arch/x86_64/mm/pageattr.c

   1 /*
   2  * Copyright 2002 Andi Kleen, SuSE Labs.
   3  * Thanks to Ben LaHaise for precious feedback.
   4  */
   5
   6 #include <linux/mm.h>
   7 #include <linux/sched.h>
   8 #include <linux/highmem.h>
   9 #include <linux/module.h>
  10 #include <linux/slab.h>
  11 #include <asm/uaccess.h>
  12 #include <asm/processor.h>
  13 #include <asm/tlbflush.h>
  14 #include <asm/io.h>
  15
  16 #ifdef CONFIG_XEN
  17 #include <asm/pgalloc.h>
  18 #include <asm/mmu_context.h>
  19
  20 LIST_HEAD(mm_unpinned);
  21 DEFINE_SPINLOCK(mm_unpinned_lock);
  22
  23 static inline void mm_walk_set_prot(void *pt, pgprot_t flags)
  24 {
  25         struct page *page = virt_to_page(pt);
  26         unsigned long pfn = page_to_pfn(page);
  27
  28         BUG_ON(HYPERVISOR_update_va_mapping(
  29                        (unsigned long)__va(pfn << PAGE_SHIFT),
  30                        pfn_pte(pfn, flags), 0));
  31 }
  32
  33 static void mm_walk(struct mm_struct *mm, pgprot_t flags)
  34 {
  35         pgd_t       *pgd;
  36         pud_t       *pud;
  37         pmd_t       *pmd;
  38         pte_t       *pte;
  39         int          g,u,m;
  40
  41         pgd = mm->pgd;
  42         /*
  43          * Cannot iterate up to USER_PTRS_PER_PGD as these pagetables may not
  44          * be the 'current' task's pagetables (e.g., current may be 32-bit,
  45          * but the pagetables may be for a 64-bit task).
  46          * Subtracting 1 from TASK_SIZE64 means the loop limit is correct
  47          * regardless of whether TASK_SIZE64 is a multiple of PGDIR_SIZE.
  48          */
  49         for (g = 0; g <= ((TASK_SIZE64-1) / PGDIR_SIZE); g++, pgd++) {
  50                 if (pgd_none(*pgd))
  51                         continue;
  52                 pud = pud_offset(pgd, 0);
  53                 if (PTRS_PER_PUD > 1) /* not folded */
  54                         mm_walk_set_prot(pud,flags);
  55                 for (u = 0; u < PTRS_PER_PUD; u++, pud++) {
  56                         if (pud_none(*pud))
  57                                 continue;
  58                         pmd = pmd_offset(pud, 0);
  59                         if (PTRS_PER_PMD > 1) /* not folded */
  60                                 mm_walk_set_prot(pmd,flags);
  61                         for (m = 0; m < PTRS_PER_PMD; m++, pmd++) {
  62                                 if (pmd_none(*pmd))
  63                                         continue;
  64                                 pte = pte_offset_kernel(pmd,0);
  65                                 mm_walk_set_prot(pte,flags);
  66                         }
  67                 }
  68         }
  69 }
  70
  71 void mm_pin(struct mm_struct *mm)
  72 {
  73         if (xen_feature(XENFEAT_writable_page_tables))
  74                 return;
  75
  76         spin_lock(&mm->page_table_lock);
  77
  78         mm_walk(mm, PAGE_KERNEL_RO);
  79         BUG_ON(HYPERVISOR_update_va_mapping(
  80                        (unsigned long)mm->pgd,
  81                        pfn_pte(virt_to_phys(mm->pgd)>>PAGE_SHIFT, PAGE_KERNEL_RO),
  82                        UVMF_TLB_FLUSH));
  83         BUG_ON(HYPERVISOR_update_va_mapping(
  84                        (unsigned long)__user_pgd(mm->pgd),
  85                        pfn_pte(virt_to_phys(__user_pgd(mm->pgd))>>PAGE_SHIFT, PAGE_KERNEL_RO),
  86                        UVMF_TLB_FLUSH));
  87         xen_pgd_pin(__pa(mm->pgd)); /* kernel */
  88         xen_pgd_pin(__pa(__user_pgd(mm->pgd))); /* user */
  89         mm->context.pinned = 1;
  90         spin_lock(&mm_unpinned_lock);
  91         list_del(&mm->context.unpinned);
  92         spin_unlock(&mm_unpinned_lock);
  93
  94         spin_unlock(&mm->page_table_lock);
  95 }
  96
  97 void mm_unpin(struct mm_struct *mm)
  98 {
