240 {
 241         struct page *pg;
 242
 243         /* So here, we only care about exec faults, as we use them
 244          * to recover lost _PAGE_EXEC and perform I$/D$ coherency
 245          * if necessary. Also if _PAGE_EXEC is already set, same deal,
 246          * we just bail out
 247          */
 248         if (dirty || (pte_val(pte) & _PAGE_EXEC) || !is_exec_fault())
 249                 return pte;
 250
 251 #ifdef CONFIG_DEBUG_VM
 252         /* So this is an exec fault, _PAGE_EXEC is not set. If it was
 253          * an error we would have bailed out earlier in do_page_fault()
 254          * but let's make sure of it
 255          */
 256         if (WARN_ON(!(vma->vm_flags & VM_EXEC)))
 257                 return pte;
 258 #endif /* CONFIG_DEBUG_VM */
 259
 260         /* If you set _PAGE_EXEC on weird pages you're on your own */
 261         pg = maybe_pte_to_page(pte);
 262         if (unlikely(!pg))
 263                 goto bail;
 264
 265         /* If the page is already clean, we move on */
 266         if (test_bit(PG_arch_1, &pg->flags))
 267                 goto bail;
 268
 269         /* Clean the page and set PG_arch_1 */
 270         flush_dcache_icache_page(pg);
 271         set_bit(PG_arch_1, &pg->flags);
 272
 273  bail:
 274         return __pte(pte_val(pte) | _PAGE_EXEC);
 275 }
 276
 277 #endif /* !(defined(CONFIG_PPC_STD_MMU) || _PAGE_EXEC == 0) */
 278
 279 /*
 280  * set_pte stores a linux PTE into the linux page table.
 281  */
 282 void set_pte_at(struct mm_struct *mm, unsigned long addr, pte_t *ptep,
 283                 pte_t pte)
 284 {
 285 #ifdef CONFIG_DEBUG_VM
 286         WARN_ON(pte_present(*ptep));
 287 #endif
 288         /* Note: mm->context.id might not yet have been assigned as
 289          * this context might not have been activated yet when this
 290          * is called.
 291          */
 292         pte = set_pte_filter(pte, addr);
 293
 294         /* Perform the setting of the PTE */
 295         __set_pte_at(mm, addr, ptep, pte, 0);
 296 }
 297
 298 /*
 299  * This is called when relaxing access to a PTE. It's also called in the page
 300  * fault path when we don't hit any of the major fault cases, ie, a minor
 301  * update of _PAGE_ACCESSED, _PAGE_DIRTY, etc... The generic code will have
 302  * handled those two for us, we additionally deal with missing execute
 303  * permission here on some processors
 304  */
 305 int ptep_set_access_flags(struct vm_area_struct *vma, unsigned long address,
 306                           pte_t *ptep, pte_t entry, int dirty)
 307 {
 308         int changed;
 309         entry = set_access_flags_filter(entry, vma, dirty);
 310         changed = !pte_same(*(ptep), entry);
 311         if (changed) {
 312                 if (!(vma->vm_flags & VM_HUGETLB))
 313                         assert_pte_locked(vma->vm_mm, address);
 314                 __ptep_set_access_flags(ptep, entry);
 315                 flush_tlb_page_nohash(vma, address);
 316         }
 317         return changed;
 318 }
 319
 320 #ifdef CONFIG_DEBUG_VM
 321 void assert_pte_locked(struct mm_struct *mm, unsigned long addr)
 322 {
 323         pgd_t *pgd;
 324         pud_t *pud;
 325         pmd_t *pmd;
 326
 327         if (mm == &init_mm)
 328                 return;
 329         pgd = mm->pgd + pgd_index(addr);
 330         BUG_ON(pgd_none(*pgd));
 331         pud = pud_offset(pgd, addr);
 332         BUG_ON(pud_none(*pud));
 333         pmd = pmd_offset(pud, addr);
 334         BUG_ON(!pmd_present(*pmd));
 335         assert_spin_locked(pte_lockptr(mm, pmd));
 336 }
 337 #endif /* CONFIG_DEBUG_VM */
 338