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/rmap.h>
  47 #include <linux/module.h>
  48 #include <linux/init.h>
  49
  50 #include <asm/pgalloc.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 #include <linux/elf.h>
  58
  59 #ifndef CONFIG_DISCONTIGMEM
  60 /* use the per-pgdat data instead for discontigmem - mbligh */
  61 unsigned long max_mapnr;
  62 struct page *mem_map;
  63
  64 EXPORT_SYMBOL(max_mapnr);
  65 EXPORT_SYMBOL(mem_map);
  66 #endif
  67
  68 unsigned long num_physpages;
  69 void * high_memory;
  70 struct page *highmem_start_page;
  71
  72 EXPORT_SYMBOL(num_physpages);
  73 EXPORT_SYMBOL(highmem_start_page);
  74 EXPORT_SYMBOL(high_memory);
  75
  76 /*
  77  * We special-case the C-O-W ZERO_PAGE, because it's such
  78  * a common occurrence (no need to read the page to know
  79  * that it's zero - better for the cache and memory subsystem).
  80  */
  81 static inline void copy_cow_page(struct page * from, struct page * to, unsigned long address)
  82 {
  83         if (from == ZERO_PAGE(address)) {
  84                 clear_user_highpage(to, address);
  85                 return;
  86         }
  87         copy_user_highpage(to, from, address);
  88 }
  89
  90 /*
  91  * Note: this doesn't free the actual pages themselves. That
  92  * has been handled earlier when unmapping all the memory regions.
  93  */
  94 static inline void free_one_pmd(struct mmu_gather *tlb, pmd_t * dir)
  95 {
  96         struct page *page;
  97
  98         if (pmd_none(*dir))
  99                 return;
 100         if (unlikely(pmd_bad(*dir))) {
 101                 pmd_ERROR(*dir);
 102                 pmd_clear(dir);
 103                 return;
 104         }
 105         page = pmd_page(*dir);
 106         pmd_clear(dir);
 107         dec_page_state(nr_page_table_pages);
 108         pte_free_tlb(tlb, page);
 109 }
 110
 111 static inline void free_one_pgd(struct mmu_gather *tlb, pgd_t * dir,
 112                                                         int pgd_idx)
 113 {
 114         int j;
 115         pmd_t * pmd;
 116
 117         if (pgd_none(*dir))
 118                 return;
 119         if (unlikely(pgd_bad(*dir))) {
 120                 pgd_ERROR(*dir);
 121                 pgd_clear(dir);
 122                 return;
 123         }
 124         pmd = pmd_offset(dir, 0);
 125         pgd_clear(dir);
 126         for (j = 0; j < PTRS_PER_PMD ; j++) {
 127                 if (pgd_idx * PGDIR_SIZE + j * PMD_SIZE >= TASK_SIZE)
 128                         break;
 129                 free_one_pmd(tlb, pmd+j);
 130         }
 131         pmd_free_tlb(tlb, pmd);
 132 }
 133
 134 /*
 135  * This function clears all user-level page tables of a process - this
 136  * is needed by execve(), so that old pages aren't in the way.
 137  *
 138  * Must be called with pagetable lock held.
 139  */
 140 void clear_page_tables(struct mmu_gather *tlb, unsigned long first, int nr)
 141 {
 142         pgd_t * page_dir = tlb->mm->pgd;
 143         int pgd_idx = first;
 144
 145         page_dir += first;
 146         do {
 147                 free_one_pgd(tlb, page_dir, pgd_idx);
 148                 page_dir++;
 149                 pgd_idx++;
 150         } while (--nr);
 151 }
 152
 153 pte_t fastcall * pte_alloc_map(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
 154 {
 155         if (!pmd_present(*pmd)) {
 156                 struct page *new;
 157
 158                 spin_unlock(&mm->page_table_lock);
 159                 new = pte_alloc_one(mm, address);
 160                 spin_lock(&mm->page_table_lock);
 161                 if (!new)
 162                         return NULL;
 163
 164                 /*
 165                  * Because we dropped the lock, we should re-check the
 166                  * entry, as somebody else could have populated it..
 167                  */
 168                 if (pmd_present(*pmd)) {
 169                         pte_free(new);
 170                         goto out;
 171                 }
 172                 inc_page_state(nr_page_table_pages);
 173                 pmd_populate(mm, pmd, new);
 174         }
 175 out:
 176         return pte_offset_map(pmd, address);
 177 }
 178
 179 pte_t fastcall * pte_alloc_kernel(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
 180 {
 181         if (!pmd_present(*pmd)) {
 182                 pte_t *new;
 183
 184                 spin_unlock(&mm->page_table_lock);
 185                 new = pte_alloc_one_kernel(mm, address);
 186                 spin_lock(&mm->page_table_lock);
 187                 if (!new)
 188                         return NULL;
 189
 190                 /*
 191                  * Because we dropped the lock, we should re-check the
 192                  * entry, as somebody else could have populated it..
 193                  */
 194                 if (pmd_present(*pmd)) {
 195                         pte_free_kernel(new);
 196                         goto out;
 197                 }
 198                 pmd_populate_kernel(mm, pmd, new);
 199         }
 200 out:
 201         return pte_offset_kernel(pmd, address);
 202 }
 203 #define PTE_TABLE_MASK  ((PTRS_PER_PTE-1) * sizeof(pte_t))
 204 #define PMD_TABLE_MASK  ((PTRS_PER_PMD-1) * sizeof(pmd_t))
 205
 206 /*
 207  * copy one vm_area from one task to the other. Assumes the page tables
 208  * already present in the new task to be cleared in the whole range
 209  * covered by this vma.
 210  *
 211  * 08Jan98 Merged into one routine from several inline routines to reduce
 212  *         variable count and make things faster. -jj
 213  *
 214  * dst->page_table_lock is held on entry and exit,
 215  * but may be dropped within pmd_alloc() and pte_alloc_map().
 216  */
 217 int copy_page_range(struct mm_struct *dst, struct mm_struct *src,
 218                         struct vm_area_struct *vma)
 219 {
 220         pgd_t * src_pgd, * dst_pgd;
 221         unsigned long address = vma->vm_start;
 222         unsigned long end = vma->vm_end;
 223         unsigned long cow;
 224
 225         if (is_vm_hugetlb_page(vma))
 226                 return copy_hugetlb_page_range(dst, src, vma);
 227
 228         cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
 229         src_pgd = pgd_offset(src, address)-1;
 230         dst_pgd = pgd_offset(dst, address)-1;
 231
 232         for (;;) {
 233                 pmd_t * src_pmd, * dst_pmd;
 234
 235                 src_pgd++; dst_pgd++;
 236
 237                 /* copy_pmd_range */
 238
 239                 if (pgd_none(*src_pgd))
 240                         goto skip_copy_pmd_range;
 241                 if (unlikely(pgd_bad(*src_pgd))) {
 242                         pgd_ERROR(*src_pgd);
 243                         pgd_clear(src_pgd);
 244 skip_copy_pmd_range:    address = (address + PGDIR_SIZE) & PGDIR_MASK;
 245                         if (!address || (address >= end))
 246                                 goto out;
 247                         continue;
 248                 }
 249
 250                 src_pmd = pmd_offset(src_pgd, address);
 251                 dst_pmd = pmd_alloc(dst, dst_pgd, address);
 252                 if (!dst_pmd)
 253                         goto nomem;
 254
 255                 do {
 256                         pte_t * src_pte, * dst_pte;
 257
 258                         /* copy_pte_range */
 259
 260                         if (pmd_none(*src_pmd))
 261                                 goto skip_copy_pte_range;
 262                         if (unlikely(pmd_bad(*src_pmd))) {
 263                                 pmd_ERROR(*src_pmd);
 264                                 pmd_clear(src_pmd);
 265 skip_copy_pte_range:
 266                                 address = (address + PMD_SIZE) & PMD_MASK;
 267                                 if (address >= end)
 268                                         goto out;
 269                                 goto cont_copy_pmd_range;
 270                         }
 271
 272                         dst_pte = pte_alloc_map(dst, dst_pmd, address);
 273                         if (!dst_pte)
 274                                 goto nomem;
 275                         spin_lock(&src->page_table_lock);
 276                         src_pte = pte_offset_map_nested(src_pmd, address);
 277                         do {
 278                                 pte_t pte = *src_pte;
 279                                 struct page *page;
 280                                 unsigned long pfn;
 281
 282                                 /* copy_one_pte */
 283
 284                                 if (pte_none(pte))
 285                                         goto cont_copy_pte_range_noset;
 286                                 /* pte contains position in swap, so copy. */
 287                                 if (!pte_present(pte)) {
 288                                         if (!pte_file(pte))
 289                                                 swap_duplicate(pte_to_swp_entry(pte));
 290                                         set_pte(dst_pte, pte);
 291                                         goto cont_copy_pte_range_noset;
 292                                 }
 293                                 pfn = pte_pfn(pte);
 294                                 /* the pte points outside of valid memory, the
 295                                  * mapping is assumed to be good, meaningful
 296                                  * and not mapped via rmap - duplicate the
 297                                  * mapping as is.
