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/gfp.h>
  26 #include <linux/mm.h>
  27 #include <linux/init.h>
  28 #include <linux/percpu.h>
  29 #include <linux/hardirq.h>
  30 #include <asm/pgalloc.h>
  31 #include <asm/tlbflush.h>
  32 #include <asm/tlb.h>
  33
  34 #include "mmu_decl.h"
  35
  36 #ifdef CONFIG_SMP
  37
  38 /*
  39  * Handle batching of page table freeing on SMP. Page tables are
  40  * queued up and send to be freed later by RCU in order to avoid
  41  * freeing a page table page that is being walked without locks
  42  */
  43
  44 static unsigned long pte_freelist_forced_free;
  45
  46 struct pte_freelist_batch
  47 {
  48         struct rcu_head rcu;
  49         unsigned int    index;
  50         unsigned long   tables[0];
  51 };
  52
  53 #define PTE_FREELIST_SIZE \
  54         ((PAGE_SIZE - sizeof(struct pte_freelist_batch)) \
  55           / sizeof(unsigned long))
  56
  57 static void pte_free_smp_sync(void *arg)
  58 {
  59         /* Do nothing, just ensure we sync with all CPUs */
  60 }
  61
  62 /* This is only called when we are critically out of memory
  63  * (and fail to get a page in pte_free_tlb).
  64  */
  65 static void pgtable_free_now(void *table, unsigned shift)
  66 {
  67         pte_freelist_forced_free++;
  68
  69         smp_call_function(pte_free_smp_sync, NULL, 1);
  70
  71         pgtable_free(table, shift);
  72 }
  73
  74 static void pte_free_rcu_callback(struct rcu_head *head)
  75 {
  76         struct pte_freelist_batch *batch =
  77                 container_of(head, struct pte_freelist_batch, rcu);
  78         unsigned int i;
  79
  80         for (i = 0; i < batch->index; i++) {
  81                 void *table = (void *)(batch->tables[i] & ~MAX_PGTABLE_INDEX_SIZE);
  82                 unsigned shift = batch->tables[i] & MAX_PGTABLE_INDEX_SIZE;
  83
  84                 pgtable_free(table, shift);
  85         }
  86
  87         free_page((unsigned long)batch);
  88 }
  89
  90 static void pte_free_submit(struct pte_freelist_batch *batch)
  91 {
  92         call_rcu_sched(&batch->rcu, pte_free_rcu_callback);
  93 }
  94
  95 void pgtable_free_tlb(struct mmu_gather *tlb, void *table, unsigned shift)
  96 {
  97         struct pte_freelist_batch **batchp = &tlb->arch.batch;
  98         unsigned long pgf;
  99
 100         if (atomic_read(&tlb->mm->mm_users) < 2) {
 101                 pgtable_free(table, shift);
 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(table, shift);
 109                         return;
 110                 }
 111                 (*batchp)->index = 0;
 112         }
 113         BUG_ON(shift > MAX_PGTABLE_INDEX_SIZE);
 114         pgf = (unsigned long)table | shift;
 115         (*batchp)->tables[(*batchp)->index++] = pgf;
 116         if ((*batchp)->index == PTE_FREELIST_SIZE) {
 117                 pte_free_submit(*batchp);
 118                 *batchp = NULL;
 119         }
 120 }
 121
 122 void pte_free_finish(struct mmu_gather *tlb)
 123 {
 124         struct pte_freelist_batch **batchp = &tlb->arch.batch;
 125
 126         if (*batchp == NULL)
 127                 return;
 128         pte_free_submit(*batchp);
 129         *batchp = NULL;
 130 }
 131
 132 #endif /* CONFIG_SMP */
 133
 134 static inline int is_exec_fault(void)
 135 {
 136         return current->thread.regs && TRAP(current->thread.regs) == 0x400;
 137 }
 138
 139 /* We only try to do i/d cache coherency on stuff that looks like
 140  * reasonably "normal" PTEs. We currently require a PTE to be present
 141  * and we avoid _PAGE_SPECIAL and _PAGE_NO_CACHE. We also only do that
 142  * on userspace PTEs
 143  */
 144 static inline int pte_looks_normal(pte_t pte)
 145 {
 146         return (pte_val(pte) &
 147             (_PAGE_PRESENT | _PAGE_SPECIAL | _PAGE_NO_CACHE | _PAGE_USER)) ==
 148             (_PAGE_PRESENT | _PAGE_USER);
 149 }
 150
 151 struct page * maybe_pte_to_page(pte_t pte)
 152 {
 153         unsigned long pfn = pte_pfn(pte);
 154         struct page *page;
 155
 156         if (unlikely(!pfn_valid(pfn)))
 157                 return NULL;
 158         page = pfn_to_page(pfn);
 159         if (PageReserved(page))
 160                 return NULL;
 161         return page;
 162 }
 163
 164 #if defined(CONFIG_PPC_STD_MMU) || _PAGE_EXEC == 0
 165
 166 /* Server-style MMU handles coherency when hashing if HW exec permission
 167  * is supposed per page (currently 64-bit only). If not, then, we always
 168  * flush the cache for valid PTEs in set_pte. Embedded CPU without HW exec
 169  * support falls into the same category.
