mm/memory.c

/* */
This source file includes following definitions.
oom
free_one_pte
free_one_pmd
free_one_pgd
clear_page_tables
free_page_tables
clone_page_tables
copy_one_pte
copy_one_pmd
copy_one_pgd
copy_page_tables
forget_pte
unmap_pte_range
unmap_pmd_range
unmap_page_range
zeromap_pte_range
zeromap_pmd_range
zeromap_page_range
remap_pte_range
remap_pmd_range
remap_page_range
put_page
put_dirty_page
do_wp_page
verify_area
get_empty_page
try_to_share
share_page
get_empty_pgtable
do_swap_page
do_no_page
   1 #define THREE_LEVEL
   2 /*
   3  *  linux/mm/memory.c
   4  *
   5  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
   6  */
   7 
   8 /*
   9  * demand-loading started 01.12.91 - seems it is high on the list of
  10  * things wanted, and it should be easy to implement. - Linus
  11  */
  12 
  13 /*
  14  * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
  15  * pages started 02.12.91, seems to work. - Linus.
  16  *
  17  * Tested sharing by executing about 30 /bin/sh: under the old kernel it
  18  * would have taken more than the 6M I have free, but it worked well as
  19  * far as I could see.
  20  *
  21  * Also corrected some "invalidate()"s - I wasn't doing enough of them.
  22  */
  23 
  24 /*
  25  * Real VM (paging to/from disk) started 18.12.91. Much more work and
  26  * thought has to go into this. Oh, well..
  27  * 19.12.91  -  works, somewhat. Sometimes I get faults, don't know why.
  28  *              Found it. Everything seems to work now.
  29  * 20.12.91  -  Ok, making the swap-device changeable like the root.
  30  */
  31 
  32 /*
  33  * 05.04.94  -  Multi-page memory management added for v1.1.
  34  *              Idea by Alex Bligh (alex@cconcepts.co.uk)
  35  */
  36 
  37 #include <linux/config.h>
  38 #include <linux/signal.h>
  39 #include <linux/sched.h>
  40 #include <linux/head.h>
  41 #include <linux/kernel.h>
  42 #include <linux/errno.h>
  43 #include <linux/string.h>
  44 #include <linux/types.h>
  45 #include <linux/ptrace.h>
  46 #include <linux/mman.h>
  47 #include <linux/mm.h>
  48 
  49 #include <asm/system.h>
  50 #include <asm/segment.h>
  51 #include <asm/pgtable.h>
  52 
  53 unsigned long high_memory = 0;
  54 
  55 /*
  56  * The free_area_list arrays point to the queue heads of the free areas
  57  * of different sizes
  58  */
  59 int nr_swap_pages = 0;
  60 int nr_free_pages = 0;
  61 struct mem_list free_area_list[NR_MEM_LISTS];
  62 unsigned char * free_area_map[NR_MEM_LISTS];
  63 
  64 #define copy_page(from,to) memcpy((void *) to, (void *) from, PAGE_SIZE)
  65 
  66 #define USER_PTRS_PER_PGD (TASK_SIZE / PGDIR_SIZE)
  67 
  68 mem_map_t * mem_map = NULL;
  69 
  70 /*
  71  * oom() prints a message (so that the user knows why the process died),
  72  * and gives the process an untrappable SIGKILL.
  73  */
  74 void oom(struct task_struct * task)
     /*  */
  75 {
  76         printk("\nOut of memory for %s.\n", current->comm);
  77         task->sigaction[SIGKILL-1].sa_handler = NULL;
  78         task->blocked &= ~(1<<(SIGKILL-1));
  79         send_sig(SIGKILL,task,1);
  80 }
  81 
  82 static inline void free_one_pte(pte_t * page_table)
     /*  */
  83 {
  84         pte_t page = *page_table;
  85 
  86         if (pte_none(page))
  87                 return;
  88         pte_clear(page_table);
  89         if (!pte_present(page)) {
  90                 swap_free(pte_val(page));
  91                 return;
  92         }
  93         free_page(pte_page(page));
  94         return;
  95 }
  96 
  97 static inline void free_one_pmd(pmd_t * dir)
     /*  */
  98 {
  99         int j;
 100         pte_t * pte;
 101 
 102         if (pmd_none(*dir))
 103                 return;
 104         if (pmd_bad(*dir)) {
 105                 printk("free_one_pmd: bad directory entry %08lx\n", pmd_val(*dir));
 106                 pmd_clear(dir);
 107                 return;
 108         }
 109         pte = pte_offset(dir, 0);
 110         pmd_clear(dir);
 111         if (pte_inuse(pte)) {
 112                 pte_free(pte);
 113                 return;
 114         }
 115         for (j = 0; j < PTRS_PER_PTE ; j++)
 116                 free_one_pte(pte+j);
 117         pte_free(pte);
 118 }
 119 
 120 static inline void free_one_pgd(pgd_t * dir)
     /*  */
 121 {
 122         int j;
 123         pmd_t * pmd;
 124 
 125         if (pgd_none(*dir))
 126                 return;
 127         if (pgd_bad(*dir)) {
 128                 printk("free_one_pgd: bad directory entry %08lx\n", pgd_val(*dir));
 129                 pgd_clear(dir);
 130                 return;
 131         }
 132         pmd = pmd_offset(dir, 0);
 133         pgd_clear(dir);
 134         if (pmd_inuse(pmd)) {
 135                 pmd_free(pmd);
 136                 return;
 137         }
 138         for (j = 0; j < PTRS_PER_PMD ; j++)
 139                 free_one_pmd(pmd+j);
 140         pmd_free(pmd);
 141 }
 142         
 143 
 144 /*
 145  * This function clears all user-level page tables of a process - this
 146  * is needed by execve(), so that old pages aren't in the way. Note that
 147  * unlike 'free_page_tables()', this function still leaves a valid
 148  * page-table-tree in memory: it just removes the user pages. The two
 149  * functions are similar, but there is a fundamental difference.
