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 /*
  70  * A number of key systems in x86 including ioremap() rely on the assumption
  71  * that high_memory defines the upper bound on direct map memory, then end
  72  * of ZONE_NORMAL.  Under CONFIG_DISCONTIG this means that max_low_pfn and
  73  * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
  74  * and ZONE_HIGHMEM.
  75  */
  76 void * high_memory;
  77 struct page *highmem_start_page;
  78 unsigned long vmalloc_earlyreserve;
  79
  80 EXPORT_SYMBOL(num_physpages);
  81 EXPORT_SYMBOL(highmem_start_page);
  82 EXPORT_SYMBOL(high_memory);
  83 EXPORT_SYMBOL(vmalloc_earlyreserve);
  84
  85 /*
  86  * We special-case the C-O-W ZERO_PAGE, because it's such
  87  * a common occurrence (no need to read the page to know
  88  * that it's zero - better for the cache and memory subsystem).
  89  */
  90 static inline void copy_cow_page(struct page * from, struct page * to, unsigned long address)
  91 {
  92         if (from == ZERO_PAGE(address)) {
  93                 clear_user_highpage(to, address);
  94                 return;
  95         }
  96         copy_user_highpage(to, from, address);
  97 }
  98
  99 /*
 100  * Note: this doesn't free the actual pages themselves. That
 101  * has been handled earlier when unmapping all the memory regions.
 102  */
 103 static inline void free_one_pmd(struct mmu_gather *tlb, pmd_t * dir)
 104 {
 105         struct page *page;
 106
 107         if (pmd_none(*dir))
 108                 return;
 109         if (unlikely(pmd_bad(*dir))) {
 110                 pmd_ERROR(*dir);
 111                 pmd_clear(dir);
 112                 return;
 113         }
 114         page = pmd_page(*dir);
 115         pmd_clear(dir);
 116         dec_page_state(nr_page_table_pages);
 117         pte_free_tlb(tlb, page);
 118 }
 119
 120 static inline void free_one_pgd(struct mmu_gather *tlb, pgd_t * dir,
 121                                                         int pgd_idx)
 122 {
 123         int j;
 124         pmd_t * pmd;
 125
 126         if (pgd_none(*dir))
 127                 return;
 128         if (unlikely(pgd_bad(*dir))) {
 129                 pgd_ERROR(*dir);
 130                 pgd_clear(dir);
 131                 return;
 132         }
 133         pmd = pmd_offset(dir, 0);
 134         pgd_clear(dir);
 135         for (j = 0; j < PTRS_PER_PMD ; j++) {
 136                 if (pgd_idx * PGDIR_SIZE + j * PMD_SIZE >= TASK_SIZE)
 137                         break;
 138                 free_one_pmd(tlb, pmd+j);
 139         }
 140         pmd_free_tlb(tlb, pmd);
 141 }
 142
 143 /*
 144  * This function clears all user-level page tables of a process - this
 145  * is needed by execve(), so that old pages aren't in the way.
 146  *
 147  * Must be called with pagetable lock held.
 148  */
 149 void clear_page_tables(struct mmu_gather *tlb, unsigned long first, int nr)
 150 {
 151         pgd_t * page_dir = tlb->mm->pgd;
 152         int pgd_idx = first;
 153
 154         page_dir += first;
 155         do {
 156                 free_one_pgd(tlb, page_dir, pgd_idx);
 157                 page_dir++;
 158                 pgd_idx++;
 159         } while (--nr);
 160 }
 161
 162 pte_t fastcall * pte_alloc_map(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
 163 {
 164         if (!pmd_present(*pmd)) {
 165                 struct page *new;
 166
 167                 spin_unlock(&mm->page_table_lock);
 168                 new = pte_alloc_one(mm, address);
 169                 spin_lock(&mm->page_table_lock);
 170                 if (!new)
 171                         return NULL;
 172
 173                 /*
 174                  * Because we dropped the lock, we should re-check the
 175                  * entry, as somebody else could have populated it..
 176                  */
 177                 if (pmd_present(*pmd)) {
 178                         pte_free(new);
 179                         goto out;
 180                 }
 181                 inc_page_state(nr_page_table_pages);
 182                 pmd_populate(mm, pmd, new);
 183         }
 184 out:
 185         return pte_offset_map(pmd, address);
 186 }
 187
 188 pte_t fastcall * pte_alloc_kernel(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
 189 {
 190         if (!pmd_present(*pmd)) {
 191                 pte_t *new;
 192
 193                 spin_unlock(&mm->page_table_lock);
 194                 new = pte_alloc_one_kernel(mm, address);
 195                 spin_lock(&mm->page_table_lock);
 196                 if (!new)
 197                         return NULL;
 198
 199                 /*
 200                  * Because we dropped the lock, we should re-check the
 201                  * entry, as somebody else could have populated it..
 202                  */
 203                 if (pmd_present(*pmd)) {
 204                         pte_free_kernel(new);
 205                         goto out;
 206                 }
 207                 pmd_populate_kernel(mm, pmd, new);
 208         }
 209 out:
 210         return pte_offset_kernel(pmd, address);
 211 }
 212 #define PTE_TABLE_MASK  ((PTRS_PER_PTE-1) * sizeof(pte_t))
 213 #define PMD_TABLE_MASK  ((PTRS_PER_PMD-1) * sizeof(pmd_t))
 214
 215 /*
 216  * copy one vm_area from one task to the other. Assumes the page tables
 217  * already present in the new task to be cleared in the whole range
 218  * covered by this vma.
 219  *
 220  * 08Jan98 Merged into one routine from several inline routines to reduce
 221  *         variable count and make things faster. -jj
 222  *
 223  * dst->page_table_lock is held on entry and exit,
 224  * but may be dropped within pmd_alloc() and pte_alloc_map().
 225  */
 226 int copy_page_range(struct mm_struct *dst, struct mm_struct *src,
 227                         struct vm_area_struct *vma)
 228 {
 229         pgd_t * src_pgd, * dst_pgd;
 230         unsigned long address = vma->vm_start;
 231         unsigned long end = vma->vm_end;
 232         unsigned long cow;
 233
 234         if (is_vm_hugetlb_page(vma))
 235                 return copy_hugetlb_page_range(dst, src, vma);
 236
 237         cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
 238         src_pgd = pgd_offset(src, address)-1;
 239         dst_pgd = pgd_offset(dst, address)-1;
 240
 241         for (;;) {
 242                 pmd_t * src_pmd, * dst_pmd;
 243
 244                 src_pgd++; dst_pgd++;
 245
 246                 /* copy_pmd_range */
 247
 248                 if (pgd_none(*src_pgd))
 249                         goto skip_copy_pmd_range;
 250                 if (unlikely(pgd_bad(*src_pgd))) {
 251                         pgd_ERROR(*src_pgd);
 252                         pgd_clear(src_pgd);
 253 skip_copy_pmd_range:    address = (address + PGDIR_SIZE) & PGDIR_MASK;
 254                         if (!address || (address >= end))
 255                                 goto out;
 256                         continue;
 257                 }
 258
 259                 src_pmd = pmd_offset(src_pgd, address);
 260                 dst_pmd = pmd_alloc(dst, dst_pgd, address);
 261                 if (!dst_pmd)
 262                         goto nomem;
 263
 264                 do {
 265                         pte_t * src_pte, * dst_pte;
 266
 267                         /* copy_pte_range */
 268
 269                         if (pmd_none(*src_pmd))
 270                                 goto skip_copy_pte_range;
 271                         if (unlikely(pmd_bad(*src_pmd))) {
 272                                 pmd_ERROR(*src_pmd);
 273                                 pmd_clear(src_pmd);
 274 skip_copy_pte_range:
 275                                 address = (address + PMD_SIZE) & PMD_MASK;
 276                                 if (address >= end)
 277                                         goto out;
 278                                 goto cont_copy_pmd_range;
 279                         }
 280
 281                         dst_pte = pte_alloc_map(dst, dst_pmd, address);
 282                         if (!dst_pte)
 283                                 goto nomem;
 284                         spin_lock(&src->page_table_lock);
 285                         src_pte = pte_offset_map_nested(src_pmd, address);
 286                         do {
 287                                 pte_t pte = *src_pte;
 288                                 struct page *page;
 289                                 unsigned long pfn;
 290
 291                                 if (!vx_rsspages_avail(dst, 1)) {
 292                                         spin_unlock(&src->page_table_lock);
 293                                         goto nomem;
 294                                 }
 295                                 /* copy_one_pte */
 296
 297                                 if (pte_none(pte))
 298                                         goto cont_copy_pte_range_noset;
 299                                 /* pte contains position in swap, so copy. */
 300                                 if (!pte_present(pte)) {
 301                                         if (!pte_file(pte))
 302                                                 swap_duplicate(pte_to_swp_entry(pte));
 303                                         set_pte(dst_pte, pte);
 304                                         goto cont_copy_pte_range_noset;
 305                                 }
 306                                 pfn = pte_pfn(pte);
 307                                 /* the pte points outside of valid memory, the
 308                                  * mapping is assumed to be good, meaningful
 309                                  * and not mapped via rmap - duplicate the
 310                                  * mapping as is.