 170  */
 171
 172 static pte_t set_pte_filter(pte_t pte, unsigned long addr)
 173 {
 174         pte = __pte(pte_val(pte) & ~_PAGE_HPTEFLAGS);
 175         if (pte_looks_normal(pte) && !(cpu_has_feature(CPU_FTR_COHERENT_ICACHE) ||
 176                                        cpu_has_feature(CPU_FTR_NOEXECUTE))) {
 177                 struct page *pg = maybe_pte_to_page(pte);
 178                 if (!pg)
 179                         return pte;
 180                 if (!test_bit(PG_arch_1, &pg->flags)) {
 181 #ifdef CONFIG_8xx
 182                         /* On 8xx, cache control instructions (particularly
 183                          * "dcbst" from flush_dcache_icache) fault as write
 184                          * operation if there is an unpopulated TLB entry
 185                          * for the address in question. To workaround that,
 186                          * we invalidate the TLB here, thus avoiding dcbst
 187                          * misbehaviour.
 188                          */
 189                         /* 8xx doesn't care about PID, size or ind args */
 190                         _tlbil_va(addr, 0, 0, 0);
 191 #endif /* CONFIG_8xx */
 192                         flush_dcache_icache_page(pg);
 193                         set_bit(PG_arch_1, &pg->flags);
 194                 }
 195         }
 196         return pte;
 197 }
 198
 199 static pte_t set_access_flags_filter(pte_t pte, struct vm_area_struct *vma,
 200                                      int dirty)
 201 {
 202         return pte;
 203 }
 204
 205 #else /* defined(CONFIG_PPC_STD_MMU) || _PAGE_EXEC == 0 */
 206
 207 /* Embedded type MMU with HW exec support. This is a bit more complicated
 208  * as we don't have two bits to spare for _PAGE_EXEC and _PAGE_HWEXEC so
 209  * instead we "filter out" the exec permission for non clean pages.
