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                                 if (!vx_rsspages_avail(dst, 1)) {
 283                                         spin_unlock(&src->page_table_lock);
 284                                         goto nomem;
 285                                 }
 286                                 /* copy_one_pte */
 287
 288                                 if (pte_none(pte))
 289                                         goto cont_copy_pte_range_noset;
 290                                 /* pte contains position in swap, so copy. */
 291                                 if (!pte_present(pte)) {
 292                                         if (!pte_file(pte))
 293                                                 swap_duplicate(pte_to_swp_entry(pte));
 294                                         set_pte(dst_pte, pte);
 295                                         goto cont_copy_pte_range_noset;
 296                                 }
 297                                 pfn = pte_pfn(pte);
 298                                 /* the pte points outside of valid memory, the
 299                                  * mapping is assumed to be good, meaningful
 300                                  * and not mapped via rmap - duplicate the
 301                                  * mapping as is.
 302                                  */
 303                                 page = NULL;
 304                                 if (pfn_valid(pfn))
 305                                         page = pfn_to_page(pfn);
 306
 307                                 if (!page || PageReserved(page)) {
 308                                         set_pte(dst_pte, pte);
 309                                         goto cont_copy_pte_range_noset;
 310                                 }
 311
 312                                 /*
 313                                  * If it's a COW mapping, write protect it both
 314                                  * in the parent and the child
 315                                  */
 316                                 if (cow) {
 317                                         ptep_set_wrprotect(src_pte);
 318                                         pte = *src_pte;
 319                                 }
 320
 321                                 /*
 322                                  * If it's a shared mapping, mark it clean in
 323                                  * the child
 324                                  */
 325                                 if (vma->vm_flags & VM_SHARED)
 326                                         pte = pte_mkclean(pte);
 327                                 pte = pte_mkold(pte);
 328                                 get_page(page);
 329                                 // dst->rss++;
 330                                 vx_rsspages_inc(dst);
 331                                 set_pte(dst_pte, pte);
 332                                 page_dup_rmap(page);
 333 cont_copy_pte_range_noset:
 334                                 address += PAGE_SIZE;
 335                                 if (address >= end) {
 336                                         pte_unmap_nested(src_pte);
 337                                         pte_unmap(dst_pte);
 338                                         goto out_unlock;
 339                                 }
 340                                 src_pte++;
 341                                 dst_pte++;
 342                         } while ((unsigned long)src_pte & PTE_TABLE_MASK);
 343                         pte_unmap_nested(src_pte-1);
 344                         pte_unmap(dst_pte-1);
 345                         spin_unlock(&src->page_table_lock);
 346                         cond_resched_lock(&dst->page_table_lock);
 347 cont_copy_pmd_range:
 348                         src_pmd++;
 349                         dst_pmd++;
 350                 } while ((unsigned long)src_pmd & PMD_TABLE_MASK);
 351         }
 352 out_unlock:
 353         spin_unlock(&src->page_table_lock);
 354 out:
 355         return 0;
 356 nomem:
 357         return -ENOMEM;
 358 }
 359
 360 static void zap_pte_range(struct mmu_gather *tlb,
 361                 pmd_t *pmd, unsigned long address,
 362                 unsigned long size, struct zap_details *details)
 363 {
 364         unsigned long offset;
 365         pte_t *ptep;
 366
 367         if (pmd_none(*pmd))
 368                 return;
 369         if (unlikely(pmd_bad(*pmd))) {
 370                 pmd_ERROR(*pmd);
 371                 pmd_clear(pmd);
 372                 return;
 373         }
 374         ptep = pte_offset_map(pmd, address);
 375         offset = address & ~PMD_MASK;
 376         if (offset + size > PMD_SIZE)
 377                 size = PMD_SIZE - offset;
 378         size &= PAGE_MASK;
 379         if (details && !details->check_mapping && !details->nonlinear_vma)
 380                 details = NULL;
 381         for (offset=0; offset < size; ptep++, offset += PAGE_SIZE) {
 382                 pte_t pte = *ptep;
 383                 if (pte_none(pte))
 384                         continue;
 385                 if (pte_present(pte)) {
 386                         struct page *page = NULL;
 387                         unsigned long pfn = pte_pfn(pte);
 388                         if (pfn_valid(pfn)) {
 389                                 page = pfn_to_page(pfn);
 390                                 if (PageReserved(page))
 391                                         page = NULL;
 392                         }
 393                         if (unlikely(details) && page) {
 394                                 /*
 395                                  * unmap_shared_mapping_pages() wants to
 396                                  * invalidate cache without truncating:
 397                                  * unmap shared but keep private pages.
 398                                  */
 399                                 if (details->check_mapping &&
 400                                     details->check_mapping != page->mapping)
 401                                         continue;
 402                                 /*
 403                                  * Each page->index must be checked when
 404                                  * invalidating or truncating nonlinear.
 405                                  */
 406                                 if (details->nonlinear_vma &&
 407                                     (page->index < details->first_index ||
 408                                      page->index > details->last_index))
 409                                         continue;
 410                         }
 411                         pte = ptep_get_and_clear(ptep);
 412                         tlb_remove_tlb_entry(tlb, ptep, address+offset);
 413                         if (unlikely(!page))
 414                                 continue;
 415                         if (unlikely(details) && details->nonlinear_vma
 416                             && linear_page_index(details->nonlinear_vma,
 417                                         address+offset) != page->index)
 418                                 set_pte(ptep, pgoff_to_pte(page->index));
 419                         if (pte_dirty(pte))
 420                                 set_page_dirty(page);
 421                         if (pte_young(pte) && page_mapping(page))
 422                                 mark_page_accessed(page);
 423                         tlb->freed++;
 424                         page_remove_rmap(page);
 425                         tlb_remove_page(tlb, page);
 426                         continue;
 427                 }
 428                 /*
 429                  * If details->check_mapping, we leave swap entries;
 430                  * if details->nonlinear_vma, we leave file entries.
 431                  */
 432                 if (unlikely(details))
 433                         continue;
 434                 if (!pte_file(pte))
 435                         free_swap_and_cache(pte_to_swp_entry(pte));
 436                 pte_clear(ptep);
 437         }
 438         pte_unmap(ptep-1);
 439 }
 440
 441 static void zap_pmd_range(struct mmu_gather *tlb,
 442                 pgd_t * dir, unsigned long address,
 443                 unsigned long size, struct zap_details *details)
 444 {
 445         pmd_t * pmd;
 446         unsigned long end, pgd_boundary;
 447
 448         if (pgd_none(*dir))
 449                 return;
 450         if (unlikely(pgd_bad(*dir))) {
 451                 pgd_ERROR(*dir);
 452                 pgd_clear(dir);
 453                 return;
 454         }
 455         pmd = pmd_offset(dir, address);
 456         end = address + size;
 457         pgd_boundary = ((address + PGDIR_SIZE) & PGDIR_MASK);
 458         if (pgd_boundary && (end > pgd_boundary))
 459                 end = pgd_boundary;
 460         do {
 461                 zap_pte_range(tlb, pmd, address, end - address, details);
 462                 address = (address + PMD_SIZE) & PMD_MASK;
 463                 pmd++;
 464         } while (address && (address < end));
 465 }
 466
 467 static void unmap_page_range(struct mmu_gather *tlb,
 468                 struct vm_area_struct *vma, unsigned long address,
 469                 unsigned long end, struct zap_details *details)
 470 {
 471         pgd_t * dir;
 472
 473         BUG_ON(address >= end);
 474         dir = pgd_offset(vma->vm_mm, address);
 475         tlb_start_vma(tlb, vma);
 476         do {
 477                 zap_pmd_range(tlb, dir, address, end - address, details);
 478                 address = (address + PGDIR_SIZE) & PGDIR_MASK;
 479                 dir++;
 480         } while (address && (address < end));
 481         tlb_end_vma(tlb, vma);
 482 }
 483
 484 /* Dispose of an entire struct mmu_gather per rescheduling point */
 485 #if defined(CONFIG_SMP) && defined(CONFIG_PREEMPT)
 486 #define ZAP_BLOCK_SIZE  (FREE_PTE_NR * PAGE_SIZE)
 487 #endif
 488
 489 /* For UP, 256 pages at a time gives nice low latency */
 490 #if !defined(CONFIG_SMP) && defined(CONFIG_PREEMPT)
 491 #define ZAP_BLOCK_SIZE  (256 * PAGE_SIZE)
 492 #endif
 493
 494 /* No preempt: go for improved straight-line efficiency */
 495 #if !defined(CONFIG_PREEMPT)
 496 #define ZAP_BLOCK_SIZE  (1024 * PAGE_SIZE)
 497 #endif
 498
 499 /**
 500  * unmap_vmas - unmap a range of memory covered by a list of vma's
 501  * @tlbp: address of the caller's struct mmu_gather
 502  * @mm: the controlling mm_struct
 503  * @vma: the starting vma
 504  * @start_addr: virtual address at which to start unmapping
 505  * @end_addr: virtual address at which to end unmapping
 506  * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
 507  * @details: details of nonlinear truncation or shared cache invalidation
 508  *
 509  * Returns the number of vma's which were covered by the unmapping.