 311                                  */
 312                                 page = NULL;
 313                                 if (pfn_valid(pfn))
 314                                         page = pfn_to_page(pfn);
 315
 316                                 if (!page || PageReserved(page)) {
 317                                         set_pte(dst_pte, pte);
 318                                         goto cont_copy_pte_range_noset;
 319                                 }
 320
 321                                 /*
 322                                  * If it's a COW mapping, write protect it both
 323                                  * in the parent and the child
 324                                  */
 325                                 if (cow) {
 326                                         ptep_set_wrprotect(src_pte);
 327                                         pte = *src_pte;
 328                                 }
 329
 330                                 /*
 331                                  * If it's a shared mapping, mark it clean in
 332                                  * the child
 333                                  */
 334                                 if (vma->vm_flags & VM_SHARED)
 335                                         pte = pte_mkclean(pte);
 336                                 pte = pte_mkold(pte);
 337                                 get_page(page);
 338                                 // dst->rss++;
 339                                 vx_rsspages_inc(dst);
 340                                 set_pte(dst_pte, pte);
 341                                 page_dup_rmap(page);
 342 cont_copy_pte_range_noset:
 343                                 address += PAGE_SIZE;
 344                                 if (address >= end) {
 345                                         pte_unmap_nested(src_pte);
 346                                         pte_unmap(dst_pte);
 347                                         goto out_unlock;
 348                                 }
 349                                 src_pte++;
 350                                 dst_pte++;
 351                         } while ((unsigned long)src_pte & PTE_TABLE_MASK);
 352                         pte_unmap_nested(src_pte-1);
 353                         pte_unmap(dst_pte-1);
 354                         spin_unlock(&src->page_table_lock);
 355                         cond_resched_lock(&dst->page_table_lock);
 356 cont_copy_pmd_range:
 357                         src_pmd++;
 358                         dst_pmd++;
 359                 } while ((unsigned long)src_pmd & PMD_TABLE_MASK);
 360         }
 361 out_unlock:
 362         spin_unlock(&src->page_table_lock);
 363 out:
 364         return 0;
 365 nomem:
 366         return -ENOMEM;
 367 }
 368
 369 static void zap_pte_range(struct mmu_gather *tlb,
 370                 pmd_t *pmd, unsigned long address,
 371                 unsigned long size, struct zap_details *details)
 372 {
 373         unsigned long offset;
 374         pte_t *ptep;
 375
 376         if (pmd_none(*pmd))
 377                 return;
 378         if (unlikely(pmd_bad(*pmd))) {
 379                 pmd_ERROR(*pmd);
 380                 pmd_clear(pmd);
 381                 return;
 382         }
 383         ptep = pte_offset_map(pmd, address);
 384         offset = address & ~PMD_MASK;
 385         if (offset + size > PMD_SIZE)
 386                 size = PMD_SIZE - offset;
 387         size &= PAGE_MASK;
 388         if (details && !details->check_mapping && !details->nonlinear_vma)
 389                 details = NULL;
 390         for (offset=0; offset < size; ptep++, offset += PAGE_SIZE) {
 391                 pte_t pte = *ptep;
 392                 if (pte_none(pte))
 393                         continue;
 394                 if (pte_present(pte)) {
 395                         struct page *page = NULL;
 396                         unsigned long pfn = pte_pfn(pte);
 397                         if (pfn_valid(pfn)) {
 398                                 page = pfn_to_page(pfn);
 399                                 if (PageReserved(page))
 400                                         page = NULL;
 401                         }
 402                         if (unlikely(details) && page) {
 403                                 /*
 404                                  * unmap_shared_mapping_pages() wants to
 405                                  * invalidate cache without truncating:
 406                                  * unmap shared but keep private pages.
 407                                  */
 408                                 if (details->check_mapping &&
 409                                     details->check_mapping != page->mapping)
 410                                         continue;
 411                                 /*
 412                                  * Each page->index must be checked when
 413                                  * invalidating or truncating nonlinear.
 414                                  */
 415                                 if (details->nonlinear_vma &&
 416                                     (page->index < details->first_index ||
 417                                      page->index > details->last_index))
 418                                         continue;
 419                         }
 420                         pte = ptep_get_and_clear(ptep);
 421                         tlb_remove_tlb_entry(tlb, ptep, address+offset);
 422                         if (unlikely(!page))
 423                                 continue;
 424                         if (unlikely(details) && details->nonlinear_vma
 425                             && linear_page_index(details->nonlinear_vma,
 426                                         address+offset) != page->index)
 427                                 set_pte(ptep, pgoff_to_pte(page->index));
 428                         if (pte_dirty(pte))
 429                                 set_page_dirty(page);
 430                         if (pte_young(pte) && !PageAnon(page))
 431                                 mark_page_accessed(page);
 432                         tlb->freed++;
 433                         page_remove_rmap(page);
 434                         tlb_remove_page(tlb, page);
 435                         continue;
 436                 }
 437                 /*
 438                  * If details->check_mapping, we leave swap entries;
 439                  * if details->nonlinear_vma, we leave file entries.
 440                  */
 441                 if (unlikely(details))
 442                         continue;
 443                 if (!pte_file(pte))
 444                         free_swap_and_cache(pte_to_swp_entry(pte));
 445                 pte_clear(ptep);
 446         }
 447         pte_unmap(ptep-1);
 448 }
 449
 450 static void zap_pmd_range(struct mmu_gather *tlb,
 451                 pgd_t * dir, unsigned long address,
 452                 unsigned long size, struct zap_details *details)
 453 {
 454         pmd_t * pmd;
 455         unsigned long end, pgd_boundary;
 456
 457         if (pgd_none(*dir))
 458                 return;
 459         if (unlikely(pgd_bad(*dir))) {
 460                 pgd_ERROR(*dir);
 461                 pgd_clear(dir);
 462                 return;
 463         }
 464         pmd = pmd_offset(dir, address);
 465         end = address + size;
 466         pgd_boundary = ((address + PGDIR_SIZE) & PGDIR_MASK);
 467         if (pgd_boundary && (end > pgd_boundary))
 468                 end = pgd_boundary;
 469         do {
 470                 zap_pte_range(tlb, pmd, address, end - address, details);
 471                 address = (address + PMD_SIZE) & PMD_MASK;
 472                 pmd++;
 473         } while (address && (address < end));
 474 }
 475
 476 static void unmap_page_range(struct mmu_gather *tlb,
 477                 struct vm_area_struct *vma, unsigned long address,
 478                 unsigned long end, struct zap_details *details)
 479 {
 480         pgd_t * dir;
 481
 482         BUG_ON(address >= end);
 483         dir = pgd_offset(vma->vm_mm, address);
 484         tlb_start_vma(tlb, vma);
 485         do {
 486                 zap_pmd_range(tlb, dir, address, end - address, details);
 487                 address = (address + PGDIR_SIZE) & PGDIR_MASK;
 488                 dir++;
 489         } while (address && (address < end));
 490         tlb_end_vma(tlb, vma);
 491 }
 492
 493 #ifdef CONFIG_PREEMPT_VOLUNTARY
 494 # define ZAP_BLOCK_SIZE (128 * PAGE_SIZE)
 495 #else
 496
 497 /* Dispose of an entire struct mmu_gather per rescheduling point */
 498 #if defined(CONFIG_SMP) && defined(CONFIG_PREEMPT)
 499 #define ZAP_BLOCK_SIZE  (FREE_PTE_NR * PAGE_SIZE)
 500 #endif
 501
 502 /* For UP, 256 pages at a time gives nice low latency */
 503 #if !defined(CONFIG_SMP) && defined(CONFIG_PREEMPT)
 504 #define ZAP_BLOCK_SIZE  (256 * PAGE_SIZE)
 505 #endif
 506
 507 /* No preempt: go for improved straight-line efficiency */
 508 #if !defined(CONFIG_PREEMPT)
 509 #define ZAP_BLOCK_SIZE  (1024 * PAGE_SIZE)
 510 #endif
 511
 512 #endif
 513
 514 /**
 515  * unmap_vmas - unmap a range of memory covered by a list of vma's
 516  * @tlbp: address of the caller's struct mmu_gather
 517  * @mm: the controlling mm_struct
 518  * @vma: the starting vma
 519  * @start_addr: virtual address at which to start unmapping
 520  * @end_addr: virtual address at which to end unmapping
 521  * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
 522  * @details: details of nonlinear truncation or shared cache invalidation
 523  *
 524  * Returns the number of vma's which were covered by the unmapping.
