mm/memory.c

   1 /*
   2  *  linux/mm/memory.c
   3  *
   4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
   5  */
   6
   7 /*
   8  * demand-loading started 01.12.91 - seems it is high on the list of
   9  * things wanted, and it should be easy to implement. - Linus
  10  */
  11
  12 /*
  13  * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
  14  * pages started 02.12.91, seems to work. - Linus.
  15  *
  16  * Tested sharing by executing about 30 /bin/sh: under the old kernel it
  17  * would have taken more than the 6M I have free, but it worked well as
  18  * far as I could see.
  19  *
  20  * Also corrected some "invalidate()"s - I wasn't doing enough of them.
  21  */
  22
  23 /*
  24  * Real VM (paging to/from disk) started 18.12.91. Much more work and
  25  * thought has to go into this. Oh, well..
  26  * 19.12.91  -  works, somewhat. Sometimes I get faults, don't know why.
  27  *              Found it. Everything seems to work now.
  28  * 20.12.91  -  Ok, making the swap-device changeable like the root.
  29  */
  30
  31 /*
  32  * 05.04.94  -  Multi-page memory management added for v1.1.
  33  *              Idea by Alex Bligh (alex@cconcepts.co.uk)
  34  *
  35  * 16.07.99  -  Support of BIGMEM added by Gerhard Wichert, Siemens AG
  36  *              (Gerhard.Wichert@pdb.siemens.de)
  37  */
  38
  39 #include <linux/kernel_stat.h>
  40 #include <linux/mm.h>
  41 #include <linux/hugetlb.h>
  42 #include <linux/mman.h>
  43 #include <linux/swap.h>
  44 #include <linux/highmem.h>
  45 #include <linux/pagemap.h>
  46 #include <linux/vcache.h>
  47 #include <linux/rmap-locking.h>
  48
  49 #include <asm/pgalloc.h>
  50 #include <asm/rmap.h>
  51 #include <asm/uaccess.h>
  52 #include <asm/tlb.h>
  53 #include <asm/tlbflush.h>
  54 #include <asm/pgtable.h>
  55
  56 #include <linux/swapops.h>
  57
  58 #ifndef CONFIG_DISCONTIGMEM
  59 /* use the per-pgdat data instead for discontigmem - mbligh */
  60 unsigned long max_mapnr;
  61 struct page *mem_map;
  62 #endif
  63
  64 unsigned long num_physpages;
  65 void * high_memory;
  66 struct page *highmem_start_page;
  67
  68 /*
  69  * We special-case the C-O-W ZERO_PAGE, because it's such
  70  * a common occurrence (no need to read the page to know
  71  * that it's zero - better for the cache and memory subsystem).
  72  */
  73 static inline void copy_cow_page(struct page * from, struct page * to, unsigned long address)
  74 {
  75         if (from == ZERO_PAGE(address)) {
  76                 clear_user_highpage(to, address);
  77                 return;
  78         }
  79         copy_user_highpage(to, from, address);
  80 }
  81
  82 /*
  83  * Note: this doesn't free the actual pages themselves. That
  84  * has been handled earlier when unmapping all the memory regions.
  85  */
  86 static inline void free_one_pmd(struct mmu_gather *tlb, pmd_t * dir)
  87 {
  88         struct page *page;
  89
  90         if (pmd_none(*dir))
  91                 return;
  92         if (pmd_bad(*dir)) {
  93                 pmd_ERROR(*dir);
  94                 pmd_clear(dir);
  95                 return;
  96         }
  97         page = pmd_page(*dir);
  98         pmd_clear(dir);
  99         pgtable_remove_rmap(page);
 100         pte_free_tlb(tlb, page);
 101 }
 102
 103 static inline void free_one_pgd(struct mmu_gather *tlb, pgd_t * dir)
 104 {
 105         int j;
 106         pmd_t * pmd;
 107
 108         if (pgd_none(*dir))
 109                 return;
 110         if (pgd_bad(*dir)) {
 111                 pgd_ERROR(*dir);
 112                 pgd_clear(dir);
 113                 return;
 114         }
 115         pmd = pmd_offset(dir, 0);
 116         pgd_clear(dir);
 117         for (j = 0; j < PTRS_PER_PMD ; j++)
 118                 free_one_pmd(tlb, pmd+j);
 119         pmd_free_tlb(tlb, pmd);
 120 }
 121
 122 /*
 123  * This function clears all user-level page tables of a process - this
 124  * is needed by execve(), so that old pages aren't in the way.
 125  *
 126  * Must be called with pagetable lock held.
 127  */
 128 void clear_page_tables(struct mmu_gather *tlb, unsigned long first, int nr)
 129 {
 130         pgd_t * page_dir = tlb->mm->pgd;
 131
 132         page_dir += first;
 133         do {
 134                 free_one_pgd(tlb, page_dir);
 135                 page_dir++;
 136         } while (--nr);
 137 }
 138
 139 pte_t * pte_alloc_map(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
 140 {
 141         if (!pmd_present(*pmd)) {
 142                 struct page *new;
 143
 144                 spin_unlock(&mm->page_table_lock);
 145                 new = pte_alloc_one(mm, address);
 146                 spin_lock(&mm->page_table_lock);
 147                 if (!new)
 148                         return NULL;
 149
 150                 /*
 151                  * Because we dropped the lock, we should re-check the
 152                  * entry, as somebody else could have populated it..
 153                  */
 154                 if (pmd_present(*pmd)) {
 155                         pte_free(new);
 156                         goto out;
 157                 }
 158                 pgtable_add_rmap(new, mm, address);
 159                 pmd_populate(mm, pmd, new);
 160         }
 161 out:
 162         return pte_offset_map(pmd, address);
 163 }
 164
 165 pte_t * pte_alloc_kernel(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
 166 {
 167         if (!pmd_present(*pmd)) {
 168                 pte_t *new;
 169
 170                 spin_unlock(&mm->page_table_lock);
 171                 new = pte_alloc_one_kernel(mm, address);
 172                 spin_lock(&mm->page_table_lock);
 173                 if (!new)
 174                         return NULL;
 175
 176                 /*
 177                  * Because we dropped the lock, we should re-check the
 178                  * entry, as somebody else could have populated it..
 179                  */
 180                 if (pmd_present(*pmd)) {
 181                         pte_free_kernel(new);
 182                         goto out;
 183                 }
 184                 pgtable_add_rmap(virt_to_page(new), mm, address);
 185                 pmd_populate_kernel(mm, pmd, new);
 186         }
 187 out:
 188         return pte_offset_kernel(pmd, address);
 189 }
 190 #define PTE_TABLE_MASK  ((PTRS_PER_PTE-1) * sizeof(pte_t))
 191 #define PMD_TABLE_MASK  ((PTRS_PER_PMD-1) * sizeof(pmd_t))
 192
 193 /*
 194  * copy one vm_area from one task to the other. Assumes the page tables
 195  * already present in the new task to be cleared in the whole range
 196  * covered by this vma.
 197  *
 198  * 08Jan98 Merged into one routine from several inline routines to reduce
 199  *         variable count and make things faster. -jj
 200  *
 201  * dst->page_table_lock is held on entry and exit,
 202  * but may be dropped within pmd_alloc() and pte_alloc_map().