 298                                  */
 299                                 page = NULL;
 300                                 if (pfn_valid(pfn))
 301                                         page = pfn_to_page(pfn);
 302
 303                                 if (!page || PageReserved(page)) {
 304                                         set_pte(dst_pte, pte);
 305                                         goto cont_copy_pte_range_noset;
 306                                 }
 307
 308                                 /*
 309                                  * If it's a COW mapping, write protect it both
 310                                  * in the parent and the child
 311                                  */
 312                                 if (cow) {
 313                                         ptep_set_wrprotect(src_pte);
 314                                         pte = *src_pte;
 315                                 }
 316
 317                                 /*
 318                                  * If it's a shared mapping, mark it clean in
 319                                  * the child
 320                                  */
 321                                 if (vma->vm_flags & VM_SHARED)
 322                                         pte = pte_mkclean(pte);
 323                                 pte = pte_mkold(pte);
 324                                 get_page(page);
 325                                 dst->rss++;
 326                                 set_pte(dst_pte, pte);
 327                                 page_dup_rmap(page);
 328 cont_copy_pte_range_noset:
 329                                 address += PAGE_SIZE;
 330                                 if (address >= end) {
 331                                         pte_unmap_nested(src_pte);
 332                                         pte_unmap(dst_pte);
 333                                         goto out_unlock;
 334                                 }
 335                                 src_pte++;
 336                                 dst_pte++;
 337                         } while ((unsigned long)src_pte & PTE_TABLE_MASK);
 338                         pte_unmap_nested(src_pte-1);
 339                         pte_unmap(dst_pte-1);
 340                         spin_unlock(&src->page_table_lock);
 341                         cond_resched_lock(&dst->page_table_lock);
 342 cont_copy_pmd_range:
 343                         src_pmd++;
 344                         dst_pmd++;
 345                 } while ((unsigned long)src_pmd & PMD_TABLE_MASK);
 346         }
 347 out_unlock:
 348         spin_unlock(&src->page_table_lock);
 349 out:
 350         return 0;
 351 nomem:
 352         return -ENOMEM;
 353 }
 354
 355 static void zap_pte_range(struct mmu_gather *tlb,
 356                 pmd_t *pmd, unsigned long address,
 357                 unsigned long size, struct zap_details *details)
 358 {
 359         unsigned long offset;
 360         pte_t *ptep;
 361
 362         if (pmd_none(*pmd))
 363                 return;
 364         if (unlikely(pmd_bad(*pmd))) {
 365                 pmd_ERROR(*pmd);
 366                 pmd_clear(pmd);
 367                 return;
 368         }
 369         ptep = pte_offset_map(pmd, address);
 370         offset = address & ~PMD_MASK;
 371         if (offset + size > PMD_SIZE)
 372                 size = PMD_SIZE - offset;
 373         size &= PAGE_MASK;
 374         if (details && !details->check_mapping && !details->nonlinear_vma)
 375                 details = NULL;
 376         for (offset=0; offset < size; ptep++, offset += PAGE_SIZE) {
 377                 pte_t pte = *ptep;
 378                 if (pte_none(pte))
 379                         continue;
 380                 if (pte_present(pte)) {
 381                         struct page *page = NULL;
 382                         unsigned long pfn = pte_pfn(pte);
 383                         if (pfn_valid(pfn)) {
 384                                 page = pfn_to_page(pfn);
 385                                 if (PageReserved(page))
 386                                         page = NULL;
 387                         }
 388                         if (unlikely(details) && page) {
 389                                 /*
 390                                  * unmap_shared_mapping_pages() wants to
 391                                  * invalidate cache without truncating:
 392                                  * unmap shared but keep private pages.
 393                                  */
 394                                 if (details->check_mapping &&
 395                                     details->check_mapping != page->mapping)
 396                                         continue;
 397                                 /*
 398                                  * Each page->index must be checked when
 399                                  * invalidating or truncating nonlinear.
 400                                  */
 401                                 if (details->nonlinear_vma &&
 402                                     (page->index < details->first_index ||
 403                                      page->index > details->last_index))
 404                                         continue;
 405                         }
 406                         pte = ptep_get_and_clear(ptep);
 407                         tlb_remove_tlb_entry(tlb, ptep, address+offset);
 408                         if (unlikely(!page))
 409                                 continue;
 410                         if (unlikely(details) && details->nonlinear_vma
 411                             && linear_page_index(details->nonlinear_vma,
 412                                         address+offset) != page->index)
 413                                 set_pte(ptep, pgoff_to_pte(page->index));
 414                         if (pte_dirty(pte))
 415                                 set_page_dirty(page);
 416                         if (pte_young(pte) && page_mapping(page))
 417                                 mark_page_accessed(page);
 418                         tlb->freed++;
 419                         page_remove_rmap(page);
 420                         tlb_remove_page(tlb, page);
 421                         continue;
 422                 }
 423                 /*
 424                  * If details->check_mapping, we leave swap entries;
 425                  * if details->nonlinear_vma, we leave file entries.
 426                  */
 427                 if (unlikely(details))
 428                         continue;
 429                 if (!pte_file(pte))
 430                         free_swap_and_cache(pte_to_swp_entry(pte));
 431                 pte_clear(ptep);
 432         }
 433         pte_unmap(ptep-1);
 434 }
 435
 436 static void zap_pmd_range(struct mmu_gather *tlb,
 437                 pgd_t * dir, unsigned long address,
 438                 unsigned long size, struct zap_details *details)
 439 {
 440         pmd_t * pmd;
 441         unsigned long end, pgd_boundary;
 442
 443         if (pgd_none(*dir))
 444                 return;
 445         if (unlikely(pgd_bad(*dir))) {
 446                 pgd_ERROR(*dir);
 447                 pgd_clear(dir);
 448                 return;
 449         }
 450         pmd = pmd_offset(dir, address);
 451         end = address + size;
 452         pgd_boundary = ((address + PGDIR_SIZE) & PGDIR_MASK);
 453         if (pgd_boundary && (end > pgd_boundary))
 454                 end = pgd_boundary;
 455         do {
 456                 zap_pte_range(tlb, pmd, address, end - address, details);
 457                 address = (address + PMD_SIZE) & PMD_MASK;
 458                 pmd++;
 459         } while (address && (address < end));
 460 }
 461
 462 static void unmap_page_range(struct mmu_gather *tlb,
 463                 struct vm_area_struct *vma, unsigned long address,
 464                 unsigned long end, struct zap_details *details)
 465 {
 466         pgd_t * dir;
 467
 468         BUG_ON(address >= end);
 469         dir = pgd_offset(vma->vm_mm, address);
 470         tlb_start_vma(tlb, vma);
 471         do {
 472                 zap_pmd_range(tlb, dir, address, end - address, details);
 473                 address = (address + PGDIR_SIZE) & PGDIR_MASK;
 474                 dir++;
 475         } while (address && (address < end));
 476         tlb_end_vma(tlb, vma);
 477 }
 478
 479 /* Dispose of an entire struct mmu_gather per rescheduling point */
 480 #if defined(CONFIG_SMP) && defined(CONFIG_PREEMPT)
 481 #define ZAP_BLOCK_SIZE  (FREE_PTE_NR * PAGE_SIZE)
 482 #endif
 483
 484 /* For UP, 256 pages at a time gives nice low latency */
 485 #if !defined(CONFIG_SMP) && defined(CONFIG_PREEMPT)
 486 #define ZAP_BLOCK_SIZE  (256 * PAGE_SIZE)
 487 #endif
 488
 489 /* No preempt: go for improved straight-line efficiency */
 490 #if !defined(CONFIG_PREEMPT)
 491 #define ZAP_BLOCK_SIZE  (1024 * PAGE_SIZE)
 492 #endif
 493
 494 /**
 495  * unmap_vmas - unmap a range of memory covered by a list of vma's
 496  * @tlbp: address of the caller's struct mmu_gather
 497  * @mm: the controlling mm_struct
 498  * @vma: the starting vma
 499  * @start_addr: virtual address at which to start unmapping
 500  * @end_addr: virtual address at which to end unmapping
 501  * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
 502  * @details: details of nonlinear truncation or shared cache invalidation
 503  *
 504  * Returns the number of vma's which were covered by the unmapping.