 150  */
 151 void clear_page_tables(struct task_struct * tsk)
     /*  */
 152 {
 153         int i;
 154         pgd_t * page_dir;
 155 
 156         if (!tsk)
 157                 return;
 158         if (tsk == task[0])
 159                 panic("task[0] (swapper) doesn't support exec()\n");
 160         page_dir = pgd_offset(tsk, 0);
 161         if (!page_dir || page_dir == swapper_pg_dir) {
 162                 printk("Trying to clear kernel page-directory: not good\n");
 163                 return;
 164         }
 165         if (pgd_inuse(page_dir)) {
 166                 pgd_t * new_pg;
 167 
 168                 if (!(new_pg = pgd_alloc())) {
 169                         oom(tsk);
 170                         return;
 171                 }
 172                 for (i = USER_PTRS_PER_PGD ; i < PTRS_PER_PGD ; i++)
 173                         new_pg[i] = page_dir[i];
 174                 SET_PAGE_DIR(tsk, new_pg);
 175                 pgd_free(page_dir);
 176                 return;
 177         }
 178         for (i = 0 ; i < USER_PTRS_PER_PGD ; i++)
 179                 free_one_pgd(page_dir + i);
 180         invalidate();
 181         return;
 182 }
 183 
 184 /*
 185  * This function frees up all page tables of a process when it exits.
 186  */
 187 void free_page_tables(struct task_struct * tsk)
     /*  */
 188 {
 189         int i;
 190         pgd_t * page_dir;
 191 
 192         if (!tsk)
 193                 return;
 194         if (tsk == task[0]) {
 195                 printk("task[0] (swapper) killed: unable to recover\n");
 196                 panic("Trying to free up swapper memory space");
 197         }
 198         page_dir = pgd_offset(tsk, 0);
 199         if (!page_dir || page_dir == swapper_pg_dir) {
 200                 printk("Trying to free kernel page-directory: not good\n");
 201                 return;
 202         }
 203         SET_PAGE_DIR(tsk, swapper_pg_dir);
 204         if (pgd_inuse(page_dir)) {
 205                 pgd_free(page_dir);
 206                 return;
 207         }
 208         for (i = 0 ; i < PTRS_PER_PGD ; i++)
 209                 free_one_pgd(page_dir + i);
 210         pgd_free(page_dir);
 211         invalidate();
 212 }
 213 
 214 /*
 215  * clone_page_tables() clones the page table for a process - both
 216  * processes will have the exact same pages in memory. There are
 217  * probably races in the memory management with cloning, but we'll
 218  * see..
 219  */
 220 int clone_page_tables(struct task_struct * tsk)
     /*  */
 221 {
 222         pgd_t * pg_dir;
 223 
 224         pg_dir = pgd_offset(current, 0);
 225         pgd_reuse(pg_dir);
 226         SET_PAGE_DIR(tsk, pg_dir);
 227         return 0;
 228 }
 229 
 230 static inline void copy_one_pte(pte_t * old_pte, pte_t * new_pte)
     /*  */
 231 {
 232         pte_t pte = *old_pte;
 233 
 234         if (pte_none(pte))
 235                 return;
 236         if (!pte_present(pte)) {
 237                 swap_duplicate(pte_val(pte));
 238                 *new_pte = pte;
 239                 return;
 240         }
 241         if (pte_page(pte) > high_memory || (mem_map[MAP_NR(pte_page(pte))] & MAP_PAGE_RESERVED)) {
 242                 *new_pte = pte;
 243                 return;
 244         }
 245         if (pte_cow(pte))
 246                 pte = pte_wrprotect(pte);
 247         if (delete_from_swap_cache(pte_page(pte)))
 248                 pte = pte_mkdirty(pte);
 249         *new_pte = pte;
 250         *old_pte = pte;
 251         mem_map[MAP_NR(pte_page(pte))]++;
 252 }
 253 
 254 static inline int copy_one_pmd(pmd_t * old_pmd, pmd_t * new_pmd)
     /*  */
 255 {
 256         int j;
 257         pte_t *old_pte, *new_pte;
 258 
 259         if (pmd_none(*old_pmd))
 260                 return 0;
 261         if (pmd_bad(*old_pmd)) {
 262                 printk("copy_one_pmd: bad page table: probable memory corruption\n");
 263                 pmd_clear(old_pmd);
 264                 return 0;
 265         }
 266         old_pte = pte_offset(old_pmd, 0);
 267         if (pte_inuse(old_pte)) {
 268                 pte_reuse(old_pte);
 269                 *new_pmd = *old_pmd;
 270                 return 0;
 271         }
 272         new_pte = pte_alloc(new_pmd, 0);
 273         if (!new_pte)
 274                 return -ENOMEM;
 275         for (j = 0 ; j < PTRS_PER_PTE ; j++) {
 276                 copy_one_pte(old_pte, new_pte);
 277                 old_pte++;
 278                 new_pte++;
 279         }