 510  *
 511  * Unmap all pages in the vma list.  Called under page_table_lock.
 512  *
 513  * We aim to not hold page_table_lock for too long (for scheduling latency
 514  * reasons).  So zap pages in ZAP_BLOCK_SIZE bytecounts.  This means we need to
 515  * return the ending mmu_gather to the caller.
 516  *
 517  * Only addresses between `start' and `end' will be unmapped.
 518  *
 519  * The VMA list must be sorted in ascending virtual address order.
 520  *
 521  * unmap_vmas() assumes that the caller will flush the whole unmapped address
 522  * range after unmap_vmas() returns.  So the only responsibility here is to
 523  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
 524  * drops the lock and schedules.
 525  */
 526 int unmap_vmas(struct mmu_gather **tlbp, struct mm_struct *mm,
 527                 struct vm_area_struct *vma, unsigned long start_addr,
 528                 unsigned long end_addr, unsigned long *nr_accounted,
 529                 struct zap_details *details)
 530 {
 531         unsigned long zap_bytes = ZAP_BLOCK_SIZE;
 532         unsigned long tlb_start = 0;    /* For tlb_finish_mmu */
 533         int tlb_start_valid = 0;
 534         int ret = 0;
 535         int atomic = details && details->atomic;
 536
 537         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
 538                 unsigned long start;
 539                 unsigned long end;
 540
 541                 start = max(vma->vm_start, start_addr);
 542                 if (start >= vma->vm_end)
 543                         continue;
 544                 end = min(vma->vm_end, end_addr);
 545                 if (end <= vma->vm_start)
 546                         continue;
 547
 548                 if (vma->vm_flags & VM_ACCOUNT)
 549                         *nr_accounted += (end - start) >> PAGE_SHIFT;
 550
 551                 ret++;
 552                 while (start != end) {
 553                         unsigned long block;
 554
 555                         if (!tlb_start_valid) {
 556                                 tlb_start = start;
 557                                 tlb_start_valid = 1;
 558                         }
 559
 560                         if (is_vm_hugetlb_page(vma)) {
 561                                 block = end - start;
 562                                 unmap_hugepage_range(vma, start, end);
 563                         } else {
 564                                 block = min(zap_bytes, end - start);
 565                                 unmap_page_range(*tlbp, vma, start,
 566                                                 start + block, details);
 567                         }
 568
 569                         start += block;
 570                         zap_bytes -= block;
 571                         if ((long)zap_bytes > 0)
 572                                 continue;
 573                         if (!atomic && need_resched()) {
 574                                 int fullmm = tlb_is_full_mm(*tlbp);
 575                                 tlb_finish_mmu(*tlbp, tlb_start, start);
 576                                 cond_resched_lock(&mm->page_table_lock);
 577                                 *tlbp = tlb_gather_mmu(mm, fullmm);
 578                                 tlb_start_valid = 0;
 579                         }
 580                         zap_bytes = ZAP_BLOCK_SIZE;
 581                 }
 582         }
 583         return ret;
 584 }
 585
 586 /**
 587  * zap_page_range - remove user pages in a given range
 588  * @vma: vm_area_struct holding the applicable pages
 589  * @address: starting address of pages to zap
 590  * @size: number of bytes to zap
 591  * @details: details of nonlinear truncation or shared cache invalidation
 592  */
 593 void zap_page_range(struct vm_area_struct *vma, unsigned long address,
 594                 unsigned long size, struct zap_details *details)
 595 {
 596         struct mm_struct *mm = vma->vm_mm;
 597         struct mmu_gather *tlb;
 598         unsigned long end = address + size;
 599         unsigned long nr_accounted = 0;
 600
 601         if (is_vm_hugetlb_page(vma)) {
 602                 zap_hugepage_range(vma, address, size);
 603                 return;
 604         }
 605
 606         lru_add_drain();
 607         spin_lock(&mm->page_table_lock);
 608         tlb = tlb_gather_mmu(mm, 0);
 609         unmap_vmas(&tlb, mm, vma, address, end, &nr_accounted, details);
 610         tlb_finish_mmu(tlb, address, end);
 611         spin_unlock(&mm->page_table_lock);
 612 }
 613
 614 /*
 615  * Do a quick page-table lookup for a single page.
 616  * mm->page_table_lock must be held.
 617  */
 618 struct page *
 619 follow_page(struct mm_struct *mm, unsigned long address, int write)
 620 {
 621         pgd_t *pgd;
 622         pmd_t *pmd;
 623         pte_t *ptep, pte;
 624         unsigned long pfn;
 625         struct page *page;
 626
 627         page = follow_huge_addr(mm, address, write);
 628         if (! IS_ERR(page))
 629                 return page;
 630
 631         pgd = pgd_offset(mm, address);
 632         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
 633                 goto out;
 634
 635         pmd = pmd_offset(pgd, address);
 636         if (pmd_none(*pmd))
 637                 goto out;
 638         if (pmd_huge(*pmd))
 639                 return follow_huge_pmd(mm, address, pmd, write);
 640         if (unlikely(pmd_bad(*pmd)))
 641                 goto out;
 642
 643         ptep = pte_offset_map(pmd, address);
 644         if (!ptep)
 645                 goto out;
 646
 647         pte = *ptep;
 648         pte_unmap(ptep);
 649         if (pte_present(pte)) {
 650                 if (write && !pte_write(pte))
 651                         goto out;
 652                 pfn = pte_pfn(pte);
 653                 if (pfn_valid(pfn)) {
 654                         page = pfn_to_page(pfn);
 655                         if (write && !pte_dirty(pte) && !PageDirty(page))
 656                                 set_page_dirty(page);
 657                         mark_page_accessed(page);
 658                         return page;
 659                 }
 660         }
 661
 662 out:
 663         return NULL;
 664 }
 665
 666 struct page *
 667 follow_page_pfn(struct mm_struct *mm, unsigned long address, int write,
 668                 unsigned long *pfn_ptr)
 669 {
 670         pgd_t *pgd;
 671         pmd_t *pmd;
 672         pte_t *ptep, pte;
 673         unsigned long pfn;
 674         struct page *page;
 675
 676         *pfn_ptr = 0;
 677         page = follow_huge_addr(mm, address, write);
 678         if (!IS_ERR(page))
 679                 return page;
 680
 681         pgd = pgd_offset(mm, address);
 682         if (pgd_none(*pgd) || pgd_bad(*pgd))
 683                 goto out;
 684
 685         pmd = pmd_offset(pgd, address);
 686         if (pmd_none(*pmd))
 687                 goto out;
 688         if (pmd_huge(*pmd))
 689                 return follow_huge_pmd(mm, address, pmd, write);
 690         if (unlikely(pmd_bad(*pmd)))
 691                 goto out;
 692
 693         ptep = pte_offset_map(pmd, address);
 694         if (!ptep)
 695                 goto out;
 696
 697         pte = *ptep;
 698         pte_unmap(ptep);
 699         if (pte_present(pte)) {
 700                 if (write && !pte_write(pte))
 701                         goto out;
 702                 pfn = pte_pfn(pte);
 703                 if (pfn_valid(pfn)) {
 704                         page = pfn_to_page(pfn);
 705                         if (write && !pte_dirty(pte) && !PageDirty(page))
 706                                 set_page_dirty(page);
 707                         mark_page_accessed(page);
 708                         return page;
 709                 } else {
 710                         *pfn_ptr = pfn;
 711                         return NULL;
 712                 }
 713         }
 714
 715 out:
 716         return NULL;
 717 }
 718
 719
 720 /*
 721  * Given a physical address, is there a useful struct page pointing to
 722  * it?  This may become more complex in the future if we start dealing
 723  * with IO-aperture pages for direct-IO.