  99         if (xen_feature(XENFEAT_writable_page_tables))
 100                 return;
 101
 102         spin_lock(&mm->page_table_lock);
 103
 104         xen_pgd_unpin(__pa(mm->pgd));
 105         xen_pgd_unpin(__pa(__user_pgd(mm->pgd)));
 106         BUG_ON(HYPERVISOR_update_va_mapping(
 107                        (unsigned long)mm->pgd,
 108                        pfn_pte(virt_to_phys(mm->pgd)>>PAGE_SHIFT, PAGE_KERNEL), 0));
 109         BUG_ON(HYPERVISOR_update_va_mapping(
 110                        (unsigned long)__user_pgd(mm->pgd),
 111                        pfn_pte(virt_to_phys(__user_pgd(mm->pgd))>>PAGE_SHIFT, PAGE_KERNEL), 0));
 112         mm_walk(mm, PAGE_KERNEL);
 113         xen_tlb_flush();
 114         mm->context.pinned = 0;
 115         spin_lock(&mm_unpinned_lock);
 116         list_add(&mm->context.unpinned, &mm_unpinned);
 117         spin_unlock(&mm_unpinned_lock);
 118
 119         spin_unlock(&mm->page_table_lock);
 120 }
 121
 122 void mm_pin_all(void)
 123 {
 124         if (xen_feature(XENFEAT_writable_page_tables))
 125                 return;
 126
 127         while (!list_empty(&mm_unpinned))
 128                 mm_pin(list_entry(mm_unpinned.next, struct mm_struct,
 129                                   context.unpinned));
 130 }
 131
 132 void _arch_dup_mmap(struct mm_struct *mm)
 133 {
 134     if (!mm->context.pinned)
 135         mm_pin(mm);
 136 }
 137
 138 void _arch_exit_mmap(struct mm_struct *mm)
 139 {
 140     struct task_struct *tsk = current;
 141
 142     task_lock(tsk);
 143
 144     /*
 145      * We aggressively remove defunct pgd from cr3. We execute unmap_vmas()
 146      * *much* faster this way, as no tlb flushes means bigger wrpt batches.
 147      */
 148     if ( tsk->active_mm == mm )
 149     {
 150         tsk->active_mm = &init_mm;
 151         atomic_inc(&init_mm.mm_count);
 152
 153         switch_mm(mm, &init_mm, tsk);
 154
 155         atomic_dec(&mm->mm_count);
 156         BUG_ON(atomic_read(&mm->mm_count) == 0);
 157     }
 158
 159     task_unlock(tsk);
 160
 161     if ( mm->context.pinned && (atomic_read(&mm->mm_count) == 1) &&
 162          !mm->context.has_foreign_mappings )
 163         mm_unpin(mm);
 164 }
 165
 166 void pte_free(struct page *pte)
 167 {
 168         unsigned long va = (unsigned long)__va(page_to_pfn(pte)<<PAGE_SHIFT);
 169
 170         if (!pte_write(*virt_to_ptep(va)))
 171                 BUG_ON(HYPERVISOR_update_va_mapping(
 172                         va, pfn_pte(page_to_pfn(pte), PAGE_KERNEL), 0));
 173         __free_page(pte);
 174 }
 175 #endif  /* CONFIG_XEN */
 176
 177 static inline pte_t *lookup_address(unsigned long address)
 178 {
 179         pgd_t *pgd = pgd_offset_k(address);
 180         pud_t *pud;
 181         pmd_t *pmd;
 182         pte_t *pte;
 183         if (pgd_none(*pgd))
 184                 return NULL;
 185         pud = pud_offset(pgd, address);
 186         if (!pud_present(*pud))
 187                 return NULL;
 188         pmd = pmd_offset(pud, address);
 189         if (!pmd_present(*pmd))
 190                 return NULL;
 191         if (pmd_large(*pmd))
 192                 return (pte_t *)pmd;
 193         pte = pte_offset_kernel(pmd, address);
 194         if (pte && !pte_present(*pte))
 195                 pte = NULL;
 196         return pte;
 197 }
 198
 199 static struct page *split_large_page(unsigned long address, pgprot_t prot,
 200                                      pgprot_t ref_prot)
 201 {
 202         int i;
 203         unsigned long addr;
 204         struct page *base = alloc_pages(GFP_KERNEL, 0);
 205         pte_t *pbase;
 206         if (!base)
 207                 return NULL;
 208         /*
 209          * page_private is used to track the number of entries in
 210          * the page table page have non standard attributes.