 525  *
 526  * Unmap all pages in the vma list.  Called under page_table_lock.
 527  *
 528  * We aim to not hold page_table_lock for too long (for scheduling latency
 529  * reasons).  So zap pages in ZAP_BLOCK_SIZE bytecounts.  This means we need to
 530  * return the ending mmu_gather to the caller.
 531  *
 532  * Only addresses between `start' and `end' will be unmapped.
 533  *
 534  * The VMA list must be sorted in ascending virtual address order.
 535  *
 536  * unmap_vmas() assumes that the caller will flush the whole unmapped address
 537  * range after unmap_vmas() returns.  So the only responsibility here is to
 538  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
 539  * drops the lock and schedules.
 540  */
 541 int unmap_vmas(struct mmu_gather **tlbp, struct mm_struct *mm,
 542                 struct vm_area_struct *vma, unsigned long start_addr,
 543                 unsigned long end_addr, unsigned long *nr_accounted,
 544                 struct zap_details *details)
 545 {
 546         unsigned long zap_bytes = ZAP_BLOCK_SIZE;
 547         unsigned long tlb_start = 0;    /* For tlb_finish_mmu */
 548         int tlb_start_valid = 0;
 549         int ret = 0;
 550         int atomic = details && details->atomic;
 551
 552         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
 553                 unsigned long start;
 554                 unsigned long end;
 555
 556                 start = max(vma->vm_start, start_addr);
 557                 if (start >= vma->vm_end)
 558                         continue;
 559                 end = min(vma->vm_end, end_addr);
 560                 if (end <= vma->vm_start)
 561                         continue;
 562
 563                 if (vma->vm_flags & VM_ACCOUNT)
 564                         *nr_accounted += (end - start) >> PAGE_SHIFT;
 565
 566                 ret++;
 567                 while (start != end) {
 568                         unsigned long block;
 569
 570                         if (!tlb_start_valid) {
 571                                 tlb_start = start;
 572                                 tlb_start_valid = 1;
 573                         }
 574
 575                         if (is_vm_hugetlb_page(vma)) {
 576                                 block = end - start;
 577                                 unmap_hugepage_range(vma, start, end);
 578                         } else {
 579                                 block = min(zap_bytes, end - start);
 580                                 unmap_page_range(*tlbp, vma, start,
 581                                                 start + block, details);
 582                         }
 583
 584                         start += block;
 585                         zap_bytes -= block;
 586                         if (!atomic && need_resched()) {
 587                                 int fullmm = tlb_is_full_mm(*tlbp);
 588                                 tlb_finish_mmu(*tlbp, tlb_start, start);
 589                                 cond_resched_lock(&mm->page_table_lock);
 590                                 *tlbp = tlb_gather_mmu(mm, fullmm);
 591                                 tlb_start_valid = 0;
 592                         }
 593                         if ((long)zap_bytes > 0)
 594                                 continue;
 595                         zap_bytes = ZAP_BLOCK_SIZE;
 596                 }
 597         }
 598         return ret;
 599 }
 600
 601 /**
 602  * zap_page_range - remove user pages in a given range
 603  * @vma: vm_area_struct holding the applicable pages
 604  * @address: starting address of pages to zap
 605  * @size: number of bytes to zap
 606  * @details: details of nonlinear truncation or shared cache invalidation
 607  */
 608 void zap_page_range(struct vm_area_struct *vma, unsigned long address,
 609                 unsigned long size, struct zap_details *details)
 610 {
 611         struct mm_struct *mm = vma->vm_mm;
 612         struct mmu_gather *tlb;
 613         unsigned long end = address + size;
 614         unsigned long nr_accounted = 0;
 615
 616         if (is_vm_hugetlb_page(vma)) {
 617                 zap_hugepage_range(vma, address, size);
 618                 return;
 619         }
 620
 621         lru_add_drain();
 622         spin_lock(&mm->page_table_lock);
 623         tlb = tlb_gather_mmu(mm, 0);
 624         unmap_vmas(&tlb, mm, vma, address, end, &nr_accounted, details);
 625         tlb_finish_mmu(tlb, address, end);
 626         spin_unlock(&mm->page_table_lock);
 627 }
 628
 629 /*
 630  * Do a quick page-table lookup for a single page.
 631  * mm->page_table_lock must be held.
 632  */
 633 struct page *
 634 follow_page(struct mm_struct *mm, unsigned long address, int write)
 635 {
 636         pgd_t *pgd;
 637         pmd_t *pmd;
 638         pte_t *ptep, pte;
 639         unsigned long pfn;
 640         struct page *page;
 641
 642         page = follow_huge_addr(mm, address, write);
 643         if (! IS_ERR(page))
 644                 return page;
 645
 646         pgd = pgd_offset(mm, address);
 647         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
 648                 goto out;
 649
 650         pmd = pmd_offset(pgd, address);
 651         if (pmd_none(*pmd))
 652                 goto out;
 653         if (pmd_huge(*pmd))
 654                 return follow_huge_pmd(mm, address, pmd, write);
 655         if (unlikely(pmd_bad(*pmd)))
 656                 goto out;
 657
 658         ptep = pte_offset_map(pmd, address);
 659         if (!ptep)
 660                 goto out;
 661
 662         pte = *ptep;
 663         pte_unmap(ptep);
 664         if (pte_present(pte)) {
 665                 if (write && !pte_write(pte))
 666                         goto out;
 667                 pfn = pte_pfn(pte);
 668                 if (pfn_valid(pfn)) {
 669                         page = pfn_to_page(pfn);
 670                         if (write && !pte_dirty(pte) && !PageDirty(page))
 671                                 set_page_dirty(page);
 672                         mark_page_accessed(page);
 673                         return page;
 674                 }
 675         }
 676
 677 out:
 678         return NULL;
 679 }
 680
 681 struct page *
 682 follow_page_pfn(struct mm_struct *mm, unsigned long address, int write,
 683                 unsigned long *pfn_ptr)
 684 {
 685         pgd_t *pgd;
 686         pmd_t *pmd;
 687         pte_t *ptep, pte;
 688         unsigned long pfn;
 689         struct page *page;
 690
 691         *pfn_ptr = 0;
 692         page = follow_huge_addr(mm, address, write);
 693         if (!IS_ERR(page))
 694                 return page;
 695
 696         pgd = pgd_offset(mm, address);
 697         if (pgd_none(*pgd) || pgd_bad(*pgd))
 698                 goto out;
 699
 700         pmd = pmd_offset(pgd, address);
 701         if (pmd_none(*pmd))
 702                 goto out;
 703         if (pmd_huge(*pmd))
 704                 return follow_huge_pmd(mm, address, pmd, write);
 705         if (pmd_bad(*pmd))
 706                 goto out;
 707
 708         ptep = pte_offset_map(pmd, address);
 709         if (!ptep)
 710                 goto out;
 711
 712         pte = *ptep;
 713         pte_unmap(ptep);
 714         if (pte_present(pte)) {
 715                 if (write && !pte_write(pte))
 716                         goto out;
 717                 if (write && !pte_dirty(pte)) {
 718                         struct page *page = pte_page(pte);
 719                         if (!PageDirty(page))
 720                                 set_page_dirty(page);
 721                 }
 722                 pfn = pte_pfn(pte);
 723                 if (pfn_valid(pfn)) {
 724                         struct page *page = pfn_to_page(pfn);
 725
 726                         mark_page_accessed(page);
 727                         return page;
 728                 } else {
 729                         *pfn_ptr = pfn;
 730                         return NULL;
 731                 }
 732         }
 733
 734 out:
 735         return NULL;
 736 }
 737
 738
 739 /*
 740  * Given a physical address, is there a useful struct page pointing to
 741  * it?  This may become more complex in the future if we start dealing
 742  * with IO-aperture pages for direct-IO.
 743  */
 744
 745 static inline struct page *get_page_map(struct page *page)
 746 {
 747         if (!pfn_valid(page_to_pfn(page)))
 748                 return NULL;
 749         return page;
 750 }
 751
 752
 753 #ifndef CONFIG_X86_4G
 754 static inline int
 755 untouched_anonymous_page(struct mm_struct* mm, struct vm_area_struct *vma,
 756                          unsigned long address)
 757 {
 758         pgd_t *pgd;
 759         pmd_t *pmd;
 760
 761         /* Check if the vma is for an anonymous mapping. */
 762         if (vma->vm_ops && vma->vm_ops->nopage)
 763                 return 0;
 764
 765         /* Check if page directory entry exists. */
 766         pgd = pgd_offset(mm, address);
 767         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
 768                 return 1;
 769
 770         /* Check if page middle directory entry exists. */
 771         pmd = pmd_offset(pgd, address);
 772         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
 773                 return 1;
 774
 775         /* There is a pte slot for 'address' in 'mm'. */
 776         return 0;
 777 }
 778 #endif
 779
 780
 781 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
 782                 unsigned long start, int len, int write, int force,
 783                 struct page **pages, struct vm_area_struct **vmas)
 784 {
 785         int i;
 786         unsigned int flags;
 787
 788         /*
 789          * Require read or write permissions.