 505  *
 506  * Unmap all pages in the vma list.  Called under page_table_lock.
 507  *
 508  * We aim to not hold page_table_lock for too long (for scheduling latency
 509  * reasons).  So zap pages in ZAP_BLOCK_SIZE bytecounts.  This means we need to
 510  * return the ending mmu_gather to the caller.
 511  *
 512  * Only addresses between `start' and `end' will be unmapped.
 513  *
 514  * The VMA list must be sorted in ascending virtual address order.
 515  *
 516  * unmap_vmas() assumes that the caller will flush the whole unmapped address
 517  * range after unmap_vmas() returns.  So the only responsibility here is to
 518  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
 519  * drops the lock and schedules.
 520  */
 521 int unmap_vmas(struct mmu_gather **tlbp, struct mm_struct *mm,
 522                 struct vm_area_struct *vma, unsigned long start_addr,
 523                 unsigned long end_addr, unsigned long *nr_accounted,
 524                 struct zap_details *details)
 525 {
 526         unsigned long zap_bytes = ZAP_BLOCK_SIZE;
 527         unsigned long tlb_start = 0;    /* For tlb_finish_mmu */
 528         int tlb_start_valid = 0;
 529         int ret = 0;
 530         int atomic = details && details->atomic;
 531
 532         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
 533                 unsigned long start;
 534                 unsigned long end;
 535
 536                 start = max(vma->vm_start, start_addr);
 537                 if (start >= vma->vm_end)
 538                         continue;
 539                 end = min(vma->vm_end, end_addr);
 540                 if (end <= vma->vm_start)
 541                         continue;
 542
 543                 if (vma->vm_flags & VM_ACCOUNT)
 544                         *nr_accounted += (end - start) >> PAGE_SHIFT;
 545
 546                 ret++;
 547                 while (start != end) {
 548                         unsigned long block;
 549
 550                         if (!tlb_start_valid) {
 551                                 tlb_start = start;
 552                                 tlb_start_valid = 1;
 553                         }
 554
 555                         if (is_vm_hugetlb_page(vma)) {
 556                                 block = end - start;
 557                                 unmap_hugepage_range(vma, start, end);
 558                         } else {
 559                                 block = min(zap_bytes, end - start);
 560                                 unmap_page_range(*tlbp, vma, start,
 561                                                 start + block, details);
 562                         }
 563
 564                         start += block;
 565                         zap_bytes -= block;
 566                         if ((long)zap_bytes > 0)
 567                                 continue;
 568                         if (!atomic && need_resched()) {
 569                                 int fullmm = tlb_is_full_mm(*tlbp);
 570                                 tlb_finish_mmu(*tlbp, tlb_start, start);
 571                                 cond_resched_lock(&mm->page_table_lock);
 572                                 *tlbp = tlb_gather_mmu(mm, fullmm);
 573                                 tlb_start_valid = 0;
 574                         }
 575                         zap_bytes = ZAP_BLOCK_SIZE;
 576                 }
 577         }
 578         return ret;
 579 }
 580
 581 /**
 582  * zap_page_range - remove user pages in a given range
 583  * @vma: vm_area_struct holding the applicable pages
 584  * @address: starting address of pages to zap
 585  * @size: number of bytes to zap
 586  * @details: details of nonlinear truncation or shared cache invalidation
 587  */
 588 void zap_page_range(struct vm_area_struct *vma, unsigned long address,
 589                 unsigned long size, struct zap_details *details)
 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         if (is_vm_hugetlb_page(vma)) {
 597                 zap_hugepage_range(vma, address, size);
 598                 return;
 599         }
 600
 601         lru_add_drain();
 602         spin_lock(&mm->page_table_lock);
 603         tlb = tlb_gather_mmu(mm, 0);
 604         unmap_vmas(&tlb, mm, vma, address, end, &nr_accounted, details);
 605         tlb_finish_mmu(tlb, address, end);
 606         spin_unlock(&mm->page_table_lock);
 607 }
 608
 609 /*
 610  * Do a quick page-table lookup for a single page.
 611  * mm->page_table_lock must be held.
 612  */
 613 struct page *
 614 follow_page(struct mm_struct *mm, unsigned long address, int write)
 615 {
 616         pgd_t *pgd;
 617         pmd_t *pmd;
 618         pte_t *ptep, pte;
 619         unsigned long pfn;
 620         struct page *page;
 621
 622         page = follow_huge_addr(mm, address, write);
 623         if (! IS_ERR(page))
 624                 return page;
 625
 626         pgd = pgd_offset(mm, address);
 627         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
 628                 goto out;
 629
 630         pmd = pmd_offset(pgd, address);
 631         if (pmd_none(*pmd))
 632                 goto out;
 633         if (pmd_huge(*pmd))
 634                 return follow_huge_pmd(mm, address, pmd, write);
 635         if (unlikely(pmd_bad(*pmd)))
 636                 goto out;
 637
 638         ptep = pte_offset_map(pmd, address);
 639         if (!ptep)
 640                 goto out;
 641
 642         pte = *ptep;
 643         pte_unmap(ptep);
 644         if (pte_present(pte)) {
 645                 if (write && !pte_write(pte))
 646                         goto out;
 647                 pfn = pte_pfn(pte);
 648                 if (pfn_valid(pfn)) {
 649                         page = pfn_to_page(pfn);
 650                         if (write && !pte_dirty(pte) && !PageDirty(page))
 651                                 set_page_dirty(page);
 652                         mark_page_accessed(page);
 653                         return page;
 654                 }
 655         }
 656
 657 out:
 658         return NULL;
 659 }
 660
 661 struct page *
 662 follow_page_pfn(struct mm_struct *mm, unsigned long address, int write,
 663                 unsigned long *pfn_ptr)
 664 {
 665         pgd_t *pgd;
 666         pmd_t *pmd;
 667         pte_t *ptep, pte;
 668         unsigned long pfn;
 669         struct page *page;
 670
 671         *pfn_ptr = 0;
 672         page = follow_huge_addr(mm, address, write);
 673         if (!IS_ERR(page))
 674                 return page;
 675
 676         pgd = pgd_offset(mm, address);
 677         if (pgd_none(*pgd) || pgd_bad(*pgd))
 678                 goto out;
 679
 680         pmd = pmd_offset(pgd, address);
 681         if (pmd_none(*pmd))
 682                 goto out;
 683         if (pmd_huge(*pmd))
 684                 return follow_huge_pmd(mm, address, pmd, write);
 685         if (pmd_bad(*pmd))
 686                 goto out;
 687
 688         ptep = pte_offset_map(pmd, address);
 689         if (!ptep)
 690                 goto out;
 691
 692         pte = *ptep;
 693         pte_unmap(ptep);
 694         if (pte_present(pte)) {
 695                 if (write && !pte_write(pte))
 696                         goto out;
 697                 if (write && !pte_dirty(pte)) {
 698                         struct page *page = pte_page(pte);
 699                         if (!PageDirty(page))
 700                                 set_page_dirty(page);
 701                 }
 702                 pfn = pte_pfn(pte);
 703                 if (pfn_valid(pfn)) {
 704                         struct page *page = pfn_to_page(pfn);
 705
 706                         mark_page_accessed(page);
 707                         return page;
 708                 } else {
 709                         *pfn_ptr = pfn;
 710                         return NULL;
 711                 }
 712         }
 713
 714 out:
 715         return NULL;
 716 }
 717
 718
 719 /*
 720  * Given a physical address, is there a useful struct page pointing to
 721  * it?  This may become more complex in the future if we start dealing
 722  * with IO-aperture pages for direct-IO.