 203  */
 204 int copy_page_range(struct mm_struct *dst, struct mm_struct *src,
 205                         struct vm_area_struct *vma)
 206 {
 207         pgd_t * src_pgd, * dst_pgd;
 208         unsigned long address = vma->vm_start;
 209         unsigned long end = vma->vm_end;
 210         unsigned long cow;
 211         struct pte_chain *pte_chain = NULL;
 212
 213         if (is_vm_hugetlb_page(vma))
 214                 return copy_hugetlb_page_range(dst, src, vma);
 215
 216         pte_chain = pte_chain_alloc(GFP_ATOMIC);
 217         if (!pte_chain) {
 218                 spin_unlock(&dst->page_table_lock);
 219                 pte_chain = pte_chain_alloc(GFP_KERNEL);
 220                 spin_lock(&dst->page_table_lock);
 221                 if (!pte_chain)
 222                         goto nomem;
 223         }
 224
 225         cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
 226         src_pgd = pgd_offset(src, address)-1;
 227         dst_pgd = pgd_offset(dst, address)-1;
 228
 229         for (;;) {
 230                 pmd_t * src_pmd, * dst_pmd;
 231
 232                 src_pgd++; dst_pgd++;
 233
 234                 /* copy_pmd_range */
 235
 236                 if (pgd_none(*src_pgd))
 237                         goto skip_copy_pmd_range;
 238                 if (pgd_bad(*src_pgd)) {
 239                         pgd_ERROR(*src_pgd);
 240                         pgd_clear(src_pgd);
 241 skip_copy_pmd_range:    address = (address + PGDIR_SIZE) & PGDIR_MASK;
 242                         if (!address || (address >= end))
 243                                 goto out;
 244                         continue;
 245                 }
 246
 247                 src_pmd = pmd_offset(src_pgd, address);
 248                 dst_pmd = pmd_alloc(dst, dst_pgd, address);
 249                 if (!dst_pmd)
 250                         goto nomem;
 251
 252                 do {
 253                         pte_t * src_pte, * dst_pte;
 254
 255                         /* copy_pte_range */
 256
 257                         if (pmd_none(*src_pmd))
 258                                 goto skip_copy_pte_range;
 259                         if (pmd_bad(*src_pmd)) {
 260                                 pmd_ERROR(*src_pmd);
 261                                 pmd_clear(src_pmd);
 262 skip_copy_pte_range:
 263                                 address = (address + PMD_SIZE) & PMD_MASK;
 264                                 if (address >= end)
 265                                         goto out;
 266                                 goto cont_copy_pmd_range;
 267                         }
 268
 269                         dst_pte = pte_alloc_map(dst, dst_pmd, address);
 270                         if (!dst_pte)
 271                                 goto nomem;
 272                         spin_lock(&src->page_table_lock);
 273                         src_pte = pte_offset_map_nested(src_pmd, address);
 274                         do {
 275                                 pte_t pte = *src_pte;
 276                                 struct page *page;
 277                                 unsigned long pfn;
 278
 279                                 /* copy_one_pte */
 280
 281                                 if (pte_none(pte))
 282                                         goto cont_copy_pte_range_noset;
 283                                 /* pte contains position in swap, so copy. */
 284                                 if (!pte_present(pte)) {
 285                                         if (!pte_file(pte))
 286                                                 swap_duplicate(pte_to_swp_entry(pte));
 287                                         set_pte(dst_pte, pte);
 288                                         goto cont_copy_pte_range_noset;
 289                                 }
 290                                 pfn = pte_pfn(pte);
 291                                 /* the pte points outside of valid memory, the
 292                                  * mapping is assumed to be good, meaningful
 293                                  * and not mapped via rmap - duplicate the
 294                                  * mapping as is.
 295                                  */
 296                                 page = NULL;
 297                                 if (pfn_valid(pfn))
 298                                         page = pfn_to_page(pfn);
 299
 300                                 if (!page || PageReserved(page)) {
 301                                         set_pte(dst_pte, pte);
 302                                         goto cont_copy_pte_range_noset;
 303                                 }
 304
 305                                 /*
 306                                  * If it's a COW mapping, write protect it both
 307                                  * in the parent and the child
 308                                  */
 309                                 if (cow) {
 310                                         ptep_set_wrprotect(src_pte);
 311                                         pte = *src_pte;
 312                                 }
 313
 314                                 /*
 315                                  * If it's a shared mapping, mark it clean in
 316                                  * the child
 317                                  */
 318                                 if (vma->vm_flags & VM_SHARED)
 319                                         pte = pte_mkclean(pte);
 320                                 pte = pte_mkold(pte);
 321                                 get_page(page);
 322                                 dst->rss++;
 323
 324                                 set_pte(dst_pte, pte);
 325                                 pte_chain = page_add_rmap(page, dst_pte,
 326                                                         pte_chain);
 327                                 if (pte_chain)
 328                                         goto cont_copy_pte_range_noset;
 329                                 pte_chain = pte_chain_alloc(GFP_ATOMIC);
 330                                 if (pte_chain)
 331                                         goto cont_copy_pte_range_noset;
 332
 333                                 /*
 334                                  * pte_chain allocation failed, and we need to
 335                                  * run page reclaim.
 336                                  */
 337                                 pte_unmap_nested(src_pte);
 338                                 pte_unmap(dst_pte);
 339                                 spin_unlock(&src->page_table_lock);
 340                                 spin_unlock(&dst->page_table_lock);
 341                                 pte_chain = pte_chain_alloc(GFP_KERNEL);
 342                                 spin_lock(&dst->page_table_lock);
 343                                 if (!pte_chain)
 344                                         goto nomem;
 345                                 spin_lock(&src->page_table_lock);
 346                                 dst_pte = pte_offset_map(dst_pmd, address);
 347                                 src_pte = pte_offset_map_nested(src_pmd,
 348                                                                 address);
 349 cont_copy_pte_range_noset:
 350                                 address += PAGE_SIZE;
 351                                 if (address >= end) {
 352                                         pte_unmap_nested(src_pte);
 353                                         pte_unmap(dst_pte);
 354                                         goto out_unlock;
 355                                 }
 356                                 src_pte++;
 357                                 dst_pte++;
 358                         } while ((unsigned long)src_pte & PTE_TABLE_MASK);
 359                         pte_unmap_nested(src_pte-1);
 360                         pte_unmap(dst_pte-1);
 361                         spin_unlock(&src->page_table_lock);
 362
 363 cont_copy_pmd_range:
 364                         src_pmd++;
 365                         dst_pmd++;
 366                 } while ((unsigned long)src_pmd & PMD_TABLE_MASK);
 367         }
 368 out_unlock:
 369         spin_unlock(&src->page_table_lock);
 370 out:
 371         pte_chain_free(pte_chain);
 372         return 0;
 373 nomem:
 374         pte_chain_free(pte_chain);
 375         return -ENOMEM;
 376 }
 377
 378 static void
 379 zap_pte_range(struct mmu_gather *tlb, pmd_t * pmd,
 380                 unsigned long address, unsigned long size)
 381 {
 382         unsigned long offset;
 383         pte_t *ptep;
 384
 385         if (pmd_none(*pmd))
 386                 return;
 387         if (pmd_bad(*pmd)) {
 388                 pmd_ERROR(*pmd);
 389                 pmd_clear(pmd);
 390                 return;
 391         }
 392         ptep = pte_offset_map(pmd, address);
 393         offset = address & ~PMD_MASK;
 394         if (offset + size > PMD_SIZE)
 395                 size = PMD_SIZE - offset;
 396         size &= PAGE_MASK;