 790          * If 'force' is set, we only require the "MAY" flags.
 791          */
 792         flags = write ? (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
 793         flags &= force ? (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
 794         i = 0;
 795
 796         do {
 797                 struct vm_area_struct * vma;
 798
 799                 vma = find_extend_vma(mm, start);
 800                 if (!vma && in_gate_area(tsk, start)) {
 801                         unsigned long pg = start & PAGE_MASK;
 802                         struct vm_area_struct *gate_vma = get_gate_vma(tsk);
 803                         pgd_t *pgd;
 804                         pmd_t *pmd;
 805                         pte_t *pte;
 806                         if (write) /* user gate pages are read-only */
 807                                 return i ? : -EFAULT;
 808                         pgd = pgd_offset_gate(mm, pg);
 809                         if (!pgd)
 810                                 return i ? : -EFAULT;
 811                         pmd = pmd_offset(pgd, pg);
 812                         if (!pmd)
 813                                 return i ? : -EFAULT;
 814                         pte = pte_offset_map(pmd, pg);
 815                         if (!pte)
 816                                 return i ? : -EFAULT;
 817                         if (!pte_present(*pte)) {
 818                                 pte_unmap(pte);
 819                                 return i ? : -EFAULT;
 820                         }
 821                         if (pages) {
 822                                 pages[i] = pte_page(*pte);
 823                                 get_page(pages[i]);
 824                         }
 825                         pte_unmap(pte);
 826                         if (vmas)
 827                                 vmas[i] = gate_vma;
 828                         i++;
 829                         start += PAGE_SIZE;
 830                         len--;
 831                         continue;
 832                 }
 833
 834                 if (!vma || (pages && (vma->vm_flags & VM_IO))
 835                                 || !(flags & vma->vm_flags))
 836                         return i ? : -EFAULT;
 837
 838                 if (is_vm_hugetlb_page(vma)) {
 839                         i = follow_hugetlb_page(mm, vma, pages, vmas,
 840                                                 &start, &len, i);
 841                         continue;
 842                 }
 843                 spin_lock(&mm->page_table_lock);
 844                 do {
 845                         struct page *map;
 846                         int lookup_write = write;
 847                         while (!(map = follow_page(mm, start, lookup_write))) {
 848                                 /*
 849                                  * Shortcut for anonymous pages. We don't want
 850                                  * to force the creation of pages tables for
 851                                  * insanly big anonymously mapped areas that
 852                                  * nobody touched so far. This is important
 853                                  * for doing a core dump for these mappings.
 854                                  *
 855                                  * disable this for 4:4 - it prevents
 856                                  * follow_page() from ever seeing these pages.
 857                                  *
 858                                  * (The 'fix' is dubious anyway, there's
 859                                  * nothing that this code avoids which couldnt
 860                                  * be triggered from userspace anyway.)
 861                                  */
 862 #ifndef CONFIG_X86_4G
 863                                 if (!lookup_write &&
 864                                     untouched_anonymous_page(mm,vma,start)) {
 865                                         map = ZERO_PAGE(start);
 866                                         break;
 867                                 }
 868 #endif
 869                                 spin_unlock(&mm->page_table_lock);
 870                                 switch (handle_mm_fault(mm,vma,start,write)) {
 871                                 case VM_FAULT_MINOR:
 872                                         tsk->min_flt++;
 873                                         break;
 874                                 case VM_FAULT_MAJOR:
 875                                         tsk->maj_flt++;
 876                                         break;
 877                                 case VM_FAULT_SIGBUS:
 878                                         return i ? i : -EFAULT;
 879                                 case VM_FAULT_OOM:
 880                                         return i ? i : -ENOMEM;
 881                                 default:
 882                                         BUG();
 883                                 }
 884                                 /*
 885                                  * Now that we have performed a write fault
 886                                  * and surely no longer have a shared page we
 887                                  * shouldn't write, we shouldn't ignore an
 888                                  * unwritable page in the page table if
 889                                  * we are forcing write access.
 890                                  */
 891                                 lookup_write = write && !force;
 892                                 spin_lock(&mm->page_table_lock);
 893                         }
 894                         if (pages) {
 895                                 pages[i] = get_page_map(map);
 896                                 if (!pages[i]) {
 897                                         spin_unlock(&mm->page_table_lock);
 898                                         while (i--)
 899                                                 page_cache_release(pages[i]);
 900                                         i = -EFAULT;
 901                                         goto out;
 902                                 }
 903                                 flush_dcache_page(pages[i]);
 904                                 if (!PageReserved(pages[i]))
 905                                         page_cache_get(pages[i]);
 906                         }
 907                         if (vmas)
 908                                 vmas[i] = vma;
 909                         i++;
 910                         start += PAGE_SIZE;
 911                         len--;
 912                 } while(len && start < vma->vm_end);
 913                 spin_unlock(&mm->page_table_lock);
 914         } while(len);
 915 out:
 916         return i;
 917 }
 918
 919 EXPORT_SYMBOL(get_user_pages);
 920
 921 static void zeromap_pte_range(pte_t * pte, unsigned long address,
 922                                      unsigned long size, pgprot_t prot)
 923 {
 924         unsigned long end;
 925
 926         address &= ~PMD_MASK;
 927         end = address + size;
 928         if (end > PMD_SIZE)
 929                 end = PMD_SIZE;
 930         do {
 931                 pte_t zero_pte = pte_wrprotect(mk_pte(ZERO_PAGE(address), prot));
 932                 BUG_ON(!pte_none(*pte));
 933                 set_pte(pte, zero_pte);
 934                 address += PAGE_SIZE;
 935                 pte++;
 936         } while (address && (address < end));
 937 }
 938
 939 static inline int zeromap_pmd_range(struct mm_struct *mm, pmd_t * pmd, unsigned long address,
 940                                     unsigned long size, pgprot_t prot)
 941 {
 942         unsigned long base, end;
 943
 944         base = address & PGDIR_MASK;
 945         address &= ~PGDIR_MASK;
 946         end = address + size;
 947         if (end > PGDIR_SIZE)
 948                 end = PGDIR_SIZE;
 949         do {
 950                 pte_t * pte = pte_alloc_map(mm, pmd, base + address);
 951                 if (!pte)
 952                         return -ENOMEM;
 953                 zeromap_pte_range(pte, base + address, end - address, prot);
 954                 pte_unmap(pte);
 955                 address = (address + PMD_SIZE) & PMD_MASK;
 956                 pmd++;
 957         } while (address && (address < end));
 958         return 0;
 959 }
 960
 961 int zeromap_page_range(struct vm_area_struct *vma, unsigned long address, unsigned long size, pgprot_t prot)
 962 {
 963         int error = 0;
 964         pgd_t * dir;
 965         unsigned long beg = address;
 966         unsigned long end = address + size;
 967         struct mm_struct *mm = vma->vm_mm;
 968
 969         dir = pgd_offset(mm, address);
 970         flush_cache_range(vma, beg, end);
 971         if (address >= end)
 972                 BUG();
 973
 974         spin_lock(&mm->page_table_lock);
 975         do {
 976                 pmd_t *pmd = pmd_alloc(mm, dir, address);
 977                 error = -ENOMEM;
 978                 if (!pmd)
 979                         break;
 980                 error = zeromap_pmd_range(mm, pmd, address, end - address, prot);
 981                 if (error)
 982                         break;
 983                 address = (address + PGDIR_SIZE) & PGDIR_MASK;
 984                 dir++;
 985         } while (address && (address < end));
 986         /*
 987          * Why flush? zeromap_pte_range has a BUG_ON for !pte_none()
 988          */
 989         flush_tlb_range(vma, beg, end);
 990         spin_unlock(&mm->page_table_lock);
 991         return error;
 992 }
 993
 994 /*
 995  * maps a range of physical memory into the requested pages. the old
 996  * mappings are removed. any references to nonexistent pages results
 997  * in null mappings (currently treated as "copy-on-access")
 998  */
