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                 if (write && !pte_dirty(pte)) {
 653                         struct page *page = pte_page(pte);
 654                         if (!PageDirty(page))
 655                                 set_page_dirty(page);
 656                 }
 657                 pfn = pte_pfn(pte);
 658                 if (pfn_valid(pfn)) {
 659                         struct page *page = pfn_to_page(pfn);
 660
 661                         mark_page_accessed(page);
 662                         return page;
 663                 }
 664         }
 665
 666 out:
 667         return NULL;
 668 }
 669
 670 /*
 671  * Given a physical address, is there a useful struct page pointing to
 672  * it?  This may become more complex in the future if we start dealing
 673  * with IO-aperture pages for direct-IO.
 674  */
 675
 676 static inline struct page *get_page_map(struct page *page)
 677 {
 678         if (!pfn_valid(page_to_pfn(page)))
 679                 return 0;
 680         return page;
 681 }
 682
 683
 684 static inline int
 685 untouched_anonymous_page(struct mm_struct* mm, struct vm_area_struct *vma,
 686                          unsigned long address)
 687 {
 688         pgd_t *pgd;
 689         pmd_t *pmd;
 690
 691         /* Check if the vma is for an anonymous mapping. */
 692         if (vma->vm_ops && vma->vm_ops->nopage)
 693                 return 0;
 694
 695         /* Check if page directory entry exists. */
 696         pgd = pgd_offset(mm, address);
 697         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
 698                 return 1;
 699
 700         /* Check if page middle directory entry exists. */
 701         pmd = pmd_offset(pgd, address);
 702         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
 703                 return 1;
 704
 705         /* There is a pte slot for 'address' in 'mm'. */
 706         return 0;
 707 }
 708
 709
 710 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
 711                 unsigned long start, int len, int write, int force,
 712                 struct page **pages, struct vm_area_struct **vmas)
 713 {
 714         int i;
 715         unsigned int flags;
 716
 717         /*
 718          * Require read or write permissions.
 719          * If 'force' is set, we only require the "MAY" flags.
 720          */
 721         flags = write ? (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
 722         flags &= force ? (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
 723         i = 0;
 724
 725         do {
 726                 struct vm_area_struct * vma;
 727
 728                 vma = find_extend_vma(mm, start);
 729                 if (!vma && in_gate_area(tsk, start)) {
 730                         unsigned long pg = start & PAGE_MASK;
 731                         struct vm_area_struct *gate_vma = get_gate_vma(tsk);
 732                         pgd_t *pgd;
 733                         pmd_t *pmd;
 734                         pte_t *pte;
 735                         if (write) /* user gate pages are read-only */
 736                                 return i ? : -EFAULT;
 737                         pgd = pgd_offset_k(pg);
 738                         if (!pgd)
 739                                 return i ? : -EFAULT;
 740                         pmd = pmd_offset(pgd, pg);
 741                         if (!pmd)
 742                                 return i ? : -EFAULT;
 743                         pte = pte_offset_kernel(pmd, pg);
 744                         if (!pte || !pte_present(*pte))
 745                                 return i ? : -EFAULT;
 746                         if (pages) {
 747                                 pages[i] = pte_page(*pte);
 748                                 get_page(pages[i]);
 749                         }
 750                         if (vmas)
 751                                 vmas[i] = gate_vma;
 752                         i++;
 753                         start += PAGE_SIZE;
 754                         len--;
 755                         continue;
 756                 }
 757
 758                 if (!vma || (pages && (vma->vm_flags & VM_IO))
 759                                 || !(flags & vma->vm_flags))
 760                         return i ? : -EFAULT;
 761
 762                 if (is_vm_hugetlb_page(vma)) {
 763                         i = follow_hugetlb_page(mm, vma, pages, vmas,
 764                                                 &start, &len, i);
 765                         continue;
 766                 }
 767                 spin_lock(&mm->page_table_lock);
 768                 do {
 769                         struct page *map;
 770                         int lookup_write = write;
 771                         while (!(map = follow_page(mm, start, lookup_write))) {
 772                                 /*
 773                                  * Shortcut for anonymous pages. We don't want
 774                                  * to force the creation of pages tables for
 775                                  * insanly big anonymously mapped areas that
 776                                  * nobody touched so far. This is important
 777                                  * for doing a core dump for these mappings.
 778                                  */
 779                                 if (!lookup_write &&
 780                                     untouched_anonymous_page(mm,vma,start)) {
 781                                         map = ZERO_PAGE(start);
 782                                         break;
 783                                 }
 784                                 spin_unlock(&mm->page_table_lock);
 785                                 switch (handle_mm_fault(mm,vma,start,write)) {
 786                                 case VM_FAULT_MINOR:
 787                                         tsk->min_flt++;
 788                                         break;
 789                                 case VM_FAULT_MAJOR:
 790                                         tsk->maj_flt++;
 791                                         break;
 792                                 case VM_FAULT_SIGBUS:
 793                                         return i ? i : -EFAULT;
 794                                 case VM_FAULT_OOM:
 795                                         return i ? i : -ENOMEM;
 796                                 default:
 797                                         BUG();
 798                                 }
 799                                 /*
 800                                  * Now that we have performed a write fault
 801                                  * and surely no longer have a shared page we
 802                                  * shouldn't write, we shouldn't ignore an
 803                                  * unwritable page in the page table if
 804                                  * we are forcing write access.