 280         return 0;
 281 }
 282 
 283 static inline int copy_one_pgd(pgd_t * old_pgd, pgd_t * new_pgd)
     /*  */
 284 {
 285         int j;
 286         pmd_t *old_pmd, *new_pmd;
 287 
 288         if (pgd_none(*old_pgd))
 289                 return 0;
 290         if (pgd_bad(*old_pgd)) {
 291                 printk("copy_one_pgd: bad page table: probable memory corruption\n");
 292                 pgd_clear(old_pgd);
 293                 return 0;
 294         }
 295         old_pmd = pmd_offset(old_pgd, 0);
 296         if (pmd_inuse(old_pmd)) {
 297                 pmd_reuse(old_pmd);
 298                 *new_pgd = *old_pgd;
 299                 return 0;
 300         }
 301         new_pmd = pmd_alloc(new_pgd, 0);
 302         if (!new_pmd)
 303                 return -ENOMEM;
 304         for (j = 0 ; j < PTRS_PER_PMD ; j++) {
 305                 int error = copy_one_pmd(old_pmd, new_pmd);
 306                 if (error)
 307                         return error;
 308                 old_pmd++;
 309                 new_pmd++;
 310         }
 311         return 0;
 312 }
 313 
 314 /*
 315  * copy_page_tables() just copies the whole process memory range:
 316  * note the special handling of RESERVED (ie kernel) pages, which
 317  * means that they are always shared by all processes.
 318  */
 319 int copy_page_tables(struct task_struct * tsk)
     /*  */
 320 {
 321         int i;
 322         pgd_t *old_pgd;
 323         pgd_t *new_pgd;
 324 
 325         new_pgd = pgd_alloc();
 326         if (!new_pgd)
 327                 return -ENOMEM;
 328         old_pgd = pgd_offset(current, 0);
 329         SET_PAGE_DIR(tsk, new_pgd);
 330         for (i = 0 ; i < PTRS_PER_PGD ; i++) {
 331                 int errno = copy_one_pgd(old_pgd, new_pgd);
 332                 if (errno) {
 333                         free_page_tables(tsk);
 334                         invalidate();
 335                         return errno;
 336                 }
 337                 old_pgd++;
 338                 new_pgd++;
 339         }
 340         invalidate();
 341         return 0;
 342 }
 343 
 344 static inline void forget_pte(pte_t page)
     /*  */
 345 {
 346         if (pte_none(page))
 347                 return;
 348         if (pte_present(page)) {
 349                 free_page(pte_page(page));
 350                 if (mem_map[MAP_NR(pte_page(page))] & MAP_PAGE_RESERVED)
 351                         return;
 352                 if (current->mm->rss <= 0)
 353                         return;
 354                 current->mm->rss--;
 355                 return;
 356         }
 357         swap_free(pte_val(page));
 358 }
 359 
 360 static inline void unmap_pte_range(pmd_t * pmd, unsigned long address, unsigned long size)
     /*  */
 361 {
 362         pte_t * pte;
 363         unsigned long end;
 364 
 365         if (pmd_none(*pmd))
 366                 return;
 367         if (pmd_bad(*pmd)) {
 368                 printk("unmap_pte_range: bad pmd (%08lx)\n", pmd_val(*pmd));
 369                 pmd_clear(pmd);
 370                 return;
 371         }
 372         pte = pte_offset(pmd, address);
 373         address &= ~PMD_MASK;
 374         end = address + size;
 375         if (end >= PMD_SIZE)
 376                 end = PMD_SIZE;
 377         do {
 378                 pte_t page = *pte;
 379                 pte_clear(pte);
 380                 forget_pte(page);
 381                 address += PAGE_SIZE;
 382                 pte++;
 383         } while (address < end);
 384 }
 385 
 386 static inline void unmap_pmd_range(pgd_t * dir, unsigned long address, unsigned long size)
     /*  */
 387 {
 388         pmd_t * pmd;
 389         unsigned long end;
 390 
 391         if (pgd_none(*dir))
 392                 return;
 393         if (pgd_bad(*dir)) {
 394                 printk("unmap_pmd_range: bad pgd (%08lx)\n", pgd_val(*dir));
 395                 pgd_clear(dir);
 396                 return;
 397         }
 398         pmd = pmd_offset(dir, address);
 399         address &= ~PGDIR_MASK;
 400         end = address + size;
 401         if (end > PGDIR_SIZE)
 402                 end = PGDIR_SIZE;
 403         do {
 404                 unmap_pte_range(pmd, address, end - address);
 405                 address = (address + PMD_SIZE) & PMD_MASK; 
 406                 pmd++;
 407         } while (address < end);
 408 }
 409 
 410 /*
 411  * a more complete version of free_page_tables which performs with page
 412  * granularity.