 999 static inline void remap_pte_range(pte_t * pte, unsigned long address, unsigned long size,
1000         unsigned long phys_addr, pgprot_t prot)
1001 {
1002         unsigned long end;
1003         unsigned long pfn;
1004
1005         address &= ~PMD_MASK;
1006         end = address + size;
1007         if (end > PMD_SIZE)
1008                 end = PMD_SIZE;
1009         pfn = phys_addr >> PAGE_SHIFT;
1010         do {
1011                 BUG_ON(!pte_none(*pte));
1012                 if (!pfn_valid(pfn) || PageReserved(pfn_to_page(pfn)))
1013                         set_pte(pte, pfn_pte(pfn, prot));
1014                 address += PAGE_SIZE;
1015                 pfn++;
1016                 pte++;
1017         } while (address && (address < end));
1018 }
1019
1020 static inline int remap_pmd_range(struct mm_struct *mm, pmd_t * pmd, unsigned long address, unsigned long size,
1021         unsigned long phys_addr, pgprot_t prot)
1022 {
1023         unsigned long base, end;
1024
1025         base = address & PGDIR_MASK;
1026         address &= ~PGDIR_MASK;
1027         end = address + size;
1028         if (end > PGDIR_SIZE)
1029                 end = PGDIR_SIZE;
1030         phys_addr -= address;
1031         do {
1032                 pte_t * pte = pte_alloc_map(mm, pmd, base + address);
1033                 if (!pte)
1034                         return -ENOMEM;
1035                 remap_pte_range(pte, base + address, end - address, address + phys_addr, prot);
1036                 pte_unmap(pte);
1037                 address = (address + PMD_SIZE) & PMD_MASK;
1038                 pmd++;
1039         } while (address && (address < end));
1040         return 0;
1041 }
1042
1043 /*  Note: this is only safe if the mm semaphore is held when called. */
1044 int remap_page_range(struct vm_area_struct *vma, unsigned long from, unsigned long phys_addr, unsigned long size, pgprot_t prot)
1045 {
1046         int error = 0;
1047         pgd_t * dir;
1048         unsigned long beg = from;
1049         unsigned long end = from + size;
1050         struct mm_struct *mm = vma->vm_mm;
1051
1052         phys_addr -= from;
1053         dir = pgd_offset(mm, from);
1054         flush_cache_range(vma, beg, end);
1055         if (from >= end)
1056                 BUG();
1057
1058         spin_lock(&mm->page_table_lock);
1059         do {
1060                 pmd_t *pmd = pmd_alloc(mm, dir, from);
1061                 error = -ENOMEM;
1062                 if (!pmd)
1063                         break;
1064                 error = remap_pmd_range(mm, pmd, from, end - from, phys_addr + from, prot);
1065                 if (error)
1066                         break;
1067                 from = (from + PGDIR_SIZE) & PGDIR_MASK;
1068                 dir++;
1069         } while (from && (from < end));
1070         /*
1071          * Why flush? remap_pte_range has a BUG_ON for !pte_none()
1072          */
1073         flush_tlb_range(vma, beg, end);
1074         spin_unlock(&mm->page_table_lock);
1075         return error;
1076 }
1077
1078 EXPORT_SYMBOL(remap_page_range);
1079
1080 /*
1081  * Do pte_mkwrite, but only if the vma says VM_WRITE.  We do this when
1082  * servicing faults for write access.  In the normal case, do always want
1083  * pte_mkwrite.  But get_user_pages can cause write faults for mappings
1084  * that do not have writing enabled, when used by access_process_vm.
1085  */
1086 static inline pte_t maybe_mkwrite(pte_t pte, struct vm_area_struct *vma)
1087 {
1088         if (likely(vma->vm_flags & VM_WRITE))
1089                 pte = pte_mkwrite(pte);
1090         return pte;
1091 }
1092
1093 /*
1094  * We hold the mm semaphore for reading and vma->vm_mm->page_table_lock
1095  */
1096 static inline void break_cow(struct vm_area_struct * vma, struct page * new_page, unsigned long address,
1097                 pte_t *page_table)
1098 {
1099         pte_t entry;
1100
1101         flush_cache_page(vma, address);
1102         entry = maybe_mkwrite(pte_mkdirty(mk_pte(new_page, vma->vm_page_prot)),
1103                               vma);
1104         ptep_establish(vma, address, page_table, entry);
1105         update_mmu_cache(vma, address, entry);
1106 }
1107
1108 /*
1109  * This routine handles present pages, when users try to write
1110  * to a shared page. It is done by copying the page to a new address
1111  * and decrementing the shared-page counter for the old page.
1112  *
1113  * Goto-purists beware: the only reason for goto's here is that it results
1114  * in better assembly code.. The "default" path will see no jumps at all.
1115  *
1116  * Note that this routine assumes that the protection checks have been
1117  * done by the caller (the low-level page fault routine in most cases).
1118  * Thus we can safely just mark it writable once we've done any necessary
1119  * COW.
1120  *
1121  * We also mark the page dirty at this point even though the page will
1122  * change only once the write actually happens. This avoids a few races,
1123  * and potentially makes it more efficient.
1124  *
1125  * We hold the mm semaphore and the page_table_lock on entry and exit
1126  * with the page_table_lock released.
1127  */
1128 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct * vma,
1129         unsigned long address, pte_t *page_table, pmd_t *pmd, pte_t pte)
1130 {
1131         struct page *old_page, *new_page;
1132         unsigned long pfn = pte_pfn(pte);
1133         pte_t entry;
1134
1135         if (unlikely(!pfn_valid(pfn))) {
1136                 /*
1137                  * This should really halt the system so it can be debugged or
1138                  * at least the kernel stops what it's doing before it corrupts
1139                  * data, but for the moment just pretend this is OOM.
1140                  */
1141                 pte_unmap(page_table);
1142                 printk(KERN_ERR "do_wp_page: bogus page at address %08lx\n",
1143                                 address);
1144                 spin_unlock(&mm->page_table_lock);
1145                 return VM_FAULT_OOM;
1146         }
1147         old_page = pfn_to_page(pfn);
1148
1149         if (!TestSetPageLocked(old_page)) {
1150                 int reuse = can_share_swap_page(old_page);
1151                 unlock_page(old_page);
1152                 if (reuse) {
1153                         flush_cache_page(vma, address);
1154                         entry = maybe_mkwrite(pte_mkyoung(pte_mkdirty(pte)),
1155                                               vma);
1156                         ptep_set_access_flags(vma, address, page_table, entry, 1);
1157                         update_mmu_cache(vma, address, entry);
1158                         pte_unmap(page_table);
1159                         spin_unlock(&mm->page_table_lock);
1160                         return VM_FAULT_MINOR;
1161                 }
1162         }
1163         pte_unmap(page_table);
1164
1165         /*
1166          * Ok, we need to copy. Oh, well..
1167          */
1168         if (!PageReserved(old_page))
1169                 page_cache_get(old_page);
1170         spin_unlock(&mm->page_table_lock);
1171
1172         if (unlikely(anon_vma_prepare(vma)))
1173                 goto no_new_page;
1174         new_page = alloc_page_vma(GFP_HIGHUSER, vma, address);
1175         if (!new_page)
1176                 goto no_new_page;
1177         copy_cow_page(old_page,new_page,address);
1178
1179         /*
1180          * Re-check the pte - we dropped the lock
1181          */
1182         spin_lock(&mm->page_table_lock);
1183         page_table = pte_offset_map(pmd, address);
1184         if (likely(pte_same(*page_table, pte))) {
1185                 if (PageReserved(old_page))
1186                         // ++mm->rss;
1187                         vx_rsspages_inc(mm);
1188                 else
1189                         page_remove_rmap(old_page);
1190                 break_cow(vma, new_page, address, page_table);
1191                 lru_cache_add_active(new_page);
1192                 page_add_anon_rmap(new_page, vma, address);
1193
1194                 /* Free the old page.. */
1195                 new_page = old_page;
1196         }
1197         pte_unmap(page_table);
1198         page_cache_release(new_page);
1199         page_cache_release(old_page);
1200         spin_unlock(&mm->page_table_lock);
1201         return VM_FAULT_MINOR;
1202
1203 no_new_page:
1204         page_cache_release(old_page);
1205         return VM_FAULT_OOM;
1206 }
1207
1208 /*
1209  * Helper function for unmap_mapping_range().
1210  */
1211 static inline void unmap_mapping_range_list(struct prio_tree_root *root,
1212                                             struct zap_details *details)
1213 {
1214         struct vm_area_struct *vma = NULL;
1215         struct prio_tree_iter iter;
1216         pgoff_t vba, vea, zba, zea;
1217
1218         while ((vma = vma_prio_tree_next(vma, root, &iter,
1219                         details->first_index, details->last_index)) != NULL) {
1220                 vba = vma->vm_pgoff;
1221                 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
1222                 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
1223                 zba = details->first_index;
1224                 if (zba < vba)
1225                         zba = vba;
1226                 zea = details->last_index;
1227                 if (zea > vea)
1228                         zea = vea;
1229                 zap_page_range(vma,
1230                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
1231                         (zea - zba + 1) << PAGE_SHIFT, details);
1232         }
1233 }
1234
1235 /**
1236  * unmap_mapping_range - unmap the portion of all mmaps
1237  * in the specified address_space corresponding to the specified
1238  * page range in the underlying file.