 805                                  */
 806                                 lookup_write = write && !force;
 807                                 spin_lock(&mm->page_table_lock);
 808                         }
 809                         if (pages) {
 810                                 pages[i] = get_page_map(map);
 811                                 if (!pages[i]) {
 812                                         spin_unlock(&mm->page_table_lock);
 813                                         while (i--)
 814                                                 page_cache_release(pages[i]);
 815                                         i = -EFAULT;
 816                                         goto out;
 817                                 }
 818                                 flush_dcache_page(pages[i]);
 819                                 if (!PageReserved(pages[i]))
 820                                         page_cache_get(pages[i]);
 821                         }
 822                         if (vmas)
 823                                 vmas[i] = vma;
 824                         i++;
 825                         start += PAGE_SIZE;
 826                         len--;
 827                 } while(len && start < vma->vm_end);
 828                 spin_unlock(&mm->page_table_lock);
 829         } while(len);
 830 out:
 831         return i;
 832 }
 833
 834 EXPORT_SYMBOL(get_user_pages);
 835
 836 static void zeromap_pte_range(pte_t * pte, unsigned long address,
 837                                      unsigned long size, pgprot_t prot)
 838 {
 839         unsigned long end;
 840
 841         address &= ~PMD_MASK;
 842         end = address + size;
 843         if (end > PMD_SIZE)
 844                 end = PMD_SIZE;
 845         do {
 846                 pte_t zero_pte = pte_wrprotect(mk_pte(ZERO_PAGE(address), prot));
 847                 BUG_ON(!pte_none(*pte));
 848                 set_pte(pte, zero_pte);
 849                 address += PAGE_SIZE;
 850                 pte++;
 851         } while (address && (address < end));
 852 }
 853
 854 static inline int zeromap_pmd_range(struct mm_struct *mm, pmd_t * pmd, unsigned long address,
 855                                     unsigned long size, pgprot_t prot)
 856 {
 857         unsigned long base, end;
 858
 859         base = address & PGDIR_MASK;
 860         address &= ~PGDIR_MASK;
 861         end = address + size;
 862         if (end > PGDIR_SIZE)
 863                 end = PGDIR_SIZE;
 864         do {
 865                 pte_t * pte = pte_alloc_map(mm, pmd, base + address);
 866                 if (!pte)
 867                         return -ENOMEM;
 868                 zeromap_pte_range(pte, base + address, end - address, prot);
 869                 pte_unmap(pte);
 870                 address = (address + PMD_SIZE) & PMD_MASK;
 871                 pmd++;
 872         } while (address && (address < end));
 873         return 0;
 874 }
 875
 876 int zeromap_page_range(struct vm_area_struct *vma, unsigned long address, unsigned long size, pgprot_t prot)
 877 {
 878         int error = 0;
 879         pgd_t * dir;
 880         unsigned long beg = address;
 881         unsigned long end = address + size;
 882         struct mm_struct *mm = vma->vm_mm;
 883
 884         dir = pgd_offset(mm, address);
 885         flush_cache_range(vma, beg, end);
 886         if (address >= end)
 887                 BUG();
 888
 889         spin_lock(&mm->page_table_lock);
 890         do {
 891                 pmd_t *pmd = pmd_alloc(mm, dir, address);
 892                 error = -ENOMEM;
 893                 if (!pmd)
 894                         break;
 895                 error = zeromap_pmd_range(mm, pmd, address, end - address, prot);
 896                 if (error)
 897                         break;
 898                 address = (address + PGDIR_SIZE) & PGDIR_MASK;
 899                 dir++;
 900         } while (address && (address < end));
 901         /*
 902          * Why flush? zeromap_pte_range has a BUG_ON for !pte_none()
 903          */
 904         flush_tlb_range(vma, beg, end);
 905         spin_unlock(&mm->page_table_lock);
 906         return error;
 907 }
 908
 909 /*
 910  * maps a range of physical memory into the requested pages. the old
 911  * mappings are removed. any references to nonexistent pages results
 912  * in null mappings (currently treated as "copy-on-access")
 913  */
 914 static inline void remap_pte_range(pte_t * pte, unsigned long address, unsigned long size,
 915         unsigned long phys_addr, pgprot_t prot)
 916 {
 917         unsigned long end;
 918         unsigned long pfn;
 919
 920         address &= ~PMD_MASK;
 921         end = address + size;
 922         if (end > PMD_SIZE)
 923                 end = PMD_SIZE;
 924         pfn = phys_addr >> PAGE_SHIFT;
 925         do {
 926                 BUG_ON(!pte_none(*pte));
 927                 if (!pfn_valid(pfn) || PageReserved(pfn_to_page(pfn)))
 928                         set_pte(pte, pfn_pte(pfn, prot));
 929                 address += PAGE_SIZE;
 930                 pfn++;
 931                 pte++;
 932         } while (address && (address < end));
 933 }
 934
 935 static inline int remap_pmd_range(struct mm_struct *mm, pmd_t * pmd, unsigned long address, unsigned long size,
 936         unsigned long phys_addr, pgprot_t prot)
 937 {
 938         unsigned long base, end;
 939
 940         base = address & PGDIR_MASK;
 941         address &= ~PGDIR_MASK;
 942         end = address + size;
 943         if (end > PGDIR_SIZE)
 944                 end = PGDIR_SIZE;
 945         phys_addr -= address;
 946         do {
 947                 pte_t * pte = pte_alloc_map(mm, pmd, base + address);