 723  */
 724
 725 static inline struct page *get_page_map(struct page *page)
 726 {
 727         if (!pfn_valid(page_to_pfn(page)))
 728                 return 0;
 729         return page;
 730 }
 731
 732
 733 #ifndef CONFIG_X86_4G
 734 static inline int
 735 untouched_anonymous_page(struct mm_struct* mm, struct vm_area_struct *vma,
 736                          unsigned long address)
 737 {
 738         pgd_t *pgd;
 739         pmd_t *pmd;
 740
 741         /* Check if the vma is for an anonymous mapping. */
 742         if (vma->vm_ops && vma->vm_ops->nopage)
 743                 return 0;
 744
 745         /* Check if page directory entry exists. */
 746         pgd = pgd_offset(mm, address);
 747         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
 748                 return 1;
 749
 750         /* Check if page middle directory entry exists. */
 751         pmd = pmd_offset(pgd, address);
 752         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
 753                 return 1;
 754
 755         /* There is a pte slot for 'address' in 'mm'. */
 756         return 0;
 757 }
 758 #endif
 759
 760
 761 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
 762                 unsigned long start, int len, int write, int force,
 763                 struct page **pages, struct vm_area_struct **vmas)
 764 {
 765         int i;
 766         unsigned int flags;
 767
 768         /*
 769          * Require read or write permissions.
 770          * If 'force' is set, we only require the "MAY" flags.
 771          */
 772         flags = write ? (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
 773         flags &= force ? (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
 774         i = 0;
 775
 776         do {
 777                 struct vm_area_struct * vma;
 778
 779                 vma = find_extend_vma(mm, start);
 780                 if (!vma && in_gate_area(tsk, start)) {
 781                         unsigned long pg = start & PAGE_MASK;
 782                         struct vm_area_struct *gate_vma = get_gate_vma(tsk);
 783                         pgd_t *pgd;
 784                         pmd_t *pmd;
 785                         pte_t *pte;
 786                         if (write) /* user gate pages are read-only */
 787                                 return i ? : -EFAULT;
 788                         pgd = pgd_offset_k(pg);
 789                         if (!pgd)
 790                                 return i ? : -EFAULT;
 791                         pmd = pmd_offset(pgd, pg);
 792                         if (!pmd)
 793                                 return i ? : -EFAULT;
 794                         pte = pte_offset_kernel(pmd, pg);
 795                         if (!pte || !pte_present(*pte))
 796                                 return i ? : -EFAULT;
 797                         if (pages) {
 798                                 pages[i] = pte_page(*pte);
 799                                 get_page(pages[i]);
 800                         }
 801                         if (vmas)
 802                                 vmas[i] = gate_vma;
 803                         i++;
 804                         start += PAGE_SIZE;
 805                         len--;
 806                         continue;
 807                 }
 808
 809                 if (!vma || (pages && (vma->vm_flags & VM_IO))
 810                                 || !(flags & vma->vm_flags))
 811                         return i ? : -EFAULT;
 812
 813                 if (is_vm_hugetlb_page(vma)) {
 814                         i = follow_hugetlb_page(mm, vma, pages, vmas,
 815                                                 &start, &len, i);
 816                         continue;
 817                 }
 818                 spin_lock(&mm->page_table_lock);
 819                 do {
 820                         struct page *map;
 821                         int lookup_write = write;
 822                         while (!(map = follow_page(mm, start, lookup_write))) {
 823                                 /*
 824                                  * Shortcut for anonymous pages. We don't want
 825                                  * to force the creation of pages tables for
 826                                  * insanly big anonymously mapped areas that
 827                                  * nobody touched so far. This is important
 828                                  * for doing a core dump for these mappings.
 829                                  *
 830                                  * disable this for 4:4 - it prevents
 831                                  * follow_page() from ever seeing these pages.
 832                                  *
 833                                  * (The 'fix' is dubious anyway, there's
 834                                  * nothing that this code avoids which couldnt
 835                                  * be triggered from userspace anyway.)
 836                                  */
 837 #ifndef CONFIG_X86_4G
 838                                 if (!lookup_write &&
 839                                     untouched_anonymous_page(mm,vma,start)) {
 840                                         map = ZERO_PAGE(start);
 841                                         break;
 842                                 }
 843 #endif
 844                                 spin_unlock(&mm->page_table_lock);
 845                                 switch (handle_mm_fault(mm,vma,start,write)) {
 846                                 case VM_FAULT_MINOR:
 847                                         tsk->min_flt++;
 848                                         break;
 849                                 case VM_FAULT_MAJOR:
 850                                         tsk->maj_flt++;
 851                                         break;
 852                                 case VM_FAULT_SIGBUS:
 853                                         return i ? i : -EFAULT;
 854                                 case VM_FAULT_OOM:
 855                                         return i ? i : -ENOMEM;
 856                                 default:
 857                                         BUG();
 858                                 }
 859                                 /*
 860                                  * Now that we have performed a write fault
 861                                  * and surely no longer have a shared page we
 862                                  * shouldn't write, we shouldn't ignore an
 863                                  * unwritable page in the page table if
 864                                  * we are forcing write access.
 865                                  */
 866                                 lookup_write = write && !force;
 867                                 spin_lock(&mm->page_table_lock);
 868                         }
 869                         if (pages) {
 870                                 pages[i] = get_page_map(map);
 871                                 if (!pages[i]) {
 872                                         spin_unlock(&mm->page_table_lock);
 873                                         while (i--)
 874                                                 page_cache_release(pages[i]);
 875                                         i = -EFAULT;
 876                                         goto out;
 877                                 }
 878                                 flush_dcache_page(pages[i]);
 879                                 if (!PageReserved(pages[i]))
 880                                         page_cache_get(pages[i]);
 881                         }
 882                         if (vmas)
 883                                 vmas[i] = vma;
 884                         i++;
 885                         start += PAGE_SIZE;
 886                         len--;
 887                 } while(len && start < vma->vm_end);
 888                 spin_unlock(&mm->page_table_lock);
 889         } while(len);
 890 out:
 891         return i;
 892 }
 893
 894 EXPORT_SYMBOL(get_user_pages);
 895
 896 static void zeromap_pte_range(pte_t * pte, unsigned long address,
 897                                      unsigned long size, pgprot_t prot)
 898 {
 899         unsigned long end;
 900
 901         address &= ~PMD_MASK;
 902         end = address + size;
 903         if (end > PMD_SIZE)
 904                 end = PMD_SIZE;
 905         do {
 906                 pte_t zero_pte = pte_wrprotect(mk_pte(ZERO_PAGE(address), prot));
 907                 BUG_ON(!pte_none(*pte));
 908                 set_pte(pte, zero_pte);
 909                 address += PAGE_SIZE;
 910                 pte++;
 911         } while (address && (address < end));
 912 }
 913
 914 static inline int zeromap_pmd_range(struct mm_struct *mm, pmd_t * pmd, unsigned long address,
 915                                     unsigned long size, pgprot_t prot)
 916 {
 917         unsigned long base, end;
 918
 919         base = address & PGDIR_MASK;
 920         address &= ~PGDIR_MASK;
 921         end = address + size;
 922         if (end > PGDIR_SIZE)
 923                 end = PGDIR_SIZE;
 924         do {
 925                 pte_t * pte = pte_alloc_map(mm, pmd, base + address);
 926                 if (!pte)
 927                         return -ENOMEM;
 928                 zeromap_pte_range(pte, base + address, end - address, prot);
 929                 pte_unmap(pte);
 930                 address = (address + PMD_SIZE) & PMD_MASK;
 931                 pmd++;
 932         } while (address && (address < end));
 933         return 0;
 934 }
 935
 936 int zeromap_page_range(struct vm_area_struct *vma, unsigned long address, unsigned long size, pgprot_t prot)
 937 {
 938         int error = 0;
 939         pgd_t * dir;