1239  * @address_space: the address space containing mmaps to be unmapped.
1240  * @holebegin: byte in first page to unmap, relative to the start of
1241  * the underlying file.  This will be rounded down to a PAGE_SIZE
1242  * boundary.  Note that this is different from vmtruncate(), which
1243  * must keep the partial page.  In contrast, we must get rid of
1244  * partial pages.
1245  * @holelen: size of prospective hole in bytes.  This will be rounded
1246  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
1247  * end of the file.
1248  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
1249  * but 0 when invalidating pagecache, don't throw away private data.
1250  */
1251 void unmap_mapping_range(struct address_space *mapping,
1252                 loff_t const holebegin, loff_t const holelen, int even_cows)
1253 {
1254         struct zap_details details;
1255         pgoff_t hba = holebegin >> PAGE_SHIFT;
1256         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1257
1258         /* Check for overflow. */
1259         if (sizeof(holelen) > sizeof(hlen)) {
1260                 long long holeend =
1261                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1262                 if (holeend & ~(long long)ULONG_MAX)
1263                         hlen = ULONG_MAX - hba + 1;
1264         }
1265
1266         details.check_mapping = even_cows? NULL: mapping;
1267         details.nonlinear_vma = NULL;
1268         details.first_index = hba;
1269         details.last_index = hba + hlen - 1;
1270         details.atomic = 1;     /* A spinlock is held */
1271         if (details.last_index < details.first_index)
1272                 details.last_index = ULONG_MAX;
1273
1274         spin_lock(&mapping->i_mmap_lock);
1275         /* Protect against page fault */
1276         atomic_inc(&mapping->truncate_count);
1277
1278         if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
1279                 unmap_mapping_range_list(&mapping->i_mmap, &details);
1280
1281         /*
1282          * In nonlinear VMAs there is no correspondence between virtual address
1283          * offset and file offset.  So we must perform an exhaustive search
1284          * across *all* the pages in each nonlinear VMA, not just the pages
1285          * whose virtual address lies outside the file truncation point.
1286          */
1287         if (unlikely(!list_empty(&mapping->i_mmap_nonlinear))) {
1288                 struct vm_area_struct *vma;
1289                 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
1290                                                 shared.vm_set.list) {
1291                         details.nonlinear_vma = vma;
1292                         zap_page_range(vma, vma->vm_start,
1293                                 vma->vm_end - vma->vm_start, &details);
1294                 }
1295         }
1296         spin_unlock(&mapping->i_mmap_lock);
1297 }
1298 EXPORT_SYMBOL(unmap_mapping_range);
1299
1300 /*
1301  * Handle all mappings that got truncated by a "truncate()"
1302  * system call.
1303  *
1304  * NOTE! We have to be ready to update the memory sharing
1305  * between the file and the memory map for a potential last
1306  * incomplete page.  Ugly, but necessary.
1307  */
1308 int vmtruncate(struct inode * inode, loff_t offset)
1309 {
1310         struct address_space *mapping = inode->i_mapping;
1311         unsigned long limit;
1312
1313         if (inode->i_size < offset)
1314                 goto do_expand;
1315         /*
1316          * truncation of in-use swapfiles is disallowed - it would cause
1317          * subsequent swapout to scribble on the now-freed blocks.
1318          */
1319         if (IS_SWAPFILE(inode))
1320                 goto out_busy;
1321         i_size_write(inode, offset);
1322         unmap_mapping_range(mapping, offset + PAGE_SIZE - 1, 0, 1);
1323         truncate_inode_pages(mapping, offset);
1324         goto out_truncate;
1325
1326 do_expand:
1327         limit = current->rlim[RLIMIT_FSIZE].rlim_cur;
1328         if (limit != RLIM_INFINITY && offset > limit)
1329                 goto out_sig;
1330         if (offset > inode->i_sb->s_maxbytes)
1331                 goto out_big;
1332         i_size_write(inode, offset);
1333
1334 out_truncate:
1335         if (inode->i_op && inode->i_op->truncate)
1336                 inode->i_op->truncate(inode);
1337         return 0;
1338 out_sig:
1339         send_sig(SIGXFSZ, current, 0);
1340 out_big:
1341         return -EFBIG;
1342 out_busy:
1343         return -ETXTBSY;
1344 }
1345
1346 EXPORT_SYMBOL(vmtruncate);
1347
1348 /*
1349  * Primitive swap readahead code. We simply read an aligned block of
1350  * (1 << page_cluster) entries in the swap area. This method is chosen
1351  * because it doesn't cost us any seek time.  We also make sure to queue
1352  * the 'original' request together with the readahead ones...
1353  *
1354  * This has been extended to use the NUMA policies from the mm triggering
1355  * the readahead.
1356  *
1357  * Caller must hold down_read on the vma->vm_mm if vma is not NULL.
1358  */
1359 void swapin_readahead(swp_entry_t entry, unsigned long addr,struct vm_area_struct *vma)
1360 {
1361 #ifdef CONFIG_NUMA
1362         struct vm_area_struct *next_vma = vma ? vma->vm_next : NULL;
1363 #endif
1364         int i, num;
1365         struct page *new_page;
1366         unsigned long offset;
1367
1368         /*
1369          * Get the number of handles we should do readahead io to.
1370          */
1371         num = valid_swaphandles(entry, &offset);
1372         for (i = 0; i < num; offset++, i++) {
1373                 /* Ok, do the async read-ahead now */
1374                 new_page = read_swap_cache_async(swp_entry(swp_type(entry),
1375                                                            offset), vma, addr);
1376                 if (!new_page)
1377                         break;
1378                 page_cache_release(new_page);
1379 #ifdef CONFIG_NUMA
1380                 /*
1381                  * Find the next applicable VMA for the NUMA policy.
1382                  */
1383                 addr += PAGE_SIZE;
1384                 if (addr == 0)
1385                         vma = NULL;
1386                 if (vma) {
1387                         if (addr >= vma->vm_end) {
1388                                 vma = next_vma;
1389                                 next_vma = vma ? vma->vm_next : NULL;
1390                         }
1391                         if (vma && addr < vma->vm_start)
1392                                 vma = NULL;
1393                 } else {
1394                         if (next_vma && addr >= next_vma->vm_start) {
1395                                 vma = next_vma;
1396                                 next_vma = vma->vm_next;
1397                         }
1398                 }
1399 #endif
1400         }
1401         lru_add_drain();        /* Push any new pages onto the LRU now */
1402 }
1403
1404 /*
1405  * We hold the mm semaphore and the page_table_lock on entry and
1406  * should release the pagetable lock on exit..
1407  */
1408 static int do_swap_page(struct mm_struct * mm,
1409         struct vm_area_struct * vma, unsigned long address,
1410         pte_t *page_table, pmd_t *pmd, pte_t orig_pte, int write_access)
1411 {
1412         struct page *page;
1413         swp_entry_t entry = pte_to_swp_entry(orig_pte);
1414         pte_t pte;
1415         int ret = VM_FAULT_MINOR;
1416
1417         pte_unmap(page_table);
1418         spin_unlock(&mm->page_table_lock);
1419         page = lookup_swap_cache(entry);
1420         if (!page) {
1421                 swapin_readahead(entry, address, vma);
1422                 page = read_swap_cache_async(entry, vma, address);
1423                 if (!page) {
1424                         /*
1425                          * Back out if somebody else faulted in this pte while
1426                          * we released the page table lock.
1427                          */
1428                         spin_lock(&mm->page_table_lock);
1429                         page_table = pte_offset_map(pmd, address);
1430                         if (likely(pte_same(*page_table, orig_pte)))
1431                                 ret = VM_FAULT_OOM;
1432                         else
1433                                 ret = VM_FAULT_MINOR;
1434                         pte_unmap(page_table);
1435                         spin_unlock(&mm->page_table_lock);
1436                         goto out;
1437                 }
1438
1439                 /* Had to read the page from swap area: Major fault */
1440                 ret = VM_FAULT_MAJOR;
1441                 inc_page_state(pgmajfault);
1442                 grab_swap_token();
1443         }
1444
1445         if (!vx_rsspages_avail(mm, 1)) {
1446                 ret = VM_FAULT_OOM;
1447                 goto out;
1448         }
1449         mark_page_accessed(page);
1450         lock_page(page);
1451
1452         /*
1453          * Back out if somebody else faulted in this pte while we
1454          * released the page table lock.