 724  */
 725
 726 static inline struct page *get_page_map(struct page *page)
 727 {
 728         if (!pfn_valid(page_to_pfn(page)))
 729                 return 0;
 730         return page;
 731 }
 732
 733
 734 #ifndef CONFIG_X86_4G
 735 static inline int
 736 untouched_anonymous_page(struct mm_struct* mm, struct vm_area_struct *vma,
 737                          unsigned long address)
 738 {
 739         pgd_t *pgd;
 740         pmd_t *pmd;
 741
 742         /* Check if the vma is for an anonymous mapping. */
 743         if (vma->vm_ops && vma->vm_ops->nopage)
 744                 return 0;
 745
 746         /* Check if page directory entry exists. */
 747         pgd = pgd_offset(mm, address);
 748         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
 749                 return 1;
 750
 751         /* Check if page middle directory entry exists. */
 752         pmd = pmd_offset(pgd, address);
 753         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
 754                 return 1;
 755
 756         /* There is a pte slot for 'address' in 'mm'. */
 757         return 0;
 758 }
 759 #endif
 760
 761
 762 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
 763                 unsigned long start, int len, int write, int force,
 764                 struct page **pages, struct vm_area_struct **vmas)
 765 {
 766         int i;
 767         unsigned int flags;
 768
 769         /*
 770          * Require read or write permissions.
 771          * If 'force' is set, we only require the "MAY" flags.
 772          */
 773         flags = write ? (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
 774         flags &= force ? (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
 775         i = 0;
 776
 777         do {
 778                 struct vm_area_struct * vma;
 779
 780                 vma = find_extend_vma(mm, start);
 781                 if (!vma && in_gate_area(tsk, start)) {
 782                         unsigned long pg = start & PAGE_MASK;
 783                         struct vm_area_struct *gate_vma = get_gate_vma(tsk);
 784                         pgd_t *pgd;
 785                         pmd_t *pmd;
 786                         pte_t *pte;
 787                         if (write) /* user gate pages are read-only */
 788                                 return i ? : -EFAULT;
 789                         pgd = pgd_offset_k(pg);
 790                         if (!pgd)
 791                                 return i ? : -EFAULT;
 792                         pmd = pmd_offset(pgd, pg);
 793                         if (!pmd)
 794                                 return i ? : -EFAULT;
 795                         pte = pte_offset_kernel(pmd, pg);
 796                         if (!pte || !pte_present(*pte))
 797                                 return i ? : -EFAULT;
 798                         if (pages) {
 799                                 pages[i] = pte_page(*pte);
 800                                 get_page(pages[i]);
 801                         }
 802                         if (vmas)
 803                                 vmas[i] = gate_vma;
 804                         i++;
 805                         start += PAGE_SIZE;
 806                         len--;
 807                         continue;
 808                 }
 809
 810                 if (!vma || (pages && (vma->vm_flags & VM_IO))
 811                                 || !(flags & vma->vm_flags))
 812                         return i ? : -EFAULT;
 813
 814                 if (is_vm_hugetlb_page(vma)) {
 815                         i = follow_hugetlb_page(mm, vma, pages, vmas,
 816                                                 &start, &len, i);
 817                         continue;
 818                 }
 819                 spin_lock(&mm->page_table_lock);
 820                 do {
 821                         struct page *map;
 822                         int lookup_write = write;
 823                         while (!(map = follow_page(mm, start, lookup_write))) {
 824                                 /*
 825                                  * Shortcut for anonymous pages. We don't want
 826                                  * to force the creation of pages tables for
 827                                  * insanly big anonymously mapped areas that
 828                                  * nobody touched so far. This is important
 829                                  * for doing a core dump for these mappings.
 830                                  *
 831                                  * disable this for 4:4 - it prevents
 832                                  * follow_page() from ever seeing these pages.
 833                                  *
 834                                  * (The 'fix' is dubious anyway, there's
 835                                  * nothing that this code avoids which couldnt
 836                                  * be triggered from userspace anyway.)
 837                                  */
 838 #ifndef CONFIG_X86_4G
 839                                 if (!lookup_write &&
 840                                     untouched_anonymous_page(mm,vma,start)) {
 841                                         map = ZERO_PAGE(start);
 842                                         break;
 843                                 }
 844 #endif
 845                                 spin_unlock(&mm->page_table_lock);
 846                                 switch (handle_mm_fault(mm,vma,start,write)) {
 847                                 case VM_FAULT_MINOR:
 848                                         tsk->min_flt++;
 849                                         break;
 850                                 case VM_FAULT_MAJOR:
 851                                         tsk->maj_flt++;
 852                                         break;
 853                                 case VM_FAULT_SIGBUS:
 854                                         return i ? i : -EFAULT;
 855                                 case VM_FAULT_OOM:
 856                                         return i ? i : -ENOMEM;
 857                                 default:
 858                                         BUG();
 859                                 }
 860                                 /*
 861                                  * Now that we have performed a write fault
 862                                  * and surely no longer have a shared page we
 863                                  * shouldn't write, we shouldn't ignore an
 864                                  * unwritable page in the page table if
 865                                  * we are forcing write access.