 948                 if (!pte)
 949                         return -ENOMEM;
 950                 remap_pte_range(pte, base + address, end - address, address + phys_addr, prot);
 951                 pte_unmap(pte);
 952                 address = (address + PMD_SIZE) & PMD_MASK;
 953                 pmd++;
 954         } while (address && (address < end));
 955         return 0;
 956 }
 957
 958 /*  Note: this is only safe if the mm semaphore is held when called. */
 959 int remap_page_range(struct vm_area_struct *vma, unsigned long from, unsigned long phys_addr, unsigned long size, pgprot_t prot)
 960 {
 961         int error = 0;
 962         pgd_t * dir;
 963         unsigned long beg = from;
 964         unsigned long end = from + size;
 965         struct mm_struct *mm = vma->vm_mm;
 966
 967         phys_addr -= from;
 968         dir = pgd_offset(mm, from);
 969         flush_cache_range(vma, beg, end);
 970         if (from >= end)
 971                 BUG();
 972
 973         spin_lock(&mm->page_table_lock);
 974         do {
 975                 pmd_t *pmd = pmd_alloc(mm, dir, from);
 976                 error = -ENOMEM;
 977                 if (!pmd)
 978                         break;
 979                 error = remap_pmd_range(mm, pmd, from, end - from, phys_addr + from, prot);
 980                 if (error)
 981                         break;
 982                 from = (from + PGDIR_SIZE) & PGDIR_MASK;
 983                 dir++;
 984         } while (from && (from < end));
 985         /*
 986          * Why flush? remap_pte_range has a BUG_ON for !pte_none()
 987          */
 988         flush_tlb_range(vma, beg, end);
 989         spin_unlock(&mm->page_table_lock);
 990         return error;
 991 }
 992
 993 EXPORT_SYMBOL(remap_page_range);
 994
 995 /*
 996  * Do pte_mkwrite, but only if the vma says VM_WRITE.  We do this when
 997  * servicing faults for write access.  In the normal case, do always want
 998  * pte_mkwrite.  But get_user_pages can cause write faults for mappings
 999  * that do not have writing enabled, when used by access_process_vm.
1000  */
1001 static inline pte_t maybe_mkwrite(pte_t pte, struct vm_area_struct *vma)
1002 {
1003         if (likely(vma->vm_flags & VM_WRITE))
1004                 pte = pte_mkwrite(pte);
1005         return pte;
1006 }
1007
1008 /*
1009  * We hold the mm semaphore for reading and vma->vm_mm->page_table_lock
1010  */
1011 static inline void break_cow(struct vm_area_struct * vma, struct page * new_page, unsigned long address,
1012                 pte_t *page_table)
1013 {
1014         pte_t entry;
1015
1016         flush_cache_page(vma, address);
1017         entry = maybe_mkwrite(pte_mkdirty(mk_pte(new_page, vma->vm_page_prot)),
1018                               vma);
1019         ptep_establish(vma, address, page_table, entry);
1020         update_mmu_cache(vma, address, entry);
1021 }
1022
1023 /*
1024  * This routine handles present pages, when users try to write
1025  * to a shared page. It is done by copying the page to a new address
1026  * and decrementing the shared-page counter for the old page.
1027  *
1028  * Goto-purists beware: the only reason for goto's here is that it results
1029  * in better assembly code.. The "default" path will see no jumps at all.
1030  *
1031  * Note that this routine assumes that the protection checks have been
1032  * done by the caller (the low-level page fault routine in most cases).
1033  * Thus we can safely just mark it writable once we've done any necessary
1034  * COW.
1035  *
1036  * We also mark the page dirty at this point even though the page will
1037  * change only once the write actually happens. This avoids a few races,
1038  * and potentially makes it more efficient.
1039  *
1040  * We hold the mm semaphore and the page_table_lock on entry and exit
1041  * with the page_table_lock released.
1042  */
1043 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct * vma,
1044         unsigned long address, pte_t *page_table, pmd_t *pmd, pte_t pte)
1045 {
1046         struct page *old_page, *new_page;
1047         unsigned long pfn = pte_pfn(pte);
1048         pte_t entry;
1049
1050         if (unlikely(!pfn_valid(pfn))) {
1051                 /*
1052                  * This should really halt the system so it can be debugged or
1053                  * at least the kernel stops what it's doing before it corrupts
1054                  * data, but for the moment just pretend this is OOM.
1055                  */
1056                 pte_unmap(page_table);
1057                 printk(KERN_ERR "do_wp_page: bogus page at address %08lx\n",
1058                                 address);
1059                 spin_unlock(&mm->page_table_lock);
1060                 return VM_FAULT_OOM;
1061         }
1062         old_page = pfn_to_page(pfn);
1063
1064         if (!TestSetPageLocked(old_page)) {
1065                 int reuse = can_share_swap_page(old_page);
1066                 unlock_page(old_page);
1067                 if (reuse) {
1068                         flush_cache_page(vma, address);
1069                         entry = maybe_mkwrite(pte_mkyoung(pte_mkdirty(pte)),
1070                                               vma);
1071                         ptep_set_access_flags(vma, address, page_table, entry, 1);
1072                         update_mmu_cache(vma, address, entry);
1073                         pte_unmap(page_table);
1074                         spin_unlock(&mm->page_table_lock);
1075                         return VM_FAULT_MINOR;
1076                 }
1077         }
1078         pte_unmap(page_table);
1079
1080         /*
1081          * Ok, we need to copy. Oh, well..