1455          */
1456         spin_lock(&mm->page_table_lock);
1457         page_table = pte_offset_map(pmd, address);
1458         if (unlikely(!pte_same(*page_table, orig_pte))) {
1459                 pte_unmap(page_table);
1460                 spin_unlock(&mm->page_table_lock);
1461                 unlock_page(page);
1462                 page_cache_release(page);
1463                 ret = VM_FAULT_MINOR;
1464                 goto out;
1465         }
1466
1467         /* The page isn't present yet, go ahead with the fault. */
1468
1469         swap_free(entry);
1470         if (vm_swap_full())
1471                 remove_exclusive_swap_page(page);
1472
1473         // mm->rss++;
1474         vx_rsspages_inc(mm);
1475         pte = mk_pte(page, vma->vm_page_prot);
1476         if (write_access && can_share_swap_page(page)) {
1477                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
1478                 write_access = 0;
1479         }
1480         unlock_page(page);
1481
1482         flush_icache_page(vma, page);
1483         set_pte(page_table, pte);
1484         page_add_anon_rmap(page, vma, address);
1485
1486         if (write_access) {
1487                 if (do_wp_page(mm, vma, address,
1488                                 page_table, pmd, pte) == VM_FAULT_OOM)
1489                         ret = VM_FAULT_OOM;
1490                 goto out;
1491         }
1492
1493         /* No need to invalidate - it was non-present before */
1494         update_mmu_cache(vma, address, pte);
1495         pte_unmap(page_table);
1496         spin_unlock(&mm->page_table_lock);
1497 out:
1498         return ret;
1499 }
1500
1501 /*
1502  * We are called with the MM semaphore and page_table_lock
1503  * spinlock held to protect against concurrent faults in
1504  * multithreaded programs.
1505  */
1506 static int
1507 do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
1508                 pte_t *page_table, pmd_t *pmd, int write_access,
1509                 unsigned long addr)
1510 {
1511         pte_t entry;
1512         struct page * page = ZERO_PAGE(addr);
1513
1514         /* Read-only mapping of ZERO_PAGE. */
1515         entry = pte_wrprotect(mk_pte(ZERO_PAGE(addr), vma->vm_page_prot));
1516
1517         /* ..except if it's a write access */
1518         if (write_access) {
1519                 /* Allocate our own private page. */
1520                 pte_unmap(page_table);
1521                 spin_unlock(&mm->page_table_lock);
1522
1523                 if (unlikely(anon_vma_prepare(vma)))
1524                         goto no_mem;
1525                 if (!vx_rsspages_avail(mm, 1))
1526                         goto no_mem;
1527
1528                 page = alloc_page_vma(GFP_HIGHUSER, vma, addr);
1529                 if (!page)
1530                         goto no_mem;
1531                 clear_user_highpage(page, addr);
1532
1533                 spin_lock(&mm->page_table_lock);
1534                 page_table = pte_offset_map(pmd, addr);
1535
1536                 if (!pte_none(*page_table)) {
1537                         pte_unmap(page_table);
1538                         page_cache_release(page);
1539                         spin_unlock(&mm->page_table_lock);
1540                         goto out;
1541                 }
1542                 // mm->rss++;
1543                 vx_rsspages_inc(mm);
1544                 entry = maybe_mkwrite(pte_mkdirty(mk_pte(page,
1545                                                          vma->vm_page_prot)),
1546                                       vma);
1547                 lru_cache_add_active(page);
1548                 mark_page_accessed(page);
1549                 page_add_anon_rmap(page, vma, addr);
1550         }
1551
1552         set_pte(page_table, entry);
1553         pte_unmap(page_table);
1554
1555         /* No need to invalidate - it was non-present before */
1556         update_mmu_cache(vma, addr, entry);
1557         spin_unlock(&mm->page_table_lock);
1558 out:
1559         return VM_FAULT_MINOR;
1560 no_mem:
1561         return VM_FAULT_OOM;
1562 }
1563
1564 /*
1565  * do_no_page() tries to create a new page mapping. It aggressively
1566  * tries to share with existing pages, but makes a separate copy if
1567  * the "write_access" parameter is true in order to avoid the next
1568  * page fault.
1569  *
1570  * As this is called only for pages that do not currently exist, we
1571  * do not need to flush old virtual caches or the TLB.
1572  *
1573  * This is called with the MM semaphore held and the page table
1574  * spinlock held. Exit with the spinlock released.
1575  */
1576 static int
1577 do_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
1578         unsigned long address, int write_access, pte_t *page_table, pmd_t *pmd)
1579 {
1580         struct page * new_page;
1581         struct address_space *mapping = NULL;
1582         pte_t entry;
1583         int sequence = 0;
1584         int ret = VM_FAULT_MINOR;
1585         int anon = 0;
1586
1587         if (!vma->vm_ops || !vma->vm_ops->nopage)
1588                 return do_anonymous_page(mm, vma, page_table,
1589                                         pmd, write_access, address);
1590         pte_unmap(page_table);
1591         spin_unlock(&mm->page_table_lock);
1592
1593         if (vma->vm_file) {
1594                 mapping = vma->vm_file->f_mapping;
1595                 sequence = atomic_read(&mapping->truncate_count);
1596         }
1597         smp_rmb();  /* Prevent CPU from reordering lock-free ->nopage() */
1598 retry:
1599         new_page = vma->vm_ops->nopage(vma, address & PAGE_MASK, &ret);
1600
1601         /* no page was available -- either SIGBUS or OOM */
1602         if (new_page == NOPAGE_SIGBUS)
1603                 return VM_FAULT_SIGBUS;
1604         if (new_page == NOPAGE_OOM)
1605                 return VM_FAULT_OOM;
1606         if (!vx_rsspages_avail(mm, 1))
1607                 return VM_FAULT_OOM;
1608
1609         /*
1610          * Should we do an early C-O-W break?
1611          */
1612         if (write_access && !(vma->vm_flags & VM_SHARED)) {
1613                 struct page *page;
1614
1615                 if (unlikely(anon_vma_prepare(vma)))
1616                         goto oom;
1617                 page = alloc_page_vma(GFP_HIGHUSER, vma, address);
1618                 if (!page)
1619                         goto oom;
1620                 copy_user_highpage(page, new_page, address);
1621                 page_cache_release(new_page);
1622                 new_page = page;
1623                 anon = 1;
1624         }
1625
1626         spin_lock(&mm->page_table_lock);
1627         /*
1628          * For a file-backed vma, someone could have truncated or otherwise
1629          * invalidated this page.  If unmap_mapping_range got called,
1630          * retry getting the page.
1631          */
1632         if (mapping &&
1633               (unlikely(sequence != atomic_read(&mapping->truncate_count)))) {
1634                 sequence = atomic_read(&mapping->truncate_count);
1635                 spin_unlock(&mm->page_table_lock);
1636                 page_cache_release(new_page);
1637                 goto retry;
1638         }
1639         page_table = pte_offset_map(pmd, address);
1640
1641         /*
1642          * This silly early PAGE_DIRTY setting removes a race
1643          * due to the bad i386 page protection. But it's valid
1644          * for other architectures too.
1645          *
1646          * Note that if write_access is true, we either now have
1647          * an exclusive copy of the page, or this is a shared mapping,
1648          * so we can make it writable and dirty to avoid having to
1649          * handle that later.
1650          */
1651         /* Only go through if we didn't race with anybody else... */
1652         if (pte_none(*page_table)) {
1653                 if (!PageReserved(new_page))
1654                         ++mm->rss;
1655                 flush_icache_page(vma, new_page);
1656                 entry = mk_pte(new_page, vma->vm_page_prot);
1657                 if (write_access)
1658                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1659                 set_pte(page_table, entry);
1660                 if (anon) {
1661                         lru_cache_add_active(new_page);
1662                         page_add_anon_rmap(new_page, vma, address);
1663                 } else
1664                         page_add_file_rmap(new_page);
1665                 pte_unmap(page_table);
1666         } else {
1667                 /* One of our sibling threads was faster, back out. */
1668                 pte_unmap(page_table);
1669                 page_cache_release(new_page);
1670                 spin_unlock(&mm->page_table_lock);
1671                 goto out;
1672         }
1673
1674         /* no need to invalidate: a not-present page shouldn't be cached */
1675         update_mmu_cache(vma, address, entry);
1676         spin_unlock(&mm->page_table_lock);
1677 out:
1678         return ret;
1679 oom:
1680         page_cache_release(new_page);
1681         ret = VM_FAULT_OOM;
1682         goto out;
1683 }
1684
1685 /*
1686  * Fault of a previously existing named mapping. Repopulate the pte
1687  * from the encoded file_pte if possible. This enables swappable
1688  * nonlinear vmas.