 413  */
 414 int unmap_page_range(unsigned long address, unsigned long size)
     /*  */
 415 {
 416         pgd_t * dir;
 417         unsigned long end = address + size;
 418 
 419         dir = pgd_offset(current, address);
 420         while (address < end) {
 421                 unmap_pmd_range(dir, address, end - address);
 422                 address = (address + PGDIR_SIZE) & PGDIR_MASK;
 423                 dir++;
 424         }
 425         invalidate();
 426         return 0;
 427 }
 428 
 429 static inline void zeromap_pte_range(pte_t * pte, unsigned long address, unsigned long size, pte_t zero_pte)
     /*  */
 430 {
 431         unsigned long end;
 432 
 433         address &= ~PMD_MASK;
 434         end = address + size;
 435         if (end > PMD_SIZE)
 436                 end = PMD_SIZE;
 437         do {
 438                 pte_t oldpage = *pte;
 439                 *pte = zero_pte;
 440                 forget_pte(oldpage);
 441                 address += PAGE_SIZE;
 442                 pte++;
 443         } while (address < end);
 444 }
 445 
 446 static inline int zeromap_pmd_range(pmd_t * pmd, unsigned long address, unsigned long size, pte_t zero_pte)
     /*  */
 447 {
 448         unsigned long end;
 449 
 450         address &= ~PGDIR_MASK;
 451         end = address + size;
 452         if (end > PGDIR_SIZE)
 453                 end = PGDIR_SIZE;
 454         do {
 455                 pte_t * pte = pte_alloc(pmd, address);
 456                 if (!pte)
 457                         return -ENOMEM;
 458                 zeromap_pte_range(pte, address, end - address, zero_pte);
 459                 address = (address + PMD_SIZE) & PMD_MASK;
 460                 pmd++;
 461         } while (address < end);
 462         return 0;
 463 }
 464 
 465 int zeromap_page_range(unsigned long address, unsigned long size, pgprot_t prot)
     /*  */
 466 {
 467         int error = 0;
 468         pgd_t * dir;
 469         unsigned long end = address + size;
 470         pte_t zero_pte;
 471 
 472         zero_pte = pte_wrprotect(mk_pte(ZERO_PAGE, prot));
 473         dir = pgd_offset(current, address);
 474         while (address < end) {
 475                 pmd_t *pmd = pmd_alloc(dir, address);
 476                 error = -ENOMEM;
 477                 if (!pmd)
 478                         break;
 479                 error = zeromap_pmd_range(pmd, address, end - address, zero_pte);
 480                 if (error)
 481                         break;
 482                 address = (address + PGDIR_SIZE) & PGDIR_MASK;
 483                 dir++;
 484         }
 485         invalidate();
 486         return error;
 487 }
 488 
 489 /*
 490  * maps a range of physical memory into the requested pages. the old
 491  * mappings are removed. any references to nonexistent pages results
 492  * in null mappings (currently treated as "copy-on-access")
 493  */
 494 static inline void remap_pte_range(pte_t * pte, unsigned long address, unsigned long size,
     /*  */
 495         unsigned long offset, pgprot_t prot)
 496 {
 497         unsigned long end;
 498 
 499         address &= ~PMD_MASK;
 500         end = address + size;
 501         if (end > PMD_SIZE)
 502                 end = PMD_SIZE;
 503         do {
 504                 pte_t oldpage = *pte;
 505                 *pte = mk_pte(offset, prot);
 506                 forget_pte(oldpage);
 507                 address += PAGE_SIZE;
 508                 offset += PAGE_SIZE;
 509                 pte++;
 510         } while (address < end);
 511 }
 512 
 513 static inline int remap_pmd_range(pmd_t * pmd, unsigned long address, unsigned long size,
     /*  */
 514         unsigned long offset, pgprot_t prot)
 515 {
 516         unsigned long end;
 517 
 518         address &= ~PGDIR_MASK;
 519         end = address + size;
 520         if (end > PGDIR_SIZE)
 521                 end = PGDIR_SIZE;
 522         offset -= address;
 523         do {
 524                 pte_t * pte = pte_alloc(pmd, address);
 525                 if (!pte)
 526                         return -ENOMEM;
 527                 remap_pte_range(pte, address, end - address, address + offset, prot);
 528                 address = (address + PMD_SIZE) & PMD_MASK;
 529                 pmd++;
 530         } while (address < end);
 531         return 0;
 532 }
 533 
 534 int remap_page_range(unsigned long from, unsigned long offset, unsigned long size, pgprot_t prot)
     /*  */
 535 {
 536         int error = 0;
 537         pgd_t * dir;
 538         unsigned long end = from + size;
 539 
 540         offset -= from;
 541         dir = pgd_offset(current, from);
 542         while (from < end) {
 543                 pmd_t *pmd = pmd_alloc(dir, from);
 544                 error = -ENOMEM;
 545                 if (!pmd)
 546                         break;
 547                 error = remap_pmd_range(pmd, from, end - from, offset + from, prot);
 548                 if (error)
 549                         break;
 550                 from = (from + PGDIR_SIZE) & PGDIR_MASK;
 551                 dir++;
 552         }
 553         invalidate();
 554         return error;
 555 }
 556 
 557 /*
 558  * sanity-check function..
