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