arch/x86/mm/pgtable.c

   1 #include <linux/mm.h>
   2 #include <asm/pgalloc.h>
   3 #include <asm/pgtable.h>
   4 #include <asm/tlb.h>
   5 #include <asm/fixmap.h>
   6
   7 #define PGALLOC_GFP GFP_KERNEL | __GFP_NOTRACK | __GFP_REPEAT | __GFP_ZERO
   8
   9 #ifdef CONFIG_HIGHPTE
  10 #define PGALLOC_USER_GFP __GFP_HIGHMEM
  11 #else
  12 #define PGALLOC_USER_GFP 0
  13 #endif
  14
  15 gfp_t __userpte_alloc_gfp = PGALLOC_GFP | PGALLOC_USER_GFP;
  16
  17 pte_t *pte_alloc_one_kernel(struct mm_struct *mm, unsigned long address)
  18 {
  19         return (pte_t *)__get_free_page(PGALLOC_GFP);
  20 }
  21
  22 pgtable_t pte_alloc_one(struct mm_struct *mm, unsigned long address)
  23 {
  24         struct page *pte;
  25
  26         pte = alloc_pages(__userpte_alloc_gfp, 0);
  27         if (pte)
  28                 pgtable_page_ctor(pte);
  29         return pte;
  30 }
  31
  32 static int __init setup_userpte(char *arg)
  33 {
  34         if (!arg)
  35                 return -EINVAL;
  36
  37         /*
  38          * "userpte=nohigh" disables allocation of user pagetables in
  39          * high memory.
  40          */
  41         if (strcmp(arg, "nohigh") == 0)
  42                 __userpte_alloc_gfp &= ~__GFP_HIGHMEM;
  43         else
  44                 return -EINVAL;
  45         return 0;
  46 }
  47 early_param("userpte", setup_userpte);
  48
  49 void ___pte_free_tlb(struct mmu_gather *tlb, struct page *pte)
  50 {
  51         pgtable_page_dtor(pte);
  52         paravirt_release_pte(page_to_pfn(pte));
  53         tlb_remove_page(tlb, pte);
  54 }
  55
  56 #if PAGETABLE_LEVELS > 2
  57 void ___pmd_free_tlb(struct mmu_gather *tlb, pmd_t *pmd)
  58 {
  59         paravirt_release_pmd(__pa(pmd) >> PAGE_SHIFT);
  60         tlb_remove_page(tlb, virt_to_page(pmd));
  61 }
  62
  63 #if PAGETABLE_LEVELS > 3
  64 void ___pud_free_tlb(struct mmu_gather *tlb, pud_t *pud)
  65 {
  66         paravirt_release_pud(__pa(pud) >> PAGE_SHIFT);
  67         tlb_remove_page(tlb, virt_to_page(pud));
  68 }
  69 #endif  /* PAGETABLE_LEVELS > 3 */
  70 #endif  /* PAGETABLE_LEVELS > 2 */
  71
  72 static inline void pgd_list_add(pgd_t *pgd)
  73 {
  74         struct page *page = virt_to_page(pgd);
  75
  76         list_add(&page->lru, &pgd_list);
  77 }
  78
  79 static inline void pgd_list_del(pgd_t *pgd)
  80 {
  81         struct page *page = virt_to_page(pgd);
  82
  83         list_del(&page->lru);
  84 }
  85
  86 #define UNSHARED_PTRS_PER_PGD                           \
  87         (SHARED_KERNEL_PMD ? KERNEL_PGD_BOUNDARY : PTRS_PER_PGD)
  88
  89 static void pgd_ctor(pgd_t *pgd)
  90 {
  91         /* If the pgd points to a shared pagetable level (either the
  92            ptes in non-PAE, or shared PMD in PAE), then just copy the
  93            references from swapper_pg_dir. */
  94         if (PAGETABLE_LEVELS == 2 ||
  95             (PAGETABLE_LEVELS == 3 && SHARED_KERNEL_PMD) ||
  96             PAGETABLE_LEVELS == 4) {
  97                 clone_pgd_range(pgd + KERNEL_PGD_BOUNDARY,
  98                                 swapper_pg_dir + KERNEL_PGD_BOUNDARY,
  99                                 KERNEL_PGD_PTRS);
 100                 paravirt_alloc_pmd_clone(__pa(pgd) >> PAGE_SHIFT,
 101                                          __pa(swapper_pg_dir) >> PAGE_SHIFT,
 102                                          KERNEL_PGD_BOUNDARY,
 103                                          KERNEL_PGD_PTRS);
 104         }
 105
 106         /* list required to sync kernel mapping updates */
 107         if (!SHARED_KERNEL_PMD)
 108                 pgd_list_add(pgd);
 109 }
 110
 111 static void pgd_dtor(pgd_t *pgd)
 112 {
 113         unsigned long flags; /* can be called from interrupt context */
 114
 115         if (SHARED_KERNEL_PMD)
 116                 return;
 117
 118         spin_lock_irqsave(&pgd_lock, flags);
 119         pgd_list_del(pgd);
 120         spin_unlock_irqrestore(&pgd_lock, flags);
 121 }
 122
 123 /*
 124  * List of all pgd's needed for non-PAE so it can invalidate entries
 125  * in both cached and uncached pgd's; not needed for PAE since the
 126  * kernel pmd is shared. If PAE were not to share the pmd a similar
 127  * tactic would be needed. This is essentially codepath-based locking
 128  * against pageattr.c; it is the unique case in which a valid change
 129  * of kernel pagetables can't be lazily synchronized by vmalloc faults.