 559  */
 560 static void put_page(pte_t * page_table, pte_t pte)
     /*  */
 561 {
 562         if (!pte_none(*page_table)) {
 563                 printk("put_page: page already exists\n");
 564                 free_page(pte_page(pte));
 565                 return;
 566         }
 567 /* no need for invalidate */
 568         *page_table = pte;
 569 }
 570 
 571 /*
 572  * This routine is used to map in a page into an address space: needed by
 573  * execve() for the initial stack and environment pages.
 574  */
 575 unsigned long put_dirty_page(struct task_struct * tsk, unsigned long page, unsigned long address)
     /*  */
 576 {
 577         pgd_t * pgd;
 578         pmd_t * pmd;
 579         pte_t * pte;
 580 
 581         if (page >= high_memory)
 582                 printk("put_dirty_page: trying to put page %08lx at %08lx\n",page,address);
 583         if (mem_map[MAP_NR(page)] != 1)
 584                 printk("mem_map disagrees with %08lx at %08lx\n",page,address);
 585         pgd = pgd_offset(tsk,address);
 586         pmd = pmd_alloc(pgd, address);
 587         if (!pmd)
 588                 return 0;
 589         pte = pte_alloc(pmd, address);
 590         if (!pte)
 591                 return 0;
 592         if (!pte_none(*pte)) {
 593                 printk("put_dirty_page: page already exists\n");
 594                 pte_clear(pte);
 595                 invalidate();
 596         }
 597         *pte = pte_mkwrite(pte_mkdirty(mk_pte(page, PAGE_COPY)));
 598 /* no need for invalidate */
 599         return page;
 600 }
 601 
 602 /*
 603  * This routine handles present pages, when users try to write
 604  * to a shared page. It is done by copying the page to a new address
 605  * and decrementing the shared-page counter for the old page.
 606  *
 607  * Goto-purists beware: the only reason for goto's here is that it results
 608  * in better assembly code.. The "default" path will see no jumps at all.
 609  *
 610  * Note that this routine assumes that the protection checks have been
 611  * done by the caller (the low-level page fault routine in most cases).
 612  * Thus we can safely just mark it writable once we've done any necessary
 613  * COW.
 614  *
 615  * We also mark the page dirty at this point even though the page will
 616  * change only once the write actually happens. This avoids a few races,
 617  * and potentially makes it more efficient.
 618  */
 619 void do_wp_page(struct vm_area_struct * vma, unsigned long address,
     /*  */
 620         int write_access)
 621 {
 622         pgd_t *page_dir;
 623         pmd_t *page_middle;
 624         pte_t *page_table, pte;
 625         unsigned long old_page, new_page;
 626 
 627         new_page = __get_free_page(GFP_KERNEL);
 628         page_dir = pgd_offset(vma->vm_task,address);
 629         if (pgd_none(*page_dir))
 630                 goto end_wp_page;
 631         if (pgd_bad(*page_dir))
 632                 goto bad_wp_pagedir;
 633         page_middle = pmd_offset(page_dir, address);
 634         if (pmd_none(*page_middle))
 635                 goto end_wp_page;
 636         if (pmd_bad(*page_middle))
 637                 goto bad_wp_pagemiddle;
 638         page_table = pte_offset(page_middle, address);
 639         pte = *page_table;
 640         if (!pte_present(pte))
 641                 goto end_wp_page;
 642         if (pte_write(pte))
 643                 goto end_wp_page;
 644         old_page = pte_page(pte);
 645         if (old_page >= high_memory)
 646                 goto bad_wp_page;
 647         vma->vm_task->mm->min_flt++;
 648         /*
 649          * Do we need to copy?
 650          */
 651         if (mem_map[MAP_NR(old_page)] != 1) {
 652                 if (new_page) {
 653                         if (mem_map[MAP_NR(old_page)] & MAP_PAGE_RESERVED)
 654                                 ++vma->vm_task->mm->rss;
 655                         copy_page(old_page,new_page);
 656                         *page_table = pte_mkwrite(pte_mkdirty(mk_pte(new_page, vma->vm_page_prot)));
 657                         free_page(old_page);
 658                         invalidate();
 659                         return;
 660                 }
 661                 free_page(old_page);
 662                 oom(vma->vm_task);
 663                 *page_table = BAD_PAGE;
 664                 invalidate();
 665                 return;
 666         }
 667         *page_table = pte_mkdirty(pte_mkwrite(pte));
 668         invalidate();
 669         if (new_page)
 670                 free_page(new_page);
 671         return;
 672 bad_wp_page:
 673         printk("do_wp_page: bogus page at address %08lx (%08lx)\n",address,old_page);
 674         send_sig(SIGKILL, vma->vm_task, 1);
 675         goto end_wp_page;
 676 bad_wp_pagemiddle:
 677         printk("do_wp_page: bogus page-middle at address %08lx (%08lx)\n", address, pmd_val(*page_middle));
 678         send_sig(SIGKILL, vma->vm_task, 1);
 679         goto end_wp_page;
 680 bad_wp_pagedir:
 681         printk("do_wp_page: bogus page-dir entry at address %08lx (%08lx)\n", address, pgd_val(*page_dir));
 682         send_sig(SIGKILL, vma->vm_task, 1);
 683 end_wp_page:
 684         if (new_page)
 685                 free_page(new_page);
 686         return;
 687 }
 688 
 689 /*
 690  * Ugly, ugly, but the goto's result in better assembly..