 397         for (offset=0; offset < size; ptep++, offset += PAGE_SIZE) {
 398                 pte_t pte = *ptep;
 399                 if (pte_none(pte))
 400                         continue;
 401                 if (pte_present(pte)) {
 402                         unsigned long pfn = pte_pfn(pte);
 403
 404                         pte = ptep_get_and_clear(ptep);
 405                         tlb_remove_tlb_entry(tlb, ptep, address+offset);
 406                         if (pfn_valid(pfn)) {
 407                                 struct page *page = pfn_to_page(pfn);
 408                                 if (!PageReserved(page)) {
 409                                         if (pte_dirty(pte))
 410                                                 set_page_dirty(page);
 411                                         if (page->mapping && pte_young(pte) &&
 412                                                         !PageSwapCache(page))
 413                                                 mark_page_accessed(page);
 414                                         tlb->freed++;
 415                                         page_remove_rmap(page, ptep);
 416                                         tlb_remove_page(tlb, page);
 417                                 }
 418                         }
 419                 } else {
 420                         if (!pte_file(pte))
 421                                 free_swap_and_cache(pte_to_swp_entry(pte));
 422                         pte_clear(ptep);
 423                 }
 424         }
 425         pte_unmap(ptep-1);
 426 }
 427
 428 static void
 429 zap_pmd_range(struct mmu_gather *tlb, pgd_t * dir,
 430                 unsigned long address, unsigned long size)
 431 {
 432         pmd_t * pmd;
 433         unsigned long end;
 434
 435         if (pgd_none(*dir))
 436                 return;
 437         if (pgd_bad(*dir)) {
 438                 pgd_ERROR(*dir);
 439                 pgd_clear(dir);
 440                 return;
 441         }
 442         pmd = pmd_offset(dir, address);
 443         end = address + size;
 444         if (end > ((address + PGDIR_SIZE) & PGDIR_MASK))
 445                 end = ((address + PGDIR_SIZE) & PGDIR_MASK);
 446         do {
 447                 zap_pte_range(tlb, pmd, address, end - address);
 448                 address = (address + PMD_SIZE) & PMD_MASK;
 449                 pmd++;
 450         } while (address < end);
 451 }
 452
 453 void unmap_page_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
 454                         unsigned long address, unsigned long end)
 455 {
 456         pgd_t * dir;
 457
 458         if (is_vm_hugetlb_page(vma)) {
 459                 unmap_hugepage_range(vma, address, end);
 460                 return;
 461         }
 462
 463         BUG_ON(address >= end);
 464
 465         dir = pgd_offset(vma->vm_mm, address);
 466         tlb_start_vma(tlb, vma);
 467         do {
 468                 zap_pmd_range(tlb, dir, address, end - address);
 469                 address = (address + PGDIR_SIZE) & PGDIR_MASK;
 470                 dir++;
 471         } while (address && (address < end));
 472         tlb_end_vma(tlb, vma);
 473 }
 474
 475 /* Dispose of an entire struct mmu_gather per rescheduling point */
 476 #if defined(CONFIG_SMP) && defined(CONFIG_PREEMPT)
 477 #define ZAP_BLOCK_SIZE  (FREE_PTE_NR * PAGE_SIZE)
 478 #endif
 479
 480 /* For UP, 256 pages at a time gives nice low latency */
 481 #if !defined(CONFIG_SMP) && defined(CONFIG_PREEMPT)
 482 #define ZAP_BLOCK_SIZE  (256 * PAGE_SIZE)
 483 #endif
 484
 485 /* No preempt: go for the best straight-line efficiency */
 486 #if !defined(CONFIG_PREEMPT)
 487 #define ZAP_BLOCK_SIZE  (~(0UL))
 488 #endif
 489
 490 /**
 491  * unmap_vmas - unmap a range of memory covered by a list of vma's
 492  * @tlbp: address of the caller's struct mmu_gather
 493  * @mm: the controlling mm_struct
 494  * @vma: the starting vma
 495  * @start_addr: virtual address at which to start unmapping
 496  * @end_addr: virtual address at which to end unmapping
 497  * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
 498  *
 499  * Returns the number of vma's which were covered by the unmapping.
 500  *
 501  * Unmap all pages in the vma list.  Called under page_table_lock.
 502  *
 503  * We aim to not hold page_table_lock for too long (for scheduling latency
 504  * reasons).  So zap pages in ZAP_BLOCK_SIZE bytecounts.  This means we need to
 505  * return the ending mmu_gather to the caller.
 506  *
 507  * Only addresses between `start' and `end' will be unmapped.
 508  *
 509  * The VMA list must be sorted in ascending virtual address order.
 510  *
 511  * unmap_vmas() assumes that the caller will flush the whole unmapped address
 512  * range after unmap_vmas() returns.  So the only responsibility here is to
 513  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
 514  * drops the lock and schedules.
 515  */
 516 int unmap_vmas(struct mmu_gather **tlbp, struct mm_struct *mm,
 517                 struct vm_area_struct *vma, unsigned long start_addr,
 518                 unsigned long end_addr, unsigned long *nr_accounted)
 519 {
 520         unsigned long zap_bytes = ZAP_BLOCK_SIZE;
 521         unsigned long tlb_start;        /* For tlb_finish_mmu */
 522         int tlb_start_valid = 0;
 523         int ret = 0;
 524
 525         if (vma) {      /* debug.  killme. */
 526                 if (end_addr <= vma->vm_start)
 527                         printk("%s: end_addr(0x%08lx) <= vm_start(0x%08lx)\n",
 528                                 __FUNCTION__, end_addr, vma->vm_start);
 529                 if (start_addr >= vma->vm_end)
 530                         printk("%s: start_addr(0x%08lx) <= vm_end(0x%08lx)\n",
 531                                 __FUNCTION__, start_addr, vma->vm_end);
 532         }
 533
 534         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
 535                 unsigned long start;
 536                 unsigned long end;
 537
 538                 start = max(vma->vm_start, start_addr);
 539                 if (start >= vma->vm_end)
 540                         continue;
 541                 end = min(vma->vm_end, end_addr);
 542                 if (end <= vma->vm_start)
 543                         continue;
 544
 545                 if (vma->vm_flags & VM_ACCOUNT)
 546                         *nr_accounted += (end - start) >> PAGE_SHIFT;
 547
 548                 ret++;
 549                 while (start != end) {
 550                         unsigned long block;
 551
 552                         if (is_vm_hugetlb_page(vma))
 553                                 block = end - start;
 554                         else
 555                                 block = min(zap_bytes, end - start);
 556
 557                         if (!tlb_start_valid) {
 558                                 tlb_start = start;
 559                                 tlb_start_valid = 1;
 560                         }
 561
 562                         unmap_page_range(*tlbp, vma, start, start + block);
 563                         start += block;
 564                         zap_bytes -= block;
 565                         if ((long)zap_bytes > 0)
 566                                 continue;
 567                         if (need_resched()) {
 568                                 tlb_finish_mmu(*tlbp, tlb_start, start);
 569                                 cond_resched_lock(&mm->page_table_lock);
 570                                 *tlbp = tlb_gather_mmu(mm, 0);
 571                                 tlb_start_valid = 0;
 572                         }
 573                         zap_bytes = ZAP_BLOCK_SIZE;
 574                 }
 575                 if (vma->vm_next && vma->vm_next->vm_start < vma->vm_end)
 576                         printk("%s: VMA list is not sorted correctly!\n",
 577                                 __FUNCTION__);
 578         }
 579         return ret;
 580 }
 581
 582 /**
 583  * zap_page_range - remove user pages in a given range
 584  * @vma: vm_area_struct holding the applicable pages
 585  * @address: starting address of pages to zap
 586  * @size: number of bytes to zap
 587  */
 588 void zap_page_range(struct vm_area_struct *vma,
 589                         unsigned long address, unsigned long size)
 590 {
 591         struct mm_struct *mm = vma->vm_mm;
 592         struct mmu_gather *tlb;
 593         unsigned long end = address + size;
 594         unsigned long nr_accounted = 0;
 595
 596         might_sleep();
 597
 598         if (is_vm_hugetlb_page(vma)) {
 599                 zap_hugepage_range(vma, address, size);
 600                 return;
 601         }
 602
 603         lru_add_drain();
 604         spin_lock(&mm->page_table_lock);
 605         tlb = tlb_gather_mmu(mm, 0);
 606         unmap_vmas(&tlb, mm, vma, address, end, &nr_accounted);
 607         tlb_finish_mmu(tlb, address, end);
 608         spin_unlock(&mm->page_table_lock);
 609 }
 610
 611 /*
 612  * Do a quick page-table lookup for a single page.