 866                                  */
 867                                 lookup_write = write && !force;
 868                                 spin_lock(&mm->page_table_lock);
 869                         }
 870                         if (pages) {
 871                                 pages[i] = get_page_map(map);
 872                                 if (!pages[i]) {
 873                                         spin_unlock(&mm->page_table_lock);
 874                                         while (i--)
 875                                                 page_cache_release(pages[i]);
 876                                         i = -EFAULT;
 877                                         goto out;
 878                                 }
 879                                 flush_dcache_page(pages[i]);
 880                                 if (!PageReserved(pages[i]))
 881                                         page_cache_get(pages[i]);
 882                         }
 883                         if (vmas)
 884                                 vmas[i] = vma;
 885                         i++;
 886                         start += PAGE_SIZE;
 887                         len--;
 888                 } while(len && start < vma->vm_end);
 889                 spin_unlock(&mm->page_table_lock);
 890         } while(len);
 891 out:
 892         return i;
 893 }
 894
 895 EXPORT_SYMBOL(get_user_pages);
 896
 897 static void zeromap_pte_range(pte_t * pte, unsigned long address,
 898                                      unsigned long size, pgprot_t prot)
 899 {
 900         unsigned long end;
 901
 902         address &= ~PMD_MASK;
 903         end = address + size;
 904         if (end > PMD_SIZE)
 905                 end = PMD_SIZE;
 906         do {
 907                 pte_t zero_pte = pte_wrprotect(mk_pte(ZERO_PAGE(address), prot));
 908                 BUG_ON(!pte_none(*pte));
 909                 set_pte(pte, zero_pte);
 910                 address += PAGE_SIZE;
 911                 pte++;
 912         } while (address && (address < end));
 913 }
 914
 915 static inline int zeromap_pmd_range(struct mm_struct *mm, pmd_t * pmd, unsigned long address,
 916                                     unsigned long size, pgprot_t prot)
 917 {
 918         unsigned long base, end;
 919
 920         base = address & PGDIR_MASK;
 921         address &= ~PGDIR_MASK;
 922         end = address + size;
 923         if (end > PGDIR_SIZE)
 924                 end = PGDIR_SIZE;
 925         do {
 926                 pte_t * pte = pte_alloc_map(mm, pmd, base + address);
 927                 if (!pte)
 928                         return -ENOMEM;
 929                 zeromap_pte_range(pte, base + address, end - address, prot);
 930                 pte_unmap(pte);
 931                 address = (address + PMD_SIZE) & PMD_MASK;
 932                 pmd++;
 933         } while (address && (address < end));
 934         return 0;
 935 }
 936
 937 int zeromap_page_range(struct vm_area_struct *vma, unsigned long address, unsigned long size, pgprot_t prot)
 938 {
 939         int error = 0;
 940         pgd_t * dir;
 941         unsigned long beg = address;
 942         unsigned long end = address + size;
 943         struct mm_struct *mm = vma->vm_mm;
 944
 945         dir = pgd_offset(mm, address);
 946         flush_cache_range(vma, beg, end);
 947         if (address >= end)
 948                 BUG();
 949
 950         spin_lock(&mm->page_table_lock);
 951         do {
 952                 pmd_t *pmd = pmd_alloc(mm, dir, address);
 953                 error = -ENOMEM;
 954                 if (!pmd)
 955                         break;
 956                 error = zeromap_pmd_range(mm, pmd, address, end - address, prot);
 957                 if (error)
 958                         break;
 959                 address = (address + PGDIR_SIZE) & PGDIR_MASK;
 960                 dir++;
 961         } while (address && (address < end));
 962         /*
 963          * Why flush? zeromap_pte_range has a BUG_ON for !pte_none()
 964          */
 965         flush_tlb_range(vma, beg, end);
 966         spin_unlock(&mm->page_table_lock);
 967         return error;
 968 }
 969
 970 /*
 971  * maps a range of physical memory into the requested pages. the old
 972  * mappings are removed. any references to nonexistent pages results
 973  * in null mappings (currently treated as "copy-on-access")
 974  */
 975 static inline void remap_pte_range(pte_t * pte, unsigned long address, unsigned long size,
 976         unsigned long phys_addr, pgprot_t prot)
 977 {
 978         unsigned long end;
 979         unsigned long pfn;
 980
 981         address &= ~PMD_MASK;
 982         end = address + size;
 983         if (end > PMD_SIZE)
 984                 end = PMD_SIZE;
 985         pfn = phys_addr >> PAGE_SHIFT;
 986         do {
 987                 BUG_ON(!pte_none(*pte));
 988                 if (!pfn_valid(pfn) || PageReserved(pfn_to_page(pfn)))
 989                         set_pte(pte, pfn_pte(pfn, prot));
 990                 address += PAGE_SIZE;
 991                 pfn++;
 992                 pte++;
 993         } while (address && (address < end));
 994 }
 995
 996 static inline int remap_pmd_range(struct mm_struct *mm, pmd_t * pmd, unsigned long address, unsigned long size,
 997         unsigned long phys_addr, pgprot_t prot)
 998 {
 999         unsigned long base, end;
1000
1001         base = address & PGDIR_MASK;
1002         address &= ~PGDIR_MASK;
1003         end = address + size;
1004         if (end > PGDIR_SIZE)
1005                 end = PGDIR_SIZE;
1006         phys_addr -= address;
1007         do {
1008                 pte_t * pte = pte_alloc_map(mm, pmd, base + address);
1009                 if (!pte)
1010                         return -ENOMEM;
1011                 remap_pte_range(pte, base + address, end - address, address + phys_addr, prot);
1012                 pte_unmap(pte);
1013                 address = (address + PMD_SIZE) & PMD_MASK;
1014                 pmd++;
1015         } while (address && (address < end));
1016         return 0;
1017 }
1018
1019 /*  Note: this is only safe if the mm semaphore is held when called. */
1020 int remap_page_range(struct vm_area_struct *vma, unsigned long from, unsigned long phys_addr, unsigned long size, pgprot_t prot)
1021 {
1022         int error = 0;
1023         pgd_t * dir;
1024         unsigned long beg = from;
1025         unsigned long end = from + size;
1026         struct mm_struct *mm = vma->vm_mm;
1027
1028         phys_addr -= from;
1029         dir = pgd_offset(mm, from);
1030         flush_cache_range(vma, beg, end);
1031         if (from >= end)
1032                 BUG();
1033
1034         spin_lock(&mm->page_table_lock);
1035         do {
1036                 pmd_t *pmd = pmd_alloc(mm, dir, from);
1037                 error = -ENOMEM;
1038                 if (!pmd)
1039                         break;
1040                 error = remap_pmd_range(mm, pmd, from, end - from, phys_addr + from, prot);
1041                 if (error)
1042                         break;
1043                 from = (from + PGDIR_SIZE) & PGDIR_MASK;
1044                 dir++;
1045         } while (from && (from < end));
1046         /*
1047          * Why flush? remap_pte_range has a BUG_ON for !pte_none()
1048          */
1049         flush_tlb_range(vma, beg, end);
1050         spin_unlock(&mm->page_table_lock);
1051         return error;
1052 }
1053
1054 EXPORT_SYMBOL(remap_page_range);
1055
1056 /*
1057  * Do pte_mkwrite, but only if the vma says VM_WRITE.  We do this when
1058  * servicing faults for write access.  In the normal case, do always want
1059  * pte_mkwrite.  But get_user_pages can cause write faults for mappings
1060  * that do not have writing enabled, when used by access_process_vm.
1061  */
1062 static inline pte_t maybe_mkwrite(pte_t pte, struct vm_area_struct *vma)
1063 {
1064         if (likely(vma->vm_flags & VM_WRITE))
1065                 pte = pte_mkwrite(pte);
1066         return pte;
1067 }
1068
1069 /*
1070  * We hold the mm semaphore for reading and vma->vm_mm->page_table_lock
1071  */
1072 static inline void break_cow(struct vm_area_struct * vma, struct page * new_page, unsigned long address,
1073                 pte_t *page_table)
1074 {
1075         pte_t entry;
1076
1077         flush_cache_page(vma, address);
1078         entry = maybe_mkwrite(pte_mkdirty(mk_pte(new_page, vma->vm_page_prot)),
1079                               vma);
1080         ptep_establish(vma, address, page_table, entry);
1081         update_mmu_cache(vma, address, entry);
1082 }
1083
1084 /*
1085  * This routine handles present pages, when users try to write
1086  * to a shared page. It is done by copying the page to a new address
1087  * and decrementing the shared-page counter for the old page.