 130  * vmalloc faults work because attached pagetables are never freed.
 131  * -- wli
 132  */
 133
 134 #ifdef CONFIG_X86_PAE
 135 /*
 136  * In PAE mode, we need to do a cr3 reload (=tlb flush) when
 137  * updating the top-level pagetable entries to guarantee the
 138  * processor notices the update.  Since this is expensive, and
 139  * all 4 top-level entries are used almost immediately in a
 140  * new process's life, we just pre-populate them here.
 141  *
 142  * Also, if we're in a paravirt environment where the kernel pmd is
 143  * not shared between pagetables (!SHARED_KERNEL_PMDS), we allocate
 144  * and initialize the kernel pmds here.
 145  */
 146 #define PREALLOCATED_PMDS       UNSHARED_PTRS_PER_PGD
 147
 148 void pud_populate(struct mm_struct *mm, pud_t *pudp, pmd_t *pmd)
 149 {
 150         paravirt_alloc_pmd(mm, __pa(pmd) >> PAGE_SHIFT);
 151
 152         /* Note: almost everything apart from _PAGE_PRESENT is
 153            reserved at the pmd (PDPT) level. */
 154         set_pud(pudp, __pud(__pa(pmd) | _PAGE_PRESENT));
 155
 156         /*
 157          * According to Intel App note "TLBs, Paging-Structure Caches,
 158          * and Their Invalidation", April 2007, document 317080-001,
 159          * section 8.1: in PAE mode we explicitly have to flush the
 160          * TLB via cr3 if the top-level pgd is changed...
 161          */
 162         if (mm == current->active_mm)
 163                 write_cr3(read_cr3());
 164 }
 165 #else  /* !CONFIG_X86_PAE */
 166
 167 /* No need to prepopulate any pagetable entries in non-PAE modes. */
 168 #define PREALLOCATED_PMDS       0
 169
 170 #endif  /* CONFIG_X86_PAE */
 171
 172 static void free_pmds(pmd_t *pmds[])
 173 {
 174         int i;
 175
 176         for(i = 0; i < PREALLOCATED_PMDS; i++)
 177                 if (pmds[i])
 178                         free_page((unsigned long)pmds[i]);
 179 }
 180
 181 static int preallocate_pmds(pmd_t *pmds[])
 182 {
 183         int i;
 184         bool failed = false;
 185
 186         for(i = 0; i < PREALLOCATED_PMDS; i++) {
 187                 pmd_t *pmd = (pmd_t *)__get_free_page(PGALLOC_GFP);
 188                 if (pmd == NULL)
 189                         failed = true;
 190                 pmds[i] = pmd;
 191         }
 192
 193         if (failed) {
 194                 free_pmds(pmds);
 195                 return -ENOMEM;
 196         }
 197
 198         return 0;
 199 }
 200
 201 /*
 202  * Mop up any pmd pages which may still be attached to the pgd.
 203  * Normally they will be freed by munmap/exit_mmap, but any pmd we
 204  * preallocate which never got a corresponding vma will need to be
 205  * freed manually.