 211          */
 212         SetPagePrivate(base);
 213         page_private(base) = 0;
 214
 215         address = __pa(address);
 216         addr = address & LARGE_PAGE_MASK;
 217         pbase = (pte_t *)page_address(base);
 218         for (i = 0; i < PTRS_PER_PTE; i++, addr += PAGE_SIZE) {
 219                 pbase[i] = pfn_pte(addr >> PAGE_SHIFT,
 220                                    addr == address ? prot : ref_prot);
 221         }
 222         return base;
 223 }
 224
 225 static void cache_flush_page(void *adr)
 226 {
 227         int i;
 228         for (i = 0; i < PAGE_SIZE; i += boot_cpu_data.x86_clflush_size)
 229                 asm volatile("clflush (%0)" :: "r" (adr + i));
 230 }
 231
 232 static void flush_kernel_map(void *arg)
 233 {
 234         struct list_head *l = (struct list_head *)arg;
 235         struct page *pg;
 236
 237         /* When clflush is available always use it because it is
 238            much cheaper than WBINVD */
 239         if (!cpu_has_clflush)
 240                 asm volatile("wbinvd" ::: "memory");
 241         list_for_each_entry(pg, l, lru) {
 242                 void *adr = page_address(pg);
 243                 if (cpu_has_clflush)
 244                         cache_flush_page(adr);
 245         }
 246         __flush_tlb_all();
 247 }
 248
 249 static inline void flush_map(struct list_head *l)
 250 {
 251         on_each_cpu(flush_kernel_map, l, 1, 1);
 252 }
 253
 254 static LIST_HEAD(deferred_pages); /* protected by init_mm.mmap_sem */
 255
 256 static inline void save_page(struct page *fpage)
 257 {
 258         list_add(&fpage->lru, &deferred_pages);
 259 }
 260
 261 /*
 262  * No more special protections in this 2/4MB area - revert to a
 263  * large page again.
 264  */
 265 static void revert_page(unsigned long address, pgprot_t ref_prot)
 266 {
 267         pgd_t *pgd;
 268         pud_t *pud;
 269         pmd_t *pmd;
 270         pte_t large_pte;
 271
 272         pgd = pgd_offset_k(address);
 273         BUG_ON(pgd_none(*pgd));
 274         pud = pud_offset(pgd,address);
 275         BUG_ON(pud_none(*pud));
 276         pmd = pmd_offset(pud, address);
 277         BUG_ON(pmd_val(*pmd) & _PAGE_PSE);
 278         large_pte = mk_pte_phys(__pa(address) & LARGE_PAGE_MASK, ref_prot);
 279         large_pte = pte_mkhuge(large_pte);
 280         set_pte((pte_t *)pmd, large_pte);
 281 }
 282
 283 static int
 284 __change_page_attr(unsigned long address, unsigned long pfn, pgprot_t prot,
 285                                    pgprot_t ref_prot)
 286 {
 287         pte_t *kpte;
 288         struct page *kpte_page;
 289         pgprot_t ref_prot2;
 290         kpte = lookup_address(address);
 291         if (!kpte) return 0;
 292         kpte_page = virt_to_page(((unsigned long)kpte) & PAGE_MASK);
 293         if (pgprot_val(prot) != pgprot_val(ref_prot)) {
 294                 if (!pte_huge(*kpte)) {
 295                         set_pte(kpte, pfn_pte(pfn, prot));
 296                 } else {
 297                         /*
 298                          * split_large_page will take the reference for this
 299                          * change_page_attr on the split page.