1689  */
1690 static int do_file_page(struct mm_struct * mm, struct vm_area_struct * vma,
1691         unsigned long address, int write_access, pte_t *pte, pmd_t *pmd)
1692 {
1693         unsigned long pgoff;
1694         int err;
1695
1696         BUG_ON(!vma->vm_ops || !vma->vm_ops->nopage);
1697         /*
1698          * Fall back to the linear mapping if the fs does not support
1699          * ->populate:
1700          */
1701         if (!vma->vm_ops || !vma->vm_ops->populate ||
1702                         (write_access && !(vma->vm_flags & VM_SHARED))) {
1703                 pte_clear(pte);
1704                 return do_no_page(mm, vma, address, write_access, pte, pmd);
1705         }
1706
1707         pgoff = pte_to_pgoff(*pte);
1708
1709         pte_unmap(pte);
1710         spin_unlock(&mm->page_table_lock);
1711
1712         err = vma->vm_ops->populate(vma, address & PAGE_MASK, PAGE_SIZE, vma->vm_page_prot, pgoff, 0);
1713         if (err == -ENOMEM)
1714                 return VM_FAULT_OOM;
1715         if (err)
1716                 return VM_FAULT_SIGBUS;
1717         return VM_FAULT_MAJOR;
1718 }
1719
1720 /*
1721  * These routines also need to handle stuff like marking pages dirty
1722  * and/or accessed for architectures that don't do it in hardware (most
1723  * RISC architectures).  The early dirtying is also good on the i386.
1724  *
1725  * There is also a hook called "update_mmu_cache()" that architectures
1726  * with external mmu caches can use to update those (ie the Sparc or
1727  * PowerPC hashed page tables that act as extended TLBs).
1728  *
1729  * Note the "page_table_lock". It is to protect against kswapd removing
1730  * pages from under us. Note that kswapd only ever _removes_ pages, never
1731  * adds them. As such, once we have noticed that the page is not present,
1732  * we can drop the lock early.
1733  *
1734  * The adding of pages is protected by the MM semaphore (which we hold),
1735  * so we don't need to worry about a page being suddenly been added into
1736  * our VM.
1737  *
1738  * We enter with the pagetable spinlock held, we are supposed to
1739  * release it when done.
1740  */
1741 static inline int handle_pte_fault(struct mm_struct *mm,
1742         struct vm_area_struct * vma, unsigned long address,
1743         int write_access, pte_t *pte, pmd_t *pmd)
1744 {
1745         pte_t entry;
1746
1747         entry = *pte;
1748         if (!pte_present(entry)) {
1749                 /*
1750                  * If it truly wasn't present, we know that kswapd
1751                  * and the PTE updates will not touch it later. So
1752                  * drop the lock.
1753                  */
1754                 if (pte_none(entry))
1755                         return do_no_page(mm, vma, address, write_access, pte, pmd);
1756                 if (pte_file(entry))
1757                         return do_file_page(mm, vma, address, write_access, pte, pmd);
1758                 return do_swap_page(mm, vma, address, pte, pmd, entry, write_access);
1759         }
1760
1761         if (write_access) {
1762                 if (!pte_write(entry))
1763                         return do_wp_page(mm, vma, address, pte, pmd, entry);
1764
1765                 entry = pte_mkdirty(entry);
1766         }
1767         entry = pte_mkyoung(entry);
1768         ptep_set_access_flags(vma, address, pte, entry, write_access);
1769         update_mmu_cache(vma, address, entry);
1770         pte_unmap(pte);
1771         spin_unlock(&mm->page_table_lock);
1772         return VM_FAULT_MINOR;
1773 }
1774
1775 /*
1776  * By the time we get here, we already hold the mm semaphore
1777  */
1778 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct * vma,
1779         unsigned long address, int write_access)
1780 {
1781         pgd_t *pgd;
1782         pmd_t *pmd;
1783
1784         __set_current_state(TASK_RUNNING);
1785         pgd = pgd_offset(mm, address);
1786
1787         inc_page_state(pgfault);
1788
1789         if (is_vm_hugetlb_page(vma))
1790                 return VM_FAULT_SIGBUS; /* mapping truncation does this. */
1791
1792         /*
1793          * We need the page table lock to synchronize with kswapd
1794          * and the SMP-safe atomic PTE updates.
1795          */
1796         set_delay_flag(current,PF_MEMIO);
1797         spin_lock(&mm->page_table_lock);
1798         pmd = pmd_alloc(mm, pgd, address);
1799
1800         if (pmd) {
1801                 pte_t * pte = pte_alloc_map(mm, pmd, address);
1802                 if (pte) {
1803                         int rc = handle_pte_fault(mm, vma, address, write_access, pte, pmd);
1804                         clear_delay_flag(current,PF_MEMIO);
1805                         return rc;
1806                 }
1807         }
1808         spin_unlock(&mm->page_table_lock);
1809         clear_delay_flag(current,PF_MEMIO);
1810         return VM_FAULT_OOM;
1811 }
1812
1813 /*
1814  * Allocate page middle directory.
1815  *
1816  * We've already handled the fast-path in-line, and we own the
1817  * page table lock.
1818  *
1819  * On a two-level page table, this ends up actually being entirely
1820  * optimized away.
1821  */
1822 pmd_t fastcall *__pmd_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
1823 {
1824         pmd_t *new;
1825
1826         spin_unlock(&mm->page_table_lock);
1827         new = pmd_alloc_one(mm, address);
1828         spin_lock(&mm->page_table_lock);
1829         if (!new)
1830                 return NULL;
1831
1832         /*
1833          * Because we dropped the lock, we should re-check the
1834          * entry, as somebody else could have populated it..
1835          */
1836         if (pgd_present(*pgd)) {
1837                 pmd_free(new);
1838                 goto out;
1839         }
1840         pgd_populate(mm, pgd, new);
1841 out:
1842         return pmd_offset(pgd, address);
1843 }
1844
1845 int make_pages_present(unsigned long addr, unsigned long end)
1846 {
1847         int ret, len, write;
1848         struct vm_area_struct * vma;
1849
1850         vma = find_vma(current->mm, addr);
1851         write = (vma->vm_flags & VM_WRITE) != 0;
1852         if (addr >= end)
1853                 BUG();
1854         if (end > vma->vm_end)
1855                 BUG();
1856         len = (end+PAGE_SIZE-1)/PAGE_SIZE-addr/PAGE_SIZE;
1857         ret = get_user_pages(current, current->mm, addr,
1858                         len, write, 0, NULL, NULL);
1859         if (ret < 0)
1860                 return ret;
1861         return ret == len ? 0 : -1;
1862 }
1863
1864 /*
1865  * Map a vmalloc()-space virtual address to the physical page.
1866  */
1867 struct page * vmalloc_to_page(void * vmalloc_addr)
1868 {
1869         unsigned long addr = (unsigned long) vmalloc_addr;
1870         struct page *page = NULL;
1871         pgd_t *pgd = pgd_offset_k(addr);
1872         pmd_t *pmd;
1873         pte_t *ptep, pte;
1874
1875         if (!pgd_none(*pgd)) {
1876                 pmd = pmd_offset(pgd, addr);
1877                 if (!pmd_none(*pmd)) {
1878                         preempt_disable();
1879                         ptep = pte_offset_map(pmd, addr);
1880                         pte = *ptep;
1881                         if (pte_present(pte))
1882                                 page = pte_page(pte);
1883                         pte_unmap(ptep);
1884                         preempt_enable();
1885                 }
1886         }
1887         return page;
1888 }
1889
1890 EXPORT_SYMBOL(vmalloc_to_page);
1891
1892 #if !defined(CONFIG_ARCH_GATE_AREA)
1893
1894 #if defined(AT_SYSINFO_EHDR)
1895 struct vm_area_struct gate_vma;
1896
1897 static int __init gate_vma_init(void)
1898 {
1899         gate_vma.vm_mm = NULL;
1900         gate_vma.vm_start = FIXADDR_USER_START;
1901         gate_vma.vm_end = FIXADDR_USER_END;
1902         gate_vma.vm_page_prot = PAGE_READONLY;
1903         gate_vma.vm_flags = 0;
1904         return 0;
1905 }
1906 __initcall(gate_vma_init);
1907 #endif
1908
1909 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
1910 {
1911 #ifdef AT_SYSINFO_EHDR
1912         return &gate_vma;
1913 #else
1914         return NULL;
1915 #endif
1916 }
1917
1918 int in_gate_area(struct task_struct *task, unsigned long addr)
1919 {
1920 #ifdef AT_SYSINFO_EHDR
1921         if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
1922                 return 1;
1923 #endif
1924         return 0;
1925 }
1926
1927 #endif