 940         unsigned long beg = address;
 941         unsigned long end = address + size;
 942         struct mm_struct *mm = vma->vm_mm;
 943
 944         dir = pgd_offset(mm, address);
 945         flush_cache_range(vma, beg, end);
 946         if (address >= end)
 947                 BUG();
 948
 949         spin_lock(&mm->page_table_lock);
 950         do {
 951                 pmd_t *pmd = pmd_alloc(mm, dir, address);
 952                 error = -ENOMEM;
 953                 if (!pmd)
 954                         break;
 955                 error = zeromap_pmd_range(mm, pmd, address, end - address, prot);
 956                 if (error)
 957                         break;
 958                 address = (address + PGDIR_SIZE) & PGDIR_MASK;
 959                 dir++;
 960         } while (address && (address < end));
 961         /*
 962          * Why flush? zeromap_pte_range has a BUG_ON for !pte_none()
 963          */
 964         flush_tlb_range(vma, beg, end);
 965         spin_unlock(&mm->page_table_lock);
 966         return error;
 967 }
 968
 969 /*
 970  * maps a range of physical memory into the requested pages. the old
 971  * mappings are removed. any references to nonexistent pages results
 972  * in null mappings (currently treated as "copy-on-access")
 973  */
 974 static inline void remap_pte_range(pte_t * pte, unsigned long address, unsigned long size,
 975         unsigned long phys_addr, pgprot_t prot)
 976 {
 977         unsigned long end;
 978         unsigned long pfn;
 979
 980         address &= ~PMD_MASK;
 981         end = address + size;
 982         if (end > PMD_SIZE)
 983                 end = PMD_SIZE;
 984         pfn = phys_addr >> PAGE_SHIFT;
 985         do {
 986                 BUG_ON(!pte_none(*pte));
 987                 if (!pfn_valid(pfn) || PageReserved(pfn_to_page(pfn)))
 988                         set_pte(pte, pfn_pte(pfn, prot));
 989                 address += PAGE_SIZE;
 990                 pfn++;
 991                 pte++;
 992         } while (address && (address < end));
 993 }
 994
 995 static inline int remap_pmd_range(struct mm_struct *mm, pmd_t * pmd, unsigned long address, unsigned long size,
 996         unsigned long phys_addr, pgprot_t prot)
 997 {
 998         unsigned long base, end;
 999
1000         base = address & PGDIR_MASK;
1001         address &= ~PGDIR_MASK;
1002         end = address + size;
1003         if (end > PGDIR_SIZE)
1004                 end = PGDIR_SIZE;
1005         phys_addr -= address;
1006         do {
1007                 pte_t * pte = pte_alloc_map(mm, pmd, base + address);
1008                 if (!pte)
1009                         return -ENOMEM;
1010                 remap_pte_range(pte, base + address, end - address, address + phys_addr, prot);
1011                 pte_unmap(pte);
1012                 address = (address + PMD_SIZE) & PMD_MASK;
1013                 pmd++;
1014         } while (address && (address < end));
1015         return 0;
1016 }
1017
1018 /*  Note: this is only safe if the mm semaphore is held when called. */
1019 int remap_page_range(struct vm_area_struct *vma, unsigned long from, unsigned long phys_addr, unsigned long size, pgprot_t prot)
1020 {
1021         int error = 0;
1022         pgd_t * dir;
1023         unsigned long beg = from;
1024         unsigned long end = from + size;
1025         struct mm_struct *mm = vma->vm_mm;
1026
1027         phys_addr -= from;
1028         dir = pgd_offset(mm, from);
1029         flush_cache_range(vma, beg, end);
1030         if (from >= end)
1031                 BUG();
1032
1033         spin_lock(&mm->page_table_lock);
1034         do {
1035                 pmd_t *pmd = pmd_alloc(mm, dir, from);
1036                 error = -ENOMEM;
1037                 if (!pmd)
1038                         break;
1039                 error = remap_pmd_range(mm, pmd, from, end - from, phys_addr + from, prot);
1040                 if (error)
1041                         break;
1042                 from = (from + PGDIR_SIZE) & PGDIR_MASK;
1043                 dir++;
1044         } while (from && (from < end));
1045         /*
1046          * Why flush? remap_pte_range has a BUG_ON for !pte_none()
1047          */
1048         flush_tlb_range(vma, beg, end);
1049         spin_unlock(&mm->page_table_lock);
1050         return error;
1051 }
1052
1053 EXPORT_SYMBOL(remap_page_range);
1054
1055 /*
1056  * Do pte_mkwrite, but only if the vma says VM_WRITE.  We do this when
1057  * servicing faults for write access.  In the normal case, do always want
1058  * pte_mkwrite.  But get_user_pages can cause write faults for mappings
1059  * that do not have writing enabled, when used by access_process_vm.
1060  */
1061 static inline pte_t maybe_mkwrite(pte_t pte, struct vm_area_struct *vma)
1062 {
1063         if (likely(vma->vm_flags & VM_WRITE))
1064                 pte = pte_mkwrite(pte);
1065         return pte;
1066 }
1067
1068 /*
1069  * We hold the mm semaphore for reading and vma->vm_mm->page_table_lock
1070  */
1071 static inline void break_cow(struct vm_area_struct * vma, struct page * new_page, unsigned long address,
1072                 pte_t *page_table)
1073 {
1074         pte_t entry;
1075
1076         flush_cache_page(vma, address);
1077         entry = maybe_mkwrite(pte_mkdirty(mk_pte(new_page, vma->vm_page_prot)),
1078                               vma);
1079         ptep_establish(vma, address, page_table, entry);
1080         update_mmu_cache(vma, address, entry);
1081 }
1082
1083 /*
1084  * This routine handles present pages, when users try to write
1085  * to a shared page. It is done by copying the page to a new address
1086  * and decrementing the shared-page counter for the old page.
1087  *
1088  * Goto-purists beware: the only reason for goto's here is that it results
1089  * in better assembly code.. The "default" path will see no jumps at all.
1090  *
1091  * Note that this routine assumes that the protection checks have been
1092  * done by the caller (the low-level page fault routine in most cases).
1093  * Thus we can safely just mark it writable once we've done any necessary
1094  * COW.
1095  *
1096  * We also mark the page dirty at this point even though the page will
1097  * change only once the write actually happens. This avoids a few races,
1098  * and potentially makes it more efficient.
1099  *
1100  * We hold the mm semaphore and the page_table_lock on entry and exit
1101  * with the page_table_lock released.
1102  */
1103 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct * vma,
1104         unsigned long address, pte_t *page_table, pmd_t *pmd, pte_t pte)
1105 {
1106         struct page *old_page, *new_page;
1107         unsigned long pfn = pte_pfn(pte);
1108         pte_t entry;
1109
1110         if (unlikely(!pfn_valid(pfn))) {
1111                 /*
1112                  * This should really halt the system so it can be debugged or
1113                  * at least the kernel stops what it's doing before it corrupts
1114                  * data, but for the moment just pretend this is OOM.
1115                  */
1116                 pte_unmap(page_table);
1117                 printk(KERN_ERR "do_wp_page: bogus page at address %08lx\n",
1118                                 address);
1119                 spin_unlock(&mm->page_table_lock);
1120                 return VM_FAULT_OOM;
1121         }
1122         old_page = pfn_to_page(pfn);
1123
1124         if (!TestSetPageLocked(old_page)) {
1125                 int reuse = can_share_swap_page(old_page);
1126                 unlock_page(old_page);
1127                 if (reuse) {
1128                         flush_cache_page(vma, address);
1129                         entry = maybe_mkwrite(pte_mkyoung(pte_mkdirty(pte)),
1130                                               vma);
1131                         ptep_set_access_flags(vma, address, page_table, entry, 1);
1132                         update_mmu_cache(vma, address, entry);
1133                         pte_unmap(page_table);
1134                         spin_unlock(&mm->page_table_lock);
1135                         return VM_FAULT_MINOR;
1136                 }
1137         }
1138         pte_unmap(page_table);
1139
1140         /*
1141          * Ok, we need to copy. Oh, well..
1142          */
1143         page_cache_get(old_page);
1144         spin_unlock(&mm->page_table_lock);
1145
1146         if (unlikely(anon_vma_prepare(vma)))
1147                 goto no_new_page;
1148         new_page = alloc_page_vma(GFP_HIGHUSER, vma, address);
1149         if (!new_page)
1150                 goto no_new_page;
1151         copy_cow_page(old_page,new_page,address);
1152
1153         /*
1154          * Re-check the pte - we dropped the lock
1155          */
1156         spin_lock(&mm->page_table_lock);
1157         page_table = pte_offset_map(pmd, address);
1158         if (likely(pte_same(*page_table, pte))) {
1159                 if (PageReserved(old_page))
1160                         ++mm->rss;
1161                 else
1162                         page_remove_rmap(old_page);
1163                 break_cow(vma, new_page, address, page_table);
1164                 lru_cache_add_active(new_page);
1165                 page_add_anon_rmap(new_page, vma, address);
1166
1167                 /* Free the old page.. */
1168                 new_page = old_page;
1169         }
1170         pte_unmap(page_table);
1171         page_cache_release(new_page);
1172         page_cache_release(old_page);
1173         spin_unlock(&mm->page_table_lock);
1174         return VM_FAULT_MINOR;
1175
1176 no_new_page:
1177         page_cache_release(old_page);
1178         return VM_FAULT_OOM;
1179 }
1180
1181 /*
1182  * Helper function for unmap_mapping_range().