1088  *
1089  * Goto-purists beware: the only reason for goto's here is that it results
1090  * in better assembly code.. The "default" path will see no jumps at all.
1091  *
1092  * Note that this routine assumes that the protection checks have been
1093  * done by the caller (the low-level page fault routine in most cases).
1094  * Thus we can safely just mark it writable once we've done any necessary
1095  * COW.
1096  *
1097  * We also mark the page dirty at this point even though the page will
1098  * change only once the write actually happens. This avoids a few races,
1099  * and potentially makes it more efficient.
1100  *
1101  * We hold the mm semaphore and the page_table_lock on entry and exit
1102  * with the page_table_lock released.
1103  */
1104 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct * vma,
1105         unsigned long address, pte_t *page_table, pmd_t *pmd, pte_t pte)
1106 {
1107         struct page *old_page, *new_page;
1108         unsigned long pfn = pte_pfn(pte);
1109         pte_t entry;
1110
1111         if (unlikely(!pfn_valid(pfn))) {
1112                 /*
1113                  * This should really halt the system so it can be debugged or
1114                  * at least the kernel stops what it's doing before it corrupts
1115                  * data, but for the moment just pretend this is OOM.
1116                  */
1117                 pte_unmap(page_table);
1118                 printk(KERN_ERR "do_wp_page: bogus page at address %08lx\n",
1119                                 address);
1120                 spin_unlock(&mm->page_table_lock);
1121                 return VM_FAULT_OOM;
1122         }
1123         old_page = pfn_to_page(pfn);
1124
1125         if (!TestSetPageLocked(old_page)) {
1126                 int reuse = can_share_swap_page(old_page);
1127                 unlock_page(old_page);
1128                 if (reuse) {
1129                         flush_cache_page(vma, address);
1130                         entry = maybe_mkwrite(pte_mkyoung(pte_mkdirty(pte)),
1131                                               vma);
1132                         ptep_set_access_flags(vma, address, page_table, entry, 1);
1133                         update_mmu_cache(vma, address, entry);
1134                         pte_unmap(page_table);
1135                         spin_unlock(&mm->page_table_lock);
1136                         return VM_FAULT_MINOR;
1137                 }
1138         }
1139         pte_unmap(page_table);
1140
1141         /*
1142          * Ok, we need to copy. Oh, well..
1143          */
1144         page_cache_get(old_page);
1145         spin_unlock(&mm->page_table_lock);
1146
1147         if (unlikely(anon_vma_prepare(vma)))
1148                 goto no_new_page;
1149         new_page = alloc_page_vma(GFP_HIGHUSER, vma, address);
1150         if (!new_page)
1151                 goto no_new_page;
1152         copy_cow_page(old_page,new_page,address);
1153
1154         /*
1155          * Re-check the pte - we dropped the lock
1156          */
1157         spin_lock(&mm->page_table_lock);
1158         page_table = pte_offset_map(pmd, address);
1159         if (likely(pte_same(*page_table, pte))) {
1160                 if (PageReserved(old_page))
1161                         // ++mm->rss;
1162                         vx_rsspages_inc(mm);
1163                 else
1164                         page_remove_rmap(old_page);
1165                 break_cow(vma, new_page, address, page_table);
1166                 lru_cache_add_active(new_page);
1167                 page_add_anon_rmap(new_page, vma, address);
1168
1169                 /* Free the old page.. */
1170                 new_page = old_page;
1171         }
1172         pte_unmap(page_table);
1173         page_cache_release(new_page);
1174         page_cache_release(old_page);
1175         spin_unlock(&mm->page_table_lock);
1176         return VM_FAULT_MINOR;
1177
1178 no_new_page:
1179         page_cache_release(old_page);
1180         return VM_FAULT_OOM;
1181 }
1182
1183 /*
1184  * Helper function for unmap_mapping_range().
1185  */
1186 static inline void unmap_mapping_range_list(struct prio_tree_root *root,
1187                                             struct zap_details *details)
1188 {
1189         struct vm_area_struct *vma = NULL;
1190         struct prio_tree_iter iter;
1191         pgoff_t vba, vea, zba, zea;
1192
1193         while ((vma = vma_prio_tree_next(vma, root, &iter,
1194                         details->first_index, details->last_index)) != NULL) {
1195                 vba = vma->vm_pgoff;
1196                 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
1197                 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
1198                 zba = details->first_index;
1199                 if (zba < vba)
1200                         zba = vba;
1201                 zea = details->last_index;
1202                 if (zea > vea)
1203                         zea = vea;
1204                 zap_page_range(vma,
1205                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
1206                         (zea - zba + 1) << PAGE_SHIFT, details);
1207         }
1208 }
1209
1210 /**
1211  * unmap_mapping_range - unmap the portion of all mmaps
1212  * in the specified address_space corresponding to the specified
1213  * page range in the underlying file.
1214  * @address_space: the address space containing mmaps to be unmapped.
1215  * @holebegin: byte in first page to unmap, relative to the start of
1216  * the underlying file.  This will be rounded down to a PAGE_SIZE
1217  * boundary.  Note that this is different from vmtruncate(), which
1218  * must keep the partial page.  In contrast, we must get rid of
1219  * partial pages.
1220  * @holelen: size of prospective hole in bytes.  This will be rounded
1221  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
1222  * end of the file.
1223  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
1224  * but 0 when invalidating pagecache, don't throw away private data.
1225  */
1226 void unmap_mapping_range(struct address_space *mapping,
1227                 loff_t const holebegin, loff_t const holelen, int even_cows)
1228 {
1229         struct zap_details details;
1230         pgoff_t hba = holebegin >> PAGE_SHIFT;
1231         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1232
1233         /* Check for overflow. */
1234         if (sizeof(holelen) > sizeof(hlen)) {
1235                 long long holeend =
1236                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1237                 if (holeend & ~(long long)ULONG_MAX)
1238                         hlen = ULONG_MAX - hba + 1;
1239         }
1240
1241         details.check_mapping = even_cows? NULL: mapping;
1242         details.nonlinear_vma = NULL;
1243         details.first_index = hba;
1244         details.last_index = hba + hlen - 1;
1245         details.atomic = 1;     /* A spinlock is held */
1246         if (details.last_index < details.first_index)
1247                 details.last_index = ULONG_MAX;
1248
1249         spin_lock(&mapping->i_mmap_lock);
1250         /* Protect against page fault */
1251         atomic_inc(&mapping->truncate_count);
1252
1253         if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
1254                 unmap_mapping_range_list(&mapping->i_mmap, &details);
1255
1256         /*
1257          * In nonlinear VMAs there is no correspondence between virtual address
1258          * offset and file offset.  So we must perform an exhaustive search
1259          * across *all* the pages in each nonlinear VMA, not just the pages
1260          * whose virtual address lies outside the file truncation point.
1261          */
1262         if (unlikely(!list_empty(&mapping->i_mmap_nonlinear))) {
1263                 struct vm_area_struct *vma;
1264                 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
1265                                                 shared.vm_set.list) {
1266                         details.nonlinear_vma = vma;
1267                         zap_page_range(vma, vma->vm_start,
1268                                 vma->vm_end - vma->vm_start, &details);
1269                 }
1270         }
1271         spin_unlock(&mapping->i_mmap_lock);
1272 }
1273 EXPORT_SYMBOL(unmap_mapping_range);
1274
1275 /*
1276  * Handle all mappings that got truncated by a "truncate()"
1277  * system call.
1278  *
1279  * NOTE! We have to be ready to update the memory sharing
1280  * between the file and the memory map for a potential last
1281  * incomplete page.  Ugly, but necessary.