 206  */
 207 static void pgd_mop_up_pmds(struct mm_struct *mm, pgd_t *pgdp)
 208 {
 209         int i;
 210
 211         for(i = 0; i < PREALLOCATED_PMDS; i++) {
 212                 pgd_t pgd = pgdp[i];
 213
 214                 if (pgd_val(pgd) != 0) {
 215                         pmd_t *pmd = (pmd_t *)pgd_page_vaddr(pgd);
 216
 217                         pgdp[i] = native_make_pgd(0);
 218
 219                         paravirt_release_pmd(pgd_val(pgd) >> PAGE_SHIFT);
 220                         pmd_free(mm, pmd);
 221                 }
 222         }
 223 }
 224
 225 static void pgd_prepopulate_pmd(struct mm_struct *mm, pgd_t *pgd, pmd_t *pmds[])
 226 {
 227         pud_t *pud;
 228         unsigned long addr;
 229         int i;
 230
 231         if (PREALLOCATED_PMDS == 0) /* Work around gcc-3.4.x bug */
 232                 return;
 233
 234         pud = pud_offset(pgd, 0);
 235
 236         for (addr = i = 0; i < PREALLOCATED_PMDS;
 237              i++, pud++, addr += PUD_SIZE) {
 238                 pmd_t *pmd = pmds[i];
 239
 240                 if (i >= KERNEL_PGD_BOUNDARY)
 241                         memcpy(pmd, (pmd_t *)pgd_page_vaddr(swapper_pg_dir[i]),
 242                                sizeof(pmd_t) * PTRS_PER_PMD);
 243
 244                 pud_populate(mm, pud, pmd);
 245         }
 246 }
 247
 248 pgd_t *pgd_alloc(struct mm_struct *mm)
 249 {
 250         pgd_t *pgd;
 251         pmd_t *pmds[PREALLOCATED_PMDS];
 252         unsigned long flags;
 253
 254         pgd = (pgd_t *)__get_free_page(PGALLOC_GFP);
 255
 256         if (pgd == NULL)
 257                 goto out;
 258
 259         mm->pgd = pgd;
 260
 261         if (preallocate_pmds(pmds) != 0)
 262                 goto out_free_pgd;
 263
 264         if (paravirt_pgd_alloc(mm) != 0)
 265                 goto out_free_pmds;
 266
 267         /*
 268          * Make sure that pre-populating the pmds is atomic with
 269          * respect to anything walking the pgd_list, so that they
 270          * never see a partially populated pgd.
 271          */
 272         spin_lock_irqsave(&pgd_lock, flags);
 273
 274         pgd_ctor(pgd);
 275         pgd_prepopulate_pmd(mm, pgd, pmds);
 276
 277         spin_unlock_irqrestore(&pgd_lock, flags);
 278
 279         return pgd;
 280
 281 out_free_pmds:
 282         free_pmds(pmds);
 283 out_free_pgd:
 284         free_page((unsigned long)pgd);
 285 out:
 286         return NULL;
 287 }
 288
 289 void pgd_free(struct mm_struct *mm, pgd_t *pgd)
 290 {
 291         pgd_mop_up_pmds(mm, pgd);
 292         pgd_dtor(pgd);
 293         paravirt_pgd_free(mm, pgd);
 294         free_page((unsigned long)pgd);
 295 }
 296
 297 int ptep_set_access_flags(struct vm_area_struct *vma,
 298                           unsigned long address, pte_t *ptep,
 299                           pte_t entry, int dirty)
 300 {
 301         int changed = !pte_same(*ptep, entry);
 302
 303         if (changed && dirty) {
 304                 *ptep = entry;
 305                 pte_update_defer(vma->vm_mm, address, ptep);
 306                 flush_tlb_page(vma, address);
 307         }
 308
 309         return changed;
 310 }
 311
 312 int ptep_test_and_clear_young(struct vm_area_struct *vma,
 313                               unsigned long addr, pte_t *ptep)
 314 {
 315         int ret = 0;
 316
 317         if (pte_young(*ptep))
 318                 ret = test_and_clear_bit(_PAGE_BIT_ACCESSED,
 319                                          (unsigned long *) &ptep->pte);
 320
 321         if (ret)
 322                 pte_update(vma->vm_mm, addr, ptep);
 323
 324         return ret;
 325 }
 326
 327 int ptep_clear_flush_young(struct vm_area_struct *vma,
 328                            unsigned long address, pte_t *ptep)
 329 {
 330         int young;
 331
 332         young = ptep_test_and_clear_young(vma, address, ptep);
 333         if (young)
 334                 flush_tlb_page(vma, address);
 335
 336         return young;
 337 }
 338
 339 /**
 340  * reserve_top_address - reserves a hole in the top of kernel address space
 341  * @reserve - size of hole to reserve
 342  *
 343  * Can be used to relocate the fixmap area and poke a hole in the top
 344  * of kernel address space to make room for a hypervisor.
 345  */
 346 void __init reserve_top_address(unsigned long reserve)
 347 {
 348 #ifdef CONFIG_X86_32
 349         BUG_ON(fixmaps_set > 0);
 350         printk(KERN_INFO "Reserving virtual address space above 0x%08x\n",
 351                (int)-reserve);
 352         __FIXADDR_TOP = -reserve - PAGE_SIZE;
 353 #endif
 354 }
 355
 356 int fixmaps_set;
 357
 358 void __native_set_fixmap(enum fixed_addresses idx, pte_t pte)
 359 {
 360         unsigned long address = __fix_to_virt(idx);
 361
 362         if (idx >= __end_of_fixed_addresses) {
 363                 BUG();
 364                 return;
 365         }
 366         set_pte_vaddr(address, pte);
 367         fixmaps_set++;
 368 }
 369
 370 void native_set_fixmap(enum fixed_addresses idx, phys_addr_t phys,
 371                        pgprot_t flags)
 372 {
 373         __native_set_fixmap(idx, pfn_pte(phys >> PAGE_SHIFT, flags));
 374 }