 613  * mm->page_table_lock must be held.
 614  */
 615 struct page *
 616 follow_page(struct mm_struct *mm, unsigned long address, int write)
 617 {
 618         pgd_t *pgd;
 619         pmd_t *pmd;
 620         pte_t *ptep, pte;
 621         unsigned long pfn;
 622         struct vm_area_struct *vma;
 623
 624         vma = hugepage_vma(mm, address);
 625         if (vma)
 626                 return follow_huge_addr(mm, vma, address, write);
 627
 628         pgd = pgd_offset(mm, address);
 629         if (pgd_none(*pgd) || pgd_bad(*pgd))
 630                 goto out;
 631
 632         pmd = pmd_offset(pgd, address);
 633         if (pmd_none(*pmd))
 634                 goto out;
 635         if (pmd_huge(*pmd))
 636                 return follow_huge_pmd(mm, address, pmd, write);
 637         if (pmd_bad(*pmd))
 638                 goto out;
 639
 640         ptep = pte_offset_map(pmd, address);
 641         if (!ptep)
 642                 goto out;
 643
 644         pte = *ptep;
 645         pte_unmap(ptep);
 646         if (pte_present(pte)) {
 647                 if (!write || (pte_write(pte) && pte_dirty(pte))) {
 648                         pfn = pte_pfn(pte);
 649                         if (pfn_valid(pfn))
 650                                 return pfn_to_page(pfn);
 651                 }
 652         }
 653
 654 out:
 655         return NULL;
 656 }
 657
 658 /*
 659  * Given a physical address, is there a useful struct page pointing to
 660  * it?  This may become more complex in the future if we start dealing
 661  * with IO-aperture pages for direct-IO.
 662  */
 663
 664 static inline struct page *get_page_map(struct page *page)
 665 {
 666         if (!pfn_valid(page_to_pfn(page)))
 667                 return 0;
 668         return page;
 669 }
 670
 671
 672 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
 673                 unsigned long start, int len, int write, int force,
 674                 struct page **pages, struct vm_area_struct **vmas)
 675 {
 676         int i;
 677         unsigned int flags;
 678
 679         /*
 680          * Require read or write permissions.
 681          * If 'force' is set, we only require the "MAY" flags.
 682          */
 683         flags = write ? (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
 684         flags &= force ? (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
 685         i = 0;
 686
 687         do {
 688                 struct vm_area_struct * vma;
 689
 690                 vma = find_extend_vma(mm, start);
 691
 692 #ifdef VSYSCALL_START
 693                 if (!vma && start >= VSYSCALL_START && start < VSYSCALL_END) {
 694                         static struct vm_area_struct fixmap_vma = {
 695                                 /* Catch users - if there are any valid
 696                                    ones, we can make this be "&init_mm" or
 697                                    something.  */
 698                                 .vm_mm = NULL,
 699                                 .vm_start = FIXADDR_START,
 700                                 .vm_end = FIXADDR_TOP,
 701                                 .vm_page_prot = PAGE_READONLY,
 702                                 .vm_flags = VM_READ | VM_EXEC,
 703                         };
 704                         unsigned long pg = start & PAGE_MASK;
 705                         pgd_t *pgd;
 706                         pmd_t *pmd;
 707                         pte_t *pte;
 708                         pgd = pgd_offset_k(pg);
 709                         if (!pgd)
 710                                 return i ? : -EFAULT;
 711                         pmd = pmd_offset(pgd, pg);
 712                         if (!pmd)
 713                                 return i ? : -EFAULT;
 714                         pte = pte_offset_kernel(pmd, pg);
 715                         if (!pte || !pte_present(*pte) || !pte_user(*pte) ||
 716                             !(write ? pte_write(*pte) : pte_read(*pte)))
 717                                 return i ? : -EFAULT;
 718                         if (pages) {
 719                                 pages[i] = pte_page(*pte);
 720                                 get_page(pages[i]);
 721                         }
 722                         if (vmas)
 723                                 vmas[i] = &fixmap_vma;
 724                         i++;
 725                         start += PAGE_SIZE;
 726                         len--;
 727                         continue;
 728                 }
 729 #endif
 730
 731                 if (!vma || (pages && (vma->vm_flags & VM_IO))
 732                                 || !(flags & vma->vm_flags))
 733                         return i ? : -EFAULT;
 734
 735                 if (is_vm_hugetlb_page(vma)) {
 736                         i = follow_hugetlb_page(mm, vma, pages, vmas,
 737                                                 &start, &len, i);
 738                         continue;
 739                 }
 740                 spin_lock(&mm->page_table_lock);
 741                 do {
 742                         struct page *map;
 743                         while (!(map = follow_page(mm, start, write))) {
 744                                 spin_unlock(&mm->page_table_lock);
 745                                 switch (handle_mm_fault(mm,vma,start,write)) {
 746                                 case VM_FAULT_MINOR:
 747                                         tsk->min_flt++;
 748                                         break;
 749                                 case VM_FAULT_MAJOR:
 750                                         tsk->maj_flt++;
 751                                         break;
 752                                 case VM_FAULT_SIGBUS:
 753                                         return i ? i : -EFAULT;
 754                                 case VM_FAULT_OOM:
 755                                         return i ? i : -ENOMEM;
 756                                 default:
 757                                         BUG();
 758                                 }
 759                                 spin_lock(&mm->page_table_lock);
 760                         }
 761                         if (pages) {
 762                                 pages[i] = get_page_map(map);
 763                                 if (!pages[i]) {
 764                                         spin_unlock(&mm->page_table_lock);
 765                                         while (i--)
 766                                                 page_cache_release(pages[i]);
 767                                         i = -EFAULT;
 768                                         goto out;
 769                                 }
 770                                 flush_dcache_page(pages[i]);
 771                                 if (!PageReserved(pages[i]))
 772                                         page_cache_get(pages[i]);
 773                         }
 774                         if (vmas)
 775                                 vmas[i] = vma;
 776                         i++;
 777                         start += PAGE_SIZE;
 778                         len--;
 779                 } while(len && start < vma->vm_end);
 780                 spin_unlock(&mm->page_table_lock);
 781         } while(len);
 782 out:
 783         return i;
 784 }
 785
 786 static void zeromap_pte_range(pte_t * pte, unsigned long address,
 787                                      unsigned long size, pgprot_t prot)
 788 {
 789         unsigned long end;
 790
 791         address &= ~PMD_MASK;
 792         end = address + size;
 793         if (end > PMD_SIZE)
 794                 end = PMD_SIZE;
 795         do {
 796                 pte_t zero_pte = pte_wrprotect(mk_pte(ZERO_PAGE(address), prot));
 797                 BUG_ON(!pte_none(*pte));
 798                 set_pte(pte, zero_pte);
 799                 address += PAGE_SIZE;
 800                 pte++;
 801         } while (address && (address < end));
 802 }
 803
 804 static inline int zeromap_pmd_range(struct mm_struct *mm, pmd_t * pmd, unsigned long address,
 805                                     unsigned long size, pgprot_t prot)
 806 {
 807         unsigned long end;
 808
 809         address &= ~PGDIR_MASK;
 810         end = address + size;
 811         if (end > PGDIR_SIZE)
 812                 end = PGDIR_SIZE;