1282  */
1283 int vmtruncate(struct inode * inode, loff_t offset)
1284 {
1285         struct address_space *mapping = inode->i_mapping;
1286         unsigned long limit;
1287
1288         if (inode->i_size < offset)
1289                 goto do_expand;
1290         i_size_write(inode, offset);
1291         unmap_mapping_range(mapping, offset + PAGE_SIZE - 1, 0, 1);
1292         truncate_inode_pages(mapping, offset);
1293         goto out_truncate;
1294
1295 do_expand:
1296         limit = current->rlim[RLIMIT_FSIZE].rlim_cur;
1297         if (limit != RLIM_INFINITY && offset > limit)
1298                 goto out_sig;
1299         if (offset > inode->i_sb->s_maxbytes)
1300                 goto out;
1301         i_size_write(inode, offset);
1302
1303 out_truncate:
1304         if (inode->i_op && inode->i_op->truncate)
1305                 inode->i_op->truncate(inode);
1306         return 0;
1307 out_sig:
1308         send_sig(SIGXFSZ, current, 0);
1309 out:
1310         return -EFBIG;
1311 }
1312
1313 EXPORT_SYMBOL(vmtruncate);
1314
1315 /*
1316  * Primitive swap readahead code. We simply read an aligned block of
1317  * (1 << page_cluster) entries in the swap area. This method is chosen
1318  * because it doesn't cost us any seek time.  We also make sure to queue
1319  * the 'original' request together with the readahead ones...
1320  *
1321  * This has been extended to use the NUMA policies from the mm triggering
1322  * the readahead.
1323  *
1324  * Caller must hold down_read on the vma->vm_mm if vma is not NULL.
1325  */
1326 void swapin_readahead(swp_entry_t entry, unsigned long addr,struct vm_area_struct *vma)
1327 {
1328 #ifdef CONFIG_NUMA
1329         struct vm_area_struct *next_vma = vma ? vma->vm_next : NULL;
1330 #endif
1331         int i, num;
1332         struct page *new_page;
1333         unsigned long offset;
1334
1335         /*
1336          * Get the number of handles we should do readahead io to.
1337          */
1338         num = valid_swaphandles(entry, &offset);
1339         for (i = 0; i < num; offset++, i++) {
1340                 /* Ok, do the async read-ahead now */
1341                 new_page = read_swap_cache_async(swp_entry(swp_type(entry),
1342                                                            offset), vma, addr);
1343                 if (!new_page)
1344                         break;
1345                 page_cache_release(new_page);
1346 #ifdef CONFIG_NUMA
1347                 /*
1348                  * Find the next applicable VMA for the NUMA policy.
1349                  */
1350                 addr += PAGE_SIZE;
1351                 if (addr == 0)
1352                         vma = NULL;
1353                 if (vma) {
1354                         if (addr >= vma->vm_end) {
1355                                 vma = next_vma;
1356                                 next_vma = vma ? vma->vm_next : NULL;
1357                         }
1358                         if (vma && addr < vma->vm_start)
1359                                 vma = NULL;
1360                 } else {
1361                         if (next_vma && addr >= next_vma->vm_start) {
1362                                 vma = next_vma;
1363                                 next_vma = vma->vm_next;
1364                         }
1365                 }
1366 #endif
1367         }
1368         lru_add_drain();        /* Push any new pages onto the LRU now */
1369 }
1370
1371 /*
1372  * We hold the mm semaphore and the page_table_lock on entry and
1373  * should release the pagetable lock on exit..
1374  */
1375 static int do_swap_page(struct mm_struct * mm,
1376         struct vm_area_struct * vma, unsigned long address,
1377         pte_t *page_table, pmd_t *pmd, pte_t orig_pte, int write_access)
1378 {
1379         struct page *page;
1380         swp_entry_t entry = pte_to_swp_entry(orig_pte);
1381         pte_t pte;
1382         int ret = VM_FAULT_MINOR;
1383
1384         pte_unmap(page_table);
1385         spin_unlock(&mm->page_table_lock);
1386         page = lookup_swap_cache(entry);
1387         if (!page) {
1388                 swapin_readahead(entry, address, vma);
1389                 page = read_swap_cache_async(entry, vma, address);
1390                 if (!page) {
1391                         /*
1392                          * Back out if somebody else faulted in this pte while
1393                          * we released the page table lock.
1394                          */
1395                         spin_lock(&mm->page_table_lock);
1396                         page_table = pte_offset_map(pmd, address);
1397                         if (likely(pte_same(*page_table, orig_pte)))
1398                                 ret = VM_FAULT_OOM;
1399                         else
1400                                 ret = VM_FAULT_MINOR;
1401                         pte_unmap(page_table);
1402                         spin_unlock(&mm->page_table_lock);
1403                         goto out;
1404                 }
1405
1406                 /* Had to read the page from swap area: Major fault */
1407                 ret = VM_FAULT_MAJOR;
1408                 inc_page_state(pgmajfault);
1409         }
1410
1411         if (!vx_rsspages_avail(mm, 1)) {
1412                 ret = VM_FAULT_OOM;
1413                 goto out;
1414         }
1415         mark_page_accessed(page);
1416         lock_page(page);
1417
1418         /*
1419          * Back out if somebody else faulted in this pte while we
1420          * released the page table lock.
1421          */
1422         spin_lock(&mm->page_table_lock);
1423         page_table = pte_offset_map(pmd, address);
1424         if (unlikely(!pte_same(*page_table, orig_pte))) {
1425                 pte_unmap(page_table);
1426                 spin_unlock(&mm->page_table_lock);
1427                 unlock_page(page);
1428                 page_cache_release(page);
1429                 ret = VM_FAULT_MINOR;
1430                 goto out;
1431         }
1432
1433         /* The page isn't present yet, go ahead with the fault. */
1434
1435         swap_free(entry);
1436         if (vm_swap_full())
1437                 remove_exclusive_swap_page(page);
1438
1439         // mm->rss++;
1440         vx_rsspages_inc(mm);
1441         pte = mk_pte(page, vma->vm_page_prot);
1442         if (write_access && can_share_swap_page(page)) {
1443                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
1444                 write_access = 0;
1445         }
1446         unlock_page(page);
1447
1448         flush_icache_page(vma, page);
1449         set_pte(page_table, pte);
1450         page_add_anon_rmap(page, vma, address);
1451
1452         if (write_access) {
1453                 if (do_wp_page(mm, vma, address,
1454                                 page_table, pmd, pte) == VM_FAULT_OOM)
1455                         ret = VM_FAULT_OOM;
1456                 goto out;
1457         }
1458
1459         /* No need to invalidate - it was non-present before */
1460         update_mmu_cache(vma, address, pte);
1461         pte_unmap(page_table);
1462         spin_unlock(&mm->page_table_lock);
1463 out:
1464         return ret;
1465 }
1466
1467 /*
1468  * We are called with the MM semaphore and page_table_lock
1469  * spinlock held to protect against concurrent faults in
1470  * multithreaded programs.