 691  */
 692 int verify_area(int type, const void * addr, unsigned long size)
     /*  */
 693 {
 694         struct vm_area_struct * vma;
 695         unsigned long start = (unsigned long) addr;
 696 
 697         /* If the current user space is mapped to kernel space (for the
 698          * case where we use a fake user buffer with get_fs/set_fs()) we
 699          * don't expect to find the address in the user vm map.
 700          */
 701         if (get_fs() == get_ds())
 702                 return 0;
 703 
 704         vma = find_vma(current, start);
 705         if (!vma)
 706                 goto bad_area;
 707         if (vma->vm_start <= start)
 708                 goto good_area;
 709         if (!(vma->vm_flags & VM_GROWSDOWN))
 710                 goto bad_area;
 711         if (vma->vm_end - start > current->rlim[RLIMIT_STACK].rlim_cur)
 712                 goto bad_area;
 713 
 714 good_area:
 715         if (type == VERIFY_WRITE)
 716                 goto check_write;
 717         for (;;) {
 718                 struct vm_area_struct * next;
 719                 if (!(vma->vm_flags & VM_READ))
 720                         goto bad_area;
 721                 if (vma->vm_end - start >= size)
 722                         return 0;
 723                 next = vma->vm_next;
 724                 if (!next || vma->vm_end != next->vm_start)
 725                         goto bad_area;
 726                 vma = next;
 727         }
 728 
 729 check_write:
 730         if (!(vma->vm_flags & VM_WRITE))
 731                 goto bad_area;
 732         if (!wp_works_ok)
 733                 goto check_wp_fault_by_hand;
 734         for (;;) {
 735                 if (vma->vm_end - start >= size)
 736                         break;
 737                 if (!vma->vm_next || vma->vm_end != vma->vm_next->vm_start)
 738                         goto bad_area;
 739                 vma = vma->vm_next;
 740                 if (!(vma->vm_flags & VM_WRITE))
 741                         goto bad_area;
 742         }
 743         return 0;
 744 
 745 check_wp_fault_by_hand:
 746         size--;
 747         size += start & ~PAGE_MASK;
 748         size >>= PAGE_SHIFT;
 749         start &= PAGE_MASK;
 750 
 751         for (;;) {
 752                 do_wp_page(vma, start, 1);
 753                 if (!size)
 754                         break;
 755                 size--;
 756                 start += PAGE_SIZE;
 757                 if (start < vma->vm_end)
 758                         continue;
 759                 vma = vma->vm_next;
 760                 if (!vma || vma->vm_start != start)
 761                         goto bad_area;
 762                 if (!(vma->vm_flags & VM_WRITE))
 763                         goto bad_area;;
 764         }
 765         return 0;
 766 
 767 bad_area:
 768         return -EFAULT;
 769 }
 770 
 771 static inline void get_empty_page(struct vm_area_struct * vma, pte_t * page_table)
     /*  */
 772 {
 773         unsigned long tmp;
 774 
 775         if (!(tmp = get_free_page(GFP_KERNEL))) {
 776                 oom(vma->vm_task);
 777                 put_page(page_table, BAD_PAGE);
 778                 return;
 779         }
 780         put_page(page_table, pte_mkwrite(mk_pte(tmp, vma->vm_page_prot)));
 781 }
 782 
 783 /*
 784  * try_to_share() checks the page at address "address" in the task "p",
 785  * to see if it exists, and if it is clean. If so, share it with the current
 786  * task.
 787  *
 788  * NOTE! This assumes we have checked that p != current, and that they
 789  * share the same inode and can generally otherwise be shared.