 813         do {
 814                 pte_t * pte = pte_alloc_map(mm, pmd, address);
 815                 if (!pte)
 816                         return -ENOMEM;
 817                 zeromap_pte_range(pte, address, end - address, prot);
 818                 pte_unmap(pte);
 819                 address = (address + PMD_SIZE) & PMD_MASK;
 820                 pmd++;
 821         } while (address && (address < end));
 822         return 0;
 823 }
 824
 825 int zeromap_page_range(struct vm_area_struct *vma, unsigned long address, unsigned long size, pgprot_t prot)
 826 {
 827         int error = 0;
 828         pgd_t * dir;
 829         unsigned long beg = address;
 830         unsigned long end = address + size;
 831         struct mm_struct *mm = vma->vm_mm;
 832
 833         dir = pgd_offset(mm, address);
 834         flush_cache_range(vma, beg, end);
 835         if (address >= end)
 836                 BUG();
 837
 838         spin_lock(&mm->page_table_lock);
 839         do {
 840                 pmd_t *pmd = pmd_alloc(mm, dir, address);
 841                 error = -ENOMEM;
 842                 if (!pmd)
 843                         break;
 844                 error = zeromap_pmd_range(mm, pmd, address, end - address, prot);
 845                 if (error)
 846                         break;
 847                 address = (address + PGDIR_SIZE) & PGDIR_MASK;
 848                 dir++;
 849         } while (address && (address < end));
 850         flush_tlb_range(vma, beg, end);
 851         spin_unlock(&mm->page_table_lock);
 852         return error;
 853 }
 854
 855 /*
 856  * maps a range of physical memory into the requested pages. the old
 857  * mappings are removed. any references to nonexistent pages results
 858  * in null mappings (currently treated as "copy-on-access")
 859  */
 860 static inline void remap_pte_range(pte_t * pte, unsigned long address, unsigned long size,
 861         unsigned long phys_addr, pgprot_t prot)
 862 {
 863         unsigned long end;
 864         unsigned long pfn;
 865
 866         address &= ~PMD_MASK;
 867         end = address + size;
 868         if (end > PMD_SIZE)
 869                 end = PMD_SIZE;
 870         pfn = phys_addr >> PAGE_SHIFT;
 871         do {
 872                 BUG_ON(!pte_none(*pte));
 873                 if (!pfn_valid(pfn) || PageReserved(pfn_to_page(pfn)))
 874                         set_pte(pte, pfn_pte(pfn, prot));
 875                 address += PAGE_SIZE;
 876                 pfn++;
 877                 pte++;
 878         } while (address && (address < end));
 879 }
 880
 881 static inline int remap_pmd_range(struct mm_struct *mm, pmd_t * pmd, unsigned long address, unsigned long size,
 882         unsigned long phys_addr, pgprot_t prot)
 883 {
 884         unsigned long base, end;
 885
 886         base = address & PGDIR_MASK;
 887         address &= ~PGDIR_MASK;
 888         end = address + size;
 889         if (end > PGDIR_SIZE)
 890                 end = PGDIR_SIZE;
 891         phys_addr -= address;
 892         do {
 893                 pte_t * pte = pte_alloc_map(mm, pmd, base + address);
 894                 if (!pte)
 895                         return -ENOMEM;
 896                 remap_pte_range(pte, base + address, end - address, address + phys_addr, prot);
 897                 pte_unmap(pte);
 898                 address = (address + PMD_SIZE) & PMD_MASK;
 899                 pmd++;
 900         } while (address && (address < end));
 901         return 0;
 902 }
 903
 904 /*  Note: this is only safe if the mm semaphore is held when called. */
 905 int remap_page_range(struct vm_area_struct *vma, unsigned long from, unsigned long phys_addr, unsigned long size, pgprot_t prot)
 906 {
 907         int error = 0;
 908         pgd_t * dir;
 909         unsigned long beg = from;
 910         unsigned long end = from + size;
 911         struct mm_struct *mm = vma->vm_mm;
 912
 913         phys_addr -= from;
 914         dir = pgd_offset(mm, from);
 915         flush_cache_range(vma, beg, end);
 916         if (from >= end)
 917                 BUG();
 918
 919         spin_lock(&mm->page_table_lock);
 920         do {
 921                 pmd_t *pmd = pmd_alloc(mm, dir, from);
 922                 error = -ENOMEM;
 923                 if (!pmd)
 924                         break;
 925                 error = remap_pmd_range(mm, pmd, from, end - from, phys_addr + from, prot);
 926                 if (error)
 927                         break;
 928                 from = (from + PGDIR_SIZE) & PGDIR_MASK;
 929                 dir++;
 930         } while (from && (from < end));
 931         flush_tlb_range(vma, beg, end);
 932         spin_unlock(&mm->page_table_lock);
 933         return error;
 934 }
 935
 936 /*
 937  * Establish a new mapping:
 938  *  - flush the old one
 939  *  - update the page tables
 940  *  - inform the TLB about the new one
 941  *
 942  * We hold the mm semaphore for reading and vma->vm_mm->page_table_lock
 943  */
 944 static inline void establish_pte(struct vm_area_struct * vma, unsigned long address, pte_t *page_table, pte_t entry)
 945 {
 946         set_pte(page_table, entry);
 947         flush_tlb_page(vma, address);
 948         update_mmu_cache(vma, address, entry);
 949 }
 950
 951 /*
 952  * We hold the mm semaphore for reading and vma->vm_mm->page_table_lock
 953  */
 954 static inline void break_cow(struct vm_area_struct * vma, struct page * new_page, unsigned long address,
 955                 pte_t *page_table)
 956 {
 957         invalidate_vcache(address, vma->vm_mm, new_page);
 958         flush_cache_page(vma, address);
 959         establish_pte(vma, address, page_table, pte_mkwrite(pte_mkdirty(mk_pte(new_page, vma->vm_page_prot))));
 960 }
 961
 962 /*
 963  * This routine handles present pages, when users try to write
 964  * to a shared page. It is done by copying the page to a new address
 965  * and decrementing the shared-page counter for the old page.
 966  *
 967  * Goto-purists beware: the only reason for goto's here is that it results
 968  * in better assembly code.. The "default" path will see no jumps at all.
 969  *
 970  * Note that this routine assumes that the protection checks have been
 971  * done by the caller (the low-level page fault routine in most cases).
 972  * Thus we can safely just mark it writable once we've done any necessary
 973  * COW.
 974  *
 975  * We also mark the page dirty at this point even though the page will
 976  * change only once the write actually happens. This avoids a few races,
 977  * and potentially makes it more efficient.
 978  *
 979  * We hold the mm semaphore and the page_table_lock on entry and exit
 980  * with the page_table_lock released.
 981  */
 982 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct * vma,
 983         unsigned long address, pte_t *page_table, pmd_t *pmd, pte_t pte)
 984 {
 985         struct page *old_page, *new_page;
 986         unsigned long pfn = pte_pfn(pte);
 987         struct pte_chain *pte_chain = NULL;
 988         int ret;
 989
 990         if (unlikely(!pfn_valid(pfn))) {
 991                 /*
 992                  * This should really halt the system so it can be debugged or
 993                  * at least the kernel stops what it's doing before it corrupts
 994                  * data, but for the moment just pretend this is OOM.
 995                  */
 996                 pte_unmap(page_table);
 997                 printk(KERN_ERR "do_wp_page: bogus page at address %08lx\n",
 998                                 address);
 999                 goto oom;
1000         }
1001         old_page = pfn_to_page(pfn);
1002
1003         if (!TestSetPageLocked(old_page)) {
1004                 int reuse = can_share_swap_page(old_page);
1005                 unlock_page(old_page);
1006                 if (reuse) {
1007                         flush_cache_page(vma, address);
1008                         establish_pte(vma, address, page_table,
1009                                 pte_mkyoung(pte_mkdirty(pte_mkwrite(pte))));
1010                         pte_unmap(page_table);
1011                         ret = VM_FAULT_MINOR;
1012                         goto out;
1013                 }
1014         }
1015         pte_unmap(page_table);
1016
1017         /*
1018          * Ok, we need to copy. Oh, well..