1082          */
1083         page_cache_get(old_page);
1084         spin_unlock(&mm->page_table_lock);
1085
1086         if (unlikely(anon_vma_prepare(vma)))
1087                 goto no_new_page;
1088         new_page = alloc_page_vma(GFP_HIGHUSER, vma, address);
1089         if (!new_page)
1090                 goto no_new_page;
1091         copy_cow_page(old_page,new_page,address);
1092
1093         /*
1094          * Re-check the pte - we dropped the lock
1095          */
1096         spin_lock(&mm->page_table_lock);
1097         page_table = pte_offset_map(pmd, address);
1098         if (likely(pte_same(*page_table, pte))) {
1099                 if (PageReserved(old_page))
1100                         // ++mm->rss;
1101                         vx_rsspages_inc(mm);
1102                 else
1103                         page_remove_rmap(old_page);
1104                 break_cow(vma, new_page, address, page_table);
1105                 lru_cache_add_active(new_page);
1106                 page_add_anon_rmap(new_page, vma, address);
1107
1108                 /* Free the old page.. */
1109                 new_page = old_page;
1110         }
1111         pte_unmap(page_table);
1112         page_cache_release(new_page);
1113         page_cache_release(old_page);
1114         spin_unlock(&mm->page_table_lock);
1115         return VM_FAULT_MINOR;
1116
1117 no_new_page:
1118         page_cache_release(old_page);
1119         return VM_FAULT_OOM;
1120 }
1121
1122 /*
1123  * Helper function for unmap_mapping_range().
1124  */
1125 static inline void unmap_mapping_range_list(struct prio_tree_root *root,
1126                                             struct zap_details *details)
1127 {
1128         struct vm_area_struct *vma = NULL;
1129         struct prio_tree_iter iter;
1130         pgoff_t vba, vea, zba, zea;
1131
1132         while ((vma = vma_prio_tree_next(vma, root, &iter,
1133                         details->first_index, details->last_index)) != NULL) {
1134                 vba = vma->vm_pgoff;
1135                 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
1136                 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
1137                 zba = details->first_index;
1138                 if (zba < vba)
1139                         zba = vba;
1140                 zea = details->last_index;
1141                 if (zea > vea)
1142                         zea = vea;
1143                 zap_page_range(vma,
1144                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
1145                         (zea - zba + 1) << PAGE_SHIFT, details);
1146         }
1147 }
1148
1149 /**
1150  * unmap_mapping_range - unmap the portion of all mmaps
1151  * in the specified address_space corresponding to the specified
1152  * page range in the underlying file.
1153  * @address_space: the address space containing mmaps to be unmapped.
1154  * @holebegin: byte in first page to unmap, relative to the start of
1155  * the underlying file.  This will be rounded down to a PAGE_SIZE
1156  * boundary.  Note that this is different from vmtruncate(), which
1157  * must keep the partial page.  In contrast, we must get rid of
1158  * partial pages.
1159  * @holelen: size of prospective hole in bytes.  This will be rounded
1160  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
1161  * end of the file.
1162  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
1163  * but 0 when invalidating pagecache, don't throw away private data.
1164  */
1165 void unmap_mapping_range(struct address_space *mapping,
1166                 loff_t const holebegin, loff_t const holelen, int even_cows)
1167 {
1168         struct zap_details details;
1169         pgoff_t hba = holebegin >> PAGE_SHIFT;
1170         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1171
1172         /* Check for overflow. */
1173         if (sizeof(holelen) > sizeof(hlen)) {
1174                 long long holeend =
1175                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1176                 if (holeend & ~(long long)ULONG_MAX)
1177                         hlen = ULONG_MAX - hba + 1;
1178         }
1179
1180         details.check_mapping = even_cows? NULL: mapping;
1181         details.nonlinear_vma = NULL;
1182         details.first_index = hba;
1183         details.last_index = hba + hlen - 1;
1184         details.atomic = 1;     /* A spinlock is held */
1185         if (details.last_index < details.first_index)
1186                 details.last_index = ULONG_MAX;
1187
1188         spin_lock(&mapping->i_mmap_lock);
1189         /* Protect against page fault */
1190         atomic_inc(&mapping->truncate_count);
1191
1192         if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
1193                 unmap_mapping_range_list(&mapping->i_mmap, &details);
1194
1195         /*
1196          * In nonlinear VMAs there is no correspondence between virtual address
1197          * offset and file offset.  So we must perform an exhaustive search
1198          * across *all* the pages in each nonlinear VMA, not just the pages
1199          * whose virtual address lies outside the file truncation point.
1200          */
1201         if (unlikely(!list_empty(&mapping->i_mmap_nonlinear))) {
1202                 struct vm_area_struct *vma;
1203                 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
1204                                                 shared.vm_set.list) {
1205                         details.nonlinear_vma = vma;
1206                         zap_page_range(vma, vma->vm_start,
1207                                 vma->vm_end - vma->vm_start, &details);
1208                 }
1209         }
1210         spin_unlock(&mapping->i_mmap_lock);
1211 }
1212 EXPORT_SYMBOL(unmap_mapping_range);
1213
1214 /*
1215  * Handle all mappings that got truncated by a "truncate()"
1216  * system call.
1217  *
1218  * NOTE! We have to be ready to update the memory sharing
1219  * between the file and the memory map for a potential last
1220  * incomplete page.  Ugly, but necessary.