1183  */
1184 static inline void unmap_mapping_range_list(struct prio_tree_root *root,
1185                                             struct zap_details *details)
1186 {
1187         struct vm_area_struct *vma = NULL;
1188         struct prio_tree_iter iter;
1189         pgoff_t vba, vea, zba, zea;
1190
1191         while ((vma = vma_prio_tree_next(vma, root, &iter,
1192                         details->first_index, details->last_index)) != NULL) {
1193                 vba = vma->vm_pgoff;
1194                 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
1195                 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
1196                 zba = details->first_index;
1197                 if (zba < vba)
1198                         zba = vba;
1199                 zea = details->last_index;
1200                 if (zea > vea)
1201                         zea = vea;
1202                 zap_page_range(vma,
1203                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
1204                         (zea - zba + 1) << PAGE_SHIFT, details);
1205         }
1206 }
1207
1208 /**
1209  * unmap_mapping_range - unmap the portion of all mmaps
1210  * in the specified address_space corresponding to the specified
1211  * page range in the underlying file.
1212  * @address_space: the address space containing mmaps to be unmapped.
1213  * @holebegin: byte in first page to unmap, relative to the start of
1214  * the underlying file.  This will be rounded down to a PAGE_SIZE
1215  * boundary.  Note that this is different from vmtruncate(), which
1216  * must keep the partial page.  In contrast, we must get rid of
1217  * partial pages.
1218  * @holelen: size of prospective hole in bytes.  This will be rounded
1219  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
1220  * end of the file.
1221  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
1222  * but 0 when invalidating pagecache, don't throw away private data.
1223  */
1224 void unmap_mapping_range(struct address_space *mapping,
1225                 loff_t const holebegin, loff_t const holelen, int even_cows)
1226 {
1227         struct zap_details details;
1228         pgoff_t hba = holebegin >> PAGE_SHIFT;
1229         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1230
1231         /* Check for overflow. */
1232         if (sizeof(holelen) > sizeof(hlen)) {
1233                 long long holeend =
1234                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1235                 if (holeend & ~(long long)ULONG_MAX)
1236                         hlen = ULONG_MAX - hba + 1;
1237         }
1238
1239         details.check_mapping = even_cows? NULL: mapping;
1240         details.nonlinear_vma = NULL;
1241         details.first_index = hba;
1242         details.last_index = hba + hlen - 1;
1243         details.atomic = 1;     /* A spinlock is held */
1244         if (details.last_index < details.first_index)
1245                 details.last_index = ULONG_MAX;
1246
1247         spin_lock(&mapping->i_mmap_lock);
1248         /* Protect against page fault */
1249         atomic_inc(&mapping->truncate_count);
1250
1251         if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
1252                 unmap_mapping_range_list(&mapping->i_mmap, &details);
1253
1254         /*
1255          * In nonlinear VMAs there is no correspondence between virtual address
1256          * offset and file offset.  So we must perform an exhaustive search
1257          * across *all* the pages in each nonlinear VMA, not just the pages
1258          * whose virtual address lies outside the file truncation point.
1259          */
1260         if (unlikely(!list_empty(&mapping->i_mmap_nonlinear))) {
1261                 struct vm_area_struct *vma;
1262                 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
1263                                                 shared.vm_set.list) {
1264                         details.nonlinear_vma = vma;
1265                         zap_page_range(vma, vma->vm_start,
1266                                 vma->vm_end - vma->vm_start, &details);
1267                 }
1268         }
1269         spin_unlock(&mapping->i_mmap_lock);
1270 }
1271 EXPORT_SYMBOL(unmap_mapping_range);
1272
1273 /*
1274  * Handle all mappings that got truncated by a "truncate()"
1275  * system call.
1276  *
1277  * NOTE! We have to be ready to update the memory sharing
1278  * between the file and the memory map for a potential last
1279  * incomplete page.  Ugly, but necessary.
1280  */
1281 int vmtruncate(struct inode * inode, loff_t offset)
1282 {
1283         struct address_space *mapping = inode->i_mapping;
1284         unsigned long limit;
1285
1286         if (inode->i_size < offset)
1287                 goto do_expand;
1288         i_size_write(inode, offset);
1289         unmap_mapping_range(mapping, offset + PAGE_SIZE - 1, 0, 1);
1290         truncate_inode_pages(mapping, offset);
1291         goto out_truncate;
1292
1293 do_expand:
1294         limit = current->rlim[RLIMIT_FSIZE].rlim_cur;
1295         if (limit != RLIM_INFINITY && offset > limit)
1296                 goto out_sig;
1297         if (offset > inode->i_sb->s_maxbytes)
1298                 goto out;
1299         i_size_write(inode, offset);
1300
1301 out_truncate:
1302         if (inode->i_op && inode->i_op->truncate)
1303                 inode->i_op->truncate(inode);
1304         return 0;
1305 out_sig:
1306         send_sig(SIGXFSZ, current, 0);
1307 out:
1308         return -EFBIG;
1309 }
1310
1311 EXPORT_SYMBOL(vmtruncate);
1312
1313 /*
1314  * Primitive swap readahead code. We simply read an aligned block of
1315  * (1 << page_cluster) entries in the swap area. This method is chosen
1316  * because it doesn't cost us any seek time.  We also make sure to queue
1317  * the 'original' request together with the readahead ones...
1318  *
1319  * This has been extended to use the NUMA policies from the mm triggering
1320  * the readahead.
1321  *
1322  * Caller must hold down_read on the vma->vm_mm if vma is not NULL.
1323  */
1324 void swapin_readahead(swp_entry_t entry, unsigned long addr,struct vm_area_struct *vma)
1325 {
1326 #ifdef CONFIG_NUMA
1327         struct vm_area_struct *next_vma = vma ? vma->vm_next : NULL;
1328 #endif
1329         int i, num;
1330         struct page *new_page;
1331         unsigned long offset;
1332
1333         /*
1334          * Get the number of handles we should do readahead io to.
1335          */
1336         num = valid_swaphandles(entry, &offset);
1337         for (i = 0; i < num; offset++, i++) {
1338                 /* Ok, do the async read-ahead now */
1339                 new_page = read_swap_cache_async(swp_entry(swp_type(entry),
1340                                                            offset), vma, addr);
1341                 if (!new_page)
1342                         break;
1343                 page_cache_release(new_page);
1344 #ifdef CONFIG_NUMA
1345                 /*
1346                  * Find the next applicable VMA for the NUMA policy.
1347                  */
1348                 addr += PAGE_SIZE;
1349                 if (addr == 0)
1350                         vma = NULL;
1351                 if (vma) {
1352                         if (addr >= vma->vm_end) {
1353                                 vma = next_vma;
1354                                 next_vma = vma ? vma->vm_next : NULL;
1355                         }
1356                         if (vma && addr < vma->vm_start)
1357                                 vma = NULL;
1358                 } else {
1359                         if (next_vma && addr >= next_vma->vm_start) {
1360                                 vma = next_vma;
1361                                 next_vma = vma->vm_next;
1362                         }
1363                 }
1364 #endif
1365         }
1366         lru_add_drain();        /* Push any new pages onto the LRU now */
1367 }
1368
1369 /*
1370  * We hold the mm semaphore and the page_table_lock on entry and
1371  * should release the pagetable lock on exit..
1372  */
1373 static int do_swap_page(struct mm_struct * mm,
1374         struct vm_area_struct * vma, unsigned long address,
1375         pte_t *page_table, pmd_t *pmd, pte_t orig_pte, int write_access)
1376 {
1377         struct page *page;
1378         swp_entry_t entry = pte_to_swp_entry(orig_pte);
1379         pte_t pte;
1380         int ret = VM_FAULT_MINOR;
1381
1382         pte_unmap(page_table);
1383         spin_unlock(&mm->page_table_lock);
1384         page = lookup_swap_cache(entry);
1385         if (!page) {
1386                 swapin_readahead(entry, address, vma);
1387                 page = read_swap_cache_async(entry, vma, address);
1388                 if (!page) {
1389                         /*
1390                          * Back out if somebody else faulted in this pte while
1391                          * we released the page table lock.
1392                          */
1393                         spin_lock(&mm->page_table_lock);
1394                         page_table = pte_offset_map(pmd, address);
1395                         if (likely(pte_same(*page_table, orig_pte)))
1396                                 ret = VM_FAULT_OOM;
1397                         else
1398                                 ret = VM_FAULT_MINOR;
1399                         pte_unmap(page_table);
1400                         spin_unlock(&mm->page_table_lock);
1401                         goto out;
1402                 }
1403
1404                 /* Had to read the page from swap area: Major fault */
1405                 ret = VM_FAULT_MAJOR;
1406                 inc_page_state(pgmajfault);
1407         }
1408
1409         mark_page_accessed(page);
1410         lock_page(page);
1411
1412         /*
1413          * Back out if somebody else faulted in this pte while we
1414          * released the page table lock.