1019          */
1020         page_cache_get(old_page);
1021         spin_unlock(&mm->page_table_lock);
1022
1023         pte_chain = pte_chain_alloc(GFP_KERNEL);
1024         if (!pte_chain)
1025                 goto no_mem;
1026         new_page = alloc_page(GFP_HIGHUSER);
1027         if (!new_page)
1028                 goto no_mem;
1029         copy_cow_page(old_page,new_page,address);
1030
1031         /*
1032          * Re-check the pte - we dropped the lock
1033          */
1034         spin_lock(&mm->page_table_lock);
1035         page_table = pte_offset_map(pmd, address);
1036         if (pte_same(*page_table, pte)) {
1037                 if (PageReserved(old_page))
1038                         ++mm->rss;
1039                 page_remove_rmap(old_page, page_table);
1040                 break_cow(vma, new_page, address, page_table);
1041                 pte_chain = page_add_rmap(new_page, page_table, pte_chain);
1042                 lru_cache_add_active(new_page);
1043
1044                 /* Free the old page.. */
1045                 new_page = old_page;
1046         }
1047         pte_unmap(page_table);
1048         page_cache_release(new_page);
1049         page_cache_release(old_page);
1050         ret = VM_FAULT_MINOR;
1051         goto out;
1052
1053 no_mem:
1054         page_cache_release(old_page);
1055 oom:
1056         ret = VM_FAULT_OOM;
1057 out:
1058         spin_unlock(&mm->page_table_lock);
1059         pte_chain_free(pte_chain);
1060         return ret;
1061 }
1062
1063 static void vmtruncate_list(struct list_head *head, unsigned long pgoff)
1064 {
1065         unsigned long start, end, len, diff;
1066         struct vm_area_struct *vma;
1067         struct list_head *curr;
1068
1069         list_for_each(curr, head) {
1070                 vma = list_entry(curr, struct vm_area_struct, shared);
1071                 start = vma->vm_start;
1072                 end = vma->vm_end;
1073                 len = end - start;
1074
1075                 /* mapping wholly truncated? */
1076                 if (vma->vm_pgoff >= pgoff) {
1077                         zap_page_range(vma, start, len);
1078                         continue;
1079                 }
1080
1081                 /* mapping wholly unaffected? */
1082                 len = len >> PAGE_SHIFT;
1083                 diff = pgoff - vma->vm_pgoff;
1084                 if (diff >= len)
1085                         continue;
1086
1087                 /* Ok, partially affected.. */
1088                 start += diff << PAGE_SHIFT;
1089                 len = (len - diff) << PAGE_SHIFT;
1090                 zap_page_range(vma, start, len);
1091         }
1092 }
1093
1094 /*
1095  * Handle all mappings that got truncated by a "truncate()"
1096  * system call.
1097  *
1098  * NOTE! We have to be ready to update the memory sharing
1099  * between the file and the memory map for a potential last
1100  * incomplete page.  Ugly, but necessary.
1101  */
1102 int vmtruncate(struct inode * inode, loff_t offset)
1103 {
1104         unsigned long pgoff;
1105         struct address_space *mapping = inode->i_mapping;
1106         unsigned long limit;
1107
1108         if (inode->i_size < offset)
1109                 goto do_expand;
1110         inode->i_size = offset;
1111         down(&mapping->i_shared_sem);
1112         if (list_empty(&mapping->i_mmap) && list_empty(&mapping->i_mmap_shared))
1113                 goto out_unlock;
1114
1115         pgoff = (offset + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1116         if (!list_empty(&mapping->i_mmap))
1117                 vmtruncate_list(&mapping->i_mmap, pgoff);
1118         if (!list_empty(&mapping->i_mmap_shared))
1119                 vmtruncate_list(&mapping->i_mmap_shared, pgoff);
1120
1121 out_unlock:
1122         up(&mapping->i_shared_sem);
1123         truncate_inode_pages(mapping, offset);
1124         goto out_truncate;
1125
1126 do_expand:
1127         limit = current->rlim[RLIMIT_FSIZE].rlim_cur;
1128         if (limit != RLIM_INFINITY && offset > limit)
1129                 goto out_sig;
1130         if (offset > inode->i_sb->s_maxbytes)
1131                 goto out;
1132         inode->i_size = offset;
1133
1134 out_truncate:
1135         if (inode->i_op && inode->i_op->truncate)
1136                 inode->i_op->truncate(inode);
1137         return 0;
1138 out_sig:
1139         send_sig(SIGXFSZ, current, 0);
1140 out:
1141         return -EFBIG;
1142 }
1143
1144 /*
1145  * Primitive swap readahead code. We simply read an aligned block of
1146  * (1 << page_cluster) entries in the swap area. This method is chosen
1147  * because it doesn't cost us any seek time.  We also make sure to queue
1148  * the 'original' request together with the readahead ones...
1149  */
1150 void swapin_readahead(swp_entry_t entry)
1151 {
1152         int i, num;
1153         struct page *new_page;
1154         unsigned long offset;
1155
1156         /*
1157          * Get the number of handles we should do readahead io to.
1158          */
1159         num = valid_swaphandles(entry, &offset);
1160         for (i = 0; i < num; offset++, i++) {
1161                 /* Ok, do the async read-ahead now */
1162                 new_page = read_swap_cache_async(swp_entry(swp_type(entry),
1163                                                 offset));
1164                 if (!new_page)
1165                         break;
1166                 page_cache_release(new_page);
1167         }
1168         lru_add_drain();        /* Push any new pages onto the LRU now */
1169 }
1170
1171 /*
1172  * We hold the mm semaphore and the page_table_lock on entry and
1173  * should release the pagetable lock on exit..
1174  */
1175 static int do_swap_page(struct mm_struct * mm,
1176         struct vm_area_struct * vma, unsigned long address,
1177         pte_t *page_table, pmd_t *pmd, pte_t orig_pte, int write_access)
1178 {
1179         struct page *page;
1180         swp_entry_t entry = pte_to_swp_entry(orig_pte);
1181         pte_t pte;
1182         int ret = VM_FAULT_MINOR;
1183         struct pte_chain *pte_chain = NULL;
1184
1185         pte_unmap(page_table);
1186         spin_unlock(&mm->page_table_lock);
1187         page = lookup_swap_cache(entry);
1188         if (!page) {
1189                 swapin_readahead(entry);
1190                 page = read_swap_cache_async(entry);
1191                 if (!page) {
1192                         /*
1193                          * Back out if somebody else faulted in this pte while
1194                          * we released the page table lock.
1195                          */
1196                         spin_lock(&mm->page_table_lock);
1197                         page_table = pte_offset_map(pmd, address);
1198                         if (pte_same(*page_table, orig_pte))
1199                                 ret = VM_FAULT_OOM;
1200                         else
1201                                 ret = VM_FAULT_MINOR;
1202                         pte_unmap(page_table);
1203                         spin_unlock(&mm->page_table_lock);
1204                         goto out;
1205                 }
1206
1207                 /* Had to read the page from swap area: Major fault */
1208                 ret = VM_FAULT_MAJOR;
1209                 inc_page_state(pgmajfault);
1210         }
1211
1212         mark_page_accessed(page);
1213         pte_chain = pte_chain_alloc(GFP_KERNEL);
1214         if (!pte_chain) {
1215                 ret = -ENOMEM;
1216                 goto out;
1217         }
1218         lock_page(page);
1219
1220         /*
1221          * Back out if somebody else faulted in this pte while we
1222          * released the page table lock.