1221  */
1222 int vmtruncate(struct inode * inode, loff_t offset)
1223 {
1224         struct address_space *mapping = inode->i_mapping;
1225         unsigned long limit;
1226
1227         if (inode->i_size < offset)
1228                 goto do_expand;
1229         i_size_write(inode, offset);
1230         unmap_mapping_range(mapping, offset + PAGE_SIZE - 1, 0, 1);
1231         truncate_inode_pages(mapping, offset);
1232         goto out_truncate;
1233
1234 do_expand:
1235         limit = current->rlim[RLIMIT_FSIZE].rlim_cur;
1236         if (limit != RLIM_INFINITY && offset > limit)
1237                 goto out_sig;
1238         if (offset > inode->i_sb->s_maxbytes)
1239                 goto out;
1240         i_size_write(inode, offset);
1241
1242 out_truncate:
1243         if (inode->i_op && inode->i_op->truncate)
1244                 inode->i_op->truncate(inode);
1245         return 0;
1246 out_sig:
1247         send_sig(SIGXFSZ, current, 0);
1248 out:
1249         return -EFBIG;
1250 }
1251
1252 EXPORT_SYMBOL(vmtruncate);
1253
1254 /*
1255  * Primitive swap readahead code. We simply read an aligned block of
1256  * (1 << page_cluster) entries in the swap area. This method is chosen
1257  * because it doesn't cost us any seek time.  We also make sure to queue
1258  * the 'original' request together with the readahead ones...
1259  *
1260  * This has been extended to use the NUMA policies from the mm triggering
1261  * the readahead.
1262  *
1263  * Caller must hold down_read on the vma->vm_mm if vma is not NULL.
1264  */
1265 void swapin_readahead(swp_entry_t entry, unsigned long addr,struct vm_area_struct *vma)
1266 {
1267 #ifdef CONFIG_NUMA
1268         struct vm_area_struct *next_vma = vma ? vma->vm_next : NULL;
1269 #endif
1270         int i, num;
1271         struct page *new_page;
1272         unsigned long offset;
1273
1274         /*
1275          * Get the number of handles we should do readahead io to.
1276          */
1277         num = valid_swaphandles(entry, &offset);
1278         for (i = 0; i < num; offset++, i++) {
1279                 /* Ok, do the async read-ahead now */
1280                 new_page = read_swap_cache_async(swp_entry(swp_type(entry),
1281                                                            offset), vma, addr);
1282                 if (!new_page)
1283                         break;
1284                 page_cache_release(new_page);
1285 #ifdef CONFIG_NUMA
1286                 /*
1287                  * Find the next applicable VMA for the NUMA policy.
1288                  */
1289                 addr += PAGE_SIZE;
1290                 if (addr == 0)
1291                         vma = NULL;
1292                 if (vma) {
1293                         if (addr >= vma->vm_end) {
1294                                 vma = next_vma;
1295                                 next_vma = vma ? vma->vm_next : NULL;
1296                         }
1297                         if (vma && addr < vma->vm_start)
1298                                 vma = NULL;
1299                 } else {
1300                         if (next_vma && addr >= next_vma->vm_start) {
1301                                 vma = next_vma;
1302                                 next_vma = vma->vm_next;
1303                         }
1304                 }
1305 #endif
1306         }
1307         lru_add_drain();        /* Push any new pages onto the LRU now */
1308 }
1309
1310 /*
1311  * We hold the mm semaphore and the page_table_lock on entry and
1312  * should release the pagetable lock on exit..
1313  */
1314 static int do_swap_page(struct mm_struct * mm,
1315         struct vm_area_struct * vma, unsigned long address,
1316         pte_t *page_table, pmd_t *pmd, pte_t orig_pte, int write_access)
1317 {
1318         struct page *page;
1319         swp_entry_t entry = pte_to_swp_entry(orig_pte);
1320         pte_t pte;
1321         int ret = VM_FAULT_MINOR;
1322
1323         pte_unmap(page_table);
1324         spin_unlock(&mm->page_table_lock);
1325         page = lookup_swap_cache(entry);
1326         if (!page) {
1327                 swapin_readahead(entry, address, vma);
1328                 page = read_swap_cache_async(entry, vma, address);
1329                 if (!page) {
1330                         /*
1331                          * Back out if somebody else faulted in this pte while
1332                          * we released the page table lock.
1333                          */
1334                         spin_lock(&mm->page_table_lock);
1335                         page_table = pte_offset_map(pmd, address);
1336                         if (likely(pte_same(*page_table, orig_pte)))
1337                                 ret = VM_FAULT_OOM;
1338                         else
1339                                 ret = VM_FAULT_MINOR;
1340                         pte_unmap(page_table);
1341                         spin_unlock(&mm->page_table_lock);
1342                         goto out;
1343                 }
1344
1345                 /* Had to read the page from swap area: Major fault */
1346                 ret = VM_FAULT_MAJOR;
1347                 inc_page_state(pgmajfault);
1348         }
1349
1350         if (!vx_rsspages_avail(mm, 1)) {
1351                 ret = VM_FAULT_OOM;
1352                 goto out;
1353         }
1354         mark_page_accessed(page);
1355         lock_page(page);
1356
1357         /*
1358          * Back out if somebody else faulted in this pte while we
1359          * released the page table lock.