 790  */
 791 static int try_to_share(unsigned long to_address, struct vm_area_struct * to_area,
     /*  */
 792         unsigned long from_address, struct vm_area_struct * from_area,
 793         unsigned long newpage)
 794 {
 795         pgd_t * from_dir, * to_dir;
 796         pmd_t * from_middle, * to_middle;
 797         pte_t * from_table, * to_table;
 798         pte_t from, to;
 799 
 800         from_dir = pgd_offset(from_area->vm_task,from_address);
 801 /* is there a page-directory at from? */
 802         if (pgd_none(*from_dir))
 803                 return 0;
 804         if (pgd_bad(*from_dir)) {
 805                 printk("try_to_share: bad page directory %08lx\n", pgd_val(*from_dir));
 806                 return 0;
 807         }
 808         from_middle = pmd_offset(from_dir, from_address);
 809 /* is there a mid-directory at from? */
 810         if (pmd_none(*from_middle))
 811                 return 0;
 812         if (pmd_bad(*from_middle)) {
 813                 printk("try_to_share: bad mid directory %08lx\n", pmd_val(*from_middle));
 814                 return 0;
 815         }
 816         from_table = pte_offset(from_middle, from_address);
 817         from = *from_table;
 818 /* is the page present? */
 819         if (!pte_present(from))
 820                 return 0;
 821 /* if it is dirty it must be from a shared mapping to be shared */
 822         if (pte_dirty(from)) {
 823                 if (!(from_area->vm_flags & VM_SHARED))
 824                         return 0;
 825                 if (pte_write(from)) {
 826                         printk("nonwritable, but dirty, shared page\n");
 827                         return 0;
 828                 }
 829         }
 830 /* is the page reasonable at all? */
 831         if (pte_page(from) >= high_memory)
 832                 return 0;
 833         if (mem_map[MAP_NR(pte_page(from))] & MAP_PAGE_RESERVED)
 834                 return 0;
 835 /* is the destination ok? */
 836         to_dir = pgd_offset(to_area->vm_task,to_address);
 837 /* is there a page-directory at to? */
 838         if (pgd_none(*to_dir))
 839                 return 0;
 840         if (pgd_bad(*to_dir)) {
 841                 printk("try_to_share: bad page directory %08lx\n", pgd_val(*to_dir));
 842                 return 0;
 843         }
 844         to_middle = pmd_offset(to_dir, to_address);
 845 /* is there a mid-directory at to? */
 846         if (pmd_none(*to_middle))
 847                 return 0;
 848         if (pmd_bad(*to_middle)) {
 849                 printk("try_to_share: bad mid directory %08lx\n", pmd_val(*to_middle));
 850                 return 0;
 851         }
 852         to_table = pte_offset(to_middle, to_address);
 853         to = *to_table;
 854         if (!pte_none(to))
 855                 return 0;
 856 /* do we copy? */
 857         if (newpage) {
 858                 /* if it's in the swap cache, it's dirty by implication */
 859                 /* so we can't use it if it's not from a shared mapping */
 860                 if (in_swap_cache(pte_page(from))) {
 861                         if (!(from_area->vm_flags & VM_SHARED))
 862                                 return 0;
 863                         if (!pte_write(from)) {
 864                                 printk("nonwritable, but dirty, shared page\n");
 865                                 return 0;
 866                         }
 867                 }
 868                 copy_page(pte_page(from), newpage);
 869                 *to_table = mk_pte(newpage, to_area->vm_page_prot);
 870                 return 1;
 871         }
 872 /*
 873  * do a final swap-cache test before sharing them: if it's in the swap
 874  * cache, we have to remove it now, as we get two pointers to the same
 875  * physical page and the cache can't handle it. Mark the original dirty.
 876  *
 877  * NOTE! Even if "from" is dirty, "to" will be clean: if we get here
 878  * with a dirty "from", the from-mapping is a shared map, so we can trust
 879  * the page contents to be up-to-date
 880  */
 881         if (in_swap_cache(pte_page(from))) {
 882                 if (!(from_area->vm_flags & VM_SHARED))
 883                         return 0;
 884                 *from_table = pte_mkdirty(from);
 885                 delete_from_swap_cache(pte_page(from));
 886         }
 887         mem_map[MAP_NR(pte_page(from))]++;
 888         *to_table = mk_pte(pte_page(from), to_area->vm_page_prot);
 889 /* Check if we need to do anything at all to the 'from' field */
 890         if (!pte_write(from))
 891                 return 1;
 892         if (from_area->vm_flags & VM_SHARED)
 893                 return 1;
 894 /* ok, need to mark it read-only, so invalidate any possible old TB entry */
 895         *from_table = pte_wrprotect(from);
 896         invalidate();
 897         return 1;
 898 }
 899 
 900 /*
 901  * share_page() tries to find a process that could share a page with
 902  * the current one.
 903  *
 904  * We first check if it is at all feasible by checking inode->i_count.
 905  * It should be >1 if there are other tasks sharing this inode.
 906  */
 907 static int share_page(struct vm_area_struct * area, unsigned long address,
     /*  */
 908         int write_access, unsigned long newpage)
 909 {
 910         struct inode * inode;
 911         unsigned long offset;
 912         unsigned long from_address;
 913         unsigned long give_page;
 914         struct vm_area_struct * mpnt;
 915 
 916         if (!area || !(inode = area->vm_inode) || inode->i_count < 2)
 917                 return 0;
 918         /* do we need to copy or can we just share? */
 919         give_page = 0;
 920         if (write_access && !(area->vm_flags & VM_SHARED)) {
 921                 if (!newpage)
 922                         return 0;
 923                 give_page = newpage;
 924         }
 925         offset = address - area->vm_start + area->vm_offset;
 926         /* See if there is something in the VM we can share pages with. */
 927         /* Traverse the entire circular i_mmap list, except `area' itself. */
 928         for (mpnt = area->vm_next_share; mpnt != area; mpnt = mpnt->vm_next_share) {
 929                 /* must be same inode */
 930                 if (mpnt->vm_inode != inode) {
 931                         printk("Aiee! Corrupt vm_area_struct i_mmap ring\n");
 932                         break;  
 933                 }
 934                 /* offsets must be mutually page-aligned */
 935                 if ((mpnt->vm_offset ^ area->vm_offset) & ~PAGE_MASK)
 936                         continue;
 937                 /* the other area must actually cover the wanted page.. */
 938                 from_address = offset + mpnt->vm_start - mpnt->vm_offset;
 939                 if (from_address < mpnt->vm_start || from_address >= mpnt->vm_end)
 940                         continue;
 941                 /* .. NOW we can actually try to use the same physical page */
 942                 if (!try_to_share(address, area, from_address, mpnt, give_page))
 943                         continue;
 944                 /* free newpage if we never used it.. */
 945                 if (give_page || !newpage)
 946                         return 1;
 947                 free_page(newpage);
 948                 return 1;
 949         }
 950         return 0;
 951 }
 952 
 953 /*
 954  * fill in an empty page-table if none exists.