1471  */
1472 static int
1473 do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
1474                 pte_t *page_table, pmd_t *pmd, int write_access,
1475                 unsigned long addr)
1476 {
1477         pte_t entry;
1478         struct page * page = ZERO_PAGE(addr);
1479
1480         if (!vx_rsspages_avail(mm, 1)) {
1481                 spin_unlock(&mm->page_table_lock);
1482                 return VM_FAULT_OOM;
1483         }
1484
1485         /* Read-only mapping of ZERO_PAGE. */
1486         entry = pte_wrprotect(mk_pte(ZERO_PAGE(addr), vma->vm_page_prot));
1487
1488         /* ..except if it's a write access */
1489         if (write_access) {
1490                 /* Allocate our own private page. */
1491                 pte_unmap(page_table);
1492                 spin_unlock(&mm->page_table_lock);
1493
1494                 if (unlikely(anon_vma_prepare(vma)))
1495                         goto no_mem;
1496                 page = alloc_page_vma(GFP_HIGHUSER, vma, addr);
1497                 if (!page)
1498                         goto no_mem;
1499                 clear_user_highpage(page, addr);
1500
1501                 spin_lock(&mm->page_table_lock);
1502                 page_table = pte_offset_map(pmd, addr);
1503
1504                 if (!pte_none(*page_table)) {
1505                         pte_unmap(page_table);
1506                         page_cache_release(page);
1507                         spin_unlock(&mm->page_table_lock);
1508                         goto out;
1509                 }
1510                 // mm->rss++;
1511                 vx_rsspages_inc(mm);
1512                 entry = maybe_mkwrite(pte_mkdirty(mk_pte(page,
1513                                                          vma->vm_page_prot)),
1514                                       vma);
1515                 lru_cache_add_active(page);
1516                 mark_page_accessed(page);
1517                 page_add_anon_rmap(page, vma, addr);
1518         }
1519
1520         set_pte(page_table, entry);
1521         pte_unmap(page_table);
1522
1523         /* No need to invalidate - it was non-present before */
1524         update_mmu_cache(vma, addr, entry);
1525         spin_unlock(&mm->page_table_lock);
1526 out:
1527         return VM_FAULT_MINOR;
1528 no_mem:
1529         return VM_FAULT_OOM;
1530 }
1531
1532 /*
1533  * do_no_page() tries to create a new page mapping. It aggressively
1534  * tries to share with existing pages, but makes a separate copy if
1535  * the "write_access" parameter is true in order to avoid the next
1536  * page fault.
1537  *
1538  * As this is called only for pages that do not currently exist, we
1539  * do not need to flush old virtual caches or the TLB.
1540  *
1541  * This is called with the MM semaphore held and the page table
1542  * spinlock held. Exit with the spinlock released.
1543  */
1544 static int
1545 do_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
1546         unsigned long address, int write_access, pte_t *page_table, pmd_t *pmd)
1547 {
1548         struct page * new_page;
1549         struct address_space *mapping = NULL;
1550         pte_t entry;
1551         int sequence = 0;
1552         int ret = VM_FAULT_MINOR;
1553         int anon = 0;
1554
1555         if (!vma->vm_ops || !vma->vm_ops->nopage)
1556                 return do_anonymous_page(mm, vma, page_table,
1557                                         pmd, write_access, address);
1558         pte_unmap(page_table);
1559         spin_unlock(&mm->page_table_lock);
1560
1561         if (vma->vm_file) {
1562                 mapping = vma->vm_file->f_mapping;
1563                 sequence = atomic_read(&mapping->truncate_count);
1564         }
1565         smp_rmb();  /* Prevent CPU from reordering lock-free ->nopage() */
1566 retry:
1567         new_page = vma->vm_ops->nopage(vma, address & PAGE_MASK, &ret);
1568
1569         /* no page was available -- either SIGBUS or OOM */
1570         if (new_page == NOPAGE_SIGBUS)
1571                 return VM_FAULT_SIGBUS;
1572         if (new_page == NOPAGE_OOM)
1573                 return VM_FAULT_OOM;
1574         if (!vx_rsspages_avail(mm, 1))
1575                 return VM_FAULT_OOM;
1576
1577         /*
1578          * Should we do an early C-O-W break?
1579          */
1580         if (write_access && !(vma->vm_flags & VM_SHARED)) {
1581                 struct page *page;
1582
1583                 if (unlikely(anon_vma_prepare(vma)))
1584                         goto oom;
1585                 page = alloc_page_vma(GFP_HIGHUSER, vma, address);
1586                 if (!page)
1587                         goto oom;
1588                 copy_user_highpage(page, new_page, address);
1589                 page_cache_release(new_page);
1590                 new_page = page;
1591                 anon = 1;
1592         }
1593
1594         spin_lock(&mm->page_table_lock);
1595         /*
1596          * For a file-backed vma, someone could have truncated or otherwise
1597          * invalidated this page.  If unmap_mapping_range got called,
1598          * retry getting the page.
1599          */
1600         if (mapping &&
1601               (unlikely(sequence != atomic_read(&mapping->truncate_count)))) {
1602                 sequence = atomic_read(&mapping->truncate_count);
1603                 spin_unlock(&mm->page_table_lock);
1604                 page_cache_release(new_page);
1605                 goto retry;
1606         }
1607         page_table = pte_offset_map(pmd, address);
1608
1609         /*
1610          * This silly early PAGE_DIRTY setting removes a race
1611          * due to the bad i386 page protection. But it's valid
1612          * for other architectures too.
1613          *
1614          * Note that if write_access is true, we either now have
1615          * an exclusive copy of the page, or this is a shared mapping,
1616          * so we can make it writable and dirty to avoid having to
1617          * handle that later.
1618          */
1619         /* Only go through if we didn't race with anybody else... */
1620         if (pte_none(*page_table)) {
1621                 if (!PageReserved(new_page))
1622                         // ++mm->rss;
1623                         vx_rsspages_inc(mm);
1624                 flush_icache_page(vma, new_page);
1625                 entry = mk_pte(new_page, vma->vm_page_prot);
1626                 if (write_access)
1627                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1628                 set_pte(page_table, entry);
1629                 if (anon) {
1630                         lru_cache_add_active(new_page);
1631                         page_add_anon_rmap(new_page, vma, address);
1632                 } else
1633                         page_add_file_rmap(new_page);
1634                 pte_unmap(page_table);
1635         } else {
1636                 /* One of our sibling threads was faster, back out. */
1637                 pte_unmap(page_table);
1638                 page_cache_release(new_page);
1639                 spin_unlock(&mm->page_table_lock);
1640                 goto out;
1641         }
1642
1643         /* no need to invalidate: a not-present page shouldn't be cached */
1644         update_mmu_cache(vma, address, entry);
1645         spin_unlock(&mm->page_table_lock);
1646 out:
1647         return ret;
1648 oom:
1649         page_cache_release(new_page);
1650         ret = VM_FAULT_OOM;
1651         goto out;
1652 }
1653
1654 /*
1655  * Fault of a previously existing named mapping. Repopulate the pte
1656  * from the encoded file_pte if possible. This enables swappable
1657  * nonlinear vmas.
1658  */
1659 static int do_file_page(struct mm_struct * mm, struct vm_area_struct * vma,
1660         unsigned long address, int write_access, pte_t *pte, pmd_t *pmd)
1661 {
1662         unsigned long pgoff;
1663         int err;
1664
1665         BUG_ON(!vma->vm_ops || !vma->vm_ops->nopage);
1666         /*
1667          * Fall back to the linear mapping if the fs does not support
1668          * ->populate:
1669          */
1670         if (!vma->vm_ops || !vma->vm_ops->populate ||
1671                         (write_access && !(vma->vm_flags & VM_SHARED))) {
1672                 pte_clear(pte);
1673                 return do_no_page(mm, vma, address, write_access, pte, pmd);
1674         }
1675
1676         pgoff = pte_to_pgoff(*pte);
1677
1678         pte_unmap(pte);
1679         spin_unlock(&mm->page_table_lock);
1680
1681         err = vma->vm_ops->populate(vma, address & PAGE_MASK, PAGE_SIZE, vma->vm_page_prot, pgoff, 0);
1682         if (err == -ENOMEM)
1683                 return VM_FAULT_OOM;
1684         if (err)
1685                 return VM_FAULT_SIGBUS;
1686         return VM_FAULT_MAJOR;
1687 }
1688
1689 /*
1690  * These routines also need to handle stuff like marking pages dirty
1691  * and/or accessed for architectures that don't do it in hardware (most
1692  * RISC architectures).  The early dirtying is also good on the i386.