1223          */
1224         spin_lock(&mm->page_table_lock);
1225         page_table = pte_offset_map(pmd, address);
1226         if (!pte_same(*page_table, orig_pte)) {
1227                 pte_unmap(page_table);
1228                 spin_unlock(&mm->page_table_lock);
1229                 unlock_page(page);
1230                 page_cache_release(page);
1231                 ret = VM_FAULT_MINOR;
1232                 goto out;
1233         }
1234
1235         /* The page isn't present yet, go ahead with the fault. */
1236
1237         swap_free(entry);
1238         if (vm_swap_full())
1239                 remove_exclusive_swap_page(page);
1240
1241         mm->rss++;
1242         pte = mk_pte(page, vma->vm_page_prot);
1243         if (write_access && can_share_swap_page(page))
1244                 pte = pte_mkdirty(pte_mkwrite(pte));
1245         unlock_page(page);
1246
1247         flush_icache_page(vma, page);
1248         set_pte(page_table, pte);
1249         pte_chain = page_add_rmap(page, page_table, pte_chain);
1250
1251         /* No need to invalidate - it was non-present before */
1252         update_mmu_cache(vma, address, pte);
1253         pte_unmap(page_table);
1254         spin_unlock(&mm->page_table_lock);
1255 out:
1256         pte_chain_free(pte_chain);
1257         return ret;
1258 }
1259
1260 /*
1261  * We are called with the MM semaphore and page_table_lock
1262  * spinlock held to protect against concurrent faults in
1263  * multithreaded programs.
1264  */
1265 static int
1266 do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
1267                 pte_t *page_table, pmd_t *pmd, int write_access,
1268                 unsigned long addr)
1269 {
1270         pte_t entry;
1271         struct page * page = ZERO_PAGE(addr);
1272         struct pte_chain *pte_chain;
1273         int ret;
1274
1275         pte_chain = pte_chain_alloc(GFP_ATOMIC);
1276         if (!pte_chain) {
1277                 pte_unmap(page_table);
1278                 spin_unlock(&mm->page_table_lock);
1279                 pte_chain = pte_chain_alloc(GFP_KERNEL);
1280                 if (!pte_chain)
1281                         goto no_mem;
1282                 spin_lock(&mm->page_table_lock);
1283                 page_table = pte_offset_map(pmd, addr);
1284         }
1285
1286         /* Read-only mapping of ZERO_PAGE. */
1287         entry = pte_wrprotect(mk_pte(ZERO_PAGE(addr), vma->vm_page_prot));
1288
1289         /* ..except if it's a write access */
1290         if (write_access) {
1291                 /* Allocate our own private page. */
1292                 pte_unmap(page_table);
1293                 spin_unlock(&mm->page_table_lock);
1294
1295                 page = alloc_page(GFP_HIGHUSER);
1296                 if (!page)
1297                         goto no_mem;
1298                 clear_user_highpage(page, addr);
1299
1300                 spin_lock(&mm->page_table_lock);
1301                 page_table = pte_offset_map(pmd, addr);
1302
1303                 if (!pte_none(*page_table)) {
1304                         pte_unmap(page_table);
1305                         page_cache_release(page);
1306                         spin_unlock(&mm->page_table_lock);
1307                         ret = VM_FAULT_MINOR;
1308                         goto out;
1309                 }
1310                 mm->rss++;
1311                 entry = pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
1312                 lru_cache_add_active(page);
1313                 mark_page_accessed(page);
1314         }
1315
1316         set_pte(page_table, entry);
1317         /* ignores ZERO_PAGE */
1318         pte_chain = page_add_rmap(page, page_table, pte_chain);
1319         pte_unmap(page_table);
1320
1321         /* No need to invalidate - it was non-present before */
1322         update_mmu_cache(vma, addr, entry);
1323         spin_unlock(&mm->page_table_lock);
1324         ret = VM_FAULT_MINOR;
1325         goto out;
1326
1327 no_mem:
1328         ret = VM_FAULT_OOM;
1329 out:
1330         pte_chain_free(pte_chain);
1331         return ret;
1332 }
1333
1334 /*
1335  * do_no_page() tries to create a new page mapping. It aggressively
1336  * tries to share with existing pages, but makes a separate copy if
1337  * the "write_access" parameter is true in order to avoid the next
1338  * page fault.
1339  *
1340  * As this is called only for pages that do not currently exist, we
1341  * do not need to flush old virtual caches or the TLB.
1342  *
1343  * This is called with the MM semaphore held and the page table
1344  * spinlock held. Exit with the spinlock released.
1345  */
1346 static int
1347 do_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
1348         unsigned long address, int write_access, pte_t *page_table, pmd_t *pmd)
1349 {
1350         struct page * new_page;
1351         pte_t entry;
1352         struct pte_chain *pte_chain;
1353         int ret;
1354
1355         if (!vma->vm_ops || !vma->vm_ops->nopage)
1356                 return do_anonymous_page(mm, vma, page_table,
1357                                         pmd, write_access, address);
1358         pte_unmap(page_table);
1359         spin_unlock(&mm->page_table_lock);
1360
1361         new_page = vma->vm_ops->nopage(vma, address & PAGE_MASK, 0);
1362
1363         /* no page was available -- either SIGBUS or OOM */
1364         if (new_page == NOPAGE_SIGBUS)
1365                 return VM_FAULT_SIGBUS;
1366         if (new_page == NOPAGE_OOM)
1367                 return VM_FAULT_OOM;
1368
1369         pte_chain = pte_chain_alloc(GFP_KERNEL);
1370         if (!pte_chain)
1371                 goto oom;
1372
1373         /*
1374          * Should we do an early C-O-W break?
1375          */
1376         if (write_access && !(vma->vm_flags & VM_SHARED)) {
1377                 struct page * page = alloc_page(GFP_HIGHUSER);
1378                 if (!page) {
1379                         page_cache_release(new_page);
1380                         goto oom;
1381                 }
1382                 copy_user_highpage(page, new_page, address);
1383                 page_cache_release(new_page);
1384                 lru_cache_add_active(page);
1385                 new_page = page;
1386         }
1387
1388         spin_lock(&mm->page_table_lock);
1389         page_table = pte_offset_map(pmd, address);
1390
1391         /*
1392          * This silly early PAGE_DIRTY setting removes a race
1393          * due to the bad i386 page protection. But it's valid
1394          * for other architectures too.
1395          *
1396          * Note that if write_access is true, we either now have
1397          * an exclusive copy of the page, or this is a shared mapping,
1398          * so we can make it writable and dirty to avoid having to
1399          * handle that later.
1400          */
1401         /* Only go through if we didn't race with anybody else... */
1402         if (pte_none(*page_table)) {
1403                 ++mm->rss;
1404                 flush_icache_page(vma, new_page);
1405                 entry = mk_pte(new_page, vma->vm_page_prot);
1406                 if (write_access)
1407                         entry = pte_mkwrite(pte_mkdirty(entry));
1408                 set_pte(page_table, entry);
1409                 pte_chain = page_add_rmap(new_page, page_table, pte_chain);
1410                 pte_unmap(page_table);
1411         } else {
1412                 /* One of our sibling threads was faster, back out. */
1413                 pte_unmap(page_table);
1414                 page_cache_release(new_page);
1415                 spin_unlock(&mm->page_table_lock);
1416                 ret = VM_FAULT_MINOR;
1417                 goto out;
1418         }
1419
1420         /* no need to invalidate: a not-present page shouldn't be cached */
1421         update_mmu_cache(vma, address, entry);
1422         spin_unlock(&mm->page_table_lock);
1423         ret = VM_FAULT_MAJOR;
1424         goto out;
1425 oom:
1426         ret = VM_FAULT_OOM;
1427 out:
1428         pte_chain_free(pte_chain);
1429         return ret;
1430 }
1431
1432 /*
1433  * Fault of a previously existing named mapping. Repopulate the pte
1434  * from the encoded file_pte if possible. This enables swappable
1435  * nonlinear vmas.