1360          */
1361         spin_lock(&mm->page_table_lock);
1362         page_table = pte_offset_map(pmd, address);
1363         if (unlikely(!pte_same(*page_table, orig_pte))) {
1364                 pte_unmap(page_table);
1365                 spin_unlock(&mm->page_table_lock);
1366                 unlock_page(page);
1367                 page_cache_release(page);
1368                 ret = VM_FAULT_MINOR;
1369                 goto out;
1370         }
1371
1372         /* The page isn't present yet, go ahead with the fault. */
1373
1374         swap_free(entry);
1375         if (vm_swap_full())
1376                 remove_exclusive_swap_page(page);
1377
1378         // mm->rss++;
1379         vx_rsspages_inc(mm);
1380         pte = mk_pte(page, vma->vm_page_prot);
1381         if (write_access && can_share_swap_page(page)) {
1382                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
1383                 write_access = 0;
1384         }
1385         unlock_page(page);
1386
1387         flush_icache_page(vma, page);
1388         set_pte(page_table, pte);
1389         page_add_anon_rmap(page, vma, address);
1390
1391         if (write_access) {
1392                 if (do_wp_page(mm, vma, address,
1393                                 page_table, pmd, pte) == VM_FAULT_OOM)
1394                         ret = VM_FAULT_OOM;
1395                 goto out;
1396         }
1397
1398         /* No need to invalidate - it was non-present before */
1399         update_mmu_cache(vma, address, pte);
1400         pte_unmap(page_table);
1401         spin_unlock(&mm->page_table_lock);
1402 out:
1403         return ret;
1404 }
1405
1406 /*
1407  * We are called with the MM semaphore and page_table_lock
1408  * spinlock held to protect against concurrent faults in
1409  * multithreaded programs.
1410  */
1411 static int
1412 do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
1413                 pte_t *page_table, pmd_t *pmd, int write_access,
1414                 unsigned long addr)
1415 {
1416         pte_t entry;
1417         struct page * page = ZERO_PAGE(addr);
1418
1419         if (!vx_rsspages_avail(mm, 1)) {
1420                 spin_unlock(&mm->page_table_lock);
1421                 return VM_FAULT_OOM;
1422         }
1423
1424         /* Read-only mapping of ZERO_PAGE. */
1425         entry = pte_wrprotect(mk_pte(ZERO_PAGE(addr), vma->vm_page_prot));
1426
1427         /* ..except if it's a write access */
1428         if (write_access) {
1429                 /* Allocate our own private page. */
1430                 pte_unmap(page_table);
1431                 spin_unlock(&mm->page_table_lock);
1432
1433                 if (unlikely(anon_vma_prepare(vma)))
1434                         goto no_mem;
1435                 page = alloc_page_vma(GFP_HIGHUSER, vma, addr);
1436                 if (!page)
1437                         goto no_mem;
1438                 clear_user_highpage(page, addr);
1439
1440                 spin_lock(&mm->page_table_lock);
1441                 page_table = pte_offset_map(pmd, addr);
1442
1443                 if (!pte_none(*page_table)) {
1444                         pte_unmap(page_table);
1445                         page_cache_release(page);
1446                         spin_unlock(&mm->page_table_lock);
1447                         goto out;
1448                 }
1449                 // mm->rss++;
1450                 vx_rsspages_inc(mm);
1451                 entry = maybe_mkwrite(pte_mkdirty(mk_pte(page,
1452                                                          vma->vm_page_prot)),
1453                                       vma);
1454                 lru_cache_add_active(page);
1455                 mark_page_accessed(page);
1456                 page_add_anon_rmap(page, vma, addr);
1457         }
1458
1459         set_pte(page_table, entry);
1460         pte_unmap(page_table);
1461
1462         /* No need to invalidate - it was non-present before */
1463         update_mmu_cache(vma, addr, entry);
1464         spin_unlock(&mm->page_table_lock);
1465 out:
1466         return VM_FAULT_MINOR;
1467 no_mem:
1468         return VM_FAULT_OOM;
1469 }
1470
1471 /*
1472  * do_no_page() tries to create a new page mapping. It aggressively
1473  * tries to share with existing pages, but makes a separate copy if
1474  * the "write_access" parameter is true in order to avoid the next
1475  * page fault.
1476  *
1477  * As this is called only for pages that do not currently exist, we
1478  * do not need to flush old virtual caches or the TLB.
1479  *
1480  * This is called with the MM semaphore held and the page table
1481  * spinlock held. Exit with the spinlock released.
1482  */
1483 static int
1484 do_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
1485         unsigned long address, int write_access, pte_t *page_table, pmd_t *pmd)
1486 {
1487         struct page * new_page;
1488         struct address_space *mapping = NULL;
1489         pte_t entry;
1490         int sequence = 0;
1491         int ret = VM_FAULT_MINOR;
1492         int anon = 0;
1493
1494         if (!vma->vm_ops || !vma->vm_ops->nopage)
1495                 return do_anonymous_page(mm, vma, page_table,
1496                                         pmd, write_access, address);
1497         pte_unmap(page_table);
1498         spin_unlock(&mm->page_table_lock);
1499
1500         if (vma->vm_file) {
1501                 mapping = vma->vm_file->f_mapping;
1502                 sequence = atomic_read(&mapping->truncate_count);
1503         }
1504         smp_rmb();  /* Prevent CPU from reordering lock-free ->nopage() */
1505 retry:
1506         new_page = vma->vm_ops->nopage(vma, address & PAGE_MASK, &ret);
1507
1508         /* no page was available -- either SIGBUS or OOM */
1509         if (new_page == NOPAGE_SIGBUS)
1510                 return VM_FAULT_SIGBUS;
1511         if (new_page == NOPAGE_OOM)
1512                 return VM_FAULT_OOM;
1513         if (!vx_rsspages_avail(mm, 1))
1514                 return VM_FAULT_OOM;
1515
1516         /*
1517          * Should we do an early C-O-W break?
1518          */
1519         if (write_access && !(vma->vm_flags & VM_SHARED)) {
1520                 struct page *page;
1521
1522                 if (unlikely(anon_vma_prepare(vma)))
1523                         goto oom;
1524                 page = alloc_page_vma(GFP_HIGHUSER, vma, address);
1525                 if (!page)
1526                         goto oom;
1527                 copy_user_highpage(page, new_page, address);
1528                 page_cache_release(new_page);
1529                 new_page = page;
1530                 anon = 1;
1531         }
1532
1533         spin_lock(&mm->page_table_lock);
1534         /*
1535          * For a file-backed vma, someone could have truncated or otherwise
1536          * invalidated this page.  If unmap_mapping_range got called,
1537          * retry getting the page.