 300                          */
 301                         struct page *split;
 302                         ref_prot2 = pte_pgprot(pte_clrhuge(*kpte));
 303                         split = split_large_page(address, prot, ref_prot2);
 304                         if (!split)
 305                                 return -ENOMEM;
 306                         set_pte(kpte, mk_pte(split, ref_prot2));
 307                         kpte_page = split;
 308                 }
 309                 page_private(kpte_page)++;
 310         } else if (!pte_huge(*kpte)) {
 311                 set_pte(kpte, pfn_pte(pfn, ref_prot));
 312                 BUG_ON(page_private(kpte_page) == 0);
 313                 page_private(kpte_page)--;
 314         } else
 315                 BUG();
 316
 317         /* on x86-64 the direct mapping set at boot is not using 4k pages */
 318         /*
 319          * ..., but the XEN guest kernels (currently) do:
 320          * If the pte was reserved, it means it was created at boot
 321          * time (not via split_large_page) and in turn we must not
 322          * replace it with a large page.
 323          */
 324 #ifndef CONFIG_XEN
 325         BUG_ON(PageReserved(kpte_page));
 326 #else
 327         if(!PageReserved(kpte_page))
 328 #endif
 329
 330         if (page_private(kpte_page) == 0) {
 331                 save_page(kpte_page);
 332                 revert_page(address, ref_prot);
 333         }
 334         return 0;
 335 }
 336
 337 /*
 338  * Change the page attributes of an page in the linear mapping.
 339  *
 340  * This should be used when a page is mapped with a different caching policy
 341  * than write-back somewhere - some CPUs do not like it when mappings with
 342  * different caching policies exist. This changes the page attributes of the
 343  * in kernel linear mapping too.
 344  *
 345  * The caller needs to ensure that there are no conflicting mappings elsewhere.
 346  * This function only deals with the kernel linear map.
 347  *
 348  * Caller must call global_flush_tlb() after this.
 349  */
 350 int change_page_attr_addr(unsigned long address, int numpages, pgprot_t prot)
 351 {
 352         int err = 0;
 353         int i;
 354
 355         down_write(&init_mm.mmap_sem);
 356         for (i = 0; i < numpages; i++, address += PAGE_SIZE) {
 357                 unsigned long pfn = __pa(address) >> PAGE_SHIFT;
 358
 359                 err = __change_page_attr(address, pfn, prot, PAGE_KERNEL);
 360                 if (err)
 361                         break;
 362                 /* Handle kernel mapping too which aliases part of the
 363                  * lowmem */
 364                 if (__pa(address) < KERNEL_TEXT_SIZE) {
 365                         unsigned long addr2;
 366                         pgprot_t prot2;
 367                         addr2 = __START_KERNEL_map + __pa(address);
 368                         /* Make sure the kernel mappings stay executable */
 369                         prot2 = pte_pgprot(pte_mkexec(pfn_pte(0, prot)));
 370                         err = __change_page_attr(addr2, pfn, prot2,
 371                                                  PAGE_KERNEL_EXEC);
 372                 }
 373         }
 374         up_write(&init_mm.mmap_sem);
 375         return err;
 376 }
 377
 378 /* Don't call this for MMIO areas that may not have a mem_map entry */
 379 int change_page_attr(struct page *page, int numpages, pgprot_t prot)
 380 {
 381         unsigned long addr = (unsigned long)page_address(page);
 382         return change_page_attr_addr(addr, numpages, prot);
 383 }
 384
 385 void global_flush_tlb(void)
 386 {
 387         struct page *pg, *next;
 388         struct list_head l;
 389
 390         down_read(&init_mm.mmap_sem);
 391         list_replace_init(&deferred_pages, &l);
 392         up_read(&init_mm.mmap_sem);
 393
 394         flush_map(&l);
 395
 396         list_for_each_entry_safe(pg, next, &l, lru) {
 397                 ClearPagePrivate(pg);
 398                 __free_page(pg);
 399         }
 400 }
 401
 402 EXPORT_SYMBOL(change_page_attr);
 403 EXPORT_SYMBOL(global_flush_tlb);