1436  */
1437 static int do_file_page(struct mm_struct * mm, struct vm_area_struct * vma,
1438         unsigned long address, int write_access, pte_t *pte, pmd_t *pmd)
1439 {
1440         unsigned long pgoff;
1441         int err;
1442
1443         BUG_ON(!vma->vm_ops || !vma->vm_ops->nopage);
1444         /*
1445          * Fall back to the linear mapping if the fs does not support
1446          * ->populate:
1447          */
1448         if (!vma->vm_ops || !vma->vm_ops->populate ||
1449                         (write_access && !(vma->vm_flags & VM_SHARED))) {
1450                 pte_clear(pte);
1451                 return do_no_page(mm, vma, address, write_access, pte, pmd);
1452         }
1453
1454         pgoff = pte_to_pgoff(*pte);
1455
1456         pte_unmap(pte);
1457         spin_unlock(&mm->page_table_lock);
1458
1459         err = vma->vm_ops->populate(vma, address & PAGE_MASK, PAGE_SIZE, vma->vm_page_prot, pgoff, 0);
1460         if (err == -ENOMEM)
1461                 return VM_FAULT_OOM;
1462         if (err)
1463                 return VM_FAULT_SIGBUS;
1464         return VM_FAULT_MAJOR;
1465 }
1466
1467 /*
1468  * These routines also need to handle stuff like marking pages dirty
1469  * and/or accessed for architectures that don't do it in hardware (most
1470  * RISC architectures).  The early dirtying is also good on the i386.
1471  *
1472  * There is also a hook called "update_mmu_cache()" that architectures
1473  * with external mmu caches can use to update those (ie the Sparc or
1474  * PowerPC hashed page tables that act as extended TLBs).
1475  *
1476  * Note the "page_table_lock". It is to protect against kswapd removing
1477  * pages from under us. Note that kswapd only ever _removes_ pages, never
1478  * adds them. As such, once we have noticed that the page is not present,
1479  * we can drop the lock early.
1480  *
1481  * The adding of pages is protected by the MM semaphore (which we hold),
1482  * so we don't need to worry about a page being suddenly been added into
1483  * our VM.
1484  *
1485  * We enter with the pagetable spinlock held, we are supposed to
1486  * release it when done.
1487  */
1488 static inline int handle_pte_fault(struct mm_struct *mm,
1489         struct vm_area_struct * vma, unsigned long address,
1490         int write_access, pte_t *pte, pmd_t *pmd)
1491 {
1492         pte_t entry;
1493
1494         entry = *pte;
1495         if (!pte_present(entry)) {
1496                 /*
1497                  * If it truly wasn't present, we know that kswapd
1498                  * and the PTE updates will not touch it later. So
1499                  * drop the lock.
1500                  */
1501                 if (pte_none(entry))
1502                         return do_no_page(mm, vma, address, write_access, pte, pmd);
1503                 if (pte_file(entry))
1504                         return do_file_page(mm, vma, address, write_access, pte, pmd);
1505                 return do_swap_page(mm, vma, address, pte, pmd, entry, write_access);
1506         }
1507
1508         if (write_access) {
1509                 if (!pte_write(entry))
1510                         return do_wp_page(mm, vma, address, pte, pmd, entry);
1511
1512                 entry = pte_mkdirty(entry);
1513         }
1514         entry = pte_mkyoung(entry);
1515         establish_pte(vma, address, pte, entry);
1516         pte_unmap(pte);
1517         spin_unlock(&mm->page_table_lock);
1518         return VM_FAULT_MINOR;
1519 }
1520
1521 /*
1522  * By the time we get here, we already hold the mm semaphore
1523  */
1524 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct * vma,
1525         unsigned long address, int write_access)
1526 {
1527         pgd_t *pgd;
1528         pmd_t *pmd;
1529
1530         __set_current_state(TASK_RUNNING);
1531         pgd = pgd_offset(mm, address);
1532
1533         inc_page_state(pgfault);
1534
1535         if (is_vm_hugetlb_page(vma))
1536                 return VM_FAULT_SIGBUS; /* mapping truncation does this. */
1537
1538         /*
1539          * We need the page table lock to synchronize with kswapd
1540          * and the SMP-safe atomic PTE updates.
1541          */
1542         spin_lock(&mm->page_table_lock);
1543         pmd = pmd_alloc(mm, pgd, address);
1544
1545         if (pmd) {
1546                 pte_t * pte = pte_alloc_map(mm, pmd, address);
1547                 if (pte)
1548                         return handle_pte_fault(mm, vma, address, write_access, pte, pmd);
1549         }
1550         spin_unlock(&mm->page_table_lock);
1551         return VM_FAULT_OOM;
1552 }
1553
1554 /*
1555  * Allocate page middle directory.
1556  *
1557  * We've already handled the fast-path in-line, and we own the
1558  * page table lock.
1559  *
1560  * On a two-level page table, this ends up actually being entirely
1561  * optimized away.
1562  */
1563 pmd_t *__pmd_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
1564 {
1565         pmd_t *new;
1566
1567         spin_unlock(&mm->page_table_lock);
1568         new = pmd_alloc_one(mm, address);
1569         spin_lock(&mm->page_table_lock);
1570         if (!new)
1571                 return NULL;
1572
1573         /*
1574          * Because we dropped the lock, we should re-check the
1575          * entry, as somebody else could have populated it..
1576          */
1577         if (pgd_present(*pgd)) {
1578                 pmd_free(new);
1579                 goto out;
1580         }
1581         pgd_populate(mm, pgd, new);
1582 out:
1583         return pmd_offset(pgd, address);
1584 }
1585
1586 int make_pages_present(unsigned long addr, unsigned long end)
1587 {
1588         int ret, len, write;
1589         struct vm_area_struct * vma;
1590
1591         vma = find_vma(current->mm, addr);
1592         write = (vma->vm_flags & VM_WRITE) != 0;
1593         if (addr >= end)
1594                 BUG();
1595         if (end > vma->vm_end)
1596                 BUG();
1597         len = (end+PAGE_SIZE-1)/PAGE_SIZE-addr/PAGE_SIZE;
1598         ret = get_user_pages(current, current->mm, addr,
1599                         len, write, 0, NULL, NULL);
1600         return ret == len ? 0 : -1;
1601 }
1602
1603 /*
1604  * Map a vmalloc()-space virtual address to the physical page.
1605  */
1606 struct page * vmalloc_to_page(void * vmalloc_addr)
1607 {
1608         unsigned long addr = (unsigned long) vmalloc_addr;
1609         struct page *page = NULL;
1610         pgd_t *pgd = pgd_offset_k(addr);
1611         pmd_t *pmd;
1612         pte_t *ptep, pte;
1613
1614         if (!pgd_none(*pgd)) {
1615                 pmd = pmd_offset(pgd, addr);
1616                 if (!pmd_none(*pmd)) {
1617                         preempt_disable();
1618                         ptep = pte_offset_map(pmd, addr);
1619                         pte = *ptep;
1620                         if (pte_present(pte))
1621                                 page = pte_page(pte);
1622                         pte_unmap(ptep);
1623                         preempt_enable();
1624                 }
1625         }
1626         return page;
1627 }