1415          */
1416         spin_lock(&mm->page_table_lock);
1417         page_table = pte_offset_map(pmd, address);
1418         if (unlikely(!pte_same(*page_table, orig_pte))) {
1419                 pte_unmap(page_table);
1420                 spin_unlock(&mm->page_table_lock);
1421                 unlock_page(page);
1422                 page_cache_release(page);
1423                 ret = VM_FAULT_MINOR;
1424                 goto out;
1425         }
1426
1427         /* The page isn't present yet, go ahead with the fault. */
1428
1429         swap_free(entry);
1430         if (vm_swap_full())
1431                 remove_exclusive_swap_page(page);
1432
1433         mm->rss++;
1434         pte = mk_pte(page, vma->vm_page_prot);
1435         if (write_access && can_share_swap_page(page)) {
1436                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
1437                 write_access = 0;
1438         }
1439         unlock_page(page);
1440
1441         flush_icache_page(vma, page);
1442         set_pte(page_table, pte);
1443         page_add_anon_rmap(page, vma, address);
1444
1445         if (write_access) {
1446                 if (do_wp_page(mm, vma, address,
1447                                 page_table, pmd, pte) == VM_FAULT_OOM)
1448                         ret = VM_FAULT_OOM;
1449                 goto out;
1450         }
1451
1452         /* No need to invalidate - it was non-present before */
1453         update_mmu_cache(vma, address, pte);
1454         pte_unmap(page_table);
1455         spin_unlock(&mm->page_table_lock);
1456 out:
1457         return ret;
1458 }
1459
1460 /*
1461  * We are called with the MM semaphore and page_table_lock
1462  * spinlock held to protect against concurrent faults in
1463  * multithreaded programs.
1464  */
1465 static int
1466 do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
1467                 pte_t *page_table, pmd_t *pmd, int write_access,
1468                 unsigned long addr)
1469 {
1470         pte_t entry;
1471         struct page * page = ZERO_PAGE(addr);
1472
1473         /* Read-only mapping of ZERO_PAGE. */
1474         entry = pte_wrprotect(mk_pte(ZERO_PAGE(addr), vma->vm_page_prot));
1475
1476         /* ..except if it's a write access */
1477         if (write_access) {
1478                 /* Allocate our own private page. */
1479                 pte_unmap(page_table);
1480                 spin_unlock(&mm->page_table_lock);
1481
1482                 if (unlikely(anon_vma_prepare(vma)))
1483                         goto no_mem;
1484                 page = alloc_page_vma(GFP_HIGHUSER, vma, addr);
1485                 if (!page)
1486                         goto no_mem;
1487                 clear_user_highpage(page, addr);
1488
1489                 spin_lock(&mm->page_table_lock);
1490                 page_table = pte_offset_map(pmd, addr);
1491
1492                 if (!pte_none(*page_table)) {
1493                         pte_unmap(page_table);
1494                         page_cache_release(page);
1495                         spin_unlock(&mm->page_table_lock);
1496                         goto out;
1497                 }
1498                 mm->rss++;
1499                 entry = maybe_mkwrite(pte_mkdirty(mk_pte(page,
1500                                                          vma->vm_page_prot)),
1501                                       vma);
1502                 lru_cache_add_active(page);
1503                 mark_page_accessed(page);
1504                 page_add_anon_rmap(page, vma, addr);
1505         }
1506
1507         set_pte(page_table, entry);
1508         pte_unmap(page_table);
1509
1510         /* No need to invalidate - it was non-present before */
1511         update_mmu_cache(vma, addr, entry);
1512         spin_unlock(&mm->page_table_lock);
1513 out:
1514         return VM_FAULT_MINOR;
1515 no_mem:
1516         return VM_FAULT_OOM;
1517 }
1518
1519 /*
1520  * do_no_page() tries to create a new page mapping. It aggressively
1521  * tries to share with existing pages, but makes a separate copy if
1522  * the "write_access" parameter is true in order to avoid the next
1523  * page fault.
1524  *
1525  * As this is called only for pages that do not currently exist, we
1526  * do not need to flush old virtual caches or the TLB.
1527  *
1528  * This is called with the MM semaphore held and the page table
1529  * spinlock held. Exit with the spinlock released.
1530  */
1531 static int
1532 do_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
1533         unsigned long address, int write_access, pte_t *page_table, pmd_t *pmd)
1534 {
1535         struct page * new_page;
1536         struct address_space *mapping = NULL;
1537         pte_t entry;
1538         int sequence = 0;
1539         int ret = VM_FAULT_MINOR;
1540         int anon = 0;
1541
1542         if (!vma->vm_ops || !vma->vm_ops->nopage)
1543                 return do_anonymous_page(mm, vma, page_table,
1544                                         pmd, write_access, address);
1545         pte_unmap(page_table);
1546         spin_unlock(&mm->page_table_lock);
1547
1548         if (vma->vm_file) {
1549                 mapping = vma->vm_file->f_mapping;
1550                 sequence = atomic_read(&mapping->truncate_count);
1551         }
1552         smp_rmb();  /* Prevent CPU from reordering lock-free ->nopage() */
1553 retry:
1554         new_page = vma->vm_ops->nopage(vma, address & PAGE_MASK, &ret);
1555
1556         /* no page was available -- either SIGBUS or OOM */
1557         if (new_page == NOPAGE_SIGBUS)
1558                 return VM_FAULT_SIGBUS;
1559         if (new_page == NOPAGE_OOM)
1560                 return VM_FAULT_OOM;
1561
1562         /*
1563          * Should we do an early C-O-W break?
1564          */
1565         if (write_access && !(vma->vm_flags & VM_SHARED)) {
1566                 struct page *page;
1567
1568                 if (unlikely(anon_vma_prepare(vma)))
1569                         goto oom;
1570                 page = alloc_page_vma(GFP_HIGHUSER, vma, address);
1571                 if (!page)
1572                         goto oom;
1573                 copy_user_highpage(page, new_page, address);
1574                 page_cache_release(new_page);
1575                 new_page = page;
1576                 anon = 1;
1577         }
1578
1579         spin_lock(&mm->page_table_lock);
1580         /*
1581          * For a file-backed vma, someone could have truncated or otherwise
1582          * invalidated this page.  If unmap_mapping_range got called,
1583          * retry getting the page.
1584          */
1585         if (mapping &&
1586               (unlikely(sequence != atomic_read(&mapping->truncate_count)))) {
1587                 sequence = atomic_read(&mapping->truncate_count);
1588                 spin_unlock(&mm->page_table_lock);
1589                 page_cache_release(new_page);
1590                 goto retry;
1591         }
1592         page_table = pte_offset_map(pmd, address);
1593
1594         /*
1595          * This silly early PAGE_DIRTY setting removes a race
1596          * due to the bad i386 page protection. But it's valid
1597          * for other architectures too.
1598          *
1599          * Note that if write_access is true, we either now have
1600          * an exclusive copy of the page, or this is a shared mapping,
1601          * so we can make it writable and dirty to avoid having to
1602          * handle that later.
1603          */
1604         /* Only go through if we didn't race with anybody else... */
1605         if (pte_none(*page_table)) {
1606                 if (!PageReserved(new_page))
1607                         ++mm->rss;
1608                 flush_icache_page(vma, new_page);
1609                 entry = mk_pte(new_page, vma->vm_page_prot);
1610                 if (write_access)
1611                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1612                 set_pte(page_table, entry);
1613                 if (anon) {
1614                         lru_cache_add_active(new_page);
1615                         page_add_anon_rmap(new_page, vma, address);
1616                 } else
1617                         page_add_file_rmap(new_page);
1618                 pte_unmap(page_table);
1619         } else {
1620                 /* One of our sibling threads was faster, back out. */
1621                 pte_unmap(page_table);
1622                 page_cache_release(new_page);
1623                 spin_unlock(&mm->page_table_lock);
1624                 goto out;
1625         }
1626
1627         /* no need to invalidate: a not-present page shouldn't be cached */
1628         update_mmu_cache(vma, address, entry);
1629         spin_unlock(&mm->page_table_lock);
1630 out:
1631         return ret;
1632 oom:
1633         page_cache_release(new_page);
1634         ret = VM_FAULT_OOM;
1635         goto out;
1636 }
1637
1638 /*
1639  * Fault of a previously existing named mapping. Repopulate the pte
1640  * from the encoded file_pte if possible. This enables swappable
1641  * nonlinear vmas.