1538          */
1539         if (mapping &&
1540               (unlikely(sequence != atomic_read(&mapping->truncate_count)))) {
1541                 sequence = atomic_read(&mapping->truncate_count);
1542                 spin_unlock(&mm->page_table_lock);
1543                 page_cache_release(new_page);
1544                 goto retry;
1545         }
1546         page_table = pte_offset_map(pmd, address);
1547
1548         /*
1549          * This silly early PAGE_DIRTY setting removes a race
1550          * due to the bad i386 page protection. But it's valid
1551          * for other architectures too.
1552          *
1553          * Note that if write_access is true, we either now have
1554          * an exclusive copy of the page, or this is a shared mapping,
1555          * so we can make it writable and dirty to avoid having to
1556          * handle that later.
1557          */
1558         /* Only go through if we didn't race with anybody else... */
1559         if (pte_none(*page_table)) {
1560                 if (!PageReserved(new_page))
1561                         // ++mm->rss;
1562                         vx_rsspages_inc(mm);
1563                 flush_icache_page(vma, new_page);
1564                 entry = mk_pte(new_page, vma->vm_page_prot);
1565                 if (write_access)
1566                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1567                 set_pte(page_table, entry);
1568                 if (anon) {
1569                         lru_cache_add_active(new_page);
1570                         page_add_anon_rmap(new_page, vma, address);
1571                 } else
1572                         page_add_file_rmap(new_page);
1573                 pte_unmap(page_table);
1574         } else {
1575                 /* One of our sibling threads was faster, back out. */
1576                 pte_unmap(page_table);
1577                 page_cache_release(new_page);
1578                 spin_unlock(&mm->page_table_lock);
1579                 goto out;
1580         }
1581
1582         /* no need to invalidate: a not-present page shouldn't be cached */
1583         update_mmu_cache(vma, address, entry);
1584         spin_unlock(&mm->page_table_lock);
1585 out:
1586         return ret;
1587 oom:
1588         page_cache_release(new_page);
1589         ret = VM_FAULT_OOM;
1590         goto out;
1591 }
1592
1593 /*
1594  * Fault of a previously existing named mapping. Repopulate the pte
1595  * from the encoded file_pte if possible. This enables swappable
1596  * nonlinear vmas.
1597  */
1598 static int do_file_page(struct mm_struct * mm, struct vm_area_struct * vma,
1599         unsigned long address, int write_access, pte_t *pte, pmd_t *pmd)
1600 {
1601         unsigned long pgoff;
1602         int err;
1603
1604         BUG_ON(!vma->vm_ops || !vma->vm_ops->nopage);
1605         /*
1606          * Fall back to the linear mapping if the fs does not support
1607          * ->populate:
1608          */
1609         if (!vma->vm_ops || !vma->vm_ops->populate ||
1610                         (write_access && !(vma->vm_flags & VM_SHARED))) {
1611                 pte_clear(pte);
1612                 return do_no_page(mm, vma, address, write_access, pte, pmd);
1613         }
1614
1615         pgoff = pte_to_pgoff(*pte);
1616
1617         pte_unmap(pte);
1618         spin_unlock(&mm->page_table_lock);
1619
1620         err = vma->vm_ops->populate(vma, address & PAGE_MASK, PAGE_SIZE, vma->vm_page_prot, pgoff, 0);
1621         if (err == -ENOMEM)
1622                 return VM_FAULT_OOM;
1623         if (err)
1624                 return VM_FAULT_SIGBUS;
1625         return VM_FAULT_MAJOR;
1626 }
1627
1628 /*
1629  * These routines also need to handle stuff like marking pages dirty
1630  * and/or accessed for architectures that don't do it in hardware (most
1631  * RISC architectures).  The early dirtying is also good on the i386.
1632  *
1633  * There is also a hook called "update_mmu_cache()" that architectures
1634  * with external mmu caches can use to update those (ie the Sparc or
1635  * PowerPC hashed page tables that act as extended TLBs).
1636  *
1637  * Note the "page_table_lock". It is to protect against kswapd removing
1638  * pages from under us. Note that kswapd only ever _removes_ pages, never
1639  * adds them. As such, once we have noticed that the page is not present,
1640  * we can drop the lock early.
1641  *
1642  * The adding of pages is protected by the MM semaphore (which we hold),
1643  * so we don't need to worry about a page being suddenly been added into
1644  * our VM.
1645  *
1646  * We enter with the pagetable spinlock held, we are supposed to
1647  * release it when done.
1648  */
1649 static inline int handle_pte_fault(struct mm_struct *mm,
1650         struct vm_area_struct * vma, unsigned long address,
1651         int write_access, pte_t *pte, pmd_t *pmd)
1652 {
1653         pte_t entry;
1654
1655         entry = *pte;
1656         if (!pte_present(entry)) {
1657                 /*
1658                  * If it truly wasn't present, we know that kswapd
1659                  * and the PTE updates will not touch it later. So
1660                  * drop the lock.