 955  */
 956 static inline pte_t * get_empty_pgtable(struct task_struct * tsk,unsigned long address)
     /*  */
 957 {
 958         pgd_t *pgd;
 959         pmd_t *pmd;
 960         pte_t *pte;
 961 
 962         pgd = pgd_offset(tsk, address);
 963         pmd = pmd_alloc(pgd, address);
 964         if (!pmd)
 965                 return NULL;
 966         pte = pte_alloc(pmd, address);
 967         return pte;
 968 }
 969 
 970 static inline void do_swap_page(struct vm_area_struct * vma, unsigned long address,
     /*  */
 971         pte_t * page_table, pte_t entry, int write_access)
 972 {
 973         pte_t page;
 974 
 975         if (!vma->vm_ops || !vma->vm_ops->swapin) {
 976                 swap_in(vma, page_table, pte_val(entry), write_access);
 977                 return;
 978         }
 979         page = vma->vm_ops->swapin(vma, address - vma->vm_start + vma->vm_offset, pte_val(entry));
 980         if (pte_val(*page_table) != pte_val(entry)) {
 981                 free_page(pte_page(page));
 982                 return;
 983         }
 984         if (mem_map[MAP_NR(pte_page(page))] > 1 && !(vma->vm_flags & VM_SHARED))
 985                 page = pte_wrprotect(page);
 986         ++vma->vm_task->mm->rss;
 987         ++vma->vm_task->mm->maj_flt;
 988         *page_table = page;
 989         return;
 990 }
 991 
 992 /*
 993  * do_no_page() tries to create a new page mapping. It aggressively
 994  * tries to share with existing pages, but makes a separate copy if
 995  * the "write_access" parameter is true in order to avoid the next
 996  * page fault.
 997  */
 998 void do_no_page(struct vm_area_struct * vma, unsigned long address,
     /*  */
 999         int write_access)
1000 {
1001         pte_t * page_table;
1002         pte_t entry;
1003         unsigned long page;
1004 
1005         page_table = get_empty_pgtable(vma->vm_task,address);
1006         if (!page_table)
1007                 return;
1008         entry = *page_table;
1009         if (pte_present(entry))
1010                 return;
1011         if (!pte_none(entry)) {
1012                 do_swap_page(vma, address, page_table, entry, write_access);
1013                 return;
1014         }
1015         address &= PAGE_MASK;
1016         if (!vma->vm_ops || !vma->vm_ops->nopage) {
1017                 ++vma->vm_task->mm->rss;
1018                 ++vma->vm_task->mm->min_flt;
1019                 get_empty_page(vma, page_table);
1020                 return;
1021         }
1022         page = get_free_page(GFP_KERNEL);
1023         if (share_page(vma, address, write_access, page)) {
1024                 ++vma->vm_task->mm->min_flt;
1025                 ++vma->vm_task->mm->rss;
1026                 return;
1027         }
1028         if (!page) {
1029                 oom(current);
1030                 put_page(page_table, BAD_PAGE);
1031                 return;
1032         }
1033         ++vma->vm_task->mm->maj_flt;
1034         ++vma->vm_task->mm->rss;
1035         /*
1036          * The fourth argument is "no_share", which tells the low-level code
1037          * to copy, not share the page even if sharing is possible.  It's
1038          * essentially an early COW detection 
1039          */
1040         page = vma->vm_ops->nopage(vma, address, page,
1041                 write_access && !(vma->vm_flags & VM_SHARED));
1042         if (share_page(vma, address, write_access, 0)) {
1043                 free_page(page);
1044                 return;
1045         }
1046         /*
1047          * This silly early PAGE_DIRTY setting removes a race
1048          * due to the bad i386 page protection. But it's valid
1049          * for other architectures too.
1050          *
1051          * Note that if write_access is true, we either now have
1052          * a exclusive copy of the page, or this is a shared mapping,
1053          * so we can make it writable and dirty to avoid having to
1054          * handle that later.
1055          */
1056         entry = mk_pte(page, vma->vm_page_prot);
1057         if (write_access) {
1058                 entry = pte_mkwrite(pte_mkdirty(entry));
1059         } else if (mem_map[MAP_NR(page)] > 1 && !(vma->vm_flags & VM_SHARED))
1060                 entry = pte_wrprotect(entry);
1061         put_page(page_table, entry);
1062 }
/* */
root/mm/memory.c

DEFINITIONS