1642  */
1643 static int do_file_page(struct mm_struct * mm, struct vm_area_struct * vma,
1644         unsigned long address, int write_access, pte_t *pte, pmd_t *pmd)
1645 {
1646         unsigned long pgoff;
1647         int err;
1648
1649         BUG_ON(!vma->vm_ops || !vma->vm_ops->nopage);
1650         /*
1651          * Fall back to the linear mapping if the fs does not support
1652          * ->populate:
1653          */
1654         if (!vma->vm_ops || !vma->vm_ops->populate ||
1655                         (write_access && !(vma->vm_flags & VM_SHARED))) {
1656                 pte_clear(pte);
1657                 return do_no_page(mm, vma, address, write_access, pte, pmd);
1658         }
1659
1660         pgoff = pte_to_pgoff(*pte);
1661
1662         pte_unmap(pte);
1663         spin_unlock(&mm->page_table_lock);
1664
1665         err = vma->vm_ops->populate(vma, address & PAGE_MASK, PAGE_SIZE, vma->vm_page_prot, pgoff, 0);
1666         if (err == -ENOMEM)
1667                 return VM_FAULT_OOM;
1668         if (err)
1669                 return VM_FAULT_SIGBUS;
1670         return VM_FAULT_MAJOR;
1671 }
1672
1673 /*
1674  * These routines also need to handle stuff like marking pages dirty
1675  * and/or accessed for architectures that don't do it in hardware (most
1676  * RISC architectures).  The early dirtying is also good on the i386.
1677  *
1678  * There is also a hook called "update_mmu_cache()" that architectures
1679  * with external mmu caches can use to update those (ie the Sparc or
1680  * PowerPC hashed page tables that act as extended TLBs).
1681  *
1682  * Note the "page_table_lock". It is to protect against kswapd removing
1683  * pages from under us. Note that kswapd only ever _removes_ pages, never
1684  * adds them. As such, once we have noticed that the page is not present,
1685  * we can drop the lock early.
1686  *
1687  * The adding of pages is protected by the MM semaphore (which we hold),
1688  * so we don't need to worry about a page being suddenly been added into
1689  * our VM.
1690  *
1691  * We enter with the pagetable spinlock held, we are supposed to
1692  * release it when done.
1693  */
1694 static inline int handle_pte_fault(struct mm_struct *mm,
1695         struct vm_area_struct * vma, unsigned long address,
1696         int write_access, pte_t *pte, pmd_t *pmd)
1697 {
1698         pte_t entry;
1699
1700         entry = *pte;
1701         if (!pte_present(entry)) {
1702                 /*
1703                  * If it truly wasn't present, we know that kswapd
1704                  * and the PTE updates will not touch it later. So
1705                  * drop the lock.
1706                  */
1707                 if (pte_none(entry))
1708                         return do_no_page(mm, vma, address, write_access, pte, pmd);
1709                 if (pte_file(entry))
1710                         return do_file_page(mm, vma, address, write_access, pte, pmd);
1711                 return do_swap_page(mm, vma, address, pte, pmd, entry, write_access);
1712         }
1713
1714         if (write_access) {
1715                 if (!pte_write(entry))
1716                         return do_wp_page(mm, vma, address, pte, pmd, entry);
1717
1718                 entry = pte_mkdirty(entry);
1719         }
1720         entry = pte_mkyoung(entry);
1721         ptep_set_access_flags(vma, address, pte, entry, write_access);
1722         update_mmu_cache(vma, address, entry);
1723         pte_unmap(pte);
1724         spin_unlock(&mm->page_table_lock);
1725         return VM_FAULT_MINOR;
1726 }
1727
1728 /*
1729  * By the time we get here, we already hold the mm semaphore
1730  */
1731 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct * vma,
1732         unsigned long address, int write_access)
1733 {
1734         pgd_t *pgd;
1735         pmd_t *pmd;
1736
1737         __set_current_state(TASK_RUNNING);
1738         pgd = pgd_offset(mm, address);
1739
1740         inc_page_state(pgfault);
1741
1742         if (is_vm_hugetlb_page(vma))
1743                 return VM_FAULT_SIGBUS; /* mapping truncation does this. */
1744
1745         /*
1746          * We need the page table lock to synchronize with kswapd
1747          * and the SMP-safe atomic PTE updates.
1748          */
1749         spin_lock(&mm->page_table_lock);
1750         pmd = pmd_alloc(mm, pgd, address);
1751
1752         if (pmd) {
1753                 pte_t * pte = pte_alloc_map(mm, pmd, address);
1754                 if (pte)
1755                         return handle_pte_fault(mm, vma, address, write_access, pte, pmd);
1756         }
1757         spin_unlock(&mm->page_table_lock);
1758         return VM_FAULT_OOM;
1759 }
1760
1761 /*
1762  * Allocate page middle directory.
1763  *
1764  * We've already handled the fast-path in-line, and we own the
1765  * page table lock.
1766  *
1767  * On a two-level page table, this ends up actually being entirely
1768  * optimized away.
1769  */
1770 pmd_t fastcall *__pmd_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
1771 {
1772         pmd_t *new;
1773
1774         spin_unlock(&mm->page_table_lock);
1775         new = pmd_alloc_one(mm, address);
1776         spin_lock(&mm->page_table_lock);
1777         if (!new)
1778                 return NULL;
1779
1780         /*
1781          * Because we dropped the lock, we should re-check the
1782          * entry, as somebody else could have populated it..
1783          */
1784         if (pgd_present(*pgd)) {
1785                 pmd_free(new);
1786                 goto out;
1787         }
1788         pgd_populate(mm, pgd, new);
1789 out:
1790         return pmd_offset(pgd, address);
1791 }
1792
1793 int make_pages_present(unsigned long addr, unsigned long end)
1794 {
1795         int ret, len, write;
1796         struct vm_area_struct * vma;
1797
1798         vma = find_vma(current->mm, addr);
1799         write = (vma->vm_flags & VM_WRITE) != 0;
1800         if (addr >= end)
1801                 BUG();
1802         if (end > vma->vm_end)
1803                 BUG();
1804         len = (end+PAGE_SIZE-1)/PAGE_SIZE-addr/PAGE_SIZE;
1805         ret = get_user_pages(current, current->mm, addr,
1806                         len, write, 0, NULL, NULL);
1807         if (ret < 0)
1808                 return ret;
1809         return ret == len ? 0 : -1;
1810 }
1811
1812 /*
1813  * Map a vmalloc()-space virtual address to the physical page.
1814  */
1815 struct page * vmalloc_to_page(void * vmalloc_addr)
1816 {
1817         unsigned long addr = (unsigned long) vmalloc_addr;
1818         struct page *page = NULL;
1819         pgd_t *pgd = pgd_offset_k(addr);
1820         pmd_t *pmd;
1821         pte_t *ptep, pte;
1822
1823         if (!pgd_none(*pgd)) {
1824                 pmd = pmd_offset(pgd, addr);
1825                 if (!pmd_none(*pmd)) {
1826                         preempt_disable();
1827                         ptep = pte_offset_map(pmd, addr);
1828                         pte = *ptep;
1829                         if (pte_present(pte))
1830                                 page = pte_page(pte);
1831                         pte_unmap(ptep);
1832                         preempt_enable();
1833                 }
1834         }
1835         return page;
1836 }
1837
1838 EXPORT_SYMBOL(vmalloc_to_page);
1839
1840 #if !defined(CONFIG_ARCH_GATE_AREA)
1841
1842 #if defined(AT_SYSINFO_EHDR)
1843 struct vm_area_struct gate_vma;
1844
1845 static int __init gate_vma_init(void)
1846 {
1847         gate_vma.vm_mm = NULL;
1848         gate_vma.vm_start = FIXADDR_USER_START;
1849         gate_vma.vm_end = FIXADDR_USER_END;
1850         gate_vma.vm_page_prot = PAGE_READONLY;
1851         gate_vma.vm_flags = 0;
1852         return 0;
1853 }
1854 __initcall(gate_vma_init);
1855 #endif
1856
1857 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
1858 {
1859 #ifdef AT_SYSINFO_EHDR
1860         return &gate_vma;
1861 #else
1862         return 0;
1863 #endif
1864 }
1865
1866 int in_gate_area(struct task_struct *task, unsigned long addr)
1867 {
1868 #ifdef AT_SYSINFO_EHDR
1869         if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
1870                 return 1;
1871 #endif
1872         return 0;
1873 }
1874
1875 #endif