 210  */
 211 static pte_t set_pte_filter(pte_t pte, unsigned long addr)
 212 {
 213         struct page *pg;
 214
 215         /* No exec permission in the first place, move on */
 216         if (!(pte_val(pte) & _PAGE_EXEC) || !pte_looks_normal(pte))
 217                 return pte;
 218
 219         /* If you set _PAGE_EXEC on weird pages you're on your own */
 220         pg = maybe_pte_to_page(pte);
 221         if (unlikely(!pg))
 222                 return pte;
 223
 224         /* If the page clean, we move on */
 225         if (test_bit(PG_arch_1, &pg->flags))
 226                 return pte;
 227
 228         /* If it's an exec fault, we flush the cache and make it clean */
 229         if (is_exec_fault()) {
 230                 flush_dcache_icache_page(pg);
 231                 set_bit(PG_arch_1, &pg->flags);
 232                 return pte;
 233         }
 234
 235         /* Else, we filter out _PAGE_EXEC */
 236         return __pte(pte_val(pte) & ~_PAGE_EXEC);
 237 }
 238
 239 static pte_t set_access_flags_filter(pte_t pte, struct vm_area_struct *vma,
 240                                      int dirty)
 241 {
 242         struct page *pg;
 243
 244         /* So here, we only care about exec faults, as we use them
 245          * to recover lost _PAGE_EXEC and perform I$/D$ coherency
 246          * if necessary. Also if _PAGE_EXEC is already set, same deal,
 247          * we just bail out
 248          */
 249         if (dirty || (pte_val(pte) & _PAGE_EXEC) || !is_exec_fault())
 250                 return pte;
 251
 252 #ifdef CONFIG_DEBUG_VM
 253         /* So this is an exec fault, _PAGE_EXEC is not set. If it was
 254          * an error we would have bailed out earlier in do_page_fault()
 255          * but let's make sure of it
 256          */
 257         if (WARN_ON(!(vma->vm_flags & VM_EXEC)))
 258                 return pte;
 259 #endif /* CONFIG_DEBUG_VM */
 260
 261         /* If you set _PAGE_EXEC on weird pages you're on your own */
 262         pg = maybe_pte_to_page(pte);
 263         if (unlikely(!pg))
 264                 goto bail;
 265
 266         /* If the page is already clean, we move on */
 267         if (test_bit(PG_arch_1, &pg->flags))
 268                 goto bail;
 269
 270         /* Clean the page and set PG_arch_1 */
 271         flush_dcache_icache_page(pg);
 272         set_bit(PG_arch_1, &pg->flags);
 273
 274  bail:
 275         return __pte(pte_val(pte) | _PAGE_EXEC);
 276 }
 277
 278 #endif /* !(defined(CONFIG_PPC_STD_MMU) || _PAGE_EXEC == 0) */
 279
 280 /*
 281  * set_pte stores a linux PTE into the linux page table.
 282  */
 283 void set_pte_at(struct mm_struct *mm, unsigned long addr, pte_t *ptep,
 284                 pte_t pte)
 285 {
 286 #ifdef CONFIG_DEBUG_VM
 287         WARN_ON(pte_present(*ptep));
 288 #endif
 289         /* Note: mm->context.id might not yet have been assigned as
 290          * this context might not have been activated yet when this
 291          * is called.
 292          */
 293         pte = set_pte_filter(pte, addr);
 294
 295         /* Perform the setting of the PTE */
 296         __set_pte_at(mm, addr, ptep, pte, 0);
 297 }
 298
 299 /*
 300  * This is called when relaxing access to a PTE. It's also called in the page
 301  * fault path when we don't hit any of the major fault cases, ie, a minor
 302  * update of _PAGE_ACCESSED, _PAGE_DIRTY, etc... The generic code will have
 303  * handled those two for us, we additionally deal with missing execute
 304  * permission here on some processors
 305  */
 306 int ptep_set_access_flags(struct vm_area_struct *vma, unsigned long address,
 307                           pte_t *ptep, pte_t entry, int dirty)
 308 {
 309         int changed;
 310         entry = set_access_flags_filter(entry, vma, dirty);
 311         changed = !pte_same(*(ptep), entry);
 312         if (changed) {
 313                 if (!(vma->vm_flags & VM_HUGETLB))
 314                         assert_pte_locked(vma->vm_mm, address);
 315                 __ptep_set_access_flags(ptep, entry);
 316                 flush_tlb_page_nohash(vma, address);
 317         }
 318         return changed;
 319 }
 320
 321 #ifdef CONFIG_DEBUG_VM
 322 void assert_pte_locked(struct mm_struct *mm, unsigned long addr)
 323 {
 324         pgd_t *pgd;
 325         pud_t *pud;
 326         pmd_t *pmd;
 327
 328         if (mm == &init_mm)
 329                 return;
 330         pgd = mm->pgd + pgd_index(addr);
 331         BUG_ON(pgd_none(*pgd));
 332         pud = pud_offset(pgd, addr);
 333         BUG_ON(pud_none(*pud));
 334         pmd = pmd_offset(pud, addr);
 335         BUG_ON(!pmd_present(*pmd));
 336         assert_spin_locked(pte_lockptr(mm, pmd));
 337 }
 338 #endif /* CONFIG_DEBUG_VM */
 339