1693  *
1694  * There is also a hook called "update_mmu_cache()" that architectures
1695  * with external mmu caches can use to update those (ie the Sparc or
1696  * PowerPC hashed page tables that act as extended TLBs).
1697  *
1698  * Note the "page_table_lock". It is to protect against kswapd removing
1699  * pages from under us. Note that kswapd only ever _removes_ pages, never
1700  * adds them. As such, once we have noticed that the page is not present,
1701  * we can drop the lock early.
1702  *
1703  * The adding of pages is protected by the MM semaphore (which we hold),
1704  * so we don't need to worry about a page being suddenly been added into
1705  * our VM.
1706  *
1707  * We enter with the pagetable spinlock held, we are supposed to
1708  * release it when done.
1709  */
1710 static inline int handle_pte_fault(struct mm_struct *mm,
1711         struct vm_area_struct * vma, unsigned long address,
1712         int write_access, pte_t *pte, pmd_t *pmd)
1713 {
1714         pte_t entry;
1715
1716         entry = *pte;
1717         if (!pte_present(entry)) {
1718                 /*
1719                  * If it truly wasn't present, we know that kswapd
1720                  * and the PTE updates will not touch it later. So
1721                  * drop the lock.
1722                  */
1723                 if (pte_none(entry))
1724                         return do_no_page(mm, vma, address, write_access, pte, pmd);
1725                 if (pte_file(entry))
1726                         return do_file_page(mm, vma, address, write_access, pte, pmd);
1727                 return do_swap_page(mm, vma, address, pte, pmd, entry, write_access);
1728         }
1729
1730         if (write_access) {
1731                 if (!pte_write(entry))
1732                         return do_wp_page(mm, vma, address, pte, pmd, entry);
1733
1734                 entry = pte_mkdirty(entry);
1735         }
1736         entry = pte_mkyoung(entry);
1737         ptep_set_access_flags(vma, address, pte, entry, write_access);
1738         update_mmu_cache(vma, address, entry);
1739         pte_unmap(pte);
1740         spin_unlock(&mm->page_table_lock);
1741         return VM_FAULT_MINOR;
1742 }
1743
1744 /*
1745  * By the time we get here, we already hold the mm semaphore
1746  */
1747 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct * vma,
1748         unsigned long address, int write_access)
1749 {
1750         pgd_t *pgd;
1751         pmd_t *pmd;
1752
1753         __set_current_state(TASK_RUNNING);
1754         pgd = pgd_offset(mm, address);
1755
1756         inc_page_state(pgfault);
1757
1758         if (is_vm_hugetlb_page(vma))
1759                 return VM_FAULT_SIGBUS; /* mapping truncation does this. */
1760
1761         /*
1762          * We need the page table lock to synchronize with kswapd
1763          * and the SMP-safe atomic PTE updates.
1764          */
1765         set_delay_flag(current,PF_MEMIO);
1766         spin_lock(&mm->page_table_lock);
1767         pmd = pmd_alloc(mm, pgd, address);
1768
1769         if (pmd) {
1770                 pte_t * pte = pte_alloc_map(mm, pmd, address);
1771                 if (pte) {
1772                         int rc = handle_pte_fault(mm, vma, address, write_access, pte, pmd);
1773                         clear_delay_flag(current,PF_MEMIO);
1774                         return rc;
1775                 }
1776         }
1777         spin_unlock(&mm->page_table_lock);
1778         clear_delay_flag(current,PF_MEMIO);
1779         return VM_FAULT_OOM;
1780 }
1781
1782 /*
1783  * Allocate page middle directory.
1784  *
1785  * We've already handled the fast-path in-line, and we own the
1786  * page table lock.
1787  *
1788  * On a two-level page table, this ends up actually being entirely
1789  * optimized away.
1790  */
1791 pmd_t fastcall *__pmd_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
1792 {
1793         pmd_t *new;
1794
1795         spin_unlock(&mm->page_table_lock);
1796         new = pmd_alloc_one(mm, address);
1797         spin_lock(&mm->page_table_lock);
1798         if (!new)
1799                 return NULL;
1800
1801         /*
1802          * Because we dropped the lock, we should re-check the
1803          * entry, as somebody else could have populated it..
1804          */
1805         if (pgd_present(*pgd)) {
1806                 pmd_free(new);
1807                 goto out;
1808         }
1809         pgd_populate(mm, pgd, new);
1810 out:
1811         return pmd_offset(pgd, address);
1812 }
1813
1814 int make_pages_present(unsigned long addr, unsigned long end)
1815 {
1816         int ret, len, write;
1817         struct vm_area_struct * vma;
1818
1819         vma = find_vma(current->mm, addr);
1820         write = (vma->vm_flags & VM_WRITE) != 0;
1821         if (addr >= end)
1822                 BUG();
1823         if (end > vma->vm_end)
1824                 BUG();
1825         len = (end+PAGE_SIZE-1)/PAGE_SIZE-addr/PAGE_SIZE;
1826         ret = get_user_pages(current, current->mm, addr,
1827                         len, write, 0, NULL, NULL);
1828         if (ret < 0)
1829                 return ret;
1830         return ret == len ? 0 : -1;
1831 }
1832
1833 /*
1834  * Map a vmalloc()-space virtual address to the physical page.
1835  */
1836 struct page * vmalloc_to_page(void * vmalloc_addr)
1837 {
1838         unsigned long addr = (unsigned long) vmalloc_addr;
1839         struct page *page = NULL;
1840         pgd_t *pgd = pgd_offset_k(addr);
1841         pmd_t *pmd;
1842         pte_t *ptep, pte;
1843
1844         if (!pgd_none(*pgd)) {
1845                 pmd = pmd_offset(pgd, addr);
1846                 if (!pmd_none(*pmd)) {
1847                         preempt_disable();
1848                         ptep = pte_offset_map(pmd, addr);
1849                         pte = *ptep;
1850                         if (pte_present(pte))
1851                                 page = pte_page(pte);
1852                         pte_unmap(ptep);
1853                         preempt_enable();
1854                 }
1855         }
1856         return page;
1857 }
1858
1859 EXPORT_SYMBOL(vmalloc_to_page);
1860
1861 #if !defined(CONFIG_ARCH_GATE_AREA)
1862
1863 #if defined(AT_SYSINFO_EHDR)
1864 struct vm_area_struct gate_vma;
1865
1866 static int __init gate_vma_init(void)
1867 {
1868         gate_vma.vm_mm = NULL;
1869         gate_vma.vm_start = FIXADDR_USER_START;
1870         gate_vma.vm_end = FIXADDR_USER_END;
1871         gate_vma.vm_page_prot = PAGE_READONLY;
1872         gate_vma.vm_flags = 0;
1873         return 0;
1874 }
1875 __initcall(gate_vma_init);
1876 #endif
1877
1878 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
1879 {
1880 #ifdef AT_SYSINFO_EHDR
1881         return &gate_vma;
1882 #else
1883         return 0;
1884 #endif
1885 }
1886
1887 int in_gate_area(struct task_struct *task, unsigned long addr)
1888 {
1889 #ifdef AT_SYSINFO_EHDR
1890         if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
1891                 return 1;
1892 #endif
1893         return 0;
1894 }
1895
1896 #endif