1661                  */
1662                 if (pte_none(entry))
1663                         return do_no_page(mm, vma, address, write_access, pte, pmd);
1664                 if (pte_file(entry))
1665                         return do_file_page(mm, vma, address, write_access, pte, pmd);
1666                 return do_swap_page(mm, vma, address, pte, pmd, entry, write_access);
1667         }
1668
1669         if (write_access) {
1670                 if (!pte_write(entry))
1671                         return do_wp_page(mm, vma, address, pte, pmd, entry);
1672
1673                 entry = pte_mkdirty(entry);
1674         }
1675         entry = pte_mkyoung(entry);
1676         ptep_set_access_flags(vma, address, pte, entry, write_access);
1677         update_mmu_cache(vma, address, entry);
1678         pte_unmap(pte);
1679         spin_unlock(&mm->page_table_lock);
1680         return VM_FAULT_MINOR;
1681 }
1682
1683 /*
1684  * By the time we get here, we already hold the mm semaphore
1685  */
1686 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct * vma,
1687         unsigned long address, int write_access)
1688 {
1689         pgd_t *pgd;
1690         pmd_t *pmd;
1691
1692         __set_current_state(TASK_RUNNING);
1693         pgd = pgd_offset(mm, address);
1694
1695         inc_page_state(pgfault);
1696
1697         if (is_vm_hugetlb_page(vma))
1698                 return VM_FAULT_SIGBUS; /* mapping truncation does this. */
1699
1700         /*
1701          * We need the page table lock to synchronize with kswapd
1702          * and the SMP-safe atomic PTE updates.
1703          */
1704         spin_lock(&mm->page_table_lock);
1705         pmd = pmd_alloc(mm, pgd, address);
1706
1707         if (pmd) {
1708                 pte_t * pte = pte_alloc_map(mm, pmd, address);
1709                 if (pte)
1710                         return handle_pte_fault(mm, vma, address, write_access, pte, pmd);
1711         }
1712         spin_unlock(&mm->page_table_lock);
1713         return VM_FAULT_OOM;
1714 }
1715
1716 /*
1717  * Allocate page middle directory.
1718  *
1719  * We've already handled the fast-path in-line, and we own the
1720  * page table lock.
1721  *
1722  * On a two-level page table, this ends up actually being entirely
1723  * optimized away.
1724  */
1725 pmd_t fastcall *__pmd_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
1726 {
1727         pmd_t *new;
1728
1729         spin_unlock(&mm->page_table_lock);
1730         new = pmd_alloc_one(mm, address);
1731         spin_lock(&mm->page_table_lock);
1732         if (!new)
1733                 return NULL;
1734
1735         /*
1736          * Because we dropped the lock, we should re-check the
1737          * entry, as somebody else could have populated it..
1738          */
1739         if (pgd_present(*pgd)) {
1740                 pmd_free(new);
1741                 goto out;
1742         }
1743         pgd_populate(mm, pgd, new);
1744 out:
1745         return pmd_offset(pgd, address);
1746 }
1747
1748 int make_pages_present(unsigned long addr, unsigned long end)
1749 {
1750         int ret, len, write;
1751         struct vm_area_struct * vma;
1752
1753         vma = find_vma(current->mm, addr);
1754         write = (vma->vm_flags & VM_WRITE) != 0;
1755         if (addr >= end)
1756                 BUG();
1757         if (end > vma->vm_end)
1758                 BUG();
1759         len = (end+PAGE_SIZE-1)/PAGE_SIZE-addr/PAGE_SIZE;
1760         ret = get_user_pages(current, current->mm, addr,
1761                         len, write, 0, NULL, NULL);
1762         if (ret < 0)
1763                 return ret;
1764         return ret == len ? 0 : -1;
1765 }
1766
1767 /*
1768  * Map a vmalloc()-space virtual address to the physical page.
1769  */
1770 struct page * vmalloc_to_page(void * vmalloc_addr)
1771 {
1772         unsigned long addr = (unsigned long) vmalloc_addr;
1773         struct page *page = NULL;
1774         pgd_t *pgd = pgd_offset_k(addr);
1775         pmd_t *pmd;
1776         pte_t *ptep, pte;
1777
1778         if (!pgd_none(*pgd)) {
1779                 pmd = pmd_offset(pgd, addr);
1780                 if (!pmd_none(*pmd)) {
1781                         preempt_disable();
1782                         ptep = pte_offset_map(pmd, addr);
1783                         pte = *ptep;
1784                         if (pte_present(pte))
1785                                 page = pte_page(pte);
1786                         pte_unmap(ptep);
1787                         preempt_enable();
1788                 }
1789         }
1790         return page;
1791 }
1792
1793 EXPORT_SYMBOL(vmalloc_to_page);
1794
1795 #if !defined(CONFIG_ARCH_GATE_AREA)
1796
1797 #if defined(AT_SYSINFO_EHDR)
1798 struct vm_area_struct gate_vma;
1799
1800 static int __init gate_vma_init(void)
1801 {
1802         gate_vma.vm_mm = NULL;
1803         gate_vma.vm_start = FIXADDR_USER_START;
1804         gate_vma.vm_end = FIXADDR_USER_END;
1805         gate_vma.vm_page_prot = PAGE_READONLY;
1806         gate_vma.vm_flags = 0;
1807         return 0;
1808 }
1809 __initcall(gate_vma_init);
1810 #endif
1811
1812 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
1813 {
1814 #ifdef AT_SYSINFO_EHDR
1815         return &gate_vma;
1816 #else
1817         return 0;
1818 #endif
1819 }
1820
1821 int in_gate_area(struct task_struct *task, unsigned long addr)
1822 {
1823 #ifdef AT_SYSINFO_EHDR
1824         if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
1825                 return 1;
1826 #endif
1827         return 0;
1828 }
1829
1830 #endif